Adaptive Statistical Superpixel Merging Method for PolSAR Image Segmentation

Abstract

Region-based segmentation plays a fundamental role in polarimetric synthetic aperture radar (PolSAR) image interpretation. In this paper, an efficient and adaptive statistical superpixel merging approach with edge penalty for PolSAR image segmentation is proposed. First, we introduce a directional span-driven adaptive (DSDA) window into the edge detection and produce the superpixels using the Pol-ASLIC with edge constraint. Based on the initial superpixel partition, pixels belonging to the same superpixel should be merged immediately. Since the statistical region merging (SRM) method exhibits promising segmentation performance in solving significant noise corruption, we choose this framework to achieve statistical superpixel merging for PolSAR data in the second stage. However, the traditional SRM method starts from the pixels and cannot be directly applied to PolSAR data with multiplicative speckle noise and polarimetric information. This research proposes to define a new dissimilarity measure between the superpixels, which takes the edge penalty into consideration. Therefore, we have a reasonable and accurate merging order for the superpixel pairs. With regard to the merging predicate of superpixels, a polarimetric homogeneity measurement is used to redefine the merging threshold, making the merging predicate and the merging threshold adaptive to PolSAR image content. More importantly, the image complexity parameter of the traditional SRM method is eliminated, making the proposed approach be free of parameters and easy to use. Experimental results shows that our proposed approach can effectively improve the computation efficiency and segmentation accuracy in comparison with the state-of-the-art merging-based methods.