Development of a Model to Identify Oil Spills Using Supervised Deep Learning for Semantic Segmentation Using DeepLabV3-MobileNetV3

Crude oil spills caused by maritime transport are major threats to marine ecosystems and require quick detection and classification for an effective response. The use of deep learning for oil spill detection using UAV/SAR imagery is making consistent advancements. There remains a gap in pixel-wise semantic segmentation to handle complex scenarios in port, generalization to heterogeneous backgrounds, and multi-class delineation (e.g., water, oil thickness, background). This paper proposes a lightweight DeepLabV3-MobileNetV3 framework with focal loss to solve the class imbalance problem, which is trained using high-resolution UAV images from working harbors (2021-2023) annotated for 4 classes: background (4.33%), oil (33.04%), others (38.17%), and water (24.46%). The model converges very fast (train loss: 0.071-0.020 and validation loss: 0.031-0.017 over 5 epochs) and results in 93.63% overall accuracy, macro F1-score of 0.8155, and weighted F1 of 0.9332 on test data. Per-class results are best for oil (F1: 0.9485) and others (F1: 0.9788) with close to perfect ROC/PR curves. Confusion matrices reveal very few inter-class errors, and qualitative predictions define spill boundaries accurately, in accordance with ground truth. Outperforming previous UAV/SAR approaches in mIoU and efficiency, this architecture allows for real-time edge-deployable monitoring to support targeted clean up, accountability, and compliance with marine protocols. Results highlight region-specific training and explainable metrics for operational resiliency in high-traffic ports.