Change Detection using TerraSAR-X data

Abstract

The objectives of this thesis is to find changed areas caused by natural disaster from two coregistered calibrated SAR images. Three methods are used in this thesis. The first method histogram thresholding uses the histogram of the SAR intensity ratio image to classify the ratio image into three classes. This technique was originally proposed by Kittler et al. (1986) and modified by Bazi et al. (2005) and Moser et al. (2006) based on the Bayesian formula. In this thesis their methods are combined together to detect three classes. The relative difference of the cost function is used to detect the number of the classes instead of the determinant of the Hessian matrix suggested by Bazi et al. (2005). The second method formulates the classification problem as a hypothesis testing problem. This idea was originally used by Touzi et al. (1988) and Oliver et al. (1996). In this thesis the analytical method by Touzi et al. (1988) is replaced by using the properties of the Gamma distribution. The third method graph-cut algorithm is a post-processing method, which improves classification results from the first and second methods. The improvement is equivalent to the global optimization of an energy function in a MRF. A modern method proposed by Kolmogorov et al. (2004) and Boykov et al. (2004) is used in this thesis. This method transforms the energy function of a MRF into an equivalent graph and solve the global optimization problem using a max-flow/min-cut algorithm. These three methods are applied to the test data on Queensland, Australia and Leipzig, Germany. The most SAR ratio images can be classified into three classes successfully. The remaining problem is that the interpretation of the changed classes is still ambiguous. Other data sources should be combined to assist or improve the interpretation of the detected change.