Speeding Up Non-Gaussian POLSAR Image Analysis

Abstract

Non-Gaussian statistical models fit SAR data better than Gaussian-based statistics, in most cases, but are complicated and time-consuming to use for unsupervised image segmentation via probabilistic clustering. The more advanced the model, the more complicated and slow the clustering. The U-distribution has been demonstrated to be one of the most flexible models, capturing the Gaussian/Wishart, the K-distribution and the G 0 models as special cases, but can take hours or days to process a full sized image, depending on the chosen sensitivity. This work explains some computational improvements that drastically reduces the processing time, whilst maintaining the segmented results. The efficiencies are obtained by some smart re-parameterisations of the distribution model parameters allowing the use of look-up-tables for the density function evaluation and parameter estimation. Real images and computation times are shown as a demonstration.