Automated Oil Spill Detection and Segmentation in Marine Ecosystems Using Drone-Captured Images and Attention-Enhanced Convolutional Architectures

Marine oil spills pose significant threats to aquatic ecosystems, affecting biodiversity, water quality, and coastal resources. Accurate and timely detection of oil spills is essential for environmental monitoring and mitigation efforts. This project presents an automated framework for oil spill detection and segmentation using drone-captured images, leveraging attention-enhanced convolutional architectures. The proposed system integrates two deep learning models, HFC‑Net and SSM‑FusionNet, to address common challenges in oil spill segmentation, such as ambiguous boundaries, irregular shapes, appearance variability, domain shifts, and class imbalance. HFC‑Net employs hierarchical feature calibration with contour-based supervision and dynamic feature aggregation to precisely identify oil-water boundaries. SSM‑FusionNet combines spectral and spatial feature fusion with adaptive gating mechanisms to capture thin sheets and emulsion patterns under varying environmental conditions. Both models are trained and evaluated on datasets from Zenodo and LADOS, ensuring robustness across different marine scenarios. The system is deployed using a Flask-based web interface, allowing users to register, log in, visualize model performance, and perform oil spill predictions. Experimental results demonstrate that the proposed models achieve high segmentation accuracy and improved mean Intersection over Union (mIoU), outperforming standard convolutional architectures while maintaining efficient computational requirements. The framework offers an effective solution for environmental agencies and marine monitoring authorities to identify oil spill extents accurately and promptly.

Keywords: Oil Spill, Segmentation, Drone Imagery, Deep Learning, HFC‑Net, SSM‑FusionNet, Attention Mechanism, Convolutional Neural Networks, Environmental Monitoring, Flask.