A deep learning based oil spill detector using Sentinel-1 SAR imagery
Yi-Jie Yang, Suman Singha, Roberto Mayerle
Abstract
The Eastern Mediterranean Sea has been known as an oil pollution hotspot due to its heavy marine traffic and an increasing number of oil and gas exploration activities. To provide automatic detection of oil pollution from not only maritime accidents but also deliberate discharges in this region, a deep learning-based object detector was developed utilizing freely available Sentinel-1 Synthetic Aperture Radar (SAR) imagery. A total of 9768 oil objects were collected from 5930 Sentinel-1 scenes from 2015 to 2018 and used for training and validating the object detector and evaluating its performance. The trained object detector has an average precision (AP) of 69.10% and 68.69% on the validation and test sets, respectively, and it could be applied for building an early-stage oil contamination surveillance system.