An Alternating Mamba Selective Mechanism GANs for Oil Spill Semantic Segmentation

The limited availability of annotated oil spill data restricts supervised segmentation models from capturing comprehensive contextual semantics, while also causing them to assimilate speckle noise characteristics during training. This noise interference blurs semantic boundaries and compromises detection accuracy. To overcome these limitations, this work introduces an alternating Mamba selective mechanism GANs for oil spill semantic segmentation (AMGANs) incorporating Dynamic Feature Processing through an alternating linear-time sequence modeling with structured state-space models (Mamba-ssm) and ECABlock. The proposed framework synergizes adversarial learning with adaptive feature refinement. Specifically, Mamba-ssm and ECABlock complement each other in global-local modeling and spatial-channel dimensions, balancing efficiency and expressiveness—Mamba-ssm leverages feature selectivity to separate oil spill signatures from complex background clutter, while ECABlock enhances boundary delineation by optimizing channel-wise feature weights. By alternating between Mamba-driven selection and ECABlock-based refinement, the model achieves coherent feature integration while minimizing interference between modules. Validation on Sentinel-1 SAR data confirms that AMGANs outperforms existing approaches, attaining a segmentation accuracy of 97.34% and a segmentation mIoU of 78.29%.