GEOPHYSICS, VOL. 73, NO. 2 共MARCH-APRIL 2008兲; P. B11–B23, 14 FIGS., 1 TABLE.

10.1190/1.2825408

Case History
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Visualization of active faults using geometric attributes of 3D GPR data:

An example from the Alpine Fault Zone, New Zealand

Alastair F. McClymont1, Alan G. Green1, Rita Streich2, Heinrich Horstmeyer1, Jens Tronicke3,
David C. Nobes4, Jarg Pettinga4, Jocelyn Campbell4, and Robert Langridge5

ABSTRACT
Three-dimensional ground-penetrating radar 共GPR兲 data are
routinely acquired for diverse geologic, hydrogeologic, archeo-
logical, and civil engineering purposes. Interpretations of these
data are invariably based on subjective analyses of reflection pat-
terns. Such analyses are heavily dependent on interpreter exper-
tise and experience. Using data acquired across gravel units over-
lying the Alpine Fault Zone in New Zealand, we demonstrate the
utility of various geometric attributes in reducing the subjectivity
of 3D GPR data analysis. We use a coherence-based technique to
compute the coherency, azimuth, and dip attributes and a gray-
level co-occurrence matrix 共GLCM兲 method to compute the tex-
ture-based energy, entropy, homogeneity, and contrast attributes.
A selection of the GPR attribute volumes allows us to highlight
key aspects of the fault zone and observe important features not
apparent in the standard images. This selection also provides in-
formation that improves our understanding of gravel deposition
and tectonic structures at the study site.Anew depositional/struc-
tural model largely based on the results of our analysis of GPR at-
tributes includes four distinct gravel units deposited in three
phases and a well-defined fault trace. This fault trace coincides
with a zone of stratal disruption and shearing bound on one side
by upward-tilted to synclinally folded stratified gravels and on
the other side by moderately dipping stratified alluvial-fan grav-
els that could have been affected by lateral fault drag. When used
in tandem, the coherence- and texture-based attribute volumes
can significantly improve the efficiency and quality of 3D GPR
interpretation, especially for complex data collected across ac-
tive fault zones.

INTRODUCTION
Attributes used extensively for interpreting 3D reflection seis-
mic data are relatively uncommon in 3D ground-penetrating radar
共GPR兲 analyses. Recent advances in acquiring and processing 3D
GPR data have created new opportunities for mapping the shallow
subsurface 共Lehmann and Green, 1999, 2000; Young and Lord,
2002; Streich et al., 2006, 2007; Grasmueck and Viggiano, 2007;
Streich and van der Kruk, 2007兲. Although correctly migrated GPR
data volumes can provide vivid images of complex geologic units
viewed from arbitrary directions, our knowledge of many subsur-
face features is limited by our inability to delineate and describe the
typically diverse reflection patterns quantitatively.
Most analyses of 3D GPR data are based on subjective assess-
ments of reflection-pattern character. Such assessments are heavily
dependent on interpreter expertise and experience. Over the past few
decades, similar issues in reflection seismology have been addressed
by the introduction of attributes 共Chopra and Marfurt, 2005兲. A seis-
mic attribute is a quantitative measure of a seismic characteristic of
interest. The extensive use of attributes has significantly improved
the quality and efficiency of 3D seismic interpretation in the petro-
leum industry.

Manuscript received by the Editor 17 May 2007; revised manuscript received 25 July 2007; published online 10 January 2008.
1
Institute of Geophysics, ETH 共Swiss Federal Institute of Technology兲, Zurich, Switzerland. E-mail: amcclymont@aug.ig.erdw.ethz.ch; alan@
aug.ig.erdw..ethz.ch; heinrich@aug.ig.erdw.ethz.ch.
2
Formerly Institute of Geophysics, ETH 共Swiss Federal Institute of Technology兲, Zurich, Switzerland. Presently SINTEF Petroleum Research, S. P. Trond-
heim, Norway. E-mail: rita.streich@iku.sintef.no.
3
Formerly Institute of Geophysics, ETH 共Swiss Federal Institute of Technology兲, Zurich, Switzerland. Presently University of Potsdam, Institute for Geo-
sciences, Golm, Germany. E-mail: jens@geo.uni-potsdam.de.
4
University of Canterbury, Department of Geological Sciences, Christchurch, New Zealand. E-mail: david.nobes@canterbury.ac.nz; jarg.pettinga@
canterbury.ac.nz; jocelyn.campbell@canterbury.ac.nz.
5
GNS Science, Lower Hutt, New Zealand. E-mail: r.langridge@gns.cri.nz.
© 2008 Society of Exploration Geophysicists. All rights reserved.

B11
 B12 McClymont et al.

Most seismic investigations are designed to achieve one or more
of the following goals: 共1兲 determine the distribution of key physical
properties, 共2兲 map the boundaries and identify the lithologies/facies
of geologic units, 共3兲 estimate the locations and quantities of fluids,
共4兲 delineate potential fluid-flow pathways, and 共5兲 reconstruct the
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
deep oil and gas deposits. Moreover, the form and appearance of typ-
ical GPR and reflection seismic data sets are very similar. Yet, at-
tributes based on 3D GPR data have only been reported in a relative-
ly small number of publications 共Grasmueck, 1996; Young et al.,
1997; Sénéchal et al., 2000; Corbeanu et al., 2002; Tronicke et al.,

geologic histories of investigated regions. Careful analysis and in-
terpretation of processed data are necessary to meet these objectives.
To complement the expertise and experience of the interpreter and to
increase the efficiency of analyzing large volumes of 3D seismic
data, a variety of mathematically defined attributes are available. At-
tributes can emphasize features distinguished by specified charac-
teristics 共e.g., reflection amplitudes, frequencies, dips, azimuths,
and continuity, and general texture and form兲 or to extract informa-
tion that is not obvious in standard images. Some attributes are deter-
mined directly from the processed data; others are based on picked
horizons. Combinations of attributes are the basis for automated or
semiautomated seismic facies classification schemes.
The use of single-trace seismic attributes increased significantly
after the publication of the Taner et al. 共1979兲 classic paper on com-
plex trace analysis, and new multitrace seismic attributes have been
introduced nearly every year since the mid-1980s 共see a comprehen-
sive review by Chopra and Marfurt, 2005兲. Two suites of multitrace
attributes that have proven particularly useful in analyzing and inter-
preting 3D seismic data are the coherence-based coherency, dip, and
azimuth attributes 共Marfurt et al., 1998兲 and the texture-based ener-
gy, entropy, homogeneity, and contrast attributes. 共Note: Some of
these attributes are given different names by different authors, e.g.,
West et al., 2002; Gao, 2003, 2004; Chopra and Alexeev, 2006.兲
The goals of GPR investigations are similar to those of seismic
studies 共points 1–5 above兲, except the key physical properties are
electromagnetic rather than elastic/aneleastic and the fluids of inter-
est are shallow-water occurrences and contaminants rather than
a) b)
c)
2006兲. Most of these papers are concerned with single-trace at-
tributes or with multitrace attributes computed from picked hori-
zons.
Young et al. 共1997兲 were the first and only group to present coher-
ency, dip, and azimuth attributes computed directly from 3D GPR
data. Although they successfully imaged the boundaries of a sand-
stone channel, the shallow ⬃8-m depth penetration, long ⬃2-m
dominant wavelength, relatively coarse 0.61⫻ 1.22-m grid spacing,
and small number 共2626兲 of traces allowed only limited structural in-
formation to be extracted from the data. Other groups of researchers
共Rea and Knight, 1998; Tercier et al., 2000; Moysey et al., 2003,
2006; Dafflon et al., 2005兲 have developed and applied a variety of
statistical and texture analysis techniques to 2D GPR data sets.
Quantitative measures of texture require sufficient data samples to
represent each facies 共Gao, 2003兲. Consequently, the resolution of
texture images derived from 2D cross sections is constrained by the
dimensions of the necessarily large analysis windows. This problem
is significantly reduced for 3D data sets as a result of the additional
information available from the volume of samples surrounding each
sample point. Furthermore, by including information from both the
inline and crossline directions, interpretational bias is reduced 共Gao,
2003兲.
We use complementary information provided by coherence- and
texture-based attributes to interpret a relatively large 共28,304 traces兲
high-resolution 3D GPR data set collected across a section of the Al-
pine Fault Zone in New Zealand, which is rich in structural detail.
Our primary geologic objective is to map shallow-sediment defor-
mation produced by past earthquakes along the
fault zone. From the multitude of attribute vol-
umes, we identify subtle features that would have
been missed using traditional approaches to 3D
GPR interpretation.
After briefly reviewing the geology of our
study site, we describe how the 3D GPR data set
was acquired across the Alpine Fault Zone and
the subsequent data processing. We then outline
the essential elements of the two suites of at-
tribute techniques and apply them to the Alpine
Fault Zone data set. Based on their information
content, we choose a reduced suite of attribute
volumes to help us interpret the data in terms of a
simple structural/tectonic model.

ALPINE FAULT ZONE STUDY SITE

Figure 1. 共a兲 New Zealand Plate boundary setting and location of the Calf Paddock survey
site. 共b兲 Sketch showing the extent and location of the 3D GPR survey and CMP lines rel-
ative to the Alpine Fault Zone and monitoring wall. D and U indicate the locally
downthrown and upthrown sides of the Alpine Fault Zone. Note also the location of the
83-m-deep drillhole and the shallow paleoseismic trenches. 共c兲 Topography of the survey
area, displaying late Holocene vertical and horizontal displacements. The trough of an
abandoned stream channel is indicated by the curved arrows.
The Alpine Fault Zone is a major continental
transform fault zone on the South Island of New
Zealand that accommodates relative motion be-
tween the Pacific and Australian Plates 共Figure
1a兲. Despite average plate-convergence rates of
30–40 mm/year, the fault has not ruptured during
the nearly 150 years of European settlement in the
country 共Wells et al., 1999兲. Ongoing paleoseis-
mic investigations attempt to explain the poten-
tial effects of future earthquakes along the fault
 Active faults and 3D GPR attributes B13

共Cooper and Norris, 1995; Yetton, 1998, 2002; Sutherland et al.,
2006兲.
At our Calf Paddock study site, the Alpine Fault Zone has a
transpressive character, rupturing various late Pleistocene and Ho-
locene fluvial deposits of the Maruia River 共Figure 1兲. A succession
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
To produce a reliable structural image of the subsurface, we had to
migrate the data using an algorithm that accounted for the topo-
graphic relief across the survey site 共Lehmann and Green, 2000;
Heincke et al., 2005兲. Of the wide range of tested velocities, the 0.08-
m/ns velocity estimate based on the CMP data analyses produced the

of offset terraces and stream channels provides direct evidence of re-
cent vertical and horizontal displacements 共Wellman, 1952; Garrick
and Hatherton, 1974; Yetton, 2002兲. Unfortunately, the interpreta-
tion of offsets is complicated by the probable presence of overlap-
ping fault strands 共Yetton, 2002兲. A wall erected across the principal
fault strand in 1962 as a rough gauge of aseismic strain 共Figure 1b兲
shows no evidence of disturbance.
An early investigation at Calf Paddock involved acquiring gravity
and seismic refraction data 共Garrick and Hatherton, 1974兲 and drill-
ing a vertical hole 39 m southeast of the monitoring wall center 共Fig-
ure 1b兲. The hole passed through 30 m of gravel before penetrating
basement schist to 83 m depth. Although the hole intersected a
steeply dipping crushed zone in schist bedrock, it did not transect the
principal fault plane. As a consequence, Garrick and Hatherton
共1974兲 inferred the fault must dip more than 65° to the southeast at
this site. Yetton 共2002兲 conducted a paleoseismic study of the site
that included excavating and face-logging trenches that crossed the
projected location of the fault 共Figure 1b兲.
most coherent migrated images. This value was used for the migra-
tion and time-to-depth conversion. Application of a 3 ⫻ 3-trace f-xy
deconvolution filter reduced minor random noise yet preserved
short-wavelength features in the migrated images. Finally, the air-
waves and ground waves were muted 共Figure 2c兲.
TOPOGRAPHY AND MIGRATED 3D GPR DATA
The gridded surface elevations determined by the self-tracking la-
ser theodolite clearly illustrate the location of the fault trace and the
vertical and lateral offsets of a fluvial terrace 共Figure 1c兲. The terrace
is offset vertically by approximately 1.5 m 共downthrown to the
northwest兲, and an abandoned stream channel is dextrally displaced
a)

GPR DATA ACQUISITION AND PROCESSING

Our 3D GPR data were recorded with a semiautomated acquisi-
tion system composed of a standard PulseEKKO GPR unit linked to
a self-tracking laser theodolite with automatic target-recognition ca-
pabilities 共Lehmann and Green, 1999兲. We used a broadside acquisi-
tion geometry with 1.0-m-offset 100-MHz antennas to record GPR
traces 共sampling rate, 0.5 ns兲 at roughly 0.25-m intervals along 61 b)
approximately 116-m-long lines separated by 0.5 m. The survey
was located on a terrace above the paleoseismic trench site, and re-
cording lines were oriented perpendicular to and centered about the
principal fault trace 共Figure 1b兲.
After applying a standard median filter to remove low-frequency
system-dependent noise 共wow兲, coordinates were assigned to the in-
dividual traces and the airwave arrivals were aligned using a cross-
correlation technique. The irregularly spaced traces were then inter-
polated onto a regular 0.25⫻ 0.5-m grid using Delaunay triangula-
tion 共e.g., Streich et al., 2006兲. Division by spatially averaged Hil- c)
bert transforms 共i.e., the amplitude envelopes; Gross et al., 2004兲 of
the data and frequency filtering yielded traces in which coherent re-
flections were observed from near the surface to at least 300 ns 共Fig-
ure 2a兲. 共For all time sections in this paper, the vertical exaggeration
is based on a velocity of 0.08 m/ns.兲
Semblance analyses of reflections observed on common-mid-
point 共CMP兲 data collected at different locations 共Figure 1b兲 yielded
velocities of 0.08 Ⳳ 0.01 m/ns. A strongly dispersive ground wave
observed on all CMPs was attributed to a thin, low-velocity
共⬃0.07 m/ns兲 surface soil layer that acted as a waveguide 共Streich et
al., 2006; van der Kruk et al., 2006兲. Diffractions from boulders
within this layer exhibited the same dispersive character in the com- Figure 2. 共a兲 Unmigrated inline GPR cross section at y ⳱ 15 m.
mon-offset data 共Figure 2a兲. They could not be collapsed using stan- Blue arrows indicate a steeply dipping diffraction tail affected by
dard migration techniques, but they were well separated from the dispersion in an overlying waveguide. 共b兲 As for 共a兲 but after 3D f-k
filtering to remove steeply dipping diffractions. 共c兲 As for 共b兲 after
more shallowly dipping events in the frequency-wavenumber 共f-k兲 3D topographic migration, depth conversion, and f-xy deconvolu-
domain. They were effectively removed by 3D f-k filtering 共Figure tion. Black arrows delineate the projected location of the Alpine
2b兲. Fault Zone. Vertical exaggeration is 3:1.
 B14 McClymont et al.
by roughly 10 m 共Figure 1c兲. These observations are consistent with
Yetton’s 共2002兲 geomorphological investigation.
The chair diagram representation of the migrated and time-to-
depth converted GPR volume in Figure 3 shows many nearly planar
reflections with variable dips and dip directions. The structural con-
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
trast generated by the fault is evident from the change in reflection
pattern at x ⳱ ⬃55 m, which coincides with the vertical offset of the
terrace.
SELECTED GEOMETRIC ATTRIBUTES
Rather than follow the normal procedure of interpreting 3D GPR
reflection facies from multiple vertical and horizontal sections
共Beres et al., 1995, 1999; Grasmueck, 1996; Green et al., 2003;
Gross et al., 2004兲, we want to classify the different reflections ob-
jectively on the basis of their 3D geometric attributes. Numerous
geometric attributes have been developed for analyzing 3D reflec-
tion seismic data 共Taner et al., 1994兲. They include coherence-based
attributes that are helpful for characterizing structures 共Bahorich and
Farmer, 1995; Marfurt et al., 1998; Cohen and Coifman, 2002兲 and
texture-based attributes that are valuable for facies discrimination
共West et al., 2002; Gao, 2003, 2004; Chopra and Alexeev, 2006兲. We
have written MATLAB codes for generating coherence- and texture-
based attributes from the processed Alpine Fault Zone GPR data vol-
ume.
Calculating coherence-based attributes
Bahorich and Farmer 共1995兲 introduce the coherence cube as a
means to locate faults within 3D reflection seismic data. Coherence
is defined by waveform similarity in both the inline and crossline di-
rections. Faults and other discontinuities are distinguished by abrupt
truncations of coherence. The original algorithm of Bahorich and
Farmer 共1995兲 uses normalized crosscorrelation to estimate coher-
ence. Marfurt et al. 共1998兲 develop a second-generation algorithm
using semblance as a coherency measure. Their algorithm estimates
the coherency, dip, and azimuth for the most coherent dipping plane
within a 3D analysis window of data surrounding each sample point.
Unlike other methods that derive dip and azimuth from picked hori-
zons, these attributes are calculated directly from the data.
Our MATLAB code for estimating coherency, dip, and azimuth
from 3D GPR data is based on Marfurt et al.’s 共1998兲 semblance al-
Figure 3. A portion of the fully processed 3D GPR volume from the
Alpine Fault survey. Top mute was applied to mask the airwave and
ground wave. The blue line identifies the ground surface. Vertical
exaggeration is 3:1.
gorithm 共Figure 4兲. For the Alpine Fault Zone data, semblance is cal-
culated for hypothetical planes with ranges of apparent dips ␪ x 共in-
line兲 and ␪ y 共crossline兲 varying between ⳮ40° and Ⳮ40° — the
minimum and maximum values corresponding to the steepest dips
observed in the topographically migrated data. The coherence of
each sample point is defined to be the maximum semblance calculat-
ed from all combinations of apparent dips 共Figure 5兲. The apparent
dip pair with the highest semblance is then used to find the true dip ␪
and azimuth ␸ of the plane as follows:
␪ ⳱ tanⳮ1共tan2␪ x Ⳮ tan2␪ y兲1/2 共1兲
and
␸ ⳱ tanⳮ1共tan ␪ y /tan ␪ x兲. 共2兲
Examples of applying the semblance algorithm to two reflections
observed in the Alpine Fault Zone GPR data set are shown in Figure
5. One reflection is practically horizontal, and the other has a notice-
able apparent southwesterly dip 共in the northwest-southeast sections
shown in many figures, the southwesterly dipping reflections have
apparent southeasterly dips兲.
Calculating texture-based attributes
In addition to determining the geometry of individual GPR reflec-
tions, we want to describe and delineate objectively any distinctive
reflection patterns associated with the various geologic units without
predefining their boundaries. To achieve this goal, we use texture
analysis techniques originally developed for the automated and
semiautomated interpretation of 2D digital images 共Haralick et al.,
1973; Weszka et al., 1976兲. Such techniques are already applied in
analyzing diverse medical, biological, satellite, airborne, and side-
scan sonar images as well as gravity, magnetic, and seismic data. Al-
though the concept of texture in everyday life is intuitively obvious,
it is not precisely defined. For many purposes, texture is a measure of
a) b)
c)
Figure 4. Sketches showing the principle of the semblance-based al-
gorithm for calculating coherence-based attributes from 3D GPR
data. 共a兲 Top 共x,y兲 view of 3 ⫻ 3 共dark gray shading兲 and 5 ⫻ 5 共light
and dark gray shading兲 rectangular-analysis windows centered on an
analysis point. 共b兲 A 2D cross section 共x,t兲 showing a coherent mod-
erately dipping reflection. The analysis window extends over five
traces in the x-direction and has a time width of ⌬t. The algorithm
searches for the most coherent dip in both the x- and y-directions 共␪ x
and ␪ y兲. These apparent dips are used to calculate true dip and azi-
muth relative to north. 共c兲 An example of a relatively low-coherency
shallow-dipping reflection.
 Active faults and 3D GPR attributes B15

the visual characteristics of objects or features. As examples, tex-
tures can be described as simple or complex, smooth or rough,
coarse or fine, linear or curvilinear, parallel or divergent, ordered or
nonordered, flat or undulating, even or uneven, etc. For 2D and 3D
digital data, texture attributes provide quantitative representations
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
sity values within the respective 3D analysis window. A range of sta-
tistics applied to the GLCMs then defines the attributes that quantita-
tively describe the textures.
For 3D reflection seismic or GPR data, the analysis window con-
sists of Nx ⫻ Ny ⫻ Nz data points in the inline, crossline, and vertical
directions, respectively 共Gao, 2003兲. After specifying the orienta-

of these qualitative characteristics.

Texture attributes computed for a given data point determine the
statistical relationships between all data within a predefined 2D or
3D analysis window centered on that point. The data are pixels or
voxels in 2D and 3D digital imaging terminology, respectively, and
the texture analysis windows are texels 共textural elements兲. The size
of the analysis window is defined according to the scale of textures
observed in the data. To convert an entire 2D or 3D data set to a suite
of texture-attribute planes or volumes, we need to compute texture
attributes for all data points.
Following the work of West et al. 共2002兲, Gao 共2003, 2004兲, and
Chopra and Alexeev 共2006兲 on the application of texture-analysis
techniques to 2D and 3D seismic data, our algorithm for computing
texture-attribute volumes from 3D GPR data is based on the gray-
level co-occurrence matrix 共GLCM兲 method of Haralick et al.
tion and distance, each GLCM is obtained by summing the co-occur-
rences 共i, j兲 throughout the 3D analysis window. Although GLCMs
can be calculated for any 3D orientation 共West et al., 2002兲, we limit
our analyses to the three orthogonal directions, x, y, and z共t兲. Prior to
computing the statistical relationships, each GLCM is normalized
such that the sum of all elements is equal to one.
Figure 6 shows GLCMs computed along the orthogonal direc-
tions for two 3D analysis windows extracted from different parts of
a)

Time (ns)
共1973兲 共see Appendix A for a review兲. The computational process
begins by discretizing GPR amplitudes to a limited number of inten-
sity or gray levels Ng. Chopra and Alexeev 共2006兲 demonstrate that
Ng ⳱ 16 levels are sufficient to obtain meaningful statistics from 3D
seismic data.AGLCM is then created at each data point by determin-
ing the spatial distribution and frequency or co-occurrence of inten- b) c)

a)
Time Time
Time (ns)

d) 140 e) 80
120 60
5 100 5
80 40
10 60 10
40 20
15 20 15
5 10 15 5 10 15
b)
f) 250 g)
5 200 5 40
150 30
10 100 10 20
Time Time
50 10
15 15
5 10 15 5 10 15

h) i) 200
200
5 5 150
150
10 100 10 100
50 50
c) 30
0.7
30
0.6
15 15
20
0.6
20 5 10 15 5 10 15
0.5
0.5
10 10
0.4 Figure 6. Computations of texture-based attributes for the Alpine
Fault Zone GPR volume. 共a兲 Cross section extracted from the topo-
0.4
θ y (degrees) 0 θ y (degrees) 0
0.3 0.3
–10 –10
0.2
graphically migrated data. Vertical exaggeration is 3:1. Analysis
0.2
–20 –20 windows centered on each analysis point are extracted from the 3D
0.1 0.1
–30 –30 volume. They should be large enough to characterize the texture of
–30 –20 –10 0 10 20 30
θ x (degrees)
–30 –20 –10 0 10 20 30
θ x (degrees)
each reflection facies. 共b兲 Analysis window extracted from a portion
of the GPR volume where relatively flat and continuous reflections
Figure 5. Coherency as measured by semblance for two typical have been rediscretized into 16 intensity levels ranging from one
points within the Alpine Fault Zone GPR volume. 共a兲 Cross section 共red兲 to 16 共blue兲. 共c兲 Second analysis window extracted from a por-
extracted from the topographically migrated data. Vertical exaggera- tion of the volume where the reflections have moderate dips and are
tion is 3:1. 共b, c兲 Volumes of data surrounding the points. 共d, e兲 Sem- only quasi-continuous. 共d, e兲 GLCMs calculated along the x-axis of
blance values determined for 196 combinations of apparent dip the analysis windows shown in 共b兲 and 共c兲. 共f, g兲As for 共d兲 and 共e兲 but
共14⫻ 14 values兲. for the y-axis. 共h, i兲 As for 共d兲 and 共e兲 but for the time axis.
 B16 McClymont et al.

the Alpine Fault Zone GPR volume. To capture the fabric of the re-
flections, we use 9 ⫻ 7 ⫻ 30 analysis windows 共i.e., nine traces in the
x-direction, seven traces in the y-direction, and thirty time samples,
equivalent to a volume of 2.0⫻ 3.0⫻ 0.6 m兲. In contrast to analysis
windows used to calculate coherence-based attributes, those used
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
disorder or complexity. It is large for nonuniform textures that gener-
ally yield low GLCM values. Although energy and entropy are relat-
ed, when used together, they can draw attention to different features
in the data.
Homogeneity yields details on the similarity of pixels and is a use-

here should be large enough to capture repeating amplitude patterns
within the data. However, overly large windows will smooth fea-
tures within the transformed attribute volume. The data are redis-
cretized to 16 intensity levels from 1 共red兲 to 16 共blue兲. One analysis
window contains continuous flat reflections 共Figure 6b, d, f, and h兲;
the other contains discontinuous moderately dipping reflections
共Figure 6c, e, g, and i兲.
The different textures in the two analysis windows control the dis-
tribution of co-occurrences in the resultant GLCMs. For example,
the strong continuous reflections within the first analysis window
produce a tight distribution along the diagonals of GLCMx and
GLCMy 共Figure 6d and f兲. In contrast, the dipping and more chaotic
reflections within the second analysis window produce a wider dis-
tribution about the diagonals 共Figure 6e and g兲.
To derive meaningful texture information from the GLCM repre-
sentations of our 3D GPR data, we must analyze them statistically.
Haralick 共1973兲 proposes 14 different statistical operations for ex-
tracting a variety of texture information from the GLCMs. Fortu-
nately, most of the useful texture information in wavefield data is
contained in four statistical parameters 共West et al., 2002; Chopra
and Alexeev, 2006; see Appendix A兲: energy, entropy, homogeneity,
and contrast 共inertia兲.
Energy supplies information on texture uniformity. It is lowest
when all elements of the GLCM are equal and is helpful for empha-
sizing reflector continuity and geometry. Entropy is a measure of
ful indicator of overall image smoothness and reflector continuity. It
is high for GLCMs that have elements located near the diagonal.
Contrast 共or inertia兲 highlights local image variations and differenc-
es between adjacent data values. It is high for GLCM elements scat-
tered away from the diagonal. When homogeneity and contrast are
used together, they provide greater discrimination capabilities than
either attribute alone.
From the three GLCMs in Figure 6 and equations A-1–A-4 in Ap-
pendix A, we derive 12 statistical measures for each analysis win-
dow. These computations are repeated for every sample point in the
original GPR volume, yielding a suite of 12 texture attribute vol-
umes.
APPLICATIONS OF COHERENCE- AND
TEXTURE-BASED ATTRIBUTES TO THE
ALPINE FAULT ZONE GPR DATA
Coherence-based attributes
The size of the coherency analysis window is a controlling factor
in the resolution of the attribute volumes 共Marfurt et al., 1998兲. De-
pending on the data, overly small windows can yield noisy attribute
volumes, and overly large windows can result in excessive smooth-
ing and loss of detail. To address this issue, we tested a range of anal-
ysis windows on the migratedAlpine Fault Zone GPR data. In Figure
7, we present the results of three such tests in the form of horizontal

a)

b) e) h)

c) f) i)

d) g) j)

Figure 7. 共a兲 Horizontal slice extracted at 6 m depth from the fully processed Alpine Fault GPR volume. 共b-d兲 Coherency, azimuth, and dip at
6 m depth calculated using a 3 ⫻ 3 共0.5⫻ 1 m兲 trace analysis window. 共e-g兲 As for 共b-d兲 but for a 5 ⫻ 3 共1 ⫻ 1 m兲 trace analysis window. 共h-j兲 As
for 共b-d兲 but for a 5 ⫻ 5 共1 ⫻ 2 m兲 trace analysis window. The algorithm performed a dip search up to a maximum value of approximately 40°.
Azimuths and dips with coherencies less than 0.5 are not plotted. Various cross sections along A-A⬘ in 共a兲 are shown in Figures 8 and 12. Arrows
identify the projected location of the Alpine Fault Zone.
 Active faults and 3D GPR attributes B17

slices extracted at 6 m depth from the coherency, azimuth, and dip-
attribute volumes. The first three attribute slices are based on analy-
sis windows containing 3 ⫻ 3 traces 共0.5⫻ 1 m; Figure 7b-d兲, the
second three are based on 5 ⫻ 3 traces 共1 ⫻ 1 m; Figure 7e-g兲, and
the final three are based on 5 ⫻ 5 traces 共1 ⫻ 2 m; Figure 7h-j兲. All
calculations were performed using a 10-ns vertical window 共equiva-
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
azimuth slice at 3 m depth, except the orientation of the stream chan-
nel changes abruptly near x ⳱ 70 m 共Figure 9f兲. This change is not
obvious in the original GPR image in Figure 9b.
On the 5.0- and 7.0-m depth slices, the reflections can be divided
into three domains: 共1兲 mostly subhorizontal reflections at x

⳱ 0–25 m, 共2兲 northwesterly dipping reflections at x ⳱ 25–55 m,

lent to about 0.4 m in the depth-converted volumes兲. Samples for and 共3兲 southwesterly dipping reflections that extend from the fault
which the analysis windows overlap the borders of the GPR volume near x ⳱ 55 m to at least the southeastern boundary of the survey.
are excluded from all figures. Although the azimuth and dip at- Close inspection of these deeper slices demonstrates that reflections
tributes can be displayed together in a single figure using hue and in- within a 20-m-wide zone are deflected approximately 30° from a
tensity to represent azimuth and dip magnitude 共Marfurt et al., general southwest to west-southwest orientation as the fault is ap-
1998兲, we find it easier to interpret the attributes using separate fig- proached from the southeast. This deflection is illustrated best in the
ures. Azimuths and dips for which the coherencies are less than 0.5
are blank in Figure 7c, d, f, g, i, and j.
Except for the region between x ⳱ 20 m and x ⳱ 50 m, the GPR a)
reflections have generally low coherencies 共Figure 7b, e, and h兲. This
is a consequence of the types of sediments being imaged; coarse flu-
vial gravels are rarely well layered. Azimuths and, to a lesser extent,
dips change abruptly at x ⳱ 55 m 共Figure 7c, d, f, g, i, and j兲, coinci-
dent with the estimated surface trace of the fault. Reflections south-
east of the fault have 210°ⳮ250° azimuths and moderate 20°ⳮ35°
dips, whereas those immediately to the northwest have ⬃340° azi-
muths and somewhat shallower 10°ⳮ25° dips. Farther to the north-
west at x ⳱ 0–30 m, the reflections are mostly subhorizontal. All of
this information can be discerned in the three suites of attributes. For b)
further analysis of our data, we use the attribute volumes based on
the larger asymmetric 共1.0⫻ 2.0 m兲 window because their resolu-
tion is sufficient and they contain less noise than those based on the
smaller asymmetric 共0.5⫻ 1.0 m兲 and symmetric 共1.0⫻ 1.0 m兲
windows.
Cross sections through the attribute volumes also provide useful
structural information. Relatively continuous reflections in the orig-
inal data of Figure 8a are emphasized in the coherency section of
Figure 8b. We interpret two bounding surfaces that separate domains
of different reflection coherency and azimuth on either side of the
fault. Southeast of the fault, a prominent shallow reflection 共s1-bs in
Table 1 and Figure 8b and c兲 separates shallow easterly dipping re- c)
flections from underlying weakly coherent southwesterly dipping
ones. Northwest of the fault, an onlap relationship exists between
shallow nearly horizontal features and a dipping reflection 共s2-bs in
Table 1 and Figure 8b and c兲. The azimuth cross section shows excel-
lent discrimination between the different reflection facies on either
side of the fault, demonstrating that the division of facies can be
traced to at least 15 m depth below the surface 共i.e., f2-tf and f1af in
Figure 8c兲. Complementary structural information supplied by the
original 3D data, the three coherence-based attribute volumes, and
the texture-based attributes 共see next section兲 can be interpreted in
terms of four reflection facies and two bounding surfaces 共Table 1兲.
A selection of horizontal azimuth attribute slices at 2.5, 3.0, 5.0,
and 7.0 m depth is sufficient to characterize the geometries of the Figure 8. 共a兲 Cross section from original GPR volume 共for location,
principal structures recorded in the 3D GPR data. In Figure 9, azi- see Figure 7a兲. 共b, c兲 Coherence-based attribute cross sections calcu-
muths with coherencies less than 0.5 are not plotted, and azimuths lated using the semblance-based coherency algorithm and 5 ⫻ 5
trace analysis windows. 共b兲 Coherency. Two bounding surfaces are
with dips less than 5° are plotted in gray. identified 共s1-bs and s2-bs described in Table 1兲. 共c兲 Azimuth. Unre-
At 2.5 m depth, we see only reflections from the topographically liable data points with coherencies less than 0.5 are not plotted.
higher 共upthrown兲 block southeast of the fault 共Figure 9a and e兲. Pre- There are three radar facies 共f1-af, f2-tf and f3-fl; see Table 1 for de-
dominantly horizontal reflections 共gray in Figure 9e兲 within radar fa- tails兲 characterized by distinctive azimuths.Afourth radar facies 共f4-
cies f3-fl 共Figure 8c, Table 1兲 are transected by the northeast- and fl in Table 1兲 is inferred above the s2-bs boundary. Arrows identify
the projected location of the Alpine Fault Zone. Dotted lines delin-
southwest-facing banks of an ancient stream channel 共blues and yel- eate a zone of reduced reflection continuity and steeper dips 共best
lows in Figure 9e兲. Sediments draped over the fault scarp produce seen on the texture-based attribute in Figure 12a兲. Vertical exaggera-
the band of pink at x ⳱ 55 m. Similar features are observed in the tion is 3:1.
 B18 McClymont et al.

Table 1. Description and interpretation of radar stratigraphy. Labeling system is after Neal (2004).

Description Geologic interpretation

Radar facies Subparallel southwesterly dipping reflections with Foreset beds of a prograding alluvial fan. Their
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

f1-af limited lateral continuity. Reflections have orientation indicates a northeasterly source. This
moderate 共20°ⳮ35°兲 dips. deposit was probably formed during a period of
late Pleistocene aggradation sourced from the
Maruia River catchment 共Suggate, 1965兲.

Radar facies Subparallel northwesterly dipping reflections that
f2-tf are moderately continuous. Reflections have
shallow to moderate 共10°ⳮ25°兲 dips.
Postseismic alluvium beds deposited in the
accommodation space provided by the repeatedly
downthrown block northwest of the principal fault
trace. The current alignment of the dipping strata
with the fault plane suggests they were tilted
upward by the fault.

Radar surface Laterally continuous undulating reflection. Erosional surface formed by the Maruia River
s1-bs downcutting into radar facies f1-af on the
southeast 共hanging wall兲 side of the fault zone.

Radar facies Subhorizontal reflections that are moderately Fluvial deposits of the Maruia River.
f3-fl continuous.

Radar surface Laterally continuous to discontinuous undulating Contact at base of late Holocene river gravels
s2-bs reflection. Subhorizontal reflections of radar resting unconformably over tilted alluvium beds
facies f4-fl onlap this surface. on the northwest 共footwall兲 side of the fault zone.

Radar facies Subhorizontal reflections that are moderately Late Holocene fluvial deposits of the Maruia
f4-fl continuous. They onlap radar surface s2-bs. River.

automatically generated rose diagrams of Figure

10, which show azimuths within 10-m-wide
a) e) strips extracted at 7 m depth from the azimuth at-
tribute volume. The ancient stream channel at
y (m)
y (m)

3 m depth also appears to be deflected by roughly

the same amount 共Figure 9f兲. We interpret the de-
b) flection of the ancient stream channel and deeper
f)
reflections as the result of displacement-induced
y (m)

y (m)

drag close to the fault. Drag in this instance refers

to deflected markers that are convex in the direc-
tion of slip 共c.f. Reches and Eidelman, 1995;
c) g) Grasemann et al., 2005兲. The channel on the shal-
y (m)

lowest slice at z ⳱ 2.5 m has not been deflected,

y (m)

indicating that it has been significantly less af-

fected by fault displacement共s兲.
d) h)
Y (m)

Y (m)

Texture-based attributes
As for the test examples shown in Figure 6, the
GLCMs we used are based on 2.0⫻ 3.0⫻ 0.6-m
Azimuth (degrees)
analysis windows. Horizontal slices extracted at
6 m depth from the 12 texture attribute volumes
are displayed in Figure 11; these attribute vol-
Figure 9. Horizontal slices extracted from 共a-d兲 the original Alpine Fault volume and
共e-h兲 the azimuth volume at depths of 2.5, 3.0, 5.0, and 7.0 m. Azimuths with coherencies umes are generated from GLCMs calculated for
less than 0.5 are not plotted, and azimuths with dips less than 5° are plotted in gray. Ar- adjacent points in the x-, y-, and z-directions. Al-
rows identify the interpreted location of the fault zone. though the numerical values of the color scales
 Active faults and 3D GPR attributes B19

vary for each plot, we highlight only relative variations 共i.e., high to
low values兲. Some attributes emphasize certain features better than
others. For example, because the dipping reflections northwest of
the fault mostly strike in the y-direction and those to the southeast
strike in other directions, all four attribute volumes based on GLCMy
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
cies identified within the gravels correspond to three stages of depo-
sition:
1兲 An episode of late Pleistocene aggradation 共Suggate, 1965兲 that
involved the progradation of an alluvial fan from northeast to

show excellent discrimination between the different reflection pat- southwest 共GPR facies f1-af in Table 1兲
terns on either side of the fault 共Figure 11兲. Reflections northwest of 2兲 A period of postseismic aggradation that led to a distinct se-
the fault have higher energy, lower entropy, higher homogeneity, and quence of alluvial sediments accumulating in the accommoda-
lower contrast than those to the southeast. Similarly, entropyx and
tion space provided by the repeatedly downthrown block north-
contrastx based on GLCMx and homogeneityz and contrastz based on
west of the principal fault trace 共GPR facies f2-tf in Table 1兲
GLCMz provide very good differentiation between the subhorizon-
tal reflections at x ⳱ 0–30 m and the moderately dipping reflections 3兲 Holocene base-level lowering, during which the Maruia River
at x ⳱ 30–55 m. incised into the aggradational valley fill and left a thin veneer of
Three of the 12 texture attributes are sufficient for objectively de- fine- to coarse-grained flood deposits blanketing the degraded
lineating the different reflection patterns recorded in theAlpine Fault river terraces 共GPR facies f3-fl and f4-fl in Table 1兲
Zone GPR data: contrastx, homogeneityy, and contrastz. Figure 12
shows cross sections extracted from these attribute volumes along Of primary importance is the steep southeasterly dip 共⬃80°兲 of
profile A-A⬘ of Figure 7a. Each attribute cross section is shown as a both the principal fault strand and the zone of shearing and stratal
semitransparent layer overlying a variable area plot of the original disruption across the main fault zone. These structures can be fol-
GPR data 共positive amplitudes are dark shades兲. Some regions of lowed from the surface to at least 15 m depth at this location 共Figure
low signal-to-noise ratio at depths greater than 13 m produce anom- 13兲. This is generally consistent with the steeply dipping crushed
alously high attribute values. zone encountered in the drillhole and the moderately to steeply
High contrastx values occur where reflections
show little continuity along the profile. The high
values at x ⳱ 45–60 m outline a distinct zone of
reduced reflection continuity across the fault and
increased reflection dip on its northwestern side
共Figure 12a兲. Assuming the disrupted reflections
are a consequence of fault movement, we inter-
pret this zone as a band of shearing and stratal dis-
ruption within a broader zone of shear-induced
folding identified in the coherence-based at-
tribute volumes. In addition, moderate and low
contrastx values improve discrimination of radar
facies f3-fl from f1-af on the southeastern side of
the fault 共compare Figures 8c and 12a兲.
The homogeneityy attribute highlights contin-
uous dipping reflections perpendicular to the
plane of the cross section 共Figure 12b兲, whereas
the contrastz attribute highlights high-frequency
Figure 10. Rose diagrams of azimuth values sampled within 10-m-wide fault-parallel
reflections within the cross section. The location corridors from northwest to southeast across the fault zone. Dashed lines define the strike
of the Alpine Fault Zone is well delineated on the of the fault zone. On the southeastern side of the fault zone, the dominant azimuth chang-
homogeneityy cross section 共Figure 12b兲. Thin es from west-southwest to southwest at x ⬎ 75 m. The number of values n used for the az-
subhorizontal reflections of radar facies f3-fl and imuth estimates are shown at the bottom of each diagram.
f4-fl yield high contrastz values, whereas the un-
derlying dipping reflections produce low values
共Figure 12c兲. This attribute differentiates be-
tween radar facies f4-fl and f2-tf on the north-
western side of the fault and does a better job of
defining radar facies f3-fl on the southeastern side
of the fault than the azimuth attribute 共Figure 8c兲.

GEOLOGIC INTERPRETATION
Cross sections extracted from the original 3D
GPR data and two of the most informative at-
tribute volumes are plotted at true scale Figure 11. Horizontal slices extracted at 6 m depth from the suite of texture-attribute vol-
umes calculated from the Alpine Fault Zone GPR data. Each column shows, in descend-
in Figure 13; a sketch of our geologic interpreta- ing order, the attributes calculated for the x-, y- and time 共or z兲 directions. Dashed lines
tion is presented in Figure 14. The four GPR fa- identify the projected location of the fault zone.
 B20 McClymont et al.
southeasterly dipping shears of the main fault zone observed in the
nearby 1.0–1.5-m-deep trenches 共Figure 1b; Yetton, 2002兲. Similar-
ly steeply dipping expressions of the Alpine Fault Zone are observed
in trenches about 12 km to the southwest and 45 km to the northeast
of our study site 共Yetton, 2002兲. These observations do not preclude
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
the Alpine Fault Zone from dipping more shallowly at greater depth
共Norris et al., 1990; Davey et al., 1995, 1998; Norris and Cooper,
1995; van Avendonk et al., 2004兲.
In the near surface 共⬍⬃15 m兲, the Alpine Fault Zone is character-
ized by a nearly 50-m-wide zone of deformation that can be parti-
tioned into three overlapping domains. A roughly 30-m-wide do-
main that forms the northwest footwall to the main fault zone is char-
acterized by asymmetric synclinal folding, within which we observe
upward tilting of older river gravel beds approaching the principal
a)
b)
c)
Figure 12. Cross sections of the most useful texture attributes ex-
tracted from the respective Alpine Fault Zone volumes along profile
A-A⬘ 共Figure 7a兲. Note the color scale at the base of the figure. The
texture attributes are plotted as semitransparent colors overlying a
variable area representation of the original GPR data. Annotations
refer to the reflection facies described in the text and Table 1. Arrows
identify the projected location of the Alpine Fault Zone. Dotted lines
define a zone where reflections have reduced continuity and steeper
dips 共highlighted by the contrastx attribute in 共a兲兲. Vertical exaggera-
tion is 3:1.
fault strand 共Figures 8a and c, 13a and b, and 14兲. A second 20-m-
wide domain that forms the southeast hanging wall is characterized
by lateral fault drag of alluvial-fan deposits and an ancient river
channel that approaches the principal fault strand 共Figures 9, 10, and
14兲. The two domains of folding are divided by an approximately15-
m-wide steeply south-southeasterly dipping zone of shearing and
stratal disruption of the main fault, which coincides with the zone of
ground surface-fault displacements 共Figures 12a, 13c, and 14兲.
Consequently, it appears that the lateral and vertical components
of transpressive strain across the fault are partitioned into quasi-in-
dependent domains of off-fault folding that affect the hanging wall
and footwall, respectively. Nicol and van Dissen 共2002兲 report a sim-
ilar mechanism of upper crustal-scale oblique slip partitioning at the
free surface for the transpressive Clarence Fault to the northeast of
the study site.
a)
b)
c)
Figure 13. The same cross sections as in Figures 8a and c and 12a but
plotted with no vertical exaggeration. The azimuth 共b兲 and contrastx
共c兲 sections show that both the principal fault trace 共arrows兲 and the
zone of shearing and stratal disruption 共defined by dotted lines兲 dip
steeply 共80°兲 to the southeast.
Figure 14. Structural and depositional model based on the original
3D GPR data recorded across the Alpine Fault Zone and various at-
tribute volumes.
 Active faults and 3D GPR attributes B21

CONCLUSIONS APPENDIX A

Our 3D GPR data have provided significant new information on GLCM METHOD AND STATISTICAL
fault-generated deformation within late Pleistocene and Holocene
PARAMETERS
fluvial deposits overlying the Alpine Fault Zone in New Zealand. A
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

variety of 3D coherence- and texture-based attribute representations GLCM method

of the data enabled us to emphasize important characteristics of the
The GLCM method is described comprehensively in a number of
fault zone, extract information not obvious in the standard GPR im-
publications 共Haralick et al., 1973; Reed and Hussong, 1989; Gao et
ages, classify the depositional facies, and describe key structures.
al., 1998; West et al., 2002兲. Here, we review briefly the method for
Reflection coherency, azimuth, and dip attributes were calculated
2D digital images using an Nx ⫻ Ny analysis window. The extension
using 1.0⫻ 2.0⫻ 0.4-m analysis windows. The coherency attribute
to 3D digital data sets is relatively straightforward 共Gao, 2003, 2004;
yielded limited new knowledge, but the azimuth and dip attributes
Chopra and Alexeev, 2006兲.
enabled us to differentiate between prograding alluvial-fan deposits
A GLCM is defined by the number of times that data with intens-
southeast of the fault and tilted alluvium beds to the northwest. In ad-
ity level i occur with a specified spatial relationship 共i.e., orientation
dition, the azimuth attribute helped us to map the principal fault trace
and distance兲 relative to data with intensity level j, where i
to roughly 15 m depth and to identify a zone of lateral fault drag on
⳱ 1,2,3, . . . ,Ng and j ⳱ 1,2,3, . . . ,Ng. The Ng ⫻ Ng matrix P can be
the southeast side of the fault.
evaluated for arbitrary orientations and distances within the analysis
A multitude of texture-based attributes, generated from GLCMs
window. According to Reed and Hussong 共1989兲, effective texture
using 2.0⫻ 3.0⫻ 0.6-m analysis windows, provided information
analyses can be performed using two GLCMs, one with adjacent
overlapping and complementary to that supplied by the coherence-
data points in the x-direction 共orientations 0° and 180°兲 and one with
based attributes. For example, a thin layer of shallowly dipping flu-
adjacent points in the y-direction 共orientations 90° and 270°兲.
vial sediments that was only partially defined by the azimuth at-
To illustrate the method, consider the following data that have
tribute was well defined by a combination of contrastx and contrastz
been discretized to Ng ⳱ 4 intensity levels in an Nx ⳱ 5 by Ny ⳱ 5
attributes. The texture attributes also highlighted subtle features not
analysis window:
recognized in the coherence-based attribute volumes. In particular,
high contrastx values revealed a 15-m-wide zone of discontinuous,
moderately dipping reflections that we interpreted to be a domain of
intense shearing centered about the principal fault trace.
Our interpretation of the 3D GPR data and associated attribute
volumes was constrained by information provided by earlier geolog-
ic, paleoseismic, and geophysical investigations. The resultant dep-
ositional/structural model of the Alpine Fault Zone at the Calf Pad-
dock study site included distinct gravel units formed during three
phases of deposition and a well-defined principal fault trace. Differ-
ential motion generated a band of intense shearing on both sides of
the fault, the upward tilting of formerly flat-lying gravels on the
northwest side and lateral fault drag on the southeast side. The GLCM for this analysis window evaluated for adjacent data
Many critical details of our analysis strategy and depositional/ points in the forward and reverse x-directions is
structural model were fundamentally dependent on the 3D nature of
the Alpine Fault Zone GPR data set. Based on this experience and
experience gained by others in analyzing seismic data, we propose
that various statistical, textural, and automated classification tech-
niques recently developed for 2D GPR data could be even more ef-
fective when applied to 3D GPR data. Nevertheless, we would argue
that pragmatic approaches to 3D GPR data analysis should continue
to include a large degree of interpreter quality control in addition to
applications of semiautomated or automated classification tech-
niques.

ACKNOWLEDGMENTS
We greatly appreciate the contributions of the ETH-University of
Canterbury field crew. This project was supported by grants from the
Swiss National Science Foundation and ETH Zurich.
The 共i, j兲 element of this GLCM is the number of times data points
with intensity i occur adjacent to data points with intensity j. For ex-
ample, the number four at i ⳱ 2 and j ⳱ 3 represents four occur-
rences of a data value with intensity level two adjacent to a data value
with intensity level three along the x-direction. The calculations are
performed in the forward and reverse directions, so adjacent data
 B22 McClymont et al.
values with equal intensities 共i.e., the diagonal values兲 count twice
and the matrix is symmetrical.
Statistical parameters
The GLCM-based statistical parameters that we use are as fol-
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
lows:
Ng Ng
兺 兺
Energy ⳱ P2i,j , 共A-1兲
i⳱1 j⳱1
Ng Ng
Entropy ⳱ 兺 兺 Pi,jlog Pi,j ,
i⳱1 j⳱1
共A-2兲
Ng Ng
1
兺兺
Homogeneity ⳱ 2 Pi,j ,
i⳱1 j⳱1 1 Ⳮ 共i ⳮ j兲
共A-3兲
and
Ng Ng
兺 兺 共i ⳮ j兲2Pi,j ,
Contrast 共inertia兲 ⳱ 共A-4兲
i⳱1 j⳱1
where Pi,j is the element in the ith row of the jth column of the
GLCM.
REFERENCES
Bahorich, M. S., and S. L. Farmer, 1995, 3-D seismic coherency for faults
and stratigraphic features: The Leading Edge, 14, 1053–1058.
Beres, M., A. G. Green, P. Huggenberger, and H. Horstmeyer, 1995, Detailed
3-D georadar studies of glaciofluvial sediments: Geology, 12, 1087–1091.
Beres, M., P. Huggenberger, A. G. Green, and H. Horstmeyer, 1999, A study
of glaciofluvial architectures using two- and three-dimensional georadar
methods: Sedimentary Geology, 129, 1–24.
Chopra, S., and V. Alexeev, 2006, Applications of texture attribute analysis
to 3D seismic data: The Leading Edge, 25, 934–940.
Chopra, S., and K. J. Marfurt, 2005, Seismic attributes — A historical per-
spective: Geophysics, 70, no. 5, 3SO–28SO.
Cohen, I., and R. R. Coifman, 2002, Local discontinuity measures for 3-D
seismic data: Geophysics, 67, 1933–1945.
Cooper, A. F., and R. J. Norris, 1995, Displacement on the Alpine Fault at
Haast River, South Westland, New Zealand: New Zealand Journal of Ge-
ology and Geophysics, 38, 509–514.
Corbeanu, R. M., G. A. McMechan, R. B. Szerbiak, and K. Soegaard, 2002,
Prediction of 3D fluid permeability and mudstone distributions from
ground-penetrating radar 共GPR兲 attributes: Example from the Cretaceous
Ferron sandstone member, east-central Utah: Geophysics, 67, 1495–1504.
Dafflon, B., J. Tronicke, and K. Holliger, 2005, Inferring the lateral subsur-
face correlation structure from georadar data: Methodological background
and experimental evidence, in P. Renard, H. Demougeot-Renard, and R.
Froidevaux, eds., Geostatistics for environmental applications: Springer
Pub. Co., Inc., 467–478.
Davey, F. J., T. Henyey, W. S. Holbrook, D. Okaya, T. A. Stern, A. Melhuish,
S. Henrys, H. Anderson, D. Eberhart-Phillips, T. McEvilly, R. Urhammer,
F. Wu, G. R. Jiracek, P. E. Wannamaker, G. Caldwell, and N. Christensen,
1998, Preliminary results from a geophysical study across a modern, conti-
nent-continent collisional plate boundary — The Southern Alps, New
Zealand: Tectonophysics, 288, 221–235.
Davey, F. J., T. Henyey, S. Kleffmann, A. Melhuish, D. Okaya, T. A. Stern,
and D. J. Woodward, 1995, Crustal reflections from the Alpine Fault Zone,
South Island, New Zealand: New Zealand Journal of Geology and Geo-
physics, 38, 601–604.
Gao, D., 2003, Volume texture extraction for 3D seismic visualization and in-
terpretation: Geophysics, 68, 1294–1302.
——–, 2004, Texture model regression for effective feature discrimination
— Application to seismic facies visualization and interpretation: Geo-
physics, 69, 958–967.
Gao, D., S. D. Hurst, J. A. Karson, J. R. Delaney, and F. N. Spiess, 1998,
Computer-aided interpretation of side-looking sonar images from the east-
ern intersection of the Mid-Atlantic Ridge with the Kane transform: Jour-
nal of Geophysical Research, 103, 20997–21014.
Garrick, R. A., and T. Hatherton, 1974, Seismic refraction profiles across the
Alpine Fault: Geophysics Division, Department of Scientific and Industri-
al Research Report 87.
Grasemann, B., S. Martel, and C. Passchier, 2005, Reverse and normal drag
along a fault: Journal of Structural Geology, 27, 999–1010.
Grasmueck, M., 1996, 3-D ground-penetrating radar applied to fracture im-
aging in gneiss: Geophysics, 61, 1050–1064.
Grasmueck, M., and D. Viggiano, 2007, Integration of ground-penetrating
radar and laser position sensors for real-time 3-D data fusion: IEEE Trans-
actions on Geoscience and Remote Sensing, 45, 130–137.
Green, A. G., R. Gross, K. Holliger, H. Horstmeyer, and J. Baldwin, 2003,
Results of 3-D georadar surveying and trenching the San Andreas Fault
near its northern landward limit: Tectonophysics, 368, 7–23.
Gross, R., A. G. Green, H. Horstmeyer, and J. H. Begg, 2004, Location and
geometry of the Wellington Fault 共New Zealand兲 defined by detailed three-
dimensional georadar data: Journal of Geophysical Research, 109, doi:
10.1029/2003JB002615.
Haralick, R. M., K. Shanmugam, and I. Dinstein, 1973, Textural features for
image classification: IEEE Transactions Systems, Man and Cybernetics,
SMC-3, 610–621.
Heincke, B., A. G. Green, J. van der Kruk, and H. Horstmeyer, 2005, Acqui-
sition and processing strategies for 3D georadar surveying a region charac-
terized by rugged topography: Geophysics, 70, no. 6, K53–K61.
Lehmann, F., and A. G. Green, 1999, Semiautomated georadar data acquisi-
tion in three dimensions: Geophysics, 64, 719–731.
——–, 2000, Topographic migration of georadar data: Implications for ac-
quisition and processing: Geophysics, 65, 836–848.
Marfurt, K. J., R. L. Kirlin, S. L. Framer, and M. S. Bahorich, 1998, 3-D seis-
mic attributes using a semblance-based coherency algorithm: Geophysics,
63, 1150–1165.
Moysey, S., J. Cears, R. Knight, and R. M. Allen-King, 2003, Stochastic esti-
mation of facies using ground penetrating radar: Stochastic Environmen-
tal Research and Risk Assessment, 17, 306–318.
Moysey, S., R. J. Knight, and H. M. Jol, 2006, Texture-based classification of
ground-penetrating radar images: Geophysics, 71, no. 6, K111–K118.
Neal, A., 2004, Ground-penetrating radar and its use in sedimentology: Prin-
ciples, problems and progress: Earth Science Reviews, 66, 261–330.
Nicol, A., and R. van Dissen, 2002, Up-dip partitioning of displacement
components on the oblique-slip Clarence Fault, New Zealand: Journal of
Structural Geology, 24, 1521–1535.
Norris, R. J., and A. F. Cooper, 1995, Origin of small-scale segmentation and
transpressional thrusting along the Alpine Fault, New Zealand: Geological
Society of America Bulletin, 107, 231–240.
Norris, R. J., P. O. Koons, and A. F. Cooper, 1990, The obliquely-convergent
plate boundary in the South Island of New Zealand: Implications for an-
cient collision zones: Journal of Structural Geology, 12, 715–725.
Rea, J., and R. Knight, 1998, Geostatistical analysis of ground-penetrating
radar data: A means of describing spatial variation in the subsurface: Water
Resources Research, 34, 329–339.
Reches, Z., and A. Eidelman, 1995, Drag along faults: Tectonophysics, 247,
145–156.
Reed, T. B., and D. Hussong, 1989, Digital image processing techniques for
enhancement and classification of SeaMARC II side scan sonar imagery:
Journal of Geophysical Research, 94, 7469–7490.
Sénéchal, P., H. Perroud, and G. Sénéchal, 2000, Interpretation of reflection
attributes in a 3D GPR survey at Vallee d’Ossau, western Pyrenees,
France: Geophysics, 65, 1435–1445.
Streich, R., and J. van der Kruk, 2007, Accurate imaging of multicomponent
GPR data based on exact radiation patterns: IEEE Transactions on Geo-
science and Remote Sensing, 45, 93–103.
Streich, R., J. van der Kruk, and A. G. Green, 2006, Three-dimensional mul-
ticomponent georadar imaging of sedimentary structures: Near Surface
Geophysics, 4, 39–48.
——–, 2007, Vector-migration of standard copolarized 3D GPR data: Geo-
physics, 72, no. 5, J65–J75.
Suggate, R. P., 1965, Late Pleistocene geology of the northern part of the
South Island, New Zealand: New Zealand Geological Survey, Department
of Scientific and Industrial Research Report 77.
Sutherland, R., K. Berryman, and R. Norris, 2006, Quaternary slip rate and
geomorphology of the Alpine Fault: Implications for kinematics and seis-
mic hazard in southwest New Zealand: Geological Society of America
Bulletin, 118, 464–474.
Taner, M. T., F. Koehler, and R. E. Sheriff, 1979, Complex seismic trace anal-
ysis: Geophysics, 44, 1041–1063.
Taner, M. T., J. S. Schuelke, R. O’Doherty, and E. Baysal, 1994, Seismic at-
tributes revisited: 64th Annual International Meeting, SEG, Expanded Ab-
stracts, 1104–1106.
Tercier, P., R. Knight, and H. Jol, 2000, Acomparison of the correlation struc-
ture in GPR images of deltaic and barrier-spit depositional environments:
 Active faults and 3D GPR attributes
Geophysics, 65, 1142–1153.
Tronicke, J., P. Villamor, and A. G. Green, 2006, Detailed shallow geometry
and vertical displacement estimates of the Maleme Fault Zone, New
Zealand, using 2D and 3D georadar: Near Surface Geophysics, 4,
155–161.
van Avendonk, H. J. A., W. S. Holbrook, D. Okaya, J. K. Austin, F. Davey,
and T. Stern, 2004, Continental crust under compression — A seismic re-
Downloaded 02/05/18 to 203.189.88.157. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
fraction study of South Island geophysical transect I, South Island, New
Zealand: Journal of Geophysical Research, 109, B06302, .
van der Kruk, J., R. Streich, and A. G. Green, 2006, Properties of surface
waveguides derived from separate and joint inversion of TE and TM GPR
data: Geophysics, 71, no. 1, K19–K29.
Wellman, H. W., 1952, The Alpine Fault in detail: River terrace displace-
ment, Maruia River: New Zealand Journal of Science and Technology,
B33, 409–414.
Wells, A., M. D. Yetton, R. P. Duncan, and G. H. Stewart, 1999, Prehistoric
dates of the most recent Alpine Fault earthquakes, New Zealand: Geology,
27, 995–998.
B23
West, B. P., S. R. May, J. E. Eastwood, and C. Rossen, 2002, Interactive seis-
mic facies classification using textural attributes and neural networks: The
Leading Edge, 21, 1042–1049.
Weszka, J. S., C. R. Dyer, and A. Rosenfeld, 1976, A comparative study of
texture measures for terrain classification: IEEE Transactions Systems,
Man and Cybernetics, SMC-6, 269–285.
Yetton, M. D., 1998, Progress in understanding the paleoseismicity of the
central and northern Alpine Fault, Westland, New Zealand: New Zealand
Journal of Geology and Geophysics, 41, 475–483.
——–, 2002, Paleoseismic investigation of the North and West Wairau sec-
tions of the Alpine Fault, South Island, New Zealand: New Zealand Earth-
quake Commission Research Foundation Report 99/353.
Young, R. A., Z. Deng, K. J. Marfurt, and S. E. Nissen, 1997, 3-D dip filtering
and coherence applied to GPR data: A study: The Leading Edge, 16, 1011–
1018.
Young, R. A., and N. Lord, 2002, A hybrid laser-tracking/GPS location meth-
od allowing GPR acquisition in rugged terrain: The Leading Edge, 21,
486–490.