Detection and Delineation of Produced Water Slicks in Sentinel-1 Synthetic Aperture Radar Images

Abstract

Near oil and gas platforms oil detection services regularly detect oil slicks that are a result of legal releases of produced water. These slicks are usually observed using SAR imagery and the important task of observing and monitoring these slicks is as of now carried out manually by human operators aggregated with reported release information. In this thesis we propose three separate approaches to simplify and improve this work through the use of image segmentation and deep learning methods. The approaches are trained and tested on a set of Sentinel-1 scenes over the Brage and Norne platforms off the coast of Norway. The best performing approach was shown to be the direct use of the deep learning algorithm Mask R-CNN on the Sentinel-1 scenes. This approach was able to detect 81\% of all slicks in the scenes and had an average user's accuracy of 78\% and an average producer's accuracy of 73\%. The approaches were also shown to have a significantly reduced ability to detect slicks when the local wind speeds were below 2 m/s or above 11.5 m/s and when the daily volume of oil released from the platforms was below around 150 kg.