Semantic segmentation of Polarimetric SAR imagery using Conditional Random Fields

The paper proposes a fast and accurate semantic segmentation approach for a large Polarimetric SAR (PolSAR) image using Conditional Random Fields (CRFs). It efficiently incorporates the polarimetric signatures, texture and intensity features into a unite CRFs model, and employs a fast max-margin training method for parameters learning. Experiments on RadarSat-2 PolSAR data in Flevoland test site demonstrate that our approach achieves precise segmentation results with a few well-selected training samples.