SAR change detection based on generalized Gamma distribution divergence and auto-threshold segmentation

Abstract

Based on the clutter statistical characteristics of SAR image, this paper takes advantage of the generalized Gamma model to fit the filtered and co-registered SAR images, in order to gain the characteristics information, such as radiation value, local texture, etc. Then, the degree of evolution between the statistical characteristics of multi temporal SAR image is measured by the definition of Kullback-Leibler Divergence in information theory. Afterwards, a combination of KS and KL test has been applied into the evaluation of fitting function for the difference map captured in the former step, which help select the best fitting function automatically for the model-based KI threshold segmentation. Experiment was carried on the multi temporal SAR images for Southern Part of Tianjin, acquired by Radarsat-1/2, as well as Shunyi District of Beijing, acquired by Envisat-ASAR. Such results confirmed the method proposed in this paper not only avoid large number of false alarms generated from the changes of surface corrugation, but also effectively detected the regions ignored by traditional methods, which have no variance in mean value, but differ in texture.