Reliable Sar oil-spill segmentation for marine monitoring Via self-supervised denoising and lightweight wavelet–state-space modeling

Reliable SAR oil-spill segmentation must simultaneously address image degradation caused by speckle and semantic confusion caused by look-alike dark features, such as low-wind areas, biofilms, and oceanic internal waves. To this end, this paper proposes the OS-AdaAttReNet-WSSNet framework, which decomposes the task into two complementary stages: self-supervised image restoration and supervised oil-spill segmentation. In the first stage, a self-supervised denoising network, AdaAttReNet, is constructed. Without requiring clean reference images, it integrates rotation-equivariant features, channel-spatial attention, and adaptive residual reconstruction to alleviate the effects of speckle on image texture, weak boundaries, and elongated structures. In the second stage, a lightweight wavelet-state-space segmentation network family, WSSNet, is developed. Multi-Scale Split Attention (MSSAtt), the Dynamic Enhanced Residual Wavelet Encoder (DE-RWE), and the Mamba-Enhanced Multi-scale Context Aggregation Upsampler (MEMCAU) are employed for multi-scale feature enhancement, wavelet-domain structural encoding, and long-range contextual fusion, respectively, thereby improving the recognition and boundary localization of irregular oil slicks. On the PALSAR and Sentinel subsets of the SOS dataset, the proposed framework achieves comparable or higher average segmentation performance than recent oil-spill-specific segmentation models. In particular, WSSNet-Tiny, with only 0.27 M parameters, substantially reduces model size and computational cost while maintaining competitive segmentation performance, demonstrating a favorable accuracy-efficiency trade-off. In addition, experiments on six external data subsets show that the WSSNet series exhibits a certain degree of direct transferability without using target-domain data for training or adaptation. The proposed framework provides a solution for reliable and efficient SAR oil-spill monitoring that balances segmentation accuracy, computational efficiency, and transferability.