Segmentation-based Change Detection Method for Remotely Sensed Images
Chi-Farn Chen, Chun-Shao Hsu
Center for Space and Remote Sensing Research
National Central University, Chung-Li, Taiwan.
Tel:(886)-3-4227151ext57659 Fax:(886)-3-4264301
Taiwan
E-mail: cfchen@csrsr.ncu.edu.tw,
92322089@cc.ncu.edu.tw
ABSTRACT
Land cover/land use change detection has been an important application of remotely sensed
images. Unlike the pixel-based c hange detection methods, which normally generate pixel-sized
noises, this study develops a segmentation-based c hange detection technique. The main idea of
the study is to apply region growing segmentation to multi-dated composite images, and divide
the images into many regions. Then a statistical significance test is used to find out the regions
that have been changed or not. An experiment is performed to detect the changed areas with
remote sensing images, and the result shows that the segmentation-based c hange detection
technique can filter out the pixel-sized noises, and offer relatively spotless and area-sized
changed map.
1. INTRODUCTION
One of the major applications of remotely sensed images obtained from earth-orbiting satellites
is land cover/land use change detection (Hazel, 2001). Because most of the satellite data is
repetitively recorded in digital and multi-spectral format, there are many conventional studies
focus on the digital techniques for image change detection, such as image differencing method
and post-classification comparison method. A critical and tricky element of the conventional
methods is the choice of the thresholds between change and no-change pixels that certainly will
influence the success of the method. Furthermore, conventional change detection methods
normally produce fragmental regions and salt-and-pepper noises due to their pixel-based
approach (Singh, 1989). In this study, a segmentation approach is used to develop change
detection algorithm for multi-dated remote sensing images. The method firstly composes the
multi-dated images as a composite image, then the region growing procedure is used to divide
the composite image into numerous regions (Yamamoto, 2001; Lee, 2002). Then each region
is statistically analyzed using significance test (Bernard, 1996). Basically, the chi-square
values calculated from the gray numbers of each region will be tested for their statistical
significance by null hypothesis. Then the region will be evaluated as changed area if the
chi-square value is greater than the threshold value under certain confidence level. The
following sections will describe the proposed method, test data and results, and conclusion.
