QOSM-U2CANet: a deep learning framework for normalized oil spill thickness and concentration mapping using multispectral satellite imagery

Marine oil spills pose significant ecological and socioeconomic threats, yet traditional remote sensing methods often fail to deliver accurate quantitative assessments of oil spill thickness and concentration across diverse environmental conditions. To address these limitations, this study presents QOSM-U2CANet, a novel deep learning framework that utilizes normalized thickness ( nT ) for non-emulsified oil spill (NEOS) and normalized concentration ( nC ) for emulsified oil spill (EOS) as physics-informed metrics derived from multispectral satellite imagery. The model incorporates a Residual Coordinate Attention U-Blocks (RCAU) to enhance spatial feature sensitivity and an L1 interval-constrained (L1-IC) loss function to improve detection precision in complex backgrounds. Evaluated across five oil spill-prone regions (Yellow Sea, South China Sea, Malacca Strait, Bay of Bengal, and Persian Gulf), QOSM-U2CANet achieves F1-scores of 0.855 for NEOS and 0.814 for EOS, with mean absolute errors (MAEs) of 0.077 for nT and 0.129 for nC , surpassing state-of-the-art models. The framework demonstrates resilience to environmental confounders and generalizes effectively across sensors with spatial resolutions of 10 to 50 m. Validation using historical Deepwater Horizon (DWH) oil spill retrospective estimates reveals a strong positive correlation ( R 2 = 0.73 ) and an Intersection over Union (IoU) of 0.74 for oil spill distribution extent, confirming the model’s ability to accurately represent real-world oil spill spatial extents and volume distributions. By transitioning from categorical classification to continuous pixel-level mapping, QOSM-U2CANet provides detailed spatial distributions of oil thickness and concentration, enhancing emergency response strategies and ecological impact assessments. This advancement bridges qualitative detection and quantitative inversion, establishing a scalable data-driven foundation for global marine oil spill monitoring, with potential for further refinement under varying sun glint conditions and dynamic spill scenarios.