Superpixel generation for polarimetric SAR using Hierarchical Energy maximization

Abstract

Superpixel provides a coherent division of an image that respects the integrity of objects and has been used in Polarimetric Synthetic Aperture Radar (PolSAR) image analysis. The generation of superpixels is sensitive to noise and cluttered objects, which is typical in SAR imagery. To address this issue, we proposed a novel PolSAR Hierarchical Energy Driven (PHED) method to handle PolSAR images. A hierarchical structure is built and an energy driven hill climbing strategy is employed. In the coarse level superpixel generation, histogram intersections of coherency matrix are used to divide the image into raw superpixels. In the fine level superpixel generation, our method evaluates each superpixel using Wishart energy. A boundary term is included in the pixel-level update to enable boundary choices. Experiments are conducted with UAVSAR, AIRSAR, and simulated datasets. Our results demonstrate that the proposed PHED outperforms Simple Linear Iterative Clustering (SLIC), Linear Spectral Clustering (LSC), TurboPixels (TP), and Superpixels Extracted via Energy-Driven Sampling (SEEDS) in terms of retaining fine boundaries of ground objects and accuracy. The robustness of PHED is also confirmed.