Unsupervised segmentation of multi-polarization SAR images based on amplitude and texture characteristics

This paper presents a new approach for the unsupervised segmentation of multi-polarization SAR images based on the statistics of both amplitude variation and texture characteristics. One co-polarized and one cross-polarized image is used in the classification. It involves two steps. In the first step, a window is used to scan the image and locate the clusters within it at each position. A merging procedure follows to combine them based on statistical similarity down to an appropriate number. Bayes maximum likelihood classification is then applied. In the second step, the authors adopt the second order Gaussian Markov random field models for image texture. Segments assigned for each class in the first step are examined and divided into sub-class groups if clear textural differences exist among them.