Pixel-level detection and classification of marine oil spills in aerial imagery with annotation uncertainty handling

Oil spills pose significant environmental threats, particularly in coastal and marine ecosystems, requiring fast and accurate detection methods to support mitigation efforts. In this study, we propose a deep learning-based framework for automatic oil spill detection and classification using convolutional neural networks. We developed two complementary RGB image datasets annotated at the pixel level: one captured from real maritime monitoring operations and another curated from online sources. These datasets are employed to train and evaluate semantic segmentation models based on the U-Net architecture. The models are capable of detecting oil presence and classifying different oil types based on appearance characteristics (e.g., silvery, rainbow, emulsified). To improve performance and generalization, a series of ablation studies have been conducted, examining the effects of network backbone and pre-training, data augmentation strategies, and the handling of visually ambiguous regions. Our best-performing model achieved a mean Intersection over Union (mIoU) of 0.738 and an accuracy of 86.9% for binary oil detection, while achieving an mIoU of 0.351 and an accuracy of 81.6% for multiclass oil type classification. These results demonstrate the feasibility of using pixel-level segmentation for oil spill monitoring in diverse marine conditions, highlighting the importance of dataset quality and model configuration in real-world deployment.