A deep learning-based method for marine oil spill detection and its application in UAV imagery

With the increasing intensity of marine resource exploitation, oil spill incidents have become more frequent and now represent a significant source of pollution threatening the stability of marine ecosystems and the security of coastal economies. Oil spills not only cause long-term deterioration in seawater quality but also exert direct toxic effects on marine animals including birds, fish, and plankton, resulting in habitat degradation, mass mortality, and trophic cascade disruptions. These ecological impacts are extensive and persistent, highlighting the urgent need for an efficient and intelligent detection mechanism to enable early warning and rapid response. To address the limitations of traditional manual inspection and radar-based identification methods-such as low temporal efficiency and limited detection accuracy-this study develops a YOLOv12-based oil spill detection framework for Unmanned Aerial Vehicle (UAV) imagery. The proposed method aims to enhance the accuracy of oil spill detection and the precision of boundary delineation, thereby offering technical support for marine ecological protection. In the study, a complete detection pipeline is designed, accompanied by the construction of a remote sensing image dataset encompassing diverse oil spill morphologies, varying marine interference conditions, and multi-scale objects. To further improve the model's capacity for representing pollution-related features, the training process integrates high-resolution input, pretrained weight initialization, and a cosine annealing learning rate scheduling strategy. During the model training, loss functions including Box_Loss, Cls_Loss, and Dfl_Loss exhibit rapid convergence, ultimately achieving an F1 Score of 0.7929, mAP@0.5 of 0.8334, and mAP@0.5-0.95 of 0.7503 on the test set. Accordingly, the Precision-Recall curves exhibit a smooth distribution, indicating favorable performance in both the accuracy of oil spill detection and the precision of boundary regression. The model also achieved stable detection of oil spill regions against complex backgrounds in static image tests and maintained consistent inter-frame detection and sustained target tracking capability in typical dynamic video scenarios. The study concludes that YOLOv12 provides foundational support for ecological risk assessment, pollution distribution monitoring, and wildlife protection responses, demonstrating promising potential for real-world applications.