Operational Oil-Spill Detection in the Framework of Digital Twin Using SAR Imagery and Deep Learning

This work presents the development and implementation of an automated early-warning system for oil-spill detection, utilizing Synthetic Aperture Radar (SAR) imagery analysis and deep learning techniques. The system is designed to enhance the timely detection and management of oil spills in the Thracian and wider North Aegean Sea, a region particularly vulnerable due to high marine traffic and proximity to oil platforms. The approach combines traditional image-processing methods, including Oil Slicks and Red Tide Monitoring index (ORM) and Gray Level Co-occurrence Matrix (GLCM), with a U-Net semantic segmentation model, extensively trained on Sentinel-1 data for oil spill detection. The model employs its segmentation capabilities to accurately identify and localize oil spills while overcoming issues related to class imbalance through custom masking and data augmentation strategies. Metrics evaluation reveals robust performance in distinguishing oil spills from similar features (look-alikes), highlighting the model's capability in operational scenarios. This integrated approach not only refines detection accuracy but also enhances the reliability of early-warning systems crucial for environmental monitoring. The system operates daily by downloading new SAR imagery, executing parallel processing approaches, and issuing warnings whenever oil spills are detected. It incorporates a numerical model for oil-spill dispersion, utilizing meteorological and sea-state forecasts to predict spill trajectories. The system is calibrated using data from the 2021 oil-spill event in the Levantine Sea, ensuring its reliability in real-world scenarios.