Accurate oil spill segmentation based on synthetic aperture radar (SAR) is crucial for near-real-time marine pollution monitoring and emergency response. However, the black-box nature of deep networks limits the reliability assessment of segmentation results, while polarization feature-based methods often involve cumbersome processing and are highly dependent on prior information. To address these issues, this paper proposes a polarization matching guided network (PMG-Net) for SAR oil spill segmentation. PMG-Net introduces polarization matching coefficient (PMC) and a polarization matching estimation head (PMEHead) based on (VV, VH) channels to characterize the differences in oil-water polarization responses under different sea states. A dual-branch encoder extracts intensity and polarization matching features separately. The intensity-physical fusion module (IPFM) adaptively utilizes PMC and improves the analyzability through cross-attention fusion and physical gating fusion. To further identify elongated oil spills and weak boundaries, a multi-scale feature enhancement module (MSFEM) is developed. Experimental results show that PMG-Net outperforms other representative segmentation models in terms of IoU_Oil, mIoU, and F1, while achieving better boundary preservation and background suppression. These results demonstrate the effectiveness of polarization matching information in SAR oil spill segmentation based on physical information.
🌊
