Abstract
Oil spill accidents are one of the major problems causing marine pollution, and thus such accidents require rapid detection for early response. In recent years, deep learning algorithms for oil spill detection have been developed for analyzing SAR images. Nevertheless, to generation of deep learning training data using visual inspection is not only a time-consuming and labor-intensive, but also cause bias and lack of diversity in oil slick training data. This has easily led to false positives in challenging SAR images, such as large scale look-alikes, biologic surface films or oil slicks surrounded by look-alikes. In order to accommodate a broader range of SAR oil spill scenes, including those under challenging conditions, it is essential to continuously enhance the performance of oil spill detection algorithms. In this study, a novel self-evolving algorithm for automatic oil spill detection is proposed, which consists of three inter-connected modules: 1) oil spill detection, 2) generation of new training data, and 3) enhancement of deep learning models. The algorithm detected oil slicks automatically while the new SAR image was added, and then the additional high-quality training data was generated by an adaptive thresholding method to increase the performance of deep learning models. In order to detect oil slicks from whole SAR image with oil look-alikes, the variation of the backscattering coefficients near the oil boundary was considered as a key parameter to distinguish oil spills from look-alikes. As a result, after 21 self-evolving training cycles, 27 new training data were generated with an improved F1-score rising from 0.8423 to 0.8896. Despite the existence of look-alikes in various magnitudes, the algorithm successfully identified oil slicks, and the parameters of deep learning models were automatically updated. It was demonstrated that the performance of the algorithm was gradually improved by self-evolving without any human intervention. Additionally, the fundamental limitations of oil spill detection algorithms were mitigated by the proposed approaches, providing more possibilities for subsequent studies.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.