SAR Image Change Detection Using SURF Algorithm
Seo Li Kang
School of Electronics and Telecomm. Eng., Korea Aerospace University, Hwajeon-Dong, Goyang, Republic of Korea
Woo Kyung Lee
School of Electronics and Telecomm. Eng., Korea Aerospace University, Hwajeon-Dong, Goyang, Republic of Korea
Keywords: SAR image fusion, geometrical correction, SURF, change detection, KI Thresholding .
Abstract: With the advent of high-resolution Synthetic Aperture Radar (SAR), applications of satellite SAR have a
growing interest in this field and change detection is of high interest in both military and civil applications.
Change detection techniques have attracted increased attentions and become a topic of major research. In
change detection procedure, geometrical correction of image is essential for effective remote sensing
applications. Unlike optical sensor, the geometrical correction of SAR images is highly complicated due to
the signal interaction within the complex geometrical properties of the target structures and the inherent
speckle noise. In this paper, we present an advanced yet efficient geometrical correction method that may be
applied to multi-resolution satellite SAR images. For this purpose, SURF(Speeded-Up Robust Feature) is
adopted and modified so as to make it fully applicable to SAR images. KI thresholding technique is
constructed and applied to multi-SAR images to verify the performance.
1 INTRODUCTION
Change detection is the process of identifying
differences in the state of an object or phenomenon
by observing it at different times (Lu et al., 2004).
The goal of change detection is to detect
"significant" changes while rejecting "unimportant"
ones (Radke et al., 2005). Generally, important
changes reflect natural phenomenon or human
activities in the Earth. In recent years, with the
development of high-resolution imaging satellites,
the remote sensing has rapidly advanced that the
minute details of the Earth surface could be
investigated. Naturally, techniques for change
detection have attracted increased attentions and
become a topic of major research. Unlike optical
image, geometrical distortions commonly occur
during image acquisition of SAR images due to the
geometrical characteristics of the targets. Therefore,
geometrical correction of images plays an important
role in this area. Geometric correction of the SAR
images could be performed by utilizing satellite’s
orbit and attitude information. However, due to the
inevitable errors in measuring SAR sensor’s orbit
and attitude information, and the acquisition errors
of Earth’s geographical parameters, the reflected
radar signals are contaminated with considerable
geometric errors. In order to acquire more accurate
geometrical corrections, error correction needs to be
performed with respect to the ground control points.
Regardless of the methodology adopted, int is
prerequisite to carry out accurate geometrical
correction for the change detection to be properly
applied, especially when the image resolution
improves below sub-meter levels.
In general, geometric correction techniques can
be classified into three categories, which are based
on intensity, transform domain and feature
characteristic respectively. Feature-based technique
is simple and steady and hence has been widely used
for this purpose in the past. A number of detectors
are known to be applicable for this purpose, which
include Harris corner detector, Forstner detector,
Moravec detector, Harris-Laplacian detector,
Gaussian-determinant detector, Hessian detector and
Fast-Hessian detector.
68
Kang S. and Lee W.
SAR Image Change Detection Using SURF Algorithm.
DOI: 10.5220/0005421400680073
In Proceedings of the Third International Conference on Telecommunications and Remote Sensing (ICTRS 2014), pages 68-73
ISBN: 978-989-758-033-8
Copyright
c
 2014 by SCITEPRESS – Science and Technology Publications, Lda. All rights reser ved

Among others, SIFT(Scale Invariant Feature
Transform) is one of the most widely used method
for point feature detection. SIFT algorithm is known
to be steady and resistant to geometric deformations
and illumination changes for partial target matching
and recognition. When the feature points have good
lamination changes for partial target matching and
recognition, then feature points has stability in terms
of geometrical variation. However SIFT algorithm
suffers from a drawback of consuming a significant
amount of time for searching and calculating the
matching points. Hence SURF algorithm has been
proposed as a means to replace the SIFT by
simplifying complicated SIFT algorithm while the
general feature of high accuracy and stability are
well preserved. It is asserted that SURF is several
times faster than SIFT and claimed by its authors to
be more robust against different image
transformations than SIFT. In this paper, we adopt
the previous SURF algorithm for SAR change
detection. Parameter estimation and optimization are
carried out to obtain the best matching results.
1.1 SURF Algorithm
SURF is a useful tool for matching points between
different images. After SURF is performed,
RANSAC is applied to obtain optimal coordinate
conversion functions using the extracted matching
points. After ground point selection is completed
and conversion function is obtained, projective
transformation and second order polynomial
transformation is used to rearrange the remaining
pixels into the new coordinates. Projective
transformation needs four or more ground points and
second order polynomial needs minimum 4 ground
points. Higher multinomial matrix might be required
for the optimal geometrical correction depending on
the level of distortion in the images and the number
of extracted ground points.
SURF algorithm has been mainly applied to
optical image cases due to its inherent nature of
requiring high contrast of pixel values. Y.Murali has
developed a MOSAIC image stitching method based
on SURF algorithm. In other case, an input query
imaging technique is proposed that SURF feature
points detector algorithm is incorporated with the
color edge matching method.(Ryo Mitsumori, 2009)
A modified SURF algorithm is developed by
adopting the color and relative location information
of interest points.(KyungSeung Lee, 2012) However
the majority of the previous researches have dealt
with optical images are SAR images have rarely a
topic of research. This is mainly attributed to the low
level of contrast in radar image pixels and inherent
speckle noises around feature points. For this reason,
it has been considered a difficult task to apply SURF
for SAR change detection purpose. In this paper, we
have adopted higher multinomial matrix equation in
the process of applying SURF to the SAR images.
By doing this, the probability of erroneous feature
point detection is reduced and the point matching
can be better performed against target movement,
rotation and scale. After desired ground points are
selected, the remaining pixels are rearranged on the
new coordinate plane using a cubic convolution. We
show that the modified SURF algorithm can be
easily applied to the conventional SAR images. The
performance of the proposed algorithm is verified
throughout medium-to-low level resolution satellite
SAR images.
1.2 KI Thresholding
For a simple change detection purpose, image
thresholding is the most straightforward technique.
There are a number of known image thresholding
techniques suitable for different applications. Sezgin
et al, has shown that the clustering-based method by
Kittler and Illingworth (1986) provides the most
reliable thresholding result in their experiments. This
technique has been called KI thresholding and is
performed as: Step 1. Choose an arbitrary initial
threshold T. Step 2. Compute priori probability,
mean value, variance value. Step 3. Compute the
updated threshold. Step 4. Compare old threshold
with updated threshold. KI thresholding technique
has been widely applied in the past since thanks to
its robust performance. In this paper, we adopted KI
technique variation for the purpose of change
detection between two different SAR images.
2 EXPERIMENT RESULTS
In Fig. 1 and Fig.2 are shown Radarsat-1,2 images
taken over Vancouver respectively and their
characteristics are described in Table 1 and 2. SAR
images can be seen totally different depending upon
the sensor position and viewing angle over the same
scene. In this case, we have attempted to detect
changes for Radarsat-1 and Radarsat-2, which are
distinguished by different resolutions and their
performances are compared later. Since the scene
coverage is different from each other, regions of
interest are extracted first. Then SURF algorithms
are applied to both images to find out the common
feature points that matched with each other.
Sar Image Change Detection using Surf Algorithm
69

Table 1: RADARSAT-1 IMAGERY
Name
Characteristics of products
Acquisition date
Orbit
description
SSG
1999.07.25.
Descending
19427
Map image
SGF
1998.07.09
Ascending
13975
Path image
Table 2: RADARSAT-2 IMAGERY
Characteristics
Description
Name of Satellite
Radarsat-2
Beam mode
FineQuad15
Product type
SLC
Figure 1: RadarSAT-1 Vancouver area, (a)SSG (b)SGFT
Figure 2: RadarSAT-2 Vancouver area, HH, HV data
Figure 3: ROI(Region Of Interest) images in Fig.2
Figure 4 presents the number of interest points
extracted from each SAR images according to the
Hessian threshold. As the Hessian threshold value is
increased, the number of extracted interest points is
reduced. Figure 5 shows the number of matched
points with respect to the varying SURF matching
threshold values. SURF matching points are selected
by calculating the Euclidean distance around
extracted interest points. The number of the selected
matching pixels increases as the SURF matching
threshold value is increased from 0.4 to 0.9.
Processing times to extract interest points and
matched points are shown in Figure 6. As Hessian
threshold is increased, processing time is reduced,
but the number of matched point is decreased.
Therefore, there should be an appropriate trade-off
between time consumption and accuracy to
guarantee the optimization of the image matching.
2000 2500 3000 3500 4000 4500 5000
400
500
600
700
800
900
1000
1100
1200
1300
1400
SURF Interest Points
Hessian threshold
Num. of Interest Points
input image
reference image
Figure 4: The number of interest point according to
Hessian Threshold (Radarsat-2)
2000 2500 3000 3500 4000 4500 5000
100
200
300
400
500
600
SURF Matched Points
Hessian threshold
Num. of SURF Matched Points
Mathcing Threshold : 0.4
Mathcing Threshold : 0.5
Mathcing Threshold : 0.6
Mathcing Threshold : 0.7
Mathcing Threshold : 0.8
Mathcing Threshold : 0.9
Figure 5: The number of matched points by Euclidean
distance calculation (Radarsat-2)
Third International Conference on Telecommunications and Remote Sensing
70

2000 2500 3000 3500 4000 4500 5000
0.35
0.4
0.45
0.5
0.55
0.6
0.65
0.7
SURF Processing Time
Hessian threshold
SURF Processing Time [sec]
Mathcing Threshold : 0.4
Mathcing Threshold : 0.5
Mathcing Threshold : 0.6
Mathcing Threshold : 0.7
Mathcing Threshold : 0.8
Mathcing Threshold : 0.9
Figure 6: SURF Processing Time according to Hessian
threshold (Radarsat-2)
Extracted interest points and matched points are
shown in Figure 7 and Figure 8 for Radarsat-1.
Similar procedure is performed against Radarsat-2
and shown in Figure 9 and Figure 10 respectively.
As mentioned earlier, the performance of the
matching algorithm depends upon the parameters in
the SURF, particularly the threshold value. In our
experiment for Radarsat-2 images, the Hessian
threshold value of 2000 and matching threshold
value of 0.5 have provided relatively good
performance while the calculation burden is
minimized. After the matching sequence is
completed, geo-correction procedure is followed
using the projective transform for image matching of
two satellite images. The performance of image
matching is measured as the distance of the
dislocation of common pixels, calculated as residual
sigma. The matching accuracy and error levels are
calculated and compared with each other for
Radarsat-1 and Radarsat-2 and summarized in Table
3. The result shows that matching accuracy in
Radarsat-1 is 42.5% and matching accuracy in
Radarsat-2 is 94.6%. Matching accuracy in low
resolution images is smaller than high resolution
image, which are as expected. It can be inferred that
the Hessian threshold in low-resolution case has to
be increased than that of the high-resolution image
case.
Figure 7: Extracted interest points and matched points
(Radarsat-1)
Figure 8: Result of SURF matching (Radarsat-1)
Figure 9: Extracted interest points and matched points
(Radarsat-2)
Figure 10: Result of SURF matching (Radarsat-2)
Table 3: Matching results between the satellite SAR
images (Hessian threshold = 2000, Matching threshold =
0.5)
Radarsat-1
Radarsat-2
SURF Interest Points
11865/23164
1319/1483
SURF Matched Points
40
56
Matching Accuracy
42.5%
94.6%
Residuals sigma_x
[pixel]
1.303
1.018
Residuals sigma_y
[pixel]
1.176
1.009
In SAR imaging mode, comparison between two
different images is difficult due to the residing
speckle noise and ratio comparison tend to be
perferred for the high noise image application.
Sar Image Change Detection using Surf Algorithm
71

As in the SURF case, it is important to select a
proper threshold value when KI thresholding method
is applied. In our experiments, the optimal threshold
value was chosen to be 0.057 for Radarsat-1 and
0.026 in Radarsat-2. Pixels with intensity values
bigger than the optimal threshold are extracted in
Figure 12. These areas are considered to implicate
changes over the scenes. Here it is seen that high
resolution SAR provide more realistic details of the
changes between different images.
image ratio
Image X [pixel]
Image Y [pixel]
Figure 11: Ratio image (Radarsat-2)
Figure 12: results of applying KI thresholding
(a)Radarsat-1 (b)Radarsat-2
3 CONCLUSION
Accurate geometrical correction of radar imagery is
essential for effective change detection procedure.
The geometrical correction of SAR images is highly
complicated due to the complex geometrical
properties of the targets, signal interaction within
target structures and speckle noise. In this paper, we
present a modified SURF algorithm to geo-
registration on SAR image. The KI thresholding
technique has been applied to detect changes over
medium resolution SAR images. It is shown that the
parameter selection of SURF algorithm needs to be
carefully designed on applying SAR image by
adjusting the Hessian threshold value according to
resolution of images.
ACKNOWLEDGEMENTS
The work is supported by KARI (Korea Aerospace
Research Institute) of Republic of Korea.
REFERENCES
H. Bay, T. Tuytelaars, and L. V. Gool, “SURF : Speeded
Up Robust Features”, European Conference on
Computer Vision, Vol. 3951, pp. 404-417, 2006
Kim, T. and Im, Y “Automatic satellite image registration
by combination of matching and random sample
consensus”, IEEE Trans. on Geoscience and Remote
Sensing, Vol. 41, No. 5, pp. 1111-1117, 2003
Goshtasby, A. A., “2-D and 3-D image registration - for
medical, remote sensing, and industrial applications”,
John Wiley & Sons Inc., New York, pp. 4-5, 2005.
B. Zitova and J. Elusser, “Image registration methods: a
survey”, Image and Vision Computing 21, pp. 977-
1000, 2003
Flora Dellinger, Julie Delon, Yann Gousseau, Julen
Michel and Florence Tupin, “SAR-SIFT: A SIFT-
LIKE ALGORITHM FOR SAR IMAGES”,
Geoscience and Remote Sensing Symposium (IGARSS),
2012 IEEE International, pp. 3478 – 3481, 2012.
Mark D Pritt, Kevin J LaTourette, “Automated
Registration of Synthetic Aperture Radar Imagery to
LiDAR, IGARSS 2011, July 2011.
T. Lindeberg, “Feature detection with automatic scale
selection”, International Journal of Computer Vision,
Vol. 30, No.3, PP. 79-116, 1998
Kittler, J. and Illingworth, J., ‘Minimum error
thresholding’, Pattern recognition 19(1), 41–47.1986
Sezgin, M. et al. ‘Survey over image thresholding
techniques and quantitative performance evaluation’,
Journal of Electronic imaging 13(1), 146–168. 2004
Y.MURALE., ‘Image Mosaic Using Speeded Up Robust
Feature Detection’, International Journal of Advanced
Research in Electronics and Communication
Engineering (IJARECE) Volume 1, Issue 3, September
2012
Ryo Mitsumori, Hideaki Uchiyama, Hideo Saito, Myriam
Servieres, and Guillaume Moreau, ‘Change Detection
Based on SURF and Color Edge Matching’,
Proc.ACCV, 2009
Kamel Besbes Rochdi Bouchiha, ‘Automatic Remote-
Sensing Image Registration Using SURF’,
International Journal of Computer Theory and
Engineering, Vol. 5, No. 1, February 2013
Third International Conference on Telecommunications and Remote Sensing
72

Woo-Kyung Lee and Ah-Leum Kim, ‘An Efficient
Automatic Geo-Registration Technique for High
Resolution Spaceborne SAR Image Fusion’, IGASS
2011, July, 2011
KyungSeung Lee, Daehoon Kim, Seungmin Rho and
Eenjun Hwang, ‘Improving Matching Performance of
SURF Using Color and Relative Position’, The Korea
Navigation Institute, Vol.16., No.2, April, 2012
Smith, J., 1998. The book, The publishing company.
London, 2
nd
edition.
Sar Image Change Detection using Surf Algorithm
73