A High-Efficiency Automatic $U$ -Distribution Segmentation Algorithm for PolSAR Images

Abstract

A fully automatic, non-Gaussian, and contextual clustering algorithm for segmentation of polarimetric synthetic aperture radar (SAR) images has been previously presented by Doulgeris. It achieved good results for both simulated and actual data sets. However, the long computation time was its main drawback. This letter discusses modifications to improve computational efficiency. The primary speed issues were rooted in the complicated probability density function (PDF) of the adopted model, for which evaluating the posterior probability of samples and estimating the parameters were both very time-consuming. We investigate the model parameters, reparametrize the model, and introduce lookup tables to speed up the processing chain. The new strategy speeds up both PDF evaluation and parameter estimation while maintaining the exactly similar visual results and now makes advanced non-Gaussian SAR image analysis a practical alternative.