Marine oil spill pollution detection in remote sensing imagery using semantic sub-prototype modeling

Remote sensing monitoring of marine oil spills supports marine environmental protection, pollution surveillance, and emergency response. However, accurate detection remains difficult because oil-water boundaries are blurred and irregular, oil slicks are often small or fragmented, and oil spill look-alikes and complex backgrounds cause false alarms and missed detections. In unmanned aerial vehicle (UAV) RGB and SAR imagery, water-surface reflections, harbor structures, and dark-spot look-alikes further complicate oil spill identification. To address these challenges, we propose SSP-Net, a semantic sub-prototype network for marine oil spill pollution detection. Specifically, we design a multi-level feature aggregation module that combines shallow spatial details with deep semantic cues to represent weak boundaries and small oil spill regions. Second, we develop a class-semantic sub-prototype modeling mechanism that uses multiple fine-grained sub-prototypes to capture diverse appearances within oil slick, look-alike, water-body, and complex-background categories. In addition, we introduce a prototype-feature interaction decoder (PFID) that iteratively refines semantic sub-prototypes and image features through cross-attention, strengthening discrimination between oil slicks and interfering categories. Finally, we design a multi-class joint loss (MCJL) that reinforces segmentation supervision and semantic constraints on sub-prototypes. Experiments on the public M4D and Oil-Spill-Drone datasets demonstrate that SSP-Net outperforms several advanced methods, achieving 72.97% mIoU on M4D and 93.42% mIoU on Oil-Spill-Drone. The segmentation masks generated by this model can provide critical spatial information for marine pollution monitoring, emergency response, and environmental management.