Automated Deep-Learning Based Oil Spill Detection in the North Sea from Spaceborne SAR

Kurzfassung

Oil pollution in the North Sea is mainly caused by shipping activities and offshore oil production. Oil leaks and discharges from offshore platforms, pipelines and ships result in surface films that threaten sensitive marine and coastal ecosystems such as the Wadden Sea. Spaceborne Synthetic Aperture Radar (SAR) sensors are widely used to continuously monitor accidental and deliberate oil spills in this large area. However, the reliable discrimination of oil spills and other oceanic and atmospheric phenomena remains challenging. Here, we present an automated system for oil spill detection from Sentinel-1 imagery based on deep learning, which was trained and validated using manually labelled image patches from the North Sea covering all seasons. Detections from the YOLOv4 object detector are post-processed using machine learning to reduce false detections while maintaining sensitivity. Each detected oil spill is segmented to determine the actual extent of the polluted area. With its potential for operational use, our oil spill monitoring could inform decision-making and support an early response to mitigate pollution.