OilSegSARFormer: Adapter-Mix Transformer Backbone with Gated Detail-Refined Semantic Decoder for multi-class oil spill segmentation in SAR imagery

Synthetic aperture radar (SAR) is widely used for marine oil spill detection, yet oil spills and look-alike phenomena such as low-wind zones, rain cells, and sea ice all appear as dark spots in SAR imagery. This ambiguity, compounded by speckle noise, makes high-accuracy multi-class segmentation a persistent challenge. To address these issues, we construct a five-class pixel-level segmentation dataset from Sentinel-1 SAR imagery covering Background, Low Wind, Oil Spill, Rain Cell, and Sea Ice, addressing the lack of fine-grained look-alike annotations in existing benchmarks. We propose OilSegSARFormer, a hybrid encoder–decoder architecture. The encoder adopts a pretrained Mix Transformer backbone with lightweight Local Adapters at shallow stages to inject SAR-specific texture and edge priors. The decoder, Gated Detail-Refined Semantic Decoder, prioritizes deep semantics for global consistency and selectively incorporates shallow detail corrections via Gated Fusion Module and Cascaded Refinement Module for progressive boundary refinement. A hybrid cross-entropy and Lovász-Softmax loss directly optimizes mIoU while alleviating class imbalance. Experiments show OilSegSARFormer achieves 73.43% mIoU and 84.14% F1-score, outperforming 11 competing methods, with Oil Spill IoU reaching 66.15%, exceeding the runner-up by 4.83 percentage points. The code and model weights are available at https://github.com/su-17-97/oilseg-open .