Superpixel Segmentation of Polarimetric SAR Images Based on Integrated Distance Measure and Entropy Rate Method

Abstract

This paper proposes to integrate two different distances to measure the dissimilarity between neighboring pixels in PolSAR images, and introduces the entropy rate method into PolSAR image superpixel segmentation. Since the Gaussian model is commonly used for homogeneous scenes and less suitable for heterogeneous scenes, we adopt the spherically invariant random vector (SIRV) model to describe the back-scattering characteristics in heterogeneous areas. Moreover, a directional span-driven adaptive (DSDA) region is proposed such that it contains independent and identically distributed samples only, thus it can obtain accurate estimation of the distribution parameters. Using the DSDA region, the Wishart distance and SIRV distance are calculated, and then combined together through a homogeneity measurement. Therefore, the integrated distance takes advantage of the SIRV model and the Gaussian model, and suits both homogeneous and heterogeneous areas. Finally, based on the integrated distance, the superpixel segments are generated using the entropy rate framework. The experimental results on ESAR and PiSAR L-band datasets show that the proposed method can generate homogeneity-adaptive segments, resulting in smooth representation of the land covers in homogeneous areas, and better preserved details in heterogeneous areas.