Professional Documents
Culture Documents
AbstractPrevious polarimetric synthetic aperture radar the most popular methods to generate the objects. In some
(PolSAR) images change detection methods are generally previous object-based change detection methods[3][4], the
undertaken in the pixel scale, resulting in overlooking the segmentation methods should be performed to each PolSAR
semantic information. To solve this problem, this paper presents image and an fusion processing is required to obtain the final
a superpixel-based PolSAR images change detection methods.
objects. Thus, these methods generally need too much
Different from some previous methods, an improved SLIC
superpixel segmentation method is introduced in polarimetric computation.
interferometric SAR (Pol-InSAR) images processing, which can The difference map (DM) is used to measure the change
segment two PolSAR images simultaneously. Moreover, along degree and is crucial in change detection. Many methods have
with the difference maps (DMs) generated based on the complex been used to generate the DM, including those based on clutter
wishart distribution and the total intensity, we also use color distribution models[5] and target decomposition[6]. However,
information to generate the DM between PolSAR images, which to the best of our knowledge, the color-based DM is not used
has long been overlooked in PolSAR images change detection. so far. Actually, the color information has long been used in
Based on the segmentation results and three DMs, two different understanding the PolSAR images, like the Pauli pesudo-color
change detection schemes are proposed. One is to perform the
majority voting to three change detection results using the above
image. Uhlmann et al. [7] have integrated color features in
mentioned DMs. Another is to fuse the three DMs and generate a PolSAR image classification, and the experimental results
refined DM, and the change detection is performed on the refined improved greatly compared with that without using color-based
DM. Two Radarsat-2 images acquired over Suzhou city, China, features.
are used in our experiments to validate the proposed methods, Inspired by previous work, an improved superpixel-based
and the experimental results show that the proposed schemes can PolSAR images change detection method is proposed. The
improve the change detection results significantly. SLIC superpixel segmentation [8] has been used in FP SAR
image segmentation. For change detection, the interferometric
Index TermsPol-InSAR segmentation, color information,
coherency matrix T6 is used instead of the T3 in [8]; therefore,
change detection
the PolSAR image pair can be segmented simultaneously.
Additionally, the gradient calculated by the edge detector is
I. INTRODUCTION also introduced into the calculation of the distance
Due to rapid urbanization, the land cover types in urban and measurement in SLIC, which can preserve the edge well. To
sub-urban areas are changing gradually, and the demands for take full advantage of PolSAR information, we also use color-
precise and real-time monitoring of urban changes are also based features to generate the DM. The Pauli images are used
increasing. Some traditional methods including field survey are for extracting the color information. Along with the DM
limited by the long time-consumption and small spatial generated based on color information, two other DMs
coverage. On the contrary, remote sensing techniques have the calculated using the complex wishart distribution model and
abilities of frequent and large-scale observation. Therefore, the total intensity are also obtained. In general, the changed
remote sensing techniques provide an irreplaceable tool for areas can be extracted using some threshold segmentation
urban change detection. Compared with the optical remote methods after the generation of DMs. Recently, some scholars
sensing sensors, the synthetic aperture radar (SAR) breaks the [9] tried to combine different DMs and generated a refined DM.
limits of cloude coverage and time of day [1]; therefore, the use According to their main ideas, two different object-based
of SAR images in urban change detection gains popularity change detection schemes are proposed in this paper, and the
during the last decade. experimental results are compared using the Radarsat-2 images.
Fully Polarimetric (FP) SAR data are relatively powerful in
remote sensing applications than the single- and dual-pol SAR II. STUDY SITE AND EXPERIMENTAL DATA
data because they contain more information. Many change Two Radarsat-2 images acquired over Suzhou city, Jiangsu
detection methods using FP SAR data have been explored, province, China, are used in our experiments. The acquisition
most of them are in the pixel scale, they generally overlook the dates are April 9, 2009 (master image) and June 15, 2010
semantic information. To solve this problem, scholars (slave image), respectively. The original spatial resolution is 10
introduced some novel methods, like Markov random field m 8 m, and the indence angle ranges from 38.37 at near
(MRF)[2], into PolSAR images change detection. Another range to 39.85 at far range. Suzhou is suitable for carrying out
strategy to use the contextual information is the object-based the change detection experiments due to its rapid urban
change detection methods[3][4]. Image segmetation is one of expansion in recent years, and we select a region of interest
978-1-4673-7297-8/15/$31.00 2015
c IEEE 792
(ROI) over Shihu areas for futher investigations, as given in parameters to modulate the weight of wishart distance and the
Figs. 1(a) and (b). The image size is 895 540. And Fig. 1(c) gradient. To speed up the SLIC segmentation, the iteration is
illustrates the real change map of the study areas. terminated when the percentage of pixels switching between
classes is lower than a given threshold Pc, in our experiments,
Pc is 5.0%. S, m and N in our experiments are 15, 1 and 15,
respectively.
Fig. 2 shows the SLIC superpixel segmentation results,
which show ability to preserve the objects edge well. Figs. 2(c)
and (d) compare the results with and without using the edge
gradient, as labeled in the yellow ellipses, the improved
distance measurement performs better in edge preserving.
Fig.1. The Pauli images of the study site acquired on (a) April 9, 2009, and (b)
June 15, 2010; and (c) the real change map of the study areas.
, k6 = [ k1 k2 ]
T images, the base maps are the Pauli images; (c) and (d) show the results with
T6 = k k *T
6 6 (1) and without using edge gradient.
where B. DM generation and evaluation
k = [ S HH + SVV S HH SVV 2 S HV ]
T
2 (2) DM calculated based on the complex wishart distribution
The <> denotes the spatial averaging operation, the model can be expressed as
superscripts * and T indicate the complex conjugate and the Tr(T3 (i) 1T3 ( j ) + T3 ( j )1 T3 (i ))
DW (T3 (i), T3 ( j )) = 3 (6)
matrix transpose, k1 and k2 are the scattering vectors of master 2
and slave images. Similarly, DM calculated using the tatal intensity Span is
In [8], the edge map is only used in the initialization of the
Spani Span j
cluster centers. To further preserve the edge in the DS (T3 (i), T3 ( j )) = + 2 (7)
segmentation result, we integrate the edge gradient into the Span j Spani
definition of distance measurement between a certain pixel p Except for these widely used DMs, in this paper, we
and the cluster center i explore the potentials of color-based features in PolSAR
2 2 images change detection. Without loss of generality, the
dp d dg
D = + s + 2 (3) following experiments are carried out based on the Pauli image,
m S N which is one of the most widely used tools to visualize the
where details of objects in PolSAR images. And the definition of
color-based DM is
T6 ( p )
d p = ln
T6 ( i ) ( )
+ Tr T6 ( p ) T6 ( i ) q, q = 6
1
(4) DC (( Rm , Gm , Bm ), ( Rs , Gs , Bs ))
(8)
= { Rm Rs + Gm Gs + Bm Bs } 3
(x x ) + ( y yp )
where (Rm, Gm, Bm) and (Rs, Gs, Bs) are the colors of master and
2 2
ds = i p j (5)
slave images.
where dp is the wishart distance between two coherency Fig. 3 shows the DMs based on different distance
matrices, ds is the euclidean distance, xi and yi indicate the row measurements. To evaluate the DMs, we use the receiver
and colume number, dg is the edge gradient obtained by the operating characteristic (ROC) curve, which has been widely
edge detector. S is the desired superpixel size, m and N are the used in the performance evaluation of DM extraction methods
Fig. 3. DMs generated based on (a) the complex wishart distribution model, (b)
the total intensity, and (c) the color information.
Fig. 5. The flowchart of the first change detection scheme.
Fig. 7. The threshold segmention results of (a) the wishart-similarity DM, (b)
the Span-based DM, and (c) the color-based DM.
Fig. 9. The result using (a) the first change detection scheme and (b) the second
change detection scheme.
Tab. I gives the statistics results of the experimental results DMs in two different schemes. The proposed method shows
in Fig. 8 and Fig. 9. The overall alarms of the three DMs are potential in segmentating the PolSAR images simultaneously.
large, and the result of the first method is very close to one of Moreover, the experimental results using the proposed scheme
the three DMs. The combined DM produce the best change perform better than that using only one DM. In our future work,
detection result, the overall alarms decrease significantly. The we will explore the ability of other color-based features in
experimental results suggest that the superpixels show great change detection, like MPEG-7 color descriptors.
potential in reducing the false alarms, and the fusion of
different DMs can improve PolSAR change detection accuracy. ACKNOWLEDGMENT
This work was jointly supported by the National Natural
IV. CONCLUSIONS Science Foundation of China (Nos. 41371352 and 41331176),
In this paper, we proposed a superpixel-based PolSAR and the open project (EMW201510) funded by Key Laboratory
images change detection method, which involves Pol-InSAR for Information Science of Electromagnetic Waves.
image segmentation, DM generation, and the fusion of the