Hybrid Deep Learning for Oil Spill Mapping Leveraging Sentinel-2 and Foundation Models

Oil spills represent a persistent environmental threat requiring fast and accurate monitoring. While SAR-based methods have traditionally been dominant, multispectral satellite imagerycan offer rich spectral data for oil detection. However, challenges persist such as low number of recorded oil spill incidents, label noise and effects including sunglint, atmosphericinterference, and spectral confusion. This work proposes a hybrid deep learning framework that integrates a water-specialized foundation model (Hydro Foundation Model) with amarine-targeted segmentation architecture (MariNeXt) to improve the accuracy and generalization capacity of oil spill detection from Sentinel-2 imagery.