Improving the accuracy of SAR-based oil slick detection

SkyTruth has created a fully automated system to detect anthropogenic oil pollution across the world’s oceans in near-real-time using Sentinel-1 data. This system is called Cerulean and is designed to provide environmental organizations, researchers, governments, journalists, and other users with a global monitoring and reporting system for oil pollution.

Synthetic Aperture Radar (SAR) data is commonly used to remotely sense the presence of oil on the surface of the ocean. However, SAR based oil slick detection is prone to false positives caused by wind shadows, sea ice, organic surfactants on the water surface, and other phenomena that cause look-alike dark patches in SAR data. We have taken a three tiered approach to reducing false positive detections from our oil slick detection model: (1) Providing a significant number of false-positive look-alikes in our training dataset, (2) Experimenting with different classes of deep learning model architectures, including a U-Net semantic segmentation model and a Mask R-CNN object segmentation model, (3) Leveraging various post-processing techniques to help distinguish true positives from false positives, including morphological characteristics of slicks, proximity to shipping lanes and offshore infrastructure, and weather data. Results of these analyses and promising avenues for improving oil slick discrimination with SAR data will be described in detail.