Near real-time oil pollution monitoring using SAR for the Southeastern Mediterranean Sea

Abstract

This study proposes a near real-time oil spill surveillance system using Sentinel-1 synthetic aperture radar (SAR) imagery. The users only need to provide the coordinates of their areas of interest. With such information, the system automatically downloads the SAR scenes located within the study area and preprocesses those scenes with several corrections, such as noise removal and calibration. Afterwards, a custom-trained You Only Look Once version 4 (YOLOv4) object detector is applied to a YOLO-based Oil Detection Algorithm (YODA) for targeting oil slicks inside the SAR scenes. These oil slicks are defined by bounding boxes and fed to a segmentation algorithm for obtaining the exact locations covered by oil. This study selected the Southeastern Mediterranean Sea as a study area, where each Sentinel-1 track includes four continuous SAR scenes. From obtaining these four scenes to delivering oil slick binary masks to the users, it takes around 17 minutes for a computer with a GPU. The object detector was trained with a collection of oil slicks in the study area; however, it is feasible to extend the usage of the current system to other regions with further training on specific types and sources of oil slicks from the regions.