Texture Segmentation of Sar Images

The image surface of synthetic aperture radar imagery (SAR) is dominated locally by peaks and clusters of peaks, especially in terrain and vegetation areas. More globally, there are extended regions dis-tinguishable by texture, trees, elds, shadows, roads, etc. We describe an algorithm which segments SAR images into a set of regions of pre-speciied classes, based on two procedures: rst, the classiication of peaks into N = 3 pre-speciied classes, and second, a segmentation of the Delaunay triangulation of peaks into connected regions. A peak detection operator is used to estimate peaks in SAR images; thresholds are determined by using the histogram of the peak amplitude of each class. Peak amplitude was found to be the most useful discriminant by far in the multi-variate distribution in peak amplitude, peak width, and peak density. A Delaunay triangulation was established on the peaks of each class. Links in the triangulation were removed if they were unlikely for a population of that class. The boundary of a texture region is the boundary of a connected component of the modiied Delaunay triangulation of the appropriate class of peaks. Linking by boundary traversal was developed to extract closed boundaries of each class. Experimental and simulation results are presented in SAR and synthesized images, respectively. Boundaries of regions can be determined to an accuracy of about 2 pixels.