Simultaneous segmentation of texture properties of K-distributed SAR images

This paper proposes a technique for refining a previous segmentation by using different properties of K-distributed SAR clutter simultaneously. We first consider approximate forms for the K distribution based on a Maximum Entropy approach which assumes that only two moments of the data can be estimated with sufficient accuracy over a small sample region. The choice of moments defines the form of the approximate probability density functions (PDF). After the initial segmentation we then propose a post-processing stage in which the values of the moments of the complete previously-identified segments are assumed exact and an optimum fit to the edge position is defined. We demonstrate that joint estimation of the edge position, based on estimates of the mean of the data and of its logarithm, provides a close approximation to the full K- distribution treatment, while being significantly simpler to implement.