DeepLab-AFOS: A marine oil spill detection method based on advantageous features from polarimetric SAR

Marine oil spills can severely damage marine ecosystems and cause significant economic losses. Accurate detection and extraction represent the most critical and fundamental requirements for marine oil spill monitoring and emergency response. While the application of SAR in marine oil spill detection has been extensively studied, most existing works rely solely on SAR backscattering information, with limited integration of polarimetric and textural features. This insufficiency often leads to difficulties in distinguishing oil spills from look-alikes. To address the challenge of differentiating oil spills from look-alikes, we constructed an advantageous features dual-polarimetric SAR dataset (AF-OS), incorporating multidimensional features such as covariance matrices, polarimetric parameters, and textural descriptors. Based on this dataset, we proposed a polarimetric SAR oil spill detection method named DeepLab-AFOS. Comparative experiments using datasets with different feature combinations demonstrate that the features included in AF-OS significantly enhance the accuracy of oil spill extraction. Furthermore, model comparison results show that DeepLab-AFOS outperforms models such as U-Net, PSP-Net, ResNet-UNet, Segmenter-B, OSDMamba and DeepLabV3+, achieving an accuracy of 97.2%, F1 score of 95.8%, and MIoU of 92.06%. Additionally, we applied DeepLab-AFOS to a real-world oil spill event in the Gulf of Mexico. The extracted results were consistent with ground-truth validation, accurately identifying the oil spill and effectively distinguishing it from look-alikes. Both AF-OS and DeepLab-AFOS provide effective solutions for the accurate extraction of marine oil spills.