EN-HAN DeepNet: a category-focused SAR oil spill detection approach tackling edge ambiguities and look-alikes interferences

Marine oil spill has caused serious harm to human life and health, making continuous and reliable monitoring of sea surface pollution a critical task. Synthetic aperture radar (SAR) image has been adopted for operational oil spill observation. However, the existing oil spill detection methods are mostly limited by the inaccurate edge recognition in the process of oil spill diffusion and the insufficient discrimination between oil spill and look-alikes. To address these issues, this study proposes a deep network based on SAR image to enhance edge perception and false target discrimination, EN-HAN DeepNet. The model uses MobileNetV2 as the backbone network, which changes the fixed expansion rate of the traditional hollow space pyramid pooling module, and effectively enhances the accurate identification of the oil spill edge through the dynamic expansion rate. At the same time, the model introduces a dual-channel attention mechanism, which can enhance the response of oil spill features while suppressing the interference of look-alikes. The effectiveness and generalization ability of the model are verified on the M4D and SOS datasets. On the M4D dataset, EN-HAN DeepNet achieves 69.26 % of the mean intersection over union (mIoU) value of oil spills and 62.80 % of the mIoU value of similar appearance categories, which is better than baseline models such as U-Net and DeepLabv3 +. On the SOS dataset, the F1 value is 94.37 %, which reduces the false alarm and missed detection rate. These results confirm that EN-HAN DeepNet provides a reliable and efficient solution for marine oil spill monitoring in complex marine environments.