Oil Spill Monitoring in the Gulf of Mexico based on Deep Learning using Spaceborne SAR Imagery from the European Copernicus Mission

Kurzfassung

Due to its large oil reserves and heavy shipping traffic, the Gulf of Mexico is subject to oil pollution from both anthropogenic spills and natural seeps. In this contribution, we present an oil spill monitoring system for the Gulf of Mexico based on synthetic aperture radar (SAR) imagery from the European Copernicus satellite Sentinel-1. Oil slicks are identified using a state-of-the-art deep learning-based object detector, which was custom-trained with manually labeled SAR image patches from the northern Gulf of Mexico, including oil slicks and typical other atmospheric and oceanic phenomena. Labeling was supported by contextual information such as wind speed, chlorophyll-a concentration and sea surface temperature. An additional dataset of labeled SAR images from a full year, covering all seasons, was then used to validate the detector. Finally, each detected oil spill is segmented to provide the actual extent of the polluted area. We present an evaluation of the detector performance and a method to reduce false alarms using local image statistics. Starting from individual SAR scenes, our monitoring system automatically generates geocoded binary masks of the detected oil spills. The method is optimized for speed and accuracy and can be adapted to upcoming SAR missions. With its potential for operational use, our oil spill monitoring system can serve as a basis for decision makers and early response.