Automatic PolSAR Segmentation with the U-distribution and Markov Random Fields

Abstract:
A novel unsupervised, non-Gaussian and contextual clustering algorithm is demonstrated for segmentation of Polarimetric SAR images. Previous works have shown the added value of both non-Gaussian modelling and contextual smoothing individually, and goodness-of-fit techniques were introduced to determine the appropriate number of statistically distinct classes. This paper extends our previous work by using the more flexible, two parameter, U-distribution model and includes a Markov Random Field approach for contextual smoothing, without losing the benefits of the goodness-of-fit testing. The proposed, fully automatic, algorithm is demonstrated with both simulated and real data-sets.