CDANet: A context-detail aware network for marine oil spill detection in SAR imagery

Marine oil spills pose a serious threat to marine ecosystems and coastal economies, highlighting the need for accurate and efficient detection technologies. Synthetic Aperture Radar (SAR) combined with deep learning offers promising solutions, yet real-world spill scenarios exhibit challenges such as diverse scales, blurred boundaries, and irregular shapes. Existing oil spill datasets lack generalization, while current deep learning methods face limitations in modeling long-range dependencies, extracting multi-scale features, and fusing encoder-decoder representations. To address these issues, this study constructs a comprehensive SAR oil spill dataset incorporating multiple years, geographical regions, spill morphologies, and polarimetric modes, based on Sentinel-1 and GF-3 data. Additionally, we propose the Contextual Detail-Aware Network (CDANet), which captures both global semantics and local details through the integration of the Mamba state-space model, a multi-scale cross-attention mechanism, and a dilated orthogonal feature fusion strategy. CDANet achieves precision scores of 95.86 % on Sentinel-1 and 92.87 % on GF-3, with corresponding F1-scores of 94.38 % and 90.43 %, surpassing existing approaches. Models trained on our dataset improve average accuracy by up to 8.3 % compared to those trained on the Deep-SAR Oil Spill (SOS) dataset in real-world scenarios. This work provides essential technical support and data resources for enhancing oil spill response. The dataset is available at https://github.com/SARDEEP1/OSLM.