Enhanced segmentation of SAR images using non-Fourier imaging

This paper demonstrates that synthetic aperture radar (SAR) images formed using modern spectral estimates can be more accurately segmented than traditional SAR images. Classical FFT based Fourier image formation algorithms produce imagery with strong speckle and sidelobe artifacts that hinder the segmentation process. We show that imagery formed using Capon's minimum variance spectral estimate changes the statistics of the SAR imagery in a way that increases the separation between the various classes of natural terrain. The increased class separation leads to more accurate segmentation. We use the MSTAR dataset to show the statistical changes and demonstrate the improvement in segmentation relative to Fourier imagery.