Solar Image Analysis and Visualization

J. Ireland C.A.
Young
Editors
Solar Image Analysis

and Visualization
Previously published in Solar Physics Volume 248, Issue 2, 2008
J. Ireland
ADNET Systems, Inc.
NASA Goddard Spaceflight Center
Greenbelt, MD, USA
C.A. Young
ADNET Systems, Inc.
NASA Goddard Spaceflight Center
Greenbelt, MD, USA
Cover illustration: Image of automated detection and segmentation of CME.

Courtesy of Oscar Olmedo, George Mason University.
All rights reserved.
Library of Congress Control Number: 2009922543
DOI: 10.1007/978-0-387-98154-3
ISBN-978-0-387-98153-6
e-ISBN-978-0-387-98154-3
Printed on acid-free paper.

2009 Springer Science+Business Media, BV
No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by any
means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without the written
permission from the Publisher, with the exception of any material supplied specifically for the purpose
of being entered and executed on a computer system, for the exclusive use by the purchaser of the work.
1
springer.com
Contents
Preface: A Topical Issue on Solar Image Analysis and Visualization
FESTIVAL: A Multiscale Visualization Tool for Solar Imaging Data

F. Auchre E. Soubri K. Bocchialini F. LeGall 3
Visualization of Distributed Solar Data and Metadata with the Solar Weather
Browser
B. Nicula C. Marqu D. Berghmans 15
Widespread Occurrence of Trenching Patterns in the Granulation Field:
Evidence for Roll Convection?
A.V. Getling A.A. Buchnev 23
Principal Components and Independent Component Analysis of Solar and Space
Data
A.C. Cadavid J.K. Lawrence A. Ruzmaikin 37
Automatic Recognition and Characterisation of Supergranular Cells
from Photospheric Velocity Fields
H.E. Potts D.A. Diver 53
Automated McIntosh-Based Classification of Sunspot Groups Using MDI Images
T. Colak R. Qahwaji 67
Multifractal Properties of Evolving Active Regions
P.A. Conlon P.T. Gallagher R.T.J. McAteer J. Ireland C.A. Young P. Kestener
R.J. Hewett K. Maguire 87
Multiscale Analysis of Active Region Evolution
R.J. Hewett P.T. Gallagher R.T.J. McAteer C.A. Young J. Ireland P.A. Conlon
K. Maguire 101
A Comparison of Feature Classification Methods for Modeling Solar Irradiance
Variation
H.P. Jones G.A. Chapman K.L. Harvey J.M. Pap D.G. Preminger M.J. Turmon
S.R. Walton 113
The Observed Long- and Short-Term Phase Relation between the Toroidal and
Poloidal Magnetic Fields in Cycle 23
S. Zharkov E. Gavryuseva V. Zharkova 129
Comparison of Five Numerical Codes for Automated Tracing of Coronal Loops
M.J. Aschwanden J.K. Lee G.A. Gary M. Smith B. Inhester 149
Segmentation of Loops from Coronal EUV Images

B. Inhester L. Feng T. Wiegelmann 169
The Pixelised Wavelet Filtering Method to Study Waves and Oscillations in Time
Sequences of Solar Atmospheric Images
R.A. Sych V.M. Nakariakov 185
A Time-Evolving 3D Method Dedicated to the Reconstruction of Solar Plumes and
Results Using Extreme Ultraviolet Data
N. Barbey F. Auchre T. Rodet J.-C. Vial 199
Automatic Detection and Classification of Coronal Holes and Filaments Based
on EUV and Magnetogram Observations of the Solar Disk
I.F. Scholl S.R. Habbal 215
Spatial and Temporal Noise in Solar EUV Observations
V. Delouille P. Chainais J.-F. Hochedez 231
Multiscale Edge Detection in the Corona
C.A. Young P.T. Gallagher 247
Automated Prediction of CMEs Using Machine Learning of CME Flare
Associations
R. Qahwaji T. Colak M. Al-Omari S. Ipson 261
Automatic Detection and Tracking of Coronal Mass Ejections in Coronagraph Time
Series
O. Olmedo J. Zhang H. Wechsler A. Poland K. Borne 275
Preface: A Topical Issue on Solar Image Analysis

and Visualization
Originally published in the journal Solar Physics, Volume 248, No 2, 211211.

DOI: 10.1007/s11207-008-9168-x Springer Science+Business Media B.V. 2008
The third Solar Image Processing Workshop (SIPWork III) was held at Trinity College
Dublin, Ireland, in September 2006. This meeting brought together researchers in solar
physics, image processing, and computer vision, and it focused on preparing for the data
analysis and processing needs of new space missions such as the Solar TErrestrial RElations Observatory (STEREO), Hinode, and the Solar Dynamics Observatory (SDO), as
well as ground-based instrumentation such as the Advanced Technology Solar Telescope
(ATST), the Swedish Solar Telescope (SST), and the Dutch Open Telescope (DOT).
As with SIPWork II (Solar Phys. 228, 2005) we have gathered papers from work
presented at SIPWork III as well as related manuscripts in this topical issue of Solar
Physics Solar Image Analysis and Visualization that the papers mutually benefit from
appearing together. The result is 19 papers that apply a range of image processing and computer vision techniques to address the scientific goals of new solar physics instrumentation.
The papers are organized into four groups by types of structures in different levels of the
solar atmosphere and visualization tools: i) software tools for the access and visualization
of solar data; ii) granulation and active-region magnetic fields; iii) coronal loops in the solar
atmosphere; and iv) coronal mass ejections.
The organizers thank NASAs Heliophysics Division, the European Office of Aerospace
Research and Development (EOARD), and Trinity College Dublin for their support of this
meeting. Finally, we thank all those who participated in SIPWork III for making it a very
stimulating and successful meeting, and all of the authors and referees who helped make
this topical issue possible.
Guest Editors: C. Alex Young and Jack Ireland, ADNET Systems, Inc., Greenbelt,
Maryland, USA
Editor: John Leibacher, National Solar Observatory, Tucson, Arizona, USA
C.A. Young and J. Ireland (eds.), Solar Image Analysis and Visualization.
DOI: 10.1007/978-0-387-98154-3_1
FESTIVAL: A Multiscale Visualization Tool for Solar

Imaging Data
F. Auchre E. Soubri K. Bocchialini F. LeGall

DOI: 10.1007/s11207-008-9163-2 Springer Science+Business Media B.V. 2008
Abstract Since 4 December 2006, the SECCHI instrument suites onboard the two STEREO
A and B probes have been imaging the solar corona and the heliosphere on a wide range of
angular scales. The EUVI telescopes have a plate scale of 1.7 arcseconds pixel1 , while
that of the HI2 wide-angle cameras is 2.15 arcminutes pixel1 , i.e. 75 times larger, with the
COR1 and COR2 coronagraphs having intermediate plate scales. These very different instruments, aimed at studying Coronal Mass Ejections and their propagation in the heliosphere,
create a data visualization challenge. This paper presents FESTIVAL, a SolarSoftware package originally developed to be able to map the SECCHI data into dynamic composite images
of the sky as seen by the STEREO and SOHO probes. Data from other imaging instruments
can also be displayed. Using the mouse, the user can quickly and easily zoom in and out and
pan through these composite images to explore all spatial scales from EUVI to HI2 while
keeping the native resolution of the original data. A large variety of numerical filters can
be applied, and additional data (i.e. coordinate grids, stars catalogs, etc.) can be overlaid on
the images. The architecture of FESTIVAL is such that it is easy to add support for other
instruments and these new data immediately benefit from the already existing capabilities.
Also, because its mapping engine is fully 3D, FESTIVAL provides a convenient environment to display images from future out-of-the-Ecliptic solar missions, such as Solar Orbiter
or Solar Probe.
Keywords Data analysis Instrumentation
F. Auchre () E. Soubri K. Bocchialini
Institut dAstrophysique Spatiale, Btiment 121, Universit Paris-Sud, 91405 Orsay, France
e-mail: frederic.auchere@ias.u-psud.fr
E. Soubri
e-mail: elie.soubrie@ias.u-psud.fr
K. Bocchialini
e-mail: karine.bocchialini@ias.u-psud.fr
F. LeGall
ITT France, Paris, France
e-mail: franck.legall@ittvis.com
DOI: 10.1007/978-0-387-98154-3_2
F. Auchre et al.
1. Introduction
The SECCHI A and B instrument suites (Howard et al., 2006) onboard the two STEREO
mission spacecraft (Kaiser, 2005) are each composed of: one Extreme Ultra-Violet Imager
(EUVI), two white-light coronagraphs (COR1 and COR2), and two wide-angle heliospheric
imagers (HI1 and HI2). Technical descriptions of EUVI, COR1 and the HIs can be found
in Wuelser et al. (2004), Thompson et al. (2003), and Defise et al. (2003), respectively.
The images produced by SECCHI represent a data visualization challenge: i) the images are
2048 2048 pixels (except for the HIs, which are usually binned onboard 2 2), thus the
vast majority of computer displays are not able to display them at full frame and full resolution, and ii) more importantly, the five instruments of SECCHI A and B were designed
to be able to track Coronal Mass Ejections from their onset (with EUVI) to their propagation in the heliosphere (with the HIs), which implies that a set of SECCHI images that
covers the propagation of a CME from its initiation site to the Earth is composed of images with very different spatial resolutions from 1.7 arcseconds pixel1 for EUVI to 2.15
arcminutes pixel1 for HI2, i.e. 75 times larger. A similar situation exists with the angular
scales of the physical objects, since the size of a CME varies by orders of magnitude as it
expands in the heliosphere. This makes it difficult to have an undistorted view of the whole
phenomenon using a fixed spatial resolution and a fixed field of view. The challenge of a
visualization software for SECCHI data is to: i) generate composite images from all five
instruments from each SECCHI suite, and ii) allow the user to easily zoom and pan across
the resulting large composite. In addition, the software has to be able to manage the various
cadences of the five instruments (from 2.5 minutes or higher for EUVI to two hours for HI2).
FESTIVAL (http://www.ias.u-psud.fr/stereo/festival/), an open-source browser written in
IDL, distributed as a SolarSoftware (SSW) package, was originally designed for simultaneous, fast, and easy manipulation of SECCHI data from STEREO, and of EIT (Delaboudinire et al., 1995) and LASCO (Brueckner et al., 1995) data from SOHO (Domingo,
Fleck, and Poland, 1995). FESTIVAL works with FITS files installed locally on the users
computer or on remote disks mounted via NFS. It automatically builds dynamic composite
images of the sky as seen by the STEREO A, B, and SOHO spacecraft by interpreting the
information stored in the FITS headers (date, plate scale, roll angle, pointing). The composites preserve the native resolution of all of the instruments and the user can zoom in and out
to explore the full range of angular scales covered by SECCHI and EIT-LASCO: from the
arcsecond (EUVI/EIT) to tens of degrees (HI1/HI2).
2. FESTIVAL Basics
2.1. Layers
The traditional way to create composite images from multiple-instrument data is to build
a large matrix, which greatly limits the range of angular scales that can be covered simply
because of the amount of computer memory required. For example, a bitmap composite image covering the fields of view from EUVI to HI2 with the resolution of EUVI would be
190 000 150 000 pixels, which represents 26 GB at eight bits. Using the mapping capabilities of modern graphics cards, FESTIVAL does not need to build such a huge matrix.
Instead, FESTIVAL composite images are made of superimposed layers. One layer is associated with each instrument: one for EUVI, one for COR1, one for COR2, etc. FESTIVAL
sends the images to the graphics card along with the pointing and scaling information read
in the FITS headers. The graphics hardware then takes care of the compositing. The full
SECCHI composite image described above therefore requires only 14 MB of RAM instead
of 26 GB. A layer can contain either a 2D textured polygon or 2D vector data. Textured polygons are used to display images, while vectors are used to display context information on the
images. At present, the available context information sources are: Carrington, Stonyhurst,
polar, and sky coordinates grids, stars from the Hipparcos catalogue (see Section 2.4), and
the position of planets. The display order of the layers associated with the instruments can be
modified. EUVI and COR1, for example, have a region of overlap. By changing the order of
the layers, one can choose to visualize EUVI or COR1 data in this region. The information
layers are always on top of the instrument layers. The visibility of the layers can be switched
on and off. Note that EUVI and EIT are imaging EUV emission lines, while LASCO, the
CORs, and the HIs take broadband visible-light images. A composite can therefore contain
information from physically different emission processes and this has to be kept in mind
when interpreting the images.
2.2. Systems of Coordinates and Projections
Considering the very large total field of view covered by the SECCHI telescopes (up to 90
from the Sun to the outer edge of HI2), the composites are not mere images of the solar
corona they can cover large portions of the sky. So whether looking at a small portion of
the solar disk or at a large-angle composite of HI images, FESTIVAL is always building a
map of a full hemisphere of the sky as seen from either STEREO A, B, or SOHO. The data
from different probes cannot be overlaid. For example, STEREO and SOHO data cannot
be overlaid because these data are not taken from the same vantage point. FESTIVAL uses
two possible frames of reference called Heliocentric Ecliptic North Up (HENU) and Heliocentric Solar North Up (HSNU). The z-axis is Sun centered, the x-axis is pointing toward
either ecliptic north or solar north, and the y-axis is completing a right-handed orthogonal
coordinate system.
The type of projection used to map the sky on the computer display can be chosen by
the user and applies to all the instruments of a given probe. Projection types can be different
for each probe. All the projection types that allow the mapping of a full hemisphere (e.g.,
Equirectangular, Mercator, Aitoff, etc.) and that are supported by IDL are available.
2.3. Data Handling
An image undergoes the following operations before it is displayed:
1. Preparation: call to the appropriate SSW routine, e.g., EIT_PREP, SECCHI_PREP. This
includes reading the corresponding FITS file.
2. Rotation: rotation to solar north up or ecliptic north up, depending upon the choices made
in the projection options.
3. Filtering: apply the user-defined sequence of filters, if any.
4. Difference: subtraction of the previous image of the same instrument if in runningdifference mode. If the previous image is not already present in memory, it is computed
by going through steps 1 3.
5. Enhancement: conversion from 16 bits to 8 bits according to the parameters defined in
the enhancement options.
6. Projection: mapping of the data on the sky according to the settings defined in the projection options.
7. Storing in the history stack.
F. Auchre et al.
Figure 1 Illustration of the data

flow in FESTIVAL.
Figure 1 illustrates this sequence. After an image is displayed, FESTIVAL stores in the
history stack the 32-bit filtered image, the 8-bit projected image and detailed information
(filter sequence, calibration options, projection parameters, etc.) about its processing through
steps 1 7. The history stack is a first in, first out (FIFO). Any modification by the user
(except panning and zooming) of the rendered scene leads to a request for new data. This
includes modifying the rotation, filter sequence, enhancement parameters, projection type,
etc. When new data are requested, FESTIVAL looks up the history stack for previously
processed data that have the right properties: roll, filter sequence, enhancement parameters,
projection type, etc. If the data are found, the projected image is displayed instantaneously. If
not, FESTIVAL will find the data closest to that requested. For example, if the user displays
an image and then changes the projection type, data that are correctly calibrated and filtered
are already present in memory. In this case, FESTIVAL does not process the data from
scratch (step 1), but uses the existing information and proceeds directly from step 4; the
resulting new data are pushed onto the history stack. Its default size is 400 MB, but with
more RAM more data can be stored in memory and the processing will execute more rapidly.
2.4. Coalignment Calibration
Stars were used to investigate the accuracy of FESTIVALs mapping engine and the coalignment of the coronagraphs onboard SOHO and STEREO.

Table 1 Pitch, yaw, and roll
corrections for HI images derived
from visual matching of stars
field in FESTIVAL.
Pitch
7
Yaw
Roll
(degrees)
HI1A
0.03
0.025
0.05
HI2A
0.20
0.11
0.09
HI1B
0.05
0.06
0.0
HI2B
0.25
0.05
0.2
FESTIVAL provides a reference frame that can be used to calibrate the accuracy of the
absolute positioning of the images on the sky. The software can therefore be used as a tool
for investigating the alignment and/or coalignment of instruments. The projected views of
the sky can contain both imaging data (e.g. SECCHI images) and vectorial overlays (e.g.
lines, polygons). A vectorial overlay is used to display a set of polygons representing stars.
Knowing the position of the stars, the orientation of the ecliptic plane, the orientation of
the solar equator, and the position of the spacecraft in a common system of coordinates,
it is possible to compute the position of the stars in the two frames of reference used by
FESTIVAL and to display them on top of SECCHI or EIT-LASCO data for comparison.
We have done this for LASCO C2 and C3, and for SECCHI COR2, HI1 and HI2 (A and B
spacecraft).
For the position of stars, we used the Hipparcos catalogue (Perryman et al., 1997). It contains 118 218 entries, is complete down to visual magnitude 7 (15 405 stars), and includes
stars down to magnitude 14. The position of stars are given in the International Celestial
Reference System (ICRS) for epoch J1991.25. The axes of the ICRS are consistent within
0.1 arcsecond with the J2000.0 equatorial celestial coordinates. For our application, we do
not distinguish between the two systems. The proper motion of stars in right ascension and
declination is used to compute their position at epoch J2000.0. The position of the spacecraft (SOHO or STEREO) is known in the J2000.0 Heliocentric Aries Ecliptic system. For
STEREO, this information is listed in the FITS headers, while for SOHO it is read in the orbit files via the GET_ORBIT IDL function. In J2000.0, the solar-rotation axis has equatorial
celestial coordinates (R.A. = + 286.11 , DEC = + 63.85 ) and the inclination of the ecliptic is 23.4392911 . With these parameters, the positions of the stars are computed in one of
the two possible frames of reference (HENU or HSNU). Note that annual aberration is not
taken into account. However, this effect is small (30 arcseconds at most for an observer on
the Earth), and since the frames of reference of FESTIVAL are Sun centered, the aberration
of the Sun should be subtracted, resulting in a minute correction.
We present results for the comparison between the theoretical and observed position
of stars. We used stars down to visual magnitude 15. The alignment of LASCO data was
visually checked for 20 dates between December 2005 and January 2007. Figure 2 shows
an example of a LASCO C2/C3 composite image for 21 November 2006 with stars overlaid
(circles). The diameter of the stars is proportional to their magnitude. The top panel is an
enlargement of the C2 field of view, while the bottom panel shows the whole composite.
Overall, the match between computed and observed stars was found to be on average within
one pixel in both C2 and C3 (i.e. 11.9 arcseconds and 56.0 arcseconds, respectively). Note
that FESTIVAL is also able to display planets and their satellites. The top right inset shows
Callisto correctly positioned at the East of Jupiter. The other satellites are masked by the
CCD blooming. Venus is also circled in the C3 field of view (bottom panel).
We performed similar tests on SECCHI HI1 and HI2 data taken between 15 March 2007
and 15 August 2007. We have not yet performed these tests on COR1 and COR2. However,
F. Auchre et al.
Figure 2 Star field and planets (circles) overlaid on a LASCO C2/C3 composite image for 21 November
2006. The observed stars are, on average, within one pixel of their computed positions.
we verified that sharp coronal structures, such as streamer edges or large coronal loops, are
well connected from EUVI to COR1 to COR2, to within the resolution of the instruments.
According to the results obtained with LASCO data, we could expect FESTIVAL to position
Figure 3 Left panels: HI1A (top) and HI2B (bottom) images with overlaid star fields (circles). Only a
17 17 degree position of the HI2 field of view is shown. The discrepancies are about one arcminute in HI1
and 15 arcminutes in HI2. Right panels: same images after correction using the angles given in Table 1.
SECCHI data to within a pixel of the theoretical star positions. The star fields were found
to be off by about 1 arcminute in HI1A and HI1B, and by about 15 arcminutes in HI2A
and HI2B. The top left and bottom left panels of Figure 3 illustrate the observed discrepancies for HI1A and HI2A, respectively. Only a fraction (17 17 degrees) of the HI2A field
of view is shown. The stars are displayed as circles. From such images, we estimated the
corrections in pitch, yaw, and roll to apply to the images in order to visually match the star
fields. Results are listed in Table 1. We estimate the accuracy of this manual procedure to be
about 0.01 degrees. The HI1B values should be used with caution because the pointing of
HI1B is known to be unstable; its boresight being erratically shifted by a still-unidentified
mechanism (one hypothesis being interplanetary dust impacts). The top right and bottom
right panels of Figure 1 show the images of the left panels after pointing correction. Future
updates of the HI headers with measured pointing information are expected to reduce the
observed discrepancies.
10
F. Auchre et al.
Figure 4 The selection GUI. Each of the three tabs is associated with a probe: STEREO A, STEREO B, and
SOHO. In each tab, the search criteria are entered in the left area, and on the right is a table that summarizes
the result of the search.
2.5. The Interface

The Graphical User Interfaces (GUI) vary slightly from one operating system/window manager to another. The screenshots shown in this paper were taken running Windows XP and
IDL 6.2. FESTIVAL presents two main kinds of GUIs to the user:
1. The selection GUI, where the user makes selective searches for the data of interest
(Figure 4).
2. The visualization GUIs, where the selected data are visualized. Up to three can be opened
simultaneously, one per probe (Figure 5).
2.6. Navigation in Time
The user can move back and forth in time either with the back and next buttons in the
visualization GUI or by clicking on an entry of the table in the selection GUI. The effect
of this action will depend on which of the two navigation modes (normal or compact)
is selected. In normal mode, any image can be composited with any other, while in compact mode, the images are always forced to be as close together in time as possible with
each other. Selecting one image in the selection GUI table (or hitting the next button)
will not only bring up the selected (or the next) data, but also the images from the other
selected instruments that are closest in time to it. The compact mode is the most used since
it guarantees the best simultaneity of the data. However, due to the different cadences of the
instruments (up to few tens of seconds for EUVI, two hours for HI2), and in order to study
the propagation of an event through the FOVs, it is convenient to be able to go backward or
forward in time with one instrument only without affecting the others. This is possible in the
normal mode in which there is no constraint on simultaneity. The added flexibility permits,
11
Figure 5 The visualization GUI. Up to three visu-GUIs can be opened simultaneously, one for each
STEREO probe and one for SOHO. This screenshot of the STEREO A GUI shows a composite of EUVI
19.5 nm, COR1, and COR2. The user can dynamically pan and zoom through the composite image with the
mouse.
for example, the creation of an EUVI/COR1/COR2 composite in which the images of the
instruments are as close as possible in time (as in the compact mode), but with the added
condition that the dates are increasing from EUVI to COR1 to COR2. Finally, the normal
and compact modes are complemented by the linked mode, in which the navigation in
time is synchronized between the selected probes. The linked mode allows for easy creation
of movies of simultaneous STEREO and SOHO composites.
2.7. Filters and Enhancements
Before they are put together as a composite, the projected images are converted from 32-bit
floating point numbers to eight-bit integers, using a choice of three response curves: linear,
square root, or logarithmic. Also, a large variety of digital filters can be applied to the data.
12
F. Auchre et al.
A filter is any IDL function, either built-in or user-written. The filters are input using a
command line in the filter dialogue box. The only constraint is for the function to accept one
matrix parameter. Built-in IDL functions that can be used as filters are, e.g., smooth (image,
10) or median (image, 3). The straightforward application is to enhance the data with, for
example, an unsharp mask, a wavelet filter, etc. Complex operations can be achieved with
user-defined filters, such as processing the input matrix, computing physical parameters, and
reading and writing files. Any number of filters can be applied sequentially to the data.
2.8. Graphical Output
2.8.1. Images
The content of a visualization GUI can be stored either in postscript format or in several
bitmap formats (JPEG, TIFF, etc.). In postscript format, each image forming the composite
is stored at its native resolution. The grid is rasterized, even though it is a vector object
within FESTIVAL. In the bitmap formats, the content of the visualization GUI (including the
coordinate grid if present) is rasterized before it is saved. By default, the scene is rasterized
to produce an output of size equal to that of the visualization area of the GUI. But the width
and height of the output can be changed. The maximum resolution is 10 000 10 000 pixels.
2.8.2. Movies
FESTIVAL is able to save at once a whole sequence of images that can later be put together
in a movie (i.e., MPEG or animated GIF) using video-encoding software. The sequence of
images saved is equivalent to what would be obtained by clicking the next button through
the list of images, as defined by current mode (normal, compact, or linked).
3. 3D Visualization
Most users will use FESTIVAL with a standard computer display, and when working with
multiple probes, it is much more comfortable to use one display per probe in order to maximize the viewing area for each. Many graphics card are able to do so. FESTIVAL can also
be used with 3D visualization hardware to display stereoscopic images. A sample setup using a passive technology system with left/right circular polarizers is illustrated by Figure 6.
In order to achieve stereoscopic viewing, FESTIVAL is run as usual, but the STEREO A
visualization GUI is simply sent to one projector, and the STEREO B GUI to the other. For
use with such devices, FESTIVAL includes a 3D mode in which the panning and zooming
are synchronized between STEREO A and B. In this mode, the user can navigate through the
image or enlarge a region of interest without losing the stereo vision effect. There is no intrinsic limitation of the stereoscopic mode with EUVI data. However, stereoscopic vision is
much easier to obtain and much more comfortable for EUVI images than for coronagraphic
data. Filaments are most prominent and easy to see above the chromosphere. Interestingly,
some faint filaments would not be (or barely) identified as such using 2D images only, but
they appear clearly in 3D. This suggests that the accuracy of the measurement of filament
altitudes by tie-point methods (see, e.g., Aschwanden et al., 1999) could be improved by
working in a 3D visualization environment. Finally, it is worth noting that even though the
resolutions of the instruments are different, it is possible to use FESTIVAL for stereoscopic
viewing of, for example, EUVI B and EIT data. This naturally doubles the duration of the
period during which stereoscopic vision is achievable.
13
Figure 6 A sample hardware setup with one standard computer display and two projectors for 3D stereoscopic viewing.
4. Perspectives
FESTIVAL is a novel tool for the visualization of multispacecraft solar imaging data. The
quick panning and zooming capability makes it easy to navigate through large images or
composite of images, and this functionality is compatible with stereoscopic displays. After
a few months of tests, several directions for improvements can be envisioned.
FESTIVAL currently supports images from SECCHI, EIT-LASCO, the Nanay Radioheliograph (NRH), and the MkIV coronagraph. The addition of other instruments is a natural evolution which we are pursuing. As is usually done, by neglecting the small parallax
with the L1 halo orbit of SOHO as seen from the Sun, images provided by spacecraft in
Earth orbit or by ground-based observatories can be displayed in the SOHO visualization
GUI. However, if a new spacecraft were to be launched on an interplanetary orbit (such as
STEREO), the images would have to be displayed in a separate visualization GUI, for they
could not be meaningfully combined with the data from other probes. One can also think of
the visualization of simulated data overlayed, or not, on top of real observations.
Today, the user needs expensive 3D display hardware to be able to view stereoscopic
images with FESTIVAL. A useful addition would be the ability to create red/cyan anaglyphs.
These are not the best way to visualize 3D images because ghost images can never be totally
eliminated, but they require only inexpensive red/cyan glasses.
At present, FESTIVAL works with local data, which means that the user has to access one of the solar databases to get the data before visualizing them. The capability to
query remote databases and download the data would render this process transparent to the
user.
Finally, FESTIVAL was designed to be able to be run from a command line, without
the graphical interface, with the possibility to set the main visualization parameters. Even
14
F. Auchre et al.
though this feature is not implemented yet, it opens a large range of perspectives. The kernel
of FESTIVAL could then become a powerful visualization segment of a pipeline processing
chain.
Acknowledgements The authors thank the users of FESTIVAL for many useful comments. FESTIVAL
is a collaborative project supported by CNES. The STEREO/SECCHI data used here are produced by an
international consortium of the Naval Research Laboratory (USA), Lockheed-Martin Solar and Astrophysics
Lab (USA), NASA Goddard Space Flight Center (USA), Rutherford Appleton Laboratory (UK), University of Birmingham (UK), Max-Planck-Institut fr Sonnensystemforschung (Germany), Centre Spatiale de
Lige (Belgium), Institut dOptique Thorique et Applique (France), and Institut dAstrophysique Spatiale
(France).
References
Aschwanden, M.J., Newmark, J.S., Delaboudinire, J.-P., Neupert, W.M., Klimchuk, J.A., Gary, G.A.,
Portier-Fozzani, F., Zucker, A.: 1999, Astrophys. J. 515, 842.
Brueckner, G.E., Howard, R.A., Koomen, M.J., Korendyke, C.M., Michels, D.J., Moses, J.D., Socker, D.G.,
Dere, K.P., Lamy, P.L., Llebaria, A., Bout, M.V., Schwenn, R., Simnett, G.M., Bedford, D.K., Eyles,
C.J.: 1995, Solar Phys. 162, 257.
Defise, J.-M., Halain, J.-P., Mazy, E., Pierre, P., Howard, R.A., Moses, J.D., Socker, D.G., Harrison, R.A.,
Simnett, G.M.: 2003, In: Keil, S.L., Avakyan, S.A. (eds.) Innovative Telescopes and Instrumentation for
Solar Astrophysics, SPIE 4853, 12.
Delaboudinire, J.-P., Artzner, G.E., Brunaud, J., Gabriel, A.H., Hochedez, J.F., Millier, F., Song, X.Y., Au,
B., Dere, K.P., Howard, R.A., Kreplin, R., Michels, D.J., Moses, J.D., Defise, J.M., Jamar, C., Rochus,
P., Chauvineau, J.P., Marioge, J.P., Catura, R.C., Lemen, J.R., Shing, L., Stern, R.A., Gurman, J.B.,
Neupert, W.M., Maucherat, A., Clette, F., Cugnon, P., van Dessel, E.L.: 1995, Solar Phys. 162, 291.
Domingo, V., Fleck, B., Poland, A.I.: 1995, Solar Phys. 162, 1.
Howard, R., Moses, D., Vourlidas, A., Davila, J., Lemen, J., Harrison, R., Eyles, C., Defise, J.-M., Bothmer,
V., Ravet, M.-F., The SECCHI Team: 2006, Adv. Space Res. 29(12), 2017.
Kaiser, M.L.: 2005, Adv. Space Res. 36(8), 1483.
Perryman, M.A.C., Lindegren, L., Kovalevsky, J., Hoeg, E., Bastian, U., Bernacca, P.L., Crz, M., Donati, F.,
Grenon, M., van Leeuwen, F., van der Marel, H., Mignard, F., Murray, C.A., Le Poole, R.S., Schrijver,
H., Turon, C., Arenou, F., Froeschl, M., Petersen, C.S.: 1997, Astron. Astrophys. 323, L49.
Thompson, W.T., Davila, J.M., Fisher, R.R., Orwig, L.E., Mentzell, J.E., Hetherington, S.E., Derro, R.J.,
Federline, R.E., Clark, D.C., Chen, P.T.C., Tveekrem, J.L., Martino, A.J., Novello, J., Wesenberg, R.P.,
St Cyr, O.C., Reginald, N.L., Howard, R.A., Mehalick, K.I., Hersh, M.J., Newman, M.D., Thomas,
D.L., Card, G.L., Elmore, D.F.: 2003, In: Keil, S.L., Avakyan, S.A. (eds.) Innovative Telescopes and
Instrumentation for Solar Astrophysics, SPIE 4853, 1.
Wuelser, J.-P., Lemen, J.R., Tarbell, T.D., Wolfson, C.J., Cannon, J.C., Carpenter, B.A., Duncan, D.W.,
Gradwohl, G.S., Meyer, S.B., Moore, A.S., Navarro, R.L., Pearson, J.D., Rossi, G.R., Springer, L.A.,
Howard, R.A., Moses, J.D., Newmark, J.S., Delaboudinire, J.-P., Artzner, G.E., Auchre, F., Bougnet,
M., Bouyries, P., Bridou, F., Clotaire, J.-Y., Colas, G., Delmotte, F., Jerome, A., Lamare, M., Mercier,
R., Mullot, M., Ravet, M.-F., Song, X., Bothmer, V., Deutsch, W.: 2004, In: Fineschi, S., Gummin, M.A.
(eds.) Telescopes and Instrumentation for Solar Astrophysics, SPIE 5171, 111.
Visualization of Distributed Solar Data and Metadata

with the Solar Weather Browser
B. Nicula C. Marqu D. Berghmans

Abstract The Solar Weather Browser (SWB) is a standalone, open-source software tool
designed to display solar images with context overlays. It was originally developed for the
space-weather forecast activities of the Solar Influence Data analysis Center (SIDC) but it is
more generally well suited to display the output of solar-feature recognition methods. The
SWB is also useful in the context of distributed solar-image archives, where it could play
the role of a quick-look viewer. The SWB allows the user to visually browse large solar
data sets and investigate the solar activity for a given date. It has a client server design
that minimizes the bandwidth from the network to the users monitor. The server processes
the data using the SolarSoft library and distributes them through a Web server to which the
SWB client connects. The client is readily available for Linux, Mac OS X, and Windows at
http://sidc.be/SWB. We discuss the software technology embedded in the SWB as well as its
use for solar physics and space weather.
Keywords Sun: software Sun: database Sun: public outreach
1. Introduction
In recent years, the advent of space-borne solar observatories such as SOHO (Domingo,
Fleck, and Poland, 1995) has led to a significant increase in the amount of solar data available through the Internet. Providing easy and pertinent access is now an important challenge
facing the scientific community in making optimum use of these data. For the SOHO mission, the plan was to distribute the data through a centralized database such as MEDOC
(Scholl, 1999) and to provide the SolarSoft IDL-based library (Freeland and Handy, 1998)
for data processing. All major recent solar missions such as STEREO (Kaiser, 2004) or Hinode (Kosugi et al., 2007) follow the same track. The data from ground-based observatories
B. Nicula C. Marqu D. Berghmans ()
SIDC Royal Observatory of Belgium, Ringlaan 3 Avenue Circulaire, 1180 Brussels, Belgium
e-mail: david@oma.be
B. Nicula
e-mail: bogdan.nicula@sidc.be
DOI: 10.1007/978-0-387-98154-3_3
15
16
B. Nicula et al.
Figure 1 Screen shot of the SWB showing a SOHO/EIT image, the position of active regions and H plages
as reported by NOAA, and minor CME activity as detected by CACTus. The arcs denote the angular width
and position angle of each eruption.
are typically less centralized but are instead served on particular institute Web sites that
are sometimes difficult to find. Even more scattered are the solar metadata such as spaceweather event reports, active region, and sunspot group numberings, as well as the output
of the operational feature-recognition methods. A logistic problem appears when users want
to combine images from different missions, ground-based observatories, and/or the various
types of solar metadata. More and more analyses indeed rely on full multiwavelength and
therefore multi-instrument data processing. How can the users, in an efficient way, find,
locate, and download to their computer the relevant information they need?
One answer to this is the concept of a virtual observatory, in which a single interface is
used to request a wide range of ground-based or space-borne solar data. (See, for example,
the EGSO (Bentley, Csillaghy, and Scholl, 2004) and VSO (Hill et al., 2004) Web sites.1 ) At
present, such infrastructures are often cumbersome to use, requiring the user to go through
a number of drop-down menus, checkboxes, and database queries. We present in this paper
a quick-look viewer, the Solar Weather Browser (SWB) that offers a more intuitive, datedriven interface (see Figure 1), allowing the user to browse large amounts of solar data,
display them, and make basic and contextual analysis, without an a priori knowledge of the
data location or without the need for further data processing.
Based on a client server design, SWB was originally developed for the space-weather
forecast activities at the Royal Observatory of Belgium, but it now has the potential to be
1 http://www.egso.org, http://virtualsolar.org.
17
a scientific tool for quick analysis as well as for public-outreach purposes. The present article describes the concept of SWB (Section 2), focusing in particular on the server design
(Section 3) and the client interface (Section 4). We conclude in Section 5. The SWB is an
open-source development, readily downloadable for all major platforms (see the Appendix).
2. SWB Concept
The SWB was designed to offer intuitive access to solar data and metadata that are otherwise
scattered over the Web. In many cases, users want a one-stop shopping application that
brings them a variety of solar images of one particular time or event on the screen without
surfing from Web site to Web site. Often, users do not want to see only one image but in
addition they want to have guiding metadata: Which active region was flaring lately? Where
did that CME come from? These users can vary from solar physicists, to space-weather
forecasters, to developers of feature recognition tools, to amateur astronomers. Any system
serving such a wide and large community must be simple and intuitive, it must be able to
run thousands of concurrent accesses, and the installation on all major computer platforms
must be straightforward.
The Solar Monitor Web site2 addresses many of these requirements through an implementation in various Web technologies (HTML, PHP, etc.). It however does not offer the
flexibility of interactively combining images with context data. Instead, all of the metadata
(active region numbers, flaring, and coordinate systems) are stored together with the image files in preprocessed PNG files. Also rescaling the field of view, changing the coloring,
or quickly looping (movie-like) through a sequence of images is not possible. In principle,
some of these functionalities could be added through dedicated subpages, as the SolarSoft
Latest Event Web site3 does. The problem is that the developer must then anticipate all possible combinations of data and metadata the user might be interested in and foresee that
within the unlimited growing Web site.
A different approach is followed by the FESTIVAL package.4 FESTIVAL visualizes
beautifully and interactively combines solar images from the SECCHI instruments on
STEREO and EIT and LASCO on SOHO. This package however cannot serve as a quicklook tool for the general public as it requires that all the scientific data be locally available
and it runs within the proprietary IDL environment.
The SWB alternative that we present in this paper has a client server architecture (see
Figure 2). The SWB client, whose interface is presented in Section 4, is installed on the
users machine. It allows the user to interactively select a desired image combination, and
it then retrieves, colorizes, and combines the images. The SWB server (see Section 3) runs
remotely and through a dedicated Web site, offering all of the information that the SWB
client requires. Asynchronously, the SWB server collects new information over the Web
through a variety of protocols (http, ftp, e-mail, etc.).
Such a split architecture, with a strong decoupling between the server and the client, has
the advantage that new data types (e.g., after the availability of a new solar telescope) can be
easily added on the server side, without an upgrade being required on the users computer.
Moreover, it can take full advantage of the computer resources of the user; since the SWB
2 http://www.solarmonitor.org.
3 http://www.lmsal.com/solarsoft/latest_events/.
4 http://www.ias.u-psud.fr/stereo/festival.
18
B. Nicula et al.
Figure 2 Layout of the SWB concept. The SWB interface is installed on the users computer (bottom panel)
and allows the user to select and combine a variety of images and context information per selected date. Once
the user has made a selection, the interface collects the data in a preprocessed format from the appropriate
servers.
is a dedicated application in native code on the user computer, no limitations appear when
multiple users are working in parallel. The SWB server is accessed only once for the original
data download. High compression (Nicula, Berghmans, and Hochedez, 2005) and careful
caching techniques guarantee that this access is as smooth as if the data existed on the users
hard disk. Since images and overlays are combined on the fly in the interface, the total
storage on the SWB server does not scale with the total number of possible combinations.
At the time of this writing, the SWB server supports SOHO/EIT and LASCO images,
SOHO/MDI magnetograms and continuum images, ground-based white-light and H images, and Nobeyama radio images. STEREO/SECCHI images and Nanay radio images
are in preparation. Overlay metadata include various longitude/latitude grids, active region,
and sunspot group information, flaring history (for example, NOAA5 ), and CACTus6 CME
detections (Robbrecht and Berghmans, 2004).
The current implementation of the SWB has potential beyond pure solar data dissemination. Given its highly optimized image compression and transmission, the SWB is also
potentially useful in the context of distributed solar-image archives, where it could play the
role of a quick-look previewer. It is also particularly useful for the presentation of results
from automated solar image recognition/processing chains.
5 http://www.sec.noaa.gov.
6 http://sidc.be/cactus.
19
Figure 3 Layout of the SWB

server. The SWB server pulls
new data from external or local
archives to the input
SWBS_data directory. The data
are then processed with routines
in the SWBS_bin directory and
transformed to a standard format.
The output is written to the
directory SWBS_out and from
there made available via an http
server.
3. Server Design
The main task of the SWB server is to collect solar data (images or event catalogs) from
various sources, either local or remote, and to process and convert them into a file format
compatible with the SWB client (see Figure 2).
All images are rebinned in the server with a 512 512 pixel resolution and are recoded in
eight bits (JPEG compression or PNG format). This reduced resolution is configurable but it
has been chosen to minimize the download time over the network while not hampering the
quick-look functionalities. Full-disk images are rebinned to a common field of view, which
for historical reasons is that of the EIT telescope onboard SOHO. Depending on the kind
of image and instrument, further processing may be needed: cosmic-ray removal, shifting
of the Suns center to the center of the image, scaling-factor estimation (for coronagraphs
for example), or histogram equalization to avoid brightness variations from one image to
the next. The image processing mostly relies on the usual SolarSoft routines provided by
the different instrument teams. In addition to these generic routines, SWB-related image
processing relies on a small set of specific programs whose main role is to produce the final
output.
In addition to solar images, contextual overlays are also created (e.g., heliographic grids,
flare and sunspot locations, or CME events). The Scalable Vector Graphics7 (SVG) format
has been chosen to provide high-quality graphics independent of the resolution or the zoom
factor chosen by the user. They are produced by IDL programs calling specific open-source
libraries (libplot,8 Poppler,9 and cairo10 ).
The server itself is a simple directory tree whose structure is schematized in Figure 3.
It can be easily set up at any location by defining a few environment variables. The whole
processing, from data collection to upload to the Web server, is done automatically using
UNIX shell scripts, controlled by cron jobs.
7 http://www.w3.org/Graphics/SVG.
8 http://www.gnu.org/software/plotutils.
9 http://poppler.freedesktop.org.
10 http://cairographics.org.
20
B. Nicula et al.
The server uses a generic compression tool (zip), available for UNIX, to produce the
output files, consisting of one ASCII file and one data file (JPEG for images and SVG for
overlays). ASCII files are currently empty but it is foreseen that they will contain information
such as the spacecraft position or the location on the Web of the original FITS file for
scientific use. When the file is produced, an entry is created in a list that keeps track of all
of the files available for that day. Each entry consists of information necessary to single out
every individual file: the kind of data (image or overlay), instrument, date of observation,
UNIX time of observation, file name, color table, and scale factor.
Ultimately the output files and the list are uploaded to a Web server to which the client
eventually connects. There is currently one main Web server, located at the Royal Observatory of Belgium, but the client (described in the next section) can connect to multiple Web
servers distributed over the Internet (see Figure 2). Each of those can be attached to one local
SWB server that processes one particular instrument. This will provide in the future both redundancy and load distribution when the amount of data significantly increases. At the time
of this writing, tests are being performed with the High Altitude Observatory in Boulder for
the Mauna Loa data, as well as with the Paris Observatory for the Nanay Radioheliograph
images.
4. Client Interface
The user interface source code of the Solar Weather Browser is written in C in a crossplatform manner and runs on Microsoft Windows and any UNIX-like operating system for which the GIMP toolkit version 2.8 or later is available (including Mac OS X).
Images can be optionally rendered using OpenGL software interfaces, allowing for
hardware-accelerated overlaying and scaling. This allows also for color mapping using the
ARB_fragment_program OpenGL extension for a significant bandwidth reduction (75%) to
the video-card memory.
Browsing through the available data is in the first instance date-driven through the calendar tool at the top left of the interface. The current day, as indicated by the users computer
clock, is automatically selected at program startup. While operating the calendar control,
the interface retrieves, in the background from the SWB server, a list of all available images
and overlays for each selected day. This information is used to populate, on the fly, the three
image and overlay type drop-down menus just below the calendar tool.
After selection of an image type, the available images list can be traversed using the
daily slider. This can be done slowly, image per image, or quickly as a movie. The cursor
keys from the keyboard can be used to move backward or forward by one image (left and
right keys), by one day (up and down keys), or by one solar rotation (page down and page
up). This last functionality is very handy to compare the evolution of coronal holes and
active regions over longer time scales.
Because the images on the server are generated in near-real time, the current day list of
available images can be refetched from the server using Refresh Day in the File menu.
The message area displays for each image and overlay the time of observation, a short
description of what is shown, and a link to the original data source Web site for image
courtesy and reference.
Finally, a (normally hidden) menu for display adjustments can be brought up to zoom in
on a certain subfield or to adjust brightness, contrast, saturation, and opacity (see Figure 4).
The display of the image area, including the overlays, can be saved by selecting Save As
in the File menu. The saved file is always a nonscaled PNG image.
21
Figure 4 Screen shot of the SWB showing a SOHO/EIT 30.4-nm image, the location of minor flaring
activity (source SolarSoft Latest Event Archive), and a 10 longitude/latitude grid. The display adjustment
menu can be seen at the bottom left.
5. Conclusions
Since it was first implemented, the Solar Weather Browser has evolved from a basic display program toward a more complete tool to browse a large amount of quick-look solar
data. Without the need for expensive proprietary software, it allows a broad community interested in solar physics to have access to a wide range of solar images as well as a basic
set of metadata (flare, sunspot locations, etc.) and combine them at will on a personal computer. Potential users of such a tool are the space-weather community, teachers and science
museum personnel, amateur astronomers, and, of course, the scientific community. With the
current amount of solar data available through the Web and the concomitant necessity to
learn new and complex processing software, we believe that a quick-look browser has some
scientific potential: It can be used for example to quickly assess the importance of a solar
flare, its location, or whether a halo CME was associated with it.
This purpose can be achieved only if a large set of solar data is available through the
Solar Weather Browser. A server can be easily set up at a PI institute of a space mission
or a ground-based observatory, and we invite those interested to contact the authors of this
article to help in this process. The client and the server side of the SWB will evolve with the
characteristics of new solar missions: nonheliocentric point of view, high cadence, or high
resolution. With automatic recognition efforts initiated at many institutes to collect, classify, and archive solar events or solar features, the Solar Weather Browser has the potential
to bring all of these valuable metadata to the community within a single and easy-to-use
interface.
22
B. Nicula et al.
Appendix
The reader is encouraged to download and test the SWB from the SIDC Web site
http://sidc.be/SWB. Different packages are freely available for different operating systems
and a few installation hints are provided. People with an interest in participating in the further development of the server or interface can contact the authors. We welcome especially
data providers who would like to have their images or metadata served through the SWB.
The Solar Weather Browser client is open software distributed under the GNU General
Public License. The Solar Weather Browser server software is available on demand. Please
note that this software comes with absolutely no warranty nor support. The full licence details are available on the http://sidc.be/SWB Web site and in the Help menu of the interface.
Acknowledgements The authors would like to thank T. Katsiyannis and I. Baumann for their contribution to the SWB server development. SWB was partially funded by the European Space Agency
(16913/03/NL/LvH) and by PRODEX Contract Nos. C90204 (Solar Drivers of Space Weather), C90193
(SWAP preparation to exploitation), and C90192 (Telescience), which are managed by the European Space
Agency in collaboration with the Belgian Federal Science Policy Office. The authors would like also to thank
the open data policy of the different instrument teams, as well as the Free Software community for the wealth
of high-quality software that it makes available.
References
Bentley, R.D., Csillaghy, A., Scholl, I.: 2004, Proc. SPIE 5493, 170.
Freeland, S.L., Handy, B.N.: 1998, Solar Phys. 182, 497.
Hill, F., Bogart, R.S., Davey, A., Dimitoglou, G., Gurman, J.B., Hourcle, J.A., Martens, P.C., Suarez-Sola, I.,
Tian, K., Wampler, S., Yoshimura, K.: 2004, Proc. SPIE 5493, 163.
Kaiser, M.L.: 2004, Adv. Space Res. 36, 1483.
Kosugi, T., et al.: 2007, Solar Phys. 243, 118.
Nicula, B., Berghmans, D., Hochedez, J.-F.: 2005, Solar Phys. 228, 253.
Robbrecht, E., Berghmans, D.: 2004, Astron. Astrophys. 425, 1097.
Scholl, I.: 1999, In: Vial, J.-C., Kaldeich-Schmann, B. (eds.) Proceedings of the 8th SOHO Workshop 446,
ESA Special Publications, Noordwijk, 611.
Widespread Occurrence of Trenching Patterns

in the Granulation Field: Evidence for Roll Convection?
A.V. Getling A.A. Buchnev

Abstract Time-averaged series of granulation images are analysed using COLIBRI,

a purpose-adapted version of a code originally developed to detect straight or curvilinear
features in aerospace images. The image-processing algorithm utilises a nonparametric statistical criterion that identifies a straight-line segment as a linear feature (lineament) if the
photospheric brightness at a certain distance from this line on both sides is stochastically
lower or higher than at the line itself. Curvilinear features can be detected as chains of
lineaments, using a modified criterion. Once the input parameters used by the algorithm
are properly adjusted, the algorithm highlights ridges and trenches in the relief of the
brightness field, drawing white and dark lanes. The most remarkable property of the trenching patterns is a nearly universally present parallelism of ridges and trenches. Since the
material upflows are brighter than the downflows, the alternating, parallel light and dark
lanes should reflect the presence of roll convection in the subphotospheric layers. If the
numerous images processed by us are representative, the patterns revealed suggest a widespread occurrence of roll convection in the outer solar convection zone. In particular, the
roll systems could form the fine structure of larger scale, supergranular and/or mesogranular
convection flows. Granules appear to be overheated blobs of material that could develop into
convection rolls owing to instabilities of roll motion.
Keywords Sun: photosphere Sun: granulation
1. Introduction
Getling and Brandt (2002) reported that images of solar granulation averaged over time
intervals as long as, e.g., two hours, are far from completely smeared but contain bright,
A.V. Getling ()
Institute of Nuclear Physics, Lomonosov Moscow State University, 119991 Moscow, Russia
e-mail: a.getling@mail.ru
A.A. Buchnev
Institute of Computational Mathematics and Mathematical Geophysics, 630090 Novosibirsk, Russia
e-mail: baa@ooi.sscc.ru
DOI: 10.1007/978-0-387-98154-3_4
23
24
A.V. Getling, A.A. Buchnev
granular-sized blotches, which may form quasiregular systems of concentric rings or parallel
strips ridges and trenches in the brightness field on a meso- or supergranular scale.
Getling (2006) implemented a detailed investigation of such long-lived patterns and found
that they do not appear unusual in images averaged over one- to two-hour time intervals.
These systems resemble some specific types of the roll patterns known from laboratory
experiments on Rayleigh Bnard convection and may reflect the fine structure of subphotospheric convection cells. If the time variations of intensity are traced at the point corresponding to a local intensity maximum in the averaged image (near the centre of a light blotch)
and at a nearby local-minimum point, a remarkable pattern of correlations between these
variations can be noted. In some cases, the correlations are periodic functions of the time
lag or a tendency to anticorrelation is observed. This fact supports our suggestion (Getling,
2000) that granules are hot blobs of the solar plasma carried by the convective circulation
and that they can even reemerge on the photospheric surface.
Since the quasiregular structures manifest themselves in time-averaged images, they
should be associated with a long-lived component of the granulation field. Getling and
Brandt (2002) and Brandt and Getling (2004) noted that the decrease in the rms contrast
of the averaged images with the averaging time [t ] is considerably slower compared to the
statistical t 1/2 law, and this fact was regarded as a particular argument for the presence of
long-lived features.
In his comment on the original paper by Getling and Brandt (2002), Rast (2002) suggested that the structures in the granulation field are merely of a statistical nature and do
not reflect the structure of real flows. He constructed a series of artificial random fields with
some characteristic parameters typical of solar granulation and found some similarities between the properties of these fields and of the real granulation patterns. On this basis, Rast
raised doubts about the existence of the long-lived component in the granulation dynamics.
A detailed discussion of Rasts criticism given by Getling (2006) invokes, in particular, the
contrast-variation analysis (Brandt and Getling, 2004). A number of counterarguments are
presented and it is shown that Rasts considerations cannot be applied to the real granulation
patterns.
As noted by Getling and Brandt (2002), signatures of the prolonged persistence of
some features in the granulation patterns have already been observed previously. Roudier
et al. (1997) detected long-lived singularities intergranular holes (dark features) in
the network of supergranular lanes. Such holes were continuously observed for more than
45 minutes, and their diameters varied from 0.24 (180 km) to 0.45 (330 km). Hoekzema,
Brandt, and Rutten (1998) and Hoekzema and Brandt (2000) also studied similar features, which could be observed for 2.5 hours in some cases. Baudin, Molowny-Horas, and
Koutchmy (1997) attributed the blotchy appearance of a 109-minute-averaged granulation
image to a sort of persistence of the granulation pattern; alternatively, this effect can be interpreted in terms of the recurrent emergence of granules at the same sites (Getling, 2006).
Some indications of a long-term spatial organisation in the granulation field have also
been revealed in observations. Dialetis et al. (1988) found that granules with longer lifetimes exhibit a tendency to form mesogranular-scaled clusters. Muller, Roudier, and Vigneau (1990) also detected such clustering in the spatial arrangement of large granules; they
emphasised a plausible relationship between the clusters and mesogranules. Roudier et al.
(2003) reported their observations of a specific collective behaviour of families (trees) of
fragmenting granules. Such families can persist for up to eight hours and appear to be related to mesogranular flows. An imprint of the supergranulation structure can also be traced
in the granulation pattern (Baudin, Molowny-Horas, and Koutchmy, 1997).
Based on their analysis of pair correlations in the supergranular and granular fields,
Berrilli et al. (2004) reported a finding that bears some similarity with ours. Specifically,
Widespread Occurrence of Trenching Patterns in the Granulation Field
25
the probability of finding a target supergranule or granule (identified by its barycentre) at

a given distance from the barycentre of a chosen reference supergranule or granule is an
oscillating function of this distance with a local amplitude decreasing from some maximum
(reached at a small distance). This reflects a specific kind of order in the supergranulation
and granulation patterns, which is consistent with the presence of concentric-ring structures.
It is remarkable that an effect of spatial ordering has also been detected on a supergranular scale. Lisle, Rast, and Toomre (2004) revealed a persistent alignment of supergranules
in a meridional direction reminiscent of the alignment of light blotches described by us.
There is also a parallelism in the interpretation of the quasiregularity in the granulation and
supergranulation fields: Whereas we are inclined to interpret the trenching patterns as the
manifestation of a fine structure of the supergranular flows, Lisle, Rast, and Toomre (2004)
attribute the alignment of supergranules to an ordering influence of giant convection cells.
Here, we employ a specific image-processing algorithm to analyse granulation patterns
and investigate their spatial organisation. In particular, the results of our study show that
the arrangement of granules may assume various forms, and the patterns of light blotches
detectable in the averaged images have a common feature a nearly universally present parallelism of alternating light and dark lanes. If the images analysed by us are representative,
the trenching patterns could naturally be interpreted as manifestations of roll convection entraining granules; this form of motion proves to be widespread in the upper subphotospheric
layers.
2. The Data
We mainly deal with excellent images of the photosphere from the well-known La Palma
series obtained by Brandt, Scharmer, and Simon (see Simon et al., 1994) on 5 June 1993
using the Swedish Vacuum Solar Telescope (La Palma, Canary Islands). Specifically, we use
a seven-hour subset of this series and a 43.5 43.5 Mm2 subfield (480 480 pixels of size
90.6 km) of the original images. The observed area was located not far from the disk centre,
and the images were produced by the telescope in the 10-nm-wide spectral band centred at a
wavelength of 468 nm. The resolution was typically no worse than about 0.5 , and the frame
cadence was 21.03 seconds.
Previously (Getling and Brandt, 2002; Getling, 2006), we already described the principal
elements of the data acquisition and preprocessing techniques employed by Simon et al.
(1994). Here, we only briefly recall that the images were aligned, destretched, and cleaned
from rapid intensity variations by means of subsonic filtering. Furthermore, all of them
were normalised to the same value of the rms contrast, and the residual large-scale intensity
gradients were removed from them.
We give here our primary attention to granulation fields averaged over one- to two-hour
intervals.
In addition to the data of the La Palma series, we use here an averaged image from
the subsonically filtered version of the 45.5-hour series obtained using the SOHO MDI
instrument in 1997, from 17 January 00:01 UT to 18 January 21:30 UT (see Shine, Simon,
and Hurlburt, 2000). This series contains white-light images with a resolution of about 1.2
taken at a one-minute interval. We present here an enlarged 87 70 Mm2 cutout (200 160
pixels) of the selected image.
26
3. Image Processing
In our analysis, we use the COntours and LIneaments in the BRIghtness field (COLIBRI)
code, a purpose-adapted version of a code constructed previously at the Institute of Computational Mathematics and Mathematical Geophysics (Novosibirsk, Russia) and intended for
the detection of linear and circular structures in aerospace images (a design that has been
successfully employed for several years). The code is an implementation of certain statistical criteria for the detection of stretched features (in our case, linear structures and contour
elements) against a random background. The underlying algorithm was developed by Salov
(1997).
The principal reason for using such techniques is as follows. The random component of
the processes under study clutters the observational data (i.e., images) with random variations, which mask the brightness differences between the points of the object sought and of
the background areas. Accordingly, reliable criteria for detection of objects should be based
on a probabilistic (statistical) approach. The presence of an object in the image manifests
itself in the fact that the random observable quantities in the object are stochastically or,
more simply, systematically larger (or smaller) than in the background areas. The so-called
nonparametric criteria, based on stochastic comparisons of the observables inside and outside the hypothetical object, do not depend on the (unknown) distributions of the observables
and prove to be efficient for our purposes.
The COLIBRI code has two operation modes that enable the detection of either lineaments or contours in photospheric images.
A lineament (linear element) is a stretched, nearly linear object, blurred and nonuniform
in brightness along its extent. To detect a lineament, the algorithm analyses (almost) all
of its possible positions. For any trial position, to decide whether or not a lineament is
present, the set of brightness values taken at some points along the hypothetical lineament
of length l is compared with the two sets taken at the points located at a certain distance
on both sides of the lineament, at the normals intersecting it at the originally chosen
points. Specifically, the algorithm checks the hypotheses that either all the sets of measured
quantities are stochastically equal (no object is present) or, alternatively, the set of the central
brightness values is stochastically greater (smaller) than the sets taken on both sides (an
object is present). The nonparametric criterion used to test these hypotheses (Salov, 1997) is
based on computing the values of the Mann Whitney statistics (Mann and Whitney, 1947;
Whitney, 1951).
A somewhat different approach is used in the detection of contours, by which are meant
curvilinear features with a priori unknown shapes. In this case, the algorithm constructs the
contours as chains of lineaments, but the criterion for their detection has a different form.
Again, sets of brightness values are taken at normals to the trial lineament at a certain distance on both sides of it. To decide whether or not an object is present, the algorithm
checks the hypotheses that either the sets of brightness values taken on the two sides are
stochastically equal (no object is present) or, alternatively, the values on one side are stochastically larger then the values on the other side (an object is present). In this case, the
nonparametric criterion used to test these hypotheses is also based on the computed values
of the Mann Whitney statistics (Salov, 1997). In essence, a trial lineament is assumed to be
a really present object provided the brightness gradient across it has some stochastically predominant sign. In this respect, the detected contours very crudely correspond to isohypses
(horizontals) in a topographic contour map; however, they do not depend on the fine topographic details of the brightness field because the intensity values at a contour are not taken
into account in constructing this contour.
27
The distance , which is everywhere measured in pixels, is an input parameter of the

algorithm. Other important parameters are the admissible probability of spurious detection
of objects [p] and the range of admissible lengths of lineaments [l] (also expressed in pixels). The determination of the actual probability of spurious detection is fairly sophisticated
(Salov, 1997), and describing it is beyond the scope of this paper.
Because of the noisy appearance of the analysed images, the detection procedure straightforwardly applied to an image may reveal a large number of spurious details. (By noise, we
mean the presence of a multitude of light blotches interspaced by darker features, any of
which, taken alone, is not representative from the standpoint of seeking quasiregular structures.) To reduce the noise level, a special filtering procedure was developed. It consists of
two steps. First, any lineament is removed if it is completely covered by another one; at the
same time, lineaments that partially overlap one another are combined into one lineament.
Second, all lineaments that have no parallel neighbours (i.e., do not belong to any bunch
of lineaments) are also removed. The effect of filtering will be illustrated in the following
(Figure 2).
4. Results
Our primary goal was to analyse structures in time-averaged series of granulation images.
In addition, we attempted to process individual images, or snapshots. Signs of regularity can
be visually noted in both cases; however, the snapshots contain many more fine details than
do the averages, so that much more spurious features emerge when snapshots are processed.
Here, we restrict ourselves to considering only the structures detected in averaged images.
The number of lineaments and contours detected in granulation images strongly depends
on the chosen parameters of the image-processing procedure. It should be kept in mind
that, generally, there is no combination of parameters that could be considered a universal
optimum, applicable in any case. The optimum really depends on both the character of the
analysed pattern and on the objective of analysis in any particular situation.
First, the optimum value depends on the width of the linear features that are sought.
Typically, we are seeking contours that correspond to parallel ridges and trenches in the
relief of the brightness field. In the contour-detection mode, the COLIBRI code draws contours along the slopes and marks them as white or black curves depending on the direction of
the slope (i.e., on the stochastically predominant sign of the difference between the brightness values at the points located on the two sides of the contour at the same distance ).
Usually, the code draws bunches of nearly parallel, closely located contours of either sort,
which merge into white or black lanes. If is small, the detected contours should gather
near the steepest slope curve. However, as our experience shows (and as confirmed by some
qualitative considerations), at larger values comparable with some optimal, resonant
distance related to the characteristic distance between ridges and trenches, the algorithm
can output white lanes highlighting the ridges and dark lanes highlighting the trenches.
Similarly, there is no universally preferable value of the parameter p, which is regarded
as the admissible probability of spurious detection. The optimum p is not universal and
should be properly chosen for any processing regime specified by the other parameters.
The range of lineament lengths [l] used to construct contours has also some effect on
the detection of structures. Generally, the algorithm can better integrate lineaments into
contours for broader l ranges. However, the inclusion of very short lineaments results in a
very high noise level. Moreover, the computation time grows dramatically with widening of
the l range. Thus, a reasonable compromising choice should be made.
28
Figure 1 (a) A two-hour-averaged image and the results of contour detection in this image at
(b) p = 2.5 104 , = 5, l = 16 43; (c) p = 2 103 , = 8, l = 10 30; and (d) p = 102 , = 2,
l = 5 20.
Let us consider the two-hour-averaged image shown in Figure 1a and the structures detected in this image (Figures 1b 1d). The last three panels are arranged in the order of
increasing p; the largest (= 8) is used in the case of panel c and the smallest (= 2) in
the case of panel d. As we can see, numerous patches of trenching patterns (not everywhere
regular) with a characteristic width of ridges comparable to the granular size are clearly
manifest in panel c. It is noteworthy that, in this figure, most white lanes are parallelled by
one or more other white neighbouring lanes and at least two black neighbouring lanes. In
other words, trenching patterns are very common in the image (although they vary widely
in their area and in the number of alternating white and black lanes). A careful inspection
of panel c shows that the black lanes really correspond to dark features in the images but
not merely to background areas where no features have been detected (such areas are shown
in grey). In panel d, the image processing everywhere highlights finer details of the pattern.
29
Figure 2 An illustration to the

effect of the lineament-filtering
procedure: the result of
processing the image shown in
Figure 1a at the same values of
the processing parameters as in
Figure 1c but without filtering.
As a rule, they are even less regular than the granular-scaled ridges; however, a system of
very narrow parallel ridges can be distinguished near the upper left corner of the frame,
within the (incomplete) white rectangle, as a pattern of very thin hatches inclined to the left.
Thus, as we would expect, the characteristic scale of detectable features decreases with the
decrease of . However, if we reduce p by an order of magnitude (see panel b for example), too many features disappear, being treated by the code as spurious. The remaining
segments of curves are only the most reliably identified fragments of the structures, most
parts of which have been filtered out. Nevertheless, even such drastically filtered patterns
may be useful: the pattern enclosed in the (incomplete) white trapezoid in the upper left part
of panel b suggests that a combination of concentric arcs and radial rays is present there,
and it can be classified as a specific form of web pattern. [The occurrence of such patterns
in averaged images was noted by Getling (2006)].
In the pattern of granular-scaled trenching, which is most pronounced in panel c, the
structure marked by the white circle deserves particular attention. It appears as a system of
concentric rings deformed and spoiled in some way and should therefore be included in
our collection of ring systems. At the same time, it provides an example of a structure that
can hardly be visually distinguished in the original image but which becomes detectable if
the image is processed with a properly chosen parameter (= 8).
To illustrate the effect of the lineament-filtering procedure, we present here in Figure 2
the result of processing the image shown in Figure 1a at the parameters used to obtain Figure 1c but without filtering. It can easily be seen that the filtering procedure efficiently removes isolated bright blotches and very short lineaments, preserving more extended features
and areas with pronounced trenching.
Figure 3 also refers to two-hour averaging. Here, panel a shows an averaged image for
which the midtime of the averaging interval is nearly the same as for Figure 3 in Getling
and Brandt (2002). This is still the best-available (most interesting) image, in terms of the
presence of pronounced and highly ordered structures. The pattern in panel b was obtained
by using the same parameters as in the case of Figure 1c. Encircled in the upper left quadrant is the well-developed system of concentric rings that appear as fairly regular circles
30
Figure 3 (a) Another two-hour-averaged image and the results of contour detection in it at (b) p = 2 103 ,
= 8, l = 10 30; (c) p = 102 , = 6, l = 10 20; and (d) p = 102 , = 6, l = 20.
in panel a. Because of the presence of dark radial gaps interrupting the circles, the algorithm in some cases combines fragments of different rings into single contours, distorting
the visual regularity of the system. A structure that may be part of another ring system, also
marked with an (incomplete) circle, can be seen near the upper right corner of the frame.
As in Figure 1c, even outside the regular structures, white lanes are accompanied by and
alternate with black ones, and vice versa. In panel c, finer details are highlighted, since a
smaller is used in this case; in addition, p is here higher than for panel b. It is important
that the main features present in the pattern do not disappear as p is reduced from 102 (c)
to 2 103 (b), although the simultaneous increase of additionally removes the narrowest
lanes.
Finally, the effect of the l range can be understood by comparing panels c and d. The patterns shown in these panels were obtained at the same p and but the range l = 10 20 was
shrunk to the single value l = 20 with passing from panel c to panel d. Accordingly, short
31
Figure 4 An image averaged over a one-hour interval centred at nearly the same time as in the case of
Figure 1 and a comparison between the regimes of lineament and contour detection: (a) original image;
(b) lineaments detected at p = 103 , = 5, l = 16 43; (c) lineaments detected at p = 2.5 104 , = 8,
l = 16 43; (d) contours detected at p = 104 , = 8, l = 16 43.
lineaments disappeared; however, objects exceeding 20 pixels in lengths were preserved

because of combining partially overlapping lineaments into longer ones in the process of
filtering and combining chains of lineaments into contours.
Thus, different combinations of parameter values can highlight different features in the
image. To form a more complete idea of the structures present in the granulation field, the
parameters at which the image is processed should be varied.
Whereas varying the image-processing parameters can be useful for the detection of features differing in their size, varying the averaging time makes it possible to reveal features
differing in their lifetime. In this respect, it appears instructive to compare Figure 3, for
32
Figure 5 Dislocation in a
roll-convection pattern
(experimental photograph by
Berdnikov and Markov).
Figure 6 Processing of a two-hour-averaged MDI image: (a) original image; (b) the results of contour detection at p = 5 103 , = 2, l = 20.
which the averaging time is two hours, with Figure 4, for which this time is one hour and
the averaging interval is centred at virtually the same time as in the case of Figure 3.1
The original averaged image is given in Figure 4a; panels b and c present the results
of processing it in a lineament-detection mode without filtering.2 In both panels, there are
many patches with a highly pronounced trenching. Panel c is considerably richer in ridge
trench systems than panel b, although the corresponding p value is one-quarter that for
panel b. This is an additional illustration of the role of the parameter . As in the examples
just considered, the value = 8 is most favourable for the detection of contours of the
fundamental width, and this fact proves here to be more important than some reduction
of p.
It is interesting that the structure marked with a black box in Figure 4c is very similar to
a fragment of a convection-roll pattern with a dislocation. Such patterns have received much
attention in studies (particularly, experimental) of thermal convection (see, e.g., Getling
1 In Figure 4, where results obtained in both the lineament-detection and contour-detection modes are pre-
sented, they, by chance, appear very similar. In no way is this the general situation.
2 The pattern shown in Figure 4b was obtained by using an older version of the code, which could not draw
lineaments near the edges of the field of view.
33
(1998), for a survey); for comparison, we present an experimental photograph of a rollconvection pattern with a dislocation in Figure 5. (Obviously, this is merely an illustrative
example; taken alone, it is insufficient to substantiate the claim that the observed feature is
actually related to roll convection.)
The white rectangular box in Figures 4c and 4d (the latter panel representing the results
of contour detection) marks an extended ridge trench system, which is almost completely
smeared in the two-hour-averaged image (Figure 3). The pattern within the white circle bears
only faint resemblance to the pattern marked with the circle in Figures 3b 3d, although
some circular arcs can apparently be associated with the well-developed system present in
Figure 3.
These observations clearly demonstrate that different features can be identified in the
averaged images depending on the length of the averaging interval. In the case at hand, the
system of concentric rings visible in Figure 3 appears to have a longer lifetime than has the
elongated system identifiable in Figure 4.
Finally, we give here an example of processing results for an averaged image of the
SOHO/MDI series (Figure 6). Panel a is an enlarged cutout of the subsonically filtered
images of the MDI series averaged over a two-hour interval. As already noted by Getling
(2006), this averaged image (also reproduced in that paper) exhibits a pattern of ridges and
trenches in the form of overall hatching inclined to the right. The pixel size in the MDI
series is about five times as large as in the La Palma series; accordingly, = 2 proves to
be the optimum value for the detection of contours. The contour-detection results obtained
at this value are shown in panel b. Generally, the 1.2 MDI resolution is insufficient for our
algorithmic processing, and the code fails to detect the concentric-ring structures visible in
panel a (and marked in Figure 1 in Getling (2006)). However, in some areas of the image,
trenching patterns can be detected with certainty. A pattern of parallel ridges and trenches
is highlighted most distinctly within the black box in the upper right quadrant of the field of
view, where it appears highly ordered.
5. Conclusion
We have seen that the processing of time-averaged granulation images by the COLIBRI
code is capable of detecting systems of ridges and trenches in the relief of the brightness
field, which vary in their characteristic scale, geometry, and topology. The white lanes that
highlight ridges are as a rule parallelled by one or more other white neighbouring lanes and
at least two black lanes that mark trenches. Thus, our most general and remarkable finding
is the fact that trenching patterns, irrespective of their particular geometries, are virtually
ubiquitous in the averaged granulation images. The patterns may be more or less regular,
and their patchy appearance is typical, but the property that they include alternating parallel
white and black contours appears to be universal (provided the images analysed by us are
representative).
The detection of objects differing in their sizes and other properties is possible if the parameters of image processing are properly chosen. Of particular importance is the adjustment
of the parameter to the characteristic width of ridges and trenches.
The patterns analysed here can most naturally be interpreted in terms of the structure of
convective flows in the subphotospheric layers. If this interpretation is correct, the light
ridges in the brightness field should correspond to material upflows, whereas the dark
trenches can be associated with downflows. Accordingly, the parallelism of alternating
ridges and trenches should be typical of roll convection, and, in the framework of our interpretation, roll motions seem to be virtually ubiquitous in the upper convection zone and
34
Figure 7 Processing of a simulated granulation field: (a) a two-hour average of the pattern obtained in
simulations by Rieutord et al. (2002) and the results of contour detection in this averaged image with filtering
at (b) p = 0.01, = 8, l = 20 and (c) at p = 0.002 0.01, = 16, l = 20.
photosphere of the Sun. However, under solar conditions, the hypothetical roll patterns are
much more intricate in their structure than the roll patterns observed in laboratory experiments.
The outer diameters of the systems of concentric rings (annular rolls) are of the order
of the supergranular or mesogranular sizes, so such closed-ring systems could be an imprint
of the fine structure of the supergranular or mesogranular flows in the subsurface layers.
The validity of our interpretation of the trenching patterns could be definitely stated only
based on a comprehensive quantitative analysis. At the moment, the amount of information
available is not yet sufficient for such an analysis and conclusion. Our aim here was merely
to demonstrate, by representative examples, that the trenching patterns are widespread and
diverse in their particular forms and that our algorithm can efficiently detect them. Nevertheless, we can now present here an illustration that appears to be an additional argument in
favour of our interpretation.
We have averaged a series of images obtained in numerical simulations of granular-scaled
solar convection by Rieutord et al. (2002). The computations covered a domain that corresponded to a 30 30 Mm2 area of the solar photosphere. Each image contains 315 315
pixels of size 95.24 km, and the interval between images corresponds to 20 seconds of real
time. A two-hour-averaged image is shown in Figure 7a; Figures 7b and 7c represent the
results of processing the average using the COLIBRI code. It can easily be seen that, for
widely ranging (from 8 to 16 pixels), the code does not highlight any clear-cut ridges or
trenches. The resulting image is not sensitive to variations in p over a range in which trenching was revealed in real solar images. It is thus fairly obvious that granular-sized polygonal
convection cells cannot produce trenching patterns by themselves, and larger scaled convection must be responsible for the arrangement of granules that results in trenching.
The local brightness of a time-averaged image of the solar granulation reflects the local
probability of the emergence of granules. The light blotches in the image indicate the sites
where granules emerge most frequently. Qualitative reasoning (Getling, 2000) and correlation analyses (Getling, 2006) suggest that granules may be overheated blobs entrained by
the convective circulation, and they can even repeatedly emerge on the solar surface. In this
case, granules appear as markers of roll convective flows, which can thus be identified in
averaged photospheric images.
Acknowledgements We are indebted to P.N. Brandt and R.A. Shine for making available the La Palma
and SOHO MDI data, to T. Roudier for putting the results of numerical simulations at our disposal, and to
35
V.S. Berdnikov and V.A. Markov for providing the experimental photograph of a dislocation. We are also
grateful to the referee and to L.M. Alekseeva for valuable comments on the manuscript. This work was
supported by the Russian Foundation for Basic Research (Project Nos. 04-02-16580-a and 07-02-01094-a).
References
Baudin, F., Molowny-Horas, R., Koutchmy, S.: 1997, Astron. Astrophys. 326, 842.
Berrilli, F., Del Moro, D., Consolini, G., Pietropaolo, E., Duvall, T.L. Jr., Kosovichev, A.G.: 2004, Solar
Phys. 221, 33.
Brandt, P.N., Getling, A.V.: 2004, In: Stepanov, A.V., Benevolenskaya, E.E., Kosovichev, A.G. (eds.) MultiWavelength Investigations of Solar Activity, IAU Symp. 223, Cambridge University Press, Cambridge,
231.
Dialetis, D., Macris, C., Muller, R., Prokakis, T.: 1988, Astron. Astrophys. 204, 275.
Getling, A.V.: 1998, Rayleigh Bnard Convection: Structures and Dynamics, World Scientific, Singapore
(Russian version: 1999, URSS, Moscow).
Getling, A.V.: 2000, Astron. Zh. 77, 64 (English transl. Astron. Rep. 44, 56).
Getling, A.V.: 2006, Solar Phys. 239, 93.
Getling, A.V., Brandt, P.N.: 2002, Astron. Astrophys. 382, L5.
Hoekzema, N.M., Brandt, P.N.: 2000, Astron. Astrophys. 353, 389.
Hoekzema, N.M., Brandt, P.N., Rutten, R.J.: 1998, Astron. Astrophys. 333, 322.
Lisle, J.P., Rast, M.P., Toomre, J.: 2004, Astrophys. J. 608, 1167.
Mann, H.B., Whitney, D.R.: 1947, Ann. Math. Stat. 18, 50.
Muller, R., Roudier, T., Vigneau, J.: 1990, Solar Phys. 126, 53.
Rast, M.P.: 2002, Astron. Astrophys. 392, L13.
Rieutord, M., Ludwig, H.-G., Roudier, T., Nordlund, ., Stein, R.: 2002, Nuovo Cimento 25, 523.
Roudier, T., Malherbe, J.M., November, L., Vigneau, J., Coupinot, G., Lafon, M., Muller, R.: 1997, Astron.
Astrophys. 320, 605.
Roudier, T., Lignires, F., Rieutord, M., Brandt, P.N., Malherbe, J.M.: 2003, Astron. Astrophys. 409, 299.
Salov, G.I.: 1997, Avtometriya, No. 3, 60 (English transl. Optoelectronics, Instrumentation and Data Processing).
Shine, R.A., Simon, G.W., Hurlburt, N.E.: 2000, Solar Phys. 193, 313.
Simon, G.W., Brandt, P.N., November, L.J., Scharmer, G.B., Shine, R.A.: 1994. In: Rutten, R.J., Schrijver,
C.J. (eds.) Solar Surface Magnetism, NATO Advanced Science Institute 433, Kluwer Academic, Dordrecht, 261.
Whitney, D.R.: 1951, Ann. Math. Stat. 22, 274.
Principal Components and Independent Component

Analysis of Solar and Space Data
A.C. Cadavid J.K. Lawrence A. Ruzmaikin

Abstract Principal components analysis (PCA) and independent component analysis (ICA)
are used to identify global patterns in solar and space data. PCA seeks orthogonal modes of
the two-point correlation matrix constructed from a data set. It permits the identification of
structures that remain coherent and correlated or that recur throughout a time series. ICA
seeks for maximally independent modes and takes into account all order correlations of
the data. We apply PCA to the interplanetary magnetic field polarity near 1 AU and to the
3.25R source-surface fields in the solar corona. The rotations of the two-sector structures of
these systems vary together to high accuracy during the active interval of solar cycle 23. We
then use PCA and ICA to hunt for preferred longitudes in northern hemisphere Carrington
maps of magnetic fields.
Keywords Methods: statistical Sun: magnetic field
1. Introduction
Principal components analysis (PCA) and independent component analysis (ICA) seek to
identify global patterns in sets of images, whether these are spatial images, such as magnetograms, or segments of time series as in solar-wind data.
PCA searches for orthogonal modes of the two-point correlation matrix constructed from
a data set and permits the identification of structures that remain coherent and linearly correlated or that recur throughout a time series. These modes, or empirical orthogonal functions
(EOFs), are ordered according to the degree of linear correlation with the whole data set.
A.C. Cadavid () J.K. Lawrence
Department of Physics and Astronomy, California State University, 18111 Nordhoff Street, Northridge,
CA 91330-8268, USA
e-mail: ana.cadavid@csun.edu
A. Ruzmaikin
Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena,
CA 91109, USA
DOI: 10.1007/978-0-387-98154-3_5
37
38
A.C. Cadavid et al.
The first benefit of the procedure is that it may allow the summary or capture of the essential nature of the data set by using a significantly reduced subset. The second benefit is that
the procedure will lead to EOFs that are amenable to physical interpretation, that is, that
it will extract some hidden meaning from the data. For example, the EOFs may be eigenmodes of the underlying physical process. This second form of success is rarer, but there
are techniques to help it along. In particular, the structures obtained from PCA in general
are not independent because of possible higher-order correlations. ICA addresses this point
by seeking maximally independent modes, taking into account all-order correlations of the
data. This approach recombines a reduced set of EOFs to produce a new set of maximally
non-Gaussian, or maximally independent, images called independent components.
Previously, we have applied PCA to the large-scale, axisymmetric magnetic field and
related this to the solar dynamo (Lawrence, Cadavid, and Ruzmaikin, 2004). Subsequently,
we applied ICA to these fields and found common periodicities between their time variations and the characteristic periodicities in the solar wind and interplanetary magnetic field
(Cadavid et al., 2005a, 2005b). In the present paper we will provide details of these methods
and apply them to new problems.
In Section 2 we will apply PCA to two different, but physically related, data sets: the
interplanetary magnetic field (IMF) polarity at 1 AU and the 3.25R source-surface fields
from magnetic images of the Sun. We will show that similar PCA modes can be found in
the two data sets and that they can be related to demonstrate a close physical connection. In
Section 3 we shall apply PCA and ICA to magnetic activity in a particular latitude band in
the Suns northern hemisphere in a search for the presence of active longitudes. The EOFs
are not easily interpretable, but the independent components focus on specific longitudes.
A close look at the time variations of the importance of these independent components in the
data shows some periodic recurrence of activity lasting from two to five hours. Our results
are summarized and conclusions are drawn in Section 4.
2. Principal Components Analysis

2.1. Mathematical Background
The purpose of PCA is to reduce the original data set of two or more sequentially observed variables by identifying a small number of meaningful components or modes
(e.g., Jackson, 2003). The method therefore permits the identification of coherent structures as dominant or recurring patterns in a time series. To start, the data are organized
in an m n matrix X T (x, t) = [u1 un ], where the ui (x) vectors describe the m pixels
in an image observed at n different times. To be precise in this discussion, and without
loss of generality, we can assume that the data matrix has dimension m n with rank
r n. The data are then centered by subtracting the average from each image. The formal procedure consists in solving
the eigenvalue problem for the two-point m m co
variance matrix C(x, x ) = t X T (x, t)X(x , t). In general the number of eigenvalues will
be the same as the rank r n of the data matrix. When dealing with linearly independent observations we obtain r = n eigenvalues l . Otherwise, the number of eigenvalues
is equal to the rank of the data matrix. The corresponding m-dimensional eigenvectors
l (x) are identified
as the spatially dependent EOFs. The alternative n n covariance
matrix C (t, t ) = x X(x, t)X T (x, t ) leads to the same eigenvalues with corresponding
n-dimensional orthogonal vectors el (t). These are identified as the principal components
Principal Components and Independent Component Analysis
39
Figure 1 Interplanetary
magnetic field polarity, in
Geocentric Solar Ecliptic
coordinates, at the Wind and ACE
spacecraft. Each of the 131
rasters represents 655 hours or
27.292 days; roughly a
Carrington period. Light (dark)
pixels indicate that the field
points inward toward (outward
from) the Sun.
(PCs) that describe the time dependence of each mode. The data can then be decomposed in
terms of the modes as
X T (x, t) =
n

l elT (t)l (x).
(1)
l=1
If the power of the sources in the data is stronger than that of the noise, the sequence (spectrum), of ordered eigenvalues will exhibit a clear drop or breakpoint at a particular mode.
The dominant EOFs associated with the large eigenvalues before the breakpoint will mainly
characterize the data and will carry a large percentage of the variance. The remaining EOFs
will contain transient fluctuations or noise. Upon identification of the leading modes it is
possible to describe the system in terms of this lower number of degrees of freedom or
coherent structures.
The n EOFs and PCs can be organized in an m n matrix (x) = [1 n ] and an
n n matrix E(t) = [e1 en ], respectively, is the n n diagonal matrix of eigenvalues
and L is an n n diagonal matrix with Ll = l in its entries. In the general case Ll > 0
for 1 l r and Ll = 0 for (r + 1) < l n. With these definitions, and working with
normalized eigenvectors, the following relations are satisfied:
C = ,
C E = E,
(2)
X T = LE T .
(3)
and
Working with centered data and normalized eigenvectors, we can identify the last equation with the singular value decomposition (SVD) of the data matrix (Elsner and Tsonis,
1996). In this context, the columns of , called the left singular vectors, form an orthonormal basis. The rows of E T are the principal components, which are now identified as the
right singular vectors.
40
A.C. Cadavid et al.
Figure 2 The leading 20

eigenvalues in the spectrum
obtained in the PCA of the
interplanetary magnetic field.
The eigenvalues give the percent
of the total variance accounted
for by each mode.
Figure 3 The first two empirical

orthogonal functions obtained in
the PCA of the IMF polarity near
1 AU. These sine-wave structures
with quarter-wave shift indicate a
precessing, two-sector mode.
2.2. Application of PCA to Heliospheric and Solar Data

2.2.1. Solar Wind IMF Longitude
As an illustration of the techniques, we will apply them to two related data sets and then
use them to make a comparison. The first data set that we consider is BL , the longitude
of the average IMF. We use the hourly averaged data (longitude angle) from the Wind and
ACE spacecraft near 1 AU (from the OMNI2 Web site http://omniweb.gsfc.nasa.gov/). In
Figure 1 we show these hourly data in rasters of m = 655 pixels, corresponding to 27.2917
days. They differ in duration by 23.616 minutes from the solar synodic Carrington period
(27.2753 days). The data are highly bimodal; the light (dark) colored points in Figure 1
indicate that the field points inward toward (outward from) the Sun.
In carrying out a principal components analysis of these data, we regard each of the
n = 131 individual 655-hour rasters as a separate one-dimensional image of the IMF polarity
near 1 AU. The PCA yields 131 eigenvalues; the first 20 are shown in Figure 2.
41
Figure 4 The first two principal

components from the PCA of the
interplanetary magnetic field. For
clarity in the presentation PC2
has been shifted vertically.
The leading two eigenvalues, each accounting for about 6% of the total variance of the
data, will be seen to represent a related pair. The EOFs corresponding to these modes are
shown in Figure 3.
These two EOFs consist of a pair of single-period sine waves with a 90 phase shift
between them. The corresponding PCs (Figure 4) give the time dependence of the strength
of each of these EOFs in the data.
The relative phasing of these two PCs indicates that the two modes, taken together, represent a two-sector IMF structure that rotates at other than the Carrington rate.
The modes corresponding to eigenvalues 3 and 4 in Figure 2 also form a related pair. The
corresponding EOFs consist of two two-cycle sine waves, and these correspond to a rotating
four-sector structure in the solar wind (SW) IMF. The phase lag of the two leading modes
can be estimated by associating PC1 and PC2 with sine and cosine time series and finding
the phase as an arctangent. When a jump of more than radians was encountered, we,
somewhat arbitrarily, added or subtracted 2 to make the jump less than in magnitude.
The phase drift of the IMF longitude from 1997 to 2004 is plotted as the solid circles in
Figure 7.
2.2.2. Wilcox Solar Observatory Coronal 3.25R Magnetic Carrington Maps
The second data set we will investigate here, to compare to the preceding, is the set of
Wilcox Solar Observatory Carrington maps (from http://wso.stanford.edu) projected from
the photosphere to a source surface at R S = 3.25R . The fields on this source surface indicate open magnetic fields extending into the heliosphere. They represent a source for the
quasi-static background IMF on which more dynamical and irregular phenomena such as
magnetic clouds occur. We have included n = 396 Carrington rotations (CR) 1642 2037.
Each map comprises 30 rows of 72 pixels. Rasters 1 15 span from the north pole to the
equator of the Sun. Rasters 16 30 span from the equator to the south pole. Each vertical
row represents a range of sine latitude of 1/15 = 0.066. Each horizontal pixel represents 5
of longitude or 9.092 hours of time.
The PCA of the two-dimensional images is handled by unwinding the image into an
m = 30 72 = 2160 pixel one-dimensional string and proceeding as in the previous section.
The leading eigenvalue accounts for 38% of the total variance in the data set. Eigenvalues 2
and 3, accounting for 11.5% and 9.9% of the variance, appear to form a related pair, similar
42
A.C. Cadavid et al.
Figure 5 Three leading EOFs of

the 3.25R coronal field
synoptic maps from WSO. EOF1
corresponds to the N S polar
dipole (n = 1, m = 0). EOFs 2
and 3 (including n = 1, |m| = 1)
have the form of dipoles lying in
the equatorial plane. Note that
EOFs have arbitrary sign.
to the IMF data discussed before. Eigenvalue 4 accounts for 5.1% of the variance, and the
rest decrease from there. The leading three EOFs are shown in Figure 5.
The leading EOF 1 in Figure 5 corresponds to a magnetic dipole oriented North South.
It is very close to a spherical harmonic Ynm (, ) with n = 1 and m = 0. EOFs 2 and 3 are
equivalent to dipoles in the equatorial plane with a quarter-wave shift. They are roughly
combinations of the spherical harmonics with n = 1 and m = 1. There is no n = 0 (monopole) contribution. It is to be expected that the n = 1 harmonics will be the most prominent at
r = 3.25R , owing to coronal filtering (Wang et al., 1988). This arises because the corona
behaves in such a way that the fields therein are approximately potential. If the potential is
expanded in spherical harmonics, then solving the Laplace equation in the space external
to the photosphere gives solutions that are sums of terms proportional to Ynm (, )r (n+1) .
Thus, the greater the index n the faster the contribution dies with distance (Schatten, Wilcox,
and Ness, 1969; Altschuler and Newkirk, 1969).
Figure 6 shows the PC time series for EOFs 1, 2, and 3 for the span 1976 to 2005. We
see that mode 1 just represents the polar dipole field and indicates its 20-year cycle. It has
reversed polarity near the beginning of 1980, 1990, and 2000. The PC2 and PC3 time series,
in Figure 6, show that modes 2 and 3 taken together represent a magnetic dipole in the solar
equatorial plane that is rotating about the polar axis at variable rates not always matching the
Carrington rate. Using these we estimate the phase lag just as we did for the IMF polarity
data, and the result is shown in Figure 7 as the open circles. In Figure 7 a horizontal plot
represents rotation at the Carrington rate. The negative slope indicates a shorter period, near
27.0 days.
The phase lags for the IMF and WSO data (Figure 7) were shifted to bring the two
plots together during the years 1998 2004. We can see that from mid-1998 to at least mid2004 the two phases of the coronal source-surface equatorial dipole and the IMF two-sector
structure track each other extremely closely and in considerable detail. In the period before
1998, the two plots do not track at all. It is no doubt relevant that during the quiet-Sun period
between activity cycles 22 and 23 the amplitude of the source surface dipole was relatively
43
Figure 6 Time series for the

PCs 1, 2, and 3 of the coronal
field synoptic maps from WSO.
For clarity in the presentation
PC1 and PC3 have been shifted
vertically.
Figure 7 Phase lag versus time

for PC1 and PC2 of the IMF
(solid circles) and for PC2 and
PC3 of the potential field
extrapolated, coronal magnetic
field (open circles).
weak: A (PC22 + PC32 )1/2 0.5. This amplitude increased beginning in 1998.2 to an
erratic A 3.
3. Independent Component Analysis

3.1. Mathematical Background
PCA, based on the linear correlation between pairs of data points, offers a way to extract
structures that remain spatially coherent or recur many times throughout a time series. Although required to be linearly uncorrelated, because of higher-order correlations these structures are not necessarily independent.
ICA (Hyvrinen, Karhunen, and Oja, 2001; Stone, 2004, and references therein) is based
on the assumption that source signals from different physical processes are statistically independent in the sense that the value of one signal gives no information about the values of
the others. Mathematically, two variables x and y are statistically independent if their joint
44
A.C. Cadavid et al.
probability density function (pdf) is the product of their individual pdfs:

p(x, y) = p(x)p(y).
(4)
This implies that the mean (expectation value) of any moment of the two variables,

E xpyq =
p(x, y)x p y q dx dy,
(5)
can be given by the product of the individual expectation values,

p q

E x y =
p(x)p(y)x p y q dx dy = E x p E y q .
(6)
If the variables are not independent, but are merely uncorrelated, this equality is valid only
for p = q = 1, that is, E[xy] = E[x]E[y]. The Gaussian, or normal, distribution is entirely
determined by its first two moments (mean value and standard deviation). It follows from
this that uncorrelated Gaussian variables are independent. For non-Gaussian variables, all
moments may be needed to specify the pdfs, and higher-order correlations must be taken
into account to establish independence.
Actual observations of a system are composed of mixtures of source signals. ICA aims
to extract those statistically independent signals from the observed mixtures and thus to get
information about the underlying physical processes. To achieve this, ICA makes use of the
central limit theorem. Given two or more signals with different pdfs, the pdf of the sum
of these signals tends to be more Gaussian than those of the constituent signals. ICA is a
method to extract from the more Gaussian signal mixtures the combinations with the most
non-Gaussian possible pdfs. These are identified as the independent components (ICs) of
the observations. ICA has been applied to varied problems, such as recovering individual
voices from mixtures recorded by arbitrarily located microphones in a crowded room, or
identifying centers of brain activity from brain waves recorded by multiple sensors on the
scalp.
As will be shown later, we will take the magnetic activity in each of 71 solar-longitude
bins as our variables and the values of these observed in 406 successive solar rotations as
the 71 observed signals. Each of these 71 signals will have a pdf, which is calculated from
a histogram of the 406 observed values of that signal. The goal of the ICA is to recombine
the 71 observed signals into a set of global modes (the ICs), such that the fluctuations of
the modes (the mixing vectors, or MVs) have maximally non-Gaussian pdfs. The most
non-Gaussian of these may be taken to represent source signals.
Mathematically, X(x, t) = [u1 un ]T are the observed data, which can be described by
the relation X = AS, where S is a matrix containing the unknown source signals and A is
an unknown mixing matrix that mixes the sources to give the observations. The problem
is finding the unmixing matrix (W ) such that S = W X. Each source signal can then be
obtained by taking the inner product of an unmixing vector, a row in W , with the mixed
signals in the data. To gain some insight into the inversion process, consider first the projection pursuit approach (Kruskal, 1969) in which an unmixing vector is found that extracts
the most non-Gaussian possible (maximally non-Gaussian) source signal. This source signal
is then removed from the set of mixed signals and the process is repeated.
A way to implement this procedure is to consider the kurtosis (K) of the candidate source
signal (y = wT u) as a measure of non-Gaussianity:
K=
E[(y y)
4]
3.
(E[(y y)
2 ])2
(7)
45
In this definition a Gaussian signal has K = 0. The process consists then of finding the
unmixing vector that maximizes |K| and thus finding the corresponding source signal. In
contrast to the projection pursuit approach, ICA extracts N source signals from N mixed
signals simultaneously.
The formal procedure decomposes the n m data matrix X(x, t) = [u1 un ]T into
X(x, t) = A(t)S(x),
(8)
where A = [a1 ak ] is the n k matrix of MVs, which give the time evolution of the
modes, and k is the unknown number of independent modes. S = [s1 s k ]T is the k m
matrix of ICs that give the spatial structure of the modes (Funaro, Oja, and Valpola, 2003).
Unlike PCA, ICA does not order the modes, and the amplitude of each IC and MV is defined
only up to a multiplicative constant.
Before implementing the matrix inversion, we must fix the number (k) of unknown independent components to be sought. This number can be determined by first performing a
PCA and finding a breakpoint in the eigenvalue spectrum. If there is no clear breakpoint
in the spectrum we can keep the number of leading components that carry some arbitrary
percentage of the variance (Hyvrinen, Karhunen, and Oja, 2001). With this information
it is possible to define a k n whitening matrix V = D 1/2 E T , where the diagonal matrix D has the k leading eigenvalues of and the matrix E = [e1 ek ] has the corresponding n-dimensional PCs. The data can then be rotated to the k m dimensional matrix
Z = V X = [z1 zk ]T in which the whitened m-dimensional zi vectors have unit variance
and are uncorrelated. The original problem is rewritten in the form
Z = V X = V AS = W S
(9)
and the goal is to solve for the k k orthogonal matrix W , which permits then the calculation
of the ICs and MVs in matrices S and A, respectively. The matrix inversion is based on the
FastICA method, which maximizes the non-Gaussianity of the solutions by a fixed-point
algorithm. The mathematical details of the algorithm are beyond the scope or interest of this
paper (Hyvrinen, Karhunen, and Oja, 2001).
3.2. Application to Search for Active Longitudes
3.2.1. Data Preparation and PCA
In this analysis we make use of the WSO set of Carrington maps of photospheric magnetic
fields. These are the maps that underlie the source-surface maps discussed in Section 2.2.2,
and they come in the same format. Because the photospheric images are structurally complicated we do not analyze two-dimensional images directly. First, the 396 Carrington maps
used (CR 1642 2037) were strung together end-to-end to make 30 latitudinal time series
some 30 years in length with a 9.092-hour cadence.
We considered the absolute value of the field and used subsets of these data composed of
averages of rasters 10 12, giving a single time series spanning latitudes +11.5 to +23.6
and averages of rasters 19 21 giving another time series spanning latitudes 11.5 to
23.6 . The limited latitude coverage is intended to further increase resolution in longitude and time. These particular rasters were selected because they span the latitudes that
contain most of the late solar cycle fields. High-latitude features early in a given cycle will
be missed.
46
A.C. Cadavid et al.
Figure 8 Left: Dark pixels indicate the presence of magnetic field in the solar latitude band from 11.5 to
23.6 North. The 71-pixel rasters correspond to a period of 26.90 days and span the period from 1976 to
2006. Right: Five-mode reconstruction of the data.
Because we wish to look for rotation periods, if any, that might extend over more than
one activity cycle, we divided all values in the two time series by a 27-day running average.
This puts any signals during quiet-Sun times on an equal footing with those during active
times.
The next step was to compute frequency spectra of the two time series. Because there
are some missing data, we used a Lomb Scargle periodogram program. For the northern
hemisphere data, the program yields three periods with signal strengths above the 99.9%
significance level. The strongest peak is at a period of 26.4 days, the second strongest is at
27.05 days, and the third is at 26.7 days. For the southern hemisphere data, interestingly, the
strongest peak is at a long, 28.2-day period. This is consistent with the result of Hoeksema
and Scherrer (1987) that the southern hemisphere field rotated more slowly than that of the
northern hemisphere during cycle 21. Other significant peaks are at periods shorter than 27
days. Because there has been interest in the literature (Neugebauer et al., 2000; Ruzmaikin
et al., 2001; Henney and Harvey, 2002) in a periodicity of 27.03 days seen in the solar wind,
we chose to pursue that value. From this point we consider only the northern hemisphere
data.
Because 27.03 days corresponds to 71.3 pixels in our time series, we cannot reproduce
this periodicity. We clipped the series into the closest integer number of 71 pixel rasters. This
corresponds to a 26.90-day periodicity, which will be the focus of our study. The 406 rasters
can be stacked sequentially, as in Figure 8 (left), to look for persistent features. Then, using
each of the 71-pixel rasters of Figure 8 (left) as an input vector, we carried out a PCA of the
data. Notice that in this case the number of spatial pixels (m = 71) is less than the number
of observations (n = 406). Although the data are still centered in the spatial direction, we
now obtain a spectrum of 71 eigenvalues. Figure 9 shows that breakpoints occur at modes 4
and 6. The actual forms of the corresponding EOFs and PCs were not especially instructive
(see Figure 11). We found that a five-mode reconstruction of the data, as in Equation (1),
gives the result shown in Figure 8 (right), in which the salient features of the original rasters
are clearly displayed. These first five modes carry 32% of the variance.
47
Figure 9 Eigenvalue spectrum

for the northern hemisphere
latitudinal average of the WSO
photospheric field. The
eigenvalues give the percent of
total variance explained by each
mode.
Figure 10 Phase drift (solid

line) in rotations versus date for
the 11.5 to 23.6 North latitude
band time series.
Because we have used 71-pixel rasters, a feature rotating around the Sun in 26.90 days
would show as a vertical streak in Figure 8. Slower rotating features with longer periods
would slope down and to the right. A feature rotating with period 27.03 days would cross
from left to right in 15.24 years. A feature rotating at the Carrington rate of 27.275 days
would cross in 5.23 years. Although it might not be visible in Figure 8 (left), such rotating
features can be revealed by the PCA as the presence of a pair of phased modes with close
eigenvalues, as for modes 3 and 4 in Figure 9. We have carried out an analysis of the phase
drift of this pair of modes, analogous to the analyses leading to Figure 7. We find that the
phase drift corresponds to abrupt changes from one rotation period to another at intervals
of about a year or two. The overall drift from the first rotation to the last gives an average
rotation rate of 26.55 days, near the highest peak in the Lomb periodogram as described
earlier.
To check this result we have applied a complex demodulation to this time series (e.g.,
(Bloomfield, 1976)). This technique tests a signal by calculating its phase changes relative to
a sine wave of selected constant frequency or period. To match our 71-pixel rasterization we
48
A.C. Cadavid et al.
Figure 11 (a) The first five EOFs for the PCA of the WSO +11.5 to +23.6 latitude average of WSO
photospheric Carrington rotation maps, absolute values of line-of-sight fields. EOFs 2 5 have been shifted
downward for clarity. (b) The ICs corresponding to four modes. ICs 2 4 have been shifted for clarity.
choose the test period to be 26.90 days. The phase plot is presented in Figure 10. A constant
phase indicates a period of 26.90 days; other nearby periods give different slopes; the overall
slope from first to last phase gives a period of 26.5 days (with the best fit from 1985 to 2005
giving 26.2 days). We thus find rough agreement on periods among the Lomb periodogram
main peak (26.4 days), the PC phase advance (26.55 days), and the complex demodulation
(26.5 days).
3.2.2. Independent Components
We have seen in the preceding section that, for our purposes, the WSO longitudinal data can
be adequately represented by keeping only the first five such modes. We thus achieve the
first purpose of PCA, namely data reduction, acquiring a compact way to summarize the key
data. However, the eigenvectors themselves, shown in Figure 11a, do not lend themselves
to any particular interpretation. To attempt an interpretation, we apply ICA to the data by
selecting the first four modes suggested by the first breakpoint at mode 4 in the eigenvalue
spectrum (Figure 9).
The time dependences or MVs of the independent components are shown in Figure 12.
The kurtosis of these four maximally independent, non-Gaussian sources are 0.91, 0.81,
0.73, and 0.28, so the first three are clearly the most non-Gaussian. The negative values
indicate that the pdfs corresponding to the signals are platykurtic, that is, more rectangular
than Gaussian with very short tails. Positive values of kurtosis would indicate leptokurtic
pdfs, with strong central peak and long tails. Note the general limitation 2 K < .
To interpret the plots in Figure 12 we employ wavelet power spectra. Details are given
in the widely used description by Torrence and Compo (1998). A Morlet wavelet power
spectrum of MV 1 is shown in Figure 13. Areas of stronger wavelet power are shown as
darker colors on a plot of time horizontally and time scale vertically. The dark contours
enclose areas of 95% significance with respect to AR(1) red noise with coefficient = 0.67.
The wavelet power at times within about one time scale of the end points of the data is not
reliable.
49
Figure 12 The normalized time

dependences of the four ICs in
Figure 11b. MVs 2 through 4
have been shifted down for
clarity.
Figure 13 Morlet wavelet

power spectrum of MV 1 in
Figure 12. The time is shown
horizontally in years; the time
scale, or period, of activity is
plotted vertically. The wavelet
power at times within about one
time scale of the end points of the
data is not reliable.
The greatest spectral power for MV 1 occurs in the interval 1980 1985 with characteristic time scale of about two years. The wavelet power spectrum for MV 2 (not shown)
shows significant power in the interval 1985 1990 at a scale of about two years, and that
for MV 3 (also not shown) shows significant power between 1980 and 1990 at a time scale
of about four years. The first three ICs all indicate some activity at longitude 75 with differing patterns otherwise. To the extent that ICA has truly isolated physically independent
processes, all seem to be most active early in the data string and to refer to a longitude near
75 . This corresponds to magnetic features seen in Figure 8. Note that MV 2 indicates two
active longitudes ( 75 and 240 ) on roughly opposite sides of the Sun, as is sometimes
reported.
4. Summary and Conclusions

To examine global-scale patterns of solar activity, we have described the main properties of
principal components analysis and independent component analysis and have demonstrated
applications of each to solar and solar terrestrial physics. In particular we have used PCA
to uncover related rotating modes in solar chromospheric magnetic fields and in the IMF
50
A.C. Cadavid et al.
polarity near 1 AU. During the active phase of solar cycle 23 the rotations of these two
modes track one another in impressive detail.
To look for evidence of preferred longitudes of solar activity, we created a time series
from the WSO synoptic Carrington maps of photospheric magnetic field. In doing this we
narrowed our focus to the absolute values of the fields averaged over the latitudes from
+11.5 to +23.6 . The time series was clipped into 71-pixel segments spanning 26.90 days.
The PCA of these segments indicated that the salient features of the data could be reproduced
with just the five leading modes. Further, examination of two related modes (3 and 4) gave
evidence of a rotation of the full pattern with a period of around 26.5 days. This result was
reinforced by a Lomb periodogram analysis and by complex demodulation.
Whereas PCA looks for global patterns correlated with the data, ICA searches for independent sources of the solar field. The first three of four ICs associated with the first
four PCs emphasize particular longitudes. The time variations of the ICs indicate bursts of
activity lasting from two to five years.
The PCA and ICA analyses we have described here examine sequences of individual
quantities or images. Any set of analogous data could be used in a similar way. For example,
synoptic maps of Fe XIV coronal green line emission or of coronal holes, and their modes,
can be related to SW IMF fluctuations.
We have treated the IMF data as vectors of 27 longitudes in a sequence of Bartels rotations. Instead, each vector could include simultaneous measurements of an array of different SW IMF parameters, for example, a SW density component, a temperature component,
or an IMF magnitude component. The vector would describe the state of the SW IMF as
measured at intervals of, say, one hour by ACE or other spacecraft. Modes of such vectors
can be connected to changes in geomagnetic or solar-activity indices. Time ordering of the
data vectors is not necessary either. For example, Fourier spectra of solar magnetograms can
be used as input vectors that are ordered, not by time, but by total magnetic flux in the image.
This illuminates how the size distributions of magnetic features differ in different regimes
of total flux.
Acknowledgements We are pleased to acknowledge and thank those who provided the data used in this
work. These included the NASA OMNIWEB data center and the Wind and ACE spacecraft teams. We also
made extensive use of photospheric and source-surface synoptic Carrington maps from the Wilcox Solar
Observatory at Stanford University. We also thank the Laboratory of Computer and Information Science
at Helsinki University of Technology, which offers the free FastICA software package in MATLAB (http:
//www.cis.hut.fi/projects/ica/fastica/).
References
Altschuler, M.D., Newkirk, G.: 1969, Solar Phys. 9, 131.
Bloomfield, P.: 1976, Fourier Analysis of Time Series: An Introduction, Wiley, New York.
Cadavid, A.C., Lawrence, J.K., McDonald, D.P., Ruzmaikin, A.: 2005a, Solar Phys. 226, 359.
Cadavid, A.C., Lawrence, J.K., McDonald, D.P., Ruzmaikin, A.: 2005b, In: Sankarasubramanian, K., Penn,
M., Pevtsov, A. (eds.) Large-scale Structures and Their Role in Solar Activity, ASP Conference Series
CS-346, Astronomical Society of the Pacific, San Francisco, 91.
Elsner, J.B., Tsonis, A.A.: 1996, Singular Spectrum Analysis: A New Tool in Time Series Analysis, Plenum,
New York.
Funaro, M., Oja, E., Valpola, H.: 2003, Neural Netw. 16, 469.
Henney, C.J., Harvey, J.W.: 2002, Solar Phys. 207, 199.
Hoeksema, J.T., Scherrer, P.H.: 1987, Astrophys. J. 318, 428.
Hyvrinen, A., Karhunen, J., Oja, E.: 2001, Independent Component Analysis, Wiley, New York.
Jackson, J.E.: 2003, A Users Guide to Principal Components, Wiley, Hoboken.
Kruskal, J.B.: 1969, In: Miton, R.C., Nelder, J.A. (eds.) Statistical Computation, Academic, New York, 427.
51
Lawrence, J.K., Cadavid, A.C., Ruzmaikin, A.: 2004, Solar Phys. 225, 1.
Neugebauer, M., Smith, E.J., Ruzmaikin, A., Feynman, J., Vaughan, A.H.: 2000, J. Geophys. Res. 105(A2),
2315.
Ruzmaikin, A., Feynman, J., Neugebauer, M., Smith, E.J.: 2001, J. Geophys. Res. 106(A5), 8363.
Schatten, K.H., Wilcox, J.M., Ness, N.F.: 1969, Solar Phys. 6, 442.
Stone, V.: 2004, Independent Component Analysis: A Tutorial Introduction, Bradford Books, The MIT Press,
Cambridge.
Torrence, C., Compo, G.P.: 1998, Bull. Am. Meteorol. Soc. 79, 61 (http://atoc.colorado.edu/research/wavelets/
bams_79_01_0061.pdf).
Wang, Y.-M., Sheeley, N.R., Nash, A.G., Shampine, L.R.: 1988, Astrophys. J. 327, 427.
Automatic Recognition and Characterisation

of Supergranular Cells from Photospheric Velocity Fields
H.E. Potts D.A. Diver

Abstract We have developed an exceptionally noise-resistant method for accurate and automatic identification of supergranular cell boundaries from velocity measurements. Because
of its high noise tolerance the algorithm can produce reliable cell patterns with only very
small amounts of smoothing of the source data in comparison to conventional methods. In
this paper we describe the method and test it with simulated data. We then apply it to the
analysis of velocity fields derived from high-resolution continuum data from MDI (Michelson Doppler Imager) on SOHO. From this, we can identify with high spatial resolution
certain basic properties of supergranulation cells, such as their characteristic sizes, the flow
speeds within cells, and their dependence on cell areas. The effect of the noise and smoothing on the derived cell boundaries is investigated and quantified by using simulated data.
We show in detail the evolution of supergranular cells over their lifetime, including observations of emerging, splitting, and coalescing cells. A key result of our analysis of cell internal
velocities is that there is a simple linear relation between cell size and cell internal velocity,
rather than the power law usually suggested.
Keywords Photosphere Supergranulation Granules Photospheric flow
1. Introduction
The convection processes in the Sun have been studied for many years, with the first observation of the solar granulation made by William Herschel in 1801. Granulation is a smallscale, rapid, convectional process (of 1 Mm diameter, few-minute lifetime, and 1 km s1
typical flow speed) that has been well described and modelled. More recently, larger scale
patterns with weaker flows have been observed superimposed on the basic granulation flow.
Electronic supplementary material The online version of this article

(http://dx.doi.org/10.1007/978-0-387-98154-3_6) contains supplementary material, which is available to
authorized users.
H.E. Potts () D.A. Diver
Department of Physics and Astronomy, University of Glasgow, Glasgow, G12 8QQ, UK
e-mail: hugh@astro.gla.ac.uk
DOI: 10.1007/978-0-387-98154-3_6
53
54
H.E. Potts, D.A. Diver
The clearest of these is the supergranular flow, first observed by Hart (1954) and recognised
as a cellular flow system by Leighton, Noyes, and Simon (1961) by analysing dopplergrams taken at the Mount Wilson Observatory. Supergranulation is a flow with a cellular
form, probably convectional in nature, with characteristic size around 15 30 Mm, 20-hour
lifetime, and typical flow speeds around 300 m s1 . The supergranular pattern is most dramatically seen as patterns in the line-of-sight velocity of the solar surface seen, for example,
in full-disc MDI dopplergrams from the SOHO satellite. The velocities in supergranulation,
as seen in the photosphere, are predominantly horizontal. A consequence of this is that the
supergranulation pattern in dopplergrams is most visible towards the limb of the Sun. Observing the supergranulation at higher resolutions and near disk centre requires analysis of
the horizontal photospheric flow fields. Although supergranulation has been observed for
nearly 50 years, remarkably little is known about it indeed it is not even certain if it driven
by convection, as no significant temperature gradient has been measured across the cells.
The best attempts to date at measuring any temperature variation have concluded that it
must be less than 3 K across a cell (Lin and Kuhn, 1992), and only one recent measurement
(Meunier, Tkaczuk, and Roudier, 2007) shows a measurable temperature change from cell
centre to cell boundary of 0.8 2.8 K, derived from a continuum intensity variation of 0.05
0.18%. Part of the difficulty in analysing such flows is that they are masked by the much
more rapidly varying granulation flows, which have large temperature variations and higher
concentration of magnetic fields at cell boundaries.
With the advent of long-time-series, high-resolution data from satellite missions such
as SOHO and TRACE we can now track the motion within these larger structures. When
looking at the large-scale but weak motions found in supergranulation, it can be hard to
interpret the data to identify coherent structures. In this paper we present a new method to
construct supergranular boundaries with confidence from flow fields and explain how such
structures can be exploited to derive further essential diagnostic information. The article
is constructed as follows: Section 2 describes the algorithm for finding and displaying the
cell boundaries; Section 3 tests the algorithm against simulated data and compares results
with a conventional divergence algorithm; Section 4 shows how the algorithm works on real
solar data; Section 5 contains notes about the potential problems when using the method and
details how to avoid them.
2. Cell Analysis Method

This section describes a noise-tolerant way of deriving the cell structure from a flow field by
following the motion of tracker particles in the time-reversed flow field.
2.1. Data Preparation
To show the supergranulation pattern and its time variation in detail, large amounts of highresolution photospheric velocity data are required. This may be obtained by tracking the motion of the granules as they are advected by the supergranular flow fields. Long, continuous
time series of data are required to see the slow evolution of the supergranular flows, and highcadence, high-resolution images are required to resolve the granulation patterns. As a result
the only possible data source at the present time is the high-resolution (0.6 arcsec/pixel, oneminute cadence) continuum data sets from the MDI instrument on SOHO (Scherrer et al.,
1995). The images were processed to get the velocity field by tracking the barely resolved
granulation signal using the Balltrack method (Potts, Barrett, and Diver, 2003).
The Properties of Supergranular Cells
55
The data set used throughout this paper is from a 64-hour, nearly continuous run by
MDI from 15 18 January 1997. The run consists of high-resolution continuum, magnetogram, and dopplergram images taken with a one-minute cadence. We use a 480 1024
pixel (4.8 10.2 arcmin), 33-hour (2001-frame) subset of the continuum data. This set was
used because it gives the longest time series of a large area corotating with the Sun that is
available at the current time. The continuum data were rigidly derotated at a rate of 0.243
pixels/minute and filtered for p modes using a spatial fourier filter; the velocity was then
derived using Balltrack. Residual differential rotation was fitted and removed. The velocity samples obtained were binned in time with bin widths of two six hours and spatially
smoothed by convolving with a 2D Gaussian, with ranging from 1 13 Mm. The effect
of the smoothing radius and time binning on the derived cell patterns is discussed in more
detail in Sections 3 and 4. Absolute calibration of the data was achieved by adding a range
of known, small offset velocities to the raw granulation data during the derotation operation,
and measuring the effect of this on the derived velocity. A discussion of this method and the
intrinsic noise within any measurement of the flow field may be found in Potts, Barrett, and
Diver (2003).
2.2. Overview
The images in Figure 1 show the steps in generating images of the supergranular pattern. Figure 1a shows the raw velocity field derived from a Balltrack analysis. Some cell structures,
Figure 1 Stages in the identification of the cell boundaries. The data shown are a 2 2 arcmin (90 90 Mm)
region near disk centre under quiet-Sun conditions. The velocity field is averaged over three hours and spatially smoothed by convolution with a 2D Gaussian with radius = 1.75 Mm.
56
particularly the strong ones, are visible, but the cell boundaries are indistinct. Producing a
divergence map (where outflow regions are in red) of the data as shown in Figure 1b helps
to clarify the situation somewhat, although the cellular structure is still not clear
Part of the problem is that the data are intrinsically noisy: aside from observational errors,
the motion itself has a stochastic element owing to the turbulent nature of the small-scale
flow. Any local method to find the inflows and outflows that requires taking the spatial derivative of the data is dominated by the small-scale features, at the scale used to smooth the
data. This can be overcome by analysing the integral effect of the flow, as is usually done
by using cork-like tracking particles. We use the fact that the flow patterns are asymmetric: mostly point-like sources going to line-like sinks. To exploit this asymmetry we take a
regular array of starting points and send tracking particles flowing in the opposite direction
to the streamlines. All trackers that end up at the same final area (corresponding to the cell
upflow region) must be part of the same convection cell.
2.3. Method in Detail
1. Take the initial velocity field (see Figure 1a) v(x, y), reverse the flow direction, and
normalise to the maximum speed of the flow field, vn = v/vmax . It is important that
the mean velocity of the data set (normally dominated by the rotation and differential
rotation of the Sun) is very much less than the flows resulting from the supergranules.
We depend on the convergence of the tracers at the upflow regions of the cells, so careful
derotation of the data set is important. It may be useful to subtract the mean velocity or a
fitted differential-rotation correction of the whole flow field from the data at this point to
avoid artifacts. See Section 5 for more details.
2. Make a regular array of starting points at whatever final resolution is required for the cell
structure. More points will make for a higher resolution but take longer to calculate.
3. Advect the test points with the reversed flow field. The tracks for a low-resolution subset
of start points are shown in Figure 1c. We use a simple second-order predictor corrector
method for efficiency. Choose a maximum step comparable to the correlation length of
the velocity data for maximum efficiency. For maximum clarity enough time steps should
be given for a test particle to travel from the edge of a cell to its centre. This process can
be made numerically more efficient by the nonlinear scaling of the velocity field, so that
the particle takes less time to escape from the slow-moving edges of the cell. One way
to do this is to raise the speed to a power less than unity, while maintaining the direction,
for example, v n = v n vn0.5 .
4. If we take the set of the final positions (xf , yf ) of tracer particles mapped onto a grid of
their initial positions (xi , yi ), all particles that lie within the same cell will all record the
same value of final position; adjacent particles that travel to different cells will record a
different final position. Hence such a grid will contain regions in which the values change
discontinuously. The gradient of this data grid will then reveal the cell boundaries. The
quantity quantifies this process:

=
dxf
dxi
2
+
dyf
dyi
2 12
.
(1)
A plot of a suitably normalised value of , shown in Figure 1e, clearly shows the cell
boundaries. It is an exceptionally low-noise measurement of the local divergence of the
flow.
57
A property of a cell derived in this manner is that all of the tracking particles end up in a
similar area, which is the centre of the upflow for the cell. This is shown by the red blobs in
Figure 1f. Notice that the distance travelled by the tracking particles is a minimum at these
points, as can be seen in Figure 1d. These blobs are a smoothed image of the spatial density
of the final positions of the tracking particles, so each one represents a separate upflow
region. To find which cell any point on the surface belongs to, simply find out to which of
these regions the tracking particle ends up nearest. The area of any cell is proportional to the
number of tracking points that travel to this final location. In a movie that can be found in
the electronic supplement (http://dx.doi.org/10.1007/s11207-007-9068-5), and also at Potts
(2007), each cell, as identified by the upflow regions, has been given a different random
colour, and their time evolution can be clearly seen. The change in area of a few selected
cells over time is shown in Figure 6, described in more detail in Section 4.3.
3. Application to Test Data

To test the accuracy of the algorithm, it was run on test data with a known cell structure. The
results were compared with the commonly used watershed basin algorithm (DeRosa and
Toomre, 2004; Hagenaar, Schrijver, and Title, 1997), which does a local minimum search
in the divergence field of the flow. The test data were made to have similar properties to
observations of photospheric velocity fields, at a resolution equivalent to that of the highresolution output from SOHO/MDI. First, a velocity potential () was made by producing
a Voronoi tessellation of cells from randomly placed generator points, with the value inside
the cells a function of the distance to the cell centre. The flow pattern was then obtained
by taking the x and y gradients of the potential field and smoothed by convolving with a
Gaussian kernel with = 3 pixels to represent instrument effects. Noise was then added
in variable proportion and the data were smoothed again by convolution with a Gaussian
kernel of variable width.
The response of our supergranulation-finding algorithm is summarised in Figures 2 and 3,
along with the performance of the watershed basin algorithm, for comparison.
In Figure 2a, the performance of the algorithm is tested as the noise:signal ratio is increased. The top row shows the divergence of the velocity field, the centre row shows the
cell structure recovered by our velocity-based algorithm, and the bottom row shows the results from the divergence-based algorithm. It is clear that our velocity-based algorithm has
very high immunity to noise, in comparison with the divergence-based methods, returning
consistent and accurate results, even when the smoothed rms noise amplitude is similar to
that of the data. Note that the left-most plots are the zero-noise case, where both algorithms
recover the true cell structure of the noise-free test data, and so acts as a reference. In Figure 2b, the effect of increasing smoothing on noisy data is presented. The test data in this
case had fixed amplitude noise, equivalent to an rms noise:signal ratio of 0.4 when smoothed
with a radius of four pixels. As the smoothing level is varied, the divergence-based algorithm
gives much better results at higher smoothing radii. These effects are compared in more detail in Figure 3, where the effect of derived cell size as a function of smoothing is analysed.
The increased sensitivity to noise of the divergence algorithm can be explained by the
effect of differentiating the noise on the signal. If the noise has a flat spatial spectrum, the
result of spatial differentiation is to produce a noise signal where the amplitude is proportional to the wavenumber (k), resulting in greatly enhanced short-wavelength noise. The
result of this is to introduce many spurious sources and sinks into the divergence field, as
can be seen in Figure 2a. The velocity-based algorithm, in contrast, integrates the flow along
streamlines, mitigating the effect of small-scale noise.
58
Figure 2 Test data with increasing noise (top) and increasing smoothing (bottom). In each figure the top row
is the divergence field; the results from our velocity-based segmentation algorithm are shown in the centre
row and those from the watershed-basin algorithm in the bottom row. The blue dots give the true outflow
centres of the cells. Where grey shading occurs on the cell boundary maps it is because the boundary is not
clearly defined, usually because the velocity is too small or noisy.
59
Figure 3 Recovered characteristic cell dimension as a function of smoothing radius for the velocity-based
method (left) and the divergence-based method (right), for different noise:signal ratios. The dashed line shows
the misleading result of a linear extrapolation from high smoothing radii to zero smoothing.
The detailed effect of smoothing is shown in Figure 3 for the two methodologies. Both
algorithms show an average increase in the returned cell size as the smoothing radius is
increased. The true result is a value of unity on the y axis. Our velocity-based algorithm
fares much better at low smoothing radii, but it becomes more than 20% inaccurate as the
smoothing radius exceeds seven pixels. The divergence algorithm, conversely, has a complementary performance, showing poor accuracy up to a smoothing radius of 7 pixels and
above 15 pixels. Both algorithms show a linear trend in derived cell size as a function of
smoothing.
All real data have an element of smoothing from a variety of unavoidable sources, for
example, instrumental effects, seeing, and noise-reduction algorithms. To mitigate these effects in the data reduction, one approach that was first used by Hagenaar, Schrijver, and Title
(1997) and was also used by DeRosa and Toomre (2004) is to smooth the data at different
smoothing scales, and then extrapolate back to infer the true result that corresponds to the
zero-smoothing case. This assumes that the effect of smoothing is linear in the returned cell
size. Since we have test data, we can assess the efficacy of this technique. The results of
this operation are shown by the thick dashed lines in Figure 3. It is clear from this that the
trend of recovered feature size being proportional to smoothing radius is only linear at large
smoothing radii, where coincidentally the watershed-basin algorithm works well. However,
using this linear regime on this test data to extrapolate to the zero-smoothing-radius case significantly underestimates the true, underlying structure size for the test data, mainly because
the linear behaviour is not valid at small smoothing radius. This result leads us to conclude
that mean cell diameters obtained by extrapolation in this way are not necessarily secure.
4. Application to Real Data

Here we show how this noise-tolerant and accurate supergranular cell finder can be applied
to real solar data and exploited to reveal additional properties of the photospheric flow field.
The data set used here is the same as that in Section 2.1.
60
Figure 4 The effect on cell size of changing smoothing radius and time applied to high-resolution velocity
fields derived from MDI continuum data.
4.1. Cell Sizes

To measure the mean cell size in the real data, the same procedure was followed as described
for the test data in Section 3, where the data were smoothed over a range of radii. The results
are shown in Figure 4, where the derived mean cell dimension is shown as a function of
smoothing radius, for the two algorithms. There is a clear difference between these results
and those of the test data: There is no sharp fall-off at small smoothing radii. This reflects the
fact that there is no definite minimum scale for the features in the real data in comparison
to the test data. The next clear effect is that there is almost no significant variation in the
results at different smoothing time scales, showing that the supergranulation time scale must
be significantly larger than our smoothing times, as observed by other authors (Del Moro et
al., 2004; DeRosa and Toomre, 2004; Hagenaar, Schrijver, and Title, 1997).
For the divergence results, at smoothing radii less than 6 Mm, the characteristic cell size
is nearly proportional to the smoothing radius. This is the result that would be obtained
from smoothing random noise that had no true cell structure, showing that the algorithm
is dominated by noise in this region. Thereafter the results show a similar linear form to
that observed by DeRosa and Toomre (2004) at these smoothing radii, and when extrapolated back to zero smoothing yields cell dimension of around 8 Mm, smaller than that from
previous analyses of lower resolution data.
The velocity algorithm yields larger cell dimensions for a given smoothing radius and
remains linear over the whole smoothing range. Unlike the divergence results, the velocity
method shows evidence of real structure at low smoothing radii (down to an unprecedented
spatial resolution of 0.9 Mm with only two hours of temporal smoothing). The short extrapolation to zero smoothing gives a cell dimension of approximately 15 Mm, in broad
agreement with the aforementioned previous studies in the literature. As the data become
61
progressively noisier at smaller smoothing radii, this result gives a lower limit for the characteristic cell dimension, since noise can only decrease the observed dimension.
4.2. Cell Internal Speeds
With an easy way to delimit the parts of the photosphere that belong to a particular cell, it is
simple to derive other useful data. One such quantity is the speed profile within supergranular cells. The data that were used in this section are the 33-hour set described in Section 2.1,
smoothed with = 1.75 Mm and three-hour time binning. This smoothing was chosen to
be small enough such that the cells recovered were much larger than the smoothing radius,
ensuring that the smoothing did not significantly affect the shape or internal velocity profile
of the cells.
Following the analysis of Meunier et al. (2007) and Krishan et al. (2002), we looked for
a power-law relation between the cell size and internal speeds. For this we selected cells
that could be tracked with consistent areas for at least three hours (see Section 4.3), to be
certain that the cells were well defined. This left us with approximately 2000 cells, with
approximately 55% growing and 45% shrinking.
When looking for a relation such as this, the effect of the errors in the measured quantities
must be taken into account. We performed the analysis on both the rms speed within the cells
(errors add in quadrature to the true data) and the peak speed (errors add linearly to increase
the peak). To estimate the errors in the speed data we use the results of the analysis in Potts,
Barrett, and Diver (2003), who examine the velocity of the individual granules that are just
resolved in MDI continuum images, and look at the statistical effect of smoothing those data.
From this we get a minimum rms noise of 75 m s1 for the smoothing parameters used here.
In reality the noise may be greater than this because this method assumes perfect tracking
of every granule. The noise in the characteristic cell size (the square root of the area) is
related to the variation in the shapes of the cells. The cells as observed range from roughly
circular (well-defined radius) to the more commonly observed cusped shapes with aspect
ratios ranging from square to around 2:1 (see Figures 1 and 6 for examples). The centre
edge distances in a 2:1 rectangle vary from 0.6 to 1.4 times that of the radius of a circle with
the same area; this variation was used as the maximum noise. The power law was then fitted
by using an iterative scheme and taking into account the point-by-point errors. To estimate
the accuracy of the fits each was performed repeatedly on a random selection of 50% of the
data.
Our results, which take into account these errors, give a considerably stronger size
1
speed relation than that observed
in the literature. For an rms speed error of 75 m s , combined with a size error of 0.3 A, we get a power law with index 0.72 for the rms velocity
and1.06 for the maximum velocity. If we choose larger errors of 100 m s1 for the speed and
0.4 A for the radius, we get indices of 0.96 for the rms speed and 1.24 for the maximum
speed. The error on the individual power-law fits was 0.025. There was no significant
difference in the results between the growing and shrinking cell populations.
Note that these values are both close to, and centred around, an index of unity, suggesting
that there may be a simple linear relation between cell size and internal speeds. We explore
this in Figure 5, which shows the maximum speed within a cell as a function of the cell
dimension and the relation between the maximum speed and the rms speed within the cell.
The error bars shown are representative for those
used for all the data points. We chose an
rms noise of 75 m s1 and a cell size error of 0.3 A as before. The lines are two-parameter,
iterative, straight-line fits and take into account the point-by-point errors. In the left-hand
plot, if the cell velocity were proportional to the cell size, it would be expected that the y
62
Figure 5 Relation between the size of supergranular cells and their velocities for approximately 2000 cells.
The left plot shows the peak speed in each cell versus the cell dimension, with a linear relation fitted, taking
into account the indicated velocity and dimension errors. The right plot shows the relation between the cell
rms velocity and the peak velocity, again with a linear fit.
intercept of the fit would be equal to the peak value of the noise, roughly 2 times larger
than the rms noise, indicated by the horizontal dashed line in both plots. The line is a very
close fit to this, with the distribution of points in line with the chosen errors, indicating that
our choice of noise amplitude is reasonable. In the right-hand plot, the relation between the
rms speed and the maximum speed within the cells is shown, fitted in the same manner.
Notice that the error bars in this case appear too large; this is because the maximum and rms
speed errors are correlated. Again, a linear fit seems reasonable.
This result substantially disagrees with the findings of Krishan et al. (2002), who found a
weak power law with index 0.34 and interpreted this in terms of the solar convective motions
obeying the Kolmogorov hypothesis for a turbulent medium. The main reason for this is
probably the bias in their subjective selection method, in which a small number of unusually
large cells were picked by eye. Our work should be free of this bias as all consistent cells
within the analysis region were chosen, resulting in a much larger sample free of selection
effects.
Our results are much closer to those of Meunier et al. (2007), where an index of 0.66
was found. This is very near to our lower limit of 0.72 0.025, although still considerably
less than our average value of unity. The probable reason for this is the effect of the error
analysis, which significantly steepens the power-law fit. Another factor is that the cells found
in this study could all be much larger than the smoothing used, owing to the noise tolerance
of the cell-finding algorithm, avoiding any systematic errors from smoothing artifacts.
The conclusion from this is that is that a simple linear relationship between cell size
and internal speed matches the data within the limitations of the measurement errors. There
is also no difference between the behaviour of growing or shrinking cells, and there are
approximately equal numbers of both. This linear behaviour suggests that cells are selfsimilar, with the same internal velocity distribution, independent of size. The implications
of this for the energetics of the convectional process need to be addressed.
63
4.3. Tracking Supergranular Cells over Time

Identifying a cell at a particular time is interesting, but even a random velocity field, appropriately smoothed, would show some cell-like structures. Confidence that the cells obtained
are real physical phenomena is gained by showing the evolution of the supergranular pattern
over consecutive times using time-separated data sets. If the data can be processed this way,
the results allow the growth and decay of the patterns to be monitored, as in the previous
section.
There are many problems with tracking supergranular cells over time, the most significant
of which is that the cells are continuously splitting and merging. We have developed an
algorithm that tracks the upflow points of the cells though time (see Figure 1f). This allows
us to track the boundary of individual cells over long time periods (limited only by the
available data) and to store all of the branching and merging events from the entire data set.
All other data about the cells are also stored, such as the cell area, velocities, centroid, and
upflow centre.
The evolution of some representative cells found by the algorithm over a 30-hour period
are shown in Figure 6. Cell A is an unusual cell that grows rapidly over a 15-hour period
from the intersection of several cell boundaries. It is worth noting that this cell is unusually
free of magnetic field as it grows. Cell B stays very stable over the 30-hour period. Cell C
slowly breaks up and shrinks over the 30-hour period. Cells D are a pair of fairly large cells
that merge to produce a single cell.
Note that although this data set was derotated so that the mean equatorial velocity was
zero, there is a clear solar westward motion of all of the cells, corresponding to a rotation
rate that is around 30 nHz faster than the rotation of the granules themselves. This is a
well-known phenomenon (Thompson et al., 2003; Beck, 2000), but as yet there is no clear
consensus in the literature as to the underlying physics.
5. Cautionary Notes
Whenever a continuous process is studied by looking at discrete time steps, the frequency of
observation and the amount of smoothing in time can greatly influence the results. Such influences can be highly significant in the sort of data processing addressed in this article: The
flow field itself is obtained by spatially smoothing the small-scale granular motions. This
means that any derived supergranulation flow field obtained has been implicitly convolved
with whatever temporal and spatial smoothing was used in the observation. This problem
is ubiquitous in any measurement of supergranulation, whether from dopplergrams or even
from the measurements of the chromospheric network, where the lifetime of small magnetic
elements imposes a natural time scale on the data.
It is also important to be very careful about how the velocity data used to derive the cells
is derotated, owing to the small values of the velocity field ( 300 m s1 ) in comparison
to the rotation rate of the Sun (equator speed 2 km s1 ). For a weak supergranule, the
outflow speed near the edges is very small, so the apparent position of the lanes can change
considerably because of a small derotation error. Small cells can completely disappear if the
rotation offset is larger than their peak velocity. One way to help prevent this is to subtract
the mean velocity with differential rotation corrections from the velocity field as we have
done here.
64
Figure 6 Evolution of a selection of supergranular cells over a 31-hour period. Each of the images is
110 100 arcsec. Each column shows a cell at different intervals after it was first observed, shown in hours
at the top left of each frame. This data set has been corotated with the Sun to set the mean plasma (granule)
velocity to be zero, yet note the significant westward motion of the supergranules, showing that they rotate
faster than the fluid that forms them. A movie showing the time evolution of the entire data set over a 36-hour
period can be found in the electronic supplement and in Potts (2007).
6. Conclusions
We have developed a method for automatically identifying supergranulation cells, including
an accurate measure of the position of lanes between cells and the upflow centre of the cells,
without resorting to numerical differentiation of the data.
Since our method can work at exceptionally small smoothing radii, extrapolation to the
zero-smoothing-radius case is more secure than for conventional algorithms, which tend to
65
need high smoothing to avoid domination by noise. This makes our method particularly well
suited to the new generation of high-resolution solar data.
Our key physical result from this paper is that there appears to be a simple linear relation
between cell size and cell internal velocity, rather than the power law usually suggested. In
addition we can track the cells over considerable times, limited only by the source data, and
from this observe in detail their evolution. Because of the noise tolerance of the algorithm,
the smoothing radius required is only 1 2 Mm, much smaller than the size of supergranular
cells, allowing structure identification in unprecedented detail.
There are many other applications for which the process outlined in this paper will be
useful. For example, the problem of measuring any temperature differential across the cells
will be greatly reduced by having accurate positions for the cell outflow and intercell lanes
(e.g., Meunier, Tkaczuk, and Roudier, 2007). Another example is the study of the smallscale energetics of magnetic elements (Potts and Diver, 2007), whose motion is dominated
by the supergranular flows. Accurate measurements of the statistical properties of supergranulation, over both the supergranular evolution time scale and the solar cycle, will help
expand our knowledge of this poorly understood process.
The study of small-scale energetics will also benefit greatly from this method, particularly small-scale magnetic interactions, which are dominated by the solar surface flows. As
high-resolution Hinode data become available this will become a very interesting area to
study.
Acknowledgements We would like to thank Lyndsay Fletcher and Hugh Hudson for helpful discussions
that have considerably improved this paper. This work was funded in the UK at Glasgow University by
PPARC rolling grant number PP/C000234/1.
References
Beck, J.: 2000, Solar Phys. 191, 47 70.
Del Moro, D., Berrilli, F., Duvall, T.L., Kosovichev, A.G.: 2004, Solar Phys. 221, 23 32.
DeRosa, M.L., Toomre, J.: 2004, Astrophys. J. 616, 1242 1260.
Hagenaar, H.J., Schrijver, C.J., Title, A.M.: 1997, Astrophys. J. 481, 988 995.
Hart, A.B.: 1954, Mon. Not. Roy. Astron. Soc. 114, 17 38.
Krishan, V., Paniveni, U., Singh, J., Srikanth, R.: 2002, Mon. Not. Roy. Astron. Soc. 334, 230 232.
Leighton, R.B., Noyes, R.W., Simon, G.: 1961, Astrophys. J. 135, 474 499.
Lin, H., Kuhn, J.R.: 1992, Solar Phys. 141, 1 26.
Meunier, N., Tkaczuk, R., Roudier, T.: 2007, Astron. Astrophys. 463, 745 753.
Meunier, N., Tkaczuk, R., Roudier, T., Rieutard, M.: 2007, Astron. Astrophys. 461, 1141 1147.
Potts, H.E.: 2007, http://www.astro.gla.ac.uk/users/hugh/supergranulation.
Potts, H.E., Diver, D.A.: 2007, Solar Phys., in press, DOI: 10.1007/s11207-007-9021-7.
Potts, H.E., Barrett, R.K., Diver, D.A.: 2003, Solar Phys. 197, 69 78.
Scherrer, P.H., Bogart, R.S., Bush, R.I., Hoeksema, J.T., Kosovichev, A.G., Schou, J., Rosenberg, W.,
Springer, L., Tarbell, T.D., Title, A., Wolfson, C.J., Zayer, I.: 1995, Solar Phys. 162, 129 188.
Thompson, M.J., Christensen-Dalsgaard, J., Miesch, M.S., Toomre, J.: 2003, Ann. Rev. Astron. Astrophys. 41,
599 643.
Automated McIntosh-Based Classification of Sunspot

Groups Using MDI Images
T. Colak R. Qahwaji

Abstract This paper presents a hybrid system for automatic detection and McIntosh-based
classification of sunspot groups on SOHO/MDI white-light images using active-region data
extracted from SOHO/MDI magnetogram images. After sunspots are detected from MDI
white-light images they are grouped/clustered using MDI magnetogram images. By integrating image-processing and neural network techniques, detected sunspot regions are classified automatically according to the McIntosh classification system. Our results show that
the automated grouping and classification of sunspots is possible with a high success rate
when compared to the existing manually created catalogues. In addition, our system can
detect and classify sunspot groups in their early stages, which are usually missed by human
observers.
1. Introduction
The observation, analysis, and classification of sunspots form an important part in furthering
knowledge about the Sun, solar weather, and its effect on Earth (Phillips, 1992). Previous
research on solar flares showed that they are related mostly to sunspots and active regions
(Knzel, 1960; Severny, 1965; Warwick, 1966; Sakurai, 1970; McIntosh, 1990). Sunspots
are part of active regions, and their local behaviour is used for the forecast of solar activity
(Hathaway, Wilson, and Reichmann, 1994).
In this study, we present a computer platform for the automated detection, grouping, and
then classification of sunspots. In daily life, sunspot classification is mostly carried out manually by experts. This is a subjective, time-consuming, and labour-intensive process, and although classification rules are well defined, there is not always 100% unanimity in the resulting classification of sunspot groups between solar physicists even when working together.
T. Colak () R. Qahwaji
Department of Electronic Imaging and Media Communications, University of Bradford, Richmond
Road, Bradford BD7 1DP, England, UK
e-mail: t.colak@bradford.ac.uk
R. Qahwaji
e-mail: r.s.r.qahwaji@bradford.ac.uk
DOI: 10.1007/978-0-387-98154-3_7
67
68
T. Colak, R. Qahwaji
Accurate, objective classification of sunspots can solve the unanimity problem faced by
various solar observatories and space-weather-prediction groups around the world. Another
argument supporting the use of such systems is the expected increase in solar data because of
new space missions. Previous attempts at the detection of sunspots were reported by Curto,
Blanca, and Sol (2003), Zharkov et al. (2004) and Nguyen, Nguyen, and Nguyen (2005).
Zharkov et al. (2004) presented an automated system for the detection of sunspots on Ca K1
and Solar and Heliospheric Observatory (SOHO)/Michelson Doppler Imager (MDI) whitelight images. That system achieved a detection rate of 98% for MDI images when compared with the Locarno Solar Observatory detection results. Nguyen, Nguyen, and Nguyen
(2005) used image processing and clustering methods on SOHO/MDI white-light images
for the recognition and classification of sunspots according to the modified Zurich class of
the McIntosh system. Testing involved 128 sunspot groups. Although 100% correct classification rate was achieved for the modified Zurich classes C and H (25% of test data), only
60%, 19%, and 21% correct classification rates were obtained for D, E, and F (73.5% of the
test data) were obtained, respectively. Also, Curto, Blanca, and Sol (2003) used full-disk
white-light images to automatically detect and cluster sunspots into groups. Sunspots were
detected using morphological image processing techniques and neural networks were used
to classify them. However, no good results were reported for grouping. Previous research
shows that accurate detection of sunspots has been achieved on white-light solar images.
However, no good results were reported for the grouping and clustering of sunspots, which
is the main reason behind the classification errors. This is the biggest challenge facing the
creation of a fully automated and accurate sunspot classification system, as highlighted by
Nguyen, Nguyen, and Nguyen (2005) and Curto, Blanca, and Sol (2003).
In this work we present a system that uses SOHO/MDI intensitygram and magnetogram
images to detect, group, cluster, and classify sunspots based on the McIntosh classification
system. This is the first time, to our knowledge, that a computer platform has been created
to carry out this process automatically and objectively. Although MDI images are used in
this work, we believe that the principles and methods described here can be used by other
researchers for processing different solar images with few modifications.
This paper is organized as follows: The types of images used are described in Section 2.
The automated detection and grouping of the sunspots is introduced in Section 3. The classification of sunspot groups is described in Section 4, while the practical implementation and
evaluation of the algorithms is reported in Section 5. Finally, concluding remarks are given
in Section 6.
2. Data Description
In this study, SOHO/MDI intensitygram images are used for sunspot detection, while
SOHO/MDI magnetogram images are used for the detection of active regions. All of the
images used are downloaded from the MDI website (http://soi.stanford.edu/) and they are
in Graphics Interchange Format (GIF) format. Unlike the FITS images, GIF images do not
contain a header file with observational information. Although using FITS images can decrease the error rate in the calculations and save processing times, we choose to use GIF
images in order to combine this system, in the very near future, with the automated flares
prediction system described by Qahwaji and Colak (2007). This hybrid system will download online MDI continuum images in GIF format, detect sunspots, classify them and feed
the classification results to the machine-learning system to predict whether a major flare is
likely to occur in the short term.
Automated McIntosh-Based Classification of Sunspot Groups
69
The MDI instrument on SOHO provides almost continuous observations of the Sun in
the white-light continuum, in the vicinity of the Ni I 6 767.8 photospheric absorption
line. White-light pictures show how the Sun appears to the naked eye and MDI intensitygram images are primarily used for sunspot observations. The MDI data are available in
several processed levels. The MDI images used in this research are level-2 images, which
are smoothed, filtered, and rotated (Scherrer et al., 1995). SOHO provides two to four MDI
intensitygram images per day and twice as many magnetogram images with continuous coverage since 1995.
MDI magnetogram images help in measuring the magnetic field strengths on the Suns
photosphere. The magnetogram images show the magnetic fields of the solar photosphere,
with black and white areas indicating opposite magnetic polarities. The dark areas are regions of south magnetic polarity (pointing toward the Sun) and the white regions have north
magnetic polarity (pointing outward). These images can be used for detecting active regions.
In daily life, magnetogram images are used by observatories to detect and cluster sunspot
groups. We believe that combining intensitygram images and magnetogram images will help
us to detect and cluster sunspot groups in a similar way to the observatories.
3. Sunspot Detection and Grouping

Several stages are involved in the detection and grouping of sunspots, such as: preprocessing, initial detection of features (sunspots from intensitygrams, active regions from
magnetograms), and clustering. All of these stages are described below.
3.1. Preprocessing of MDI Images
We divided preprocessing into two stages. The first stage is applied to intensitygram and
magnetogram images and is called Stage-1 processing. This stage involves detecting the
solar disk, determining its centre and radius, calculating the solar coordinates, and filtering
irrelevant information (i.e., direction and date marks). Stage-2 processing is applied to
magnetogram images only and it is important because it enables us to correlate both MDI
images. Usually there is a time difference (usually less than 30 minutes) between magnetogram and intensitygram images, and the size of the solar disk on both images could differ.
The time difference problem has to be tackled in order to align these images and hence to
correlate them. To achieve this, magnetogram images need to be resized to have the same
centre and radius as the intensitygrams, and their rotation across the solar disk corrected.
This is very important because different magnetogram and white-light images from different observatories can then be used for sunspot grouping and classification by applying the
same conversion principle. These stages can be summarized as follows:
Stage-1:
Apply the filtering process reported by Qahwaji and Colak (2006a, 2006b). Detect the
solar disk, determine its radius and centre and create a mask.
Remove any information or marks (i.e., date and direction) from the image using the
mask created.
Calculate the Julian date by parsing the date and time information of the image from
its name (Meeus, 1998).
Using the Julian date, calculate solar coordinates (the position angle, heliographic latitude, heliographic longitude) for the image using the equations in Meeus (1998). Although the images are from the SOHO satellite, in this work the solar coordinates (the
70
position angle, heliographic latitude, and heliographic longitude) are calculated for the
Earth view. We have carried out empirical studies and this will cause less than 1% error
in our calculations which does not seem to have significant impact on the outcomes of
this research.
Stage-2:
Map the magnetogram image from heliocentric Cartesian coordinates to Carrington
heliographic coordinates.
Remap the magnetogram image from Carrington heliographic coordinates to heliocentric Cartesian coordinates by using the centre, radius, and solar coordinates of the
intensitygram image as the centre, radius, and solar coordinates.
Figure 1 shows the Stage-2 processing example for a magnetogram image that was
processed using Stage-1 processing (Figure 1a), which is first mapped to the heliographic
coordinates (Figure 1b) and then remapped to the heliocentric Cartesian coordinates using
a new radius but the same solar coordinates (Figure 1c). Figure 1b, which is represented
in heliographic coordinates, is shifted in this example for better view. The difference (Fig-
Figure 1 Images showing Stage-2 processing. a is the cleaned (Stage-1) magnetogram image in heliocentric Cartesian coordinates, b is the image in Carrington heliographic coordinates, c is the image
mapped back to heliocentric Cartesian coordinates with new values, d is the difference between a
and c.
71
ure 1d) shows the data change, which is visible especially near the solar limb. This change
is caused by the fact that the solar disk is remapped with a smaller radius.
Figure 2 shows the Stage-2 processing example for two original magnetogram images
marked as 2a, created on 27 July 2002, at 23:59, and 2b, created on 29 July 2002, at
01:35. Stage-1 and Stage-2 processing are applied to both images and the resulting images
are shown as images 2c and 2d, respectively. Figure 2a is remapped to the heliocentric
Cartesian coordinates with its previous solar coordinates and a new radius, while Figure 2b
is remapped with the solar coordinates of Figure 2a and a new radius. The result of this time
shift can be seen clearly in Figure 2d, which has an information loss towards the West of the
solar limb caused by the rotation of Sun during the 25-hour time difference.
Figure 2 Images showing the results of Stage-2 processing. a and b are magnetogram images with 25
hours difference. c is the resulting image when Stage-1 and Stage-2 processing is applied to a. d is the
resulting image when Stage-1 and Stage-2 processing are applied to b using the solar coordinate and radius
values from a. The white line going through the images is provided for showing the rotation on images.
72
3.2. Initial Detection of Solar Features

Initial detection of sunspots from intensitygram images and active regions from magnetogram images is carried out using intensity filtering and region growing methods, in a manner similar to Qahwaji and Colak (2006a).
The threshold value (Tf ) for each image is found automatically using Equation (1), where
is the mean, represents the standard deviation, and is a constant that is determined
empirically based on the type features to be detected and images:
Tf = ( ).
(1)
In order to detect sunspot candidates from intensitygram images, a threshold value is calculated using Equation (1) with the minus () sign and 2.7 as the value of . All of the solar
disk pixels are compared with this threshold value. If the intensity value of the pixel is less
than the threshold value, it is marked as a sunspot candidate.
Two threshold values have to be determined to detect the active region candidates in
magnetogram images. The first threshold is used for detecting seeds with north magnetic
polarity and the second is used for detecting seeds with south magnetic polarity. The value
of the first threshold is determined using Equation (1) with a plus (+) sign and equals two.
All pixels that have intensity values larger than this threshold are marked as active region
seeds with north polarity. In the same manner, the second threshold is determined using
Equation (1) with the minus () sign and equals two. Any pixel with an intensity value
less than this threshold is marked as an active region seed with south polarity.
To find the optimum vale of , intensive experiments are carried out by applying the
initial detection algorithm, with different values of , on many intensitygram and magnetogram images. The performance of the algorithm was subjectively analysed for each image. By changing the value, the number of candidate pixels can be increased or decreased,
which can affect the outcome of the feature-detection process. For intensitygram images an
increase in the value of will decrease the number of sunspot candidates and can cause the
missed detection of some sunspots. Also a decrease in this value will increase the number
of sunspot candidates and can increase false detections. Changing the value of affects the
detection of active regions in magnetogram images in a similar manner.
After deciding the seeds for active regions, a simple region growing algorithm is applied.
A 9 9 window is placed on every seed and every pixel inside this window that has a similar
intensity to the seeds intensity (20%) is marked as an active region candidate.
The input and output images in this stage are shown in Figure 3. In Figure 3c active
region candidates with the south polarity are marked with dark pixels and candidates with
the north polarity are marked with light pixels.
3.3. Deciding Active Regions and Grouping of Sunspots
After detecting initial candidates for sunspots and active regions, the resulting images are
combined to cluster sunspots into groups. Using this method the exact locations of the active regions and sunspots are determined and grouped. This method can be summarized as
follows:
1. Get a pixel marked as a candidate (Pspotcan ) on the sunspot candidate image (Figure 4b).
2. If the active region candidate image (Figure 4a) has an active region candidate (Pactcan )
at the same location, create a new image for active regions and mark it as an active
region (Pact ) with the same pixel value (dark or bright) of Pactcan and continue processing,
otherwise return to step 1 for processing another Pspotcan .
73
Figure 3 The results of initial detections for sunspots and active regions. a and b are magnetogram
and intensitygram images, respectively. c is the image showing active region candidates. d is the image
showing sunspot candidates.
3. On the active-region candidate image place a circle on Pact with degree radius and
mark all the Pactcan within this circular region as Pact on the newly created active region
image.
In this work, the value of is determined empirically by applying the sunspotgrouping algorithm to five solar image pairs (intensitygrams and magnetograms) that are
taken close in time. The value of is increased gradually from 1 to 15, and it is found,
using manual inspection, that the best grouping performance is achieved when = 10.
4. After processing all of the Pspotcan , the created image will show the active regions divided
into different polarity regions (Figure 4c). By training and applying a neural network
(NN) similar to the one described below, we can decide which polarity regions are coupled with each other and are part of the same active region. Using NN, the different
polarity regions that belong to the same active region will be given the same colours and
if they are not part of the same group they will be given different colours (Figure 4d).
We used a NN to combine regions of opposite magnetic polarities in order to determine the exact boundaries of sunspot groups. The NN is applied to two opposite-polarity
74
Figure 4 Deciding active regions and sunspot groups. a and b are selected areas from active region and
sunspot candidate images, c is the resulting image after growing sunspot candidates inside active region
candidates, d shows coloured regions after applying NN, e is the final active regions, and f is the image
showing final sunspots where the ones belonging to same group have the same intensity values.
Table 1 Inputs and output for NN training vector for active region decision.
Inputs
Description
Min(Aa, Ab )/Max(Aa, Ab ) Ratio of the smallest area to biggest area of regions

Iab /Aa
Ratio of intersecting area to area of first region
Iab /Ab
Ratio of intersecting area to area of second region
dlon /d
Ratio of the longitude difference between regions to distance between regions
dlat /d
Ratio of the latitude difference between regions to distance between regions
d/180
Ratio of distance between regions to 180
0.1 or 0.9
If two regions are intersected by same sunspot candidate it is 0.9 otherwise 0.1
Output
Description
0.1 or 0.9
If two regions are part of the same active regions it is 0.9 otherwise 0.1
regions to decide if they are part of the same active region or not. The NN training vector
consists of seven inputs and one output showing the relation between opposite polarity
magnetic field pairs.
In order to construct the NN training vector, first we calculate the boundaries, area
in pixels (Aa , Ab ), and centre of each region in heliographic degrees. We also calculate
the distance between the two regions in heliographic degrees (d), longitude and latitude
difference between the two regions (dlon , dlat ), and the intersecting area between the two
regions in pixels (Iab ). The calculations for input and output members of the training
vector are given in Table 1. Figure 5 shows visual descriptions for some of the terms used
in this table. Figure 5e is the final image, which is obtained by ANDing the magnified
Figures 5c and d (the corresponding area on sunspot candidate image).
The training vector is constructed using nearly 100 examples. Several experiments are
carried out to optimise the NN in a manner similar to Qahwaji and Colak (2007). It was
75
Figure 5 Visual descriptions of terms used for constructing NN training vector. a is the resulting image
after growing sunspot candidates inside active region candidates, b shows some terms used, c is one of
the magnified regions from a, and d is the corresponding area on a sunspot candidate image, e is the
result of ANDing c and d; sunspots that intersect two opposite polarity regions are shown on this image.
found that the best learning performance is obtained with a back-propagation training
algorithm and using the following NN topology: Seven input nodes, one hidden layer
with eight nodes, and one output node. For more information on NNs please refer to
Appendix.
5. Marked regions with the same colour will be counted as a part of same active region and
these regions will be combined by filling the gaps between them by marking the spaces
with the associated active region colour horizontally and vertically (Figure 4e).
6. Finally, this image will be ANDed with the original sunspot candidate image to group
the detected sunspots. In this final image every sunspot belonging to the same group will
have the same intensity values (Figure 4f).
After deciding on the active regions and sunspots, the spots belonging to same groups
are marked as detected groups (Figure 6g). All of the stages after preprocessing are shown
in Figure 6. The detected groups are then further processed for determining their McIntosh classes.
76
Figure 6 Stages in detecting and grouping sunspots. a and b are magnetogram and intensitygram images, respectively. c is the image showing active region candidates. d is the image showing sunspot
candidates. e is the resulting image when d and c are combined using region growing. f is the image created by applying NN to regions on image e, and g is the final image that shows sunspot groups
detected.
4. McIntosh Classification of Sunspot Regions

After grouping the detected sunspots, each sunspot group is classified based on the McIntosh
classification system which is the standard for the international exchange of solar geophysical data. The classification depends on the size, shape and spot density of sunspots. It is
a modified version of the Zrich classification system, which has improved definitions and
added indicators of size, stability, and complexity (McIntosh, 1990). The general form of
the McIntosh classification is Zpc, where Z is the modified Zrich class, p is the type
of penumbra on the largest spot and c is the degree of compactness in the interior of the
group.
In McIntosh (1990) the logical sequence for determining the McIntosh classification and
the type of classes for sunspot groups is explained below:
Computing the modified Zrich class Z:
77
Determine if the group is unipolar or bipolar.

Determine if a penumbra exist in any of the spots.
Determine if the spots with the penumbra are located on one end or both ends.
Calculate the length of the group in absolute heliographic degrees.
Computing the type of penumbra (largest spot) p:
Decide if the penumbra of the largest spot is rudimentary or not.
Decide if the penumbra of the largest spot is symmetric or not.
Calculate the value of the North-to-South diameter in heliographic degrees.
Computing the distribution of the sunspot c:
Determine the compactness of sunspots within the group.
Determine if there is a spot with mature penumbra in the group besides the leader and
follower.
In this research the same logical sequence is used for determining the McIntosh classification
of the sunspot groups.
4.1. Computing the Modified Zrich Class Z
As illustrated earlier, to determine the modified Zrich class we have to find the polarity,
penumbra status, and the length of the group.
The polarity of the sunspot groups is determined based on the separation between sunspots
within the group. The largest separation distance between the sunspots within the group
is calculated in heliographic coordinates. If there is a single spot or compact cluster of
spots in a group and the greatest separation is smaller than 3, the group is considered to
be unipolar; if the separation is greater, the group is considered to be bipolar.
In white-light images, large sunspots have a dark central umbra surrounded by the brighter
penumbra. In order to decide if a sunspot has penumbra or not, the mean (), standard
deviation ( ) of the detected sunspots on the original image is found and a threshold
value (Tp ) is calculated using Equation (2). Then the detected sunspot pixel values are
compared with this threshold value. If the sunspot pixel value is smaller than Tp , it is
considered to be part of the umbra; otherwise it is considered to be part of the penumbra
Tp = .
(2)
Figure 7c shows the detected umbra and penumbra areas for sunspots. After detecting the
umbra and penumbra regions, smaller sunspots within the sunspot group are searched to
determine whether they have a penumbra or not.
The length of the group is calculated by finding the distance separating both ends of the
group (i.e., longitudinal extent) in absolute heliographic degrees,
After finding all the necessary information, they are applied to a decision tree to determine
the modified Zrich class for the sunspot group.
4.2. Determining the Type of the Largest Spot p
The largest spot in a sunspot group can be classified depending on its type, size, and symmetry of its penumbra (McIntosh, 1990). The penumbra can either be rudimentary (partially
surrounds the umbra) or mature (completely surrounds the umbra) and its size is the value
of the North-to-South diameter. A rudimentary penumbra usually denotes a spot that is either forming or decaying. The symmetry of the penumbra depends on the irregularity of the
78
Figure 7 Deciding penumbra and umbra of spots on detected sunspot regions. a is the image showing
detected sunspot groups, b is the magnified area from original image and c is the same area showing
penumbra and umbra areas detected by our algorithms.
outline associated with this penumbra. A symmetric penumbra is mostly either circular or
elliptical in shape. The size of the spot can be easily calculated by finding the difference
between its north and south latitudes. However, finding the symmetry and type of penumbra
is a real challenge because it depends mostly on subjective judgment.
In this research we used another NN to determine the symmetry and maturity of each
sunspot. The NN training vector consists of nine inputs: kurtosis, standard deviation ( ),
mean (), skewness, heliographic length (Lheli ), heliographic diameter (Dheli ), heliographic
area (Aheli ), penumbra ratio, and umbra ratio, and consists of two outputs: symmetry and
maturity. Most of these features were used by the authors for the verification of solar features
in Qahwaji and Colak (2006a).
The input and output parameters of the training vector are determined as explained in Table 2. For this work we have used the back-propagation neural network because this learning
algorithm provides high degrees of robustness and generalisation in classification (Kim et
al., 2000). To find the optimum NN topology, a large number of learning experiments, in
a manner similar to Qahwaji and Colak (2007), were carried out. The performance of the
NN is tested after each experiment using the jack-knife technique. This technique randomly
divides the learning data into two sets: a training set containing 80% of the data and a testing set containing the remaining 20%, as explained by Qahwaji and Colak (2007). In this
work we have used 100 samples of learning data, each sample consists of nine inputs and
two outputs, as explained in Table 2. We found that the best performance is obtained for
the following topology: nine input nodes, one hidden layer with five hidden nodes, and two
output nodes. For more information on NNs please refer to the Appendix.
After optimisation, the NN is trained. A successful training is achieved if the normalised
system error falls below 0.001. After training is completed, the NN is tested with new inputs
that were not part of its training examples, in a manner similar to the jack-knife technique
(Fukunaga, 1990). The output of the NN is analysed to determine the maturity and symmetry
for each sunspot. If the first output of the NN is higher than 0.5, the sunspot under consideration is assumed to be symmetric; otherwise it is considered to be asymmetric. Similarly, if
79
Table 2 The input and output parameters involved in the NN training to determine the sunspot penumbra
type.
Inputs
Description
Kurtosis
The distribution measurement that shows the peakedness (broad or narrow)
/255
Normalized standard deviation value of the sunspot
/255
Normalized mean value of the sunspot
Skewness
The distribution measurement that shows distortion in a positive or negative direction
Lheli
The heliographic length of the sunspot
Dheli
N S diameter of the sunspot in heliographic degrees
Aheli
Area of sunspots in heliographic degrees
Ppen /Apixel
Ratio of number of pixels that are part of the penumbra to total number of pixels on sunspot
Pumb /Apixel
Ratio of number of pixels that are part of the umbra to total number of pixels on sunspot
Output
Description
0.1 or 0.9
If the sunspot is symmetric output is 0.9, if it is asymmetric output is 0.1
0.1 or 0.9
If the sunspot is mature output is 0.9, if it is rudimentary output is 0.1
the second output of the NN is higher than 0.5 the sunspot under consideration is assumed
to be mature; otherwise it is assumed to be rudimentary.
In addition we determined the North South diameter of the largest sunspots by calculating the longitude and latitude of the uppermost and lowermost pixels and then calculating
their distances. Depending on the output from the NN, the second class of the McIntosh
classification system is determined.
4.3. Determining the Sunspot Distribution c
The sunspot distribution depends on the compactness of the sunspot group (McIntosh,
1990). In order to analyze the sunspot distribution within the group, the following steps
are followed:
Find the boundaries of the sunspot group.

Calculate the area of the group in pixels within the calculated boundaries.
Calculate the total area of the individual spots in pixels.
Find the ratio (R) of the total spot area to the group area.
Calculate the number of spots with mature penumbra.
The sunspot distribution type for all the unipolar sunspot groups are X. As for the bipolar
sunspots, the classification depends on R. If R is less than 50%, then the sunspot group
is assumed to be open (McIntosh, 1990). If R is higher than 50% and the number of
spots with mature penumbra is greater than two, the sunspot is assumed to be compact;
otherwise it is intermediate.
5. Implementation and Evaluation
5.1. Practical Implementation of the System
A computer platform using C++.Net was created for the automated detection and classification of sunspots using SOHO/MDI intensitygram and magnetogram images in the GIF
80
format. A publicly available library corona.dll1 is used for reading all of the GIF images.
The program for training and applying the NN is also created and implemented in C++.
The whole system works with 1 024 1 024 images and the detection of sunspots, detection of active regions and classification of sunspot groups takes approximately four seconds
per image depending on the complexity of features. The processing time is measured on
P4-2.8 GHz PC with 1 GB RAM.
Our system was tested on a total of 103 intensitygram and 103 magnetogram images
available from 1 May 2001 until 31 May 2001. Using these images, we created our own
catalogue that consists of sunspot groups and their classifications, which will be referred to
as the Automated Sunspot Catalogue (ASC) for the rest of this text.
5.2. The Evaluation of the ASC
ASC is compared with the publicly available sunspot catalogues from the National Geophysical Data Center (NGDC).2 NGDC keeps records of data from several observatories
around the world and holds one of the most comprehensive publicly available databases for
solar features. Different observatories provide sunspot classification data at different times.
Sometimes there could be three or four SETs of data within a single day in a NGDC
catalogue, which are provided by several observatories. This makes the NGDC sunspot catalogue suitable for comparison with our ASC. We refer to SET as all of the sunspots
grouping data that are provided for a specific time in a day. Approximately four SETs are
available on ASC per day. Its frequency depends on the availability of MDI images. As ASC
is formed by processing 103 images with different dates or times, it has 103 SETs.
For testing the accuracy of ASC, we created a testing program in C++ that will read
both catalogues and compare sunspots group data sets according to date, time, location, and
classification. This testing program allows us to increase the amount of comparison data, by
controlling the time difference for comparing the SETs available in the NGDC catalogue
and ASC. This program work as follows:
Read the first SET available from ASC (SETASC ) and calculate its time.
Calculate the time difference between every SET available on NGDC catalogue
(SETNGDC ) and SETASC .
If the time difference between SETASC and SETNGDC is less than the desired time difference (DT ) continue to the next step. Otherwise return to the beginning and do not
take this SETASC into account for comparison.
Get sunspot group data from the SETASC and compare its location with all of the sunspots
grouping data in SETNGDC .
If any of the sunspot group location SETNGDC and SETASC matches, mark this group
and compare the classifications.
If none of the locations match, mark the sunspot group on SETASC as unmatched.
Repeat the previous step for the sunspot groups within SETASC .
Repeat all of the steps for all of the SETASC in ASC.
We run our testing program by setting the times of DT to 30 minutes, 1 hour, 1 hour and
30 minutes, 2 hours, 3 hours, 6 hours, 12 hours, and 1 day. Ideally, the DT between SETs
from ASC and NGDC catalogue should be zero for an accurate comparison but, as can be
1 http://corona.sourceforge.net/.
2 ftp://ftp.ngdc.noaa.gov/STP/SOLAR_DATA/,lastaccess:2007.
81
Table 3 Evaluation of sunspot grouping for different DT values.

DT
Total number
Number of
Total sunspot
Total sunspot
Total number of
(hour)
of SETs
matched
groups on
groups on
matched sunspot
SETs
NGDC SETs
ASC SETs
groups
FRR
FAR
22.9%
0.5
103
19
155
179
138
10.9%
103
25
195
225
174
10.8%
22.7%
1.5
103
37
293
350
261
10.9%
25.4%
103
47
389
439
330
15.2%
24.8%
103
57
529
535
413
21.9%
22.8%
103
70
738
656
505
31.6%
23.0%
12
103
96
847
898
605
28.6%
32.6%
24
103
103
948
957
618
34.8%
35.4%
Table 4 Evaluation of McIntosh subclasses for different DT values.

DT
(hour)
Total number
Number of
Correct Z
Correct p
Correct c
of matched sunspot
matched SETs
ratio
ratio
ratio
72.5%
groups
0.5
138
19
64.5%
47.1%
174
25
63.8%
43.7%
73.0%
1.5
261
37
62.8%
47.1%
73.2%
330
47
63.3%
47.0%
75.2%
413
57
62.0%
45.5%
75.8%
505
70
58.8%
44.6%
73.3%
12
605
96
56.2%
42.6%
69.6%
24
618
103
53.9%
42.4%
67.6%
seen from Table 3, even when DT is made equal to 30 minutes, the number of matching
SETs is 19 out of 103. These 19 SETs, corresponding to 179 individual sunspot groups, are
not enough for an accurate evaluation. Table 3 shows the results for the evaluation of sunspot
grouping. In order to evaluate the grouping performance, the following two error rates are
introduced (Hong and Jain, 1997):
The false acceptance rate (FAR), which is the percentage of a non-sunspot group being
detected as a sunspot group.
The false rejection rate (FRR), which is the percentage of a sunspot group not being
detected because it is considered to be a non-sunspot group.
Table 3 shows that the best results for FRR and FAR are achieved when DT is set equal to
1 hour and 30 minutes. After 2 hours difference, FRR and FAR rates increase dramatically.
Also, Table 4 shows the evaluation results for our automated McIntosh classification for
each DT setting. In this table, Z represents the modified Zurich class, p represents the type
of largest sunspot, and c represents the distribution of the group. The best classification
results are achieved up to a maximum of two hours which is logical when we take into
account that the change of classification usually takes a few hours.
82
6. Discussions, Conclusions and Future Work

6.1. Discussions and Concluding Remarks
To the best of our knowledge, this is the first time that a complete automated system for the
detection, grouping, and then classification of sunspots is presented. The system provides
reliable and speedy performance. This system processes two types of images simultaneously: SOHO/MDI intensitygram images and magnetogram images. Intensitygram images
are processed to detect and gather information about sunspots, while magnetogram images
are processed to provide the active-region information that is used later to group the detected
sunspots.
The system is tested on 103 MDI intensitygram images for the month of May 2001,
with a total of 957 sunspot groups and compared with the NGDC sunspot catalogue that
are created by solar physicists from different observatories. A program is created using C++
to provide correct evaluation for our system by comparing the sunspots reported in NGDC
catalogue with the ones generated in our ASC within the time difference specified. The
time difference is increased gradually with the program and the results for comparison are
recorded and shown in Tables 3 and 4. These tables show that an accurate evaluation for
sunspot grouping can be achieved for a time difference that extends up to 1 hour and 30
minutes, and an accurate classification can be achieved for a time difference that extends up
to 2 hours.
If we take into account the 1 hour and 30 minute time difference for analysing our algorithms, this means that the number of images in our test pool is reduced to 37 with a total
of 350 sunspot groups. For the exact period of time, the NGDC sunspots catalogue contains
293 recorded sunspot groups and 261 of them are matched with our 350 detected sunspot
groups by simply comparing their locations and timing information. This means that there
is nearly a 90% correct match for sunspot groups between the two catalogues and 25% of
the groups detected by our algorithms are not reported on the NGDC catalogue. More than
85% of the sunspot groups that are not available on the NGDC catalogue are the sunspot
groups with one or two sunspots. We believe that this difference can be caused by:
(a) Wrong sunspot detection by our algorithms.
(b) Wrong grouping of sunspots by our algorithms.
(c) Missed detections of sunspots by solar experts.
Although almost 99% of the detected sunspots are correct, we found that there are some
missed detections of very small sunspots (smaller than three pixels). All of the initial sunspot
candidates are compared with their corresponding magnetic activity on magnetograms images. This reduces the probability for wrong detection of sunspot candidates. This also
shows that most of the errors are caused by wrong grouping of our algorithms and/or missed
detections of sunspots by observatories.
Our algorithms clustered some sunspots into separate groups despite the fact that they
belong to the same group. This applies in particular to sunspots that are separated by large
distances compared to their areas. This causes their magnetic traces to be separated from
each other and as a result the NN clusters them as separate groups. Sometimes two or three
small sunspots that are part of the same group can be clustered as two or three different
sunspot groups.
Also, lack of visibility by ground observatories at the time of sunspots detection and human error (some small sunspots are very hard to determine by human eye) can cause the
missed detection of sunspots. Furthermore, sunspots forming or decaying can be hard to detect. We came across some examples where some sunspot groups are detected by our system
83
Figure 8 Comparison of grouping and classification results on ASC and NGDC catalogue.
in their early development stage and are not reported in the NGDC Sunspot Catalogue until
they have matured a little.
An example for incorrect grouping and missed detection on NGDC catalogue is shown in
Figure 8. Figure 8 gives the detected groups and classification results on ASC for the images
on 1 May 2001, at 00:00 (Figure 8a) and at 06:24 (Figure 8b) and also their corresponding
SETs in the NGDC catalogue. The groups marked as 1a and 1b, which are detected as separate groups on ASC, are actually one group. Our algorithms have not managed to connect
these groups. As a result, one of the groups is counted as a wrong group in our test results
and the one that is closest to group 1 on the NGDC catalogue is counted as the matched
group. If we look at the group 6 detected and classified as AXX at 00:00 and as BXO at
06:24 on ASC, we can see that this group is mentioned only as CRO at 07:00 in the NGDC
catalogue and there is no information about this group at 00:20 in the NGDC catalogue. This
group will be counted as a wrong grouping in our test results, although it is not.
As can be seen in Table 4, for a maximum of 1 hour and 30 minutes time difference,
out of the 261 matched sunspot groups, the correct classification rates for modified Zrich
class (Z), type of largest spot (p), and group distribution (c) are 63%, 47%, and 73%, respectively. Also individual matching rates for each of these McIntosh classes are given in
84
Table 5 Evaluation of modified Zrich class classification for 1 hour 30 minutes time difference.
Modified Zurich class
Matched
14
10
21
25
18
13
63
Total
19
24
38
51
36
17
76
Distribution rate
7.3%
9.2%
14.6%
19.5%
13.8%
6.5%
29.1%
Matching rate
73.7%
41.7%
55.3%
49.0%
50.0%
76.5%
82.9%
Table 6 Evaluation of largest spot class classification for 1 hour 30 minutes time difference.
Largest spot class
Matched
37
44
32
Total
43
10
109
85
12
Distribution rate
16.5%
3.8%
41.8%
32.6%
0.8%
4.6%
Matching rate
86.1%
10.0%
40.4%
37.7%
0.0%
75.0%
Table 7 Evaluation of sunspot distribution class classification for 1 hour 30 minutes time difference.
Sunspot distribution class
Matched
85
Total
95
100
132
32
Distribution rate
36.4%
50.6%
12.3%
0.8%
Matching rate
89.5%
75.8%
15.6%
50.0%
Tables 5, 6, and 7. From these tables, we can find the total number of individual classes,
their distribution rate among the total number of test groups (261), and also the number and
matching rates for individual classes.
Although the modified Zrich class ratio and group distribution ratio results are satisfactory, we cannot say the same thing for the type of the largest spot ratio. Deciding the
type of the largest spot is a very hard task, even for an experienced solar physicist, because
it involves subjective judgment on the degree of symmetry and maturity. Our system has
a classification rate of 47% for this class, but it is hard at this stage to judge whether this
is caused by the misclassification of our algorithms, which seems to be more likely, or the
misjudgement of observers. In either case, this has to be improved by adding more training
examples to the related neural networks or by applying imaging algorithms to detect the
geometry of the largest spot (i.e., Hough transform, etc.).
6.2. Future Work
For future work, we plan on improving the grouping and classification rates. Sunspot grouping can be improved by using statistical clustering algorithms for grouping in addition to
grouping with the help of the detected active regions from magnetogram images. Classification, especially for determining the type of the largest spot, has to be improved. This can
be achieved by a better training of the NN used for deciding the symmetry and type (mature
or rudimentary) of penumbra or using other machine-learning techniques, such as support
vector machines, in a manner similar to Qahwaji and Colak (2007).
85
We also plan to classify the sunspot groups according to the Mt. Wilson classification.
This can be done with higher matching ratios when we take into account that the polarity of
each sunspot can be easily determined from the magnetogram images. Our major aim is to
combine the output data from this system with a machine-learning system, as described by
Qahwaji and Colak (2007) to provide an automated platform for the short-term prediction of
major solar flares using neural networks and/or support vector machines. More information
on this can be found at http://spaceweather.inf.brad.ac.uk/.
Acknowledgements The authors thank Dr. Christopher Balch from NWS Space Environment Center,
Forecast and Analysis Branch for his advice on sunspot classification, which helped us to improve the quality
of this paper. Images used in this paper are courtesy of the SOHO/MDI project of the Stanford Lockheed
Institute for Space Research. SOHO is a project of international cooperation between ESA and NASA.
This work is supported by an EPSRC Grant (GR/T17588/01), which is entitled Image Processing and
Machine Learning Techniques for Short-Term Prediction of Solar Activity.
Appendix: Neural Networks

The term neural networks is used to describe a number of different computational models
intended to imitate biological neurons. These models consist of artificial neurons (Figure 9)
connected to each other, where the connections, also known as synaptic weights, are used to
store the knowledge.
A neural network consists of numerous artificial neurons that are arranged into layers.
Each layer is interconnected with the layer before and after it (Figure 10). The input layer is
the first layer and it receives external inputs, while the outputs are provided by the last layer.
The layers between the input and output layers are called hidden layers. There are two basic
NN topologies feed-forward and feed-backward. In the feed-forward model, information
is fed from the input layer toward the output layer, and the output of each layer is used as
the input to next layer. In feed-backward model, the output from a layer can be used as an
input to itself or to previous layers.
NNs can be trained using supervised and unsupervised learning algorithms. In unsupervised learning, the network is provided with the inputs only and the system decides how
to cluster the input data. The training of the network using inputs and their corresponding
outputs is called supervised learning. In supervised learning, each sample in the training set
Figure 9 An artificial neuron, where i represents inputs, w represents weights.
86
Figure 10 A multilayered neural network.
specifies all inputs, as well as their desired outputs. A set of examples used for training is
called the training set and samples from the training set are chosen and presented to the
network one at a time. For each sample, the outputs generated by the network and the desired outputs are compared. After processing all of the samples in the training set, the neural
weights are updated to reduce the error.
References
Curto, J.J., Blanca, M., Sol, J.G.: 2003, Solar Image Recognition Workshop. http://sol.oma.be/SIRW/
SIRWpres/Curto.PDF.
Fukunaga, K.: 1990, Introduction to Statistical Pattern Recognition, Academic Press, New York, 220.
Hathaway, D., Wilson, R.M., Reichmann, E.J.: 1994, Solar Phys. 151, 177.
Hong, L., Jain, A.: 1997, IEEE Trans. Pattern Anal. Mach. Intel. 20, 1295.
Knzel, H.: 1960, Astron. Nachr. 285, 271.
Kim, J., Owat, A., Poole, P., Kasabov, N.: 2000, Chemometr. Intel. Lab. Syst. 51, 201.
McIntosh, P.S.: 1990, Solar. Phys. 125, 251.
Meeus, J.: 1998, Astronomical Algorithms, 2nd edn. Willmann-Bell, Richmond.
Nguyen, S.H., Nguyen, T.T., Nguyen, H.S., 2005, Rough Sets, Fuzzy Sets, Data Mining, and Granular Computing 3642, Springer, Heidelberg, 263.
Phillips, K.J.H.: 1992, Guide to the Sun, Cambridge University Press, Cambridge.
Qahwaji, R., Colak, T.: 2006a, Int. J. Imaging Syst. Technol. 15, 199.
Qahwaji, R., Colak, T.: 2006b, Int. J. Comput. Appl. 13, 9.
Qahwaji, R., Colak, T.: 2007, Solar Phys. 241, 195.
Sakurai, K.: 1970, Planet Space Sci. 18, 33.
Springer, L., Tarbell, T.D., Title, A., Wolfson, C.J., Zayer, I., Akin, D., Carvalho, B., Chevalier, R.,
Duncan, D., Edwards, C., Katz, N., Levay, M., Lindgren, R., Mathur, D., Morrison, S., Pope, T., Rehse,
R., Torgerson, D.: 1995, Solar Phys. 162, 129.
Severny, A.B.: 1965, In: Lust, R. (ed.) Stellar and Solar Magnetic Fields, IAU Symp. No. 22, North-Holland,
Amsterdam, 358.
Warwick, C.S.: 1966, Astrophys. J. 145, 215.
Zharkov, S., Zharkova, V., Ipson, S., Benkhalil, A.: 2004, In: Negoita, M.G., Howlett, R.J., Jain, L.C. (eds.)
Knowledge-Based Intelligent Information and Engineering Systems, Pt 3, Proceedings, Lecture Notes
in Computer Science 3215, 446.

P.A. Conlon P.T. Gallagher R.T.J. McAteer
J. Ireland C.A. Young P. Kestener R.J. Hewett
K. Maguire

Abstract Magnetohydrodynamic turbulence is thought to be responsible for producing

complex, multiscale magnetic field distributions in solar active regions. Here we explore
the multiscale properties of a number of evolving active regions using magnetograms from
the Michelson Doppler Imager (MDI) on the Solar and Heliospheric Observatory (SOHO).
The multifractal spectrum was obtained by using a modified box-counting method to study
the relationship between magnetic-field multifractality and region evolution and activity.
The initial emergence of each active region was found to be accompanied by characteristic
changes in the multifractal spectrum. Specifically, the range of multifractal structures (Ddiv )
was found to increase during emergence, as was their significance or support (Cdiv ). Following this, a decrease in the range in multifractal structures occurred as the regions evolved to
become large-scale, coherent structures. From the small sample considered, evidence was
found for a direct relationship between the multifractal properties of the flaring regions and
their flaring rate.
Keywords Active regions Magnetic fields Turbulence
P.A. Conlon () P.T. Gallagher
Astrophysics Research Group, School of Physics, Trinity College Dublin, Dublin 2, Ireland
e-mail: conlon.paul@gmail.com
R.T.J. McAteer
Catholic University of America, NASA Goddard Space Flight Center, Greenbelt, MD 20771, USA
J. Ireland C.A. Young
ADNET Systems Inc., NASA Goddard Space Flight Center, Greenbelt, MD 20771, USA
P. Kestener
DSM/DAPNIA/SEDI, CEA Saclay, Gif-sur-Yvette, France
R.J. Hewett
Computer Science Department, University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA
K. Maguire
School of Physics, University College Dublin, Belfield, Dublin 4, Ireland
DOI: 10.1007/978-0-387-98154-3_8
87
88
P.A. Conlon et al.
1. Introduction
Solar active regions are complex concentrations of kilogauss magnetic fields in the solar
photosphere. They are highly dynamic surface features and are the source of many extreme
solar events, such as flares and coronal mass ejections (CMEs; Gallagher et al., 2007). One
of the main goals of modern solar and space-weather research is to understand the fundamental physics responsible for the evolution and activity of these regions.
Magnetic flux ropes are thought to be formed in the tachocline near the base of the Suns
convective zone (Miesch, 2005). Magnetic buoyancy causes these flux ropes to rise to the
surface and form -shaped loops (Abbett and Fisher, 2003). Turbulent flows near the surface
then cause these flux ropes to become twisted and distorted. As the flux ropes emerge into
the corona, the decreasing atmospheric pressure causes the flux ropes to expand rapidly to
form active-region coronal loops. As the smaller scale flux ropes emerge, they coalesce to
form larger scale structures (Aschwanden, 2005). The resulting magnetic-flux distribution
exhibits self-similar properties consistent with a highly turbulent and nonlinear formation
process (McAteer, Gallagher, and Ireland, 2005; Abramenko, 2005b).
The motion of magnetic elements within active regions is governed by two processes:
flux emergence and decay and bulk mass motions. Using the magnetic-induction equation,
one can obtain an expression for the magnetic Reynolds number of the form Rm = LV /,
where L, V , and are the length scales, velocities, and magnetic diffusivity on the photosphere. For typical photospheric length scales and velocities, the magnetic Reynolds number
is 106 1010 (Parker, 1979). The photosphere is therefore highly turbulent and chaotic in
nature.
Solar active-region magnetic fields are known to display power-law distributions, identifying them as nonlinear, coherent processes (Parnell, 2001; Vlahos, 2002; Abramenko,
2005b; Hewett et al., 2008). The maintenance of a power-law distribution points to selforganized criticality (SOC) in active-region emergence. The SOC nature of active regions
is a result of the turbulent nature of the flux tubes rooted in the photosphere. Under SOC,
active regions build up energy until some threshold is surpassed, and the energy is released.
With the knowledge that the regions are highly chaotic and turbulent systems, tools from
nonlinear analysis can be applied to understand the dynamics behind these energy releases.
With this motivation, several authors have investigated the self-similar nature of solar active regions (Lawrence, Cadavid, and Ruzmaikin, 1996; Abramenko et al., 2002;
Abramenko, 2005a; Georgoulis, 2005; McAteer, Gallagher, and Ireland, 2005). Numerous
methods have been used to extract parameters associated with the turbulent properties of
active regions, and some have found a relation between multifractal parameters and flaring (Abramenko, 2005b). A more extensive statistical study by McAteer, Gallagher, and
Ireland (2005) found that flaring activity was only weakly related to a regions fractal dimension. Our work expands on the methods of McAteer, Gallagher, and Ireland (2005), by
using a box-counting method to extract the multifractal spectrum of evolving active regions.
Multifractal methods are sensitive to changes across a range of scales and are excellent
tools for characterizing the complexity of active region magnetic fields. This allows for the
detection of topological changes in the distribution of active region magnetic fields during region evolution. Georgoulis (2005) concluded that fractal and multifractal parameters
involving scales across the measure space had limited ability in space-weather prediction.
Conversely, Abramenko (2005a) has shown the power of multifractal methods to detect the
difference between flaring and nonflaring regions.
In this paper, multifractal methods are used to analyze the distribution of magnetic flux in
a number of evolving active regions observed using MDI. These observations are described
89
in Section 2. In Section 3, the fractal and multifractal methods are discussed, and our results
are given in Section 4. Our conclusions and future directions are then given in Section 5.
2. Observations and Data Reduction

Magnetic-field measurements were obtained by the MDI instrument, which images the Sun
on a 1 024 1 024 pixel CCD camera through a series of increasingly narrow filters. The
final elements, a pair of tunable Michelson interferometers, enable MDI to record filtergrams
with a FWHM bandwidth of 94 m. This paper makes use of 96-minute magnetograms of
the solar disk, which had a pixel size of 2 .
For the purposes of this work, four active regions were analyzed as they evolved on the
solar disk. NOAA 10727 emerged as a region on 24 January 2005 and experienced significant flux emergence on 29 January, before rotating off disk on 3 February 2005. NOAA
10763 rotated onto the disk on 11 May 2005 as a region. It subsequently formed a
in its trailing portions and went into decay from 16 May 2005. The most active region in
the sample was the rapidly emerging active region NOAA 10488. This region emerged on
26 October 2003 to become a region on the subsequent day. The final region in the
sample was NOAA 10798, which emerged on 15 August 2005 and slowly grew in size and
complexity over a ten-day period.
Flux densities were corrected by using the algorithm suggested by Berger and Lites
(2003) and implemented in Green et al. (2003). Magnetic-field projection effects were corrected by assuming a radial field at each point on the solar disk and implementing a cosinecorrection algorithm. Image projection effects were corrected by using an equal-area cylindrical projection method (Bugayevskiy and Snyder, 1995; McAteer et al. 2005).
3. Fractals and Multifractals

Classically, the fractal dimension is calculated by covering an object with boxes of varying
size . The fractal dimension of an object is given by the scaling relation
N () D ,
(1)
where D is the fractal dimension, and N () is the number of boxes of size .

Although fractals are useful for describing the complexity of images and regions, it has
been found that most systems are a convolution of different fractal processes. To properly
characterize such systems, the idea of a multifractal was introduced. Multifractals, like fractals, relate the number of boxes (N ()) to the size of each box () and are characterized by
an equation similar to Equation (1):
N () f () ,
(2)
where instead of a single exponent, there is a spectrum of exponents f (), each with a
relative strength or significance . Information about the f () spectrum is calculated by
using a scaling exponent q. During the calculation of each -f () pair, q is used to extract
a different moment scale. In essence q acts as a microscope into the behavior at each scale of
the magnetic flux. When q is positive it magnifies the larger measures, dwarfing the smaller
ones. The opposite happens when q is negative; it inverts the measure, thus enhancing the
90
P.A. Conlon et al.
smaller measures and dwarfing the larger ones. The reader is referred to McAteer et al.
(2007) for a detailed discussion of all of the parameters.
The measure of the line-of-sight magnetograms is given by

B(k, m) k, m i of size ,
B (i) =
(3)
allowing for oppositely signed contributions to cancel within each box (Cadavid et al., 1994;
Lawrence, Ruzmaikin, and Cadavid, 1993). Thus B (i) is always positive and contains information about the mixing of oppositely signed flux elements (i.e., neutral lines). The range
of box sizes examined, from 5 to 47 Mm, allows the examination of small-scale neutral lines
and structures and restricts the analysis from larger bipolar regions that would skew the statistics.
The normalized flux (B (i)) is then given by
q
Bq (i)
q (i) = B (i)
,
B
=
N() q
z (q)
i=1 B (i)
(4)

q
where z (q) = N()
i=1 B (i) is known as the partition function.
The strength or significance of a measure is then
(q) = lim
= lim
N()
q (i) log B (i)

B
(5)
q (i) (i),
B
(6)
i=1
N()

i=1
where B (i) = . Here (q) measures the concentration strength of the range of measure
scales given by q (Chappell and Scalo, 2001). For = 2 we have a two-dimensional image,
for a sharp spike = 0.
The distribution (fractal dimension) of all the points in the image of a given concentration
strength is then given by f ():
f (q) = lim
N()
q (i) log B
(i),
B
(7)
i=1
f () = q lim
= q ,
N()
q (i) log z (q)

B
(8)
i=1
(9)
where = lim0 log E[z (q)]. E[z (q)] is the expectation value of the partition function.
The Hausdroff, or fractal, dimension is given by f [(q = 0))]; this corresponds to the peak
of the f () versus curve. Positive q corresponds to the left side of the f () versus
curve and the larger measures; the right side is the smaller measure. Smaller measures have
larger values of as they occupy or contribute more of the image.
The algorithm used was a modified box-counting method as developed by Mach and
Mas (1997). The basic method involves covering the image with a grid of inter-grid spacing and counting the number of locations that contain part of the detail (Mandelbrot and
91
Whitrow, 1983). A modification of this is to randomly sample the image and renormalize
the calculations (Cadavid et al., 1994).
Because of errors in the calculation of the multifractal spectrum for negative q, we decided to characterize the changing complexity by only using the results for positive q (the
poor resolution of the fine-scale structures being enhanced by the discrete and numerical
nature of the method; Alber and Peinke, 1998). The terms contribution diversity and dimensional diversity (Chappell and Scalo, 2001) were to used describe changes in the multifractal
spectrum; these are defined as
Cdiv = max min ,
Ddiv = f ()max f ()min ,
(10)
for q > 0.
(11)
This allows us to parameterize changes in the system over time. A broadening of the
spectrum would result from a decrease in the contribution of larger measures to an active
region. While the region grows, the larger measure will become more significant and the
contribution diversity will increase.
4. Results
4.1. Theoretical fractals
The methods described in Section 3 were tested by using a known fractal and multifractal.
The Sierpinski carpet was chosen as the fractal for testing. This was constructed by covering
the domain with a three three grid and setting the middle section to zero. This process was
continued until the smallest scale was reached (one pixel). This created a self-similar image
with a fractal dimension of D = log(8)/log(3) 1.8928. The top left of Figure 1 shows an
image of the Sierpinski carpet, with its corresponding theoretical singularity spectrum below. Using the methods outlined in the previous section, we obtained a point-like singularity
spectrum with a fractal dimension of D or f () = 1.89 0.01, with a corresponding value
of = 1.87 0.01. A monofractal has the same fractal dimension at all scales with equal
strength.
A multifractal image and its theoretical singularity spectrum were created by using the
methods of Cadavid et al. (1994). The multifractal image is shown in the center-top panel of
Figure 1. Its corresponding theoretical spectrum (dashed lines) is given in the panel below.
Also shown is the spectrum obtained by using our methods. It can be seen that there is
good agreement between the theoretical and calculated spectra for q > 0 (i.e., the left-hand
side of the spectrum). Results for the right side of the spectrum (q < 0) deviate because of
errors in the discretization of the data and numerical issues with small measures. The final
set of panels on the far right of Figure 1 shows an MDI magnetogram for NOAA 10030. Its
singularity spectrum is shown in the panel below. Because of the errors already mentioned,
only the left side of the spectrum (q > 0) could be reliably calculated.
4.2. NOAA 10488
The evolution of NOAA 10488 is shown in Figures 2 and 3. The region emerged rapidly over
three days, starting on 26 October 2005. From Figure 2, the emergence was characterized
by a significant change in the regions f () spectrum. Most notably, the contribution and
dimensional diversities become smaller as the region forms a large-scale coherent structure.
This is due to large field measures becoming more significant (i.e., moving to higher ) and
more complex (i.e., moving to larger f ()).
92
P.A. Conlon et al.
Figure 1 (Top) A monofractal image, multifractal image, and magnetogram of NOAA 10030. (Bottom) The
corresponding f () spectra.
Figure 2 The emergence of NOAA 10488 on 26 30 October 2003. (Top) MDI magnetograms. (Bottom)
The corresponding f () spectra.
93
Figure 3 The evolution of NOAA 10488. (Top panel) Contribution diversity. (Second panel) Dimensional
diversity. The total field strength (gauss) and area (Mm) of the region are shown in the third and fourth panels,
respectively. Associated C-class flares are indicated by thin arrows; bolder arrows indicate M-class flares.
Figure 3 shows the evolution of NOAA 10488s area, total flux, and contribution and
dimensional contributions over five days, from 27 October to 1 November 2007. The initial
flux emergence is characterized by a small increase in the multifractal parameters on 27
October 2005. This results from the emergence of isolated small-scale flux elements, which
subsequently coalesce to form a number of large-scale structures with reduced significance.
The multifractal properties of the region then remain constant for a number of hours as the
region continues to grow in size and total flux.
At the end of 27 October 2003, the multifractal nature of the region again begins to
change. Although the region continues to evolve rapidly, the contribution and diversity dimensions show a continued decrease, until reaching a plateau on 29 October 2005. This
gradual decrease in Ddiv and Cdiv results from an increase in the significance and fractal
dimension of larger scales as the region becomes a large-scale coherent active region. This
gradual evolution of the regions multifractal properties is accompanied by an increase in
the flaring activity of the region. The region then remains relatively stable from 29 October
2003 to 1 November 2003, during which its fractal properties do not change significantly.
4.3. NOAA 10798
The emergence of NOAA 10798 is seen in Figures 4 and 5. The initial emergence of flux
is characterized by an increase in the contribution and dimensional diversity, owing to the
dust nature of the large flux elements at this time. Similarly to NOAA 10488, the contribution diversity and dimensional diversity decrease as the region develops a coherent structure.
The region is observed to flare at 10:06 UT on 21 August 2005. This corresponds to an increased fractal dimension at larger scales.
94
P.A. Conlon et al.
Figure 4 (Top) MDI magnetogram images of NOAA 10798 on 18 August 2005 at 04:47 UT, 18 August
2005 at 14:23 UT, and 19 August 2005 at 09:35 UT. (Bottom) The corresponding f () spectrum for each
date.
diversity. (Third panel) Total field strength (gauss). (Bottom panel) Area (Mm). Associated C-class flares are
indicated by thin arrows; bolder arrows indicate M-class flares.
95
Figure 6 (Top) MDI magnetogram images for NOAA 10763 on 14 May 2005 at 19:15 UT, 15 May 2005 at
09:39 UT, and 16 May 2005 at 12:47 UT. (Bottom) f () spectra for each date.
4.4. NOAA 10763

The emergence of large flux measures in NOAA 10763 is shown in Figures 6 and 7. As the
region grows from 19:15 UT on 14 May 2005, an increase in the significance of the larger
scales is seen (min ). The dimensional diversity decreases over this time as the complexity of the larger scales increases from a dust-like nature (f () 1.0) to a line-like nature
(f () 1.0). The region flares on 15 May 2005 at 17:35 UT. This corresponds to the formation of a coherent structure with increased fractal dimensions across all scales. An M3.5 flare
is observed on 15 May 2005 at 22:27 UT. Unlike the emergence of flux in NOAA 10488
and 10798, an increase in the multifractal parameters was not present. The emergence in
this case occurred in the center of an already developed region. With this there was no initial
increase in the contribution and dimensional diversity. Instead the multifractal parameters
decrease steadily as the larger scale flux elements grow in significance and complexity.
4.5. NOAA 10727
The formation of NOAA 10727 is shown in Figures 8 and 9. As flux elements emerge and
form on the surface, the peak of the singularity spectrum (q = 0) can be seen to move to
the right (Figure 8). This is due to the increased significance of the medium scales across
the domain of interest. Similar to the formation of NOAA 10488, the initial emergence is
characterized by an increase in the contribution diversity (Cdiv ) and dimensional diversity
(Ddiv ) followed by their decrease as the region forms a coherent structure. The emergence of
additional large-scale flux on 24 January 2005 is shown in Figure 9. The fractal dimensions
of larger scales (f (min )) are seen to decrease as the regions grow, a characteristic that differs
96
P.A. Conlon et al.
Figure 7 Evolution of NOAA 10763. (Top panel) Contribution diversity. (Second panel) Dimensional diversity. (Third panel) Total field strength (gauss). (Fourth panel) Area (Mm) Associated C-class flares are
indicated by thin arrows; bolder arrows indicate M-class flares.
from NOAA 10488. NOAA 10727 failed to flare during the time of observation, which may
be due to the small fractal dimension of the larger scales (f (min )).
5. Discussion and Conclusions

The methods of Section 3 have been shown to characterize changes in the distribution of
active-region magnetic fields, during region formation and flaring. The multifractal spectrum
allows us to monitor the complexity and significance across different scales. The emergence
of small flux measurers causes a change in the singularity spectrum, as the significance of
larger flux measures is reduced. A decrease in the fractal dimension is also seen in the min
region of the f () spectrum. As such, sudden changes in the multifractal parameters are an
excellent mathematical tool for detecting characteristic changes in active regions.
A correlation between region flaring and decreases in the multifractal parameters was
evident. This corresponds with the results of Abramenko (2005a), where a significant difference in the multifractality (
h) was found between flaring and nonflaring regions. For
nonflaring regions, the larger moments of magnetic flux have a significantly lower fractal
dimension (f ()), corresponding to a reduced complexity or fill factor at this scale. A sudden change in the structure of active regions, as evident in the increased fractal dimensions
across all scales, might represent a critical transition for a active region as it begins to flare.
This further indicates that active regions have to grow in complexity and size before flaring,
corresponding with the SOC nature of active regions.
97
Figure 8 The emergence of NOAA 10727. (Top panel) MDI magnetogram images on 24 January 2005 at
06:27 UT, 24 January 2005 at 03:11 UT, and 25 January 2005 at 22:23 UT. (Second panel) The corresponding
f () spectra for each image. (Third panel) MDI magnetogram images on 29 January 2005 at 03:12 UT,
29 January 2005 at 11:15 UT, and 30 January 2005 at 06:24 UT. (Bottom panel) The f () spectra for each
magnetogram.
98
P.A. Conlon et al.
diversity. Total field strength (gauss) and area (Mm) are shown in the third and fourth panels, respectively.
There was no flare associated with NOAA 10727.
The work of McAteer, Gallagher, and Ireland (2005) found a weak correlation between
active-region flaring and monofractal scaling. As shown, active regions are highly complex
structures, with several processes across a spectrum of scales contributing to their multifractal structure. By using multiscale and multifractal methods we can separate the different
process within the range of scales given by q. Monofractal analysis of active regions returns only the Hausdroff dimension within this multifractal structure. Therefore the previous
study of McAteer, Gallagher, and Ireland (2005) was restricted in its analysis by looking at
only one scale. We plan to perform a similar study to that of McAteer, Gallagher, and Ireland (2005), on the correlation between active region flaring and multifractal properties.
Such large-scale statistical studies are needed to justify the inclusion of any parameter of
solar active regions in space-weather predictions.
We have shown that multifractal methods have the ability to detect characteristic changes
in active regions. We plan to further our understanding of the evolution of active region
with the more stable Wavelet Modulus Maximum Method (WTMM; Kestener and Arneodo,
2004; McAteer et al., 2007). The use of wavelets as fuzzy boxes removes the errors inherent
in the box-counting method. This allows us to investigate a greater range of q values and
study the evolution of small-scale magnetic moments during active region formation and
flaring. Our current method has been shown to indicate criteria favorable for region flaring;
however, by itself the method is incapable of detecting preciously the location and time of
solar flares and CMEs. Coupling this technique with the results of Hewett et al. (2008), into
power-law spatial scaling, would provide a more complete analysis tool for the detection
of extreme solar events. We plan to automate this method and incorporate it as a feature
99
on SolarMonitior (www.solarmonitor.org; Gallagher, Moon, and Wang, 2002), allowing for

real-time analysis of the multiscale properties of solar active regions.
Acknowledgements The authors thank the SOHO/MDI consortia for their data. SOHO is a joint project
by ESA and NASA. This work is supported by the NASA Living With a Star (LWS) program and the USAF
European Office of Aerospace Research and Development. Paul Conlon is an Ussher Fellow at Trinity College
Dublin. James McAteer is grateful to the NRC for the award of a Research Associateship for part of this work.
We are also grateful to the anonymous referee, whose comments helped improve this paper.
References
Abbett, W.P., Fisher, G.H.: 2003, Astrophys. J. 582, 475.
Abramenko, V.I.: 2005a, Astrophys. J. 629, 1141.
Abramenko, V.I.: 2005b, Solar Phys. 228, 29.
Abramenko, V.I., Yurchyshyn, V.B., Wang, H., Spirock, T.J., Goode, P.R.: 2002, Astrophys. J. 577, 487.
Alber, M., Peinke, J.: 1998, Phys. Rev. E 57, 5489.
Aschwanden, M.J.: 2005, Physics of the Solar Corona, Praxis, Chichester, 257.
Berger, T.E., Lites, B.W.: 2003, Solar Phys. 213, 213.
Bugayevskiy, L.M., Snyder, J.P.: 1995, Map Projections, Taylor, Philadelphia.
Cadavid, A.C., Lawrence, J.K., Ruzmaikin, A.A., Kayleng-Knight, A.: 1994, Astrophys. J. 429, 391.
Chappell, D., Scalo, J.: 2001, Astrophys. J. 551, 712.
Gallagher, P.T., Moon, Y.-J., Wang, H.: 2002, Solar Phys. 209, 171.
Gallagher, P.T., McAteer, R.T.J., Young, C.A., Ireland, J., Hewett, R.J., Conlon, P.: 2007, In: Lilensten, J.
(ed.) Space Weather: Research Towards Applications in Europe, Springer, Dordrecht, 15.
Green, L.M., Dmoulin, P., Mandrini, C.H., Van Driel-Gesztelyi, L.: 2003, Solar Phys. 215, 307.
Georgoulis, M.K.: 2005, Solar Phys. 228, 5.
Hewett, R.J., Gallagher, P.T., McAteer, R.T.J., Young, C.A., Ireland, J., Conlon, P.A., Maguire, K.: 2008,
Solar Phys. in press.
Kestener, P., Arneodo, A.: 2004, Phys. Rev. Lett. 93, 044501.
Lawrence, J.K., Ruzmaikin, A.A., Cadavid, A.C.: 1993, Astrophys. J. 417, 805.
Lawrence, J.K., Cadavid, A.C., Ruzmaikin, A.A.: 1996, Astrophys. J. 465, 425.
Mach, J., Mas, F.: 1997, http://www.qf.ub.es/d2/jordi/mfrac.html.
Mandelbrot, B.B., Whitrow, G.J.: 1983, Brit. Astron. Assoc. 93, 238.
McAteer, R.T.J., Gallagher, P.T., Ireland, J.: 2005, Astrophys. J. 631, 628.
McAteer, R.T.J., Gallagher, P.T., Ireland, J., Young, C.A.: 2005, Solar Phys. 228, 55.
McAteer, R.T.J., Young, C.A., Ireland, J., Gallagher, P.T.: 2007, Astrophys. J. 662, 691.
Miesch, M.S.: 2005, Living Rev. Solar Phys. 2(1). http://solarphysics.livingreviews.org/Articles/lrsp-2005-1/.
Parker, E.N.: 1979, Cosmical Magnetic Fields: Their Origin and Their Activity, Clarendon Press, Oxford,
43.
Parnell, C.E.: 2001, Solar Phys. 200, 23.
Vlahos, L.: 2002, In: Sawaya-Lacoste, H. (ed.) SOLMAG 2002: Proceedings of the Magnetic Coupling of the
Solar Atmosphere Euroconference, SP 505, ESA, Noordwijk, 105.

R.J. Hewett P.T. Gallagher R.T.J. McAteer
C.A. Young J. Ireland P.A. Conlon K. Maguire

Abstract Flows in the photosphere of solar active regions are turbulent in nature. Because
magnetic fields are frozen into the plasma on the solar surface, magnetograms can be used
to investigate the processes responsible for structuring active regions. Here, a continuous
wavelet technique is developed, analyzed, and used to investigate the multiscale structure of
an evolving active region using magnetograms obtained by the Michelson Doppler Imager
(MDI) onboard the Solar and Heliospheric Observatory (SOHO). The multiscale structure
was measured using a 2D continuous wavelet technique to extract the energy spectrum of
the region over the time scale of 13 days. Preliminary evidence of an inverse cascade in
active region NOAA 10488 is presented as well as a potential relationship between energy
scaling and flare productivity.
Keywords Sun: magnetic field Turbulence Wavelets Energy spectrum
R.J. Hewett ()

University of Illinois at Urbana-Champaign, Siebel Center for Computer Science, 201 N. Goodwin Ave,
Urbana, IL, USA
e-mail: rhewett2@uiuc.edu
P.T. Gallagher P.A. Conlon
Astrophysics Research Group, School of Physics, Trinity College Dublin, College Green, Dublin 2,
Ireland
R.T.J. McAteer
Institute for Astrophysics and Computational Sciences, Department of Physics, Catholic University of
America, Washington, DC 20064, USA
R.T.J. McAteer C.A. Young J. Ireland
ADNET Systems, 164 Rollins Avenue, Suite 303, Rockville, MD 20852, USA
K. Maguire
School of Physics, University College Dublin, Belfield, Dublin 4, Ireland
DOI: 10.1007/978-0-387-98154-3_9
101
102
R.J. Hewett et al.
1. Introduction
Turbulence is the result of the nonlinear components in the dynamics of fluid flows (Farge
and Schneider, 2006). The magnetic Reynolds number of a system (Rm = V L1 , where V
is the velocity scale, L is the length scale, and is the magnetic diffusivity) is the classical
measure of how turbulent that system is, with higher Reynolds numbers implying more
turbulence in a system. The energy in turbulent systems is known to scale as E(k) k ,
where k is the wavenumber or reciprocal length scale.
A turbulent system generally consists of three different regimes, all of which tend to
scale with different . In the first regime, the injective range, energy is injected into largescale structure from some external process. In the second regime, the inertial range, energy
cascades downward from large scales to small scales, following a power law, = 5/3 (Kolmogorov, 1941) or = 3/2 (Iroshnikov, 1963; Kraichnan, 1965), in an organized fashion.
In the third regime, the dissipative range, small-scale structure dissipates energy via another
external process, such as viscosity. In addition to the standard energy-cascade phenomenon,
in certain situations in 2D turbulence the inertial range is known to undergo an inverse cascade, where small-scale structure evolves to form large-scale coherent structure (Chertkov et
al., 2006). This results in energy cascading to larger scales over time, which is the opposite
of a regular cascade.
Typically, Fourier-transform methods are used to extract the energy spectrum from measurements of turbulent systems. However, there are significant drawbacks to Fourier methods. Because of the nonlocal nature of the Fourier transform, spatial information is smeared
across the phase components of all Fourier modes (Farge and Schneider, 2006). As such,
spatially localized structure is not easily extracted from the resulting transform. A wavelet
method combines the frequency-analyzing characteristics of the Fourier transform with the
ability to maintain spatial locality, thus allowing significant localized structure to be extracted from the resulting spectrum (Farge et al., 1996). Wavelet methods extract the same
energy spectrum as the Fourier transform, perhaps with some smoothing, if certain conditions are met (Farge et al., 1996). For a detailed treatise on the relation between the global
Fourier spectrum and local and global wavelet spectra, the reader is directed to Perrier,
Philipovitch, and Basdevant (1995).
Active-region formation and evolution is governed by the equations of hydrodynamics
and magnetohydrodynamics (MHD). This, coupled with the solar photospheres Reynolds
number of Rn 108 , implies that active regions are expected to show characteristics of a
turbulent system. A wavelet technique is an ideal method for extracting information critical
to the understanding of the turbulent nature of the Sun. The spectrum of MHD turbulence
is described in Boldyrev (2005, 2006). Simulations by Tao, Rosner, and Caltaneo (1995) investigate the multiscalar distribution of structure on the solar surface and conclude that this
structure should show a multifractal spectrum. Lawrence, Cadavid, and Ruzmaikin (1995)
successfully measured this multifractal spectrum with a box-counting method. McAteer,
Gallagher, and Ireland (2005) investigated the fractal nature of active-region borders and
the multifractal nature of active regions has been measured by analyzing the scaling of
structure functions (Abramenko, 2005b). It was found that flare-quiet regions tend to be
less multifractal than flare-active regions. Multifractality is also being investigated through
box-counting methods by Conlon et al. (2007). The inertial range for active regions has
been measured to at least contain scales ranging from 3 to 10 Mm, or 2 to 20 Mm under
looser restrictions (Abramenko et al., 2001). Within this range it was shown, by applying
Fourier techniques for energy-spectrum extraction to high-resolution Michelson Doppler
Imager magnetograms, that regions that do not flare maintain a spectrum near that predicted
103
by Kolmogorov (1941) whereas regions with significant flare activity exhibit a significantly
steeper spectrum of 2 < < 3 (Abramenko, 2005a). Ishizawa and Hattori (1998) used a
discrete wavelet transform method to extract the energy spectrum from 2D MHD turbulence
simulations and found two different characteristic regions of the spectrum: an = 3/2 scaling law for the turbulent region and an = 2 scaling law for the coherent region.
In this paper, a continuous wavelet method is used to extract and analyze the energy
spectrum from the Michelson Doppler Imager magnetograms. The data are described in
Section 2. The wavelet transform and energy-spectrum extraction technique are discussed
in Section 3. Results are presented in Section 4, and conclusions and future directions are
found in Section 5.
2. Observations and Data Reduction

The Michelson Doppler Imager (MDI; Scherrer et al., 1995) on the Solar and Heliospheric
Observatory (SOHO; Domingo, Fleck, and Poland, 1995) spacecraft provides solar magnetogram images. A total of 10242 pixel full-disk images are available with a cadence of one
minute throughout most of the mission. Full-disk MDI images have 4 resolution and each
pixel covers approximately 1.4 1.4 Mm.
The data used in the study are from 200 full-disk magnetograms from 22 October 2003
at 01:35:03 UT through 3 November 2003 at 09:35:03 UT. The effective cadence was 96
minutes. From these images, subimages of active region NOAA 10488, as well as the area
of quiet Sun it emerged from, were extracted and corrected for projection effects (McAteer
et al., 2005).
NOAA 10488 is perhaps the best-known example of a rapidly emerging active region.
While in the field of view, the area of the Sun goes from a quiet state to full-sized region in
a matter of a few days. Throughout this evolution, another region, NOAA 10493, appears
nearby and merges with NOAA 10488. As NOAA 10488 approaches its maximum size,
significant outflow of magnetic fields is visible on the limbward side of the region.
In addition to NOAA 10488, we used a cropped, calibrated high-resolution MDI magnetogram of NOAA 09077 taken on 12 July 2000 at 17:00:01 UT for comparison with the
results of Abramenko (2005a).
3. A Wavelet Method for Energy Spectrum Extraction

A wavelet, denoted (x) with L2 (Rn ) and x Rn , is a compactly supported squareintegrable function that satisfies the admissibility condition of Daubechies (1992):

(x)
2 x1
(1)
C =
2 dx < ,
|Rn |
where (x)
is the Fourier transform of (x) and C is the admissibility constant. This
implies that a wavelet has finite energy, and with some work, a wavelet can be shown to
have zero mean (Addison, 2002). One common real-valued wavelet, and the one used in this
paper, that satisfies these conditions is the Mexican Hat or Marr wavelet. For the Gaussian
function,
1
g(x) = e 2 x ,
(2)
104
R.J. Hewett et al.
the Mexican Hat wavelet is its Laplacian,

1 2
g(x) = n x2 e 2 (x ) = (x).
(3)
The wavelet transform requires dilation and translation of by a scale factor a R+ and by
a shift factor x Rn . In the common shorthand notation, the scaled and translated version
of is written as

x x
n
a,x (x) = a 2
,
(4)
a
where the leading term is an energy-normalization factor. The wavelet transform of a squareintegrable function, f (x), x Rn , is the inner product (convolution) of a,x (x) with f (x),

a,x (x)f (x) dx.
(5)
w(a, x ) =
Rn
The wavelet transform, like the Fourier transform, can be used to extract an energy spectrum from an image. Each point in an energy spectrum represents the sum total energy in
the result of the wavelet transform,

1
w(a, x )2 dx ,
(6)
E(a) =
C R n
over all translations, but at only a single scale. E(a) is also typically normalized by the size
or length of the signal.
The relationship between frequency and scale must also be understood. Scale a is inversely proportional to frequency f by a critical frequency factor fc (Addison, 2002). This
factor can be computed by taking the wavelet transform of a sinusoid with known period
and finding the scale at which the energy of the result of the transform is maximized (Farge
and Schneider, 2006). For the 2D Mexican Hat, fc 0.278. Then f can be converted to
wavenumber, so the resulting plot is of the standard form for energy-spectrum analysis.
The algorithm for computing the energy spectrum of an image, using the continuous
wavelet transform is as follows:
1. Let I be the target image and A be the desired set of scales.
2. For each a A,
(a) wa = a,0 I , where is the convolution operator,
(b) compute E(a) as per Equation (5) and normalize by the image size and C , and
(c) compute f = fc /a and Ef = Ea /fc .
3. Plot Ef against f to yield the energy spectrum of I .
It is here that the major advantages of the wavelet transform come into play. First, one
can select any scale (within resolution criteria) at which to analyze the signal. This is aided
by the fact that all scaled wavelets should have the same total energy (Daubechies, 1992)
so it is easy to test whether a particular scale is within the resolution of the system. This
can be used to create an automatic system for extracting coefficients for valid scales from a
signal. Second, one can specifically select scales to analyze that encompass known interesting length scales of the signal; that is, if it is known that a process occurs at specific length
scales, for instance the inertial range of a turbulent fluid, extra care can be taken to ensure
enough measurements of that region are taken. Finally, it is possible to view the individual
coefficients of the wavelet transform, which allows further analysis of specific structures at
specific length scales.
105
It should be noted that this continuous wavelet transform (CWT) technique is not as
computationally or memory efficient as some discrete wavelet transform (DWT) techniques,
but its advantages far outweigh the costs. To decrease computation time, one can explicitly
compute the wavelet function in the Fourier domain, which saves the computation of one
FFT for each desired scale. The DWT is also, typically, restricted to dyadic scales, so for an
n n image only log2 (n) data points can be extracted without appealing to more complex
algorithms. Also, again because CWT methods are not restricted to dyadic scales like most
DWT methods, the same range of physical length scales can easily be extracted from data
with different resolutions. This is important because high-resolution data are not always
readily available, but it is good to use them when they are, though higher resolution is not
required to extract useable measurements.
4. Results
Three different experiments show the power of this method. First, we illustrate the relationship between the energy spectrum of the wavelet and the energy spectrum of a signal.
Second, we show that the method behaves as expected when applied to a simulated region.
Last, we show two examples of the method applied to calibrated MDI data.
4.1. Application to a Simple 1D Signal
Figure 1 shows a simple sinusoidal signal with period 2 , its scalogram or wavelet transform
plot, and its energy spectrum. The theoretical peak in the spectrum, 1/2 , is marked with
a vertical line. Strong agreement between this and the actual peak of the energy spectrum
(dashed line) can be seen. Also, it is clear that when the energy spectrum of the wavelet is
constant, the energy spectrum is that of the sinusoid. But when the energy in the wavelet is
not constant, that is, when the Nyquist criterion is violated or the desired scale makes the
wavelets support exceed the time window, the energy spectrum is that of the wavelet, not
the signal.
4.2. Application to Simulated Magnetograms
Simulated magnetograms, created by adding Gaussian blobs to an empty image, were used
to verify that the energy spectrum was correctly reacting to variations in energy at different
scales. Simulated region and regions are shown in Figure 2, and their resulting energy
spectra are plotted in Figure 3. As expected, the region has a steeper spectrum because,
although it has nearly the same amount of large-scale energy as the , the contains
much more small-scale energy, owing to the various small Gaussian blobs.
4.3. Application to MDI Magnetograms
4.3.1. NOAA 09077: A Comparison
NOAA 09077 was a large, complex region analyzed in Abramenko (2005a). Figure 4
shows the 12 July 2000 17:00:01 UT calibrated high-resolution MDI magnetogram of this
region, cropped to 512 512 pixels. Figure 5 shows the energy spectrum extracted from
this image. The best fit in the inertial range of 3 to 10 Mm yields a slope of 2.315, which
is within the error of the slope of 2.263 0.126 measured by Abramenko (2005a) for
106
R.J. Hewett et al.
Figure 1 (a) The input signal, a simple sinusoid with period 2 . (b) The scalogram, or wavelet transform
plot, for this signal. Each row of pixels is the wavelet coefficient at the associated scale. (c) The energy
spectrum of the signal and the dilated wavelet function. The vertical line marks the location of the theoretical
peak of the energy spectrum.
the same image. Also, in the range 2 20 Mm, the best fit yields a slope of 2.225, which
again is within the error of the Abramenko (2005a) measurement of 2.281 0.15. This is
promising, as wavelet techniques should produce similar results when compared to Fourier
techniques.
4.3.2. NOAA 10488: Evidence of an Inverse Cascade
Figure 6 shows the evolution of NOAA 10488. Each column shows the calibrated image of
the region, as well as the wavelet coefficients corresponding to the upper and lower boundaries of the inertial range. The first column shows the region just after it emerged. It is
clear that there is minimal large-scale structure and significant small-scale structure. As the
region evolves, the significance of the large-scale structure clearly increases, as it is more
prominant in the 10-Mm images. It is worth noting that NOAA 10493 emerges near NOAA
10488 and begins to contribute energy to the overall spectrum. This can be seen in the
third column, where the very small region to the left contributes some energy to the smallscale coefficient but very little to the large scale. As this smaller region evolves and merges
with NOAA 10488, again, we see the influence of the large-scale-structure increase on the
10-Mm coefficient.
Throughout the evolution of NOAA 10488, Figure 6, the slope of its magnetic-energy
spectrum gradually steepens from 1 to 3 (Figure 7). The value of is only greater
107
Figure 2 (a) A simulated region, created by placing two Gaussian functions on an empty image. (b) A
simulated region, created by placing two large Gaussian blobs and numerous small Gaussian blobs on
an empty image.
Figure 3 The energy spectrum of the two images in Figure 2, plotted as a function of spatial frequency.
than three when limb effects dominate, near the end of the regions passage across the visible
disk. Also, as NOAA 10488 evolves, a distinct bulge in the energy spectrum (Figure 8)
shifts from large k (small scales) on 26 October 2003 at 04:47 UT, through the inertial range,
and stops and stays at small k (large scales) or moves out of the resolution of the data from
27 October 2003 at 14:23 UT onward. This is a migration of large amounts of energy from
108
R.J. Hewett et al.
Figure 4 A calibrated
high-resolution MDI
magnetogram of NOAA 09077
taken on 12 July 2000 at 17:00:01
UT, cropped to 512 512 pixels.
Figure 5 The energy spectrum of the image in Figure 4, plotted as a function of spatial frequency.
small to large scales that coincides with the evolution of the region as it emerges and selforganizes into a coherent structure. Hence, this is evidence of the existence of an inverse
cascade.
109
Figure 6 Top row: Calibrated, projection-corrected images of NOAA 10488. Middle row: The wavelet coefficients of these images at a length scale of 3 Mm. Bottom row: The wavelet coefficients of these images
at a length scale of 10 Mm. Columns from right to left show the evolution of NOAA 10488 over a four-day
time period.
Figure 7 The slope of the inertial range of the energy spectrum plots for 200 images of NOAA 10488 over
two weeks.
The physics of active regions implies a fundamental maximum size of a sunspot. It

is when the region grows to this size that the bulge in the energy spectrum appears
to halt. According to Chertkov et al. (2006), such a fundamental size barrier leads to
a buildup in energy at large scales and can force a spectrum that would typically fol-
110
R.J. Hewett et al.
Figure 8 The evolution of the energy spectrum of NOAA 10488 over eight measurements. This progression
shows a bulge, marked with arrows, that shifts from small scales to large scales as the region evolves. The
vertical lines in each plot represent the inertial range of 3 to 10 Mm.
low the Kolmogorov (1941) model to approach an = 3 spectrum, which is seen in this
investigation.
During the time period of 27 October 2003 at 19:11 UT to 28 October 2003 at 12:47
UT there is a sharp flattening and then sudden steeping in the slope of the inertial range.
This is visible in Figure 7. There are many events that occur in this time period that could
be the root of this. One M class and two C class flares occur in the region during this time.
The more likely explanation is the emergence of NOAA 10493 and later its merger with
10488. Also, during this time period, a period of constant energy appears at small scales and
propagates forward in time and upward in scale at a speed consistent with that of the flow of
photospheric plasma. This time period also coincides with the beginning of a large outflow
of material on the eastern side of the region.
Figures 6 and 9 show the evolution of NOAA 10488 through images and energy spectrum
plots. In the first column of Figure 6, it is clear that most of the energy is in small scales and
there is almost no energy even at the large-scale end the inertial range. This corresponds to a
very flat spectrum in Figure 9. As the region evolves and grows, more detail is apparent in the
large-scale coefficients and, as expected, the spectra grow significantly steeper. The fourth
column of Figure 6 also shows a large increase in medium-scale energy, which corresponds
to NOAA 10493s rapid growth and merger with NOAA 10488.
Figure 10 shows the energy at scales of 3 and 10 Mm for NOAA 10488. During the
period of rapid growth, the larger scale is seen to grow faster, and it continues its rapid
growth even after the rapid growth of the small scale has stopped. This is also evidence for
an inverse cascade, as energy continues to enter the large scales but the small scale does not
grow, as one would expect with a direct cascade.
111
Figure 9 The evolution of the energy spectrum of NOAA 10488 for the four times shown in Figure 6, as a
function of spatial frequency. The inertial range is marked with vertical lines.
Figure 10 The change in energy of NOAA 10488 in two length scales over two weeks time.
112
R.J. Hewett et al.
5. Conclusions
It has been shown that the energy spectrum of a signal, for example an MDI magnetogram,
can be extracted by using a continuous wavelet transform technique. Analysis of NOAA
09077 shows that this method produces results similar to that of Abramenko (2005a), while
allowing the extraction of images of particular length scales, which can be used for further
analysis. Also, analysis of simulated active regions and NOAA 10488 shows that this method
correctly preserves the sensitivity of the energy spectrum to changes in the active regions
field strength over time. Finally, the evolution of the energy spectrum gives a unique view
of the inverse energy cascade as NOAA 10488 emerges and matures.
Because this method does not seem to respond sharply to small-time-scale events such
as flares and CMEs, it alone is not sufficient for prediction of such events. In the future, statistical studies may yield a better understanding of how this method may be used to perhaps
predict such behavior. Because of these deficiencies, this method must be used in conjunction with others, such as the work of Conlon et al. (2007), to build a better picture of dynamic
events in the photosphere. As such, to create a more complete analysis package, we plan to
incorporate this work into http://SolarMonitor.org (Gallagher, Moon, and Wang, 2002) as a
real-time analysis technique.
Acknowledgements We would like to thank the SOHO/MDI consortia for their data. SOHO is a joint
project by NASA and ESA. This research was funded by NASAs Living With a Star program. Additionally,
Russell J. Hewett completed a significant portion of this work while an undergraduate at Virginia Polytechnic
Institute and State University (Virginia Tech) and would like to dedicate this paper to faculty and students
whose lives were lost during the tragedy on 16 April 2007.
References
Abramenko, V.I.: 2005a, Astrophys. J. 629, 1141.
Abramenko, V.I.: 2005b, Solar Phys. 228, 29.
Abramenko, V., Yurchyshyn, V., Wang, H., Goode, P.R.: 2001, Solar Phys. 201, 225.
Addison, P.: 2002, The Illustrated Wavelet Transform Handbook, Institute of Physics Publishing, Bristol.
Boldyrev, S.: 2005, Astrophys. J. 626, L37.
Boldyrev, S.: 2006, Phys. Rev. Lett. 96, 115002.
Chertkov, M., Connaughton, C., Kolokolov, I., Lebedev, V.: 2006, Phys. Rev. Lett. 99, 084501.
Conlon, P.A., Gallagher, P.T., McAteer, R.T.J., Ireland, J., Young, C.A., Hewett, R.J., Maguire, K.: 2007.
Solar Phys., submitted.
Daubechies, I.: 1992, Ten Lectures on Wavelets, Society for Industrial and Applied Mathematics, Philadelphia.
Farge, M., Kevlahan, N., Perrier, V., Goirand, E.: 1996, Proc. IEEE 84, 639.
Farge, M., Schneider, K.: 2006, In: Franoise, J.P., et al. (eds.) Encyclopedia of Mathematics Physics, Elsevier, Amsterdam.
Gallagher, P.T., Moon, Y.-J., Wang, H.: 2002, Solar Phys. 2, 171.
Iroshnikov, P.S.: 1963, Astron. Zh. 40, 742.
Ishizawa, A., Hattori, Y.: 1998, J. Phys. Soc. Japan 67, 441.
Kolmogorov, A.N.: 1941, R. Acad. Sci. U.S.S.R. 30, 301; 30, 538 (translated in R. Soc. Lond. Proc. Ser. A
434).
Kraichnan, R.H.: 1965, Phys. Fluids 8, 1385.
Lawrence, J.K., Cadavid, A.C., Ruzmaikin, A.A.: 1995, Phys. Rev. E 51, 316.
McAteer, R.T.J., Gallagher, P.T., Ireland, J.: 2005, Astrophys. J. 631, 628.
McAteer, R.T.J., Gallagher, P.T., Ireland, J., Young, C.A.: 2005, Solar Phys. 228, 55.
Perrier, V., Philipovitch, T., Basdevant, C.: 1995, J. Math. Phys. 36, 1506.
Scherrer, P.H., et al.: 1995, Solar Phys. 162, 129.
Tao, L., Du, Y., Rosner, R., Cattaneo, F.: 1995, Astrophys. J. 443, 434.
A Comparison of Feature Classification Methods

for Modeling Solar Irradiance Variation
H.P. Jones G.A. Chapman K.L. Harvey J.M. Pap
D.G. Preminger M.J. Turmon S.R. Walton

Abstract Physical understanding of total and spectral solar irradiance variation depends
upon establishing a connection between the temporal variability of spatially resolved solar structures and spacecraft observations of irradiance. One difficulty in comparing models
derived from different data sets is that the many ways for identifying solar features such
as faculae, sunspots, quiet Sun, and various types of network are not necessarily consistent. To learn more about classification differences and how they affect irradiance models,
feature masks are compared as derived from five current methods: multidimensional histogram analysis of NASA/National Solar Observatory/Kitt Peak spectromagnetograph data,
statistical pattern recognition applied to SOHO/Michelson Doppler Imager photograms and
magnetograms, threshold masks allowing for influence of spatial surroundings applied to
NSO magnetograms, and one-trigger and three-trigger algorithms applied to California State University at Northridge Cartesian Full Disk Telescope intensity observations. In
general all of the methods point to the same areas of the Sun for labeling sunspots and
active-region faculae, and available time series of area measurements from the methods correlate well with each other and with solar irradiance. However, some methods include larger
label sets, and there are important differences in detail, with measurements of sunspot area
K.L. Harvey and S.R. Walton are deseased, to whom this paper is dedicated.
H.P. Jones ()
National Solar Observatory, P.O. Box 26732, Tucson, AZ 85726, USA
e-mail: hjones@nso.edu
G.A. Chapman D.G. Preminger S.R. Walton
Department of Physics and Astronomy, California State University Northridge, Northridge, CA, USA
K.L. Harvey
Solar Physics Research Corporation, Tucson, AZ, USA
J.M. Pap
Goddard Earth Science and Technology Institute/UMBC, NASAs Goddard Space Flight Center,
Greenbelt, MD, USA
M.J. Turmon
NASAs Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA, USA
DOI: 10.1007/978-0-387-98154-3_10
113
114
H.P. Jones et al.
differing by as much as a factor of two. The methods differ substantially regarding inclusion of fine spatial scale in the feature definitions. The implications of these differences for
modeling solar irradiance variation are discussed.
Keywords Solar irradiance Sunspots Active regions Photosphere Data analysis
1. Introduction
Solar irradiance variation has been measured with high relative accuracy with spacecraft
radiometers since late 1978 (Pap, 2003; Frhlich, 2006; Lean, 2001). These instruments
have shown that the Suns total irradiance decreases in the presence of sunspots and increases in the presence of continuum faculae (Willson et al., 1981). The most important
discovery of these measurements was that total irradiance changes over the solar cycle
with an amplitude of 0.1%, being higher during maximum activity conditions (Willson
and Hudson, 1988). However, despite convincing demonstrations (Foukal and Lean, 1988;
Chapman, Cookson, and Dobias, 1997; Fligge, Solanki, and Unruh, 2000; Preminger, Walton, and Chapman, 2002; Krivova et al., 2003; Wenzler et al., 2006) that sunspots and faculae are the dominant sources of total solar irradiance variation on time scales of up to several
years, many questions remain.
One major problem (Pap et al., 2002) is that the uncertainty in the absolute calibration of the various radiometers exceeds the observed magnitude of solar variability. Since
there are gaps in the observing periods of various spacecraft, there is no unambiguous way
to ensure that attempts to combine the various observations into a long-term continuous
record do not confuse errors in absolute calibration and unmeasured instrumental degradation with true solar variability. Indeed, major reconstructions of spacecraft radiometer
measurements (Frhlich, 2006; Frhlich and Lean, 1998; Willson and Mordvinov, 2003;
Lean et al., 2005) differ in important details, such as whether the average irradiance at solar
minimum is approximately constant from cycle to cycle. Moreover, although the reconstructions agree that total irradiance at the maxima of cycles 21, 22, and 23 is about the
same, many observations (de Toma et al. 2001, 2004; Pap et al. 2002) suggest that photospheric magnetic flux, the contribution of photospheric faculae, and sunspot number and
area are less at the maximum for cycle 23 than they were at the corresponding phase of
cycle 22. Finally, the explained variance ( 80%) of even the best long-term models (Wenzler et al., 2006), while consistent with ground-based measurement errors, easily allows for
global or other sources of irradiance variation outside the sunspot faculae model. The uncertainties are not large over the period spanned by spacecraft observations, but they can lead
to very different interpretations of the historical sunspot record (Fligge and Solanki, 2000;
Lean, Beer, and Bradley, 1995; Hoyt and Schatten, 1993). These differences in turn lead to
large uncertainties in the solar forcing of climate variations and complicate interpretation of
the magnitude of anthropogenic causes of global terrestrial warming.
One essential ingredient for improving our understanding of solar irradiance variation in
the era of spacecraft observations is the accurate identification of narrowband or broadband
(Foukal et al., 2004) bright and dark solar features and their relation to historically observed
indicators of solar activity. Although solar physicists can easily identify many kinds of solar
features, visual agreement on precise spatial structure is data dependent and generally difficult or impossible to attain, especially for intrinsically low-contrast features such as faculae.
Such difficulties together with the large volume of available data have motivated many automatic or computer-aided feature recognition methods. In the following sections, we examine
Comparison of Solar Feature Classification Methods
115
Figure 1 (b) Red and (c) Ca K-line images from San Fernando Observatory on 9 September 1997. Mapping
of intensity contrast into color is shown by the color bar (a) and applies to all intensity contrast images.
and compare several published methods for identifying solar features. We give no absolute
answers on the quality or correctness of any of the methods but instead develop methodology for quantitative comparison and discuss several important issues pertaining to feature
identification.
2. Observations and Feature Identification Methods

The feature identification methods we discuss here label areas of interest on solar images
obtained with different instruments. The observations from all these instruments are preprocessed to remove artifacts and solar characteristics such as center-to-limb variation as
discussed in the cited references. We will not duplicate these discussions but note that this
preprocessing is crucial for all the methods since residual systematics and artifacts are a
source of error that will produce differences between labelings even if the methods would
otherwise agree perfectly.
The one-trigger(1T) method (Chapman et al., 1992) uses simple thresholds to distinguish dark sunspots (ss) in red continuum contrast images (ic ) and bright faculae (f) in Ca
K core-wing images (icw ) from the quiet-Sun background; observations are obtained with
the Cartesian Full-Disk Telescope-1 (CFDT1) at the San Fernando Observatory (SFO) of
California State University Northridge (CSUN). As in all contrast images discussed in this
paper, contrast is determined by comparing the actual observations to the mean center-tolimb variation. CSUN/SFO images from 9 September 1997 used for comparative analysis
in this paper are shown in Figure 1. Sunspots and faculae derived by this method have been
compared extensively with spacecraft observations of irradiance variation in a number of
studies (e.g., Chapman, Cookson, and Dobias, 1996, 1997). The three-trigger (3T) algorithm (Preminger, Walton, and Chapman, 2001) is a more sophisticated application of
contrast thresholds that requires that the contrast of three contiguous pixels exceed a prespecified value to initiate a search for other neighboring pixels that deviate by more than
a (possibly less restrictive) second threshold. The authors show evidence that the method
identifies real larger feature areas than the one-trigger method without much confusion
with instrumental noise (Preminger, Walton, and Chapman, 2002).
116
H.P. Jones et al.
Figure 2 (b) SOHO/MDI magnetogram and (c) intensity contrast image on 9 September 1997. Mapping
of line-of-sight magnetic flux to color is shown by the color bar (a) and applies to both MDI and SPM
magnetograms. Intensity color mapping is as in Figure 1a.
Turmon, Pap, and Mukhtar (2002) avoid the use of thresholds by applying a statistical
method to SOHO Michelson Doppler Imager (MDI) line-of-sight magnetograms (B ) and
continuum intensity (ic ) contrast images (photograms) taken close together in time. Models
for class-conditional probabilities that a pixel will have observed characteristics given that
it is a member of a specified feature class are developed from distributions of magnetic flux
and intensity in an independently segmented training set of images. These class-conditional
models are inverted with Bayes rule to produce probability images for sunspots, faculae,
and quiet Sun given the actual observed properties of images other than the training set.
The MDI images analyzed here are shown in Figure 2. A simulated-annealing algorithm
then searches for the image labeling with the maximum global (over the entire image) a
posteriori probability. Contiguity of pixels is modeled in the Bayesian prior probability. We
refer to this technique as the maximum likelihood (ML) method.
Harvey and White (1999) apply thresholds sequentially to strictly cotemporal and cospatial line-of-sight magnetograms and continuum contrast images (868.8 nm). The observations were obtained with the NASA/National Solar Observatory (NSO) Spectromagnetograph (SPM) at the NSO/Kitt Peak Vacuum Telescope (KPVT), and they are also the basis
for the Jones et al. (2000, 2003) method (Figure 3). The Harvey and White algorithm begins
with separation of magnetized and unmagnetized regions with a threshold of magnetic
flux comparable to the noise level of the SPM. In the magnetized regions, they first extract
sunspot umbrae and penumbrae using thresholds in both magnetic flux and intensity. They
then proceed to use polarity inversion, fill factors, and thresholds to extract active regions,
decaying regions, and various kinds of network. They account for contiguity of pixels in
each feature class and smooth with Gaussian filters at various stages of the process. We use
a summary classification of their method that labels sunspots (ss; umbrae + penumbrae),
active regions (ar), decaying regions (dr), enhanced network (en), and quiet network (qn;
weak + quiet network in the nomenclature of Harvey and White, 1999). We refer to this
technique as the sequential thresholds (ST) method.
Jones et al. (2000, 2003) apply thresholds to isolate potentially interesting subdomains
of multidimensional histograms of SPM observations [line-of-sight flux, intensity contrast,
equivalent width contrast (w), line depth contrast (d), unipolarity (u), and heliocentric an-
117
Figure 3 NASA/NSO Spectromagnetograph images on 9 September 1997: (a) magnetogram, (b) intensity
contrast, (c) equivalent width contrast, and (d) unipolarity. Magnetogram color mapping is as in Figure 2a;
intensity color mapping is as in Figure 1a. Equivalent width contrast mapping is shown by color bar (e), and
unipolarity mapping by color bar (f).
gle]. Unipolarity for each pixel is calculated as the absolute value of the fractional difference
between positive and negative flux over approximately a 30 30 arcsecond surrounding
area. The authors apply factor analysis over extensive time series of these observations to
118
H.P. Jones et al.
Table 1 Characteristics of the one-trigger (1T), three-trigger (3T), maximum likelihood (ML), sequential
threshold (ST), and histogram factor (HF) feature identification methods. Abbreviations for observed quantities and feature definitions are described in the text.
Method
1T
3T
ML
ST
Instrument
SFO/CFDT1
SOHO/MDI
NASA/NSO SPM
672.3 10 nm
676.8 nm (Ni I)
868.8 nm (Fe I)
B , ic (868.8)
HF
393.4 1 nm
Observables
ic (672.3), icw (Ca II K)
B , ic (676.8)
B , ic , w, d, u
Features
ss (672.3 nm)
ss
ss
ss
f (Ca II K)
ar
uni
qs
qs
dr
mx
en
wk
qn
unas
Comparison with total solar irradiance

Interval
1988 2005
1988 2005
1996 2000
NA
1992 2000
R2
0.79
0.79
0.85
NA
0.77
find the orthogonal (uncorrelated) linear combinations of the original subdomains that explain most of the variance of the SPM observations. Three factors (sunspots umbrae +
penumbrae; strong unipolar + bright limb regions, and magnetically strong mixed polarity
regions) account for most of the SPM variance but only the first two are strongly correlated
with total irradiance variation. Although there is no strictly accurate way to establish a oneto-one correspondence with factor values and pixel position, the gross factor structure is
clear enough that a reasonably accurate spatial labeling can be established. Here, we label
pixels as weak-field (wk), (strong) mixed polarity (mx), unipolar + bright limb (uni), or
sunspots (ss) and leave an unassigned category (unas) because the histogram subdomains
defined by these labelings do not span the entire variable space. We refer to this technique
as the histogram factor (HF) method.
The observations from SFO and NSO were rotated differentially to correspond to the
time of the MDI magnetogram, and common features in the intensity images of all three
instruments as well as on the NSO and MDI magnetograms were visually identified and
their positions were marked. The SFO and NSO images were then warped by using the
IRAF geomap and geotran procedures to complete detailed spatial registration with the MDI
observations. Corresponding transformations were applied to the feature labelings, using
nearest neighbor interpolation. Errors associated with spatial registration will appear in the
comparisons that follow, but we believe they are negligible compared to the differences
arising from the different methods of feature classification.
For convenience in the following discussion, salient features of the five methods are
summarized in Table 1. Also included is multiple R 2 for linear regressions of TSI on feature
statistics over time. Harvey and White (1999) made no such comparison for ST. Despite
119
their many differences, all the methods correlate comparably with TSI and perform better
for shorter periods.
3. Comparison of Feature Labelings

The image labelings from the various methods are shown in Figure 4; many obvious differences are apparent. For example, the labeling schemes for the ST and the HF methods are
more complicated than the simple (quiet Sun, faculae, and sunspot) classification for the 1T,
3T, and ML methods. The HF labelings, which allow for contiguity only through the use of
the unipolarity information, show many more fine-scale structures than the other methods.
More quantitative cross-tabulations of the ten possible pairs of label comparisons are
shown as confusion matrices in Tables 2 5. Each table shows, from a pixel-by-pixel
comparison of two segmentations, the fraction of pixels labeled by both the indicated row
and column features. For example, the first column and second row of Table 2 indicates
that 0.021676 of the total number of pixels were labeled quiet Sun by the 1T method but
faculae by the 3T method. Diagonal entries show the fraction of pixels that were classified
the same by both methods, and perfect agreement, where the categories are putatively the
same, would be shown with nonzero values only along the diagonals. As a figure of merit, we
also show (along with asymptotic standard errors, since large numbers of pixels are involved
even for sunspots) Cohens (Cohen, 1960), which gives an assessment of agreement based
on the assumption that expected random values in the tables are the simple products of the
corresponding marginal probabilities
(the row and
column totals in Tables 2 5). If Oi,j are
the observed entries and Ei,j ( j Oi,j ) ( i Oi,j ) are the expected values,
=

i
Oi,i

i

Ei,i

Ei,i .
(1)
By comparing with expected random values, avoids giving an exaggerated impression

of agreement where,
as is the case here, one class (quiet Sun) dominates the others. For
perfect agreement, i Oi,i = 1, so that = 1, whereas = 0 for random agreement. Given
the observed marginals, however, the maximum possible value (max ) is achieved when the
minimum of the two marginals for category i is substituted for Oi,i in Equation (1). In
Tables 2 5 we also show /max , which gives a less conservative measure of agreement,
particularly when the labelings of one method are largely inclusive of the labelings of the
other. A requirement is that the categories be the same between two methods. We discuss, in
the following, groupings of the categories for the HF and ST methods that allow approximate
comparison with each other and with the (quiet Sun, faculae, and sunspot) classification of
1T, 3T, and ML. We use Cohens only as a convenient summary metric, which at least
approximately orders the quality of the various comparisons, and suggest this measure may
also be useful for such purposes as tuning parameters of a given segmentation method. More
rigorous statistical inferences based on the metric are beyond the scope of this paper.
There are of course many ways in which differences in segmentations can arise, and
there are other ways of presenting the information. Inspection of Figure 4 shows that for
the most part the methods agree on the location of sunspots and faculae or the approximate
equivalents in the ST and HF methods. The differences arise in the details of the segmentation definitions. We have also examined but, for clarity and compactness, do not show ten
detailed pixel-by-pixel comparison images that show the spatial structure of the categories
summarized by Tables 2 5. Although interesting details are apparent, those relevant to the
120
H.P. Jones et al.
Figure 4 Feature labelings by the (a) 1T, (b) 3T, (c) ML, (d) ST, and (e) HF methods. Color tables are shown
in panel (f). Feature abbreviations are described in the text.
121
Table 2 Fraction of one-trigger labelings for each label category of the three-trigger, maximum likelihood,
histogram factor, and sequential thresholds methods. Also shown for each comparison are Cohens (see
text) the asymptotic standard error and /max .
One-trigger
Quiet Sun
Faculae
Sunspot
Total
Three-trigger
Quiet Sun
0.959813
0.000000
0.000000
0.959813
Faculae
0.021676
0.013105
0.000000
0.034782
Sunspot
0.001725
0.000496
0.003185
0.005405
Total
0.983214
0.013601
0.003185
1.000000
= 0.57 0.003; /max = 0.98

Maximum likelihood
Quiet Sun
0.975289
0.005166
0.000022
0.980477
Faculae
0.006545
0.007922
0.000232
0.014699
Sunspot
0.001234
0.000631
0.002959
0.004825
Total
0.983068
0.013719
0.003212
1.000000
Weak field
0.813898
0.001625
0.000045
0.815568
Strong mixed polarity
0.052326
0.000066
0.000004
0.052396
Unipolar + bright limb
0.076339
0.009770
0.000243
0.086353
= 0.61 0.003; /max = 0.66

Histogram factor
Sunspot
0.000479
0.000197
0.002178
0.002854
Unassigned
0.040053
0.002039
0.000737
0.042829
Total
0.983095
0.013697
0.003207
1.000000
= 0.21 0.002; /max = 0.70

Sequential thresholds
Quiet network
0.884605
0.000286
0.000000
0.884891
Enhanced network
0.080937
0.002406
0.000000
0.083344
Active region
0.016362
0.010633
0.000412
0.027407
Sunspot
0.001086
0.000457
0.002815
0.004358
Total
0.982990
0.013782
0.003227
1.000000
= 0.56 0.003; /max = 0.81
purposes of this paper can also be seen in Figure 4 and are reasonably summarized in the
tables.
Table 2 compares the 1T labeling with each of the other methods. The 3T method, as
discussed in more detail by Preminger, Walton, and Chapman (2001), finds more facular
and sunspot pixels than the 1T method, with most of the additional area for both cases
coming from areas identified as quiet Sun in the one-trigger labeling. Although the facular
122
H.P. Jones et al.
Table 3 Comparison of 3T with ML, HF, and ST as in Table 2.

Three-trigger
Quiet Sun
Faculae
Sunspot
Total
Maximum likelihood
Quiet Sun
0.957087
0.023074
0.000316
0.980477
Faculae
0.002168
0.011478
0.001053
0.014699
Sunspot
0.000210
0.000531
0.004084
0.004825
Total
0.959464
0.035084
0.005452
1.000000
Weak field
0.805260
0.010199
0.000108
0.815568
0.051375
0.001016
0.000005
0.052396
0.064918
0.020079
0.001355
0.086353
= 0.53 0.003; /max = 0.83

Histogram factor
Sunspot
0.000055
0.000150
0.002649
0.002854
Unassigned
0.037920
0.003583
0.001326
0.042829
Total
0.959529
0.035028
0.005443
1.000000
= 0.32 0.002; /max = 0.56

Quiet network
0.879247
0.005644
0.000000
0.884891
Enhanced network
0.072215
0.011128
0.000000
0.083344
Active region
0.007634
0.018103
0.001670
0.027407
Sunspot
0.000181
0.000370
0.003807
0.004358
Total
0.959277
0.035245
0.005477
1.000000
= 0.62 0.002; /max = 0.71
areas identified by the 1T and ML labelings are similar, only about half of the facular pixels
identified by either method come from the same areas on the Sun; the remainder are labeled
quiet Sun by the other method. The sunspot areas identified by the ML method are larger
than those for the 1T method, with most of the difference coming from 1T quiet-Sun areas.
Comparison with both the HF and ST labelings is more complicated. A rough correspondence between the HF and 1T labelings can be made by combining the weak-field, strong
mixed polarity, and unassigned categories as corresponding to quiet Sun and unipolar regions as corresponding to faculae. Indeed, in the Jones et al. papers, unipolar + bright limb
regions correlated strongly and positively with total solar irradiance variation, as would be
expected for faculae. As in the comparison with the ML method, however, there is then still
considerable confusion between the quiet-Sun and facular regions. The HF method identifies a far larger area corresponding to unipolar regions than the facular area labeled by
the 1T method, with most of the discrepancy arising in 1T quiet-Sun regions. Finally, the
sunspot area labeled by the HF method is somewhat similar to but slightly smaller than the
corresponding 1T area.
123
Table 4 Comparison of ML with HF and ST labelings.

Maximum likelihood
Quiet Sun
Faculae
Sunspot
Total
Weak field
0.815227
0.000335
0.000086
0.815648
0.052096
0.000174
0.000007
0.052277
0.073739
0.012047
0.000636
0.086421
Histogram factor
Sunspot
0.000004
0.000147
0.002697
0.002848
Unassigned
0.039487
0.001939
0.001381
0.042807
Total
0.980552
0.014642
0.004806
1.000000
= 0.26 0.002; /max = 0.80

Quiet network
0.883910
0.000984
0.000038
0.884932
Enhanced network
0.080293
0.003061
0.000012
0.083365
Active region
0.016146
0.010440
0.000768
0.027353
Sunspot
0.000076
0.000254
0.004020
0.004350
Total
0.980425
0.014738
0.004837
1.000000
= 0.58 0.003; /max = 0.77
Table 5 Comparison of ST and HF labelings.

ST
Q. Network
E. Network
Active region
Sunspot
Total
Weak field
0.758463
0.053000
0.005294
0.000057
0.816813
Mixed polarity
0.049992
0.002112
0.000513
0.000000
0.052618
Unipolar
0.047937
0.020560
0.017854
0.000314
0.086664
Sunspot
0.000000
0.000000
0.000187
0.002679
0.002866
Unassigned
0.028542
0.007693
0.003505
0.001300
0.041039
Total
0.884933
0.083364
0.027353
0.004350
1.000000
HF
= 0.31 0.002; /max = 0.64
Similarly, a rough correspondence between the ST and 1T labelings is seen if the ST

quiet and enhanced network categories are combined and considered as quiet Sun. The ST
active region area is roughly twice as large as the 1T facular area, with most of the difference
coming from one-trigger quiet Sun, and the ST sunspot area is about 50% larger than that of
the one-trigger case.
Table 3 shows similar comparisons for the 3T with the ML, HF, and ST methods. Here,
the 3T facular area is more than twice that of ML, with most of the difference coming from
ML quiet Sun, whereas the sunspot areas are very similar. The discrepancy between the
3T facular and the HF unipolar areas is less than but qualitatively similar to that in the 1T
124
H.P. Jones et al.
comparison, and the HF sunspot area is only about half that identified by the 3T method.
Both the facular and sunspot areas for the 3T and ST methods agree better than for the 1T
comparison, although roughly one-third of the ST active-region pixels are identified as either
quiet-Sun or sunspot pixels by the 3T method.
Table 4 compares the ML labels with both the HF and ST classifications. The same
combinations of categories are identified as quiet Sun for the HF and ST methods as in
the previous two tables. Again, the HF unipolar area is much larger than the ML facular
area, whereas the HF sunspot area is smaller. The comparison of ST active regions and ML
faculae is similar to that for the ST 1T case, whereas the ST and ML sunspot areas agree
fairly well.
Finally, Table 5 compares the HF and ST labelings. To facilitate this comparison, we
combine the HF weak-field and unassigned categories as best corresponding to ST quiet network, strong-field mixed polarity regions as ST enhanced network, and unipolar as ST active
regions because the strong majority of ST active-region pixels correspond to HF unipolar
pixels. However, this correspondence is not very good as the majority of HF unipolar pixels
originate in ST quiet and enhanced network. The HF sunspot area is only about two-thirds
the ST area, with most of the discrepant pixels being in the HF unassigned category. Many
of these unassigned pixels are for large negative intensity contrasts with weak line-of-sight
fields such as might occur in some areas of penumbra and near the limb where the field is
nearly perpendicular to the line of sight.
4. Discussion
From the previous section one can see by visual inspection of the feature labels, from the
confusion matrices, and from that agreement is poorest when any of the methods are compared with HF. This is perhaps not surprising because the HF method was not designed to
identify spatial features. For example, the total sunspot area is lower for the HF method than
for any of the other algorithms. This is mostly because the histogram subdomains do not
completely span the variable space. In particular, dark features that lack strong line-of-sight
field, as will occur most often in penumbrae near the limb, and strong-field features that are
in sunspots but are not especially dark are missed by the HF technique and are counted as
unassigned. This is reflected both in the large number of unassigned pixels near the southeast
spot in Figure 4e and in the summary tables. However, correlation of the histogram factors
with total irradiance is comparable to that achieved with spatial features, and the analyses
of Jones et al. (2000, 2003) imply that many features identified by histogram subdomains
are temporally correlated with the features identified by the other techniques. The fact that
/max generally considerably exceeds for HF comparisons also indicates that HF features
either include or are included in features identified by other methods. Moreover, comparison
of HF and ML feature areas over time (J.M. Pap, private communication) explicitly shows
strong temporal correlation between ML sunspots and faculae with the corresponding HF
features even though the actual areas differ substantially. Another important point of the HF
technique is that the temporal variation of one magnetically well-determined factor (primarily strong mixed polarity fields) is not well correlated with irradiance variation.
Even where agreement between methods is comparatively good, there are still important differences, especially for faculae, but even in some instances for sunspots. Much of the
disagreement for faculae results from their low contrast at the spatial resolution of the instruments producing the data for this paper. For this reason, many feature identification methods
rely on higher contrast proxy measurements of faculae from Ca K (as in the CSUN observations), Mg k, or magnetograms (as in the ML, ST, and HF methods). The proxies are
125
usually formed higher in the solar atmosphere than faculae, tend to have larger but spatially
correlated areas, and have center-to-limb visibility that is markedly different from faculae,
which are best seen near the solar limb.
The methods also differ in their use of contiguity (guilt by association) to determine
class membership. The effect is on the spatial scale of the feature identifications. Solar physicists clearly differ in what they expect in this regard, and this may be largely a function of
the purpose for which the identifications are intended. For example, a more spatially detailed classification scheme may be needed for irradiance comparison than is required for
comparison with global dynamo models.
The comparisons between feature identification methods discussed here are far from
complete but we feel that they are representative of methods in actual use. Given the diversity of observations, the variety of approaches, and lack of clear standards, improvement
may be slow. However, we do suggest that there are areas where improvement is possible.
Perhaps the most important is to establish agreement on what classes of features should be
identified for a given purpose. If such agreement can be achieved, iterative development
of standards becomes possible. Observations can be the basis for good physical models or
physics-based feature definitions that allow prediction of feature appearance for different instruments under a variety of observing conditions. Iterative comparison between theory and
observation then can in principle lead to more rigorous absolute standards against which
various labeling algorithms can be tested. The beginnings of such development can be seen
in the modeling of faculae by Keller et al. (2004) and the corresponding observations of
Berger, Rouppe van der Voort, and Lofdahl (2005). Here, the link between accurate models
and the spatial resolution of both observations and labeling algorithms becomes especially
important since no instrument can presently image the full solar disk at the resolution of
either the above model or observations.
It is perhaps worth noting that the very notion of features is dependent on human perception and is thus subjective. Presumably, some physical reality underlies the recognition of
common image characteristics by educated observers, and all the methods attempt to reduce
uncertainties introduced by differing individual perception by relying extensively on computer algorithms, which are in some sense objective. However, different rationales are used
to set parameters. In the 1T and 3T methods, quantitative arguments are used to establish
thresholds that minimize effects of seeing and confusion with instrumental noise, but these
are dependent on instrument and observing site. The ML method requires an independent
method for labeling features in a training set but does not otherwise depend on thresholds.
The ST and HF methods use thresholds established by the knowledge and experience of the
investigators in analyzing the SPM data. All are subjective in the selection of which features
are of physical interest. A major goal of this paper is to begin testing the effects of these
many choices by direct comparison.
Acknowledgements The authors are pleased to acknowledge P. Jones for help with quantitative comparison of categorical data and an anonymous referee for suggestions that substantially improved the original
manuscript. H.P.J. carried out much of his early contribution to this research as a member of the Laboratory for Astronomy and Solar Physics at NASAs Goddard Space Flight Center (GSFC). This research was
supported by several NSF and NASA Supporting Research and Technology and Guest Investigator grants
over many years. SOHO is a mission of international cooperation between ESA and NASA, and the authors
gratefully acknowledge the effort of the MDI team. NSO/KPVT data used here were produced cooperatively
by AURA/NSO, NASAs GSFC, and NOAA/SEC.
126
H.P. Jones et al.
References
Berger, T.E., Rouppe van der Voort, L., Lofdahl, M.G.: 2005, High resolution magnetogram measurements
of solar faculae. In: AGU Spring Meeting Abstracts, SP31 A02.
Chapman, G.A., Cookson, A.M., Dobias, J.J.: 1996, Variations in total solar irradiance during solar cycle 22.
J. Geophys. Res. 101, 13541 13548. doi:10.1029/96JA00683.
Chapman, G.A., Cookson, A.M., Dobias, J.J.: 1997, Solar variability and the relation of facular to sunspot
areas during solar cycle 22. Astrophys. J. 482, 541 545. doi:10.1086/304138.
Chapman, G.A., Herzog, A.D., Lawrence, J.K., Walton, S.R., Hudson, H.S., Fisher, B.M.: 1992, Precise
ground-based solar photometry and variations of total irradiance. J. Geophys. Res. 97, 8211 8219.
Cohen, J.: 1960, A coefficient of agreement for nominal scales. Educ. Psychol. Meas. 20, 37 46.
de Toma, G., White, O.R., Chapman, G.A., Walton, S.R., Preminger, D.G., Cookson, A.M., Harvey, K.L.:
2001, Differences in the Suns radiative output in cycles 22 and 23. Astrophys. J. 549, L131 L134.
doi:10.1086/319127.
de Toma, G., White, O.R., Chapman, G.A., Walton, S.R., Preminger, D.G., Cookson, A.M.: 2004, Solar cycle
23: an anomalous cycle? Astrophys. J. 609, 1140 1152. doi:10.1086/421104.
Fligge, M., Solanki, S.K.: 2000, The solar spectral irradiance since 1700. Geophys. Res. Lett. 27, 2157 2160.
doi:10.1029/2000GL000067.
Fligge, M., Solanki, S.K., Unruh, Y.C.: 2000, Modelling irradiance variations from the surface distribution of
the solar magnetic field. Astron. Astrophys. 353, 380 388.
Foukal, P., Lean, J.: 1988, Magnetic modulation of solar luminosity by photospheric activity. Astrophys. J.
328, 347 357. doi:10.1086/166297.
Foukal, P., Bernasconi, P., Eaton, H., Rust, D.: 2004, Broadband measurements of facular photometric contrast using the solar bolometric imager. Astrophys. J. 611, L57 L60. doi:10.1086/423787.
Frhlich, C.: 2006, Solar irradiance variability since 1978. Space Sci. Rev. 125, 53 65.
doi:10.1007/s11214-006-9046-5.
Frhlich, C., Lean, J.: 1998, The Suns total irradiance: cycles, trends and related climate change uncertainties
since 1976. Geophys. Res. Lett. 25, 4377 4380. doi:10.1029/1998GL900157.
Harvey, K.L., White, O.R.: 1999, Magnetic and radiative variability of solar surface structures. I. Image
decomposition and magnetic-intensity mapping. Astrophys. J. 515, 812 831. doi:10.1086/307035.
Hoyt, D.V., Schatten, K.H.: 1993, A discussion of plausible solar irradiance variations, 1700 1992. J. Geophys. Res. 98, 18895 18906.
Jones, H.P., Branston, D.D., Jones, P.B., Wills-Davey, M.J.: 2000, Analysis of NASA/NSO spectromagnetograph observations for comparison with solar irradiance variations. Astrophys. J. 529, 1070 1083.
doi:10.1086/308315.
Jones, H.P., Branston, D.D., Jones, P.B., Popescu, M.D.: 2003, Comparison of total solar irradiance with
NASA/National Solar Observatory spectromagnetograph data in solar cycles 22 and 23. Astrophys. J.
589, 658 664. doi:10.1086/374413.
Keller, C.U., Schssler, M., Vgler, A., Zakharov, V.: 2004, On the origin of solar faculae. Astrophys. J. 607,
L59 L62. doi:10.1086/421553.
Krivova, N.A., Solanki, S.K., Fligge, M., Unruh, Y.C.: 2003, Reconstruction of solar irradiance variations in cycle 23: is solar surface magnetism the cause? Astron. Astrophys. 399, L1 L4.
doi:10.1051/0004-6361:20030029.
Lean, J.: 2001, The variability of the Sun: from the visible to the X-rays (CD-ROM directory: contribs/lean).
In: Garcia Lopez, R.J., Rebolo, R., Zapaterio Osorio, M.R. (eds.) 11th Cambridge Workshop on Cool
Stars, Stellar Systems and the Sun, ASP 223, Astron. Soc. Pac., San Francisco, 109 116.
Lean, J., Beer, J., Bradley, R.: 1995, Reconstruction of solar irradiance since 1610: implications for climate
change. Geophys. Res. Lett. 22, 3195 3198. doi:10.1029/95GL03093.
Lean, J., Rottman, G., Harder, J., Kopp, G.: 2005, SORCE contributions to new understanding of global
change and solar variability. Solar Phys. 230, 27 53. doi:10.1007/s11207-005-1527-2.
Pap, J.M.: 2003, Total solar and spectral irradiance variations from near-UV to infrared. In: Rozelot, J.P. (ed.)
The Suns Surface and Subsurface: Investigating Shape, Lecture Notes in Physics 599, Springer, Berlin,
129 155.
Pap, J.M., Turmon, M., Floyd, L., Frhlich, C., Wehrli, C.: 2002, Total solar and spectral irradiance variations
from solar cycles 21 to 23. Adv. Space Res. 29, 1923 1932.
Preminger, D.G., Walton, S.R., Chapman, G.A.: 2001, Solar feature identification using contrasts and contiguity. Solar Phys. 202, 53 62.
Preminger, D.G., Walton, S.R., Chapman, G.A.: 2002, Photometric quantities for solar irradiance modeling.
J. Geophys. Res. 107, 1354 1362. doi:10.1029/2001JA009169.
Turmon, M., Pap, J.M., Mukhtar, S.: 2002, Statistical pattern recognition for labeling solar active regions:
application to SOHO/MDI imagery. Astrophys. J. 568, 396 407. doi:10.1086/338681.
127
Wenzler, T., Solanki, S.K., Krivova, N.A., Frhlich, C.: 2006, Reconstruction of solar irradiance variations in cycles 21 23 based on surface magnetic fields. Astron. Astrophys. 460, 583 595.
doi:10.1051/0004-6361:20065752.
Willson, R.C., Hudson, H.S.: 1988, Solar luminosity variations in solar cycle 21. Nature 332, 810 812.
doi:10.1038/332810a0.
Willson, R.C., Mordvinov, A.V.: 2003, Secular total solar irradiance trend during solar cycles 21 23. Geophys. Res. Lett. 30, 1199 1202.
Willson, R.C., Gulkis, S., Janssen, M., Hudson, H.S., Chapman, G.A.: 1981, Observations of solar irradiance
variability. Science 211, 700 702.
The Observed Long- and Short-Term Phase Relation

between the Toroidal and Poloidal Magnetic Fields
in Cycle 23
S. Zharkov E. Gavryuseva V. Zharkova

Abstract The observed phase relations between the weak background solar magnetic
(poloidal) field and strong magnetic field associated with sunspots (toroidal field) measured
at different latitudes are presented. For measurements of the solar magnetic field (SMF) the
low-resolution images obtained from Wilcox Solar Observatory are used and the sunspot
magnetic field was taken from the Solar Feature Catalogues utilizing the SOHO/MDI fulldisk magnetograms. The quasi-3D latitudinal distributions of sunspot areas and magnetic
fields obtained for 30 latitudinal bands (15 in the northern hemisphere and 15 in the southern hemisphere) within fixed longitudinal strips are correlated with those of the background
SMF. The sunspot areas in all latitudinal zones (averaged with a sliding one-year filter) reveal a strong positive correlation with the absolute SMF in the same zone appearing first
with a zero time lag and repeating with a two- to three-year lag through the whole period of
observations. The residuals of the sunspot areas averaged over one year and those over four
years are also shown to have a well defined periodic structure visible in every two three
years close to one-quarter cycle with the maxima occurring at 40 and + 40 and drifts
during this period either toward the equator or the poles depending on the latitude of sunspot
occurrence. This phase relation between poloidal and toroidal field throughout the whole cycle is discussed in association with both the symmetric and asymmetric components of the
background SMF and relevant predictions by the solar dynamo models.
Keywords Solar cycle: observations Sunspots: magnetic fields, statistics
S. Zharkov ()
Department of Applied Mathematics, University of Sheffield, Sheffield, UK
e-mail: s.zharkov@sheffield.ac.uk
E. Gavryuseva
Arcetri Observatory, University of Florence, Florence, Italy
V. Zharkova
Department of Computing, University of Bradford, Bradford, UK
DOI: 10.1007/978-0-387-98154-3_11
129
130
S. Zharkov et al.
1. Introduction
Understanding the origin of a magnetic cycle in solar activity is a very important and still
unsolved problem in solar physics. The solar cycle is defined by sunspot occurrences, or
their areas, which change with a period of 11 or 22 years as observed for more than a few
hundred years (Hathaway and Wilson, 2004). The solar cycle is believed to result from a
hydromagnetic dynamo that operates from the bottom of the solar convective zone (SCZ) in
the solar interior where magnetic flux tubes are formed and travel for about half the cycle
time of 11 years to the solar surface to appear as sunspots at the beginning of a new cycle
(see reviews by Tobias, 2002; Ossendrijver, 2003; Krivodubskij, 2005). The mechanisms
affecting the travel time of the magnetic tubes from the SCZ to the solar surface and their
migration in latitude and longitude inside the solar interior define the duration of the solar
cycle and the sunspot distribution on the solar surface. Sunspots are also known to migrate
in latitude during the solar cycle, with a minority of sunspots migrating toward the poles
and the majority migrating toward the equator as shown in Maunders butterfly diagrams
(Maunder, 1904).
At the middle of the previous solar cycle a new poloidal field is assumed to be
formed by the dynamo mechanism at the poles. Once this field is brought to the base
of the SCZ, it becomes affected by differential rotation of the solar plasma and buoyancy instability, collapsing to form strong toroidal field as magnetic flux tubes. These
tubes travel through the solar interior and rise to the surface in the form of -shaped
loops (Parker, 1955) to form sunspots and active regions. As the cycle progresses, this
toroidal field becomes transformed back into a poloidal one by convection and the Coriolis force ( effect; Krause and Radler, 1980). The more recent dynamo models predict that the latitudinal distributions of the sunspot area and magnetic fields at mid-tolow latitudes where the sunspots emerge (called the royal zone; Krivodubskij, 2005)
are governed by magnetic-field advection and buoyancy (Kitchatinov and Rudiger, 1999;
Dikpati and Gilman, 2001; Nandi and Choudhuri, 2001) allowing magnetic flux tubes to
travel upward and appear on the solar surface.
However, despite the many solar dynamo models, there remain many unanswered questions about the dynamo mechanisms. One of them is the phase relation between the poloidal
and toroidal fields during the solar cycle. This relation was investigated for oscillatory axisymmetric mean-field dynamo models, which predicted a quadruple symmetry about the
equatorial plane and the observed shape of latitudinal distributions in the toroidal magnetic
field (Stix, 1976).
The Mt. Wilson Observatory measurements for cycles 20 and 21 have shown that the
polarities for sunspot magnetic fields deduced from butterfly diagrams were opposite
to the polarity of the background solar magnetic field in each hemisphere (Stix, 1976;
Vainstein, Zeldovich, and Ruzmaikin, 1980), exhibiting a full-cycle lag between the poloidal
and toroidal fields. Then, by restoring the radial magnetic-field components, it was con< 0) can recluded that only models with angular velocity increasing with depth (i.e.,
r
produce closely enough the shape of the butterfly diagrams (Stix, 1976). This implies that
the helicity parameter has to be positive to account for meridional flows of dynamo waves,
< 0; Krivodubskij, 2005). However, the positive value of contradicts other stud(i.e.,
r
ies where this parameter is found to change from positive near the surface to negative at the
bottom of the SCZ to allow the flux tubes to travel to the surface (Yoshimura, 1981).
Recent observations show that the latitudes where sunspots emerge on the surface in the
northern and southern hemisphere vary differently at different phases of the solar cycle. The
recent observations of the magnetic tilts of sunspot groups and active region areas and the
131
total areas covered by sunspots and active regions in the northern and southern hemispheres
(counted as a sum of the areas for all the sunspots/active regions detected in a given image) are found to be very asymmetric and to vary periodically during the cycle (Temmer,
Veronig, and Hanslmeier, 2002; Zharkova and Zharkov, 2007). Moreover, there is an observed distinct domination of one or another hemisphere with the two basic periods: a long
one of 11 years and a shorter one of about 2.5 3 years (or approximally one-quarter cycle)
(Zharkov, Zharkova, and Ipson, 2005; Zharkov and Zharkova, 2006).
In addition, strong negative correlation is detected between the average daily latitudinal
motions of sunspot groups and their tilt angles: The groups with smaller tilts are normally
associated with equatorward drifts (Howard, 1991). Also, toward the maximum of activity
the signs of the tilt angle in sunspot groups during their lifetime reverses (Howard, 1991).
The groups with tilt angles ( ) near the average show a positive change of daily polarity
separation (d) between leading and following spots depending on the latitude as /d
0.021 0.123 (Holder et al., 2004; Zharkova and Zharkov, 2007).
At the same time, low-resolution observations of the solar magnetic field (SMF) from
the Wilcox Solar Observatory (WSO) have shown that the SMF has a well-defined fourzone structure: two polar zones with latitudes above 25 and two near-equatorial ones for
the latitudes ranging from zero to 25 . The near-equatorial zones have opposite magnetic
field polarities in the northern and southern hemispheres that are swapped every two three
years (Gavryuseva, 2006), similar to those detected for sunspots and active regions (Zharkov,
Zharkova, and Ipson, 2005; Zharkov and Zharkova, 2006). Therefore, further investigation
is required to establish the correlation between the latitude of appearance of magnetic flux
tubes, or sunspots, and the variations of the SMF in the opposite hemispheres during different phases of the cycle. This will allow tracking of the relation between the poloidal and
toroidal field in cycle 23.
The data for sunspot and active region areas and total fluxes, as well as for the lowresolution solar magnetic field, are described in Section 2. The results of correlation of these
two time series at various latitudes and the phase of the solar cycle are discussed in Section 3.
Conclusions are drawn in Section 4.
2. Data Description and Statistical Tool

2.1. Solar Feature Catalogue
The searchable Solar Feature Catalogues (SFCs) with nine-year coverage (1996 2005) are
developed from digitized solar images by using automated pattern-recognition techniques
(Zharkova et al., 2005). All full-disk images were first automatically standardized in intensity and to a circular shape. Then automated techniques were applied for the detection of
sunspots, active regions, filaments, and line-of-sight magnetic neutral lines in full-disk solar
images in Ca II K1, Ca II K3, and H lines taken at the Paris-Meudon Observatory and for
the white-light images and magnetograms from SOHO/MDI. The results of the automated
recognition were verified with manual synoptic maps and available statistical data, revealing
a high detection accuracy of 92 96% (Zharkov, Zharkova, and Ipson, 2005).
Based on the recognized parameters, a relational database of Solar Feature Catalogues
was built for every feature and published with various predesigned search pages on the Bradford University Web site (http://solar.inf.brad.ac.uk/). At the moment, the catalogue contains
368 676 sunspot features extracted from 10 082 SOHO/MDI continuum observations and
around 100 000 active region features starting from 19 May 1996 at 19:08:35 to 31 May
132
S. Zharkov et al.
2005. The catalogue is generally accessible via http://solar.inf.brad.ac.uk (ASCII and XML
files) and through the EGSO-registered Web services and SolarSoft vobs/egso branch.
For detected sunspots the following information is extracted: number of umbrae, Carrington and planar coordinates of the sunspot center of gravity, heliographic area, heliographic diameter, maximum magnetic field, minimum field, excess magnetic flux (EMF),
total (absolute) flux, maximum umbra field, minimum umbra field, excess umbra flux, total
(absolute) umbra flux, the pixel and planar coordinates of the bounding rectangle, its raster
scan marking the penumbral and umbral pixels, and the observational parameters (Zharkov,
Zharkova, and Ipson, 2005).
Note that Carrington rotation (CR) number 1910 corresponds to the start of 1996 and CR
number 2030 to May 2005. Also note that gaps in the graphs are caused by the absence of
MDI data because of technical problems with SOHO in 1998; these periods were excluded
from the correlation analysis.
2.2. Wilcox Solar Observatory Data
To derive large-scale solar magnetic field variations, we use the low-resolution magnetograms captured daily at the WSO site (http://wso.stanford.edu/synoptic.html). The lineof-sight component of the background solar magnetic field (hereafter, solar magnetic field,
SMF) is measured from the photosphere by using the Zeeman splitting of the 525.02-nm
Fe I spectral line with the WSOs Babcock solar magnetograph. Each longitudinal value is
the weighted average of all of the observations made in the longitudinal zone within 55
around the central meridian (Hoeksema, 1985).
The observations of the photospheric magnetic field (SMF) were carried out from the
year 1976.405, corresponding to the beginning of the CR 1642, up to the year 2005.16, or
CR 2027. The data cover the solar activity cycles 21, 22, and 23 (Scherrer et al., 1977;
Hoeksema, 1985). For the purpose of the current study, we consider the averaged magnetic
field in cycle 23 only. From this data set we then select a subset within the 30 -hemispheres
in the heliographic latitude we then select a subset within the 30 -hemispheres from 72.5
North to 72.5 South binned in uniform intervals of the sine of the latitude.
2.3. Some Elements of Statistics
The correlation analysis can only establish an approximate estimate of the real dependencies
between two time series since they do not include many physical influences (e.g., differential rotation, convection, diamagnetic advection, and magnetic buoyancy) that affect the
measured data.
Therefore, the compared time series entities might not have a linear dependence and,
thus, their correlation coefficients are not expected to be very high. However, their magnitudes above or below zero will indicate the level of relation between them and a time lag,
if detected, can suggest a similarity in their periodicities, or phase relation. This statistical
analysis is sufficient for understanding the links between the two time series that can provide
some input to future models of the solar dynamo.
As is well known, a correlation coefficient indicates the strength and direction of a linear
relationship between two random variables, or their departure from being independent. The
correlation coefficient r is defined by the formula
N1
(xi x)(y
i y)
=
r =

N1
N1
1
2 1
2
(x
x)
(y
y)
i
i
i=0
i=0
N1
N1
1
N1
i=0
1
N
N1
i=0
(xi x)(y
i y)
,
x y
(1)
133
where N is the number of measurements, x and y are the expected values (means) of the
series X and Y, and x and y are their standard deviations. Depending on the type of correlation we try to obtain, the measurements in both series have to be preprocessed, or averaged,
with the averaging filter sliding across the data during the full period of measurements. The
filter has a width of a quarter or a half of the period that one tries to detect.
If one is looking for some periodic patterns in the investigated series then a crosscorrelation analysis is required to measure a level of similarity of the sunspot characteristics to those of the SMF. This cross-correlation is a function of relative time between the
measurements:

(2)
(X Y )(z) = X (t)Y (z + t) dt,
where the integral is taken over the range of the values of t and the asterisk refers to the
complex conjugate.
Another very important characteristic that one needs to check with these series is an
autocorrelation, or a measure of how well the measurements match a time-shifted version of
itself. The autocorrelation function R(t, s) is defined as follows:
R(t, s) =
E[(Xt )(Xs )]
,
2
(3)
where Xt and Xs are the measurements at different times, E is the expected-value operator,
and is the mean of X. As a result of the definition, an autocorrelation is always peaked at
a zero lag.
According to the dynamo models, one can expect a nonlinear relation between the
toroidal and poloidal magnetic fields. Since these are assumed to be associated with the
observed sunspot and SMF entities, respectively, one needs to evaluate the residuals of the
measured averaged magnitudes and estimate a confidence in their variations. The residual
for the ith measurement of X is defined as
Res(i) =
X i
S
N
N
(4)
where SN is the variance of X, X i is the measured mean of X, and is the estimated

mean with a four-year averaging filter. This is defined on the basis of the confidence interval
calculations relying on Students t -distribution which both the measured series (EMF and
ESMF) satisfy.
3. Results and Discussion

3.1. Latitudinal Distributions of Sunspot and Active Region Areas and Excess Magnetic
Flux
3.1.1. Quasi-3D Butterfly Diagrams of the Sunspot Areas and Excess Magnetic Field
Let us investigate latitudinal variations of the total sunspot areas (upper plot) and excess
magnetic flux (EMF; bottom plot) presented in Figure 1, which were calculated in narrow
latitude strips for the whole longitude range in different Carrington rotations during the
whole solar cycle. This plot is similar to the classic Maunder butterfly diagrams for sunspot
134
S. Zharkov et al.
Figure 1 Butterfly diagrams for cycle 23 for the total sunspot daily areas (upper plot) and excess magnetic
fields (lower plot) summed over longitude within two-degree-wide latitudinal strips. The areas are measured
in squared degrees and magnetic field in gauss. Note that CR number 1910 corresponds to the start of 1996
and CR number 2030 to May 2005.
135
appearances (Maunder, 1904) but includes not only the locations of sunspots but also their
total area variations in different latitudinal intervals. The total areas of sunspots for a given
Carrington rotation are marked with the color scale from a dark black/blue color (corresponding, respectively, to low areas/negative flux) to a yellow white/red (corresponding,
respectively, to high areas/positive flux) with the scales presented on the right-hand side of
the plot.
Similarly to the previous cycles, the locations of sunspots are confined within the royal
zone between the latitudes of 40 and + 40 , as can be seen from Figure 1. At the beginning of the cycle the sunspots emerge at higher latitudes. As the cycle progresses, the
sunspot emergence zone slowly moves toward the equator, showing a strong increase in
sunspot area during the years of maximum activity.
Additionally, in these two butterfly diagrams one can see fine structure appearing during
the activity period as vertical strips in the total areas and magnetic flux. These strips reveal
a quasi-periodic mosaic structure changing from one hemisphere to another (i.e., within a
given time interval the areas of a particular latitudinal strip in one hemisphere are larger than
areas in the symmetric strip in the other hemisphere). Then, for another period, the larger
area shifts to another latitude or even to the opposite hemisphere. Let us explore whether a
period of these changes can be clearly extracted and, if so, what are the mechanisms causing
such quasi-periodic changes.
3.1.2. The North South Asymmetry in the Areas and Excess Magnetic Flux
The total daily areas of sunspots and active regions (AR) and their cumulative areas for
northern and southern hemispheres, averaged with the one-year filter, are found to be
strongly asymmetric during the whole solar cycle 23 (Figure 2; Zharkov, Zharkova, and
Ipson, 2005; Zharkova, Zharkov, and Benkhalil, 2005). This asymmetry is stronger at the
beginning of the cycle, approaching the values 0.4 for sunspots and 0.6 for active regions.
The North South asymmetry in the total area variations for both sunspots and ARs also reveals well-defined periodicities with the two basic periods: a longer one covering the whole
period of observations (over nine years) and a smaller one of two three years (close to
one-quarter cycle 2.75 years).
There is also a visible asymmetry in the sunspot cumulative areas (Figure 3, upper plot),
which is also presented as the asymmetry function (Figure 3, lower plot). Cumulative areas
provide additional information about the dynamo mechanism not supplied by the total areas.
They can show whether the oscillatory patterns of North South asymmetries seen in total
areas will still be present in cumulative areas and, if so, how long they exist and when they
equalize, or whether the dynamo wave decays.
Obviously, at the start of the cycle, when the solar dynamo starts, there are fewer sunspots
and the North South asymmetry is the strongest, as seen in Figure 3. If the solar dynamo
does not produce any asymmetry in the interior conditions of both hemispheres affecting the
rate of sunspot emergence, as the sunspot number increases we would expect that the symmetry would quickly be restored. However, this is not the case, as one can see in Figure 3.
Similarly to total areas, the cumulative areas at the cycle start in 1996 are dominant in
the northern hemisphere, and then, in 2.5 3 years, the southern hemisphere takes over. This
pattern repeats again toward the cycle end. Therefore, both periods detected in the total areas
are also seen in the cumulative areas. The only difference is that the cumulative areas show
the asymmetry decrease, or decay, as ekt toward the cycle end, where k can be determined
from Figure 3 as k 3.2, where t is the time after the cycle start. This decay can characterize
some turbulent processes, which slow down the primary solar dynamo wave to 11 years, and
the decay constant k can be compared with those predicted by solar dynamo models.
136
S. Zharkov et al.
Figure 2 The North South area asymmetries in cycle 23 for the total sunspot (upper plot) and active region
daily areas (lower plot) averaged over one year.
Even stronger North South asymmetry is seen in the excess magnetic flux; that is, the
magnetic flux added with its sign for the leading and trailing polarities, contained in all
sunspots in the northern and southern hemispheres (Figure 4), which reveals a nearly complete polarity separation over the hemispheres similar to those reported for the previous
cycles 20 and 21 by Stix (1976). This separation supports the phase delay of about a whole
period of 11 years between the poloidal and toroidal fields found by Stix (1976). However,
as can be seen the full separation is sometimes violated during the periods coinciding with
increased local activity such as the Halloween flares in 2003.
The nature of the larger period was assumed to be related to the dynamo wave governing
the whole solar activity suggested by Parker (1955) and refined by many other authors (see
the reviews by Tobias, 2002; Ossendrijver, 2003; Krivodubskij, 2005). The nature of the
shorter term periodicity still needs to be investigated.
137
Figure 3 Upper plot: The cumulative areas in square degrees 103 for sunspots in northern (solid line) and
southern (dashed line) hemispheres. Lower plot: North South asymmetry in the cumulative areas for the
sunspots in cycle 23.
3.2. Correlation with the Background Solar Magnetic Field

Because in this study we are interested in periods of the sunspot area and flux variations
longer than one year, the measured means (X i ) can be selected for every measurement by
averaging with a one-year filter. Then if someone wishes to define the deviations from the
estimated mean within a longer period, say three four years, then this one-year mean needs
to be compared with the four-year estimated mean (see also Section 2.3). These two approaches to the mean estimations (one and four years) were utilized in the studies presented
below.
138
S. Zharkov et al.
Figure 4 The line-of-sight excess magnetic flux in sunspots measured separately in the northern and southern hemispheres from the SFC as described by Zharkov and Zharkova (2006).
Because the SMF is assumed to carry the poloidal magnetic field, whereas sunspots or
active regions are associated with the toroidal one, then, similarly to the analysis carried
out by Stix (1976), a comparison of these time series will allow the relations between these
fields for different phases of cycle 23 to be established.
3.2.1. Latitudinal Variations of the Background SMF
To understand the distribution of sunspot areas and magnetic fluxes and their dependence on
the activity cycle, it is necessary to study the latitudinal distribution of the background SMF.
The latter have been deduced and statistically processed (Gavryuseva and Kroussanova,
2003; Gavryuseva, 2006) from the data obtained by the WSO for the past 29 years as described in Section 2.2.
We calculate the SMF averaged over one year and their residuals from the estimated
means for four years for the symmetric part [SMF( ) + SMF( ); Figure 6, upper plot] and
for the asymmetric one [SMF( ) SMF( ); Figure 6, lower plot). The asymmetric part
variations have a period close to 10 11 years for a single polarity exchange or 22 years for
the whole polarity exchange. The symmetric SMF component changes its sign every 2.5
3 years, either coinciding with the sign of the leading polarity in a given hemisphere or being
opposite to it. As the cycle progresses, both the symmetric and asymmetric components
show a constant shift from the equator to the poles with their own phases.
The SMF clearly reveals the four-zonal (4Z) latitudinal structure with the opposite polarity zones with the boundaries located in 25 , 0 (two near-equatorial zones and two polar
zones), the signs of which also vary with the 22-year period. The polarities of the SMF in
139
Figure 5 The one-year-mean SMF for the symmetric component [SMF( ) + SMF( ); upper plot] and for
the antisymmetric components [SMF() SMF( ); lower plot; see Figure 1 and Section 2 for the relation
between CR and years]. The green-brown colors correspond to positive polarity and blue-navy to negative
polarity; the outer contours correspond to a 3 interval and the inner contours to a 2 interval about the
relevant maximum magnetic field (brown or blue).
the near-equatorial zones in both solar hemispheres are opposite and change in the each subsequent cycle (Hathaway, 2005; Gavryuseva, 2006). Also there is a 5.5-year lag between the
SMF polarity change in the two near-equatorial and two polar zones (Gavryuseva, 2006).
The background magnetic fields in the near-equatorial zones have opposite polarity to the
leading parts of most activity regions within. In the polar zone after five six years after
the cycle start, the polarity reverses (Gavryuseva, 2006), coinciding in time with the SMF
reversal in the subpolar zones.
A noticeable difference seems to occur between the periods of polarity change in the
opposite hemispheres: The longer period appears in the northern hemisphere and shorter
one in the southern hemisphere (PNorth 3 years and PSouth 2.5 years; Gavryuseva, 2006).
Because of this difference, the polarity waves in the hemispheres at different moments can
be either in phase or antiphase. The magnitude of the mean latitudinal magnetic field in 1996
(at the minimum of the solar activity, CR 1920 1922) was unusual (i.e., positive polarity
spreading from the north pole to latitudes of about 45 50 ).
3.2.2. The Long-Term Phase Relation between the Sunspot Excess Magnetic Flux and the
SMF
We compare these SMF variations with those of the sunspot excess magnetic flux averaged
over one Carrington rotation (Figure 6). The comparison shows that, similar to cycles 20
140
S. Zharkov et al.
Figure 6 The latitudinal distributions of the excess solar magnetic field (SMF) (upper plot) and sunspot
excess fluxes (bottom plot) in cycle 23 averaged over one Carrington rotation. Blue denotes positive polarity
and orange-yellow negative polarity.
and 21 measured by Stix (1976), these fluxes have nearly opposite polarities in the same
hemisphere during the whole cycle. This confirms that a large-scale phase relation between
the weak poloidal (SMF) and strong toroidal (sunspots) magnetic field is close to a whole
cycle period (11 years) found by Stix (1976). However, this polarity parity is not fully
fulfilled in equatorial zones or, especially, in the polar zones.
These variations fit the axisymmetric dynamo models rather well, with the phase of a full
cycle period between the poloidal and toroidal fields (see Figures 6 and 7 in Stix, 1976) and
with the upward magnetic advection in the royal zone (the orange line in Figure 7) supporting the upward buoyancy of flux tubes from the solar convective zone and their appearance on the solar surface as sunspots and active regions (Rudiger and Brandenburg, 1995;
Kitchatinov and Rudiger, 1999; Belvedere, Kuzanyan, and Sokoloff, 2000; Krivodubskij,
2005).
3.2.3. The Short-Term Phase Relation between the SMF and the Sunspot EMF
To understand the nature of short-term oscillations in sunspot characteristics, let us rebuild
the quasi-butterfly diagrams of the daily EMF and SMF variations with one-year and fouryear filters (see also Section 2.3). Then by calculating the residuals of the sunspot areas,
one-year minus four-year, one can reveal periodicities within a scale of less than three years
(see Figure 8).
The residuals of the sunspot areas, similar to those in the SMF, vary quasi-periodically,
revealing either much lower (blue colors) or much higher (green and brown colors) areas
141
Figure 7 Meridional cross

section of the solar convective
zone with the distributions of
radial velocity of the toroidal
field advection in depth and
latitude (with the equator marked
by the horizontal axis). The
arrows denote the advection
directions: black downward
pumping suppressing the
magnetic buoyancy;
brown upward pumping
supporting the magnetic
buoyancy within the royal zone
(orange line). (Courtesy of
Krivodubskij.)
than the areas averaged over four years. Each pattern lasts for 2.5 3 years over the whole
period of the observations, with its sign reversing in the next 2.5- to 3-year period. This
indicates that such a periodicity in both EMF and the SMF could be related. Let us establish
the phase of such a relation.
To check how stable the period of short-term oscillations is (two three years, or 1/4 cycle) let us cross-correlate the latitudinal distributions in the northern and southern hemispheres for sunspot areas of the EMF and the SMF averaged over one year; the results are
shown in Figure 9. As expected, the correlation of sunspot areas in both hemispheres is
positive for a time lag of less than three years (Figure 9, upper plot) and becomes negative
for larger time lags, whereas the North South correlation of the sunspot EMF (Figure 9,
middle plot) is negative for the first 2.5 3 years and becomes positive in the next three
years.
The additional feature appearing for the excess SMF (Figure 9, bottom plot) is the polarity of each periodic structure being opposite in each hemisphere and changing signs in
2.5 3 years, like changing colors on a chess board. Basically, in Carrington rotations 1930
1940 (years 1996 1997) the polarity is negative in the northern hemisphere and positive in
the southern one; after 2.5 3 years, the situation is opposite. Then this pattern with changing polarities repeats every 2.5 3 years. Obviously, in some of these periods the polarity
of the SMF will be opposite to the leading polarity of sunspots whereas in others it will
coincide with it.
Let us correlate each of these two series related to sunspots (areas and EMF) directly with
the background SMF (see Figure 10). The variations of the total sunspot areas in cycle 23
142
S. Zharkov et al.
Figure 8 The residuals in the latitudinal distributions in cycle 23 averaged over one year minus those over
four years for excess SMF (upper plot) and excess sunspot flux (bottom plot). The green-brown color presents
the positive residuals and the blue one denotes the negative ones (see Figure 1 and Section 2 for the relation
between CR and years); the residuals for inner contours are twice as high as those for outer ones.
averaged over one year with one CR step (upper plot) and their residuals from the areas averaged over four years (bottom plot) were correlated with the SMF data for different time lags.
The one-year-averaged areas show a strong positive correlation with the background SMFs
around zero time lag and a strong negative one at about 50 Carrington rotations (about three
years, or 1/4 cycle). The time lag increases from the polar zone in the southern hemisphere
through the equator to the polar zone in the northern one.
The residuals for the 1 y 4 y sunspot areas and the SMF (Figure 10, bottom plot) reveal
that, similarly to SMF variations, the correlation also has a four-zone structure: The time
lag increases in the two zones in each hemispheres, from the top of the royal zones toward
either the equator or the poles. The former corresponds to the SMF shift and the migration
of sunspot locations shown in the butterfly diagrams in Section 3.1.1; the latter corresponds
to the SMF migration toward the poles shown in Section 3.2.1. This correlation changes
its sign every 2.5 3 years, similarly to the symmetric variations of the background SMF
presented in Figure 5 in Section 3.2.1.
Moreover, correlation of the one-year minus four-year sunspot EMF distribution with
the excess SMF one (the bottom plot in Figure 11) also splits into four zones, similar to the
four zones of the background SMF reported in Section 3.2.1. In the southern near-equatorial
zone there is a positive correlation that increases in time toward the equatorial latitudes with
a time lag up to three years and in the northern near-equatorial zone there is a negative
correlation that increases toward the equator with a similar time lag. Then in the next three
143
Figure 9 Phase relation between

the SMF and sunspot areas
(upper plot), excess magnetic
flux (middle plot), and the excess
SMF (lower plot) in the northern
and southern hemispheres for the
time lags in years. Numbers
labeling the contours show the
correlation coefficients.
years the signs of the correlation change in each near-equatorial zone to the opposite one.
In the zones from the poles toward the top of the royal zones in each hemisphere at the
144
S. Zharkov et al.
Figure 10 Correlation between sunspot areas and the SMF averaged over one year with one CR step (top
plot) and their residuals from those averaged over four years (bottom plot). The green and brown (blue) colors
indicate positive (negative) correlation; the numbers labeling the contours denote the correlation coefficients
(0.6, 0.4, and 0.2 from inner to outer contours, respectively).
145
Figure 11 Correlation between sunspot excess magnetic field and the SMF averaged over one year with one
CR step (top plot) and their residuals from those averaged over four years (bottom plot). The green and brown
(blue) colors indicate positive (negative) correlation; the numbers labeling the contours denote the correlation
coefficients (0.6, 0.4, and 0.2).
146
S. Zharkov et al.
beginning of the cycle a negative correlation is established with a four- to five-year lag,
changing to the positive one lasting for next five six years of the cycle.
This correlation together with the correlation of the absolute magnetic flux (magnetic
field strength confined in sunspots) with sunspot areas and EMF (top and bottom plots in
Figure 11, respectively) lends support to the suggestion that, in addition to a long phase
relation of about a whole cycle (11 years) between the poloidal and toroidal fields, there is
also a shorter phase of about 1/4 cycle 2.75 years when the signs of the SMF polarity
in each hemisphere are changed. These changes seem to modulate the borders of the royal
zone where magnetic pumping allows flux tubes to float to the solar surface (the red line
in Figure 7). This, in turn, leads to the asymmetric fluctuations of the symmetric component
of the poloidal field in the northern and the southern hemispheres (see Figure 6 in Stix,
1976). If the sign of the SMF in this period coincides with the leading polarity sign, the
SMF suppresses flux-tube buoyancy and the frequency of sunspot appearances; if this sign
is opposite to the leading polarity, it increases sunspot appearances; and when it coincides
with the leading polarity, it suppresses them.
4. Conclusions
We have established a correlation between the appearances of magnetic flux tubes as
sunspots at different latitudes of each hemisphere and the latitudinal variations of solar magnetic field during different phases of solar cycle 23.
We have found that there are four zones in the SMF distributions, leading to four similar
zones in sunspot areas and EMF distributions: two near-equatorial zones located between
the latitudes of 40 and + 40 , called the royal zone (Figure 7), and two polar zones
located above these latitudes. Similarly to the previous cycles, the locations of sunspots are
confined within the royal zone, as expected from a magnetic pumping effect producing
the upward velocities of toroidal field advection (Krivodubskij, 2005). At the beginning of
the cycle the formation of sunspots is located at higher latitudes; as the cycle progresses the
formation zone slowly moves toward the equator. In both subpolar zones the sunspots and
their EMF move toward the poles.
Additionally, in the butterfly diagrams for sunspot areas and EMF built in the royal
zone, one can see a striplike fine structure in latitude appearing during the activity period
as vertical threads. This feature points to a strong increase in the solar surface areas covered by sunspots at some latitudes in one hemisphere or another. Then over the subsequent
time interval the sunspot area/EMF increase shifts to the opposite hemisphere, revealing a
repeating striplike, or a quasi-periodic, temporal structure.
A period of about 2.5 3.0 years can be extracted from these strip appearances in sunspot
areas and excess magnetic flux if the values are averaged over one year and their residuals
are defined from those averaged over four years. These periods are also confirmed by strong
positive correlation with the symmetric part of the background solar magnetic field, which
itself reveals a well-defined four-zonal latitudinal structure in each hemisphere with the
boundaries located along the tops of the royal zones. Since the SMF is assumed to carry the
poloidal magnetic field whereas sunspots or active regions are associated with the toroidal
magnetic field, then, similar to the analysis carried out by Stix (1976), a comparison of
these time series allowed us to establish the relations between them for different phases of
cycle 23.
The overall magnetic polarities of EMF and SMF in opposite hemispheres averaged over
one Carrington rotation are opposite throughout the whole cycle. Although their magnetic
147
polarities averaged by one year in the symmetric latitudinal zones of the northern and southern hemispheres are also opposite, they change their signs periodically every 2.5 3 years,
either coinciding with the leading polarity in a given hemisphere or being opposite to it. The
correlation between the SMF and EMF is either negative when the SMF coincides with the
leading magnetic polarity and suppresses the magnetic flux tube buoyancy to the surface or
positive when the SMF is opposite to the leading polarity, thus reducing the total magnetic
field and supporting the flux tube appearance on the surface.
This, in turn, leads to a period of 2.5 3 years when in one hemisphere its SMF favors
sunspot formation, resulting in higher sunspot areas and EMF averaged by one year than in
the other hemisphere; then the polarities swap and the sunspot areas and EMF in the other
hemisphere become higher. Therefore, the magnetic field confined in the sunspot EMF and
background SMF reveals two periods of their phase relation: a long-term one of 11 years,
similar to those reported by Stix (1976), and a short-term period of about 2.5 3 years (or
close to one-quarter cycle) established in this paper.
These modulations of sunspot characteristics by the SMF can explain the persistent periodic North South asymmetries in sunspot areas and excess magnetic flux detected from the
SFCs for cycle 23. Evidently, to determine the dynamo mechanisms capable of accounting
for the short-term ( one-quarter cycle) oscillations in the SMF that lead to the observed
variations of sunspot areas and EMF reported in the present study, the qualitative explanation of the periodicity presented in this paper requires the support of theoretical simulations
with nonaxisymmetric solar dynamo models. This topic will be discussed in a future paper.
Acknowledgements The authors thank the referee for the constructive comments from which the paper
strongly benefited. E.G. thanks the University of Bradford for hospitality and inspiration during her visit
when the paper was initiated. S.Z. acknowledges the funding from STFC, UK, that partially supported this
research.
References
Belvedere, G., Kuzanyan, K.M., Sokoloff, D.: 2000, Mon. Not. Roy. Astron. Soc. 315, 778.
Dikpati, M., Gilman, P.A.: 2001, Astrophys. J. 559, 428.
Gavryuseva, E.: 2006, News Acad. Sci. Izv. RAN Ser. Phys. 70(1), 102.
Gavryuseva, E., Kroussanova, N.: 2003, In: Proc. of the 10th International Solar Wind Conference, Conf.
Proc. 679, Am. Inst. Phys., Melville, 242.
Hathaway, D.H.: 2005, In: Sankarasubramanian, K., Penn, M., Pevtsov, A. (eds.) How Large-Scale Flows
May Influence Solar Activity? Conference Series 346, Astron. Soc. Pac., San Francisco, 19.
Hathaway, D.H., Wilson, R.M.: 2004, Solar Phys. 224, 5.
Hoeksema, J.T.: 1985, http://sun.stanford.edu.
Holder, Z., Canfield, R.C., McMullen, R.A., Nandy, D., Howard, A.F., Pevtsov, A.A.: 2004, Astrophys. J.
611, 1149.
Howard, R.F.: 1991, Solar Phys. 136, 251.
Kitchatinov, L.L., Rudiger, G.: 1999, Astron. Astrophys. 344, 911.
Krause, F., Radler, K.H.: 1980, Mean Field Magnetohydodynamics and Dynamo Theory, Academic, Berlin.
Krivodubskij, V.N.: 2005, Astron. Nachr. 326, 61.
Maunder, E.W.: 1904, Mon. Not. Roy. Astron. Soc. 64, 747.
Nandi, N., Choudhuri, A.R.: 2001, Astrophys. J. 551, 576.
Ossendrijver, M.: 2003, Astron. Astrophys. Rev. 11, 287.
Parker, E.N.: 1955, Astrophys. J. 121, 491.
Rudiger, G., Brandenburg, A.: 1995, Astron. Astrophys. 296, 557.
Scherrer, P.H., Wilcox, J.M., Svalgaard, L., Duvall, T.L., Dittmer, Ph.H., Gustafson, E.K.: 1977, Solar Phys.
54, 353.
Stix, M.: 1976, Mon. Not. Roy. Astron. Soc. 47, 243.
Temmer, M., Veronig, A., Hanslmeier, A.: 2002, Astron. Astrophys. 309, 707.
Tobias, S.M.: 2002, Phil. Trans. Roy. Soc. Lond. 360, 2741.
148
S. Zharkov et al.
Vainstein, S.I., Zeldovich, Ya.B., Ruzmaikin, A.A.: 1980, Turbulent Dynamo in Astrophysics, Nauka,
Moscow.
Yoshimura, H.: 1981, Astrophys. J. 247, 1102.
Zharkov, S.I., Zharkova, V.V.: 2006, Adv. Space Res. 38(N5), 868.
Zharkov, S.I., Zharkova, V.V., Ipson, S.S.: 2005, Solar Phys. 228, 401.
Zharkova, V.V., Zharkov, S.I.: 2007, Adv. Space Res. doi:10.1016/j.asr.2007.02.082.
Zharkova, V.V., Zharkov, S.I., Benkhalil, A.K.: 2005, Mem. Soc. Astron. Ital. 75, 1072.
Zharkova, V.V., Ipson, S.S., Zharkov, S.I., Benkhalil, A., Aboudarham, J., Fuller, N.: 2005, Solar Phys. 228,
134.
Comparison of Five Numerical Codes for Automated

Tracing of Coronal Loops
Markus J. Aschwanden Jong Kwan Lee
G. Allen Gary Michael Smith Bernd Inhester

Abstract The three-dimensional (3D) modeling of coronal loops and filaments requires algorithms that automatically trace curvilinear features in solar EUV or soft X-ray images. We
compare five existing algorithms that have been developed and customized to trace curvilinear features in solar images: i) the oriented-connectivity method (OCM), which is an
extension of the Strous pixel-labeling algorithm (developed by Lee, Newman, and Gary);
ii) the dynamic aperture-based loop-segmentation method (developed by Lee, Newman, and
Gary); iii) unbiased detection of curvilinear structures (developed by Steger, Raghupathy,
and Smith); iv) the oriented-direction method (developed by Aschwanden); and v) ridge detection by automated scaling (developed by Inhester). We test the five existing numerical
codes with a TRACE image that shows a bipolar active region and contains over 100 discernable loops. We evaluate the performance of the five codes by comparing the cumulative
distribution of loop lengths, the median and maximum loop length, the completeness or
detection efficiency, the accuracy, and flux sensitivity. These algorithms are useful for the
M.J. Aschwanden ()
Solar and Astrophysics Laboratory, Lockheed Martin Advanced Technology Center, Department
ADBS, Building 252, 3251 Hanover Street, Palo Alto, CA 94304, USA
e-mail: aschwanden@lmsal.com
J.K. Lee
Department of Computer Science, University of Alabama in Huntsville, Huntsville, AL 35899, USA
e-mail: jlee@cs.uah.edu
G.A. Gary
National Space Science and Technology Center, University of Alabama in Huntsville, Huntsville, AL
35899, USA
e-mail: Allen.Gary@nasa.gov
M. Smith
Mullard Space Science Laboratory, Holmbury St. Mary, Dorking, Surrey RH5 6NT, UK
e-mail: ms2@mssl.uc.ac.uk
B. Inhester
Max-Planck Institute for Solar System Research, 37191 Katlenburg-Lindau, Germany
e-mail: binhest@mps.mpg.de
DOI: 10.1007/978-0-387-98154-3_12
149
150
M.J. Aschwanden et al.
reconstruction of the 3D geometry of coronal loops from stereoscopic observations with the
STEREO spacecraft, or for quantitative comparisons of observed EUV loop geometries with
(nonlinear force-free) magnetic field extrapolation models.
Keywords Corona: structures Sun: EUV Methods: pattern recognition
1. Introduction
Objective data analysis of solar images requires automated numerical codes that supplement arbitrary visual/manual pattern recognition by rigorous mathematical rules. Generally,
solar images display area-like features (active region plages, sunspots, flare ribbons, partialhalo coronal mass ejections, etc.) as well as curvilinear features (coronal loops, filaments,
prominence fine structure, etc.). In this paper, we focus on automated detection of curvilinear features that appear in solar images recorded in extreme ultraviolet (EUV), soft X rays,
white light, or H (Aschwanden, 2004).
A recent review on two-dimensional (2D) feature recognition and three-dimensional (3D)
reconstruction in solar EUV images is given in Aschwanden (2005). Solar EUV imaging
started with the Extreme-ultraviolet Imaging Telescope (EIT) and Coronal Diagnostic Spectrometer (CDS) on the Solar and Heliospheric Observatory (SOHO), the Transition and
Coronal Explorer (TRACE), the Extreme Ultraviolet Imager (EUVI) on the STEREO spacecraft, and the Extreme ultraviolet Imaging Spectrometer (EIS), and now includes also X rays
with the X-ray Telescope (XRT) on the Hinode spacecraft and the future Atmospheric Imaging Assembly (AIA) on the Solar Dynamics Observatory (SDO).
The basic premise of an automated tracing algorithm is to achieve an objective method
to identify and measure the geometry of curvilinear features. For coronal loops for instance,
such a method can identify all detectable loops in an image and measure their spatial coordinates [x(s), y(s)], their lengths L, and their curvature radii R(s), as a function of a length
coordinate s. Ideally, we want to trace loops from one to the other footpoint, but automated
tracing may often be restricted to partial segments, missing parts with low densities (near
loop tops or the cooler footpoints), or be misled by crossing loops in the foreground and
background. Nevertheless, such a capability allows one to also count detectable loop structures in an image in an automated way, to track the same loop structure in a time sequence of
images, and to detect time-dependent spatial variations, such as eigen-motions, velocities,
oscillatory motions, and wave motions along loops. Such an algorithm can then be employed
for blind searches of structures with selected velocities, oscillation periods, or amplitudes.
For large data sets in particular, such as the upcoming SDO mission, automated detection
methods will be inevitable to find structures of interest at all for a particular science study.
Automated loop-tracing algorithms will be particularly useful for 3D reconstruction of loop
geometries using STEREO data, as well as for constraining theoretical magnetic field extrapolation models (e.g., see Wiegelmann and Inhester, 2006).
2. Automated Loop-Tracing Codes

There exist a large number of numerical, automated, pattern recognition codes, specialized
either for area-like features or on curvilinear features. However, there exists no general pattern recognition code that has a superior performance in all kinds of data, so the best code
for a particular data set needs to be tailored and customized to the particular morphological
Comparison of Five Numerical Codes for Automated Tracing
151
properties of the given data set. Here we are mostly interested in the automated detection of
soft-edge curvilinear features in solar images, and thus we largely ignore numerical codes
that work on nonsolar images, such as used for analysis of medical, geological, or geographical images. To our knowledge, there exist so far five codes that are specialized for
solar images, all of which were developed in recent years by four independent groups. We
briefly describe the numerical algorithms of these five codes in this section and provide a
comparison of their performance in Section 3.
2.1. The Oriented-Connectivity Method
The oriented connectivity-based method (OCM) (Lee, Newman, and Gary, 2004, 2006b) is
the first automated coronal-loop segmentation method that performs its loop segmentation
via a constructive edge linkage method. Since the coronal loops are the vestiges of the solar
magnetic field, OCMs edge linkage is based on model-guided processing, which exploits
external estimates of the magnetic fields local orientation derived from a simple dipolar
magnetic field model. OCM includes three processing steps: i) preprocessing, ii) modelguided linkage, and iii) postprocessing.
The goal of the preprocessing step is to remove pixels that are very unlikely to be coronal loop pixels and to precompute the estimates of the magnetic fields local orientation.
A median filtering and unsharp masking (i.e., contrast enhancement) are applied to remove
the image noise and to sharpen coronal loop structures. Strous loop pixel labeling algorithm (Strous, 2000) is also used to determine the possible loop pixels followed by medianintensity thresholdings to reduce the number of falsely labeled loop pixels. In addition to
these image cleaning steps, the magnetic fields local orientation is restricted by a dipolar
magnetic field model by using a magnetogram, e.g., from the Michelson Doppler Imager
(MDI) onboard the SOHO spacecraft. In particular, the local orientation estimation exploits
the multipolar field model by considering a set of estimates of the 3D-to-2D-projected magnetic fields orientation.
Using the potential loop pixels and the magnetic fields local orientation determined in
the preprocessing step, the OCM performs coronal-loop segmentation via a constructive
edge linkage guided by the magnetic fields orientation (i.e., magnetic field local orientation
estimates are used to progressively link loop pixels with consistent orientation). Starting
from a loop pixel determined in a preprocessing step, the OCM segments loop pixels by
forming a clustering of all the other pixels that define the same loop structure. This forming
of a clustering of loop pixels is a stepwise process, which at each step adds one loop pixel
to the current loop. The selection of the best loop pixel exploits a weighting scheme (i.e.,
distance-, intensity-, angular-, and tangent-based weighting) among all the candidate pixels
that are determined using the local orientation estimates.
In the postprocessing step, a (coronal loop) spline fitting and a linkage step followed
by another spline fitting are applied to join the disconnected subsegments of the loops and
remove the aliasing. The first B-spline fitting is designed to produce smooth loop structures
and a simple linkage step and the second B-spline fitting is applied to merge disconnected
loop segments smoothly.
2.2. The Dynamic Aperture-based Loop Segmentation Method
Carcedo et al. (2003) reported that the coronal loops have a cross-sectional intensity profile
following a Gaussian-like distribution. The dynamic aperture-based method (DAM) (Lee,
152
Newman, and Gary, 2006b) exploits the Gaussian-like shape of loop cross-sectional intensity profiles and constructively segments credible loop structures. In particular, DAM segments coronal loops via a search through the image for regions whose intensity profiles
are well-fitted by a ruled Gaussian surface (RGS). Since nearby RGSs that are on the same
loop appear to have similar Gaussian shape parameters, DAM forms loops through a linkage
process that clusters adjacent fitted RGSs if their shape parameters are similar. In addition,
the clustering joins fitted RGSs only if they have similar orientation, which is determined
by applying principal component analysis (e.g., the arctangent of the maximum eigenvector
components is used to estimate the loop angular direction) on the RGSs. DAM also includes
preprocessing steps to remove nonloop structures and postprocessing steps to remove the
aliasing and to join the disconnected loop segments. The preprocessing step, in particular,
is designed to remove the image noise, to enhance the contrast between the loops and the
background, and to remove nonloop structures by median filtering and high-boosting followed by a global mean-intensity thresholding. (The postprocessing steps used in DAM are
the same as the postprocessing steps used in OCM.)
2.3. Unbiased Detection of Curvilinear Structures Method (UDM)
This code for unbiased detection of curvilinear structures, developed by Steger (1996), detects a curvilinear feature from the local edge contrast, while also taking the geometry of
its surroundings into account. Essentially, the centroid position of a curvilinear structure is
determined from the second derivative in the direction perpendicular to the curvilinear structure, which yields a stable position with subpixel accuracy that is independent of the width
and asymmetric profile of the structure (and thus unbiased). Such tracings of curvilinear
structures have been successfully applied to photogrammetric and remote-sensing images
from satellites and airplanes (tracing roads, railways, or rivers), as well as to medical imaging (tracing blood vessels from X-ray angiograms or bones in the skull from CAT-scan or
magnetic resonance images) (Steger, 1996). The code from Steger was further explored in
the work of Raghupathy (2004), who optimizes the connection (linkage) or loop segments
at junctions, crossovers, or other interruptions, using the eigenvalues and eigenvectors of
the Hessian matrix of second-order derivatives, as well as a generalized Radon transform
(essentially the integral of the image along a traced curve). Thus, Raghupathys approach
breaks the loop-tracing problem down into i) a local search of loop segments using the gradients among neighbored pixels and ii) a global search of identifying connections between
disconnected curve segments using the generalized Radon transform. The algorithm of Steger (without Raghupathys modification) has been rewritten in the computer language C
by Smith (2005) for the SDO team at the Mullard Space Science Laboratory (MSSL) and
applied to TRACE images.
2.4. Oriented-Directivity Loop Tracing Method
The oriented-directivity loop tracing method (ODM) is currently being developed by Aschwanden, with testing still in progress. The ODM code uses directional information for
guiding the tracing of coronal loops, similar to the OCM code, but makes use only of the
local directivity, whereas the OCM code estimates the global connectivity between the endpoints of the loops (by using a priori information derived from magnetic field models). The
code consists essentially of four parts; i) preprocessing of the image to render the highest
contrast for fine loops, ii) finding starting points of loops, iii) bidirectional tracing of the loop
segments to the two opposite endpoints, and iv) reducing data noise by smoothing of loop
153
curvatures. In the preprocessing step, several high-pass filters are applied over the range of
loop widths of interest (say, approximately three to seven TRACE pixels here). In the second step, an arbitrary starting point of a candidate loop is identified by searching for local
(intensity) flux maxima in the high-pass-filtered image in each macropixel (say, with a size
of ten pixels) of the full image. In a third step, a loop is traced by moving a search box (e.g.,
of size ten ten pixels) along the expected local direction in small steps (say, five pixels).
The local direction is determined by performing a linear regression fit to the flux maxima
positions inside the search box, where local maxima with weak flux below some threshold
and with positions outside the expected direction (within some tolerance, say, three pixels)
are ignored. The direction of a traced loop is then updated at each subsequent position and
used for extrapolating the next loop position. If no suitable direction is found or if the flux is
below the threshold, the loop segment is terminated. The loop tracing is conducted in both
directions from the initial starting point to obtain a full loop segment. In the last step, the
curvature of each loop segment is smoothed with a median filter to reduce the inherent data
noise. The ODM code was applied to TRACE images and compared with visual/manual
loop tracings.
2.5. Ridge Detection by Automated Scaling
The ridge detection by automated scaling (RAS) method can be considered as an extension of the OCM by Lee, Newman, and Gary (2006a) and was developed by Inhester. The
improvements consist of replacing Strous ridge point detection scheme by a modified multiscale approach of Lindeberg (1998), which automatically adjusts to varying loop thickness
and also returns an estimate of the reliability of the ridge point location and orientation (Inhester, Feng, and Wiegelmann, 2007). The connection of loop segments is accomplished by
geometrical principles that include the orientation of the loop at the ridge points with the
co-circularity constraint proposed by Parent and Zucker (1989). The code consists of three
modules: i) search for ridgel location, ii) ridgel connection to chains or more complete loops,
and iii) curve fits to the ridgel chains to obtain smooth, spline-represented loop curves. In the
first step, the determination of the characteristic points at ridge centers is computed from the
Taylor coefficients of the local regularized differentiation, where the optimum regularization
parameter is determined from a maximum of the quality function q = d (|h1 | |h2 |). This
function q depends on the window size d and the eigenvalues hi of the second-derivative
matrix, which for ridges has to satisfy h2 < h1 . The Taylor coefficients interpolated at the
center of the ridge yield the position, orientation, and quality of the ridgel. In the second step, ridgels are combined into chains, where all possible ridgel ridgel connections are
weighted according to their binding energy, specified by the mutual distance and orientation, the latter being quantified by the co-circularity measure of Parent and Zucker (1989).
In the third step, ridgel chains are smoothly connected by polynomial fits, which balance
curvature, distance, and orientations. The mathematical framework of this RAS code and an
application to first EUVI/SECCHI images has been described in Inhester, Feng, and Wiegelmann (2007).
3. Test Comparisons of Loop-Tracing Codes

In this section we compare the five different codes described here by applying them to the
same test image, which was chosen to be the same as used in some earlier code demonstrations (Lee, Newman, and Gary, 2004, 2006a, 2006b; Smith, 2005).
154
Figure 1 Top: Original 171

TRACE image of 19 May 1998,
22:21 UT, with a size of
1024 1024 pixels. Black means
high fluxes. The circular white
boundary is caused by the
vignetting of the telescope (not to
be confused with the solar limb).
Bottom: High-pass-filtered
TRACE image, where a
7 7-boxcar smoothed image
was subtracted from the original
image. The background also
reveals some residual nonsolar
spikes, JPEG compression
artifacts, and diagonal
quasiperiodic ripples caused by
electromagnetic interference
(EMI) in the CCD readout.
3.1. Test Image and High-Pass Filtering

We show the test image in Figure 1, which is an EUV image observed with the TRACE
telescope on 19 May 1998, 22:21:43 UT, with an exposure time of 23.172 seconds in
the wavelength of 171 . The image consists of 1024 1024 pixels, with a pixel size of
0.5 . The pointing of the TRACE telescope was near disk center (i.e., the Sun center is at
XCEN = 422.027 and YCEN = 401.147 pixels), which is slightly southeast of the image
center [512, 512]. For count statistics and contrast we report the average flux value in the
image, which is 146 62 DN (data numbers), with a minimum value of 56 DN and a max-
155
imum of 2606 DN. The image has been processed for flat-fielding and spike removal (with
the standard TRACE_PREP procedure).
The original image is shown on a logarithmic gray scale in the top frame of Figure 1,
using an inverted gray scale (where black means high fluxes). A high-pass-filtered version
is shown in the bottom frame of Figure 1, where a smoothed image (with a seven seven
boxcar) was subtracted from the original image to enhance the fine structure of coronal
loops. The image reveals an active region with a fairly dipolar magnetic field, where over
100 individual coronal loops (or segments) are visible, mostly bright in the lowest vertical
density scale height (of 50 000 km at a plasma temperature of T = 1.0 MK) above the
solar surface.
3.2. Manual Tracing of Loops
The 171 and 195 TRACE images are most suitable for coronal loop tracing. To enhance
the finest loop strands we apply first a high-pass filter by subtracting a smoothed image
IS (x, y) from the original image IO (x, y), a method that is also called unsharp-masking.
The high-pass-filtered image IF (x, y) is defined as
IF (x, y) = IO (x, y) IS (x, y),
(1)
where a boxcar smoothing is applied with typically Nsm = 5, 7, 9, or 11 pixels to IS (x, y)

(see the example in Figure 1, bottom, or the enlargement in Figure 2, bottom). The filtered
image IF (x, y) is then enlarged by a factor of two or three on the computer screen to enable visual/manual tracing with subpixel accuracy (i.e., the position of the spline points).
Using an interactive cursor function, one then selects a number of spline points [xi , yi ],
i = 1, . . . , n, on the enlarged screen display, with typically n = 5 for short loops or n = 10
for long loops. A higher resolution of the loop coordinate points is then obtained by interpolating the spline coordinate points [xi , yi ], i = 1, . . . , n, with a two-dimensional spline
function (called SPLINE_P in IDL) with typically N = 8n times higher resolution, yielding
N 40 coordinate points for short loops, or N 80 coordinate points for longer loops.
The two-dimensional cubic spline fit usually follows the curvature of the traced loops much
more smoothly than manual clicking. Since short loops are nearly semicircular, three spline
points would be the absolute minimum to interpolate a circle with constant curvature radius,
and five to ten spline points allow us to follow steady changes of the curvature radius with
sufficient accuracy without having redundant spline points. In the TRACE image shown in
Figure 2 (bottom), we identified some 210 individual loops (Figure 2, top).
This method has been used to trace coronal loops in EIT 171, 195, and 284 data
(Figure 4 in Aschwanden et al., 1999a; Figure 1 in Aschwanden et al., 2000a) and in TRACE
171, 195, and 284 TRACE data (Figure 1 in Aschwanden et al., 1999b, 2002; Figures
3a 3d in Aschwanden, Nightingale, and Alexander, 2000b; Figure 8 in Aschwanden and
Alexander, 2001; Figures 4 8 in Yan et al., 2001; and in Aschwanden and Nightingale,
2005).
3.3. Automated Tracing of Loops
In Figures 3 5, we show the results of automated loop tracings with the five codes described
in Sections 2.1 2.5, with the same field of view as in Figure 2, which covers a subimage
with pixel ranges of x = 200 1000 and y = 150 850 from the original 1024 1024 image
shown in Figure 1. The same TRACE image (Figure 1, top) was given to the four originators
of the five codes, with the only instruction being to provide the loop coordinates (xiL , yiL )
156
Figure 2 Top: Enlarged central part of the TRACE image (x-pixel range of 200 1000, y-pixel range of
150 850) showing the 210 manually traced loops. Bottom: High-pass filtered image of the partial image
used for manual tracing. White means enhanced flux. The regions A and B are shown in more detail in Figures
7 and 8.
157
Figure 3 Top: Automated

tracing of 76 loops with the OCM
code. Bottom: Automated tracing
of 82 loops with the DAM code.
The coordinate grid has a step of
100 image pixels.
for a suitable number of loops (L), ignoring the shortest loop segments with lengths of 20
pixels, to suppress unwanted nonloop structures (such as the reticulated moss structure seen
in the central part of the TRACE image).
3.3.1. Automated Tracing with the OCM Code
The result of the OCM code is shown in Figure 3 (top), which in addition to the TRACE
image made use of a near-simultaneous magnetogram from SOHO/MDI. The OCM code
finds 76 loops, which appear fairly smooth and are mostly nonintersecting.
3.3.2. Automated Tracing with the DAM Code
The result of the DAM code, shown in Figure 3 (bottom), is very similar to the result of the
OCM code (Figure 3, top). The DAM code also finds a similar number of loops (n = 82),
located at almost the same locations as those of the OCM code.
158

tracing of 210 loops with the
UDM code. Bottom: Automated
ODM code.
3.3.3. Automated Tracing with the UDM Code

The result of the UDM code, shown in Figure 4 (top), comprises n = 210 loop segments.
The algorithm (written in C) was run with a of 1.0 (where pertains to the Gaussian
kernel used in convolving the test image) and a minimum loop length of Lloop > 20 pixels.
(We add also a note of caution to future users of this code that the C code has a column row
indexing of the image matrix that is opposite to that of IDL, and thus it produces an output
with inverted x and y coordinates.)
3.3.4. Automated Tracing with the ODM Code
A result of the ODM code, shown in Figure 4 (bottom), gives a total of n = 330 loop segments. The preprocessing was done with high-pass filtering with boxcars of sizes three,
five, and seven pixels. The run of the ODM code had the following parameter settings:
Nmacro = 10 pixels for the box size of the search of starting points, Nbox = 10 for the size of
159

RAS code. Bottom: Overlay of
OCM + DAM + UDM + ODM
codes.
the moving box, Nstep = 5 pixels for the stepping along loops, Ndist = 2 pixels for the minimum distance to an adjacent loop, max = 1.0 pixel for the maximum deviation of peaks
considered in the (directional) linear regression fit, Lloop = 30 pixels for the minimum length
of loop segments, and Fthresh = 0.3 DN for the minimum flux threshold. The output of the
ODM code is similar to the UDM code, showing many more short loop segments than the
OCM and DAM codes (shown in Figure 3).
3.3.5. Automated Tracing with the RAS Code
The result of the RAS code, shown in Figure 5 (top), gives a total of n = 347 loop segments.
The parameters used for this run were hmax = 5 pixels for the maximum distance between
ridgels, rmin = 30 pixels for the minimum curvature radius, and amax = 25 for the maximum
deviation of fit normal to the ridgel orientation. The output of this code reveals the largest
number of details, identifying a similar number of loop structures in the central region but
more segments in the outer (northwestern) part of the active region than the other four codes
combined together (Figure 5, bottom).
160
Figure 6 The cumulative size distribution N (> Lloop ) of detected loop lengths L shown for all five codes,
plus the manual method. The maximum detected lengths are indicated on the right side, from Lm = 244 pixels
for the ODM code to Lm = 567 pixels for the DAM code. The relative number of detected loop segments
with a length longer than L = 70 pixels is shown on the left, ranging from N (L > 70) = 30 for the UDM
code to N (L > 70) = 91 for the RAS code.
3.4. Quantitative Comparison of Automated Tracing Codes

We compare now some quantitative criteria among the different codes, such as the cumulative distribution of loop lengths, the maximum and median detected loop lengths, and the
completeness of detection.
3.4.1. Cumulative Distribution of Loop Lengths
A first characterization is the length L of the identified loops. In Figure 6 we show the
cumulative distribution N (L > Lloop ) of the number of loops that are longer than a given
loop length Lloop . This cumulative distribution is simply constructed by counting all loops
within the given length ranges.
The obtained cumulative distributions show different functional forms; some are closer
to a power-law distribution (e.g., the UDM code), whereas others are almost exponential
161
Table 1 Parameters of cumulative loop length distributions (Figure 6).

Numerical
Power-law
Ratio of
Ratio of
Ratio of
code
slope
maximum
median
number of
detected
detected
detected
loop length
loop length
loops
Lm /Lm,0
Lmed /Lmed,0
N (> 70)/N0 (> 70)

1.00
Manual
2.8 0.04
1.00
1.00
OCM
2.3 0.1
0.97
0.93
0.35
DAM
2.0 0.1
1.22
0.67
0.27
UDM
2.5 0.1
0.62
0.31
0.19
ODM
3.2 0.2
0.53
0.37
0.36
RAS
2.8 0.04
0.95
0.47
0.59
(e.g., the ODM code). The difference in the cumulative distributions is mostly at short loop
lengths, say at L 100 pixels, whereas they are nearly power-law-like at larger lengths
of L 100 pixels. The manually traced loops exhibit an almost exponential distribution
(dashed curve in Figure 6). It would be premature to interpret the functional shape of these
distributions, because they contain a large number of incomplete loop segments and thus
may be quite different from the corresponding distributions of complete loop lengths. Most
distributions have a similar power-law slope of 2.0, . . . , 3.2 (Table 1) in the loop
length range of L > 100 pixels, which corresponds to slopes of = 1 3.0, . . . , 4.2
for the differential frequency distribution N (Lloop ) L of loop sizes. The automatically
traced loops have a distribution with a similar slope as the manually traced selection of loops
( = 2.8, = 3.8). However, our loop-length distributions obtained here are somewhat
steeper than those inferred in earlier works [e.g., 2.1 for nanoflare loops with lengths
in the range of L 2 20 Mm (Aschwanden et al., 2000b) and 2.9 0.2 for TRACE
171 loops with lengths of L 5 50 Mm, 2.6 0.2 for TRACE 195 loops with
lengths of L 6 70 Mm, and 2.3 0.3 for Yohkoh/SXT loops with lengths of L 2
10 Mm (Figure 5 in Aschwanden and Parnell, 2002)], probably because of the detection
of partial, and thus incomplete, loop segments. Regarding the absolute number of loops
detected in an image, we note that it depends on a number of variables in the pre- and
postprocessing stage, as well as on the particular algorithms of connecting loop segments.
3.4.2. Maximum Detected Loop Lengths
The maximum detected loop lengths (indicated on the right side of Figure 6) are an indication of the robustness of the automated tracing codes to trace long loops, despite the
unavoidable interruptions or intersections caused by crossing background loops, or resulting from the weaker fluxes near loop tops, since the electron density in the upper corona (in
particular above one hydrostatic density scale height) fades out exponentially with altitude.
The longest loop was traced with the DAM code with a length of Lm = 567 pixels ( 205 Mm), followed by OCM with Lm = 447 pixels ( 162 Mm) and RAS with
Lm = 440 pixels ( 159 Mm), whereas the UDM code produces only Lm = 287 pixels
( 104 Mm) and the ODM yields Lm = 244 pixels ( 88 Mm). If we take the longest manually traced loop as a normalization, with L0 = 463 Mm ( 167 Mm), the various codes
achieved the following relative ratios: Lm /Lm,0 = 1.22 for DAM, 0.97 for OCM, 0.95 for
RAS, 0.62 for UDM, and 0.53 for ODM. These ratios are a good measure of the codes
162
ability to trace the longest loops. Figure 3 (bottom panel) actually reveals that the longest
loop traced with DAM probably consists of two loops that have been erroneously connected
(i.e., the falsely connected loops had end points that were close to each other and had similar
slopes). The second-longest loop traced with DAM (see the distribution in Figure 6) actually matches closely the longest traced manual loop. Thus, three of the codes (OCM, DAM,
and RAS) are capable of tracing loops as long as the manually traced ones, and two codes
fall a bit short (UDM and ODM). If we determine the median length of detected loop segments we obtain the following ratios with regard to the manually detected loops (Table 1):
Lmed /Lmed,0 = 0.93 (OCM), 0.67 (DAM), 0.60 (ODM), 0.47 (RAS), and 0.31 (UDM).
3.4.3. Completeness of Loop Detection
The number of detected loops obtained here cannot directly be compared to evaluate the
completeness of the various codes, because different criteria have been used for the minimum length. To evaluate the relative completeness of the various codes, we compare the
number of loops above an identical minimum length, say at an intermediate size of L = 70
pixels ( 25 Mm). Comparing the cumulative distributions at this intermediate value (see
the vertical dashed line at L = 70 pixels in Figure 6), we find that the number of manually
traced loops is N0 (L > 70) = 154, whereas the various codes detected 91 (RAS), 55 (OCM),
48 (ODM), 41 (DAM), and 30 (UDM) loops, which varies from 59% down to 19%. These
relative ratios N (L > 70)/N0 (> 70) (Table 1) provide an approximate measure of the relative completeness of loop detection for the various codes. The completeness of the various
codes can probably be adjusted by lowering the flux threshold and the minimum length of
detected loop segments, as long as the code does not start to pick up random structures in
the noise. The fact that all codes pick up no significant structures in the lower left corner of
the image indicates that all codes have clearly been adjusted to a safe value above the noise
threshold.
3.4.4. Accuracy and Sensitivity of Loop Detection
To investigate the accuracy of automated loop detection, we compare two selected regions in
more detail. The two regions are marked in Figure 2, where region A (Figure 7) comprises
the x-range 300 500 (pixels) and the y-range 600 800 of the original image, and region
B (Figure 8) comprises the x-range 725 925 and the y-range 350 550. We visualize the
fine structure of the loops with a gray-scale image that includes a superposition of three
high-pass filters (using smoothing with a boxcar of three, five, and seven pixels). The highpass-filtered image regions are shown in the bottom right panels of Figures 7 and 8, and a
contour map of the high-pass-filtered image is also overlaid in each of the other panels. The
automated curve tracings of the five different codes are displayed with thick black curves in
Figures 7 and 8. From the contour maps one can clearly see the noisy ridges of the loops
and compare in detail which ridges have been successfully traced, and which ones have been
missed or erroneously connected. It appears that each code can be improved to some extent.
The OCM, DAM, and UDM codes seem to trace a smaller number of structures than the
other codes, and thus the sensitivity could be lowered. The OCM and DAM codes also have
a tendency to misconnect the ridges of closely spaced near-parallel loops. The UDM code
seems to follow the ridges fairly exactly but has tends to stop at shorter segments of the loops
than the other codes (see also the median loop length of Lmed /Lmed,0 0.31 in Table 1). The
ODM code seems to be more complete in tracing all ridges, but it appears to pick up a few
spurious loop-unrelated, noisy structures. The RAS code seems to score somewhere between
163
Figure 7 Detailed tracing in image section A (marked in Figure 2: x-range 300 550, y-range 600 800)
by the five different codes. Each panel shows a contour plot of the high-pass-filtered image with contours at
0.5, 1.5, . . . , 4.5 DN; a gray-scale image is shown in the bottom right panel.
the UDM and ODM codes regarding completeness, but it misconnects a few spurious loopunrelated, strongly curved structures. We hope that such detailed comparisons, as shown in
Figures 7 and 8, stimulate further improvements of the various codes. Some improvements
probably can already be achieved with the existing codes by adjusting the built-in control
parameters.
164
Figure 8 Detailed tracing in image section B (marked in Figure 2: x-range 725 925, y-range 350 550) by
the five different codes and manual tracing. Each panel shows a contour plot of the high-pass-filtered image
with contours at 0.5, 1.5, . . . , 4.5 DN.
3.4.5. Computation Speed of Automated Tracing Codes

The computation times of the described codes for processing a single (1000 1000) image
vary between a few seconds and a few tens of seconds for the described cases. However,
it is not meaningful to compare individual performance times at this point, because each
code was run on a different machine and most codes are still in a development phase with
intermediate test displays, and thus have not yet been optimized for speed. The running
165
time also scales linearly with the volume of the parameter space that a code is processing.
For instance, the ODM code had a running time of 5.9 seconds for the case described in
Section 3.3.4 for the given macropixel size (for searching of loop starting points), but it
took four times longer if the macropixel size was halved (compared with a running time
of 0.93 seconds for the standard TIME_TEST2 in IDL, with a Mac Power PC 970FX v3.1
processor). We expect that the optimized codes will reach processing times of a few seconds
per image.
4. Discussion and Conclusions

Every quantitative analysis of the geometry of a coronal loop requires an accurate measurement of its 2D coordinates in solar images. If we could localize an isolated loop in a solar
image, the problem would be trivial, because the cross section along the loop axis could be
fitted with some generic function (e.g., a Gaussian), and this way the loop centroid positions (xi , yj ) could accurately be measured (with subpixel accuracy) for a number of spline
points i = 1, . . . , N along the loop length, with the two endpoints i = 0 and i = N . Such accurate loop coordinates (xi , yi ), which merely mark the projected position of the loop, could
then be used to reconstruct the 3D geometry, e.g., by triangulating the same loop with two
stereoscopic spacecraft, which yield two different projections, (xiA , yiA ) and (xiB , yiB ), from
which the 3D geometry (xi , yi , zi ) can be computed by simple trigonometry. Of course, this
assumes that projection effects resulting from transparent loop structures are insignificant
(i.e., that each loop in the image projection represents a physical coronal loop).
In reality, however, there is no such thing as an isolated loop, but each loop is observed on top of a background that consists of some 103 other coronal loops along each
line of sight. [See the statistical model of Composite and Elementary Loop strands in a
Thermally Inhomogeneous Corona (CELTIC) that specifies loop structures and their coronal background in a self-consistent way (Aschwanden, Nightingale, and Boerner, 2007).]
This means that the identification of a single loop becomes very ambiguous, and a visual
definition may not be sufficient. Mostly for this reason, there is a demand for an automated
loop-tracing code that is based on mathematical criteria rather than visual judgment.
In this paper, we explored for the first time systematically the performance of five such
automated loop-tracing codes, developed by four independent groups. Because all five codes
work fully automatically without human interaction, they should be able to recover the existing information on loop coordinates in the most objective way and, ideally, should converge
to the same result within the uncertainty of the data noise. The comparison made here revealed significant differences in the performance of these five codes; for instance, the maximum loop length was detected between 53% and 122% of that obtained from manual/visual
tracing, or the median length varied between 31% and 93% of that obtained from manual
tracing. Also the detection efficiency or completeness varies substantially (i.e., the number
of detected loops with intermediate to large sizes varied between 19% and 93% of the manually traced reference set). Of course, the manual/visual tracing should not be the ultimate
arbiter in the evaluation of automated numerical codes, but it provides at least an educated
guess of how many structures are to be expected, based on the quantitative output from visual
pattern recognition. One could construct an artificial test image with a well-defined number
and distribution of loops, but such artificial test data are only useful if they accurately mimic
the real data regarding morphological structures, the distributions of parameters, and data
noise.
These experiments, therefore, can be used to adjust the control parameters of flux sensitivity, minimum length, minimum separation, minimum curvature, etc., of each code, so
166
that they can be run more consistently versus each other. Once all codes are tuned to the
same sensitivity, they should produce about the same number of detected structures, or at
least a consistent cumulative distribution of loop lengths. The control parameters can then
iteratively be adjusted until all codes produce the same results within the uncertainties of the
data noise, unless some codes have an inherent incapability to achieve the same task, or use
external (a priori) information from physical models.
In the end, even when all codes converge to the same optimum solution, we might learn
what the true limitations of automated loop recognition are. We know already that the top
parts of large coronal loops are untracable, in particular when only one single temperature
filter is used, because the emission measure drops below the noise threshold as a consequence of gravitational stratification (or other physical mechanisms). Also the footpoints of
the loops may not be visible because of the temperature drop from the corona toward the
transition region, but the cooler part of the loop footpoints may be only a negligible small
fraction of the entire loop length. Another limitation is the complexity of the background,
which can disrupt loop tracing at countless locations beyond repair. The SDO/AIA data will
greatly improve the detection on nonisothermal loops owing to its wider temperature coverage. Also, the use of time sequences of images will render loop detection more redundant,
and thus more robust. Nevertheless, even if we are only able to produce reliable measurement of partial loop segments, we will have stringent constraints for testing of theoretical
(e.g., nonlinear force-free) magnetic field models (e.g., Schrijver et al., 2006), for correcting
magnetic field solutions (e.g., Gary and Alexander, 1995), and for stereoscopic reconstruction of the 3D geometry of loops and magnetic field lines (e.g., Wiegelmann and Inhester,
2006).
Acknowledgements We acknowledge helpful discussions with David Alexander, Karel J. Schrijver, Marc
DeRosa, Paulett Liewer, and Eric DeJong. The work benefitted from the stimulating discussions at the 3rd
Solar Image Processing Workshop in Dublin, Ireland, 6 8 September 2006 and at the 5th SECCHI Consortium Meeting in Orsay, Paris, 5 8 March 2007. Part of the work was supported by NASA TRACE Contract
No. NAS5-38099 and NASA STEREO/SECCHI Contract No. N00173-02-C-2035 (administrated by NRL).
References
Aschwanden, M.J.: 2004, Physics of the Solar Corona An Introduction, Praxis, Chichester, and Springer,
Berlin.
Aschwanden, M.J.: 2005, Solar Phys. 228, 339.
Aschwanden, M.J., Alexander, D.: 2001, Solar Phys. 204, 93.
Aschwanden, M.J., Nightingale, R.W.: 2005, Astrophys. J. 633, 499.
Aschwanden, M.J., Parnell, C.E.: 2002, Astrophys. J. 572, 1048.
Aschwanden, M.J., Nightingale, R.W., Alexander, D.: 2000, Astrophys. J. 541, 1059.
Aschwanden, M.J., Nightingale, R.W., Boerner, P.: 2007, Astrophys. J. 656, 577.
Aschwanden, M.J., Newmark, J.S., Delaboudinire, J.P., Neupert, W.M., Klimchuk, J.A., Gary, G.A., PortierFornazzi, F., Zucker, A.: 1999a, Astrophys. J. 515, 842.
Aschwanden, M.J., Fletcher, L., Schrijver, C., Alexander, D.: 1999b, Astrophys. J. 520, 880.
Aschwanden, M.J., Alexander, D., Hurlburt, N., Newmark, J.S., Neupert, W.M., Klimchuk, J.A., Gary, G.A.:
2000a, Astrophys. J. 531, 1129.
Aschwanden, M.J., Tarbell, T., Nightingale, R., Schrijver, C.J., Title, A., Kankelborg, C.C., Martens, P.C.H.,
Warren, H.P.: 2000b, Astrophys. J. 535, 1047.
Aschwanden, M.J., DePontieu, B., Schrijver, C.J., Title, A.: 2002, Solar Phys. 206, 99.
Carcedo, L., Brown, D., Hood, A., Neukirch, T., Wiegelmann, T.: 2003, Solar Phys. 218, 29.
Gary, G.A., Alexander, D.: 1995, Solar Phys. 186, 123.
Inhester, B., Feng, L., Wiegelmann, T.: 2007, Solar Phys. doi:10.1007/S11207-007-9027-1.
Lee, J.K., Newman, T.S., Gary, G.A.: 2004, In: Proc. 17th Int. Conf. on Pattern Recognition (ICPR), Cambridge, United Kingdom, 315. doi:10.1007/S11207-007-9027-1.
167
Lee, J.K., Newman, T.S., Gary, G.A.: 2006a, Pattern Recognit. 39, 246.
Lee, J.K., Newman, T.S., Gary, G.A.: 2006b, In: Proc. 7th IEEE SSIAI, Denver, CO, 91
Lindeberg, T.: 1998, Int. J. Comput. Vision 30, 117.
Parent, P., Zucker, S.: 1989, IEEE Trans. Pattern Anal. Mach. Intel. 11, 823.
Raghupathy, K.: 2004, Ph.D. thesis, Cornell University.
Schrijver, C.J., DeRosa, M., Metcalf, T.R., Liu, Y., McTiernan, J., Regnier, S., Valori, G., Wheatland, M.S.,
Wiegelmann, T.: 2006, Solar Phys. 235, 161.
Smith, M.: 2005, http://www.mssl.ucl.ac.uk/twiki/bin/view/SDO/LoopRecognition.
Steger, L.H.: 1996, Technical Report FGBV-96-03, Forschungsgruppe Bildverstehen (FG BV), Informatik
IX, Technische Universitt Mnchen, Germany.
Strous, L.H.: 2000, http://www.lmsal.com/~aschwand/stereo/2000_easton/cdaw.html.
Wiegelmann, T., Inhester, B.: 2006, Solar Phys. 236, 25.
Yan, Y., Aschwanden, M.J., Wang, S.J., Deng, Y.Y.: 2001, Solar Phys. 204, 29.

B. Inhester L. Feng T. Wiegelmann

Abstract We present a procedure to extract bright loop features from solar EUV images. In
terms of image intensities, these features are elongated ridge-like intensity maxima. To discriminate the maxima, we need information about the spatial derivatives of the image intensity. Commonly, the derivative estimates are strongly affected by image noise. We therefore
use a regularized estimation of the derivative, which is then used to interpolate a discrete
vector field of ridge points; these ridgels are positioned on the ridge center and have the
intrinsic orientation of the local ridge direction. A scheme is proposed to connect ridgels to
smooth, spline-represented curves that fit the observed loops. Finally, a half-automated user
interface allows one to merge or split curves or eliminate or select loop fits obtained from this
procedure. In this paper we apply our tool to one of the first EUV images observed by the
SECCHI instrument onboard the recently launched STEREO spacecraft. We compare the
extracted loops with projected field lines computed from near-simultaneous magnetograms
measured by the SOHO/MDI Doppler imager. The field lines were calculated by using a
linear force-free field model. This comparison allows one to verify faint and spurious loop
connections produced by our segmentation tool and it also helps to prove the quality of the
magnetic-field model where well-identified loop structures comply with field-line projections. We also discuss further potential applications of our tool such as loop oscillations and
stereoscopy.
Keywords EUV images Coronal magnetic fields Image processing
1. Introduction
Solar EUV images offer a wealth of information about the structure of the solar chromosphere, transition region, and corona. Moreover, these structures are in continuous moB. Inhester () L. Feng T. Wiegelmann
Max-Planck-Institut fr Sonnensystemforschung, Max-Planck-Str. 2, 37191 Katlenburg-Lindau,
Germany
e-mail: inhester@mps.mpg.de
L. Feng
Purple Mountain Observatory, Chinese Academy of Sciences, Nanjing, China
DOI: 10.1007/978-0-387-98154-3_13
169
170
B. Inhester et al.
tion so that the information collected by EUV images of the Sun is enormous. For many
purposes this information must be reduced. A standard task for many applications (e.g., for
the comparison with projected field lines computed from a coronal magnetic-field model or
for tie-point stereoscopic reconstruction) requires the extraction of the shape of bright loops
from these images.
Solar physics shares this task of ridge detection with many other disciplines in physics
and also in other areas of research. A wealth of different approaches for the detection and
segmentation of ridges has been proposed ranging from multiscale filtering (Koller et al.,
1995; Lindeberg, 1998) and curvelet and ridgelet transforms (Starck, Donoho, and Cands, 2003) to snake and watershed algorithms (Nguyen, Worring, and van den Boomgaard,
2000) and combining detected ridge points by tensor voting (Medioni, Tang, and Lee, 2000).
These general methods however always need to be modified and optimized for specific applications. Much work in this field has been motivated by medical imaging (e.g., Jang and
Hong, 2002; Dimas, Scholz, and Obermayer, 2002) and also by the application in more
technical fields such as fingerprint classification (Zhang and Yan, 2004) and the detection of
roads in areal photography (Steger, 1998).
For the automated segmentation of loops, a first step was made by Strous (2002, unpublished) who proposed a procedure to detect pixels in the vicinity of loops. This approach was
further extended by Lee, Newman, and Gary (2006) by means of a connection scheme that
makes use of a local solar-surface magnetic-field estimate to obtain a preferable connection
orientation. The procedure then leads to spline curves as approximations for the loop shapes
in the image. The method gave quite promising results for artificial and also for observed
trace EUV images.
The program presented here can be considered an extension of the work by Lee, Newman, and Gary (2006). The improvements we propose are to replace Strouss ridge-point
detection scheme by a modified multiscale approach of Lindeberg (1998), which automatically adjusts to varying loop thicknesses and also returns an estimate of the reliability of the
ridge point location and orientation. When connecting the ridge points, we would prefer not
to use any magnetic-field information as this prejudices a later comparison of the extracted
loops with field lines computed from an extrapolation of the surface magnetic field. As
we consider this comparison a validity test for the underlying-field extrapolation, it would
be desirable to derive the loop shapes independently. Our connectivity method is therefore
based only on geometrical principles and combines the orientation of the loop at the ridge
point with the cocircularity constraint proposed by Parent and Zucker (1989).
The procedure is performed in three steps, each of which could be considered a module
of its own and performs a very specific task. In Section 2 we explain these individual steps
in some detail. In Section 3 we apply the scheme to one of the first images observed by
the SECCHI instruments onboard the recently launched STEREO spacecraft (Howard et al.,
2007) to demonstrate the capability of our tool. Our procedure offers alternative subschemes
and adaptive parameters to be adjusted to the contrast and noise level of the image under
consideration. We discuss how the result depends on the choice of some of these parameters.
In Section 4 we discuss potential applications of our tool.
2. Method
Our approach consists of three modular steps, each of which is described in one of the
following subsections. The first is to find points that presumably are located on the loop
axis. At these positions, we also estimate the orientation of the loop for these estimates.
171
Each item with this set of information is called a ridgel. The next step is to establish probable
neighborhood relations between them, which yields chains of ridgels. Finally, each chain is
fitted by a smoothing spline, which approximates the loop that gave rise to the ridgels.
2.1. Ridgel Location and Orientation
In terms of image intensities, loop structures are elongated ridge-like intensity maxima. To
discriminate the maxima, we need information about the spatial derivatives of the image intensity. Commonly, these derivatives are strongly affected by image noise. In fact, numerical
differentiation of data is an ill-posed problem and calls for proper regularization.
We denote by i I2 the integer coordinate values of the pixel centers in the image and
by x R2 the 2D continuous image coordinates with x = i at the pixel centers. We further
assume that the observed image intensity I (i) varies sufficiently smoothly so that a Taylor
expansion at the cell centers is a good approximation to the true intensity variation I (x) in
the neighborhood of i, that is,
I (x) I(x) = c + gT (x i) + (x i)T H(x i).
(1)
Pixels close to a ridge in the image intensity can then be detected on the basis of the local
derivatives g and H (the factor 1/2 gets absorbed in H). We achieve this by diagonalizing
H; that is, we determine the unitary matrix U with
UT HU = diag(h , h ),
where U = (u , u ),
(2)
where we assume that the eigenvector columns u and u of U associated with the eigenvalues h and h , respectively, are ordered so that h h .
We have implemented two ways to estimate the Taylor coefficients. The first is a local fit
of Equation (1) to the image within a (2m + 1) (2m + 1) pixel box centered around each
pixel i:

2
w(i j) I(j) I (j) .
(3)
(c, g, H)(i) = argmin
ji[m,m][m,m]
We use different weight functions w with their support limited to the box size such as triangle, cosine, or cosine2 tapers.
The second method commonly used is to calculate the Taylor coefficients (1) not from
the original but from a filtered image:
I(x) =
wd (x j)I (j).
(4)
As window-function wd we use a normalized Gaussian of width d. The Taylor coefficients

can now be explicitly derived by differentiation of I, which however acts on the window
function wd instead of on the image data. We therefore effectively use a filter kernel for
each Taylor coefficient that relates to the respective derivatives of the window function wd .
One advantage of the latter method over the local fit described here is that the window
width d can be chosen from R+ while the window size for the fit procedure must be an
odd integer 2m + 1. Both of these methods regularize the Taylor coefficient estimate by
the finite size of their window function. In fact, the window size could be considered as a
regularization parameter.
172
B. Inhester et al.
A common problem of regularized inversions is the proper choice of the regularization

parameter. Lindeberg (1998) has devised a scheme for how this parameter can be optimally
chosen. Our third method is a slightly modified implementation of his automated scaleselection procedure. The idea is to apply method two here for each pixel repeatedly with
increasing scales d and thereby obtain an approximation of the ridges second derivative
eigenvalues h and h , each as a function of the scale d.
Since the h and h are the principal second-order derivatives of the image after being filtered with wd , they depend on the width d of the filter window roughly in the following way.
As long as d is much smaller than the intrinsic width of the ridge, d = |(uT )2 log I |1/2 ,
the value in h will be a (noisy) estimate of the true principal second derivative (uT )2 I
of the image, independent of d. Hence, h Imax /d2 for d 2 d2 . To reduce the noise
and enhance the significance of the estimate, we would however like to choose d as large as
possible. For d 2
d2 , the result obtained for h will reflect the shape of the window rather
than that of the width of the ridge, h Imax /d 2 = Imax d /d 3 for d 2
d2 . Roughly,
details near d d depend on the exact shape of the ridge, and we have
h (d)
d
.
(d2 + d 2 )3/2
For each pixel we consider in addition a quality function

q(d) = d |h | |h | , (0, 3),
(5)
(6)
which will vary as d for small d and decrease asymptotically to zero for d
d as d 3 .
In between, q(d) will reach a maximum approximately where the window width d matches
the local width d of the ridge (which is smaller than the scale along the ridge). The choice
of the right width d has now been replaced by a choice for the exponent . The result,
however, is much less sensitive to variations in than to variations in d. Smaller values
of shift the maximum of q only slightly to smaller values of d and hence tend to favor
more narrow loop structures. Although is a constant for the whole image in this automated
scale selection scheme, the window width d is chosen individually for every pixel from the
respective maximum of the quality-factor q.
In Figure 1 we show as an example a = 171 image of active region NOAA 10930
observed by STEREO/SECCHI on 12 December 2006 at 20:43 UT and the corresponding
image of q obtained with = 0.75 and window sizes d in the range of 0.6 to 4 pixels.
Clearly, the q factor has a maximum in the vicinity of the loops. The distribution of the scales
d for which the maximum q was found for each pixel is shown in Figure 2. About 1/3 of the
pixels had optimal widths < one pixel, many of which originate from local elongated noise
and moss features of the image. The EUV moss is an amorphous emission that originates in
the upper transition region (Berger et al., 1999) and is not associated with loops. For proper
loop structures, the optimum width found was about 1.5 pixels with, however, a widely
spread distribution.
In the case that i is located exactly on a ridge, u is the direction across and u the
direction along the ridge, and h and h are the associated second derivatives of the image
intensity in the respective direction. A positive ridge is identified from the Taylor coefficients
by means of the following conditions (Lindeberg, 1998):
uT I = uT g = 0
T 2
u I = h < 0
(a vanishing gradient across the ridge),
(7)
(a negative second-order derivative across the ridge),
(8)
173
Figure 1 Original image of active region NOAA 10930 observed by STEREO/SECCHI (left) and the corresponding image of the quality factor q (6) obtained from the third ridgel determination method by automated
scale selection (right). This image was taken at = 171 on 12 December 2006 at 20:43 UT and was not
processed by the SECCHI_prep routine.
Figure 2 Distribution of widths
d of the window function wd for
which Equation (6) was found to
reach a maximum when applied
to the data in Figure 1. The
maximum was independently
determined for each image pixel
for which h < |h |.
T 2 T 2
u I > u I
or
|h | > |h |
(a second-order derivative magnitude

across the ridge larger than along).
(9)
The latter two inequalities are assumed to also hold in the near neighborhood of the ridge
and are used to indicate whether the pixel center is close to a ridge.
In the vicinity of the ridge, along a line x = i + u t , the image intensity (1) then varies
as
I (t) c + uT gt + uT u h t 2 .
(10)
According to the first ridge criterion (7), the precise ridge position is where I (t) has its
maximum. Hence the distance to the ridge is
tmax =
uT g
2h uT u
(11)
174
B. Inhester et al.
and a tangent section to the actual ridge curve closest to i is

r(s) = i
u uT g
+ su
2h uT u
for s R.
(12)
Note that uT u = 1 for a unitary U.

We have implemented two methods for the interpolation of the ridge position from the
Taylor coefficients calculated at the pixel centers. One is the interpolation of the ridge center
with the help of Equation (12). The second method interpolates the zeros of uT g in between
neighboring pixel centers i and j if its sign changes. Hence the alternative realization of
condition (7) is

if |c| = uT (i)u (j) > cmin

and sign(c) uT g (j) uT g (i) < 0 then
r = i + t (j i),
(13)
where
t=
(gT u )(i)
.
(gT u )(i) sign(c)(gT u )(j)
The first condition ensures that u (i) and u (j) are sufficiently parallel or antiparallel. Note
that the orientation of u of neighboring pixels may be parallel or antiparallel because an
eigenvector u has no uniquely defined sign.
In general, the interpolation according to Equation (12) yields fewer ridge points along
a loop but they have a fairly constant relative distance. With the second method (13), the
ridge points can only be found at the intersections of the ridge with the grid lines connecting
the pixel centers. For ridges directed obliquely to the grid, the distances between neighboring ridge points produced may vary by some amount. Another disadvantage of the second
method is that it cannot detect faint ridges that are just about one pixel wide. It needs at least
two detected neighboring pixels in the direction across the ridge to properly interpolate the
precise ridge position. The advantage of the second method is that it does not make use of
the second-order derivative h , which unavoidably is more noisy than the first-order derivative g. In Figure 3 we compare the ridgels obtained with the two interpolation methods for
the same image.
The final implementation of identifying ridge points in the image comprises two steps:
first the Taylor coefficients (1) are determined for every pixel and saved for those pixels that have an intensity above a threshold value Imin , for which the ridge shape factor
(h2 h2 )/(h2 + h2 ) exceeds a threshold smin in accordance with condition (9) and that
also satisfy condition (8). The second step is then to interpolate the precise subpixel ridge
point position from the derivatives at these pixel centers by either of the two methods. This
interpolation complies with the third ridge criterion (7). The ridgel orientations u are also
interpolated from the cell centers to the ridgel position. The information retained for every
ridge point n in the end consists of its location rn and the ridge normal orientation u,n
defined modulo .
2.2. Ridgel Connection to Chains
The connection algorithm we apply to the ridgels to form chains of ridgels is based on the
cocircularity condition of Parent and Zucker (1989).
175
Figure 3 Comparison of the resulting ridgels from interpolation method (12, left) and (13, right). The images
show an enlarged portion of the original data in Figure 1. The short sticks denote the local orientation of u
on the loop trace.
Figure 4 Illustration of angles

and distances of a connection
element between two ridgels
according to the cocircularity
condition of Parent and Zucker
(1989). The ridgel positions are
indicated by small dots; the ridge
normal orientation is shown by
the line centered at the ridgel
position. The axis units are image
pixels.
For two ridgels at rn and rn+1 a virtual center of curvature can be defined that forms
an isosceles triangle with the ridgels as shown in Figure 4. One edge is formed by the
connection between the two ridgels of mutual distance hn,n+1 . The two other triangle edges
in this construction connect one of the two ridgels with the center of curvature, which is
176
B. Inhester et al.
chosen so that these two symmetric edges of the isosceles triangle make angles n,n+1
and n+1,n as small as possible with the respective ridgel orientation u,n and u,n+1 ,
respectively. It can be shown that
2

2
+ n+1,n
min n,n+1
(14)
requires equal magnitudes for the angles n,n+1 and n+1,n . The distance rn,n+1 = rn+1,n
is the local radius of curvature and can be calculated from
rn,n+1 = rn+1,n =
1
h
2 n,n+1
cos(12 (n,n+1 + n+1,n ))
(15)
where n,n+1 is the angle between rn+1 rn and u,n , with the sign being chosen so that
|n,n+1 | < /2.
With each connection between a pair of ridgels we associate a binding energy that depends on the three parameters just derived in the form

en,n+1 =
n,n+1
max
2
+
rmin
rn,n+1
2
hn,n+1
+
hmax
2
3.
(16)
2
2
Note that n,n+1
= n+1,n
according to the cocircularity construction and hence en,n+1 is
symmetric in its indices. The three terms measure three different types of distortions and can
be looked upon as the energy of an elastic line element. The first term measures the deviation
of the ridgel orientation from strict cocircularity, the second the bending of the line element,
and the third term its stretching. The constants max , rmin , and hmax give us control on the
relative weight of the three terms. If we only accept connections with a negative value for the
energy (16), then rmin is the smallest acceptable radius of curvature and hmax is the largest
acceptable distance.
In practical applications, the energy (16) is problematic since it puts nearby ridgel pairs
with small distances hn,n+1 at a severe disadvantage because small changes of their u
easily reduces the radius of curvature rn,n+1 below acceptable values. We therefore allow
for measurement errors in r and u and the final energy considered is the minimum of
Equation (16) within these given error bounds.
The final goal is to establish a whole set of connections between as many ridgels as
possible so that the individual connections add up to chains. Note that each ridgel has two
sides defined by the two half-spaces that are separated by the ridgel orientation u . We
only allow at most one connection per ridgel in each of these sides. This restriction avoids
junctions in the chains that we are going to generate. The sum of the binding energies (16)
of all accepted connections ideally should attain a global minimum in the sense that any
alternative set of connections that complies with this restriction should yield a larger energy
sum.
We use the following approach to find a state that comes close to this global minimum.
The energy en,n+1 is calculated for each ridgel pair less than hmax apart, and those connections that have a negative binding energy are stored. These latter are the only connections
we expect to contribute to the energy minimum. Next we order the stored connections according to their energy and connect the ridgels to chains starting from the lowest energy
connection. Connections to one side of a ridgel that has already been occupied by a lower
energy connection before are simply discarded.
177
2.3. Curve Fits to the Ridgel Chains

In this final section we calculate a smooth fit to the chains of ridgels obtained. The fit curve
should level out small errors in the position and orientation of individual ridgels. We found
from experiments that higher order spline functions are far too flexible for the curves we aim
at. We expect that magnetic-field lines in the corona do not rapidly vary their curvature along
their length and we assume this also holds for their projections on EUV images. We found
that parametric polynomials of third or fifth degree are sufficient for our purposes. Hence for
each chain of ridgels we seek polynomial coefficients qn that generate a two-dimensional
curve
p(t) =
5
qn t n
for t [1, 1]
(17)
n=0
that best approximates the chain of ridgels. What we mean by best approximation will be
defined more precisely in the following. The relevant parameters of this approximation are
sketched in Figure 5.
The polynomial coefficients q of a fit (17) are determined by minimizing

T
dTi di + p p (ti ),
(18)
ichain
where
di = ri p(ti )
with respect to qn for a given . Initially, we distribute the curve parameters ti in the interval [1, 1] such that the differences ti tj of neighboring ridgels are proportional to the
geometric distances |ri rj |. The p are the second-order derivatives of (17). Hence, the
Figure 5 Sketch of the curve-fit
parameters. The ridgels are
represented by their location and
the two-pixel-long bar of the
ridgel orientation. For each
ridgel i, the proximity to the
smooth-fit curve is expressed by
its distance di and the angle i
between the distance direction of
di to the curve and the ridgel
orientation. Another measure of
the quality of the curve is the
inverse radius of curvature r.
178
B. Inhester et al.
second term increases with increasing curvature of the fit whereas a more strongly curved fit
is required to reduce the distances di between ri and the first-order closest curve point p(ti ).
The minimum coefficients qn () can be found analytically in a straightforward way.
Whenever a new set of qn () has been calculated, the curve nodes ti are readjusted by
2

(19)
ti = argmint ri p(t) ,
so that p(ti ) is always the point along the curve closest to the ridgel.
For different , this minimum yields fit curves with different levels of curvature. The
local inverse radius of curvature can at any point along the curve be calculated from Equation (17) by
|p (t) p (t)|
1
=
.
r(t)
|p (t)|3
(20)
1
2
di2
1
i
2
+
+ rmin
dt
Echain () =
2
2
2
d
r(t)
1
ichain max
ichain max
(21)
The final is then chosen so that

is a minimum where n is the angle between the local normal direction of dn of the fit
curve and the ridgel orientation u,n , with the sign again chosen to yield the smallest
possible |i |. The meaning of the terms is obvious, and clearly the first two terms in
general require a large curvature, which is limited by the minimization of the last term.
Expression (21) depends nonlinearly on the parameter , which we use to control the
overall curvature. The minimum for Equation (21) is found by iterating starting from a
large numerical value (i.e., a straight-line fit). The parameters rmin , max , and dmax can be
used to obtain fits with a different balance between the mean square spatial and angular
deviation of the fit from the observed chain of ridgels and the curvature of the fit. Unless
these parameters are chosen reasonably (e.g. rmin not too small), we have always found a
minimum for Equation (21) after a few iteration steps.
In the left part of Figure 6 we show the final fits obtained. For this result, the ridgels
were found by automated scaling and interpolated by method (12), the parameters in Equa-
Figure 6 Fit curves obtained for those chains that involve ten or more ridgels (left) and curves remaining
after cleaning of those curves resulting from moss (right).
179
tions (16) and (21) were hmax = 3.0 pixels, rmin = 15.0 pixels, and amax = 10.0 degrees. The
fits are represented by fifth-degree parametric polynomials.
Obviously, the image processing cannot easily distinguish between structures resulting
from moss and bright surface features and coronal loops. Even the observer is sometimes
misled and there are no rigorous criteria for this distinction. Roughly, coronal loops produce
longer and smoother fit curves, but there is no strict threshold because it may appear that
the fit curve is split along a loop where the loop signal becomes faint. As a rule of thumb, a
restriction to smaller curvature by choosing a higher parameter rmax and discarding shorter
fit curves tends to favor coronal loops. Eventually, however, loops are suppressed, too. We
have therefore appended a user interactive tool as the last step of our processing to allow
us to eliminate unwanted curves and merge or split curves when smooth fits result with an
energy (21) of the output fits not much higher than the energy of the input. The left part of
Figure 6 shows the result of such a cleaning step.
3. Application
In this section we present an application of our segmentation tool to another EUV image
of active region NOAA 10930 taken by the SECCHI instrument onboard STEREO A. This
EUV image was observed at = 195 on 12 December 2006 at 23:43:11 UT. At that
time the STEREO spacecraft were still close together so stereoscopy could not be applied.
We therefore selected an image that was taken close to the MDI magnetogram observed
at 23:43:30 UT on the same day. It is therefore possible to calculate magnetic-field lines
from an extrapolation model and project them onto the STEREO view direction to compare
them with the loop fits obtained with our tool. In Figure 7 the MDI contour lines of the
line-of-sight field intensity were superposed on the EUV image.
Figure 7 MDI contours overlaid on the STEREO/SECCHI EUV image for NOAA 10930. The EUV and
MDI data were recorded on 12 December 2006 at 23:43:11 UT and 23:43:30 UT, respectively. The color code
on the right indicates the field strength at the contour lines in gauss.
180
B. Inhester et al.
Figure 8 The left diagram shows as red lines the loops identified by the segmentation tool for the EUV
image in Figure 7. To show the loops more clearly, the image in the background was contrast enhanced by an
unsharp mask filter. The right diagram displays in yellow field lines calculated from the MDI data. The field
lines were selected so that they are located closest to the extracted loops in the left part of the image. See the
text for more details on how the field lines were determined.
The loop fits here were obtained by applying the automated scaling with d up to two
pixels (i.e. window sizes up to 2d + 1 = 5 pixels) to identify the ridgels. Pixels with maximum quality q below 0.4 were discarded and we applied method (12) to interpolate the
local ridge maxima. hmax = 5.0 pixels, rmin = 15.0 pixels, and amax = 10.0 . The fits are
fifth-degree parametric polynomials. In Figure 8, we show some of the fits obtained that are
most likely associated with a coronal loop. They are superposed onto the EUV image as
red lines. Those loops that were found close to computed magnetic-field lines are displayed
again in the left part of Figure 9 with loop numbers so that they can be identified.
The magnetic-field lines were computed from the MDI data by an extrapolation based
on a linear force-free field model (see Seehafer, 1978, and Alissandrakis, 1981, for details).
This model is a simplification of the general nonlinear, force-free magnetic-field model:
B = B,
where B = 0,
(22)
and may vary on different field lines. An extrapolation of magnetic surface observations
based on this model requires boundary data from a vector magnetograph. The linear forcefree field model treats as a global constant. The advantage of the linear force-free field
model is that it requires only a line-of-sight magnetogram, such as MDI data, as input.
A test of the validity of the linear force-free assumption is to determine different values
of from a comparison of field lines with individual observed loop structures (e.g., Carcedo
et al., 2003). The range of obtained then indicates how close the magnetic field can be
described by the linear model. Since the linear force-free field has the minimum energy
for given normal magnetic boundary field and magnetic helicity, a linear force-free field is
supposed to be much more stable than the more general nonlinear field configuration (Taylor,
1974).
We calculated about 5000 field lines with the linear force-free model with the value
varied in the range from 0.0427 to 0.0356 Mm1 . These field lines were then projected
onto the EUV image for a comparison with the detected loops.
181
Figure 9 The left panel shows the loops identified by the segmentation tool that corresponds to the closed
magnetic filed lines. In the right panel are the loops (solid lines) with their best-fitting field lines (dotted
lines). The x and y axes are in units of EUV pixels.
Table 1 Identified loops, the
averaged distances in units of
pixels between the loop and the
best-fitting field line, and
values of the best-fitting field
lines.
Loop
Cl (b) (pixel)
(103 Mm1 )
3.1957
10.680
3.7523
35.600
2.3432
11
10.7692
13
0.4864
2.1360
14
1.3636
9.2560
15
4.2386
17.088
17
4.8912
16.376
18
2.4256
14.240
19
2.5388
32.752
9.2560
35.600
For each coronal loop li , we calculate the average distance of the loop to every projected
field line bj discarding those field lines that do not fully cover the observed loop li . This
distance is denoted by Cli (bj ). For details of this distance calculation see Feng et al. (2007).
In the end we could find a closest field line for every coronal loop by minimizing Cli (bj ).
The detected loops and their closest field lines are plotted in the right diagram of Figure 9.
An overplot of the closest field lines onto the EUV image is shown in Figure 8.
In Table 1 we list the distance measure C along with the loop number and the linear forcefree parameter for the closest field line found. We find that our values are not uniform
over this active region; that is, the linear force-free model is not adequate to describe the
magnetic properties of this active region. This is also seen by the characteristic deviation
at their upper right end in Figure 8 between the eastward-inclined loops (solid) and their
closest, projected field lines (dotted). With no value of the shape of these loops could be
satisfactorily fitted. Further evidence for strong and inhomogeneous currents in the active
region loops is provided by the fact that only 2.5 hours later, at 02:14 UT on 13 December,
182
B. Inhester et al.
a flare occurred in this active region and involved the magnetic structures associated with
loops 2, 4, and 17.
4. Discussion
EUV images display a wealth of structures and there is an important need to reduce this
information for specified analyses. For the study of the coronal magnetic field, the extraction
of loops from EUV images is a particularly important task. Our tool intends to improve
earlier work in this direction. Whether we have achieved this goal can only be decided from
a rigorous comparison, which is underway elsewhere (Aschwanden et al., 2007). At least
from a methodological point of view, we expect that our tool should yield improved results
compared to those of Strous (2002, unpublished) and Lee, Newman, and Gary (2006).
From the EUV image alone it is often difficult to decide which of the features are associated with coronal loops and which are due to moss or other bright surface structures.
A final comparison of the loops with the extrapolated magnetic field and its field line shapes
is therefore very helpful for this distinction. Yet we have avoided involving the magneticfield information in the segmentation procedure that extracts the loops from the EUV image
because this might bias the loop shapes obtained.
For the case we have investigated, we find a notable variation of the optimal values
and also characteristic deviation of the loop shapes from the calculated field lines. These
differences are evidence of the fact that the true coronal magnetic field near this active
region is not close to a linear force-free state. This is in agreement with earlier findings.
Wiegelmann et al. (2005), for example, have shown for another active region that a nonlinear
force-free model describes the coronal magnetic field more accurately than linear models.
The computation of nonlinear models is, however, more involved because of the nonlinearity
of the mathematical equations (e.g., Wiegelmann, 2004; Inhester and Wiegelmann, 2006).
Furthermore, these models require photospheric vector magnetograms as input, which were
not available for the active region investigated.
Coronal loop systems are often very complex. To access them in three dimensions, the
new STEREO/SECCHI telescopes now provide EUV images that can be analyzed with
stereoscopic tools. We plan to apply our loop-extraction program to EUV images from different viewpoints and undertake a stereoscopic reconstruction of the true 3D structure of
coronal loops along the lines described by Inhester (2006) and Feng et al. (2007). Knowledge of the 3D geometry of a loop allows us to estimate more precisely its local EUV emissivity. From this quantity we hope to be able to derive more reliably the plasma parameters
along the length of the loop.
Other applications can be envisaged. One interesting application of our tool will be the
investigation of loop oscillations. Here, the segmentation tool will be applied to times series
of EUV images. We are confident that oscillation modes can be resolved and, in the case of
a STEREO/SECCHI pairwise image sequence, the polarization of the loop oscillation can
also be discerned.
Acknowledgements B.I. thanks the International Space Institute, Bern, for its hospitality and the head
of its STEREO working group, Thierry Dudoc de Wit, and also Jean-Francois Hochedez for stimulating
discussions. L.F. was supported by the IMPRESS graduate school run jointly by the Max Planck Society
and the Universities of Gttingen and Braunschweig. The work was further supported by DLR Grant No.
50OC0501. The authors thank the SOHO/MDI and the STEREO/SECCHI consortia for their data. SOHO
and STEREO are joint projects of ESA and NASA. The STEREO/SECCHI data used here were produced
by an international consortium of the Naval Research Laboratory (USA), Lockheed Martin Solar and Astrophysics Lab (USA), NASA Goddard Space Flight Center (USA), Rutherford Appleton Laboratory (UK),
183
University of Birmingham (UK), Max-Planck-Institut for Solar System Research (Germany), Centre Spatiale
de Lige (Belgium), Institut dOptique Thorique et Applique (France), and Institut dAstrophysique Spatiale (France). The USA institutions were funded by NASA, the UK institutions by the Particle Physics and
Astronomy Research Council (PPARC), the German institutions by Deutsches Zentrum fr Luft- und Raumfahrt e.V. (DLR), the Belgian institutions by the Belgian Science Policy Office, and the French institutions
by the Centre National dEtudes Spatiales (CNES) and the Centre National de la Recherche Scientifique
(CNRS). The NRL effort was also supported by the USAF Space Test Program and the Office of Naval
Research.
References
Alissandrakis, C.E.: 1981, Astron. Astrophys. 100, 197.
Aschwanden, M., Lee, J.K., Gary, G.A., Smith, M., Inhester, B.: 2007, Solar Phys., accepted.
Berger, T., De Pontieu, B., Fletcher, L., Schrijver, C., Tarbell, T., Title, A.: 1999, Solar Phys. 190, 409.
Carcedo, L., Brown, D.S., Hood, A.W., Neukirch, T., Wiegelmann, T.: 2003, Solar Phys. 218, 29.
Dimas, A., Scholz, M., Obermayer, K.: 2002, IEEE Trans. Image Process. 11(7), 790.
Feng, L., Wiegelmann, T., Inhester, B., Solanki, S., Gan, W.Q., Ruan, P.: 2007, Solar Phys. 241, 235.
Howard, R., Moses, J., Vourlidas, A., Newmark, J., Socker, D., Plunckett, S., Korendyke, C., Cook, J., Hurley, A., Davila, J., Thompson, W., St. Cyr, O., Mentzell, E., Mehalick, K., Lemen, J., Wuelser, J., Duncan, D., Tarbell, T., Harrison, R., Waltham, N., Lang, J., Davis, C., Eyles, C., Halain, J., Defise, J.,
Mazy, E., Rochus, P., Mercier, R., Ravet, M., Delmotte, F., Auchre, F., Delaboudinire, J., Bothmer, V.,
Deutsch, W., Wang, D., Rich, N., Cooper, S., Stephens, V., Maahs, G., Baugh, R., McMullin, D.: 2007,
Space Sci. Rev., in press (http://secchi.lmsal.com/EUVI/DOCUMENTS/howard.pdf).
Inhester, B.: 2006, Int. Space Sci. Inst., submitted (astro-ph/0612649).
Inhester, B., Wiegelmann, T.: 2006, Solar Phys. 235, 201.
Jang, J.H., Hong, K.S.: 2002, Pattern Recognit. 35, 807.
Koller, T., Gerig, B., Szkely, G., Dettwiler, D.: 1995, In: Proceedings Fifth Int. Conf. on Computer Vision
(ICCV95), IEEE Computer Society Press, Washington, 864.
Lee, J.K., Newman, T.S., Gary, G.A.: 2006, Pattern Recognit. 39, 246.
Lindeberg, T.: 1998, Int. J. Comput. Vis. 30(2), 117.
Medioni, G., Tang, C.K., Lee, M.S.: 2000, http://citeseer.ist.psu.edu/medioni00tensor.html.
Nguyen, H., Worring, M., van den Boomgaard, R.: 2000, http://citeseer.ist.psu.edu/article/
nguyen00watersnakes.html.
Parent, P., Zucker, S.: 1989, IEEE Trans. Pattern Anal. Mach. Intell. 11, 823.
Seehafer, N.: 1978, Solar Phys. 58, 215.
Starck, J.L., Donoho, D.L., Cands, E.J.: 2003, Astron. Astrophys. 398, 785.
Steger, C.: 1998, IEEE Trans. Pattern Anal. Mach. Intell. 20(2), 113.
Taylor, J.B.: 1974, Phys. Rev. Lett. 33, 1139.
Wiegelmann, T.: 2004, Solar Phys. 219, 87.
Wiegelmann, T., Lagg, A., Solanki, S.K., Inhester, B., Woch, J.: 2005, Astron. Astrophys. 433, 701.
Zhang, Q., Yan, H.: 2004, Pattern Recognit. 37, 2233.
The Pixelised Wavelet Filtering Method to Study Waves

and Oscillations in Time Sequences of Solar Atmospheric
Images
R.A. Sych V.M. Nakariakov

Abstract Pixelised wavelet filtering (PWF) for the determination of the spatial, temporal,
and phase structure of oscillation sources in temporal sequences of 2D images, based upon
the continuous wavelet transform, has been designed and tested. The PWF method allows
us to obtain information about the presence of propagating and nonpropagating waves in the
data and localise them precisely in time and in space. The method is tested on the data sets
obtained in microwaves with the Nobeyama Radioheliograph and in the EUV with TRACE.
The method reveals fine spatial structuring of the sources of 3-, 5-, and 15-minute periodicities in the microwave and EUV emission generated in sunspot atmospheres. In addition,
the PWF method provides us with unique information about the temporal variability of the
power, amplitude, and phase narrowband maps of the observed oscillations and waves. The
applicability of the method to the analysis of coronal wave phenomena is discussed.
1. Introduction
Wave and oscillatory phenomena are believed to play a crucial role in a number of physical processes operating in the atmosphere of the Sun and Sun-like stars (e.g., Nakariakov
and Verwichte, 2005). However, the analysis of solar data to determine parameters of atmospheric waves and oscillations is a nontrivial task, mainly because these phenomena lie
near the very threshold instrumental detectability, necessitating the analysis of spatial, temporal, and phase variability at the highest possible resolution. The commissioning of the
modern generation of high spatial- and temporal-resolution solar observational instruments
(e.g., the Nobeyama Radioheliograph (NoRH), the Transition Region and Coronal Explorer
(TRACE), and instruments onboard the Solar and Heliospheric Observatory (SOHO)) stimulated the development of novel data analysis methods, such as multiscale wavelet analysis
R.A. Sych ()
Institute of Solar-Terrestrial Physics, 126 Lermontov St., Irkutsk, Russia
e-mail: sych@iszf.irk.ru
V.M. Nakariakov
Physics Department, University of Warwick, Coventry CV4 7AL, UK
e-mail: V.Nakariakov@warwick.ac.uk
DOI: 10.1007/978-0-387-98154-3_14
185
186
R.A. Sych, V.M. Nakariakov
(MWA) (Meyer and Roques, 1993), the empirical mode decomposition method (EMD), the
complex empirical orthogonal function (CEOF) method (Huang, Shen, and Long, 1999),
singular spectrum analysis (SSA) (Golyandina, Nekrutkin, and Zhiglyavsky, 2001), and
multifractal analysis (Milovanov and Zelenyj, 1993). The interest of these methods is motivated by the transient and localised nature of the solar atmospheric oscillations.
One of the intrinsic properties of solar coronal and low-atmospheric wave and oscillatory phenomena is the pronounced temporal modulation of the wave parameters, e.g. of the
amplitude and period, and also their transient nature. A popular tool allowing researchers
to study wave phenomena of this kind is the wavelet-transform technique. This method
has been intensively used in solar atmospheric physics for the past decade. In particular,
Baudin, Bocchialini, and Koutchmy (1996) applied wavelets to the study of oscillations in
the chromospheric network; Ireland et al. (1999) studied intensity oscillations in active regions observed with SOHO/CDS. Recently, wavelet analysis was applied to the analysis of
longitudinal waves observed over sunspots with TRACE and other instruments (De Moortel,
Hood, and Ireland, 2002; King et al., 2003; Marsh et al., 2003). Also, Christopoulou et al.
(2003) considered temporal evolution of the velocity and intensity oscillations in sunspot
umbrae at the chromospheric level.
The study presented here proposes and tests a novel method of pixelised wavelet filtering
(PWF) for the study of spatially-distributed oscillation sources in the solar atmosphere. In
this work, we follow Torrence and Compo (1998) and use the continuous wavelet transform
with the Morlet mother function. The PWF method is based upon the application of the
wavelet transform to the construction of narrowband and broadband static and dynamical
spectral 2D maps of the analysed temporal sequences of 2D images (data cubes). One of
the main aims of this work is the development and testing of an analytical tool for the determination of the fine spatial structure of transient oscillation sources and interrelations of
the spectral peaks revealed by the analysis of the spatially-integrated signal. To the temporal
signal of each pixel of the spatial field of view, we apply wavelet filtration to decompose the
signal in the temporal domain (temporal frequencies) and determine the spatial distribution
of the power of different spectral components in a way to allow study of the temporal evolution of the spatialfrequency information. This analysis allows us to obtain narrowband
maps (associated with a certain chosen periodicity), to determine the wave type (propagating
or nonpropagating in the field of view), and to study their temporal evolution.
The developed method is similar to two other previously-used approaches to the analysis of wave phenomena in solar atmospheric imaging data sets, the wavelet-based oscillation detection technique developed by De Moortel and McAteer (2004) and the EMD- and
CEOF-based methods applied by Terradas, Oliver, and Ballester (2004). The PWF method
discussed in this paper differs from these methods in the following aspects:
i) In PWF, one can prescribe precisely the frequency boundaries of the analysed spectral
band. It is possible to analyse either a chosen discrete harmonic or a prescribed band of
continuous frequencies.
ii) A possible outcome of PWF is a 2D map showing the spatial distribution of the amplitude, power, and phase of selected narrowband signals (narrowband maps).
iii) It is possible to study the temporal evolution of the narrowband maps, by considering the
variation of the amplitude, power, and phase of selected narrowband signals (dynamical
narrowband maps). In particular, this allows one to estimate the projected phase speed
of the waves.
iv) The method, in contrast with the EMD and COEF methods, provides us with the flexibility in the choice of the mother wavelet function as well as in the frequency time
resolution of the wavelet filtering.
187
v) The PWF technique can be readily developed to the calculation of the cross-correlation
spectra, allowing for the study of the interaction of wave and oscillatory processes contained in different data cubes. In particular, these data cubes can be obtained with different instruments and/or in different wavelengths.
Some of these features can be obtained with the methods of De Moortel and McAteer
(2004) and Terradas, Oliver, and Ballester (2004), but not simultaneously with the same
analytical tool. The unique feature of the PWF method is the construction of the dynamical
narrowband maps (item iii). In particular, the method of De Moortel and McAteer, designed
for the automated detection of temporal periodicities and of their spatial distribution, allows
us to identify the presence of a periodicity in the signal associated with a given spatial
spectrum, but it does not give us its temporal evolution. For example, the PWF method
allows us to determine the number of periods in the detected oscillation and the evolution of
the signal in the spectral band of interest.
We apply the PWF method to the analysis of sunspot and coronal loop oscillations, using
the well-known data cubes obtained with NoRH (17 GHz, already discussed by Gelfreikh
et al., 1999) and with TRACE (171 , discussed by De Moortel, Ireland, and Walsh, 2000,
and by Terradas, Oliver, and Ballester, 2004). The aim was to reproduce the previous findings and to demonstrate the applicability of the PWF method to the analysis of coronal
imaging data cubes.
2. Test Signals
The input signal is a data cube that consists of a sequence of 2D images sampled with a specified temporal cadence. In particular, it can be an EUV data cube obtained with SOHO/EIT
or TRACE EUV imagers, Hinode/XRT or Yohkoh/SXT imagers, or with NoRH. The data
cubes obtained with different instruments have their own intrinsic features; thus it is worth
testing the general applicability of the PWF method with an artificial synthetic data cube.
In the test, analogous to the test signal applied by Terradas, Oliver, and Ballester (2004)
for the CEOF method, the simulated data cube consists of 300 21 21 pixel frames. The
temporal cadence is one second. The synthetic data cube contains a superposition of a standing and upwards and downwards propagating monochromatic waves of different periods,
wavelengths, and amplitudes:
X (x, y, t) = XSW (x, y, t) + XUTW (x, y, t) + XDTW + N (x, y, t),
(1)
where x and y are the spatial coordinates, t is time,

2x
2t
y 10 2
,
sin
exp
15
100
4

t
x
y 17 2
XUTW (x, y, t) = 4 sin 2
,
exp
12 60
2

t
x
y 3 2
XDTW = 4 sin 2
+
,
exp
4 20
2
XSW (x, y, t) = 2 sin
(2)
(3)
(4)
and N (x, y, t) is a high-frequency noise. The standing wave has a period of 100 arbitrary
units, an amplitude of 2, and a wavelength of 15. In the following we assume that the temporal units in the test signal are seconds; hence the period is 100 seconds. The propagating
188
Figure 1 Left: A 2D snapshot of the test signal, which contains a standing and two propagating monochromatic waves and noise. Right: The variance map showing the spatial distribution of the oscillation power
for the whole time sample. It is possible to see two vertical strips of Gaussian shape, corresponding to the
propagating waves, and three maxima, corresponding to the standing wave.
waves have amplitude 4, periods of 60 and 20 seconds, and wavelengths of 12 and 4 units,
respectively, and are localised at different parts of the image. A snapshot of the signal is
shown in the left panel of Figure 1.
3. The Scheme of the Pixelised Wavelet Filtering Method

The PWF method is based upon the wavelet transform described by detail in Weng and Lau
(1994) and Torrence and Compo (1998). In this work, as the mother wavelet function, we
use the complex Morlet wavelet, which is well localised in both the spectral domain characterised by the spectral coordinate (the frequency) and the physical domain characterised
by the coordinate t (the time):

(5)
(t) = exp(it) exp t 2 /2 ,
which is a monochromatic wave modulated by a Gaussian of unit width.
As a first step, we identify the region of interest (ROI) for the subsequent analysis. In particular, we may obtain information about the spatial distribution of the temporal variations
including all spectral frequencies by constructing a broadband variance map based upon the
temporal variance of each pixel. For this purpose, we calculate mean-square variations from
the mean value for each pixel and make a contour-plot map of the analysed field of view
(FOV). The map shows the spatial distribution of the power of all spectral components in
the FOV (a similar technique was developed by Grechnev, 2003). The variance maps allow
us to select the ROI for further analysis. The variance map of test signal (1) is shown in the
right panel of Figure 1. We point out that the variance map does not contain any information
about the localisation of the event of interest in time.
As the next step, we determine the significant frequencies of the studied oscillations and
the time intervals showing the oscillations with these frequencies. We calculate the wavelet
spectrum of the temporal signal integrated over the whole ROI. Figure 2 shows the integrated
189
Figure 2 Upper panel: Time signal integrated over the field of view. Bottom panel: Amplitude wavelet of
the signal given in the upper panel. The horizontal dashed lines show the built-in periods.
wavelet spectrum of the synthetic data cube described in Section 2. The built-in periodicities
are well seen in the spectrum.
Similar wavelet spectra are constructed for all individual pixels, obtaining a 4D data cube
given by two spatial axes, the time axis and the frequency axis. Then, for each frequency of
interest (built-in, identified in the integrated wavelet spectrum, or just assumed to be in the
signal), for each pixel, we make the inverse wavelet transform. Thus, we obtain temporal
variations of the narrowband signals in the vicinity of the frequency of interest at every
pixel, with all other frequencies filtered out. Repeating this for each frequency of interest
(or for all frequencies), we obtain a set of temporal variations for each pixel. The number of
curves is determined by the number of frequencies in the discrete spectrum, which in turn is
determined by the number of frames in the analysed data cube.
Then, we determine the power of each of the obtained narrowband curves as an integral
over time of the mean-square deviation (variance) from its mean. Repeating this procedure
for each pixel we obtain the spatial distribution of the power associated with this particular
frequency. The obtained data cube formed by the frequency and two spatial axes (broadband
power cube) can be used for the visualisation of the spatial distribution of the spectrum of
the oscillation sources. The left panel of Figure 3 shows the broadband power cube of the
analysed synthetic data set, highlighting the built-in wave motions in the analysed test signal.
190
Figure 3 Left: The spatial distribution of the signal power of different built-in waves (the broadband power
cube). The vertical axis (not shown) corresponds to the periods and shows the detected periodicity: 20-,
60-, and 100-second monochromatic signals and the noise. Right: Spatially integrated global wavelet power
spectrum of the analysed time sequences of 2D images.
The morphological features of the waves are clearly seen in Figure 3: the standing waves
have well-defined periodically spaced shapes corresponding to their extremes, whereas the
propagating waves are cylinders of elliptical cross section, resulting from averaging the wave
profiles. The high-frequency noise is uniformly distributed over the space.
Taking the spatial distribution of the power associated with a certain frequency (or integrated over a certain frequency band) we obtain a narrowband power map for the frequency
or band of interest. In Figure 3 (left panel), this corresponds to slicing the 3D frequency
spatial cube in the horizontal direction, perpendicular to the vertical axis representing the
frequency.
Using this spatial distribution and integrating the narrowband temporal signals over the
spatial domain for each frequency we obtain the global spatially integrated wavelet power
spectrum (an analogue of the Fourier power spectrum) of the analysed data cube, which is
shown in the right panel of Figure 3. There are three distinct peaks in the spectrum, corresponding to the built-in 20-, 60-, and 100-second periodicities, as well as high-frequency
noise.
Similarly, using the phase spectrum instead of the amplitude one, we get information
about the phase of the signal. The phase is obtained by calculating separately the real and
imaginary parts of the signal.
One of the disadvantages of the broadband and narrowband power maps is the lack of
temporal information, as the signal is integrated over time. Using the values of the significant periods found in the global wavelet spectrum (Figure 3, right panel) we can study the
temporal evolution of the spatial sources of these oscillations, making movies of the narrowband maps calculated at different instants of time (dynamical narrowband maps). The
duration of the analysed temporal interval is determined by the width of the mother wavelet
function. A snapshot of the dynamical map gives us the spatial distribution of the amplitude
or phase of the signal in the frequency band of interest (or at a prescribed harmonic) at a
particular instant of time (a narrowband map). Sequences of narrowband maps give us the
temporal evolution of the periodicity of interest.
Figure 4 shows amplitude and phase narrowband maps of the analysed data cube at three
different instants of time. The crosses on the image show the positions of the extrema in the
Figure 4 Sequences of amplitude narrowband maps of the synthetic data cube, showing the temporal variation of the amplitude and phase of the narrowband oscillation sources
of upward and downward propagating waves with periods of 60 and 20 second (upper and middle panels, respectively) and of standing waves with a period of 100 second
(bottom panel). The corresponding spatial distribution of the wave phases (the phase narrowband maps) are shown in the right panels. The crosses show the wave extrema. The
arrows show the direction of the wave propagation.

191
192
signal. It is seen that the positions of the extrema of the standing waves (Figure 4, bottom
panel) do not change in time, whereas their amplitudes do. Also, the phase of standing waves
is seen to change in a period (Figure 4, right panel). The extrema of the propagating waves
(with periods of 20 and 60 seconds) are seen to move in time (Figure 4, upper and middle
panels). A movie constructed with the use of all frames shows the upwards and downwards
propagating waves, clearly in agreement with the test signal.
Thus, the test demonstrates the applicability of the PWF method to the identification of
the nature of harmonic or quasi-harmonic oscillations and waves in imaging data cubes. In
particular, the method allows one to distinguish confidently between standing and propagating waves. We point out that in the analysis of natural data cubes, the projection effect
should be taken into account. The phase speed of the propagating waves is the speed projected on the FOV plane. The term standing is applied to the waves that do not propagate
in the FOV plane. However, the wave can be propagating in the direction of the line of sight.
4. Application to Sunspot Oscillations

We process the microwave observations of sunspots, obtained with NoRH at 17 GHz on
22 July 1994, 30 June 1993, and 16 August 1992, with a temporal cadence of 10 seconds
and a spatial resolution 10 in the circular-polarisation channel V = R L. In those data
sets, the shortest detectable period is 30 seconds, and the longest detectable period is about
1000 s. The data have already been analysed by Gelfreikh et al. (1999) and were found to
contain oscillations of the integrated radio flux with periods of 120 220 and 300 seconds.
Here, we apply the PWF method for determining the spatial distribution and evolution of
the dominating spectral harmonics in this data set.
The data sets were preprocessed with the use of the Solar Soft (SSW) IDL routines
norh_trans and norh_synth. The images were co-aligned with each other to get rid of highfrequency jitter and their calibration. For this purpose, we use align_cube_correl.
The broadband variance maps of the events, aimed at localising the regions with maximum variance of the emission in the FOV, demonstrate the presence of the signal over both
umbral and penumbral regions of analysed sunspots, providing us with the ROI for further
analysis.
The global spatially integrated wavelet power spectra of the signals coming from the
selected ROI (Figure 5) show the presence of five-minute oscillations. There are also 3and 15-minute peaks in the spectra. (We do not need to estimate the probability of false
Figure 5 The spatially integrated global wavelet power spectra of the analysed temporal data cube (NoRH
circular polarisation channel V = R L) for 22 July 1994, 30 June 1993, and 16 August 1992.
193
Figure 6 The broadband power cube of the data cube obtained on 30 June 1993 in the circular polarisation
channel (V = R L) with NoRH.
detection, as the presence of the signal in the data will be confirmed or disproved by the
following analysis.)
The spatial distribution of the oscillation power in the data set of 30 June 1993 is shown
by the broadband power cube (Figure 6). It is clearly seen that the 3- and 15-minute oscillations are located at the centre of the microwave source, which lies over the sunspots umbra,
whereas the five-minute oscillations are offset from the centre of the microwave source.
However, the broadband power cube does not give us information about the temporal evolution of the oscillations. Applying the PWF method to the analysed datasets, we construct
dynamical narrowband maps of the data cubes for the three significant periodicities found.
Narrowband power maps showing the spatial distribution of the 3-, 5-, and 15-minute
oscillations in the analysed data cubes are given in Figure 7. According to the maps, there
is an obvious difference in the position and shapes of the spatial sources of the detected
periodicities. The size, shape, and positioning of all three sources coincide in the data of
22 July 1994 (Figure 7, upper panels) only. For the data set of 30 June 1993 (Figure 7, middle
panels), the source of three-minute oscillations is situated at the central part of the integrated
microwave image. The source of the 15-minute oscillations is situated at the centre of the
integrated image too. However, the source of five-minute oscillations forms two patches of
about 15 in size. The five-minute oscillation sources seem to be situated at the umbra
penumbra boundary.
The phase narrowband maps of the event on 30 June 1993 reveal that the oscillations of
each periodicity are in phase. This suggests that the observed oscillations are either standing
or propagating along the line of sight. The analysis of the data of 16 August 1992 reveals
similar morphology: The 3- and 15-minute oscillation sources are situated in the central
part of the microwave source, but the five-minute oscillation source is found to be split into
several patches situated at the umbra penumbra boundary. The five-minute oscillations in
some patches are antiphase with each other. This antiphase character of the five-minute os-
194
Figure 7 Narrowband power maps of the 3-, 5-, and 15-minute oscillation sources for the data of 22 July
1994, 30 June 1993, and 16 August 1992. The contours show the integrated microwave image. The arrows
point out the patches of the antiphase five-minute oscillations.
cillations may reduce the observed power of the corresponding spectral peak in the spectrum
of the signal integrated over the whole microwave source. This possibly explains the domination of the three-minute component in the spatially integrated signal.
5. Application to Coronal Loop Oscillations

In this section, we apply the PWF method to the analysis of the EUV data obtained with
TRACE on 23 March 1999 for the active region 8496. The data were obtained in the 171
bandpass with a temporal cadence of about nine second and a 1 pixel size. This data cube
contains propagating longitudinal waves analysed by De Moortel, Hood, and Ireland (2002)
with the use of the timedistance plots. Later on, Terradas, Oliver, and Ballester (2004) used
these data to demonstrate the applicability of the EMD and CEOF methods. There were two
195
Figure 8 Left: The original image of the coronal loop. Right: Spatial variance of EUV emission with position
of the loop details where observed standing (about 12 minutes) and travelling (about 5 minutes) waves are
indicated.
distinct periodicities found in the data, about 3 and 11 minutes, in the form of both standing
and propagating waves. Our aim is to confirm this finding with the PWF method.
The scheme of the application of the PWF method to this data cube is similar to that
described earlier. First, the data are preprocessed with the use of the SSW IDL routines
trace_prep, trace_align_cube, and align_cube_correl. A snapshot of the data cube is shown
in the left panel of Figure 8. The broadband variance map is shown in the right panel of
Figure 8. The strongest signal is seen to be localised along fine structures extended in the
radial direction along the expected direction of the magnetic field. The prevailing periods
of the oscillations present in the data were determined with the use of the global spatially
integrated wavelet power spectrum and are found to be 5, 8, and 12 minutes. In addition, the
analysed date cube was found to have strong high-frequency noise. For each of the 5- and
12-minute oscillations we constructed a sequence of the amplitude narrowband maps shown
in Figure 9. It is evident that five-minute oscillations are amplitude variations propagating
outwards along the same fine structures that are highlighted in Figure 8. The longer period
oscillations, with a period of about 12 minutes, are seen to be situated near the outer edges
of the loop as a nonpropagating wave, either standing or propagating along the line of sight.
In those regions, the phase changes periodically without any movement of the nodes and
extrema. A similar effect has already been found by Terradas, Oliver, and Ballester (2004)
with the use of EMD and CEOF methods.
6. Conclusions
The main aim of this work was the development and testing of a method for the analysis of
astronomical (in particular, solar) dynamical imaging data sets: the pixelised wavelet filtering method. The PWF method is a development of the wavelet technique proposed by De
Moortel and McAteer (2004). The main novel element of this technique is the possibility of
obtaining dynamical amplitude and phase narrowband maps (or movies), which provide us
with information about the wave type (propagating or standing in the FOV) and allow us to
study the temporalspatial structure of quasi-harmonic signals. The dynamical narrowband
196
Figure 9 The snapshots of the amplitude distribution of the propagating (with periods of about five minutes)
and nonpropagating (with periods of about 12 minutes) waves along an EUV coronal loop at different instants
of time. The arrows point out the oscillation locations. The contours show the location of the loop.
maps can also be visualised by a sequence of narrowband maps calculated at certain instants
of time.
Testing the PWF method with artificial test data sets that contained standing and propagating harmonic waves demonstrated its efficiency and robustness in the detection of the
periodicities, identification of their type, and the determination of the spatial location of their
sources.
The application of the PWF method to the analysis of sunspot oscillations gives us narrowband maps of temporal sequences of 2D images (data cubes) obtained in the microwave
band with NoRH. The PWF analysis revealed that the 3-, 5-, and 15-minute periodicities,
dominating the global spectrum of the oscillations, have a quite different spatial distribution
over the sunspot. The 3- and 15-minute oscillations are situated in the umbra, whereas the
sources of 5-minute oscillations are localised at small-size patches at the umbra penumbra
boundary. A similar spatial distribution of sunspot oscillations was observed by Zhugzhda,
Balthasar, and Staude (2000) and Nindos et al. (2002).
197
Applying the PWF method to the coronal data sets obtained with TRACE in the EUV
171 band, we obtained the spatial structure of the previously found 5- and 12-minute
oscillations of the EUV intensity. The five-minute waves, propagating along the FOV, are
situated in narrow threads extended along the magnetic field. The 12-minute oscillations
are observed to be nonpropagating in the FOV and also situated in the narrow magnetic
field-aligned threads near the external boundaries of the magnetic fan coronal structure. Our
results confirm the findings of Terradas, Oliver, and Ballester (2004) with the EMD and
CEOF methods. As those methods are not well understood and have not become popular
in the solar physics community, the confirmation of this finding with a well-understood and
popular Morlet wavelet technique is of particular importance.
The possible interpretation of the observed phenomena will be discussed elsewhere.
In comparison with the alternative techniques already applied in coronal seismology
and other solar atmospheric wave studies, such as the EMD and CEOF methods (Terradas,
Oliver, and Ballester, 2004), the PWF method has the following advantages: We know the
exact boundaries of the frequency bands of the narrowband signal (as a multiple of two, and
as any other, including an arbitrarily chosen one); we can prescribe a certain frequency (or a
band) to be extracted and analysed. Also, by varying the wavenumber in the mother wavelet
function (and, possibly, changing the mother wavelet function itself) the PWF method allows
us to vary the frequencytemporal resolution of the wavelets as a frequency filter. Such flexibility is absent from the aforementioned alternative methods, where the internal frequencies
of the obtained elementary functions should be determined with other spectral methods, e.g.
with the Fourier transform.
Thus, we believe that the PWF method discussed here has a number of important advantages for the analysis of imaging data. Possible applications of the PWF method include
the study of the interaction of various wave modes, revealing the time-dependent energy
exchange between the modes. The PWF method can easily be modified to show the spatial
distribution of phase speeds (projected on the FOV plane) of detected wave processes. The
PWF method can work in combination with the pre-analysis period-mapping technique developed by Nakariakov and King (2007) for the automated detection of wave and oscillatory
processes in coronal EUV imaging data cubes.
Acknowledgements The authors thank Dr. Uralov for useful discussions. R.A.S. is grateful to the Russian
Fund of Fundamental Research for the funding provided under Grant Nos. N04-02-39003 and N05-07-90147.
The authors thank the Royal Society for the support through the International Incoming Short Visit scheme.
Wavelet software was provided by C. Torrence and G. Compo (http://paos.colorado.edu/research/wavelets).
References
Baudin, F., Bocchialini, K., Koutchmy, S.: 1996, Astron. Astrophys. 314, L9.
Christopoulou, E.B., Skodras, A., Georgakilas, A.A., Koutchmy, S.: 2003, Astrophys. J. 591, 416.
De Moortel, I., Hood, A.W., Ireland, J.: 2002, Astron. Astrophys. 381, 311.
De Moortel, I., Ireland, J., Walsh, R.W.: 2000, Astron. Astrophys. 355, L23.
De Moortel, I., McAteer, R.T.J.: 2004, Solar Phys. 223, 1.
Gelfreikh, G.B., Grechnev, V., Kosugi, T., Shibasaki, K.: 1999, Solar Phys. 185, 177.
Golyandina, N., Nekrutkin, V., Zhiglyavsky, A.: 2001, Analysis of Time Series Structure. SSA and Related
Techniques, Chapman and Hall/CRC, London/Boca Raton.
Grechnev, V.V.: 2003, Solar Phys. 213, 103.
Huang, N.E., Shen, Z., Long, S.R.: 1999, Annu. Rev. Fluid Mech. 31, 417.
Ireland, J., Walsh, R.W., Harrison, R.A., Priest, E.R.: 1999, Astron. Astrophys. 347, 355.
King, D.B., Nakariakov, V.M., Deluca, E.E., Golub, L., McClements, K.G.: 2003, Astron. Astrophys. 404,
L1.
198
Marsh, M.S., Walsh, R.W., De Moortel, I., Ireland, J.: 2003, Astron. Astrophys. 404, L37.
Meyer, Y., Roques, S. (eds.): 1993, Proceedings of the International Conference: Wavelets and Applications,
Editions Frontieres, Gif-sur-Yvette.
Milovanov, A.V., Zelenyj, L.M.: 1993, Phys. Fluids B 5, 2609.
Nakariakov, V.M., King, D.B.: 2007, Solar Phys. 241, 397.
Nakariakov, V.M., Verwichte, E.: 2005, Living Rev. Solar Phys. 2, 3. http://www.livingreviews.org/lrsp-2005-3.
Nindos, A., Alissandrakis, C.E., Gelfreikh, G.B., Bogod, V.M., Gontikakis, C.: 2002, Astron. Astrophys. 386,
658.
Terradas, J., Oliver, R., Ballester, J.L.: 2004, Astrophys. J. 614, 435.
Torrence, C., Compo, G.P.: 1998, Bull. Am. Meteorol. Soc. 79, 61.
Weng, H., Lau, K.-M.: 1994, J. Atmos. Sci. 51, 2523.
Zhugzhda, Y.D., Balthasar, H., Staude, J.: 2000, Astron. Astrophys. 355, 347.
A Time-Evolving 3D Method Dedicated

to the Reconstruction of Solar Plumes and Results
Using Extreme Ultraviolet Data
Nicolas Barbey Frdric Auchre Thomas Rodet
Jean-Claude Vial

Abstract An important issue in the tomographic reconstruction of the solar poles is the relatively rapid evolution of the polar plumes. We demonstrate that it is possible to take into
account this temporal evolution in the reconstruction. The difficulty of this problem comes
from the fact that we want a four-dimensional reconstruction (three spatial dimensions plus
time) whereas we only have three-dimensional data (two-dimensional images plus time).
To overcome this difficulty, we introduce a model that describes polar plumes as stationary
objects whose intensity varies homogeneously with time. This assumption can be physically
justified if one accepts the stability of the magnetic structure. This model leads to a bilinear inverse problem. We describe how to extend linear inversion methods to these kinds of
problems. Studies of simulations show the reliability of our method. Results for SOHO/EIT
data show that we can estimate the temporal evolution of polar plumes to improve the reconstruction of the solar poles from only one point of view. We expect further improvements
from STEREO/EUVI data when the two probes will be separated by about 60.
Keywords Tomography Plumes EIT SOHO Solar corona Temporal evolution
N. Barbey () F. Auchre J.-C. Vial

Institut dAstrophysique Spatiale, Universit Paris-Sud, Orsay, France
e-mail: nicolas.barbey@ias.u-psud.fr
F. Auchre
e-mail: frederic.auchere@ias.u-psud.fr
J.-C. Vial
e-mail: jean-claude.vial@ias.u-psud.fr
N. Barbey T. Rodet
Laboratoire des Signaux et Systmes, Suplc, Gif-sur-Yvette, France
N. Barbey
e-mail: nicolas.barbey@lss.supelec.fr
T. Rodet
e-mail: thomas.rodet@lss.supelec.fr
DOI: 10.1007/978-0-387-98154-3_15
199
200
N. Barbey et al.
1. Introduction
A method known as solar rotational tomography has been used to retrieve the 3D geometry
of the solar corona (Frazin, 2000; Frazin and Janzen, 2002). In this method the structures
are assumed to be stable during the time necessary to acquire the data. Since we generally
have only one point of view at our disposal, about 15 days are required to have data for
half a solar rotation at the poles. Here, we focus our study on solar polar plumes. They are
bright, radial, coronal ray structures located at the solar poles in regions of open magnetic
field. The study of plumes is of great interest because it may be the key to understanding
the acceleration of the fast component of the solar wind (Teriaca et al., 2003). However, the
three-dimensional (3D) shape of these structures is poorly known and different assumptions
have been made (e.g., Gabriel et al. 2005; Llebaria, Saez, and Lamy, 2002). The plumes
are known to evolve with a characteristic time of approximately 24 hours on spatial scales
typical of Extreme ultra-violet Imaging Telescope (SOHO/EIT) data (2400 km) (DeForest, Lamy, and Llebaria, 2001). Consequently, the stability assumption made in rotational
tomography fails. Fortunately, the Solar TErestrial RElations Observatory (STEREO) mission consists of two identical spacecraft, STEREOA and STEREOB , which take pictures of
the Sun from two different points of view. With the SOHO mission still operating, this results
in three simultaneous points of view. Three viewpoints help to improve the reconstruction
of the plumes, but they are still insufficient for use in standard tomographic algorithms. The
problem is underdetermined and consequently one has to add an a priori information to
overcome the lack of information. This leads to challenging and innovative signal analysis
problems.
There are different ways to deal with underdetermination depending on the kind of object
to be reconstructed. Interestingly, the field of medical imaging faces the same kind of issues.
In cardiac reconstruction, use is made of the motion periodicity in association with a high
redundancy of the data (Grass et al., 2003; Kachelriess, Ulzheimer, and Kalender, 2000). If
one can model the motion as an affine transformation, and if one assumes that we know this
transformation, one can obtain an analytic solution (Ritchie et al., 1996; Roux et al., 2004).
In solar tomography, the proposed innovative approaches involve the use of additional
data such as magnetic-field measurements in the photosphere (Wiegelmann and Inhester,
2003) or data fusion (Frazin and Kamalabadi, 2005). Attempts have been made by Frazin et
al. (2005) to treat temporal evolution by using Kalman filtering.
Since polar plumes have apparently a local, rapid, and aperiodic temporal evolution, we
developed, as in the previously referenced work, a model based on the specifics of the object
we intend to reconstruct (preliminary results of which can be found in Barbey et al., 2007).
Plumes have an intensity that evolves rapidly with time, but their position can be considered
as constant. This hypothesis is confirmed by previous studies of the plumes such as that of
DeForest, Lamy, and Llebaria (2001). The model is made up of an invariant morphological
part (x) multiplied by a gain term ( t ) that varies with time. Only one gain term is associated
with each plume to constrain the model. So we assume that the position of each plume in the
scene is known. This model is justified if we consider polar plumes to be slowly evolving
magnetic structures in which plasma flows.
Thanks to this model we can perform time-evolving 3D tomography of the solar corona
using only extreme ultraviolet images. Furthermore, there is no complex, underlying physical model. The only assumptions are the smoothness of the solution, the area-dependent
evolution model, and the knowledge of the plume position. These assumptions allow us
to consider a temporal variation of a few days, while assuming only temporal smoothness
would limit variations to the order of one solar rotation (about 27 days). To our knowledge,
A Time-Evolving 3D Method Dedicated to the Reconstruction of Solar
201
the estimation of the temporal evolution has never been undertaken in tomographic reconstruction of the solar corona.
We first explain our reconstruction method in a Bayesian framework (Section 2). We then
test the validity of our algorithm with simulated data (Section 3). An example of a reconstruction on real SOHO/EIT data is shown in Section 4. Results are discussed in Section 5.
We conclude in Section 6.
2. Method
Tomographic reconstruction can be seen as an inverse problem, the direct problem being the
acquisition of data images knowing the emission volume density of the object (Section 2.1).
If the object is evolving during the data acquisition, the inverse problem is highly underdetermined. So our first step is to redefine the direct problem by a reparametrization that
defines more constraints (Section 2.2). Then, we place ourselves in the Bayesian inference
framework in which data and unknowns are considered to be random variables. The solution
of the inverse problem is chosen to be the a posteriori maximum (Section 2.3). This leads
to a criterion that we minimize with an alternate optimization algorithm (Section 2.4).
2.1. Direct Problem
The geometrical acquisition is mathematically equivalent to a conical beam data acquisition
with a virtual spherical detector (see Figure 1). In other words, the step between two pixels
vertically and horizontally is constant in angle. The angle of the full field of view is around
45 arcminutes. To obtain an accurate reconstruction, we take into account the exact geometry, which means the exact position and orientation of the spacecraft relatively to Sun center.
We approximate integration of the emission in a flux tube related to a pixel by an integration
along the line of sight going through the middle of that pixel. We choose to discretize the
object in the usual cubic voxels. x is a vector of size N containing the values of all voxels.
In the same way, we define the vector of data yt of size M at time t . Since the integration
operator is linear, the projection can be described by a matrix Pt . We choose nt to be an
additive noise:
yt = Pt xt + nt ,
t [1, . . . , T ].
(1)
Pt is the projection matrix at time t of size M N, which is defined by the position and
the orientation of the spacecraft at this time. Its transpose is the back-projection matrix.
Note that a uniform sampling in time is not required. To be able to handle large problems
with numerous well-resolved data images and a large reconstruction cube, we chose not
to store the whole projection matrix. Instead, we perform the projection operation (P x)
or its transpose each time it is needed at each iteration. Thus, we need a very efficient
algorithm. We developed a code written in C that performs the projection operation. It makes
use of the geometrical parameters given in the data headers to take into account the exact
geometry (conicity, position, and orientation of the spacecraft). To keep this operation fast,
we implemented the Siddon algorithm (Siddon, 1985), which allows a fast projection or back
projection in the case of cubic voxels (Cartesian grid). Since we focus on a small region at
the poles, we consider that we do not need to use a spherical grid, which would require a
more time-consuming projection algorithm.
We take into account the fact that the field of view is conical. Despite the acquisition
being very close to the parallel acquisition geometry, it is sufficient to introduce an error
of several voxels of size 0.01 solar radius from one side to the other of a three-solar-radii
reconstructed cube.
202
N. Barbey et al.
Figure 1 Scheme of the data-acquisition geometry. (O; x, y, z) defines the Carrington heliocentric frame of
reference; S is the spacecraft considered, is the latitude and the longitude of this spacecraft, and V is the
virtual detector.
2.2. Modeling of the Temporal Evolution

With this model, the inverse problem is underdetermined since we have at most three images
at one time and we want to reconstruct the object with its temporal evolution. To do so,
we first redefine our unknowns to separate temporal evolution from spatial structure. We
introduce a new set of variables gt of size N describing the temporal evolution and require
that x does not depend on time:
yt = Pt (x gt ) + nt ,
(2)
with being the term-by-term multiplication of vectors. This operator is clearly bilinear.
However, this model would increase the number of variables excessively. So, we need to
introduce some other kind of a priori information into our model. We make the hypothesis
that all of the voxels of one polar plume have the same temporal evolution:
gt = L t .
(3)
The matrix L of size N P (where P is the number of areas) localizes areas where the
temporal evolution is identical. Each column of L is the support function of one of the
plumes. We would like to stress that in our hypothesis, those areas do not move relative to
the object. In other words, L does not depend on time. Localizing these areas defines L and
only leaves P T variables to estimate. We redefined our problem in a way that limits the
number of parameters to estimate but still allows many solutions. Furthermore, the problem
is linear in x if we know and linear in if we know x. This will simplify the inversion of
the problem, as we shall see later. Note, however, that the uniqueness of a solution (x, ) is
not guaranteed with bilinearity despite its being guaranteed in the linear case. This example
shows that A can be chosen arbitrarily without changing the closeness to the data: x g =
(Ax) (A1 g), where A is a real constant. Introducing an a priori feature of closeness
to 1 for would allow us to deal with this indeterminacy in principle. But note that this
indeterminacy is not critical since the physical quantity of interest is only the product x g.
Fron, Duchne, and Mohammad-Djafari (2005) present a method that solves a bilinear
inversion problem in the context of microwave tomography.
We do not deal with the estimation of the areas undergoing evolution, but we assume in
this paper that the localization is known. This localization can be achieved by using other
sources of information (e.g., stereoscopic observations). We expect to be able to locate the
areas using some other source of information.
203
We can regroup the equations of the direct problem. We have two ways to do so, each
emphasizing the linearity throughout one set of variables. First, we can write
y = Ux + n,

1
y1
0
P1 XL
n1
.. ..
..
..
. + . ,
. =
.
yT
T
nT
0
PT XL
(4)
with X = diag(x), the diagonal matrix defined by x, where x is of size N , y and n are of
size MT , is of size P T , and Ux is of size M T P T .
Similarly,
y = V x + n,
with V =
P1 diag(L 1 )
0
Id
..
..
. ,
.
Id
PT diag(L T )
0
(5)
with Id the identity matrix of size M M. V is of size MT N .

2.3. Inverse Problem
In Bayes formalism, solving an inverse problem consists in knowing the a posteriori information (the conditional probability density function of the parameters, with the data being
given). To do so we need to know the likelihood (the conditional probability density function of the data knowing the parameters) and the a priori information (the probability density
function of the parameters). An appropriate model is a Gaussian, independent, identically
distributed (with the same variance) noise n. The likelihood function is deduced from the
noise statistic:

y Ux 2
,
(6)
f (y|x, , n , M) = K1 exp
2n2
where M = [P , L] describes our model (the projection algorithm and parameters and the
choice of the plume position). We assume that the solution is smooth spatially and temporally, so we write the a priori information as follows:

Dr x2
Dt 2
and f ( | ) = K3 exp
,
(7)
f (x|x ) = K2 exp
2x2
22
Dr and Dt are discrete differential operators in space and time. Bayes theorem gives us the
a posteriori law if we assume that the model M is known as well as the hyperparameters
H = [n , x , ]:
f (x, |y, H, M) =
f (y|x, , n , M)f (x|x )f ( | )

.
f (y|H, M)
(8)
We need to choose an estimator. This allows us to define a unique solution instead of having
a whole probability density function. We then choose to define our solution as the maximum
a posteriori. which is given by
MAP MAP
(9)
= arg max f (y|x, , n , M)f (x|x )f ( | ),
x ,
x,
204
N. Barbey et al.
since f (y|M) is a constant. Equation (9) can be rewritten as a minimization problem:
MAP MAP
(10)
= arg min J (x, ),
x ,
x,
with
J (x, ) = 2n log f (x, |y, M, H) = y Ux 2 + Dr x2 + Dt 2 ,
(11)
where = n2 /x2 and = n2 /2 are user-defined hyperparameters.

The equivalence of Equations (9) and (10) has been proved by Demoment (1989).
Note that the solution does not have to be very smooth. It mostly depends on the level of
noise, since noise increases the underdetermination of the problem, as has been shown by
the definition of and .
2.4. Criterion Minimization
The two sets of variables x and are very different in nature. However, because of the
problems bilinearity, one can easily estimate one set while keeping the other fixed. Consequently, we perform an iterative minimization of the criterion, and we alternate minimization
of x and . At each step n we perform

and xn+1 = arg min J x, n+1 .
(12)
n+1 = arg min J xn ,
x
The two subproblems are formally identical, but is much smaller than x. This is of
the utmost practical importance since one can directly find the solution on by using the
pseudo-inverse method (Appendix A). However, x is too big for this method, and we have to
use an iterative scheme such as the conjugate-gradient method to approximate the minimum
(Appendix B).
2.5. Descent Direction Definition and Stop Threshold
We choose to use an approximation of the conjugate-gradient method that is known to converge much more rapidly than the simple gradient method (Nocedal and Wright, 2000;
Polak and Ribire, 1969):
dp + bp x J |x=xp ,
x J |x=xp , x J |x=xp1
=
.
x J |x=xp1 2
dp+1 =
bp
(13)
Since the minimum is only approximately found, we need to define a threshold that we
consider to correspond to an appropriate closeness to the data to stop the iterations. Since the
solution is the point at which the gradient is zero, we choose this threshold for updating x:
meanx[xp ,xp1 ,xp2 ] x J 2 < Sx .
For the global minimization, the gradient is not computed, so we choose

2
mean[n,n1,n2] (xn , n ) (xn1 , n1 ) < SG .
(14)
(15)
Note that this way of stopping the iteration allows one to define how close one wants to be
to the solution: If the difference between two steps is below this threshold, it is considered
negligible. The algorithm can be summarized as shown in Figure 2.
205
initialize : x = 0 and = 1
while Equation (15) is satisfied
x minimization:
while Equation (14) is satisfied
compute gradient at xn with Equation (20)

compute descent direction with Equation (13)
compute optimum step with Equation (22)
update x with Equation (23)
endwhile
minimization:
compute the matrix UxTn Uxn and the vector UxTn y
inverse the matrix UxTn Uxn + DrT Dr
compute Equation (19)
endwhile
Figure 2 Tomographic reconstruction with temporal evolution algorithm.
Table 1 Simulation definition:
plume parameters.
Plume
Semimajor
Semiminor
number
axis a
axis b
4.8
4.2
1.2
29
29
329
5.6
3.3
1.1
23
33
430
5.2
4.8
0.1
40
42
723
x0
y0
Intensity
(A)
3. Method Validation
To validate the principle of our method and test its limits, we simulate an object containing
some plumes with temporal evolution and try to extract them from the data.
3.1. Simulation Generation Process
We generate an emission cube with randomly placed, ellipsoidal plumes with a Gaussian
shape along each axis:

1 r.u ) 2 1 r.u+ 2 2
.
Ep = A exp
2
a
2
b
(16)
The plumes evolve randomly but smoothly by interpolating over a few randomly generated
points. Once the object is generated, we compute a typical set of 60 images equally spaced
along 180 using our projector algorithm. A Gaussian random noise is added to the projections with a signal-to-noise ratio (SNR) of five. The simulation parameters are summarized
in Table 1.
3.2. Analysis of Results
We now compare our results (Figure 3) with a filtered back-projection (FBP) algorithm. This
method is explained by Natterer (1986) and Kak and Slaney (1987).
206
N. Barbey et al.
Table 2 Simulation definition: geometric parameters.

Cube size
Cube number
Pixel
Projection
(solar radii)
of voxels
size (radians)
number of pixels
1 1 0.05
64 64 4
5 105 5 105
128 8
Table 3 Simulation definition:

other parameters.
SNR
5
Sx
SG
2 102
100
2 102
1 102
Figure 3 Comparison of a standard FBP method (a), the real simulated object (b), and the object reconstructed with our method (c). The object is reconstructed by using 60 projections regularly spaced over 180 .
The areas of homogeneous temporal evolution (d) are the same in the simulation and the reconstruction. We
associated one time per projection to define in the simulation (e) and our reconstruction (f). The time scale
is in days, under the assumption of a rotation speed of half a rotation in 14 days. x is the spatial distribution
of the emission density volume. is a gain representing the emission variation over time. Except for the FBP
reconstruction, only the product x has physical dimensions. The spatial scales are given in solar radii and
centered on the solar axis of rotation. (a), (b), and (c) are slices of 3D cubes at the same z = 0.1R
. Emission
densities (arbitrary units) are scaled by the color bars on the right side of (a), (b), and (c).
By comparing the simulation and the reconstruction in Figure 3, we can see the quality
of the temporal evolution estimation. The shape of the intensity curves is well reproduced
except for the first plume in the first ten time steps where the intensity is slightly underestimated. This corresponds to a period when plume 1 is hidden behind plume 2. Thus, our
algorithm attributes part of the plume 1 intensity to plume 2. Let us note that this kind of
ambiguity will not arise in the case of observations from multiple points of view such as
STEREO/EUVI observations. The indeterminacy of the problem is due to its bilinearity, as
discussed in Section 2.2. This allows the algorithm to attribute larger values to the parameters and to compensate by decreasing the corresponding x. This is not a drawback of the
207
Figure 4 Comparison of x g simulated and reconstructed at different times. T is the time between two
data images (5.6 hours). Distances are in solar radii. Values represent the volume emission density. All of
these images are slices of 3D cubes at the same z = 0.1R
.
method since it allows discontinuities between plumes and interplumes. The only physical
value of interest is the product x g.
Figure 4 shows the relative intensity of the plumes at different times. One can compare
these results with the reconstruction. One way to quantify the quality of the reconstruction
is to compute the distance (quadratic norm of the difference) between the real object and
the reconstructed one. Since the FBP reconstruction does not actually correspond to a reconstruction at one time, we evaluate the minimum of the distances at each time. We find it
to be 3000. This is to be compared with a value of 700 with our algorithm, which is much
better.
3.3. Choice of Evolution Areas
One can think that the choice of the evolution areas is critical to the good performance
of our method. We show in this section that it is not necessarily the case by performing a
reconstruction based on simulations with incorrect evolution areas. All parameters and data
are exactly the same as in the previous reconstruction. The only difference is in the choice
of the areas (i.e., the L matrix). These are now defined as shown in Figure 5(a).
Although approximately 50% of the voxels are not associated with their correct area, we
can observe that the algorithm still performs well. The emission map of Figure 5(b) is still
better than the emission reconstructed by an FBP method. Moreover, the estimation of the
temporal evolution in Figure 5(c) corresponds to the true evolution, Figure 3(e), even if less
precisely than in Figure 3(f).
208
N. Barbey et al.
Figure 5 Reconstruction with smaller areas (to be compared with Figure 3): (a) areas no longer corresponding to the ones used to generate the data, (b) emission map, and (c) temporal evolution estimated with our
algorithm. (b) and (c) are slices of 3D cubes at the same z = 0.1 R
. Emission densities (arbitrary units) are
scaled by the color bars on the right side of (b).
Table 4 EIT data reconstruction: geometric parameters.
Cube size
Cube number
(solar radii)
of voxels
3 3 0.15
256 256 8
Pixel
size (radians)
2.55 105 2.55 105
Projection
number of pixels
512 38
4. Reconstruction of SOHO/EIT Data

4.1. Data Preprocessing
We now perform reconstruction using SOHO/EIT data. We have to be careful when applying
our algorithm to real data. Some problems may arise from phenomena not taken into account
in our model (e.g., cosmic rays or missing data).
Some of these problems can be handled with simple preprocessing. We consider pixels
hit by cosmic rays as missing data. They are detected with a median filter. These pixels and
missing blocks are labeled as missing data and the projector and the back projector do not
take them into account (i.e., the corresponding rows in the matrices are removed).
4.2. Analysis of Results
In Figures 6 and 7, we present results from 17.1 nm EIT data between 1 and 14 November
1996. This period corresponds to the minimum of solar activity when one can expect to have
less temporal evolution. The 17.1 nm data are chosen because it is the wavelength where the
contrast of the plumes is the strongest. Some images are removed, resulting in a sequence
of 57 irregularly spaced projections for a total coverage of 191 . We assume that we know
the position of four evolving plumes as shown in Figure 6(b). For each reconstructed image,
we present 64 64 subareas of the reconstructed cube centered on the axis of rotation. We
assume the rotation speed to be the rigid-body Carrington rotation. All of the parameters
given in Tables 4 and 5 are shared by the different algorithms provided they are required by
the method. The computation of this reconstruction on a 3.0-GHz Intel Pentium 4 CPU took
13.5 hours.
The presence of negative values indicates poor behavior of the tomographic algorithm
since it does not correspond to actual physical values. We can see in Figure 6 that our reconstruction has considerably fewer negative values in the x map than the FBP reconstruction.
209
Figure 6 A comparison of FBP (a), the chosen areas (b), a gradient-like algorithm without temporal evolution (c), and our algorithm (d) with real EIT data. x is the spatial distribution of the volume emission density
integrated over the EIT 17.1 nm passband. is a gain representing the emission variation during time (e). The
time scale is in days. In the case of our algorithm, only the product x has physical meaning. The spatial
scales are given in solar radii and centered on the solar axis of rotation. (a), (b), (c), and (d) are slices of 3D
cubes at the same z = 1.3R
. Emission densities (arbitrary units) are scaled by the color bars on the right
sides of (a), (c), and (d).
210
Table 5 EIT data
reconstruction: other parameters.
N. Barbey et al.
Sx
SG
2 102
1 104
0.1
0.05
Figure 7 Reconstruction of x g at different times. Distances are in solar radii. Values represent the volume
emission density integrated over the EIT 17.1 nm passband. All of these images are slices of 3D cubes at the
same z = 1.3R
.
In the FBP reconstruction cube, 50% of the voxels have negative values; in the gradient-like
reconstruction without temporal evolution 36% of the voxels are negative whereas in our
reconstruction only 25% are negative. This still seems like a lot but most of these voxels
are in the outer part of the reconstructed cube. The average value of the negative voxels is
much smaller also. It is 120 for the FBP, 52 for the gradient-like method without temporal evolution, and only 19 for our reconstruction with temporal evolution. However, we
notice that the gain coefficients present a few slightly negative values.
In the reconstructions without temporal evolution, plumes three (upper right) and four
(lower right) correspond to a unique elongated structure, which we choose to divide. Note
how our algorithm updated the x map, reducing the emission values between these two
plumes and showing that what was seen as a unique structure was an artifact resulting from
temporal evolution, which tends to validate the usefulness of our model in estimating the
emission map. We note the disappearance of a plume located around (0.2, 0.15) solar
radii on the FBP reconstruction. This shows the utility of gradient-like methods in eliminating artifacts arising from the nonuniform distribution of images. Another plume at (0.2, 0.2)
solar radii has more intensity in the reconstruction without temporal evolution than with our
algorithm, illustrating how temporal evolution can influence the spatial reconstruction.
5. Discussion
The major feature of our approach is the quality of our reconstruction, which is much improved with respect to FBP reconstruction, as demonstrated by the smaller number of negative values and the increased closeness to the data. Let us now discuss the various assumptions that have been made through the different steps of the method.
The strongest assumption we made, to estimate the temporal evolution of polar plumes,
is the knowledge of the plume position. Here, we choose to define the plumes as being the
brightest points in a reconstruction without temporal evolution. The choice is not based on
any kind of automatic threshold. The areas are chosen by the user by looking at a reconstruction. It is possible that these areas do not correspond to the actual physical plumes;
211
they could correspond to areas presenting increased emission during half a rotation. Note
that this is biased in favor of plumes closer to the axis of rotation since, along one slice of
the reconstructed Cartesian cube, their altitude is lower and thus their intensity is higher. To
have constant altitude maps one would have to carry out the computation on a spherical grid
or to interpolate afterward onto such a grid. For this reconstruction example we are aware
that we did not locate all of the plumes but only tried to find a few. It would be interesting
to try to locate the plumes using other data or with a method estimating their positions and
shapes.
The method involves hyperparameters, which we choose to set manually. There are methods to estimate hyperparameters automatically such as the L-curve method, the cross-validation method (Golub, Heath, and Wahba, 1979), or the full-Bayesian method (Higdon et al.,
1997; Champagnat, Goussard, and Idier, 1996). We performed reconstructions using different hyperparameter values. We then looked at the reconstruction to see whether the smoothness seemed exaggerated or whether the noise became amplified in the results. This allowed
us to reduce the computational cost and does not really put the validity of the method into
question.
One possible issue with this algorithm is the nonconvexity of our criterion. This can lead
to the convergence to a local minimum that does not correspond to the desired solution
defined as the global minimum of the criterion. One way to test this would be to change the
initialization many times.
We chose the speed of rotation of the poles to be the Carrington rotation speed. But the
speed of the polar structures has not been measured precisely to our knowledge and could
drastically affect the reconstruction. This is an issue shared by all tomographic reconstructions of the Sun.
In the current approach, we need to choose on our own the position of the time-evolving
areas, which are assumed to be plumes. This is done by choosing the more intense areas
of a reconstruction without temporal evolution. A more rigorous way would be to try to
use other sources of information to try to localize the plumes. Another, self-consistent way
would be to develop a method that jointly estimates the position of the plumes in addition
to the emission (x) and the time evolution ( ). We could try to use the results of Yu and
Fessler (2002), who propose an original approach to reconstruct a piecewise homogeneous
object while preserving edges. The minimization is alternated between an intensity map
and boundary curves. The estimation of the boundary curves is made by using level sets
techniques (Yu and Fessler, 2002, and references therein). It would also be possible to use a
Gaussian mixture model (Snoussi and Mohammad-Djafari, 2007).
6. Conclusion
We have described a method that takes into account the temporal evolution of polar plumes
for tomographic reconstruction near the solar poles. A simple reconstruction based on simulations demonstrates the feasibility of the method and its efficiency in estimating the temporal evolution by assuming that parameters such as plume position or rotation speed are
known. Finally, we show that it is possible to estimate the temporal evolution of the polar
plumes with real data.
In this study we limited ourselves to reconstruction of images at 17.1 nm but one can
perform reconstructions at 19.5 and 28.4 nm as well. This would allow us to estimate the
temperatures of the electrons, as in Frazin, Kamalabadi, and Weber (2005) or Barbey et al.
(2006).
212
N. Barbey et al.
Acknowledgements Nicolas Barbey acknowledges the support of the Centre National dtudes Spatiales
and the Collecte Localisation Satellites. The authors thank the referee for useful suggestions for the article.
Appendix A: Pseudo-Inverse Minimization

We want to minimize
2

J = y Uxn 2 + Dr xn + Dt 2 .
(17)
The second term does not depend on . Owing to the strict convexity of the criterion, the solution is a zero of the gradient. Since the criterion is quadratic, one can explicitly determine
the solution:

J | = n+1 = 2UxTn Uxn n+1 y + 2DtT Dt n+1 = 0,
(18)
from which we conclude

1
n+1 = UxTn Uxn + DtT Dt UxTn y.
(19)
Appendix B: Gradient-Like Method

In this method we try to find an approximation of the minimum by decreasing the criterion
iteratively. The problem is divided into two subproblems: searching for the direction and
searching for the step of the descent. In gradient-like methods, the convergence is generally
guaranteed ultimately to a local minimum. But because the criterion is convex, the minimum
is global. To iterate, we start at an arbitrary point (x0 ) and go along a direction related to the
gradient. The gradient at the pth step is

x J |x=xp = 2VTn+1 V n+1 xp y + 2DrT Dr xp .
(20)
Once the direction is chosen, searching for the optimum step is a linear minimization problem of one variable:

p+1
aOPT = arg min J xp + adp+1 ,
(21)
a
which is solved by
p+1
aOPT =
dp+1 x J |x=xp
1
.
2 V n+1 dp+1 2 + Dr dp+1 2
(22)
We can write the iteration

p+1
xp+1 = xp + aOPT dp+1 .
(23)
213
References
Barbey, N., Auchre, F., Rodet, T., Bocchialini, K., Vial, J.C.: 2006, Rotational tomography of the solar
corona-calculation of the electron density and temperature. In: Lacoste, H. (ed.) SOHO 17 10 Years of
SOHO and Beyond, ESA, Noordwijk, 66 68.
Barbey, N., Auchre, F., Rodet, T., Vial, J.C.: 2007, Reconstruction tomographique de squences dimages
3D application aux donnes SOHO/STEREO. In: Flandrin, P. (ed.) Actes de GRETSI 2007, Association
GRETSI, Troyes, 709 712.
Champagnat, F., Goussard, Y., Idier, J.: 1996, Unsupervised deconvolution of sparse spike trains using stochastic approximation. IEEE Trans. Signal Process. 44(12), 2988 2998.
DeForest, C.E., Lamy, P.L., Llebaria, A.: 2001, Solar polar plume lifetime and coronal hole expansion: Determination from long-term observations. Astrophys. J. 560, 490 498.
Demoment, G.: 1989, Image reconstruction and restoration: Overview of common estimation structure and
problems. IEEE Trans. Acoust. Speech Signal Process. 37(12), 2024 2036.
Fron, O., Duchne, B., Mohammad-Djafari, A.: 2005, Microwave imaging of inhomogeneous objects made
of a finite number of dielectric and conductive materials from experimental data. Inverse Probl. 21(6),
S95 S115.
Frazin, R.A.: 2000, Tomography of the solar corona. I. A robust, regularized, positive estimation method.
Astrophys. J. 530, 1026 1035.
Frazin, R.A., Janzen, P.: 2002, Tomography of the solar corona. II. Robust, regularized, positive estimation
of the three-dimensional electron density distribution from LASCO-C2 polarized white-light images.
Astrophys. J. 570, 408 422.
Frazin, R.A., Kamalabadi, F.: 2005, Rotational tomography for 3D reconstruction of the white-light and EUV
corona in the post-SOHO era. Solar Phys. 228, 219 237.
Frazin, R.A., Kamalabadi, F., Weber, M.A.: 2005, On the combination of differential emission measure
analysis and rotational tomography for three-dimensional solar EUV imaging. Astrophys. J. 628, 1070
1080.
Frazin, R.A., Butala, M.D., Kemball, A., Kamalabadi, F.: 2005, Time-dependent reconstruction of nonstationary objects with tomographic or interferometric measurements. Astrophys. J. 635, L197 L200.
Gabriel, A.H., Abbo, L., Bely-Dubau, F., Llebaria, A., Antonucci, E.: 2005, Solar wind outflow in polar
plumes from 1.05 to 2.4 Rsolar . Astrophys. J. 635, L185 L188.
Golub, G.H., Heath, M., Wahba, G.: 1979, Generalized cross-validation as a method for choosing a good
ridge parameter. Technometrics 21(2), 215 223.
Grass, M., Manzke, R., Nielsen, T., KoKen, P., Proksa, R., Natanzon, M., Shechter, G.: 2003, Helical cardiac
cone beam reconstruction using retrospective ECG gating. Phys. Med. Biol. 48, 3069 3083.
Higdon, D.M., Bowsher, J.E., Johnson, V.E., Turkington, T.G., Gilland, D.R., Jaszczak, R.J.: 1997, Fully
Bayesian estimation of Gibbs hyperparameters for emission computed tomography data. IEEE Med.
Imag. 16(5), 516 526.
Kachelriess, M., Ulzheimer, S., Kalender, W.: 2000, ECG-correlated imaging of the heart with subsecond
multislice spiral CT. IEEE Med. Imag. 19(9), 888 901.
Kak, A.C., Slaney, M.: 1987, Principles of Computerized Tomographic Imaging, IEEE Press, New York.
Llebaria, A., Saez, F., Lamy, P.: 2002, The fractal nature of the polar plumes. In: Wilson, A. (ed.) From Solar
Min to Max: Half a Solar Cycle with SOHO, SP-508, ESA, Noordwijk, 391 394.
Natterer, F.: 1986, The Mathematics of Computerized Tomography, Wiley, New York.
Nocedal, J., Wright, S.J.: 2000, Numerical Optimization, Series in Operations Research, Springer, New York.
Polak, E., Ribire, G.: 1969, Note sur la convergence de mthodes de directions conjugues. Rev. Fr. Inf.
Rech. Oprat. 16, 35 43.
Ritchie, C.J., Crawford, C.R., Godwin, J.D., King, K.F., Kim, Y.: 1996, Correction of computed tomography
motion artifacts using pixel-specific backprojection. IEEE Med. Imag. 15(3), 333 342.
Roux, S., Debat, L., Koenig, A., Grangeat, P.: 2004, Exact reconstruction in 2d dynamic CT: Compensation
of time-dependent affine deformations. Phys. Med. Biol. 49(11), 2169 2182.
Siddon, R.L.: 1985, Fast calculation of the exact radiological path for a three-dimensional CT array. Med.
Phys. 12, 252 255.
Snoussi, H., Mohammad-Djafari, A.: 2007, Estimation of structured Gaussian mixtures: The inverse EM
algorithm. IEEE Trans. Signal Process. 55(7), 3185 3191.
Teriaca, L., Poletto, G., Romoli, M., Biesecker, D.A.: 2003, The nascent solar wind: Origin and acceleration.
Astrophys. J. 588, 566 577.
Wiegelmann, T., Inhester, B.: 2003, Magnetic modeling and tomography: First steps towards a consistent
reconstruction of the solar corona. Solar Phys. 214, 287 312.
Yu, D., Fessler, J.: 2002, Edge-preserving tomographic reconstruction with nonlocal regularization. IEEE
Med. Imag. 21, 159 173.
Automatic Detection and Classification of Coronal Holes

and Filaments Based on EUV and Magnetogram
Observations of the Solar Disk
Isabelle F. Scholl Shadia Rifai Habbal

Abstract A new method for the automated detection of coronal holes and filaments on the
solar disk is presented. The starting point is coronal images taken by the Extreme Ultraviolet
Telescope on the Solar and Heliospheric Observatory (SOHO/EIT) in the Fe IX / X 171 , Fe
XII 195 , and He II 304 extreme ultraviolet (EUV) lines and the corresponding full-disk
magnetograms from the Michelson Doppler Imager (SOHO/MDI) from different phases of
the solar cycle. The images are processed to enhance their contrast and to enable the automatic detection of the two candidate features, which are visually indistinguishable in these
images. Comparisons are made with existing databases, such as the He I 10830 NSO/Kitt
Peak coronal-hole maps and the Solar Feature Catalog (SFC) from the European Grid of
Solar Observations (EGSO), to discriminate between the two features. By mapping the features onto the corresponding magnetograms, distinct magnetic signatures are then derived.
Coronal holes are found to have a skewed distribution of magnetic-field intensities, with
values often reaching 100 200 gauss, and a relative magnetic-flux imbalance. Filaments,
in contrast, have a symmetric distribution of field intensity values around zero, have smaller
magnetic-field intensity than coronal holes, and lie along a magnetic-field reversal line. The
identification of candidate features from the processed images and the determination of their
distinct magnetic signatures are then combined to achieve the automated detection of coronal
holes and filaments from EUV images of the solar disk. Application of this technique to all
three wavelengths does not yield identical results. Furthermore, the best agreement among
all three wavelengths and NSO/Kitt Peak coronal-hole maps occurs during the declining
phase of solar activity. The He II data mostly fail to yield the location of filaments at solar
I.F. Scholl ()
International Space University, 1 rue J.D. Cassini, 67400 Illkirch-Graffenstaden, France
e-mail: isa.scholl@gmail.com
I.F. Scholl
Laboratoire dtudes Spatiales et dInstrumentation en Astrophysique (LESIA),
Observatoire de Paris-Meudon, Meudon, France
S.R. Habbal
University of Hawaii, 2680 Woodlawn Drive, Honolulu, HI 96822, USA
e-mail: shadia@ifa.hawaii.edu
DOI: 10.1007/978-0-387-98154-3_16
215
216
I.F. Scholl, S.R. Habbal
minimum and provide only a subset at the declining phase or peak of the solar cycle. However, the Fe IX / X 171 and Fe XII 195 data yield a larger number of filaments than the
H data of the SFC.
Keywords Sun: coronal holes Sun: filaments Sun: magnetic field Automatic detection
and classification
1. Introduction
Despite the absence of coronal magnetic-field measurements, the conventional wisdom is
that coronal holes are the site of unipolar magnetic-field lines that escape into interplanetary
space along with the solar wind. Coronal holes received their name when an almost total absence of emission was first noticed in solar-disk images taken with high-temperature spectral
lines in the visible, extreme ultraviolet (EUV), and X rays (e.g., Reeves and Parkinson, 1970;
Munro and Withbroe, 1972; Bell and Noci, 1976; Vaiana et al., 1976). Their association with
the source of the solar wind was first established when the longitude of a high-speed solarwind stream measured in interplanetary space in the ecliptic plane was mapped back to the
central meridian passage of a coronal hole at the Sun (Krieger, Timothy, and Roelof, 1973;
see also Sheeley, Harvey, and Feldman, 1976). If coronal holes are indeed the only source
of open magnetic flux at the Sun, then identifying them remains an important task.
Hand-drawn images of coronal holes, spanning almost three decades from 1974 through
September 2003, were produced by combining He I images taken with the National Solar
Observatory (NSO) Kitt Peak Vacuum Telescope (KPVT), with measurements of the He I
equivalent width and photospheric magnetic-field measurements (Harvey, Harvey, and Sheeley, 1982; Harvey and Recely, 2002; Henney and Harvey, 2005). The hand-drawn technique
often involved the use of two-day averages of He I 10830 spectroheliograms and magnetograms to produce the coronal-hole maps. The unique formation of the He I 10830 line in
the chromosphere is such that the intensity of this multiplet strongly depends on photoionizing radiation from the overlying upper transition region and corona (Goldberg, 1939). Any
reduction of this overlying radiation will lead to a weakening of the He II absorption (Andretta and Jones, 1997; Malanushenko and Jones, 2005), as is believed to occur primarily
in coronal holes where the electron density and temperature are lower than the rest of the
corona. To date, ground-based observations of the He I 10830 line remain the most widely
used proxy for estimating the location of coronal holes on the solar disk.
As the size of images and the frequency of observations grow, the automated detection
of coronal holes is becoming increasingly important. A number of approaches using the
He I ground-based observations have been recently developed (e.g., Henney and Harvey,
2005; Malanushenko and Jones, 2005; de Toma and Arge, 2005). Henney and Harvey (2005)
combined He I 10830 observations with photospheric magnetograms and developed a
scheme to parametrize, a priori, the properties of coronal holes as determined originally
from the hand-drawn images by Harvey and Recely (2002). The resulting coronal holes
were generally in better agreement with EIT Fe XII 195 line images than with the handdrawn coronal-hole maps. Using He I imaging spectroscopy, notably the half-widths and
central intensities of the He I 10830 line, Malanushenko and Jones (2005) found that the
outline of coronal holes on the solar disk agrees with those visually determined from EUV
measurements from SOHO/EIT but was not always in agreement with the NSO/Kitt Peak
coronal-hole maps. de Toma and Arge (2005) also defined a set of criteria for coronal hole
identification, a priori, and applied them to synoptic maps. They found that their coronal
holes formed a subset of the NSO/Kitt Peak coronal holes.
217
The advent of coordinated magnetic-field and EUV measurements from the Solar and
Heliospheric Observatory (SOHO), which cover different phases of the current solar activity cycle, from early 1996 until the present, offers a new opportunity for the automated
detection of coronal holes and filaments on the solar disk. Although enticing, the use of EUV
lines for the determination of coronal-hole boundaries is not without ambiguity. In fact, the
early EUV space observations showed that the intensity of many EUV emission lines is also
low near filaments, owing to the low electron densities in the coronal cavity surrounding the
filament and the absorption of the surrounding EUV radiation by the cool filament material
(e.g., Munro and Withbroe, 1972). Filaments appear as dark, thin, and elongated regions on
the solar disk in the H 6563 line. They are also dark but larger when observed in EUV
lines. Hence, any technique developed to determine the location of coronal holes on the
solar disk from EUV measurements must contend with the potential ambiguity introduced
by the presence of filaments. Consequently, the approach will include criteria to distinguish
between the two. Fortunately, coronal holes and filaments have distinct magnetic characteristics. Filaments develop along a polarity-reversal boundary, whereas coronal holes fall
within regions with a dominant magnetic polarity.
The goal of this study is to develop an automated approach to detect and classify coronal
holes (and by extension also filaments) on the solar disk using EUV data from the Extreme Ultraviolet Telescope (EIT) and photospheric magnetic-field measurements from the
Michelson Doppler Imager (MDI) on SOHO. The approach relies on the processing of EUV
images to enhance their contrast and on a classification method that relies on their distinct
magnetic characteristics.
2. Overview of Approach and Data Selection

The SOHO/EIT data encompass several spectral lines. Consequently, they reflect the behavior of different temperature and density structures in the solar atmosphere. By supplementing the EUV data with the corresponding magnetic-field measurements from SOHO/MDI, it
will be shown that the distinct magnetic signatures of coronal holes and filaments contribute
significantly to the establishment of a reliable set of criteria. The detection of these candidate features thus defined will be compared with existing feature catalogs. These consist
of i) the catalog of coronal-hole maps from NSO/Kitt Peak synoptic maps generated from
He I 10830 observations and ii) the Solar Feature Catalog (SFC), an on-line database
built in the framework of the European Grid of Solar Observations (EGSO) project (Bentley, Csillaghy, and Scholl, 2004) and implemented at Bradford University (Zharkova et al.,
2005). It contains detailed descriptions of filaments, sunspots, and active regions, detected
automatically from the Observatoire de Paris/Meudon and MDI observations.
The following SOHO keywords have been used to select the EUV data and magnetograms needed. The EUV/EIT data (Delaboudinire et al., 1995) selected from the
SOHO archive correspond to i) level zero data, with ii) science objective as FULL SUN
171/284/195/304, and iii) object as full FOV. The data thus consist of four synoptic fulldisk and full-resolution images taken, in general, every six hours at the four wavelengths,
one at the chromosphere/transition region interface (He II 304 ) and three corresponding
to coronal temperatures (Fe IX / X 171, Fe XII 195, and Fe XV 284 ). Because the Fe XV
284 data lack the necessary contrast for detection purposes, they will not be considered in
this study.
The magnetograms from SOHO/MDI (Scherrer et al., 1995) are used to compute the
magnetic-field intensity and polarity. Criteria to select these data from the archive are
218
i) level 1.8, with the most recent calibrated algorithm applied, and ii) the observing mode
as FD_Magnetogram_Sum. The data thus retrieved correspond to synoptic magnetograms
taken every 96 minutes with an average of five magnetograms. The noise level for this type
of data is 9.2 G (Liu, Zhao, and Hoeksema, 2004). Prior to any analysis, the quality of the
selected data, such as absence of telemetry gaps and instrumental problems, is checked.
Since comparisons will be made among the features determined from the combined
EIT/MDI data, the NSO/Kitt Peak coronal-hole maps, and the SFC, it is important to consider the time in selecting the data to be analyzed. The observing times must be as close
as possible to one another. They are therefore selected in the following order: i) Kitt Peak,
ii) SFC, iii) EIT, and iv) MDI, to minimize the assumptions made regarding the evolution of
different features and the corresponding magnetic field.
3. Data Preparation and Image Processing

Given that the data originate from different sources, that they are not identically co-temporal,
and that the images are not the same size (i.e., the pixel resolution is not the same in the EIT
and MDI images), the following steps must first be taken: i) a given time of observation is
identified, and all observations are scaled to match in size and time (i.e., rotation). ii) Imageprocessing techniques are then applied to the data thus matched to extract candidate features.
In what follows, the details of these two steps are presented.
To be able to compute magnetic-field intensity and polarity on extracted features, EIT
images and MDI magnetograms must be first co-aligned in spatial resolution (i.e., pixel size,
center of the disk, and radius) and rotation. To keep the magnetic-field values intact, the EIT
data are aligned with respect to the MDI data.1 Therefore, using the strict data selection
process given in Section 2 minimizes the rotation to be applied to each observation. It is
found that when overlapping SFC contours, their coordinates have to be rotated to match
the observing time of the magnetogram. Finally, NSO/Kitt Peak maps are also co-rotated
when needed.
A suite of image-processing operations (Gonzalez and Woods, 2002) has been implemented for the different wavelengths in the EUV EIT images. Once the images are calibrated
using the standard EIT routines, the following steps, which are summarized in Table 1, are
implemented: i) normalize the images (i.e., eliminate effects from the variability of the Sun,
such as nonuniform illumination and center-to-limb variation, as well as detector-aging effects), ii) apply filters for noise cleaning (i.e., apply smoothing and spatial linear filters and
very bright pixel removal), iii) enhance the contrast through histogram equalization (Gonzalez and Woods, 2002; Pizer et al., 1987), and iv) apply an image-segmentation technique
based on a combination of region-based and edge-based methods. The regions of interest are
subsequently extracted, and a final cleanup is performed, which consists of removing small
features. The candidate features are identified so that the next step involving comparison
with the magnetograms, and thus final classification, can be initiated.
These steps are not performed in exactly the same manner for all wavelengths. Although
Fe IX / X 171 and Fe XII 195 images are treated in an identical manner, the He II
304 images are not. For the latter, the global contrast needs to be enhanced to enable
the image segmentation and contour extraction because of the higher level of spatial detail
1 See the IDL Map Software for Analyzing Solar Images package for SolarSoft (http://orpheus.nascom.
nasa.gov/~zarro/idl/maps/maps.html).
219
Table 1 Summary of image-processing operations.

Step Operations
Detailed operations per wavelength

304
Calibration
Normalization
171
195
Standard EIT routine

Smoothing spatial linear filter
(filter out pixels brighter than their surrounding area)
Limb darkening removal
Contrast enhancement
Log scaling
n/a
Low-pass filter
n/a
Histogram equalization
Local histogram equalization
(range parameters
depend on the solar
cycle period)
5
Image segmentation:
histogram-based and
region-growing methods
Smoothing to connect
Smoothing to connect closed dark
closed dark small features

Thresholding
small features
Histogram equalization to isolate
(binary transformation)
darkest features
Thresholding (binary transformation)
morphological erode operation
Limb and disk center are

processed with different
parameters
Steps 4 and 5 are executed three

times for 171 and twice for 195
to extract feature of different sizes
Results are
Results are combined together
combined together
6
Regions of interest
Based on contour method
extraction
7
Final cleanup
Remove small features

<300 pixels
Region label
<1000 pixels
Identification of all features for next step
(magnetic-profile computation and classification)
observed at this wavelength. For the 171 and 195 images, the local-contrast enhancement is needed to reveal candidate features that could be hidden either because of excessive
brightness, such as for features close to an active region, or because of poor contrast. In the
former case, the features in the neighborhood of an active region can only be revealed when
images are processed locally. In the latter case, for large dark regions generally believed by
default to be coronal holes, slight changes in brightness are flattened and become invisible
without a local enhancement. In both cases, the local-contrast enhancement is such that the
dynamic range in brightness is reduced, and the contrast is increased (Stark, 2000). The
main side effect of this technique is the enhancement of noise or small spatial details, but no
artificial features are created (Gonzalez and Woods, 2002).
4. Candidate Features: Detection and Classification Method

The image-processing detection technique described in the previous section enables the extraction of candidate features, namely coronal holes and filament channels. Without process-
220
Figure 1 Example of SOHO/EIT data from solar minimum, on 27 May 1997. The top row corresponds to
He II 304 , the middle row to Fe IX / X 171 , and the bottom row to Fe XII 195 data from SOHO/EIT. The
left panels show the original data, the middle panels show the processed images, and the right panels show
the detected candidate features (i.e., coronal holes and filaments).
ing, the images do not have enough contrast to provide means for an automated detection
process (see Scholl, 2003). Three examples were selected from dates coinciding with solar
minimum (Figure 1), solar maximum (Figure 2), and the declining phase of the solar cycle
(Figure 3). In each of these figures there are three rows consisting of three panels each. The
top row corresponds to the He II 304 data, the middle to Fe IX / X 171 , and the bottom to Fe XII 195 . The left panel in each row shows the original EIT image, the middle
one gives the EIT image after processing, and the right panel shows the candidate features
extracted from the image processing, that is, filaments and coronal holes without distinction, placed on a Stonyhurst disk. Upon comparison of the left (original images) and middle
panels (processed images) in each row, it is clear that the processing brings out far more
detail.
At this point there is no other way to distinguish between coronal holes and filaments,
as their characteristics in all of the EIT EUV images, both unprocessed and processed, are
the same since they all appear as dark regions. There is also a significant variation in the
number, size, and shape of these features depending on the spectral line considered, as seen
in the examples of Figures 1 3.
221
Figure 2 Same as Figure 1 for solar maximum on 6 November 1999.
Once all candidate features have been isolated from each observation, the surface of
the feature is mapped onto the corresponding magnetogram to extract the corresponding
magnetic-field intensity and net polarity. From this step, two categories of objects emerge:
one is characterized by a dominant polarity, supported by a significant percentage of strong
values, mostly of one polarity; the other category has an average intensity close to zero, with
no strong values. The skewness of the distribution is also very different between the two.
To improve on these findings, a set of statistical estimators has been used. To establish
the magnetic polarity within a given feature, the weighted mean (net flux), the sum of the
field strengths in both polarities, the average of the two maximal values in each polarity, the
skewness (and its sign) of the real magnetic-field distribution, and the relative magnetic-flux
imbalance (Bz /|Bz |) are used. Furthermore, to analyze the distribution of the magneticfield intensity underlying a given feature, a simplified histogram with bins of < 170,
120, 70, 20, 20, 70, 120 and > 170 G is computed. This simplified distribution reveals
the nature of most filament channels. The noise level within the magnetograms is taken into
account in two different ways: each estimator is computed with and without the noise and
both results must agree if the percentage of values within the noise is not too excessive (i.e.,
< 70%).
Typical magnetic-field profiles for features selected from Figures 1 3 are shown in Figures 4 6. The top row in each figure corresponds to a filament channel, and the bottom one
222
Figure 3 Same as Figure 1 for the declining phase of the solar cycle on 23 March 2003.
to a coronal hole. The left panel in each row shows the candidate feature. The middle panel
is a histogram of the line-of-sight magnetic-field intensity, Bz , within the given area of the
feature. The right panel is the same histogram but with the magnetic-field values binned.
Among all these estimators (see the summary in Table 2), the most significant and discriminating one is the relative magnetic-flux imbalance. Other discriminators, which are not
used during the classification itself (but only during the manual analysis of the features),
allow us to specify two typical magnetic profiles: a coronal hole is always associated with a
relative magnetic-flux imbalance > 0.15, a skewness of the intensity distribution 0.7, and
a mean > 1. All estimators have the same sign, including the sum of the field strength (i.e.,
the number of pixels of the dominant polarity is always greater than the number of pixels of
the other polarity), and the minimum field intensity is > 70 G. For most filament channels,
the relative magnetic-flux imbalance is < 0.15, the skewness of the intensity distribution is
< 0.7, and the mean is around 0. Not all estimators have the same sign, and the maximum
intensity is < 70 G (with very few values above 70 G).
Although these estimators lead to a clear separation between coronal holes and filament
channels, a few outstanding issues remain in this classification scheme: i) features located
close to the limb can lead to a wrong classification owing to the poor resolution in these
areas. ii) Features too close to an active region can have strong magnetic-field values, probably coming from the active region itself. Such is the case, for example, of two features in
223
Figure 4 Left panels: Sample candidate filament (top) and coronal hole (bottom) for solar minimum, on 27
May 1997. Middle panels: Histogram of the line-of-sight magnetic-field intensity within the corresponding
feature. Also shown in these panels is the skewness, the size in pixels, the relative magnetic-flux imbalance
(RMFI), the maximum and minimum intensities for each polarity, and the mean. Right panels: Corresponding
histogram of the line-of-sight magnetic-field intensity for magnetic-field values binned as < 170, 120,
70, 20, 20, 70, 120, and > 170 G. The percentage of pixels for each polarity is also shown as well as the
percentage of pixels within the noise level.
Table 2 Summary of magnetic profiles for candidate features.

Magnetic profiles
Coronal holes
Filament channels
Relative magnetic-flux imbalance (RMFI)
> 0.15 G
< 0.15 G
Skewness
> 0.17
< 0.17
Weighted mean
>1
Intensity range
70 170 G
< 70 G
Estimators sign
All > 0 or all < 0
Mix of both signs
Figure 8, at coordinates ( 250, 300) and (600, 400), where in He II and Fe XII the
features are identified as filaments, but as a coronal hole in Fe IX / X. This is likely due to
the lack of precision in the contour detection process itself, especially for Fe IX / X 171 and
Fe XII 195 lines corresponding to hot plasmas, but not for the cooler plasma emitting at
He II 304 . iii) Some filament channels exhibit the same profile as coronal holes. iv) For a
coronal hole close to the limb the histogram may have the same shape as that of a filament
channel. v) The percentage of high values in all cases may not be significant for very small
features with sizes less than 2000 pixels, even though the mean can still be strong.
224
5. Candidate Features: Final Product

Figures 7 9 show the results of the application of these described criteria to the classification of the features shown in Figures 1 3 for the three wavelengths He II 304 (top row),
Fe IX / X 171 (middle rows), and Fe XII 195 (bottom rows). Figure 7 is an example from
solar minimum, Figure 8 from solar maximum, and Figure 9 from the declining phase of
225
Figure 7 Example of feature classification at solar minimum, on 27 May 1997. The top row corresponds to
He II 304 , the middle row to Fe IX / X 171 , and the bottom row to Fe XII 195 data from SOHO/EIT. In
the left panels, all candidate features, both coronal holes and filaments, detected with the method presented
here, are shown in black. Shown in red are active regions, the blue dots indicate sunspots, and the green
contours are filaments, all taken from the SFC. The background, in this and the middle panel, is the result of
the time-corresponding processed magnetogram from SOHO/MDI, as described in Scholl (2003), where the
white areas represent the dominant negative polarity, and the gray the dominant positive polarity. In the middle
panels, the filament channels detected are shown in black together with the same filaments (green) and active
regions (red) from the SFC also shown in the left panels. In the right panels, the detected coronal holes are
shown as orange contours plotted on a Stonyhurst grid together with the NSO/Kitt Peak coronal holes given
as black contours. The double solid-dashed blue line is the demarcation line between the opposite-polarity
magnetic-field areas shown as gray and white in the two left panels.
the solar cycle. In the left panels of each row, the candidate features are shown in black.
(Note that these are the same features shown in the right panels of Figures 1 3, which were
derived from the image processing step.) The red contours correspond to active regions, the
blue dots to sunspots, and the green contours to filament channels, all taken from the SFC.
The white and gray backgrounds in the left and middle panels represent the dominant negative and positive polarities, respectively, in the MDI processed magnetogram. As described
by Scholl (2003), the magnetogram is treated as an image. It is smoothed to eliminate small
scales. A line is subsequently drawn that separates pixels with opposite signs (i.e., polarities).
226
In these figures, the filament channels and coronal holes have been separated in the middle and right panels. The middle panels show the filament channels as identified from the
application of the magnetic-field criteria, plotted together with the active regions (red) and
filament channels (green) given by the SFC. The yellow contours in the right panels are
the detected coronal holes plotted on a Stonyhurst disk together with the black contours of
the NSO/Kitt Peak coronal holes. In the right panels the polarity inversion line from the
processed magnetogram is given as a thick blue line.
Several expected results emerge from these figures, together with some unexpected ones.
In general, there are more candidate features, both filaments and coronal holes, detected
with this approach than those appearing in the SFC or NSO/Kitt Peak coronal-hole maps.
Furthermore, the numbers, sizes, and distribution of candidate features are not the same at
all wavelengths.
Consider first the filaments, which are shown separately in the middle panels of Figures 7 9. At solar minimum (Figure 7), only one small filament is detected in He II 304 ,
which does not correspond spatially with any of the SFC filaments. However, the Fe IX / X
171 and Fe XII 195 images yield more filaments than present in the SFC, albeit not necessarily all the same at these two wavelengths. At solar maximum (Figure 8), there are more
filaments detected in He II 304 than at solar minimum, with some corresponding very well
with those in the SFC. Overall, the correspondence between filaments detected in 171 and
227
195 and the SFC is much better than the solar minimum case, with a few minor exceptions.
This is also the case during the declining phase of the solar cycle (Figure 9). Comparison
with the underlying magnetograms shows that the shape of filaments follows the magnetic
reversal line very nicely in many places. This is particularly striking at solar maximum and
the declining phase of solar activity. There are a few exceptions at solar minimum.
For coronal holes, there seems to be a much better correspondence with the NSO/Kitt
Peak coronal holes, than the correspondence between candidate filaments and the SFC filaments, as previously noted. In some cases, there are more coronal holes than NSO/Kitt
Peak ones. All detected holes lie in magnetic regions with a dominant polarity, as seen from
comparison with the magnetograms in the left panels. It is not surprising that the He II
304 coronal hole contours are the closest to those defined in the He I 10830 NSO/Kitt
Peak coronal-hole maps, as originally reported by Harvey and Sheeley (1977). It also seems
that the Fe XII 195 maps yield finer, narrower, and more numerous coronal holes than
the other wavelengths. The best correspondence between the coronal holes detected in all
wavelengths and the NSO/Kitt Peak coronal-hole maps is found during the declining phase
of the solar cycle (see right panels of Figure 9).
Finally, we note that the location of active regions is given in Figures 7 9 because of
their impact on the statistical estimators of the magnetic field when a coronal hole or a
filament is in their vicinity.
228
6. Discussion and Conclusion

The automated approach described here, for the detection of, and the distinction between,
coronal holes and filaments in EUV images of the solar disk, is based on the processing of
the SOHO/EIT coronal images and the identification of distinct magnetic signatures from
the corresponding magnetograms. Although sharing the same apparent visual signatures in
all the EUV wavelengths considered, this study shows that coronal holes and filaments are
characterized by well-defined and distinct magnetic signatures.
Comparison with other databases for coronal holes and filaments, namely the NSO/Kitt
Peak catalog of coronal-hole maps and the Solar Feature Catalog for filaments, shows a good
correspondence between the candidate features detected in this study and those present in
these catalogs. There are however notable discrepancies. In general, more features are detected in this approach than are present in the catalogs, with the exception of examples from
solar minimum. This could likely be due to the low contrast in the EIT images during that
time period. It is not evident, a priori, whether this problem will be resolved in the future
with better-quality data, as expected from the STEREO mission. As originally noted by Harvey and Sheeley (1977), the comparison between He I 10830 and He II 304 images does
not always yield consistent results, and the same holds true for the other EUV wavelengths
considered here. For features lying close to Sun center, when line-of-sight effects are minimal, a one-to-one correspondence between the different data sets is found for both coronal
holes and filaments.
Interestingly, this automated technique reveals features close to active regions that would
not have been detected otherwise. In those cases, the detection can sometimes be ambiguous
as filaments are often found to have large magnetic-field intensity values, making them
indistinguishable from coronal holes.
As is amply evident from the examples presented here, the richness of the EUV data can
be a drawback. It is not clear whether a unique set of criteria can be defined to design a
robust automated detection scheme. Why the process leads to differences among the three
wavelengths is not evident. Nonetheless, there is no metric, or physical reason, to assume
that the He I 10830 line, which has been used as a diagnostic tool for coronal holes for
over three decades, is indeed the most reliable metric. There is no doubt that the variety of
temperatures, densities, and sizes of magnetic structures extending in the corona, which lead
to differences in the EUV images, impact the detection process, in particular the distinction
between open and closed magnetic structure an ambiguity that is further aggravated by
line-of-sight effects.
There are several considerations that might lead to a better and more consistent automation in future studies. One can take advantage of a time sequence for ambiguous cases to see
if a feature (i.e., either a coronal hole or filament channel) maintains its magnetic characteristics as it traverses the solar disk, and as the effect of the line of sight varies. This next step
will lead to the construction of a coronal-hole catalog that will use the time tracking method
described in Aboudarham et al. (2007).
In summary, this scheme differs from other automated coronal-hole detection techniques
(e.g., Henney and Harvey, 2005; de Toma and Arge, 2005) as no criteria for their detection
are identified a priori here. Although robust, the proposed automated approach is by no
means complete, and it should be considered complementary to other published approaches.
One of its strengths is that it is the first to consider the role of filament channels. With
the advent of the STEREO mission, it should become possible to define additional criteria
for the automated detection of coronal holes and filaments, as these features might become
visually distinguishable in higher spatial resolution images.
229
Acknowledgements We would like to thank Dr. Huw Morgan for valuable comments. I. Scholl thanks the
International Space University for supporting her research and the Institute for Astronomy of the University
of Hawaii for supporting her visit in 2006 when this project was initiated. SOHO is a mission of international collaboration between ESA and NASA. NSO/Kitt Peak data used here are produced cooperatively by
NSF/NSO, NASA/GSFC, and NOAA/SEL.
References
Aboudarham, J., Scholl, I., Fuller, N., Fouesneau, M., Galametz, M., Gonon, F., Maire, A., Leroy, Y.: 2007,
Automatic detection and tracking of filaments to fill-in a solar feature database. Ann. Geophys., in press.
Andretta, V., Jones, H.P.: 1997, On the role of the solar corona and transition region in the excitation of the
spectrum of neutral helium. Astrophys. J. 489, 375 394.
Bell, B., Noci, G.: 1976, Intensity of the Fe XV emission line corona, the level of geomagnetic activity, and
the velocity of the solar wind. J. Geophys. Res. 81, 4508 4516.
Bentley, R.D., Csillaghy, A., Scholl, I.: 2004, The European grid of solar observations. In: Quinn, P.J.,
Bridger, A. (eds.) Optimizing Scientific Return for Astronomy through Information Technologies, Proc.
SPIE 5493, 170 177.
Delaboudinire, J.P., Artzner, G.E., Brunaud, J., Gabriel, A.H., Hochedez, J.F., Millier, F., Song, X.Y., Au, B.,
Dere, K.P., Howard, R.A., Kreplin, R., Michels, D.J., Moses, J.D., Defise, J.M., Jamar, C., Rochus, P.,
Chauvineau, J.P., Marioge, J.P., Catura, R.C., Lemen, J.R., Shing, L., Stern, R.A., Gurman, J.B., Neupert, W.M., Maucherat, A., Clette, F., Cugnon, P., van Dessel, E.L.: 1995, EIT: extreme-ultraviolet
imaging telescope for the SOHO mission. Solar Phys. 162, 291 312.
de Toma, G.D., Arge, C.N.: 2005, Multi-wavelength observations of coronal holes. In: Sankarasubramanian,
K., Penn, M., Pevtsov, A. (eds.) Large-scale Structures and Their Role in Solar Activity, CS-346, Publ.
Astron. Soc. Pac., San Francisco, 251 260.
Goldberg, L.: 1939, Transition probabilities for He I. Astrophys. J. 90, 414 428.
Gonzalez, R.C., Woods, R.E.: 2002, Digital Image Processing, Addison-Wesley Longman, Boston.
Harvey, J.W., Sheeley, N.R. Jr.: 1977, A comparison of He II 304 and He I 10830 spectroheliograms.
Solar Phys. 54, 343 351.
Harvey, K.L., Recely, F.: 2002, Polar coronal holes during cycles 22 and 23. Solar Phys. 211, 31 52.
Harvey, K.L., Harvey, J.W., Sheeley, N.R. Jr.: 1982, Magnetic measurements of coronal holes during 1975
1980. Solar Phys. 79, 149 160.
Henney, C.J., Harvey, J.W.: 2005, Automated coronal hole detection using He 1083 nm spectroheliograms
and photospheric magnetograms. In: Sankarasubramanian, K., Penn, M., Pevtsov, A. (eds.) Large-scale
Structures and Their Role in Solar Activity, CS-346, Publ. Astron. Soc. Pac., San Francisco, 261 268.
Krieger, A.S., Timothy, A.F., Roelof, E.C.: 1973, A coronal hole and its identification as the source of a high
velocity solar wind stream. Solar Phys. 29, 505 525.
Liu, Y., Zhao, X., Hoeksema, T.: 2004, Correction of offset in MDI/SOHO magnetograms. Solar Phys. 219,
39 53.
Malanushenko, O.V., Jones, H.P.: 2005, Differentiating coronal holes from the quiet Sun by He 1083 nm
imaging spectroscopy. Solar Phys. 226, 3 16.
Munro, R.H., Withbroe, G.L.: 1972, Properties of a coronal hole derived from extreme-ultraviolet observations. Astrophys. J. 176, 511 520.
Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., Romeny, B.T.H., Zimmerman, J.B.: 1987, Adaptive histogram equalization and its variation. Comput. Vision Graphics Image
Process. 39(3), 355 368.
Reeves, E.M., Parkinson, W.H.: 1970, An atlas of extreme-ultraviolet spectroheliograms from OSO-IV. Astrophys. J. Suppl. Ser. 21, 405 409.
Springer, L., Tarbell, T.D., Title, A., Wolfson, C.J., Zayer, I., MDI Engineering Team: 1995, The solar oscillations investigation Michelson Doppler imager. Solar Phys. 162, 129 188.
Scholl, I.: 2003, Conception, ralisation et utilisation darchives de donnes solaires spatiales. Ph.D. thesis,
Universit Paris 6, France.
Sheeley, N.R. Jr., Harvey, J.W., Feldman, W.C.: 1976, Coronal holes, solar wind streams, and recurrent
geomagnetic disturbances 1973 1976. Solar Phys. 49, 271 278.
Stark, J.A.: 2000, Adaptive image contrast enhancement using generalizations of histogram equalization.
IEEE Trans. Image Process. 9(5), 889 896.
Vaiana, G.S., Zombeck, M., Krieger, A.S., Timothy, A.F.: 1976, ATM observations X-ray results. Astrophys.
Space Sci. 39, 75 101.
Zharkova, V.V., Aboudarham, J., Zharkov, S., Ipson, S.S., Benkhalil, A.K., Fuller, N.: 2005, Solar feature
catalogues in EGSO. Solar Phys. 228, 361 375.

V. Delouille P. Chainais J.-F. Hochedez

DOI: 10.1007/s11207-008-9131-x Springer Science+Business Media B.V. 2008
Abstract Solar telescopes will never be able to resolve the smallest events at their intrinsic
physical scales. Pixel signals recorded by SOHO/(CDS, EIT, SUMER), STEREO/SECCHI/
EUVI, TRACE, SDO/AIA, and even by the future Solar Orbiter EUI/HRI contain an inherent spatial noise since they represent an average of the solar signal present at subpixel
scales. In this paper, we aim at investigating this spatial noise, and hopefully at extracting
information from subpixel scales. Two paths are explored. We first combine a regularity
analysis of a sequence of EIT images with an estimation of the relationship between mean
and standard deviation, and we formulate a scenario for the evolution of the local signal-tonoise ratio (SNR) as the pixel size becomes smaller. Second, we use an elementary forward
modeling to examine the relationship between nanoflare characteristics (such as area, duration, and intensity) and the global mean and standard deviation. We use theoretical distributions of nanoflare parameters as input to the forward model. A fine-grid image is generated
as a random superposition of those pseudo-nanoflares. Coarser resolution images (simulating images acquired by a telescope) are obtained by rebinning and are used to compute the
mean and standard deviation to be analyzed. Our results show that the local SNR decays
more slowly in regions exhibiting irregularities than in smooth regions.
Keywords Solar corona Signal-to-noise ratio Subpixel scale Aliasing
V. Delouille () J.-F. Hochedez

SIDC Royal Observatory of Belgium, Brussels, Belgium
e-mail: v.delouille@oma.be
J.-F. Hochedez
e-mail: hochedez@oma.be
P. Chainais
Universit Blaise Pascal Clermont II, Clermont-Ferrand, France
e-mail: pchainai@isima.fr
DOI: 10.1007/978-0-387-98154-3_17
231
232
V. Delouille et al.
1. Introduction
Future missions such as Solar Orbiter (SO) aim at studying the Sun from closer than ever.
It is expected that the High Resolution Imagers (HRI) onboard SO will have a pixel size
of 0.5 arcsec and a cadence better than one second (Hochedez et al., 2001). When the SO
mission is at perihelion (i.e., at a distance of 0.2 AU from the Sun), one pixel of HRI will
represent approximately (80 km2 ) on the Sun. Note that 80 km corresponds to a pixel size
of 0.1 arcsec for an instrument located at 1 AU.
Under these conditions it is unclear whether there will be enough photons per pixel to
illuminate the detector and whether the signal-to-noise ratio (SNR) will be sufficient. It is
thus necessary to quantify the expected level of SNR, and more generally to provide tools
for extracting information from subpixel scales. We propose two ways to get new insights
into this issue.
First, the analysis of a high-cadence data set recorded by the EIT instrument (Delaboudinire et al., 1995) on 4 October 1996 allows us to describe how the local SNR evolves as
the scale of observation (or pixel size) becomes smaller. We show that the decrease in SNR
as the resolution gets finer is not the same when the spatial fluctuations of the radiance are
taken into account or when a uniform radiance is considered. Indeed, the photon emission
process can be modeled by a Poisson distribution Po(L), where L represents the radiance.
If the emission of the solar corona were uniform over the entire Sun, we would have an
homogeneous Poisson process. Denoting by a the mean intensity recorded and by a the
corresponding standard deviation (STD), we have that the SNR would simply linearly decrease as a function of the scale (a) of observation: a /a a. Fortunately, different parts
of the corona have different levels of emission; that is, the photon flux that hits the detector
at a particular location (x) during an exposure time (T ) follows an inhomogeneous Poisson
distribution Po[L(x, T )], where the radiance L is space and time dependent. Let us consider
that L(x, T0 ) = f (x)L0 for some given exposure time (T0 ), with L0 a constant and f (x) a
function describing the inhomogeneities in a solar coronal image. Our purpose is to study
how the local variations of f (x) influence the local SNR.
In the second part of the paper, we propose a basic forward modeling technique that
takes as input different distributions of flare characteristics proposed in the literature
(Crosby et al., 1993; Isliker et al., 2001; Krucker and Benz, 1998; Paczuski et al., 2005;
Vlahos et al., 1995). We emulate the instrument response through rebinning. The forward
model gives as output the average (a ) and standard deviation (a ), computed either over
b
space or over time. We confirm that the relationship a = b0 a1 prevails, similarly to what
is observed in real data, and we investigate how the coefficient b1 is influenced by pseudonanoflare distribution parameter values.
This paper is organized as follows: Section 2 recalls the various sources of noise that
impact the quality of EIT images and explains how to access subpixel information by using a high-cadence EIT data set. Section 3 introduces our forward modeling approach and
presents the results of the corresponding simulation study. Finally, Section 4 places our work
in perspective with other studies on noise in solar images and gives prospects for future research.
2. EIT Data Set Analysis

The aim of this section is to estimate subpixel variability within a high-cadence EIT sequence. We begin by recalling the main sources of error present in EIT images.
233
2.1. Sources of Noise in EIT Images

The incident EUV flux of interest on EIT is converted in digital numbers (DN) through a
series of steps. At each step, some noise may be introduced. In brief, the beam of photons
impinges the optical system where the optical point-spread-function (PSF) acts as a blurring
operator. Simultaneously, a spectral selection is performed on the signal before it reaches
the CCD detector. The latter has a heterogeneous response across its surface. Finally, the
camera electronics convert photon counts into DN; this conversion adds the read-out noise.
2.1.1. Poisson Noise
The photon emission processes in the solar atmosphere are random and incoherent in nature. They are usually modeled by a Poisson process so that the number of incident photons
between t and t + T obeys a Poisson law Po[L(x, t, T )] of mean L(x, t, T ), where the radiance L is space and time dependent. When L is sufficiently large (L > 1000), the Poisson
distribution may be approximated by a Gaussian. Lower photon counts give higher variance
with respect to the mean and hence smaller SNR. Let N (x) denote the number of photons
that hit the detector at location x, time t0 , and during an exposure time T0 . N (x) is modeled by a Poisson random variable distributed as Po[L(x, t0 , T0 )]. The recorded signal S(x)
measured in DN at location x is equal to (Defise, 1999; Janesick et al., 1985)
S(x) = qN (x),
q = QE i
1
,
G
(1)
where QE is the inner quantum efficiency (in number of detected photons per incident photon), i is the ideal quantum yield (in electrons per detected photon), and G is the electronic
gain (in electrons per DN).
Since N (x) can be viewed as a realization from a random variable, so is S(x). Its mathematical expectation denoted by E[S(x)] is equal to (x) E[S(x)] = qL(x). The variance
of S(x) is given by

(x)2 Var S(x) = q 2 Var N (x) = q 2 L(x) = q(x).
(2)
In summary, if one assumes that the detector does not introduce any other fluctuations besides the Poisson randomness, the local mean and standard deviation are related by
(x) = q 1/2 (x)1/2 .
(3)
2.1.2. Blurring from the PSF

The EIT telescope has a pixel size of 2.6 arcsec (1.8 Mm at the center of the Sun). The total
PSF of the instrument is the combination of the optical PSF and the pixel shape. The optical PSF of EIT has a full-width-at-half-maximum (FWHM) of less than one pixel (Defise,
1999), and by convolution with the square pixel shape, one can deduce the angular resolution of the instrument on the solar disk. This convolution is close to the nominal 2.6 arcsec
angular size of the pixel; see Delaboudinire et al. (1995). The signal recorded by one pixel
thus corresponds to an average of the photon counts on a well-defined area. Two distinct
structures separated by a distance smaller than the instrument PSF width will be mixed.
With EIT, the smallest detectable wavelength for a periodic spatial feature is 3.6 Mm. This
is the origin of the discussed spatial noise. DeForest (2007) made a precise study of the
averaging effect on loop structures observed with EIT and TRACE. The goal of this section is to use a high-cadence data set to try and recover some information about the spatial
modulation of the signal that has been averaged by the PSF.
234
V. Delouille et al.
2.1.3. Flat Field

The CCD detector exhibits inhomogeneities in the response. A map of these inhomogeneities, called the flat field (FF), has been estimated, and the eit_prep procedure of the
Solar Software library ssw: http://www.msal.com/solarsoft/ corrects for this nonuniformity.
If, however, this correction is biased, the true signal S will be recorded as S = S + S ,
where is the relative bias in the FF correction estimate. This bias = (x) fluctuates in
space, so it can be considered as a random noise. Because of this bias, the recorded signal
[S(x)] corresponding to an ideal flat, uniform source [S (x) = S0 ] will exhibit a linear relationship between its spatial mean and standard deviation. This can be shown as follows. Let
us consider a small neighborhood [N (x)] of some point (x). The spatial variance of S over
N (x) is
VarN (x) (S) = S02 VarN (x) ().
(4)
Hence the relationship between spatial mean () and standard deviation ( ) of the signal
(S) over N (x) is equal to
= kFF (x) ,
(5)
where kFF (x) = VarN (x) () is called here the flat-field noise factor, which varies from one
pixel to another. This reasoning generalizes to nonuniform S = S0 as long as the intensity
in the neighborhood stays fairly uniform. For highly varying photon counts, the influence of
becomes negligible with respect to solar variability.
2.1.4. Read-out Noise
The electronics that converts photon counts to digital number generate a read-out noise of
Gaussian nature. Its standard deviation can be estimated but its influence becomes important
only for small values of the radiance. Finally, the thermal noise (variability of the dark
current) originates from the fluctuations of charges generated electronically. However, at
EIT operational temperature, it is negligible.
2.2. Method
We consider the Joint Observing Program (JOP-020) recorded by EIT on 4 October 1996 in
the 19.5-nm bandpasses. This JOP was dedicated to the observation of nanoflares. It consists
of 89 images of size 128 192 pixels with one-minute cadence. One pixel corresponds to a
surface of (1800 km)2 on the Sun. We consider level-1 images, preprocessed and calibrated
through the eit_prep procedure of the ssw library.
We make use of this high-cadence data set to estimate subpixel variability (sometimes
called an aliasing effect). To this end, we use time-average estimates of the mean and
standard deviation computed at each pixel. Indeed, proceeding on a pixel-by-pixel basis
ensures that the subsequent results are not contaminated by the flat-field noise. Moreover,
the one-minute cadence gives access to subpixel information that is masked by the averaging
effect of the PSF in individual images: The solar rotation induces a displacement of one pixel
in the horizontal direction every 15 minutes in EIT images; there is thus a displacement of
about 1/15 of a pixel per minute. Hence the mean and STD computed on temporal windows
of 15 minutes give us insight into the subpixel spatial statistics, as well as into the temporal
variability.
235
The JOP sequence shows bright points together with more uniform areas (see Figure 1).
These two types of structures are likely to behave differently at subpixel scales: Bright points
will typically evolve more quickly than uniform regions showing few signs of activity. Hence
we consider separately these two types of regions, as explained in Section 2.3. Section 2.4
proposes a regularity analysis of the data set and shows how it allows us to give scenarios
for the evolution of the local SNR and meanSTD relationship across scales.
2.3. Regularity Analysis Using Pointwise Hlder Exponents
To distinguish between regions of high and low regularity in a signal (g), it is customary
to compute the pointwise Hlder exponent (h) at each point (x0 ). This exponent precisely
characterizes the local regularity: a function g is Hlder regular with exponent h(x0 ) at some
point x0 if |g(x) g(x0 )| |x x0 |h(x0 ) for x in the neighborhood of x0 .1 If the function is
locally C in a neighborhood of x0 , then h(x0 ) = +. If the function is locally similar to
a -function, then h(x0 ) = 0. If h(x0 ) 1, the function is at least first-order differentiable.
In broad terms, the smaller h(x0 ) is, the more singular the image around x0 will be. The
Appendix provides a precise definition of h(x0 ).
In Delouille et al. (2005), we used a local scale measure computed with the Mexican Hat
wavelet transform to estimate the local Hlder exponent. In the present paper, we choose
to compute each h(x0 ) through a series of bivariate regressions on spatial windows of increasing sizes. The signal g considered is not the gray-level pixel value of the image, but
rather the cardinal of the largest subset of pixels having the same gray level. With this implementation, values for the Hlder exponents range between zero (where the signal behaves
like a -function) and two (where the signal is twice differentiable). The FracLab software
(http://www2.irccyn.ec-nantes.fr/FracLab/; Vhel and Legrand, 2004) was used to compute
the values of h.
Figure 1 shows the image recorded by EIT on 4 October 1996 at 07:39:10 in the 19.5-nm
bandpass, together with its map of pointwise Hlder exponents. In our analysis to follow,
we discard pixels with Hlder exponent less than 0.2. Therefore cosmic-ray hits, which
typically behave like discontinuous -functions, are not taken into account.
Figure 1 (a) Image from JOP nanoflares data set taken on 4 October 1996 at 07:39:10 (in logarithmic scale).
(b) Map of local Hlder exponents computed from the corresponding linear image.
1 Here denotes an asymptotic proportionality as x x .
0
236
V. Delouille et al.
2.4. Signal-to-Noise Ratio and Relationship between Mean and Standard Deviation
We distinguish two sets of pixels with different regularities: one set of pixels where the
image is somewhat smooth and the other where it is more singular. The first set contains all
pixels for which the Hlder exponent is greater than 1.6 (smooth regions), and the second
set contain all pixels for which the Hlder exponent is smaller than 0.6 (singular regions)
but larger than 0.2 (to avoid cosmic rays).
We are interested in the fluctuations of the intensity measured by a pixel of size a located at x0 . For an homogenous Poisson process of intensity L observed at scale a, it is a
classical result (Snyder and Miller, 1991) that the mean intensity measured in one pixel is
1/2
a = La 2 and the variance is a2 = La 2 and
so a = a . In this case, one usually defines
a signal-to-noise-ratio by SNR = a /a = L a a. By analogy, we focus on the relationship between the temporal mean [a (x0 )] and the standard deviation [a (x0 )] of the function
f at some given position (x0 ) observed at resolution a. The purpose of the present analysis is to get insight into the evolution of the SNRa (x0 ) = a (x0 )/a (x0 ) as the resolution a
changes. As already stated, we estimate a and a for each pixel separately by averaging
the observed pixel values over nonoverlapping temporal windows of 15 minutes. Since the
rotation induces a displacement of one pixel in 15 minutes, our statistics actually cover an
area equivalent to two pixels. Hence this technique allows us to see what happens at scales
below two EIT pixels, that is, at subresolution scale.
To extract the predominant relationship between the observed a and a , we compute
the two-dimensional (2D) histogram of (log a , log a ). In Figure 2, this histogram is represented in gray levels: Brighter values at coordinate (log 0 , log 0 ) indicate more pixels
in the original data set for which a = 0 and a = 0 . In other words, Figure 2 provides
an estimate of the 2D probability density function of the vector (log a , log a ). In Figure 2a the 2D density estimation of (log , log ) corresponds to pixels in singular regions
[h(x0 ) 0.6], whereas Figure 2b shows the 2D density for pixels belonging to smooth regions [h(x0 ) 1.6]. We computed these 2D densities as a succession of 1D density function
estimations of the standard deviation (represented in the y-axis). Each 1D density estimation
is carried out on a slice of width equal to log() = 0.016 (in DN s1 ). For each slice,
we compute the mode (i.e., the maximum) of the 1D density function of the standard deviation. Next, we estimate the line (log a = k + b1 log a ) that fits these modes. The slope
Figure 2 Density estimation for (log , log ) for (a) pixels having a Hlder exponent less than 0.6 (representing singular regions) and (b) pixels having a Hlder exponent larger than 1.6 (representing smooth
regions). The value of the slope is equal to 1.3 in (a) and to 0.7 in (b).
237
(b1 ) obtained by the least-squares method is equal to 1.3 in the case of singular regions versus 0.7 for smooth regions. Note that it is not possible to disentangle spatial from temporal
variability in the interpretation of these values. Although we keep in mind that both types of
variability might be present, for the sake of clarity we omit in the notation in the following
the temporal dependence of the intensity f .
The standard deviation, or intensity fluctuations, at scale a and location x0 can be approximated as

(6)
a (x0 ) f (x0 + au) f (x0 ),
where u is any unitary vector. As a direct consequence of the definition of the Hlder exponent h(x0 ), we then have a (x0 ) a h(x0 ) . The local Hlder regularity helps us in providing
a scenario for the evolution of the local SNR as follows.
In smooth regions, the image is rather regular and neighboring pixels have similar values,
close to a constant f (x0 ) times the pixel area. Therefore, the mean value observed at scale a
around x0 is roughly proportional to the surface of the pixel, so a (x0 ) a 2 or equivalently
a a (x0 )1/2 . As a consequence, for h(x0 ) close to 1.6 [recall that 1.6 h(x0 ) < 2] we
obtain a relation of the type
a (x0 ) a 1.6 a (x0 )1.6/2 a (x0 )0.8 .
(7)
The 0.8 exponent is close to the observed a 0.7

a and is greater than the 1/2 exponent
that would be observed for a homogeneous Poisson noise. Notice that the Hlder exponent
was computed on the original image by using scales above the pixel size, whereas the slope
b1 is representative of the evolution at subresolution scale. Under the approximation (6), the
value of the Hlder exponent can be related to the slope b1 ; this suggests that there is some
consistency between the evolution at super- and subpixel resolution.
Considering now the SNR, one would expect that
SNRa (x0 ) =
a
a 21.6 a 0.4
a
(8)
in such regions. As a consequence, the SNR defined above would go to zero as a 0.4 (as
a 0). This is yet slower than the usual property of Poisson noise for which SNRa a.
In singular regions for which h(x0 ) < 0.6, the singularity corresponds to either a local
minimum or a local maximum. We observe that the value of a is about twice as large
(log10 2 0.3) in singular regions (h < 0.6) as in smooth regions (h > 1.6) (see the values
of log a in Figure 2a, which range from 2 to 2.5, compared to Figure 2b, where they range
from 1.8 to 2.3 for smooth regions). Thus these singular regions correspond mostly to local
maxima.
As a consequence, the intensity in regions where h(x0 ) < 0.6 is peaked: The extreme case
would be that of a -function, where the intensity would be infinitely concentrated at some
unique position. Therefore, the mean (a ) is dominated by some extreme value associated
with a small region of emission. Then one would expect that a a (x0 ) with (x0 ) < 2
(if the image is locally a -function, a is constant, and = 0); equivalently, one then has
1/
a a . Hence for h(x0 ) 0.6 we get
a a 0.6 0.6/
a1+ ,
a
(9)
where > 0 as soon as < 0.6. This argument is consistent with the observed value of
b1 = 1.3 = 1 + 0.3 [which would typically correspond to h(x0 ) = 0.6 and = 0.46].
238
V. Delouille et al.
For the SNR, one would expect for such singular regions that
SNRa =
a
a 0.6 .
a
(10)
Therefore, in the particular case when < 0.6, one would even get that SNRa as
a 0. Locally, at places where there are -function structures, the SNR would increase as
the scale becomes smaller. In practice, the limit a 0 needs to be considered with caution
in the present approach. Our argument remains valid only when the scale of observation
remains much greater than the typical scale of the singular object we focus on. As soon as
the resolution becomes sufficient to resolve an elementary object, it will appear as smooth
rather than singular. The singularity or smoothness of a region is a notion relative to the
scale of observation. The argument here simply tells us that one may expect some gain, or at
least a slow decrease, in SNR as the resolution gets finer around regions that appear singular
at the present scale of observation; for example, if h(x0 ) = 0.6 and < 1, the SNR will
decrease following a power law with exponent smaller than 0.4 as a 0.
This qualitative study suggests that modeling images of the Sun by some homogeneous
Poisson process leads to a pessimistic prediction of the evolution of the local SNR when
decreasing the scale of observation, that is, when observing at a finer resolution. The use
of a uniform intensity for the Poisson process neglects the spatial fluctuations of the emission coming from the Sun. Taking into account the local regularity of the image by using
the Hlder exponent allows us to make more precise predictions about this evolution: The
resulting prediction is less pessimistic since we get SNRa a with < 1 in place of the
usual SNRa a ( = 1) predicted by the Poisson-noise model. The more singular the region of interest is, the slower will be the decrease in SNR as the resolution gets finer. This is
important information for assessing the quality of forthcoming high-resolution observations,
since these might be threatened by the SNR decreasing too fast as the resolution gets higher.
In particular, in the context of the SO mission, radiometric models assessing the number
of photons available per pixel at HRI resolution should take into account the spatial and
temporal fluctuations of the radiance.
3. Forward Modeling Approach

In this section, we model the solar corona as a superposition in space and time of a large
number of flarelike events, characterized by their surface, duration, intensity, and localization. Having fixed a probability distribution function for these quantities, we generate a time
series of images. We then compute the mean () and standard deviation ( ), either in space
(over the whole image) or in time (over the whole sequence). In both cases, we estimate the
parameters (b0 , b1 ) of the model = b0 b1 . We repeat this operation for several values of
the power-law index characterizing the distributions of flarelike events. We then relate the
value of b1 to these indices. This allows us to identify which flare characteristics (area, duration, and/or intensity) most influence the relationship between mean and standard deviation.
The spatial and temporal resolution of the fine-grid data sequence corresponds to the
expected resolution of HRI at perihelion, namely a pixel size of (80 km) and a cadence of
one second. This is approximately five times better spatial resolution than what the TRACE
instrument provides and a 25 times enhancement as compared to the EIT telescope.
239
3.1. Generation of the Data Sequence

We begin by describing how to generate a flarelike event. To handle the simulation within
a reasonable time scale, we simplify the physics and especially the behavior of the cooling
phase: An event is a cube with a constant intensity within a given spatial and temporal
window and a zero value outside. In future studies, we plan on using more sophisticated
models for the time evolution of an event.
We need to choose a probability distribution for the area, duration, and intensity of
an event. Past measurements of bright events point at scale invariance of these three
characteristics; their distribution is therefore usually modeled by a power law (see, e.g.,
Crosby et al., 1993). In our simulation, the area (A), duration (D), and intensity (I ) characterizing an event follow such a power-law distribution:
p(A) A ,
p(D) D ,
p(I ) I ,
(11)
where p(x) denotes the probability density function of the quantity x.

Different values for (, , ) are proposed in the literature. Values for the power-law
index of the peak flux ( ), or emission measure, differ the most: Aschwanden and Parnell
(2002) observe values ranging between 1.75 and 1.94 for the peak flux in TRACE and
Yohkoh/SXT, whereas Crosby et al. (1993) provide a value of 1.59 for the peak HXR
flux power slope. However, Parkers hypothesis conveying the idea that the solar corona
could be heated by a multitude of nanoflares needs a slope of at most 2 for the thermal
energy content. Krucker and Benz (1998) give a value of ranging between 2.6 and
2.3; avalanche models produce a slope of 2.4 (Viticchi et al., 2006) and even 3.5 for
the smallest flare (Vlahos et al., 1995). Studies exploring the values for the power law of
the area, , include that of Aschwanden and Parnell (2002), who derived values of 2.45
and 1.86. Finally, values for the duration index range from 1.95 (Crosby et al., 1993)
to 8 (Paczuski et al., 2005) and even 11 (Vlahos et al., 1995) for small flares. To be
consistent with these different results, we explore all combinations of values for , ,
given in Table 1.
We now specify the range of values for A, D, and I , as well as their relationships. Berghmans et al. (1998) used a one-minute cadence data set recorded by EIT on 28 December
2006 and found that the area and duration of a brightening are almost linearly correlated in
a log log plot. However, the typical duration of an 80-km event is unclear. Lin et al. (1984)
reported X-ray microflares lasting from a few seconds to several tens of seconds and having
a power-law energy spectra. Golub et al. (1989) computed ionization times of the order of
two to four seconds for the Fe IX, Fe X, and Fe XI ions, whereas the cooling time was of the
order of several minutes. We decided to choose intervals of values for the area and duration
that are compatible with a straightforward extrapolation of the results in Berghmans et al.
(1998). Areas range between 0.0064 and 100 Mm2 (covering from 1 to 1252 pixels in the fine
grid), and the duration of an event may last between one second and five minutes. This simplified setting is easily implemented, but in a further study, we plan to consider other types
of extrapolations of the distribution of events at smaller scales and to take into account the
exponential decay of the cooling phase. Finally, the intensity (in arbitrary units) is allowed
to span five orders of magnitude, in agreement with the results of Aschwanden and Parnell
(2002) on the relationship between area and intensity. Table 1 summarizes these ranges.
The high-resolution data sequences contain 300 temporal frames, each with a spatial
extent of 500 500 pixels. They are simulated as the superposition of a large number (N )
of events, so that the entire space is filled with events. An event is modeled as a cube in
240
V. Delouille et al.
Table 1 Power-law indices and ranges for the area, duration, and intensity considered in the simulation.
Power-law index
Range
Area
{1.5, 1.6, 1.9, 2.9}
[0.0064 100 Mm2 ] or [1 1252 ] pixels
Duration
{2.1; 2.5; 4, 8}
[1 300 s] or [1 300] frames
Intensity
{1.6; 1.8; 1.95; 2.3; 3.5}
[1 105 ] (arbitrary units)
Table 2 Coarsening factors operated in the simulation.

Data cube
Spatial dimension (in pixels)
Time dimension
: size generated/size rebin
(number of frames)
Generated
500 500
300
Rebin 1
100 100
1500 ( TRACE)
Rebin 2
20 20
37500 ( EIT, 1-min cadence)
the spatio-temporal dimension: It has a constant intensity within this data cube and a zero
intensity outside. The data sequences are generated as follows:
1. Choose a particular combination of values for , , and .
2. Using the prescribed power-law distributions (index value and range), generate N values
for A, D, and I .
3. Rank the N values generated for A, D, and I by increasing order. A flarelike event is
characterized by the triple (A(k) , D(k) , I(k) ), k = 1, . . . , N , where X(i) denotes the ith
element of the order statistics of X (i.e., X(1) X(2) ). With this ordering, an event
with small area will have also a small duration and intensity.
4. Generate the localization in space and time of the N events as independent samples from
a uniform distribution.
5. Obtain the data sequence by superpositioning the N events.
Note that, in the third step, we do not impose a deterministic relationship among A, D,
and I . Indeed, there is a large dispersion around the linear fit between area and duration
of observed brightenings (cf. Berghmans et al., 1998). This means that there is a range of
possible durations for a given event size.
We consider three values for the number of events generated: N = 105 , 107 , and 5 107 .
For each of these three values, we generate two data sequences. In total, for a given set of
parameters (, , ), we thus generate six data sets.
From the high-resolution data cube, we derive two coarser resolution data sets that reflect,
respectively, TRACE and EIT resolution: A rebin by a factor five in space and 60 in time is
close to the specifications of TRACE, whereas a rebin by a factor 25 in space emulates EITs
spatial resolution. These two rebinning factors () are displayed in Table 2. Recall that the
spatial rebin emulates the PSF operator, which is assumed here to be a step function. The
temporal rebinning simulates the integration over time during an image acquisition.
A pixel in a rebinned image is termed a macro-pixel. Because the localization of each
event is generated randomly, there will be a different number of events in each macro-pixel.
This introduces another source of variability: When averaging to obtain the rebinned image, the sum is performed over a random number of events in each macro-pixel. Figure 3a
shows one realization with N = 105 events and the smallest power-law indices in absolute
value ( = 1.5, = 2.1, = 1.6). This setting generates values for the area, duration,
241
Figure 3 Example of one realization, with N = 105 , = 1.5, = 2.1, and = 1.6. (a) A high resolution image; (b) the corresponding image rebinned by a factor of 25 25 in space and 60 in time. The values
of the intensity are in arbitrary units.
and intensity that cover a large range. Structured regions appear as a consequence of the
superposition of events of diverse sizes. These fine structures are largely smoothed out in
the rebinned version displayed in Figure 3b.
3.2. Phenomenological Model
For each coarse-resolution data set, the following quantities are considered: S and S are
an average (over time) of the mean and STD computed over space on each frame of the
sequence; T and T are an average (over space) of the temporal mean and STD computed
for each pixel. We summarize the six values of the same combination of (, , , ) by
estimating the slope and intercept of a linear regression between log( ) and log():
log S = b0,S + b1,S log S ,
(12)
log T = b0,T + b1,T log T .
(13)
All the linear regressions done to estimate (b0,S , b1,S ) and (b0,T , b1,T ) have a R 2 goodness of
fit larger than 0.96.2 We may thus consider that the model = b0 b1 is valid.
We model the slopes b1,S and b1,T as a function of , , , and to see which parameters
have the most influence on these slopes. We do not analyze the behavior of the intercept b0
since this quantity is not based on any physical model and cannot be physically related to
observations. We consider a simple linear model for b1,T and b1,S :
b1 = c0 + c1 + c2 + c3 + c4 + .
(14)
Variables are standardized before entering the model: They are centered and divided
by their half-range. This facilitates the interpretation of coefficient values c: The constant
parameter c0 gives the response value at the center of the hypercube (i.e., for a mean value
of , , , and ). The coefficient before a variable X in the model indicates an average
increase of the response when X increases from the center of the domain to its maximum.
2 The R 2 statistic, R 2 [0, 1], is the ratio between the variance explained by the model and the total variation
the observations. R 2 = 1 means that the model fits the data perfectly (Chatterjee and Hadi, 1986).
242
V. Delouille et al.
Table 3 Coefficient values and standard deviations for the linear models explaining the responses b1,S and
b1,T , respectively. NS stands for nonsignificant: A star in this column indicates a parameter that can be
omitted in the model at the 0.05 level. The slope b1,S is mostly driven by the power-law index for the area,
whereas the duration influences b1,T the most.
Variable
b1,S
STD(b1,S )
NS
b1,T
STD(b1,T )
CONST
0.756
0.0055
0.709
0.0100
0.156
0.0070
0.056
0.0126
0.172
0.0126
NS
0.022
0.0070
0.001
0.0078
0.016
0.0140
0.008
0.0055
0.012
0.0100
R 2 statistics
0.57
0.91
3.3. Results of the Simulation Study

Table 3 shows the estimation of the regression coefficients ci in Equation (14) together with
an estimate of their uncertainty. We entered in the model the values = {15, 375} for the
rebin factor.
A linear model for the slope b1,S seems satisfactory since R 2 = 0.77. The constant parameter represents the value of b1,S for a mean value of , , , and among the range of
values considered. This constant is equal to 0.75 and is thus above the 0.5 value that we
would obtain in case of a homogeneous Poisson process. The variable that influences b1,S
the most is the power-law index for the area; the duration has a smaller influence, whereas
both the power-law index for the intensity ( ) and the rebin parameter () are statistically
nonsignificant. With a value of = 3 (and all other parameters kept at their mid-values),
mainly small events are generated. The slope b1,S then decreases down to 0.6, close to the
situation of a homogeneous Poisson process. When grows to = 1.5, larger events are
generated as well, and the superposition of large and small events produces a more inhomogeneous process: The value of b1,S then increases up to 0.9.
If one now considers the slope b1,T , the results of Table 3 indicate that the duration
mainly influences the evolution of the mean and STD computed over time. In this case, the
goodness of fit is relatively low (R 2 = 0.57). Figure 4 shows how b1,T and b1,S evolve as
a function of and , with kept fixed ( = 1.6), and for a rebinned data set of size
100 100 5. A quadratic model seems more appropriate for the slope b1,T . If we fit a
quadratic model of the type
b1,T = c0 + c1 + c2 + c3 + c4 + c11 2 + c22 2 + c33 2 +
(15)
we obtain an acceptable R 2 = 0.81. Within the quadratic terms, only the coefficient for 2
is significant. The parameter influences b1,T similarly to how influences b1,S : When
is equal to 2.1, the slope b1,T increases to 0.87; when = 8 the slope decreases to 0.63
(with all other parameters kept fixed at their mid-values). The parameter also influences
b1,T , but to a lesser extent. The intensity and rebin factor are again statistically nonsignificant.
In summary, we observe that i) the relationship between spatial mean and standard deviation is mostly affected by the distribution of areas of the events, ii) temporal mean and
standard deviation are driven mostly by the temporal dynamics, iii) the intensity distribution
does not seem to play a significant role, iv) we observe no strong effect of the coarsening
factor.
243
Figure 4 Variation of the slopes b1,S (a) and b1,T (b) as a function of and . The value for is fixed
and equal to 1.6. Here the rebin factor was equal to = 1500 (i.e., the coarsened data set is of size
100 100 5).

We proposed a way to extract subpixel information in a one-minute-cadence EIT data set,
and we derived a scenario for the variation of the SNR as the scale of observation becomes
smaller. Taking into account the variability in the emission process, we showed that the local
SNR decays much more slowly in singular regions (where the variability is larger) than in
smooth regions (characterized by a more uniform emission). We observed in both cases that
the slope between standard deviation and mean is above 0.5 and that the SNR degrades more
slowly than in the case of a uniform Poisson process.
Next, we investigated how the mean variance relationship evolves with different theoretical distribution of nanoflare parameters. A small index for the power-law distribution of
the area favors small events and generates a process close to a homogeneous Poisson distribution. When a large disparity of areas is allowed in the simulation of events (by choosing
a large value for the power-law index), more inhomogeneous structures appear. These are
averaged at coarser resolution and create spatial noise. The slope b1,S in a graph representing the spatial standard deviation against the mean has in this case a value close to one. In
a parallel way, when the distribution (D) generates a large range of durations, the temporal
mean STD relationship exhibits a slope larger than what is observed for a homogeneous
Poisson process.
The determination of solar variability (such as, e.g., microflaring or the determination of
coronal loop width) at the limit of instrument measure often requires a careful analysis to
separate noise components from true solar variability. We now relate our work to several
other methods presented in the literature.
Aschwanden et al. (2000) evaluate the level of different sources of instrumental noise
in the TRACE instrument: read-out, digitization, compression, and dark-current subtraction
noise. In addition to these noises, whose levels are independent of the flux, they estimate
two other noise components that depend on the flux level: the Poisson noise and the error
coming from the removal of cosmic-ray hits. However, they did not apply a flat-field correction to pixels, and hence their estimates are not subject to the flat-field noise described in
Section 2.1.
Our data analysis did not assume a particular photon-noise distribution; it was meant to
be as generic as possible. Katsukawa and Tsuneta (2001) took another approach in their
study of the time profile of X-ray intensities using Yohkoh/SXT. They first derive for each
244
V. Delouille et al.
pixel its time profile and mean background intensity (I0 ), and they used this value to estimate
the standard deviation of the photon noise (P ). Next, they estimate the standard deviation
(I ) of the time-series profile, assuming the core component has a Gaussian distribution.
The ratio I /P increases with the mean profile I0 , and they estimate the relation between
the two quantities. They observed that the fluctuations in darker regions (I0 100 DN)
are almost entirely due to photon noise, whereas there are significant fluctuations of solar
origin in bright regions (I0 100 DN). Although the methodology is different, the results
in Katsukawa and Tsuneta (2001) parallel our study: Darker regions look typically more
uniform, and we observe there a behavior closer to a homogeneous Poisson process than in
bright (more irregular) areas.
Finally, DeForest (2007) investigates the effect of random noise and telescope PSF on
compact linear structures featuring coronal loops. A forward model of the TRACE PSF
indicates that structures with apparent size less than two pixels wide cannot be distinguished
visually from structures of zero width. He also studied some particular loops observed by
TRACE. Assuming a loop is composed of a set of faint threads, he found an estimate for
the size of elementary structures in the lower corona. Similarly, Figure 3 shows how fine
structures disappear when observed at lower resolution.
In a forthcoming paper, we will analyze images of the quiet corona using multifractal
tools. It is also possible to use multifractal processes to synthesize images similar to the
quiet-Sun corona at higher resolution than what current telescopes offer. This allows us
to make more precise predictions about the SNR that would be available at a given high
resolution.
Acknowledgements The authors would like to thank the anonymous reviewer for valuable comments and
suggestions. Funding of V.D. and J.-F.H. by the Belgian Federal Science Policy Office (BELSPO) through the
ESA/PRODEX program is hereby appreciatively acknowledged. V.D. thanks the Universit Blaise Pascal of
Clermont-Ferrand for the one-month stay during which this work was initiated. This work has been supported
by a France Belgium grant Tournesol (Hubert Curien grant).
Appendix: Pointwise Hlder Exponent

The Hlder exponent at x0 , denoted h(x0 ), provides a way to quantify the strength of a
singularity of a function g at the point x0 . It is defined in a rigorous way as the largest
exponent such that there exists a polynomial of degree n h(x0 ) and a constant C > 0 with
the property that for any x in a neighborhood of x0 the following inequality is verified:

g(x) P (x x0 ) C|x x0 |h(x0 ) .
(A1)
When g is n-times differentiable at x0 , the polynomial P (x x0 ) is simply the Taylor expansion polynomial of g(x) at x0 ; in this case, h(x0 ) > n. If h(x0 ) < 1 the polynomial P (x x0 )
simplifies to g(x0 ). A well-known example is given by the function g(x) = a + b|x x0 | ,
whose Hlder exponent at x0 is given by (when is not an even integer). In general,
the higher h is, the more regular is the function g. Conversely, the smaller h is, the more
singular is the function g.
References
Aschwanden, M.J., Parnell, C.E.: 2002, Nanoflare statistics from first principles: Fractal geometry and temperature synthesis. Astrophys. J. 572, 1048 1071.
245
Aschwanden, M.J., Nightingale, R.W., Tarbell, T.D., Wolfson, C.J.: 2000, Time variability of the quiet sun
observed with TRACE. I. Instrumental effects, event detection, and discrimination of extreme-ultraviolet
microflares. Astrophys. J. 535, 1027 1046.
Berghmans, D., Clette, F., Moses, D.: 1998, Quiet Sun EUV transient brightenings and turbulence. A
panoramic view by EIT on board SOHO. Astron. Astrophys. 336, 1039 1055.
Chatterjee, S., Hadi, A.S.: 1986, Influential observations, high leverage points, and outliers in linear regression. Stat. Sci. 1, 379 416.
Crosby, N.B., Aschwanden, M.J., Dennis, B.R.: 1993, Frequency distributions and correlations of solar X-ray
flare parameters. Solar Phys. 143, 275 299.
Defise, J.M.: 1999, Analyse des performances instrumentales du tlscope spatial EIT. Ph.D. thesis, Universit de Lige.
DeForest, C.E.: 2007, On the size of structures in the Solar corona. Astrophys. J. 661, 532 542.
Delaboudinire, J.P., Artzner, G.E., Brunaud, J., Gabriel, A.H., Hochedez, J.F., Millier, F., Song, X.Y., Au, B.,
Dere, K.P., Howard, R.A., Kreplin, R., Michels, D.J., Moses, J.D., Defise, J.M., Jamar, C., Rochus, P.,
Chauvineau, J.P., Marioge, J.P., Catura, R.C., Lemen, J.R., Shing, L., Stern, R.A., Gurman, J.B., Neupert, W.M., Maucherat, A., Clette, F., Cugnon, P., van Dessel, E.L.: 1995, EIT: Extreme-ultraviolet
imaging telescope for the SOHO mission. Solar Phys. 162, 291 312.
Delouille, V., Patoul, J., Hochedez, J.F., Jacques, L., Antoine, J.P.: 2005, Wavelet spectrum analysis of
EIT/SOHO images. Solar Phys. 228, 301 321.
Golub, L., Hartquist, T.W., Quillen, A.C.: 1989, Comments on the observability of coronal variations. Solar
Phys. 122, 245 261.
Hochedez, J.F., Lemaire, P., Pace, E., Schhle, U., Verwichte, E.: 2001, Wide bandgap EUV and VUV imagers for the Solar Orbiter. In: Battrick, B., Sawaya-Lacoste, H., Marsch, E., Martinez Pillet, V., Fleck,
B., Marsden, R. (eds.) Solar Encounter. Proceedings of the First Solar Orbiter Workshop 493. ESA,
Noordwijk, 245 250.
Isliker, H., Anastasiadis, A., Vlahos, L.: 2001, MHD consistent cellular automata (CA) models. II. Applications to Solar flares. Astron. Astrophys. 377, 1068 1080.
Janesick, J., Klaasen, K., Elliott, T.: 1985, CCD charge collection efficiency and the photon transfer technique.
In: Dereniak, E.L., Prettyjohns, K.N. (eds.) Solid State Imaging Arrays (SPIE) 570, 7 19.
Katsukawa, Y., Tsuneta, S.: 2001, Small fluctuation of coronal X-ray intensity and a signature of nanoflares.
Astrophys. J. 557, 343 350.
Krucker, S., Benz, A.O.: 1998, energy distribution of heating processes in the quiet Solar corona. Astrophys.
J. Lett. 501, 213 216.
Lin, R.P., Schwartz, R.A., Kane, S.R., Pelling, R.M., Hurley, K.C.: 1984, Solar hard X-ray microflares. Astrophys. J. 283, 421 425.
Paczuski, M., Boettcher, S., Baiesi, M.: 2005, Interoccurrence times in the Bak Tang Wiesenfeld sandpile
model: A comparison with the observed statistics of Solar flares. Phys. Rev. Lett. 95(18), 181102
181105.
Snyder, D.L., Miller, M.I.: 1991, Random Point Processes in Time and Space, Springer, New York.
Vhel, J.L., Legrand, P.: 2004, Signal and image processing with fraclab. In: Novak, M. (ed.) Thinking in
Patterns, World Scientific, Singapore, 321 323.
Viticchi, B., Del Moro, D., Berrilli, F.: 2006, Statistical properties of synthetic nanoflares. Astrophys. J. 652,
1734 1739.
Vlahos, L., Georgoulis, M., Kluiving, R., Paschos, P.: 1995, The statistical flare. Astron. Astrophys. 299,
897 911.

C. Alex Young Peter T. Gallagher

Abstract Coronal Mass Ejections (CMEs) are challenging objects to detect using automated techniques, due to their high velocity and diffuse, irregular morphology. A necessary step to automating the detection process is to first remove the subjectivity introduced
by the observer used in the current, standard, CME detection and tracking method. Here
we describe and demonstrate a multiscale edge detection technique that addresses this step
and could serve as one part of an automated CME detection system. This method provides
a way to objectively define a CME front with associated error estimates. These fronts can
then be used to extract CME morphology and kinematics. We apply this technique to a CME
observed on 18 April 2000 by the Large Angle Solar COronagraph experiment (LASCO)
C2/C3 and a CME observed on 21 April 2002 by LASCO C2/C3 and the Transition Region
and Coronal Explorer (TRACE). For the two examples in this work, the heights determined
by the standard manual method are larger than those determined with the multiscale method
by 10% using LASCO data and 20% using TRACE data.
Keywords Sun: corona Sun: coronal mass ejections (CMEs) Techniques: image
processing
1. Introduction
Currently, the standard method for detection and tracking of Coronal Mass Ejections
(CMEs) is by visual inspection (e.g., the CUA CDAW CME catalog, available at http:
//cdaw.gsfc.nasa.gov). A human operator uses a sequence of images to visually locate a
CME. A feature of interest is marked interactively so that it may be tracked in the sequence
C.A. Young ()
ADNET Systems Inc., NASA/GSFC, Greenbelt, MD 20771, USA
e-mail: c.alex.young@nasa.gov
P.T. Gallagher
Astrophysics Research Group, School of Physics, Trinity College Dublin, Dublin 2, Ireland
DOI: 10.1007/978-0-387-98154-3_18
247
248
C.A. Young, P.T. Gallagher
of images. From these manual measurements, the observer can then plot height time profiles of the CMEs. These height measurements are used to compute velocities and accelerations (e.g., Gallagher, Lawrence, and Dennis, 2003). Although the human visual system is
very sensitive, there are many problems with this technique. This methodology is inherently
subjective and prone to error. Both the detection and tracking of a CME is dependent upon
the experience, skill, and well being of the operator. Which object or part of the CME to
track varies, from observer to observer and is highly dependent on the quality of the data.
Also, there is no way to obtain statistical uncertainties, which is particularly important for
the determination of velocity and acceleration profiles. Lastly, the inability to handle large
data volumes and the use of an interactive, manual analysis do not allow for a real-time
data analysis required for space-weather forecasting. A visual analysis of coronagraph data
is a tedious and labor-intensive task. Current data rates from the Solar and Heliospheric
Observatory (SOHO) are low enough to make an interactive analysis possible (< 1 Gigabyte/day). This will not be the case for recent missions such as the Solar TErrestrial RElations Observatory (STEREO) and new missions such as the Solar Dynamics Observatory
(SDO). These missions have projected data rates that make an interactive analysis infeasible
(> 1 Terabyte/day). For these reasons, it is necessary to develop an automatic, real-time
CME detection and tracking system. Current digital imaging processing methods and paradigms provide the tools needed for such a system. In the next section we discuss in general,
the parts needed in an automated system, as well as some of the current work on this topic.
2. A System for Automatic CME Detection

The general design for an automated CME detection system should basically follow the digital image-processing paradigm described by Gonzalez and Woods (2002). Digital image
processing as they describe can basically be broken into three parts: i) image preprocessing and segmentation; ii) image representation and description, and; iii) object recognition.
Image preprocessing includes standard image preparation such as calibration, cosmic-ray removal but it also includes noise reduction based on the statistics of the image (e.g., Gaussian
or Poisson; Starck and Murtagh, 2002). Image segmentation is the extraction of individual features of interest in an image (e.g., edges, boundaries, and regions). Some methods
used for this part include filtering, edge detection, and morphological operations. Image
representation and description converts the extracted features into a form such as statistical
moments or topological descriptors (e.g., areas, lengths, etc.) that are easier to store and manipulate computationally. Object recognition includes techniques such as neural networks
and support vector machines (SVMs) to characterize and classify descriptors determined in
the previous step.
Determining the complex structure of CMEs is complicated by the fact that CMEs are
diffuse objects with ill-defined boundaries, making their automatic detection with many
traditional image processing techniques a difficult task. To address this difficulty, new image
processing methods were employed by Stenborg and Cobelli (2003) and Portier-Fozzani et
al. (2001), who were the first to apply a wavelet-based technique to study the multiscale
nature of coronal structures in LASCO and EIT data, respectively. Their methods employed
a multilevel decomposition scheme via the so-called trous wavelet transform.
Robbrecht and Berghmans (2004), developed a system to autonomously detect CMEs in
image sequences from LASCO. Their software, Computer Aided CME Tracking (CACTus)
(http://sidc.oma.be/cactus/), relies on the detection of bright ridges in CME height time
maps using the Hough transform. The main limitation of this method is that the Hough transform (as implemented) imposes a linear height time evolution, therefore forcing constant
249
Table 1 Data sets for the 18 April 2000 and 21 April 2002 CMEs used in this work.
18 April 2000
21 April 2002
C2
C3
C2
C3
TRACE
Start time (UT)
16:06
17:18
01:27
01:42
00:54
Wavelength ()
5400 6400
4000 8500
5400 6400
4000 8500
195
no. images
30
pix. size (arcsec)
11.9
56
11.9
56
0.5
min/max cadence
16/37 min
24/60 min
16/37 min
24/60 min
20 sec
velocity profiles for each bright feature. This method is therefore not appropriate to study
CME acceleration. Other autonomous CME detection systems include ARTEMIS (Boursier
et al., 2005) (http://lascor.oamp.fr/lasco/index.jsp), the Solar Eruptive Event Detection Systems (SEEDS) (Olmedo et al., 2005) (http://spaceweather.gmu.edu/seeds/), and the Solar
Feature Monitor (Qu et al., 2006) (http://filament.njit.edu/detection/vso.html).
In this work, a multiscale edge detector is used to objectively identify and track CME
leading edges. In Section 3, Transition Region and Coronal Explorer (TRACE; Handy et
al., 1999) and Large Angle and Spectrometric COronagraph experiment (SOHO/LASCO;
Brueckner et al., 1995) observations and initial data-reduction is discussed, while the
multiscale-based edge detection techniques are presented in Section 4. Our results and conclusions are then given in Sections 5 and 6.
3. Observations
To demonstrate the use of multiscale edge detection on CMEs in EUV imaging and whitelight coronagraph data, two data sets were used. The first data set contains a CME observed
on 18 April 2000 with the LASCO C2 and C3 telescopes. The data set contains six C2
images and two C3 images. The second data set of TRACE and LASCO observations contains a CME observed on 21 April 2002. The data set contains one C2 image, three C3
images, and 30 TRACE images. TRACE observed a very faint loop-like system propagating
away from the initial flare brightening, which was similar in shape to a CME (or number of
CMEs) observed in LASCO. The appearance of the features remained relatively constant as
they passed through the TRACE 195 passband and LASCO fields of view.
The TRACE observations were taken during a standard 195 bandpass observing campaign that provides 20-second cadence and an image scale of 0.5 arcsec per pixel. Following
Gallagher et al. (2002), standard image corrections were first applied before pointing offsets
were accounted for by cross-correlating and shifting each frame. The white-light LASCO
images were obtained using a standard LASCO observing sequence. The C2 images were
taken using the orange filter (5400 6400 ) with a variable cadence between 16 and 37
minutes and an image scale of 11.9 arcsec per pixel. The C3 images were taken using the
clear filter (4000 8500 ) with a variable cadence between 24 and 60 minutes and an image scale of 56 arcsec per pixel. Both the C2 and C3 images were unpolarized. Table 1
summarizes the details of these two data sets.
250
4. Methodology
4.1. Edge Detection
Sharp variations or discontinuities often carry the most important information in signals
and images. In images, these take the form of boundaries described by edges that can be
detected by taking first and second derivatives (Marr, 1982). The most common choice of
first derivative for images is the gradient (Gonzalez and Woods, 2002). The gradient of an
image I (x, y) at a point (x, y) is the vector,
I
Gx
=
.
(1)
I (x, y) = x
I
G
y
y
The gradient points in the direction of maximum change of I at (x, y). The magnitude of
this change is defined by the magnitude of the gradient,

I (x, y) = G2 + G2 1/2 .
x
(2)
The direction of the change at (x, y) measured with respect to the x-axis is
(x, y) = arctan(Gx /Gy ).
(3)
The edge direction is perpendicular to the gradients direction at (x, y).

The partial derivatives Gx and Gy are well approximated by the Sobel and Roberts gradient operators (Gonzalez and Woods, 2002), although these operators cannot easily be
adapted to multiscale applications. In this work, scale is considered to be the size of the
neighborhood over which the changes are calculated.
Most edges are not steep, so additional information to that returned by the gradient is
needed to accurately describe edge properties. This can be achieved using multiscale techniques. First proposed by Canny (1986), this form of edge detection uses Gaussians of different width ( ) as a smoothing operator ( ). The Gaussians are convolved with the original
image, so that Gaussians with smaller width correspond to smaller length scales. Equation
(1) for the Canny edge detector can be written as

( I (x, y))

x
I (x, y) =
.
(4)
( I (x, y))
y
The image is smoothed using a Gaussian filter with a selected . Then a derivative of the
smoothed image is computed for the x and y-direction. The local-gradient magnitude and
direction are computed at each point in the image. An edge point is defined as a point whose
gradient-magnitude is locally maximum in the direction defined by . These edge points
form ridges in the gradient-magnitude image. The process of non-maximal suppression is
performed by setting to zero all pixels not lying along the ridges. The ridge pixels are then
thresholded using two thresholds (Ta and Tb with Tb > Ta ). Ridge pixels with values between
Ta and Tb are defined as weak edges. The ridge pixels with values greater than Tb are called
strong edges. The edges are linked by incorporating the weak edges that are 8-connected
with the strong pixels (Gonzalez and Woods, 2002). Figure 1 shows a comparison of the
Roberts, Sobel, and Canny edge detectors applied to an unprocessed LASCO C2 image of
the 18 April 2000 CME from 16:54 UT. The Roberts (Figure 1d) and the Sobel (Figure 1e)
251
Figure 1 (a) A LASCO C2 image of a 18 April 2000 CME at 16:54 UT processed with a monthly background model. (b) A running-difference image made by subtracting the previous C2 image from the 16:54 UT
image. (c) The unprocessed C2 image. (d) Edges from applying the Roberts edge detector to the unprocessed
image. (e) Edges from applying the Sobel edge detector to the unprocessed image. (f) Edges from applying
the Canny edge detector ( = 5) to the unprocessed image.
detectors pick up a small piece of the CME core and noise. Using the multiscale nature of
the Canny detector (Figure 1f), choosing a larger scale size ( = 5), edges corresponding to
streamers and the CME front can be seen. Unfortunately, the Canny method has two main
limitations: i) it is slow because it is based on a continuous transform and, ii) there is no
natural way to select appropriate scales.
4.2. Multiscale Edge Detection
Mallat and Hwang (1992), showed that the maximum modulus of the continuous wavelet
transform ( MMWT ) is equivalent to the multiscale Canny edge detector described in the
previous section. The wavelet transform converts a 2D image into a 3D function, where two
of the dimensions are position parameters and the third dimension is scale. The transform
decomposes the image into translated and dilated (scaled) versions of a basic function called
a wavelet: (x, y). A wavelet dilated by a scale factor s is denoted as

1
x y
,
(5)
= s 2 s 1 x, s 1 y .
s (x, y) = 2
s
s s
The wavelet is a function that satisfies a specific set of conditions, but these functions
have the key characteristic that they are localized in position and scale (Mallat, 1998). The
minimum and maximum scales are determined by the wavelet transform, addressing the first
problem of the Canny edge detector mentioned in the previous section.
252
Mallat and Zhong (1992) refined the wavelet transform described by Mallat and Hwang,
creating a fast, discrete transform. This addresses the computational speed problem of the
Canny edge detector making this fast transform more suited to realtime applications. As with
the Canny edge detector, this wavelet starts with a smoothing function: s 2 (s 1 x, s 1 y).
The smoothing function is a cubic spline, a discrete approximation of a Gaussian. The
smoothing function is also separable, i.e., (x, y) = (x) (y). The wavelets are then the
first derivative of the smoothing function. This allows the wavelets to be written as
sx (x, y) = s 2
(s 1 x) 1
s y and
x

(s 1 y)
.
sy (x, y) = s 2 s 1 x
y
(6)
Another factor that adds to the speed of the wavelet transform algorithm is the choice
of a dyadic scale factor (s). Dyadic means that s = 2j where j = 1, 2, 3, . . ., J or s =
21 , 22 , 23 , . . ., 2J , smallest scale to largest scale. The index J is determined by the largest
dimension of the image, N , i.e., N = 2J . The wavelet transforms of I (x, y) with respect to
x and y at scale s can then be written
Wsx I = Wsx I (x, y) = sx (x, y) I (x, y) and
y
Ws I = Ws I (x, y) = s (x, y) I (x, y),
(7)
where denotes a convolution. Substituting these into Equations (2) and (3) gives the following expression for the gradient of an image at scale s in terms of the wavelets:

s I (x, y) = W x I 2 + Wsy I 2 1/2
s

y
s (x, y) = arctan Wsx I /Ws I .
and
(8)
The detailed steps associated with implementing Equation (6) are shown in Figure 2. The
rows from top to bottom are scales one to five, respectively. Column (a) displays the horizontal wavelet components Wsx I (x, y). Column (b) shows the vertical wavelet components
y
Ws I (x, y). The next two columns show the magnitude (c) of the multiscale gradient and
the angle (d) of the multiscale gradient. The edges calculated from the multiscale gradient
are displayed in column (e).
4.3. Edge Selection and Error Estimation
Once the gradient was found using the wavelet transform, the edges were calculated using
the local maxima at each scale, as described in the end of Section 4.1. Closed or stationary
edges due to features such as coronal loops, ribbons, and cosmic rays were then removed,
thus leaving only moving, open edges visible. Currently not all available information is
used so there were still some open spurious edges that were removed manually. Finally,
only the edges from expanding, moving features were left. It was these edges that were used
to characterize the temporal evolution of the CME front.
The multiscale edge detector can objectively define the CME front but it is also important
to estimate the statistical uncertainty in the edges and to obtain errors in position or height.
A straightforward way to do this is by using a bootstrap (Efron and Tibshirani, 1993). To
do this we must create statistical realizations of the data, then apply the multiscale edge
detection method to each realization. The realizations of the data are created by estimating
a true, non-noisy image then applying a noise model to the true image. Applying the noise
model means using a random number generator (random deviate) for our particular noise
model to generate noise and adding it to the non-noisy image estimate. In our case the
253
Figure 2 The steps calculating the multiscale edges of the unprocessed C2 image (18 April 2000 CME
at 16:54 UT). The rows from top to bottom are scales j = 1 to j = 5, respectively. (a) Horizontal wavelet
coefficients. (b) Vertical wavelet coefficients. (c) The magnitude of the multiscale gradient. (d) The angle of
the multiscale gradient. (e) The edges calculated from the multiscale gradient.
noise model is well approximated by Gaussian noise. Estimation of the noise and the true
image is described by Starck and Murtagh (2002). The noise model is applied 1000 times
to the true image with each application creating a new realization. Doing this 1000 times,
a mean and standard deviation is calculated for the edge location or in our case for each
height point used. The steps for creating the estimate are i) estimate the noise in the original
image, ii) compute the isotropic wavelet transform of the image ( trous wavelet transform),
iii) threshold the wavelet coefficients based on the estimated noise, and iv) reconstruct the
image. The reconstructed image is the estimate of the true image. The noise estimate from
the original image is used in the noise model applied to the estimated true image.
254
Figure 3 Illustration of a CME edge detection in subsequent images. (a) The original LASCO C2 images,
(b) running difference images of the LASCO C2 images, and (c) application of the multiscale edge detection
algorithm to the sequence of the original images. The edges are the black lines displayed over the running
difference images. The CME erupted on 18 April 2000, the times for the frames are (from left to right) 16:06
UT, 16:30 UT, 16:54 UT, and 17:06 UT.
5. Results
The first example of application of the multiscale edge detection is illustrated in Figure 3.
Figure 3a shows four of the original, unprocessed LASCO C2 images for the 18 April 2000
data set. The times for the frames from left to right are 16:06 UT, 16:30 UT, 16:54 UT, and
17:06 UT. Figure 3b shows running difference images for the sequence of C2 images. The
results of the multiscale edge detection method applied to the original images are shown in
Figure 3c. The edges of the CME front are displayed as black contours over the difference
images shown in Figure 3b. Once the method is applied to the entire data set of C2 and C3
images, one point from each edge (all at the same polar angle) is selected. The distance of
each point from Sun center is plotted against time to create a height-time profile. A bootstrap
(described in Section 4.3) was performed for each edge calculation so that each point in the
height time profile has an associated error in height.
The resulting height time profile is shown in Figure 4. The data plotted using + symbols are those obtained via the multiscale method. One- errors in height are also shown.
For comparison, data from the CUA CDAW CME catalog are plotted in the figure using
symbols. The points determined with the multiscale method are systematically lower in
height than the points from the CUA CDAW points. This is because the CUA CDAW points
are selected using difference images and, as can be seen in Figure 3c, the true CME front
(black edge) is inside of the edge in the difference images. This illustrates a drawback of
255
Figure 4 The height time

profile for the 18 April 2000
CME. The data plotted using +
symbols were obtained via the
multiscale methods described
herein, while the data plotted
using symbols are from the CU
CDAW CME catalog. The first
six points are for LASCO C2 and
the last two are for LASCO C3.
Figure 5 (a) TRACE 195 difference image from 21 April 2002, created by subtracting an image taken
at 00:42:30 UT from an image at 00:54:54 UT, together with running difference images from (b) LASCO
C2 (01:50:05 UT) and (c) LASCO C3 (02:13:05 UT). The same (d) TRACE and (e), (f) LASCO images
(as shown in (a), (b), and (c) respectively, but now overlaid with multiscale edges from a scale (j = 8) that
isolates the leading-edge of the CME. (The C2 images are not cut off on the right side. In order for the C2
images to have a similar FOV as the C3 images and for all images to be square, white space was added to the
C2 images.)
using difference images to determine the CME front. The difference image height estimates
from the CUA CDAW catalog are larger than the multiscale edge estimates by 10%.
Figure 5 displays data from the 21 April 2002 data set. Figure 5a shows a TRACE difference image created by subtracting an image at 00:42:30 UT from an image at 00:54:54 UT.
256
Figure 6 (a) TRACE 195 difference image from 21 April 2002, created by subtracting an image taken at
00:42:30 UT from an image at 00:54:54 UT. (b) The same TRACE images, but now overlaid with multiscale
edges from a scale (j = 8) that isolates the leading-edge of the CME. (c) The set of multiscale edges at scale
j = 8, from all 30 TRACE images (00:54:54 UT to 01:05:19 UT) superimposed on the first image of the
TRACE sequence. (d) Same as (c), but with unwanted edges removed.
A very faint loop-like feature was only visible in the difference image after it was smoothed
and scaled to an extremely narrow range of intensities. Both these operations were arbitrarily decided upon, and are therefore likely to lead to the objects true morphology being
distorted. Figure 5d shows the same TRACE difference image as Figure 5a but overlaid
with multiscale edges at scale j = 8. The edge of the faint loop-like features is clearly visible as a result of decomposing the image using wavelets, and then searching for edges in the
resulting scale maps using the gradient-based techniques described in the previous section.
Figures 5b and 5c show LASCO C2 and C3 running difference images, respectively. Figures
5e and 5f show the LASCO difference images overlaid with the multiscale edges for scale
j = 8.
Figure 6 displays on the TRACE data for the 21 April 2002 data set. Figure 6a displays
the processed difference image (same as Figure 5a) and Figure 6b is the difference image
257
Figure 7 The height time profile for the 21 April 2002 CME, together with fits assuming an exponential
acceleration of the form given in Equation (9). The data plotted using open symbols were obtained via the
multiscale methods described here, while the data plotted using solid symbols are from Gallagher, Lawrence,
and Dennis (2003). The solid line is a fit to the multiscale data, the dashed to the latter. Error bars on the
LASCO data have been multiplied by a factor of two to improve their visibility and the mean uncertainty in
TRACE-derived heights are equal to the diameter of the open and filled circles ( 3 Mm).
overlaid with multiscale edges at scale j = 8 (same as Figure 5d). The underlying TRACE
image in Figure 6c is the original image from 00:54:54 UT. The multiscale edges at scale
j = 8 were calculated for all 30 TRACE images from 00:54:54 UT to 01:05:19 UT. All 30
sets of edges were overlaid upon the original base image. Figure 6d contains the same set of
edges but by using size and shape information, the expanding front is isolated. The leading
edge, only partially visible in the original TRACE difference image in Figure 6a, is now
clearly visible and therefore more straightforward to characterize in terms of its morphology
and kinematics. The multiscale edges reveal the existence of two separate moving features.
Using these edges, the expansion and motion of the CME from the Sun is now clearly
visible and therefore characterizing it in terms of its morphology and kinematics is more
straightforward. The resulting height time plot is shown in Figure 7, together with data
from Gallagher, Lawrence, and Dennis (2003), for comparison. We again find the heights
determined manually are larger by 10% using LASCO data and 20% using TRACE
data.
Following a procedure similar to that of Gallagher, Lawrence, and Dennis, the height
time curve was fitted assuming a double exponential acceleration profile of the form:

a(t) =
1
1
+
ar exp(t/r ) ad exp(t/d )
1
,
(9)
where ar and ad are the initial accelerations and r and d give the e-folding times for the rise
and decay phases. A best fit to the height time curve was obtained with h0 = 17 3 Mm,
258
v0 = 40 4 km s1 , ar = 1 1 m s2 , r = 138 26 s, ad = 4950 926 m s2 , and d =

1100 122 s, and is shown in Figure 7.
6. Conclusions and Future Work

CMEs are diffuse, ill-defined features that propagate through the solar corona and inner
heliosphere at large velocities ( 100 km s1 ) (Yashiro et al., 2004), making their detection and characterization a difficult task. Multiscale methods offer a powerful technique by
which CMEs can be automatically identified and characterized in terms of their shape, size,
velocity, and acceleration.
Here, the entire leading edge of the 18 April 2000 and 21 April 2002 CMEs has been
objectively identified and tracked using a combination of wavelet and gradient based techniques. We have shown that multiscale edge detection successfully locates the front edge for
both well-defined events seen in LASCO as well as very faint structures seen in TRACE.
Although height time profiles were only calculated for one point, this method allows us to
objectively calculate height time profiles for the entire edge. This represents an advancement over previous point-and-click or difference-based methods, which only facilitate the
CME apex to be tracked. Comparing height time profiles determined using standard methods with the multiscale method shows that for these two CMEs the heights determined
manually are larger by 10% using LASCO data and 20% using TRACE data.
Future work is needed to fully test the use of this technique in an automated system.
Application of this edge detection method to a large, diverse set of events is necessary. An
important improvement would be better edge selection. This will be accomplished by better incorporating scale information. By chaining the edges together as a function of scale
we can distinguish false edges from true edges. This information can also be used to better distinguish different-shaped edges. Another improvement can be made by using image
enhancement. During the denoising stage, wavelet-based image enhancement (such as in
Stenborg and Cobelli, 2003) can be performed at the same time that noise in the images
is estimated and reduced. More sophisticated multiscale methods using transforms such as
curvelets will be studied. In order to distinguish between edges such as those due to streamers from those due to a CME front angle information can be incorporated. This multiscale
edge detection has been shown to have potential as a useful tool for studying the structure
and dynamics of CMEs. With a few more improvements this method could prove to be an
important part of an automated system.
Acknowledgements The authors thank the referee for their suggestions and comments and NASA for its
support. C.A.Y. is supported by the NASA SESDA contract. P.T.G. is supported by a grant from Science
Foundation Irelands Research Frontiers Programme.
References
Boursier, Y., Llebaria, A., Goudail, F., Lamy, P., Robelus, S.: 2005, Automatic detection of coronal mass
ejections on LASCO-C2 synoptic maps. SPIE 5901, 13 24.
C.J.: 1995, The large angle spectroscopic coronagraph (LASCO). Solar Phys. 162, 357 402.
Canny, J.: 1986, A computational approach to edge detection. IEEE Trans. Pattern Recognit. Mach. Intell. 8,
679 698.
Efron, S., Tibshirani, R.J.: 1993, An Introduction to the Bootstrap, Chapman and Hall, New York.
259
Gallagher, P.T., Lawrence, G.R., Dennis, B.R.: 2003, Rapid acceleration of a coronal mass ejection in the low
corona and implications for propagation. Astrophys. J. 588, L53 L56.
Gallagher, P.T., Dennis, B.R., Krucker, S., Schwartz, R.A., Tolbert, A.K.: 2002, RHESSI and TRACE observations of the 21 April 2002 X1.5 flare. Solar Phys. 210, 341 356.
Gonzalez, R.C., Woods, R.E.: 2002, Digital Image Processing, 2nd edn., Prentice-Hall, Upper Sadle River.
Handy, B.N., Acton, L.W., Kankelborg, C.C., Wolfson, C.J., Akin, D.J., Bruner, M.E., Caravalho, R., Catura,
R.C., Chevalier, R., Duncan, D.W., Edwards, C.G., Feinstein, C.N., Freeland, S.L., Friedlaender, F.M.,
Hoffmann, C.H., Hurlburt, N.E., Jurcevich, B.K., Katz, N.L., Kelly, G.A., Lemen, J.R., Levay, M., Lindgren, R.W., Mathur, D.P., Meyer, S.B., Morrison, S.J., Morrison, M.D., Nightingale, R.W., Pope, T.P.,
Rehse, R.A., Schrijver, C.J., Shine, R.A., Shing, L., Strong, K.T., Tarbell, T.D., Title, A.M., Torgerson,
D.D., Golub, L., Bookbinder, J.A., Caldwell, D., Cheimets, P.N., Davis, W.N., Deluca, E.E., McMullen,
R.A., Warren, H.P., Amato, D., Fisher, R., Maldonado, H., Parkinson, C.: 1999, The transition region
and coronal explorer. Solar Phys. 187, 229 260.
Mallat, S.: 1998, A Wavelet Tour of Signal Processing, Academic, San Diego.
Mallat, S., Hwang, W.L.: 1992, Singularity detection and processing with wavelets. IEEE Trans. Inf. Theory
38, 617 643.
Mallat, S., Zhong, S.: 1992, Characterization of signals from multiscale edges. IEEE Trans. Pattern Anal.
Mach. Intell. 14, 710 732.
Marr, D.: 1982, Vision, W.H. Freeman, New York.
Olmedo, O., Zhang, J., Wechsler, H., Borne, K., Poland, A.: 2005, Solar eruptive event detection system
(SEEDS). Bull. Am. Astron. Soc. 37, 1342.
Portier-Fozzani, F., Vandame, B., Bijaoui, A., Maucherat, A.J.: 2001, A multiscale vision model applied to
analyze EIT images of the solar corona. Solar Phys. 201, 271 287.
Qu, M., Shih, F.Y., Jing, J., Wang, H.: 2006, Automatic detection and classification of coronal mass ejections.
Solar Phys. 237, 419 431.
Robbrecht, E., Berghmans, D.: 2004, Automated recognition of coronal mass ejections (CMEs) in near-realtime data. Astron. Astrophys. 425, 1097 1106.
Starck, J.-L., Murtagh, F.: 2002, Handbook of Astronomical Data Analysis, Springer, Berlin.
Stenborg, G., Cobelli, P.J.: 2003, A wavelet packets equalization technique to reveal the multiple spatial-scale
nature of coronal structures. Astron. Astrophys. 398, 1185 1193.
Yashiro, S., Gopalswamy, N., Michalek, G., St. Cyr, O.C., Plunkett, S.P., Rich, N.B., Howard, R.A.: 2004,
A catalog of white light coronal mass ejections observed by the SOHO spacecraft. J. Geophys. Res.
109(A7), A07105.
Automated Prediction of CMEs Using Machine Learning

of CME Flare Associations
R. Qahwaji T. Colak M. Al-Omari S. Ipson

Abstract Machine-learning algorithms are applied to explore the relation between significant flares and their associated CMEs. The NGDC flares catalogue and the SOHO/LASCO
CME catalogue are processed to associate X and M-class flares with CMEs based on timing
information. Automated systems are created to process and associate years of flare and CME
data, which are later arranged in numerical-training vectors and fed to machine-learning algorithms to extract the embedded knowledge and provide learning rules that can be used
for the automated prediction of CMEs. Properties representing the intensity, flare duration, and duration of decline and duration of growth are extracted from all the associated
(A) and not-associated (NA) flares and converted to a numerical format that is suitable for
machine-learning use. The machine-learning algorithms Cascade Correlation Neural Networks (CCNN) and Support Vector Machines (SVM) are used and compared in our work.
The machine-learning systems predict, from the input of a flares properties, if the flare is
likely to initiate a CME. Intensive experiments using Jack-knife techniques are carried out
and the relationships between flare properties and CMEs are investigated using the results.
The predictive performance of SVM and CCNN is analysed and recommendations for enhancing the performance are provided.
Keywords CMEs prediction Machine learning Solar flares Space weather CME
Neural networks Support vector machines
1. Introduction
The term space weather refers to adverse conditions on the Sun, in the solar wind, and
in the Earths magnetosphere, ionosphere, and thermosphere that may affect space-borne
or ground-based technological systems and endanger human health or life (Koskinen et al.,
R. Qahwaji () T. Colak M. Al-Omari S. Ipson
Department of Electronic Imaging and Media Communications, University of Bradford, Richmond
Road, Bradford BD7 1DP, England, UK
e-mail: r.s.r.qahwaji@brad.ac.uk
DOI: 10.1007/978-0-387-98154-3_19
261
262
R. Qahwaji et al.
2001). The importance of space weather is increasing as more human activities take place
in space, and as more reliance is placed on communications and power systems.
The most dramatic solar events affecting the terrestrial environment are solar flares and
Coronal Mass Ejections (CMEs) (Koskinen et al., 2001). These are two types of solar eruptions that spew vast quantities of radiation and charged particles into space (Lenz, 2004).
Earth environment and geomagnetic activity are affected by the solar wind. The interplanetary magnetic field (IMF) creates storms by injecting plasma into the Earths magnetosphere.
Geomagnetic storms are correlated with CMEs (Wilson and Hildner, 1984) and predicting
CMEs can be useful in forecasting space weather (Webb, 2000). Major solar flares can also
seriously disrupt the ionosphere and in order to guarantee that humans can work safely
and effectively in the space, the forecast of strong solar flares is also important (Kurokawa,
2002).
There is a lack of clear definitions for solar features, which increases the difficulty of
designing automated detection and processing systems. Data volumes will soon increase
1000 to 10 000 times because of recent space mission launches (Hinode and STEREO).
Extracting useful knowledge from this vast amount of data and trying to establish useful
connections between data relating to different time periods is very challenging. In addition,
large-scale and automated data mining and processing techniques that integrate advanced
image processing and machine-learning techniques have not been fully exploited to look for
accurate correlations between the occurrences of solar activities (e.g. flares and CMEs) and
solar features observed in various wavelengths.
Despite recent advances in solar imaging, machine-learning and data mining have not
been widely applied to solar data. However, very recently, several learning algorithms (i.e.
neural networks (NNs), support vector machines (SVMs), and radial basis functions (RBFs))
were optimised for the automated short-term prediction of solar flares and the results compared (Qahwaji and Colak, 2007). These machine-learning-based systems accept two sets of
inputs: The McIntosh classification of sunspot groups and real-time simulation of the solar
cycle. Fourteen years of data from the sunspot and flare catalogues of the National Geophysical Data Center (NGDC) were explored to associate sunspots with their corresponding
flares based on timing information and NOAA numbers. Borda et al. (2002) described a
method for the automatic detection of solar flares using the multi-layer perceptron (MLP)
with back-propagation training rule, where a supervised learning technique that required a
large number of iterations was used. The classification performance for features extracted
from solar flares was compared by Qu et al. (2003) using RBF, SVM, and MLPF methods.
Each flare is represented using nine features. However, these features provide no information about the position, size and verification of solar flares. Qahwaji and Colak (2006) used
NNs after image segmentation to verify the regions of interest as solar filaments.
The aim of this paper is to provide a platform for large-scale analysis, association, and
knowledge extraction for CME and flare catalogues data. Data from the publicly available
solar flare catalogue, which are provided by the NGDC,1 are used in this study. NGDC
keeps records of data from several observatories around the world and holds one of the most
comprehensive publicly available databases for solar features and activities. The CME data
are obtained from the SOHO/LASCO CME catalogue.
This paper is organised as follows: Section 2 explores the associations between CMEs
and other solar activities or features. Section 3 describes the design of the Computer Platform for CME Prediction. The practical implementation and evaluation of the system using
1 ftp://ftp.ngdc.noaa.gov/STP/SOLAR_DATA/, last access: 2006.
263
machine-learning algorithms is discussed in Section 4. Concluding remarks and recommendations for future work are presented in Section 5.
2. CMEs and their Associations with Solar Activities and Features

CMEs are bursts of plasma ejected from the Sun. For years, solar flares were thought to be
responsible for major interplanetary (IP) particle events and geomagnetic storms. However,
space-based coronagraphs revealed the existence of CMEs (Tousey, 1973). Since then there
have been many studies to investigate how CMEs are initiated and triggered. A pioneering
and controversial work by Gosling (1995), argues that CMEs, not flares, were the critical element for large geomagnetic storms, interplanetary shocks, and major solar-energetic-particle
events. This contradicts the findings of Lin and Hudson (1976) where the particles accelerated in big flares are thought to provide the energy for all of the later activities (i.e., CMEs
and large energetic particles events). It is not clear whether there is a cause and effect relation between flares and CMEs and this uncertainty has driven most of the solar flare myth
controversy (Cliver and Hudson, 2002). We have conducted an extensive survey of CME associations with other solar features and have concluded that there have been limited studies
containing large-scale processing and analysis of years of solar data to explore the associations between CMEs and other solar activities and/or features. From previous research it
can also be shown that there is a degree of association between CMEs on the one hand and
flares and erupting filaments/prominences on the other. The exact degree of association is
not clear though because most of the available studies were carried out on a few years of
data or on limited cases and using physics-based modelling. In some cases, contradicting
findings are reported. Also data mining and machine learning have not been implemented
before to verify this association and to represent it using computer-based learning rules that
can be used to extract knowledge and provide predictions by analysing recent data in realtime. In this work the aim is to provide, for the first time, a machine-learning-based study to
provide a new perspective on this long-standing problem in solar physics.
3. The Computer Platform Design for CME Predictions

We present a computer platform that analyses all of the available data from flare and CME
catalogues to extract learning rules and then provide automated predictions for CMEs. Several different stages are involved in this work, as shown in Figure 1 and explained in the
following sections.
Figure 1 The hybrid prediction
computer system.
264
R. Qahwaji et al.
Table 1 The levels of associations based on values of and .

Flares
Total
NA
15
389
5554
3355
PA ( = 150)
89
926
6770
2066
9851
104
1315
12 324
5421
19 164
A ( = 30)
57
318
1181
246
1802
A ( = 60)
71
510
2229
526
3336
A ( = 90)
77
592
3016
764
4449
A ( = 120)
78
654
3757
1018
5507
Total
9313
3.1. Associating Flares and CMEs

A C++ platform was created to automatically associate CMEs in the SOHO/LASCO CMEs
catalogue with flares in the NGDC X-ray flares catalogue. The association is determined
based on their timing information; the date and time of every CME is compared with date
and time of every flare.
Two criteria are used for comparison:
If a CME is not recorded minutes before or after the time a flare reaches its peak, then
this flare is marked as not-associated (NA), otherwise it is marked as possibly-associated
(PA).
If a CME is recorded minutes after the time a PA flare reaches its peak, then this flare
is marked as an associated (A) flare.
After finding all of the associations, a numerical dataset was created for the machinelearning algorithms using associated and not-associated flares.
3.2. Creating the Associated Numerical Data Set
All of the CME and flare data for the period from January 1996 through December 2004
were processed, analysing data relating to 9297 CMEs and 19 164 flares.
To determine the NA flares, the value of was made equal to 150 minutes in all of
the experiments. It is easier to determine if a CME is not associated with any flares rather
than determine the level of association between every CME with flares based on timing
information. To explore the different levels of associations, the association algorithm was
applied with different values of , as shown in Table 1.
We associate CMEs with significant flares (i.e. X- and M-class flares) only. In previous work (Qahwaji and Colak, 2007), an automated machine-learning system that provides
short-term predictions for the occurrences of these significant flares was introduced. Our
long-term goal is to determine the level of associations between CMEs and flares using
machine-learning so that a hybrid system that integrates both systems can be designed. Associating CMEs with significant flares seems to be supported by the findings of Yashiro
et al. (2005), that all CMEs associated with X-class flares are detected by LASCO, while
almost half the CMEs associated with C flares are invisible. They also concluded that the
CME association rate increases with the increase of the X-ray brightness for flares starting
from 20% for C-class flares (between C3 and C9 levels) and reaching 100% for huge flares
(above X3 level). In addition, they found that faster (median 1556 km s1 ) and wider (median 244) CMEs are associated with X-class flares while slower (432 km s1 ) and narrower
(68) CMEs are associated with disk C-class flares.
265
Table 2 Description of each property that is used as an input node in the machine-learning algorithms.
Name
Description
Intensity
The normalised numerical value of intensity of the flare (Intensity 1000)
Flare duration
The normalised numerical value of the time difference in minutes between the
ending and the starting times of the flare (Difference/120)
Decline duration
ending and the peak times of the flare (Difference/120)
Incline duration
peak and the starting times of the flare (Difference/120)
As shown in Table 1, more CMEs are associated with flares as the value of increases. The rate of increase in the number of associations is higher when increases from
30 minutes to 60 minutes. The rates of increase are equal to 85%, 33%, and 23% when
increases from 30 to 60, from 60 to 90, and from 90 to 120, respectively. Since the increase in the association rate drops from 85% to 33% over a 60-minute difference, the value
= 60 was found to be most suitable for the experiments. By applying the association algorithm, with = 150 minutes and = 60 minutes, an associated data set consisting of
985 flares with 581 A flares and 404 NA flares was created. Possibly associated flares were
not included in any of the data sets. As shown in Table 1, 71 X-class flares and 510 Mclass flares were associated with CMEs, while 15 X-class flares and 389 M-class flares were
not associated. Because machine-learning algorithms deal mainly with numbers, it was essential that appropriate numerical representations for A and NA flares were proposed and
implemented. Properties such as intensity, starting time, peak and ending times of the flares
can be extracted from the NGDC flares catalogue. However, it was hoped to include additional properties such as flare locations. Unfortunately a large number of the associated
flares do not have locations included in the NGDC catalogues. Hence, it was decided to
use the properties shown in Table 2. Numerical representations of these properties are used
later to construct input parameters for the training and testing stages of the machine-learning
system. As it was not clear which properties are more important for machine-learning and
for the prediction of CMEs, it was decided to carry out extensive experiments in order to
determine the significance of each property for this application.
4. Practical Implementation and Results

After creating the associated data set, the training and testing experiments for the machinelearning algorithms were started. These experiments and the evaluation of prediction performance are explained below.
4.1. The Learning Algorithms and Techniques
The present study has compared the performances of Cascade Correlation Neural Networks
(CCNNs) and Support Vector Machines (SVMs) which have both proven to be very effective learning algorithms for similar applications; see Qahwaji and Colak (2007), where
more information on the theory and implementation of these learning algorithms is given.
All of the machine-learning/training and testing experiments were carried out with the aid of
the Jack-knife technique (Fukunaga, 1990). This technique is usually implemented in order
266
R. Qahwaji et al.
to provide a correct statistical evaluation of the performance of the classifier when implemented on a limited number of samples. The technique divides the total number of samples
into two sets: a training set and a testing set. In practice, a random number generator is used
to decide which samples are used for training the classifier and which are used for testing
it. The classification error depends mainly on the training and testing samples used. For a
finite number of samples, an error-counting procedure can be used to estimate the performance of the learning algorithms (Fukunaga, 1990). In this work, all of the experiments
were carried out using 80% randomly selected samples for training and the remaining 20%
for testing. The CME-prediction performance is evaluated using ROC curves, as explained
in Fawcett (2006). Two performance indicators are used: the True Positive (TP) rate and the
False Positive (FP) rate, calculated from Equations (1) and (2) respectively:
correct positive predictions
,
total positives
(1)
incorrect negative predictions

.
total negatives
(2)
TP rate =
FP rate =
In these equations: correct positive predictions is the total number of cases for which the
system correctly predicts that a flare produces a CME; incorrect negative predictions is
the total number of cases for which the system incorrectly predicted that a flare does not
produce a CME; total positives is the sum of cases for which a flare produces a CME
(number of associated cases used in testing); total negatives is the sum of cases for which
a flare did not produce a CME (number of un-associated cases used in testing).
4.2. Optimising the Learning Algorithms
The prediction performances of CCNN and SVM are compared in order to determine which
of these machine-learning algorithms is more suitable for this application. The learning
algorithms were both optimised to ensure that their best performances were achieved. In
order to find the best parameters and/or topologies for the learning algorithms, initial training
and testing experiments were applied using the Jack-knife technique as explained previously.
A total of 788 associated and not-associated flares were used for training. This constituted
80% of the total number of associated cases available. The remaining 197 associated and
not-associated flares were used for testing. The results obtained are evaluated using the ROC
analysis technique. It is also worth mentioning that all the reported TP and FP rates were
obtained by averaging the rates from five Jack-knife experiments.
4.2.1. Optimising the CCNN
Colak and Qahwaji (2007) showed that CCNNs provided the optimum neural-network performance for processing solar data in catalogues. However, a CCNN should be optimised
for the current application before it is used. The CCNNs used here consist of input, hidden,
and output layers. The output layer consists of one output node which has a numerical value
of 0.9 if a CME is predicted to occur and 0.1 if not. The number of input parameters/nodes
and the number of hidden nodes in each experiment were changed to find the best inputs
and their related topologies. The number of input parameters/nodes was varied from one to
three and the number of hidden nodes was varied from one to twenty. The test results were
recorded to provide an indication of the prediction rates for CMEs.
The MATLAB neural network toolkit was used for these experiments. The optimisation
experiments proceeded as follows:
267
Figure 2 ROC graph showing the best CCNN topologies with different inputs.
The number of input features was varied from one to three.

For each input feature the CCNN topology was varied by changing the number of hidden
nodes from 1 to 20 and arranging the training and testing data based on the Jack-knife
technique.
For every topology, five experiments were carried out based on the Jack-knife technique
and the average TP and FP values found.
At the end of these experiments, 60 CCNN topologies resulting in 60 average TP and 60
average FP values were compared. The relations between the calculated values of TPs and
FPs for different topologies are shown in Figure 2. The optimum topology for each input
feature was then determined from this graph by determining the point with the maximum
perpendicular distance from the diagonal line in the northwest direction (Fawcett, 2006).
In order to find the optimum classification threshold values that provided the best predictions, the threshold values were changed from 0 to 1 in steps of 0.01 for each input and
their associated optimum topologies.
For each threshold value, five experiments were carried out using the Jack-knife technique
and average TP and FP values obtained.
At the end of these experiments, a ROC curve was drawn and the optimum threshold
values were found, as shown in Figure 3.
The three best topologies are shown in Figure 2 and have three input nodes with three hidden
nodes, two input nodes with 19 hidden nodes and one input node with two hidden nodes.
The classification thresholds were found for these three topologies to draw their ROC curves,
which are shown in Figure 3. As shown in this figure, a CCNN with three input nodes and
268
R. Qahwaji et al.
Figure 3 ROC graph showing the best CCNN topologies with different inputs and variable threshold values.
three hidden nodes with a classification threshold of 0.56 gives the best results for CME
prediction as it provides 0.63 TP rate and 0.43 FP rate.
4.2.2. Optimising the SVM
Again, the SVM classifier should be optimised before comparing its performance to that
of the CCNN. For the reasons mentioned in Qahwaji and Colak (2007), it was decided to
use the ANOVA kernel. This requires optimising the values of Gamma and Degree and the
classification threshold afterwards.
The MySVM2 software was used for the experiments. The optimisation experiments
were started by arranging the learning data into one, two, and three inputs. The optimisation
process for SVM proceeded as follows:
The number of input features was varied from one to three.
For each new input, the degree value was varied from one to ten in steps of one and for
each degree value the gamma value was varied from ten to 100 in steps of ten.
For each of these 100 iterations, five experiments were carried out using the Jack-knife
technique and the average TP and FP values recorded.
At the end of these experiments, the average TP and FP values obtained for 300 SVM
configurations were compared. These values were plotted to find the optimum degree
2 http://www-ai.cs.uni-dortmund.de/SOFTWARE/MYSVM.
269
Figure 4 ROC graph showing the best SVM topologies with different inputs.
and gamma values to use with the ANOVA kernel. The relation between TPs and FPs for
different inputs and different topologies are shown in Figure 4 and were used to determine
the optimum SVM configuration.
In order to find the optimum classification thresholds that provide the best prediction for
the optimum SVM topologies, the threshold values were changed from 0 to 1 in steps of
0.01 for every input and their selected optimum topologies.
For each threshold value, five experiments were carried out using the Jack-knife technique
and the average TP and FP values calculated.
At the end of these experiments the ROC curve was drawn and the optimum threshold
values found as shown in Figure 5.
As can be seen by inspection of Figures 4 and 5, an SVM classifier that accepts three inputs
with Degree and Gamma values of 8 and 90 respectively and a classification threshold value
of 0.83 provides the best prediction performance. This SVM configuration provides TP and
FP rates of 0.73 and 0.53 respectively.
4.3. Comparing the Prediction Performances
The results show that both classifiers perform better when three inputs are used. In general,
there has been an increase in the prediction rate with the addition of more discriminative
input features. However, the initial experiments indicate that adding the incline duration of
flare (property D in Table 2) does not improve the prediction performance. It should also be
noted that property D can be calculated by subtracting the decline duration from the total
270
R. Qahwaji et al.
Figure 5 ROC graph showing the best SVM topologies with different inputs variable threshold values.
duration (i.e., D = B C in Table 2). From the perspective of machine-learning, this could
be the reason behind its insignificance. It can be also said that the time needed for a flare to
reach its peak intensity is not very important in terms of CME predictions using machinelearning. The experiments also indicate that the decline duration for the flare (property C)
is more important for CME prediction than the total flare duration (property B). This means
that decline duration of the flare is very important for determining the probability of CME
occurrence and this coincides with the findings of Yashiro et al. (2006).
As shown in Figure 3, a CCNN with three input nodes and three hidden nodes with a
classification threshold of 0.56 gives the best results for CME predictions as it provides 0.63
TP rate and 0.43 FP rate.
As can be seen from Figures 4 and 5, a SVM classifier that accepts three inputs with
Degree and Gamma values of 8 and 90 respectively and a classification threshold value of
0.83 provides the best prediction performance. This SVM configuration provides TP and FP
rates of 0.73 and 0.53 respectively.
It has been concluded that the optimum SVM classifier provides TP and FP rates of
0.73 and 0.53 respectively. This is a more liberal performance (Fawcett, 2006) compared
to the optimum CCNN. On the other hand, a more conservative performance is provided
by the optimum CCNN with TP and FP rates of 0.63 and 0.43 respectively. SVM classifier
generates higher positive predictions compared to CCNN, but it also produces higher rates
of false alarm predictions. If a real-time system is to be designed, then choosing the right
classifier will depend mainly on the objectives and domain of application for this system.
This is something that we intend to investigate further in the near future.
271
4.4. Further Investigation of Catalogue Data

It was decided to conduct further learning experiments with the classifiers to improve the
prediction performance. Due to limitations of the data available in the flare catalogues, it is
not practical to input more flare properties to the SVM and hence to modify the learning data.
Because the current work deals only with catalogued data, the obvious remaining option is to
modify the association rules to reduce the number of falsely associated CMEs by exploring
other features provided in the CME catalogue. The Measurement Position Angle (MPA) in
the CME catalogue was used in further experiments to provide indications of the locations
of associated flares. The rules of association, which are explained in Section 3, have been
modified to include MPA as a second criterion of comparison besides timing. Applying
this extra feature has reduced the number of associated flares. The new set, obtained using
parameter values = 150 minutes and = 60 minutes, consists of 405 A flares and 404
NA flares. The number of NA flares did not change because the previously associated flares
were treated, just based on timing, as possibly associated. The optimisation and learning
experiments were carried out as explained in Section 4. At the end of these experiments
the optimum configuration obtained for a three input SVM was 8, 90, and 0.72 for Degree,
Gamma, and Classification Threshold, respectively. This configuration provides TP and FP
rates of 0.74 and 0.59 respectively. On the other hand the optimum topology for a CCNN
is three input nodes, with three hidden nodes and a classification threshold of 0.47. This
topology generates TP and FP rates of 0.71 and 0.46 respectively.
Its obvious from these results that the prediction performance for both classifiers have
been improved and the FP rate has been reduced despite the fact that MPA provides a very
limited indication of the location of the associated flare and is a very coarse measure to use.
Using this feature enabled the association sets to be refined and hence eliminate some of the
false associations, which produced some improvement in the prediction performance.
5. Conclusions and Future Research

In this paper, a machine-learning-based system that analyses flare and CME data is introduced. This system analyses all data records in the NGDC flare catalogues and the
SOHO/LASCO CME catalogue and applies an association algorithm to associate flares with
their corresponding CMEs based on timing information. In this work, the CCNN has been
used because of its efficient knowledge extraction and generalisation performance for the
processing of solar data (Qahwaji and Colak, 2007). The SVM was also used because of its
outstanding classification performance, which has been reported in Acir and Guzelis (2004),
Pal and Mather (2004), Huang et al. (2004), and Distante, Ancona, and Siciliano (2003) to
outperform neural networks.
To determine the optimum configurations for the CCNN and SVM classification systems
that were used in this work, many experiments were carried out changing the numbers of
input nodes and hidden nodes for the CCNN and the values of Gamma and Degree for the
SVM. Different classification thresholds were tested for both classifiers to determine the
optimum configurations based on ROC curves. The Jack-knife technique was used in all
these experiments as described in Section 4.1.
All the reported flares and CMEs between 01 January 1992 and 31 December 2004 have
been investigated. The association software has associated 581 M and X soft X-ray flares
with corresponding CME groups and highlighted another 404 significant flares as not associated with any CMEs. These associations are for parameter values = 150 minutes and
272
R. Qahwaji et al.
= 60 minutes. After finding the optimum configurations for SVM and CCNN, it was
found that SVM provides a more liberal performance as it provides a better CME prediction
performance with a higher number of false alarms.
It is believed that this work is the first to introduce a fully automated computer platform
that can verify the associations between significant flares and CMEs using machine-learning.
This work enables the representation of this association using computerised learning rules
and is a first step towards constructing a fully automated computer platform that would
provide short-term prediction for the possible eruptions of CMEs. These learning rules are
a computerised representation of the experts knowledge that is embedded in the CME and
flare catalogues. Nevertheless, two promising areas for future work are suggested for the
following reasons.
1. The current work has associated only a small percentage of CMEs with significant flares
(M and X-class flares). However, the largest association rate is for CMEs and C-class
flares, as shown in Table 1. In future work the associations for B and C-class flares will
be investigated as well.
2. The predictive performance is not as high as we think it could be because:
(a) From Section 2, it is clear that CMEs can be associated with either flares or erupting filaments/prominences. However, in this study, CMEs were associated only with
flares, and erupting filaments/prominence are not considered. To enhance the prediction accuracy, the CMEs that are associated with eruptive filaments have to be
considered as well. For example, on 21 March 1999 a filament erupted from the
southern boundary of NOAA AR 8494. This filament erupted between 12:35 and
14:30 UT. Its associated CME first appeared in the field of view of LASCO C2 at
15:54 UT, and later in LASCO C3 at 17:42 UT. This CME was not associated with
any significant X-ray flare or H flare, according to the study by Yun-Chun, Le-Ping,
and Li-Heng (2006). This confirms the result provided by the association algorithm
presented here, which identifies this case as a not associated CME.
(b) The initial association between flares and CMEs in the present work depends mainly
on temporal analysis. As explained in Yashiro et al. (2006) this may lead to false
associations. Adding the MPA to refine the associations has improved the CME prediction rates. This suggests that more discriminating features are required to improve
the performance further and to enhance the accuracy of the learning data. Most of the
data provided by the flare and CME catalogues have been considered in the present
study but other sources of data such as images, especially lower corona images obtained by EIT and SXT, should be investigated as well. There is only a small difference in the visibility of frontside and backside CMEs, which makes it very hard to
distinguish between them using only coronagraph observations (Yashiro et al., 2006).
To overcome this difficulty it would be necessary to confirm that a CME originates
from the frontside by checking the lower corona images obtained by EIT and SXT.
This will be investigated in future work. GOES X-ray images will be investigated as
well.
Acknowledgements This work is supported by an EPSRC Grant (GR/T17588/01), which is entitled Image Processing and Machine-learning Techniques for Short-Term Prediction of Solar Activity.
We are grateful to the following providers of data: ESA-NASA SOHO spacecraft team (SOHO is a project
of international cooperation between ESA and NASA), National Geophysical Data Centre (NGDC), Center
for Solar Physics and Space Weather at the Catholic University of America, Naval Research Laboratory, and
the Solar Data Analysis Center (SDAC) at the Goddard Space Flight Center.
273
References
Acir, N., Guzelis, C.: 2004, Expert Syst. Appl. 27, 451.
Borda, R.A.F., Mininni, P.D., Mandrini, C.H., Gomez, D.O., Bauer, O.H., Rovira, M.G.: 2002, Solar Phys.
206, 347.
Cliver, E.W., Hudson, H.S.: 2002, J. Atmos. Solar-Terr. Phys. 64, 231.
Colak, T., Qahwaji, R.: 2007, Adv. Soft Comput. 39, 316.
Distante, C., Ancona, N., Siciliano, P.: 2003, Sens. Actuator B-Chem. 88, 30.
Fawcett, T.: 2006, Pattern Recognit. Lett. 27, 861.
Fukunaga, K.: 1990, Introduction to Statistical Pattern Recognition, Academic, New York.
Gosling, J.T.: 1995, J. Geophys. Res. 100, 7921.
Huang, Z., Chen, H.C., Hsu, C.J., Chen, W.H., Wu, S.S.: 2004, Decis. Support Syst. 37, 543.
Koskinen, H., Tanskanen, E., Pirjola, R., Pulkkinen, A., Dyer, C., Rogers, D., Cannon, P., Menville, J.,
Boscher, D.: 2001, Space Weather Effects Catalogue, Finnish Meteorological Inst., Helsinki.
Kurokawa, H.: 2002, Commun. Res. 49, 49.
Lenz, D.: 2004, Ind. Phys. 9, 18.
Lin, R.P., Hudson, H.S.: 1976, Solar Phys. 50, 153.
Pal, M., Mather, P.M.: 2004, Fut. Gener. Comput. Syst. 20, 1215.
Qahwaji, R., Colak, T.: 2006, Comput. Appl. 13, 9.
Qu, M., Shih, F.Y., Jing, J., Wang, H.M.: 2003, Solar Phys. 217, 157.
Tousey, R.: 1973, Adv. Space Res. 13, 713.
Webb, D.F.: 2000, J. Atmos. Solar-Terr. Phys. 62, 1415.
Wilson, R.M., Hildner, E.: 1984, Solar Phys. 91, 169.
Yashiro, S., Gopalswamy, N., Akiyama, S., Michalek, G., Howard, R.A.: 2005, J. Geophys. Res. 110,
A12S05.
Yashiro, S., Gopalswamy, N., Akiyama, S., Howard, R.A.: 2006, In: 36th COSPAR Scientific Assembly, Beijing, China, 1778.
Yun-Chun, J., Le-Ping, L., Li-Heng, Y.: 2006, Chin. J. Astron. Astrophys. 6, 345.
Automatic Detection and Tracking of Coronal Mass

Ejections in Coronagraph Time Series
O. Olmedo J. Zhang H. Wechsler A. Poland
K. Borne

Abstract We present the current capabilities of a software tool to automatically detect coronal mass ejections (CMEs) based on time series of coronagraph images: the solar eruptive
event detection system (SEEDS). The software developed consists of several modules: preprocessing, detection, tracking, and event cataloging. The detection algorithm is based on
a 2D to 1D projection method, where CMEs are assumed to be bright regions moving radially outward as observed in a running-difference time series. The height, velocity, and
acceleration of the CME are automatically determined. A threshold-segmentation technique
is applied to the individual detections to automatically extract an approximate shape of the
CME leading edge. We have applied this method to a 12-month period of continuous coronagraph images sequence taken at a 20-minute cadence by the Large Angle and Spectrometric
Coronagraph (LASCO) instrument (using the C2 instrument only) onboard the Solar and
Heliospheric Observatory (SOHO) spacecraft. Our automated method, with a high computational efficiency, successfully detected about 75% of the CMEs listed in the CDAW
CME catalog, which was created by using human visual inspection. Furthermore, the tool
picked up about 100% more small-size or anomalous transient coronagraph events that were
ignored by human visual inspection. The output of the software is made available online
at http://spaceweather.gmu.edu/seeds/. The parameters of scientific importance extracted
by the software package are the position angle, angular width, velocity, peak, and average
brightness. Other parameters could easily be added if needed. The identification of CMEs is
known to be somewhat subjective. As our system is further developed, we expect to make
the process significantly more objective.
Keywords Coronal mass ejection Automatic detection
O. Olmedo () J. Zhang H. Wechsler A. Poland K. Borne

George Mason University, Fairfax, VA 22030, USA
e-mail: oolmedo@gmu.edu
J. Zhang
e-mail: jzhang7@gmu.edu
DOI: 10.1007/978-0-387-98154-3_20
275
276
O. Olmedo et al.
1. Introduction
Coronal mass ejections (CMEs) are the largest and most energetic eruptive events that occur at the Sun. They have been observed and studied since the 1970s by several groundbased and space-based coronagraphs. CMEs occur with varying sizes and velocities, and
they are widely accepted to be caused by large-scale magnetic instabilities in the corona
that release a huge amount of energy. Since the launch of the Solar and Heliospheric Observatory (SOHO) spacecraft in 1995, CME observations have mainly been made with the
Large Angle and Spectrometric Coronagraph (LASCO; Brueckner et al., 1995) instrument
onboard. The identification and cataloging of LASCO CMEs is an important task that provides the basic knowledge for further scientific studies. Currently two prominent catalogs
exist: the Naval Research Lab (NRL) catalog (currently online at http://lasco-www.nrl.navy.
mil/index.php?p=content/cmelist under the Preliminary List section of this page) and the
Coordinated Data Analysis Workshop (CDAW) Data Center catalog (currently online at
http://cdaw.gsfc.nasa.gov/CME_list/). The NRL catalog is compiled by LASCO observers
who look through the sequence of LASCO coronagraph images and report on events that
have taken place, on a daily basis. This is a preliminary catalog and provides information on
CME time and approximate position as well as a brief description of the event. The CDAW
catalog provides some measurements of CME properties, including the position angle, the
angular width, and the height of the CME in each individual image. This catalog combines
the height measurements in the time sequence to determine the velocity and acceleration,
which are also provided in the catalog (Yashiro et al., 2004). These measurements are made
by dedicated human operators who look at and then choose the CME height and position on
coronagraph images displayed one by one on a computer screen. This human-based process
is rather time consuming and the events provided and parameters measured are subject to
human bias.
To date several automated CME detection schemes have been described in the literature.
The first of its kind, the Computer Aided CME Tracking (CACTus) software package was
introduced in 2002 (Berghmans, Foing, and Fleck, 2002; Robbrecht and Berghmans, 2004).
It implements the image processing technique of the Hough transform, which finds lines
within a 2D image. CACTus utilizes this transform to identify CMEs as a bright streak in 2D
time height images composed along a specific position angle from a series of coronagraph
images. Currently CACTus has compiled an online catalog that spans from 1996 to the
present day using LASCO C2 and C3 observations (currently online at http://sidc.oma.be/
cactus/). This system operates in near-real time, meaning that detections of CMEs can be
found online on a daily basis.
Several other automated CME detection methods have been proposed but have not
yielded a full CME catalog. One method utilizes LASCO C2 synoptic maps and looks for
signatures of CMEs (Borsier et al., 2005). Liewer et al. (2005) have proposed a scheme that
tracks arc-like features in coronagraph image pairs for use by the Solar TErrestrial RElations Observatory (STEREO)/Sun Earth Connection Coronal and Heliospheric Investigation
(SECCHI) coronagraph instrument to preferentially downlink data containing CMEs. Qu
et al. (2006) use image-segmentation techniques to find CMEs in running-difference and
running-ratio images and further classify CMEs into different categories using machinelearning techniques. We want to emphasize here the importance of the automated detection
in a general sense. As the technologies of detectors and communication advance, the capabilities for observing at higher resolutions and higher cadences yield data sets that are
tremendous in volume. As more NASA missions are flown with advanced capabilities, the
data acquisition rate begins to overcome the rate at which human operators can analyze and
Automatic Detection and Tracking of Coronal Mass Ejections
277
interpret the data. An automated method to analyze data would be valuable, not only for the
present ongoing missions but also and probably more critically for future missions.
In this paper, we present our own software package that automatically detects CMEs.
This is the core of the Solar Eruptive Event Detection System (SEEDS) that we have been
developing. Unlike CACTus, which operates on a time height stack of running-difference
coronagraph images, the SEEDS system uses advanced segmentation techniques that have
the capacity of automatically detecting a CME based on individual LASCO C2 runningdifference coronagraph images in the earliest stage and then tracking the CME in the subsequent frames. In Section 2, we present the methodology. The detection results and a preliminary event catalog are presented in Section 3. A discussion and conclusions are provided in
Section 4.
2. Methodology
Before describing the computational algorithm, we first must state a physics-based CME
recognition criteria. A CME, as observed in white-light coronagraphs, is commonly defined
as a radially outward-propagating structure, with higher density than the surrounding background solar wind. This higher density causes an enhancement in the observed white light,
and in running-difference images the morphological changes can best be observed. The algorithm presented takes advantage of this fact and uses the enhancement to track the CME
as it propagates radially from the Sun. There are three basic modules involved in the algorithm: The first is the preprocessing, which optimizes the input image data for detection; the
second is to make the initial detection by looking for the brightness enhancement; and the
third module is to track the CME in the subsequent running-difference images.
2.1. Preprocessing
The input to the detection system is LASCO C2 LEVEL 0.5 images essentially the raw
data. The preprocessing module involves several steps. First, the input 1024 1024 pixel
image is normalized by the exposure time. It is then passed through a noise filter to suppress
sharp noise features such as stars and cosmic rays. Then a mask is made to indicate areas
of missing blocks in the telemetry. This mask is useful because the missing data blocks will
cause anomalous false signals in the difference image and thus the false detection. For the
same reason, the planets seen in the images are also masked to avoid any false signals in
CME detection. Finally, the image is transformed to a polar coordinate system and then put
through a running-difference filter.
The telemetry of the SOHO spacecraft is such that data are sent in packets. On occasion,
packets will not be transmitted properly and hence create missing blocks in the coronagraph
images. These data blocks within the images are represented with zero pixel value. A bivalence mask of equal size to the input image is created, such that the area of missing data
blocks is represented with zero value and everything else with value one.
An important step in the preprocessing procedure is the application of a noise filter to
the input images. There are two general purposes for this filter: to remove features such as
cosmic rays and random noise, and to remove background stars and correct for planets and
comets. The basis to this filter is the procedure developed by Llebaria, Lamy, and Malburet (1998), which discriminates point-like objects in astronomical images by using surface
curvature. There are three primary steps to this procedure. In the first step the log of the
image is taken and the surface curvature, which parameterizes how the brightness gradient
278
O. Olmedo et al.
changes within the 2D spatial coordinates of the image, is calculated. The surface curvature is parameterized by two principle curvature coefficients (k1 and k2 ) that characterize
the curvature on the surface for each pixel location. Within the 2D [k1 ,k2 ] space, different
areas represent spike-like features, such as stars and planets, or elongated features, such as
cosmic rays. This leads to the second step in the procedure: finding the location of valid
and invalid pixels and making a mask of ones and zeros, equal in size to the input image.
Spike-like and elongated features, found in the input image within the 2D [k1 ,k2 ] space, are
represented with zeros in the mask whereas the rest are represented with ones. The third, and
final, step is to correct the invalid pixels such that they are representative of the background.
This is done with the aid of the mask created and by a local, pyramidal interpolation using a
nonlinear, multiresolution method, where the end result is an image that is devoid of background stars, cosmic rays, and random noise. For the removal of features such as planets and
comets, which tend to be larger than background stars and cosmic rays, a two-step process
is followed, where the end product is a mask that identifies the feature. First, a copy of the
original input image is made and a median filter followed by a smoothing filter is applied to
it; this essentially smoothes out the small features (cosmic rays and background stars) and
leaves the larger ones. The next step involves again calculating the surface curvature, finding
the spike-like and elongated features in the 2D [k1 , k2 ] space, and creating a mask of ones
and zeros. This new mask will contain the location of the larger features (planets, comets,
etc.) and the invalid pixels will be represented with zeros. To more fully cover the larger
features, morphological dilation is applied to this mask. Morphological dilation is an image
processing technique where a binary image is used to dilate or grow objects in size based
on a smaller structuring element (see, for example, Qu et al., 2004, for the mathematical
formulation). Hence, the area covered by the invalid pixels is slightly increased.
To make image processing efficient, the input images, which are in [x, y] Cartesian coordinates, are transformed into a [, r] polar coordinate system, because the features of interest
are intrinsically in polar coordinates owing to the spherical structure of the Sun. This kind
of transformation has been used in other CME-detection algorithms (Robbrecht and Berghmans, 2004; Qu et al., 2006). The transformation takes the [x, y] field of view (FOV) of
the LASCO C2 image, and, starting from the North of the Sun going counterclockwise,
transforms each angle onto a [, r] FOV so that each column in the resulting 360 360
image corresponds to one degree in angle. The radial FOV in this image corresponds to 360
discrete points between 2.2 and 6.2 solar radii.
To enhance moving features, a running-difference sequence of the processed images is
made using the following equation:

,
ui = ni ni1 (n i /n i1 )
t
(1)
where ui is the running-difference image, n is the mean of the pixels in the entire FOV of
the image n, t is the time difference between the images (in minutes), is a constant, and
the subscript i denotes the current image and i 1 denotes the previous image. The constant
is set to approximately the smallest possible time difference (t ) between any pair of
images, typically equal to a value between five and ten minutes. This normalized difference
ensures that the mean of the new image (ui ) will be effectively zero. After the differencing
is performed, the masks of missing blocks and planets are applied. The masks are first polar
transformed, then the difference image is multiplied by them such that the missing block
areas as well as the planets become zero-value pixels. Figure 1 shows the preprocessing of
an input LASCO C2 image. In this figure, panel (a) shows the input LASCO C2 input image
processed with the exposure-time normalization and the noise filer; panel (b) shows the
279
Figure 1 An image of the CME that occurred on 12 September 2002 and its polar representation. (a) The
input LASCO C2 image processed with the exposure time normalization and the noise filtering; the white
arrow points to the CME in the FOV. (b) The polar representation of the running-difference image made from
the image in panel (a). The bright region seen in panel (b) is the corresponding CME seen in panel (a).
corresponding polar representation of the running-difference image made from the image in
panel (a).
2.2. Initial Detection
The initial CME detection is made on the polar-transformed running-difference sequence
images. First, the 2D images are projected to one dimension along the angular axis such
that the values of the radial pixels in each degree column in the image are summed; this
effectively measures the total brightness of signals along one particular angular degree. Observationally, a CME is best observed in running-difference images. This type of image
essentially removes static or quasi-static background features such as coronal streamers and
enhances features that change at faster time scales. A CME in a running-difference image
will appear as a bright leading-edge enhancement (positive pixel values) followed by a dark
area deficient in brightness (negative pixel values), and the background will appear gray,
indicating zero-change pixels. Since the detection concerns the bright leading edge, only the
positive pixels are counted when making the 1D projection. Hence, the 2D enhancement is
seen as a peak in the 1D intensity profile. Also, by excluding the negative enhancement, the
positive enhancement becomes more outstanding in the projection profile. The following is
a mathematical formulation of the projection:
p =
1
u(r, ),
c r
u(r, ) > 0,
(2)
where c is the number of positive pixels for each given angle . An example of the projection can be seen in Figure 2.
To make the detection, a threshold (T 1) is chosen to identify the peaks in the 1D projection based on the standard deviation (p ) and mean (p ) of the projection (p ):
T 1 = N1 p + p .
(3)
Here N1 is a number chosen to set the threshold level. This value remains constant throughout the detection and is determined through experimental methods. The identified peak angles above the threshold are called the core angles. These core angles correspond to the
central and the brightest part of a CME.
N 1 (and thus the threshold T 1) is a critical number in the performance of the detection
algorithm. It is often chosen to be between two and four. The lower the number, the more
280
O. Olmedo et al.
Figure 2 The intensity profile

along the angular axis showing
the 1D projection of the CME
image. Only positive pixels along
the radial axis are used. This
profile effectively indicates the
angular positions of a CME when
it is present.
sensitive the detection and the more false detections occur. In contrast, the higher the number, the lower the sensitivity of the detection, which results in more missed events. Currently
we are still experimenting to find an optimal number. An important issue that arises at this
point, but is not fully addressed in our detection algorithm, is the emergence of multiple
CMEs at the same time, especially if there were a bright and a faint CME. Presently, because the threshold (T 1) is calculated with the standard deviation and mean of the whole
projection (p ), what will happen is that the signal of the faint CME will be overpowered
by the bright CME and not detected. This is not a big problem: If the brightness of multiple emerging CMEs are close in magnitude, they will in general all be detected. We have
been experimenting with a few other methods to try to overcome the issue of bright and
faint CMEs simultaneously emerging but have not yet successfully come to a conclusion as
to which may be the best. For example, we could take the log of the projection, which is
similar to what Llebaria, Lamy, and Malburet (1998) have done for discriminating pointlike sources, where using the log scale is justified because the range of curvatures become
narrower and improves the detection of faint signals. Another option could be to use local thresholds, splitting the 360 points of the projection into several sections, computing a
threshold for each, then piecing all of the detections together to make a final determination
of the core angles. A promising solution proposed by Qu et al. (2006) is to use not only
running-difference images but also running-ratio images. This has the advantage that weak
CMEs in the ratio images will appear closer in brightness to the bright CMEs, but it has the
disadvantage that noise will also be enhanced, causing overflow errors. To overcome this,
Qu et al. (2006) have proposed that only the pixels greater than the median in the reference
image (ni ) be used to make the ratio image. Future investigation is needed to find the best
method to overcome this issue.
Figure 3a shows, with two vertical lines, the core angles found. These angles only yield
a small fraction of the CMEs total angular width. To find the full angular width of the CME
region growing is applied. Region growing is the procedure used to expand a small region
found in a multidimensional space into a larger region. In our case, region growing takes two
inputs: the maximum and minimum values for the core region to be grown to. This means
that the CME angles are widened to include all of the pixels that are within the maximum
and the minimum range; hence the starting and ending position angles (PAs) of the full-sized
CME are determined (Figure 3b). For our application, we have chosen the maximum input
281
Figure 3 Determination of
CME angular position, angular
size, and heights. (a) The CME
core (or brightest) angels within
the two vertical lines. (b) The full
CME angles after the
region-growing method is
applied. (c) The CME heights at
the half-max-lead, maximum,
and the half-max-follow
indicated by the three horizontal
lines, respectively; the white box
at the bottom corresponds to the
trailing box, where exponential
suppression is applied. The right
side of panel (c) shows the
intensity profile of the 1D
projection along the radial
direction.
value to be the maximum value in the core-angle range and the minimum is set to a value
similar to a reduced T 1, T 2, where instead of the mean and standard deviation of the whole
projection, the mean and standard deviation of only the values outside of the core-angle
range are used. In many cases, especially with limb CMEs, it is observed that the CME will
deflect the streamer structure of the Sun. This type of disturbance has the appearance of a
wispy-like structure and is caused by the CME pushing the material of the streamer away
from it as it expands during its radial propagation. In the region-growing step just discussed,
it is difficult to distinguish between what is and what is not part of the CME structure in
the 1D profile; therefore on some occasions the streamer deflection will become part of the
detected angular width of the CME. Another disadvantage to be discussed is the emergence
of two CMEs at approximately the same time and angular position such that the full angular
widths of both CMEs overlap. After the region-growing step, since the full angular widths
overlap, what will happen is that the algorithm will mistake the multiple CMEs as one CME
and attempt to track it as though it were a single CME. Currently, no solution to this problem
has been found and further investigation is needed.
2.3. Tracking
To identify a newly emerging CME, the CME must be seen to move outward in at least two
running-difference images. This condition is also set by Yashiro et al. (2004) to define a
newly emerging CME. This condition is very useful because very often it is found that when
only one detection is found with SEEDS it is most likely a noise feature and not a feature
that is evolving with time. Currently SEEDS cannot distinguish between a noise feature or
an object with CME-like features in a single detection. Also, the actual number of CMEs
that are found with such a high velocity as to only be seen in one LASCO C2 frame is
very small. In the near future, once the SEEDS system has been adapted to accept STEREO
coronagraph images (from the COR2 coronagraph onboard STEREO), this should no longer
be a problem for extremely fast CMEs, because the FOV for COR2 is much larger than that
of C2.
282
O. Olmedo et al.
To determine the presence of a newly emerging CME, we look at running-difference

images i and i + 1, assuming that a detection has been seen in image i. If the detection in
image i overlaps with a detection in image i + 1 then it is assumed that image i is the initial
detection of a newly emerging CME, where overlap here refers to the overlapping of PAs
detected between the two images. Once the PAs of the initial detection are established, the
leading edge and height can be found.
Determination of the leading edge begins with performing another projection along the
radial direction. This projection is made within the starting and ending PAs for each CME
seen in the FOV of the image. To do this projection, the pixels within the PAs are averaged
along the component such that one is left with a 1D array of 360 points corresponding to
the 360 discrete points in the radial FOV of the image. This is written as
Pr =
=PA end
1
u(r, ),
w =PA start
(4)
where w is the angular width of the CME detected (equal to the end PA minus the start PA).
The peak of this array is found and is defined as the height of CME maximum brightness,
max-height (see Figure 3c). The two half-maximum points that are radially higher and lower
than the max-height point are calculated and are called the half-max-lead for the point that
is higher and half-max-follow for the point that is lower. At present, the half-max-lead is
thought of as the leading edge of the CME and is the point used for making calculations of
velocity and acceleration. With the two half-max points and the starting and ending PAs we
define a region of interest, which we call the CME box. Figure 3c shows a detection where
the two half-maximas and the max-height of the radial projection are shown.
At this point, the initial detection of the CME has been established; what follows is
the tracking of its path in subsequent running-difference images. The running-difference
image that follows the image with the initial detection now becomes the subject in focus and
becomes the current image. Starting and ending PAs have already been determined for this
image through the initial detection module. They are compared with the starting and ending
PAs of the previous image and are combined such that the starting PA is the minimum of the
two starting PAs and the ending PA is the maximum of the two ending PAs. The next thing
to determine is the leading edge of the CME. Radial projection is again performed, as was
previously done, but for this projection the lower limit becomes the half-max-follow that was
formerly determined in the previous image instead of the lower radial boundary of the FOV
of the image. The max-height within this projection is found, as well as the half-max-lead
and half-max-follow. Two criteria are applied at this point to the position of the max-height
and half-max-lead. First, the max-height is not allowed to be below the max-height of the
previous detection. Second, the distance between the max-height and half-max-lead is to
be greater than or equal to the distance between the max-height and half-max-lead of the
previous detection. The second criterion is used to avoid the max-height or half-max-lead
being found below the previous detection. This assures that the leading-edge detection in
a sequence of images always increases in height with time and never backtracks, as that
would not be realistic in the tracking of a CME, which is a radially outward moving object.
This is not to say that backtracking material, known as inflow, does not exist it can be seen
in LASCO C2 coronagraphs up to 5.5 solar radii (Sheeley and Wang, 2002). This project
only focuses on outward-moving CMEs. The criteria also assumes that the CME is a radially
expanding object where it is thought that the distance between the max-height and half-maxlead should increase with time.
283
To continue tracking the CME, the steps just described in the previous paragraph are
applied to successive images in chronological order. The current running-difference image
becomes the previous running-difference image and the next running-difference image becomes the current running-difference image and so forth. These steps continue until the
half-max-lead height is found to be outside of the FOV of the image or until no detection
is found in the current image because the CME becomes too diffused. Finding the halfmax-lead height outside of the FOV is possible because of the assumption that the distance
between the max-height and half-max-lead height in the current detection must be greater
than the distance in the previous detection. In the case that a temporal gap is found, considered to be more than about two hours, the algorithm stops all tracking, outputs the data for
any CME tracked, and starts over at the time after the gap.
Following a major CME, very often there exists a period of continuous outflow of small
blob-like outflow features, which could be detected as other CMEs. To suppress this effect
we first define a trailing box (see the white box in Figure 3). Within the trailing box, an
exponential function is applied. The following equation shows the function used to suppress
CME outflow features:

p0 = p0 1 e(T T0 )/Tc ,
(5)
where p0 are the pixel values within the trailing area, p0 indicates the exponentially suppressed pixel values, T0 is the onset time of the CME, and Tc is a time constant. This function
is applied for several hours after the onset of the CME at T0 , and an optimal time was found
to be between five and six hours. Further investigation into the Tc parameter is needed to
find an optimal value; with current experimentation we find a value in the range between
three and five. The value of Tc , which determines the rate of the exponential suppression,
is highly correlated with the velocity with which the CME leaves the FOV as well as the
extent, or the radial length, of the CME.
2.4. Leading-Edge Determination and Visualization
In this section we describe the method used to find an approximate leading-edge shape of
the CME. This is done for visualization purposes to help guide the eye and show how the
CME is evolving; it could also be used to measure velocity. The biggest difficulty in this
task is the fact that CMEs tend to appear as diffuse clouds that have no clear boundary. In
our application we implement a simple segmentation technique based on a threshold to find
the approximate shape of the leading edge.
There are three main steps in this calculation. We begin with the CME detected within
the CME box. The first step is to extend the box and set a new search area for the leading
edge. The new upper boundary is set to a height that is half the height of the CME box (half
the distance between the half maximums) above the half-max-lead position. Within this new
region, a CME area is segmented by scrolling through each position angle within the box
and setting all the pixels that are half of the maximum within each angle to value one and
the rest to value zero (Figure 4a). Mathematical morphology operators opening and closing,
based on set theory, as used in image processing (e.g., Qu et al., 2004, 2006), are applied to
the segmented image to remove noise and connect disjoint pieces. The pixels on the upper
border of the area are chosen as the leading-edge outline of the CME; these leading-edge
points are linearly smoothed and projected back to the coronagraph image (Figure 4b).
2.5. Example Case
As an example case, the event that occurred on 12 September 2002 at 21:12 UT is shown
in Figure 5. We show the CME evolution in four difference images with the leading edge
284
O. Olmedo et al.
Figure 4 The CME leading

edge. (a) The segmented CME
within the segmentation area.
The segmentation area is
contained within the vertical lines
and the horizontal lines excluding
the half-max-lead line, which is
shown only for reference
purposes. (b) The pixels that
define the leading edge of the
CME along the position angles.
Figure 5 The image sequence showing the automatic tracking of the event on 12 September 2002 at 21:12
UT. The black dotted lines indicate the detected leading edges of the CME overlaid on the running-difference
images.
overplotted with a black dotted line. Figure 6 shows the time height profile for this example CME measured with the automatic detection as compared with the CDAW catalog
entry from human measurement. The + line represents the height time plot from CDAW
but using only C2 observations, the is the highest peak found in the leading-edge segmentation as found with SEEDS, and the is the half-max-lead. We calculate a velocity using the leading-edge segmentation and half-max-lead and find values of 543.4 and
474.2 km s1 , respectively. For the C2 measurements in the CDAW catalog we found a
velocity of 484.9 km s1 . The angular width found by SEEDS was 105 and for CDAW
123 . We can see that the SEEDS automatic measurements for this event are very similar
to the manual measurements in terms of measuring the angular width and velocity of the
CME. The height, however, seems to fall slightly short of the manual measurement. This
discrepancy shows the difference between a human and computer choosing the position of
the height for the CME. But both capture the fact that the CME is propagating in a radial
direction and they find a velocity that is approximately the same. It can also be noticed that
the leading-edge-segmentation and the half-max-lead heights are diverging with increasing
time. This discrepancy may be due to the expansion of the CME and could possibly be a
measure of how the CME expands with time; this finding merits further investigation outside
of the scope of this paper.
3. Results and Validation

3.1. Results
The output of the SEEDS system that quantifies CMEs includes the PA, angular width,
and height, which are further used to calculate the velocity and acceleration as a function
285
Figure 6 The comparative

height time plots of the CME on
12 September 2002 at 21:12 UT.
The measurements of the
manual-based CDAW, the
automated SEEDS at the highest
leading-edge-segmentation
position, and at the half-max-lead
are denoted with +, , and
symbols, respectively. The
straight lines show the linear fit
to the height time measurement,
yielding the CME velocity.
of time. There are many other measurements that are made with this system besides the
basic CME parameters. As mentioned before, the parameters that make up the CME box
include the PAs and the half-max-lead and half-max-follow. Five more parameters that are
associated with the area contained within the CME box are the mean, standard deviation,
maximum, and minimum brightness values within the CME box. The final parameter comes
from taking the difference between the detection box as seen in the current difference image
and that in the previous difference image, and then taking the mean of the pixels within this
difference. We call this last parameter the mean box difference, which effectively indicates
the brightness variation of a given CME. Some of these proposed measurements may not
have obvious scientific value but they have been discussed to show that our technique is
flexible and has the capability to extract many measurements. We believe that many of these
measurements may be useful as CME signatures and for the possible classification of events.
This type of classification, for example, has been explored by Qu et al. (2006), who used a
list of measurements and classified CMEs into three categories: strong, medium, and weak.
3.2. Validation and Comparison
The SEEDS CME detection system was tested by using data for 2002 and compared with
the CDAW catalog for the same period. Statistics were compiled on latitude (converted from
central PA), angular width, and velocity. The data were run at monthly intervals to simplify
the comparison between the two catalogs. The monthly text version of the CDAW catalog
for 2002 was used. Because of the large number of events, a manual comparison between
the catalogs would be difficult, so an automated method was implemented to compare our
results with those of the CDAW catalog. A criterion was established to compare entries
between the two catalogs. It involved looking at the starting and ending PAs as well as the
initial time of detection. Because starting and ending PAs are not available from the CDAW
286
O. Olmedo et al.
catalog they must be calculated. This was accomplished by subtracting one-half the angular
width from the central PA, to find the starting PA, and adding one-half the angular width
to the central PA, to find the ending PA. To do the comparison we first look at the time of
the initial detection of the entry in the SEEDS catalog and find all detections in the CDAW
catalog that fall within a range of one and a half hours either before or after the time of
detection, giving us a three-hour window. The reason for this window has to do with the
nature of the SEEDS scheme in that it may sometimes detect an event either several frames
early or several frames late. So we stipulate that a three-hour window is adequate for making
our comparison. We next look at the PAs and find the unique entry in the CDAW catalog
within the time window that overlaps the SEEDSs detection. If no entry exists in the CDAW
catalog then it is thought of as either a new CME detection that had gone unreported in the
CDAW catalog or an erroneous detection that may have been caused by a large streamer
disturbance or noise that had not been removed properly by the detection processes.
This analysis yields the following results. For 2002, SEEDS finds a total of 4045 events,
whereas the CDAW catalog lists 1389 CMEs. However, this number of CDAW catalog
CMEs was limited to those seen with C2 measurements and that had a quality index of
greater than or equal to one, where the quality index is a visual measure of how well defined the CME leading edge is. This value ranges between zero and five, with zero being
ill-defined and five begin excellent. It is found that 1306 SEEDS events corresponded to
events in the CDAW catalog. On occasion SEEDS will not detect a CME as whole; for
example, a very wide angle or halo CME may be detected in several pieces and therefore
SEEDS will report that multiple CMEs have been detected when in fact the pieces belong to
a single event. Another example of when this occurs is if a CME in the middle of tracking
is not detected in one frame but is then redetected in the next frame and then retracked.
Further study is needed to associate these two events temporally such that SEEDS will only
report one event. We found that, of the 1306 SEEDS events that correspond to events in the
CDAW catalog, 281 were disjoint events that should have corresponded to a single event,
leaving 1025 events that mapped from SEEDS to the CDAW catalog. This leaves us with
a true positive rate of approximately 74%, assuming the CDAW catalog is completely correct. Nonetheless, SEEDS detects over twice as many events as listed in the CDAW catalog,
which is not unusual since the automated system picks up many outflow, narrow, or small
events that are ignored by human observers. Upon inspection of the CMEs that were missed,
it is concluded that they tend to be ones that are weak or faint. This is not uncommon with automated CME detection; for example, Qu et al. (2006) reported that in their case-study time
period, using their CME detection algorithm, all of the CMEs missed were weak events.
They also made a comparison with the CACTus catalog, using the same case-study time
period, and found that the CMEs missed by CACTus were also weak events.
Figure 7 shows scatter plots of the basic parameters for the associated CMEs between
SEEDS and CDAW. Figure 7a shows the comparison between angular widths, indicating
that the widths from SEEDS are in general narrower than CDAW measurements, especially
with increasing angular extent. This tells us that perhaps the region-growing maximum and
minimum thresholds need to be further investigated since the region-growing step in the
detection is what determines the full angular width of the CME. Figure 7b shows that latitude measurements of SEEDS very closely follow CDAW measurements. Finally, Figure 7c
shows the velocity comparison, which shows a broad scatter. This type of scatter indicates
that much more work is needed on how CMEs are tracked and how the position of the
leading edge of the event is determined. This scatter could also be due to events that were
detected as disjoint events. For example, one detection may be part of the core of the CME,
expanding in the radial direction, and another disjoint detection, of the same CME, is expanding more in the latitudinal direction; therefore it would appear to have a lower velocity.
287
Figure 7 Comparison of CME parameters measured in the automated SEEDS system and the manual CDAW
measurements. Panels (a), (b), and (c) show the width, latitude, and velocity, respectively. The straight line in
the panels is a diagonal line with slope of one.
Figure 8 Histograms of CME
parameters from the SEEDS
method for all LASCO C2
observations in 2002. From top to
bottom the three panels are for
CME width, latitude
(equivalently position angle), and
velocity, respectively. The two
vertical lines in panel (b) show
the critical latitudes that bound
the area where 80% of the CMEs
are to be found.
Histograms of the SEEDS detections of angular width, latitude, and velocity can be seen
in Figure 8. The mean and median of the angular width for normal CMEs within the distribution were found to be 40 and 34 , respectively; normal CMEs are the ones that are
found to have an angular width greater than 20 and less than 120 (Yashiro et al., 2004).
The mean and median of the velocity distribution were found to be 292 and 233 km s1 , re-
288
O. Olmedo et al.
spectively. The critical latitude angles were calculated for the latitude distribution and were
found to be 68 and 51 . The critical latitudes represent the latitudinal positions on the Sun
where 80% of all CME emerge in a given time period (Yashiro et al., 2004). Comparing the
statistical properties of SEEDS for 2002 with those reported by the CDAW catalog we find
similar results. The mean and median of the angular width for normal CMEs were found to
be 53 and 49 , respectively, the mean and median for the velocity distribution were found
to be 521 and 468 km s1 , respectively, and the critical latitude angles were found to be
59 and 51 . The largest discrepancies in the comparison of statistical results are the mean
and median of the velocity. The discrepancies could be attributed to the fact that SEEDS has
detected many anomalous small objects with much slower velocities. At this point it is important to note the sources of error in the histograms of the CME parameters. As previously
stated, many small anomalous detections are made; these anomalous features may include
things such as streamer deflections or streamers that are bright enough that their signal can
be seen in the 1D projection (p ). Other objects may include blob-like features that may or
may not be associated with either pre or post CME outflow but that are brighter than the
background. Finally, especially during times when no CME is erupting, areas of increased
brightness may be observed in the 1D projection (p ). These bright regions typically may
not be associated with any type of solar feature, but if such a region in the same angular position is seen in multiple frames, the algorithm will output an erroneous detection. Because
of these anomalous features that may be detected, future statistical studies using the output
of SEEDS must be handled with extreme care.

We have implemented a novel system that is capable of detecting, tracking, and cataloging
CMEs. The basic algorithm steps are: i) preprocessing, ii) making the initial detection, iii)
tracking the CME in coronagraph time series, iv) calculating the velocity and acceleration
and outputting the measured parameters, and v) finally, with the measurements made, compiling a catalog of CME events and making it available online.
Automated methods such as CACTus and SEEDS have a tendency to report more than
twice as many CMEs as are identified in catalogs made by a human observer. This major
difference is primarily due to a sensitive balance between internal thresholds of the algorithm and the goal of trying to include every CME detected by a human operator. The excess CMEs detected are primarily CME outflow, blob-like features, streamer disturbance, or
noise. These kinds of small features have been reported and studied (e.g., Wang et al., 2000;
Bemporad et al., 2005). It is important to note that because of the diffused, cloud-like nature
of CMEs, any detection, be it by human or computer algorithm, will have some subjectivity.
For example, a CME, as detected by a human or computer algorithm, may be detected differently by another human or computer algorithm, such that the CME may be seen as two
CMEs, have different angular width and velocity, or have different starting and ending times
(Qu et al., 2006). Although a computer algorithm attempts to reduce human subjectivity, the
algorithm itself is based on a set of rules and internal thresholds. Through experimentation
with the rules and thresholds it may be possible create a catalog, although it may not contain
100% of the CMEs that emerge from the Sun, which is at least consistent with a given set of
rules. Currently no automated system is capable of detecting 100% of all CMEs reported in
the CDAW catalog. The events missed tend to be weak in nature. Further investigation into
these weak events should be made to more fully understand how to detect and track them.
Our automated system is capable of creating a catalog with a much larger set of parameters that describe a CME than human observers currently do. At each time of a CME,
289
we will provide the CME height of the leading edge, the starting position angle, the central
position angle, the ending position angle, velocity, and acceleration, as well as the various
measurements mentioned in Section 3.1. The output of our detection system is available online at http://spaceweather.gmu.edu/seeds/. On the site one will find a calendar that contains
links to all of the data that have been processed. Height time plots with linear and secondorder fits are also available as image files. Finally, we also provide Java script movies of
each detection that contains the running-difference images, with the projected leading edge
as shown in Figure 5 as well as a second panel that contains the direct coronagraph image.
We are currently working on many improvements to the system. For example, preliminary tests using data-mining techniques have yielded promising results in being able to distinguish a detection from an erroneous detection. This type of detection is analogous to the
work done by Fernandez-Borda et al. (2002) and Qu et al. (2003), where data-mining techniques were used in the detection of solar flares, by Qahwaji and Colak (2007), who used
machine learning to make short-term predictions of solar flares, and by Qu et al. (2006),
who used data mining to detect and classify CMEs. We are also working toward developing SEEDS to make detections in near-real time, meaning that as soon as the data from the
LASCO C2 instrument are made available, SEEDS will automatically retrieve them, apply
the detection algorithm, and finally make the results available in our online catalog on a
daily basis. We also plan to apply the SEEDS system to STEREO/SECCHI coronagraph
data.
Acknowledgements SOHO is a project of international cooperation between ESA and NASA. The
SOHO/LASCO data used here are produced by a consortium of the Naval Research Laboratory (USA), MaxPlanck-Institute fr Aeronomie (Germany), Laboratoire dAstronomie Spatiale (France), and the University
of Birmingham (UK). We acknowledge use of the CME catalog generated and maintained at the CDAW Data
Center by NASA and The Catholic University of America in cooperation with the Naval Research Laboratory.
O.O. thanks Robin Colaninno for valuable discussions on LASCO data. The authors acknowledge support
from NASA AISRP Grant No. NNG05GB02G. J.Z., and O.O. acknowledge support from NSF SHINE Grant
No. ATM-0454613. J.Z. also acknowledges NASA Grant Nos. NNG04GN36G and NNG05GG19G.
References
Bemporad, A., Sterling, A.C., Moore, R.L., Poletto, G.: 2005, Astrophys. J. 635, L189.
Berghmans, D., Foing, B.H., Fleck, B.: 2002. In: SOHO 11, From Solar Min to Max: Half a Solar Cycle with
SOHO SP-508, ESA, Noordwijk, 437.
Borsier, Y., Llebariaf, A., Goudail, F., Lamy, P., Robelus, S.: 2005, SPIE 5901, 13.
C.J.: 1995, Solar Phys. 162, 357.
Fernandez-Borda, R.A., Mininni, P.D., Mandrini, C.H., Gomez, D.O., Bauer, O.H., Rovira, M.G.: 2002, Solar
Phys. 206, 347.
Liewer, P.C., DeJong, E.M., Hall, J.R., Lorre, J.J.: 2005, Eos Trans. AGU 86(52), Fall Meet. Suppl., Abstract
SH14A-01.
Llebaria, A., Lamy, P., Malburet, P.: 1998, Astron. Astrophys. Suppl. Ser. 127, 587.
Qu, M., Shih, F.Y., Jing, J., Wang, H.: 2003, Solar Phys. 217, 157.
Robbrecht, E., Berghmans, D.: 2004, Astron. Astrophys. 425, 1097.
Sheeley, N.R. Jr., Wang, Y.-M.: 2002, Astrophys. J. 579, 874.
Wang, Y.-M., Sheeley, N.R. Jr., Socker, D.G., Howard, R.A., Rich, N.B.: 2000, J. Geophys. Res. 105, 25133.
Yashiro, S., Gopalswamy, N., Michalek, G., St. Cyr, O.C., Plunkett, S.P., Rich, N.B., Howard, R.A.: 2004,
J. Geophys. Res. 109, A7105.

Solar Image Analysis and Visualization

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Solar Image Analysis and Visualization

Uploaded by

Copyright:

Available Formats

J. Ireland C.A.

Solar Image Analysis

Previously published in Solar Physics Volume 248, Issue 2, 2008

Cover illustration: Image of automated detection and segmentation of CME.

Printed on acid-free paper.

Preface: A Topical Issue on Solar Image Analysis and Visualization

FESTIVAL: A Multiscale Visualization Tool for Solar Imaging Data

Segmentation of Loops from Coronal EUV Images

Preface: A Topical Issue on Solar Image Analysis

Originally published in the journal Solar Physics, Volume 248, No 2, 211211.

FESTIVAL: A Multiscale Visualization Tool for Solar

Originally published in the journal Solar Physics, Volume 248, No 2, 213224.

FESTIVAL: A Multiscale Visualization Tool for Solar Imaging Data

Figure 1 Illustration of the data

FESTIVAL: A Multiscale Visualization Tool for Solar Imaging Data

FESTIVAL: A Multiscale Visualization Tool for Solar Imaging Data

2.5. The Interface

FESTIVAL: A Multiscale Visualization Tool for Solar Imaging Data

FESTIVAL: A Multiscale Visualization Tool for Solar Imaging Data

Visualization of Distributed Solar Data and Metadata

Originally published in the journal Solar Physics, Volume 248, No 2, 225232.

Visualization of Distributed Solar Data and Metadata

Visualization of Distributed Solar Data and Metadata

Figure 3 Layout of the SWB

Visualization of Distributed Solar Data and Metadata

Widespread Occurrence of Trenching Patterns

Originally published in the journal Solar Physics, Volume 248, No 2, 233245.

Abstract Time-averaged series of granulation images are analysed using COLIBRI,

A.V. Getling, A.A. Buchnev

Widespread Occurrence of Trenching Patterns in the Granulation Field

the probability of finding a target supergranule or granule (identified by its barycentre) at

A.V. Getling, A.A. Buchnev

Widespread Occurrence of Trenching Patterns in the Granulation Field

The distance , which is everywhere measured in pixels, is an input parameter of the

A.V. Getling, A.A. Buchnev

Widespread Occurrence of Trenching Patterns in the Granulation Field

Figure 2 An illustration to the

A.V. Getling, A.A. Buchnev

Widespread Occurrence of Trenching Patterns in the Granulation Field

lineaments disappeared; however, objects exceeding 20 pixels in lengths were preserved

A.V. Getling, A.A. Buchnev

lineaments near the edges of the field of view.

Widespread Occurrence of Trenching Patterns in the Granulation Field

A.V. Getling, A.A. Buchnev

Widespread Occurrence of Trenching Patterns in the Granulation Field

Principal Components and Independent Component

Originally published in the journal Solar Physics, Volume 248, No 2, 247261.

A.C. Cadavid et al.

2. Principal Components Analysis

Principal Components and Independent Component Analysis

l elT (t)l (x).

A.C. Cadavid et al.

Figure 2 The leading 20

Figure 3 The first two empirical

2.2. Application of PCA to Heliospheric and Solar Data

Principal Components and Independent Component Analysis

Figure 4 The first two principal

A.C. Cadavid et al.

Figure 5 Three leading EOFs of

Principal Components and Independent Component Analysis

Figure 6 Time series for the

Figure 7 Phase lag versus time

3. Independent Component Analysis

A.C. Cadavid et al.