Polarimetric Sar Image Segmentation Using Affinity Function from Probabilistic Boundaries and Patch Features

We investigate the segmentation of high resolution polarimetric Synthetic Aperture Radar (PolSAR) images of urban areas. The segmentation strategy in [1] is applied in this paper. Spectral graph segmentation has the advantage of capturing non-local property. The probabilistic boundaries and patch features are integrated for spectral graph segmentation. Accurate boundaries extraction and efficient patch features improve the segmentation. On the other hand, a better segmentation corresponds to a refined binary boundary map. Gradients of amplitude, texture, PolSAR CFAR edges and gradient magnitude are incorporated to produce an accurate probabilistic boundary map. These gradient features are combined in a supervised manner. The combination rules are learned from ground truth data using a logistic regression classifier. For a test image, a soft boundary map is generated by the classifier using all the gradient features. Probabilistic boundaries and patch features are integrated into affinity matrix construction [1] in spectral graph segmentation. Probabilistic boundary map provides an estimate of intervening contours. Learning of affinity function is treated as a supervised classification problem. Eigen decomposition of the affinity matrix results in spectral segmentations [2].