Automated Oil Spill Detection System Using Spaceborne Synthetic Aperture Radar with Deep Learning-Based Algorithm

Oil spills are one of the major sources of marine pollution. They cause environmental damage and affect marine wildlife. In order to take early action in oil spill incidents, an automated system for the regular monitoring and detection of oil spills can be beneficial. In the frame of this thesis, such an automated oil spill detection system has been developed and its potential for integration into an early warning system has been demonstrated. The system identifies oil slicks in synthetic aperture radar (SAR) images with a custom trained deep learning-based object detector and provides binary masks of oil as output using a segmentation algorithm. The system has been tested in the Southeastern Mediterranean Sea for a year of acquisitions in 2019, demonstrating the ability of using such an automated system in operation to reduce the effort of manual interpretation.

Complete records of oil spills, especially those from illegal operations, are often missing or unavailable to the public, so collecting and annotating oil slicks is usually the first task when developing a deep learning-based method for oil slick detection. As a result, different published methods use different datasets, making it difficult to compare the diverse approaches. Therefore, a ready-to-use dataset has been published as part of this thesis, which contains oil slicks and look-alikes that originate from other phenomena but may be misinterpreted as oil slicks in SAR data. The performance of the object detector used in the developed system was evaluated on image patches from different sources, providing a baseline for other studies to use for comparison. In addition, this thesis provides a comprehensive explanation of how various oceanic and atmospheric phenomena manifest in SAR scenes, thus bridging the knowledge gap between readers with different backgrounds, such as remote sensing, oceanography, and machine learning.

Rights

Use and reproduction:


CC BY-NC-ND 4.0

Please note that individual components of the publication may be subject to other licensing or copyright conditions.

Cite

Citation style:
Could not load citation form.