PetroAnalyticNet: Applying separable dilated convolutions for drone-captured oil spill segmentation in ports

Oil spill monitoring in port and maritime environments is essential for environmental protection and emergency response. However, precise segmentation from UAV-captured optical imagery is still difficult due to specular reflections on water surfaces, complex port infrastructures, irregular spill shapes, and significant visual ambiguity. To overcome these limitations, we develop PetroAnalyticNet, a lightweight semantic segmentation framework tailored for UAV-based oil spill detection in challenging aerial scenes. The model combines depthwise separable convolution for efficient local representation learning, dilated convolution to expand contextual perception, residual connections to enhance feature stability, and an attention-based refinement mechanism to mitigate interference from reflective backgrounds. A standardized experimental protocol is constructed using two publicly available datasets, Oil Spill Drone and LADOS, ensuring consistent and reproducible evaluation. Experimental evaluations demonstrate that PetroAnalyticNet delivers superior segmentation accuracy on both benchmarks. On the Oil Spill Drone dataset, it attains an F1 score of 0.91 and an IoU of 0.84, while on the more challenging LADOS dataset, it achieves an F1 score of 0.83 and an IoU of 0.73. These results highlight the effectiveness of the proposed approach for fine-grained UAV-based oil spill segmentation and suggest its potential for deployment in intelligent maritime monitoring systems under optical aerial imaging conditions.