Reducing Manual Workload in SAR-Based Oil Spill Detection Through Uncertainty-Aware Deep Learning

Dina Svendsen Solskinnsbakk, Sigurd Almli Hanssen, Harald Lykke Joakimsen, Vilde B. Gjærum, Elisabeth Wetzer, Kristoffer Knutsen Wickstrøm

Keywords: remote sensing, oil spill detection, uncertainty, SAR

TL;DR: Uncertainty-aware deep learning models can reduce manual workload in SAR-based detection of oil spills

Abstract: Constant monitoring of the oceans is required to detect oil spills and reduce environmental damage associated with spills. Synthetic Aperture Radar (SAR) imaging is a critical tool for oil spill detection, but is complex and requires time-consuming manual labor for analysis. Deep learning has shown encouraging performance in automatic classification of oil spills on these images, but the performance is still not sufficient for a deep learning classifier to act autonomously, making manual assessment essential. However, if only a reduced subset of uncertain samples had to be analyzed by human experts while the remaining samples could be automatically classified, it could greatly reduce the manual workload. In this study, we investigate if uncertainty estimates can identify which samples should be prioritized for manual inspection. Specifically, we propose a pipeline of defining a user-specified error tolerance and identifying an uncertainty threshold that filters out samples for automatic/manual thresholding. We evaluate the proposed pipeline on challenging real-world data. The results show that our proposed uncertainty-based ranking technique can reduce the manual workload by 41%, paving the way for new and more efficient ways to detect marine oil spills.

Serve As Reviewer: ~Kristoffer_Knutsen_Wickstrøm1, ~Elisabeth_Wetzer1, ~Harald_Lykke_Joakimsen1

Submission Number: 26