Sar Image Change Detection Based On Generalized Gamma Distribution

For the change detection technique of synthetic aperture radar (SAR) images based statistical model, its detection performance depends on whether the selected model can accurately characterize the statistical properties of the images of interest. Limited to the accuracy of traditional distribution model, we make use of the more flexible and powerful generalized Gamma distribution (GFD) as the modeling tool to carry out the change detection of multitemporal SAR images.The main research contents of this thesis are two-fold in the following:1) The generalized Kittler-Illingworth threshold (GKIT) algorithm, proposed by G. Moser, constructs statistical probability distribution function of ratio image by ratio probability density function (PDF) deduced from Nakagami, Weibull and Lognormal distribution, with the assumption of the same equivalent number of looks (ENL) for two original SAR images. In view of the flexibility and effectiveness of the GΓD, and considering the assumption of the same ENL cannot be satisfied in general because of the de-correlation between multitemporal SAR images, this paper proposes an improved GKIT algorithm based on GΓD by relaxing the ENLs to be different for two SAR images acquired in the same geographical area but at two different time instances. The algorithm models two original SAR images by GFD, and deduces the corresponding PDF of the ratio images. Based on the high degree of freedom characteristics of the ratio PDF, we estimate the corresponding parameters in the original-image domain by mapping the pixel coordinates of the changed and unchanged classes of ratio image to two original SAR images, rather than directly in the ratio-image domain. Further, the method of logarithmic cumulants (MoLC) is introduced to estimate the corresponding parameters of the involving distributions from the class-dependent data in original two SAR images to obtain the PDF of ratio image. Finally, the GKIT algorithm is used to adaptively determine the optimal threshold for obtaining the results of change detection.2) To reduce the influence of speckle and to preserve the edge and detail information, the change detection of SAR images based on the divergence of GFD is investigated in the wavelet domain. The Kullback-Leibler (KL) distance is introduced as the change factor to measure the degree of difference between two SAR image data in the wavelet domain. Specifically, multiscale wavelet decomposition is implemented for two original SAR images, and then the wavelet coefficients at different scales are modeled as the generalized Gamma distribution, respectively, to calculate the KL distances between wavelet coefficients in the same directions at different scales. Considering the inter-and intra-correlation of wavelet decomposition, the products of KL divergences at same direction but different scales are calculated for three detailed subbands, whose sum is proposed as the final difference between the two SAR images in the wavelet domain. The receiver-operating characteristic (ROC) curve is used to adaptively determine the optimal threshold for binary change detection.