Automated image segmentation for synthetic aperture radar feature extraction

Abstract

Automated segmentation routines may be used to extract scattering features in synthetic aperture radar (SAR) images. The watershed transform segments real-valued images into regions associated with a local minima. Watershed algorithms suffer from over-segmentation which, for SAR image segmentation, results in many more regions than scatterers. We consider an algorithm called Peak Region Segmentation (PRS). PRS is an inverted version of the watershed transform that seeks to group pixel regions associated with a local maxima. We implement the algorithm to segment one, two, and three-dimensional images. We extend PRS to include region merging to avoid over-segmentation. Threshold settings allow the user to strike a balance between region merging and separation of closely-spaced scatterers. Image segmentation examples are shown for 1D, 2D, and 3D SAR images.