A Hybrid Framework With Scattering Distribution Perception and Edge Refinement for PolSAR Image Segmentation

Polarimetric Synthetic Aperture Radar (PolSAR) provides a more comprehensive and detailed characterization of target scattering, offering rich information both within and between image pixels. However, existing segmentation methods face challenges such as the difficulty in matching polarization characteristics with deep networks, underutilization of features, and complexity in edge segmentation. To address these issues, a segmentation framework is proposed that integrates Diffusion-based polarization parameter optimization and distribution perception, multi-scale and multi-pol feature fusion, and edge enhancement. First, it leverages Transformer and CNN to adaptively extract, analyze, and fuse features from co-polarization and cross-polarization data. Second, a POA regression and compensation module is incorporated to obtain compensated features. The segmentation results derived from these features are combined with the general results through weighted averaging, enabling mutual supervision and improving final precision by reducing misclassification in complex scattering regions. Additionally, a polarization edge-aware enhancement strategy is introduced to refine segmentation boundaries and reclassify uncertain points. Overall, the framework establishes a hybrid mapping relationship between polarization, texture, statistical properties, and label distributions, enabling effective PolSAR image segmentation. Extensive experimental results and discussion demonstrate the superior performance of our method.