Reliable reduction of manual workload for oil spill detection in SAR images using uncertainty estimation and deep learning

Abstract

Marine oil spills require constant monitoring as they can cause severe environmental damage. Synthetic Aperture Radar (SAR) images are often used for oil spill detection, but they are complex and the analysis is a time-consuming process as there are a lot of areas to monitor. Therefore, the manual analysis of the images will inevitably lead to some errors. Deep learning models can be deployed for automatic classification of these images, but they fail to provide reliable confidence estimation, which can have major consequences in the case of misclassification. This thesis explores how uncertainty estimation can be used for filtering of uncertain images for manual review by an operator, while high-confidence images are automatically classified by a ResNet-50 model. For uncertainty estimation, Test-Time Augmentation (TTA) is used with dropout on the images, Pixel-Value Shift (PVS) and elastic transformation as data augmentations. The elastic transformation has the best performance, with PVS performing nearly as well. The results show that to achieve a total error of 5%, operators only need to manually analyze 41% of the dataset. Since operators regularly process numerous images, the proposed uncertainty-filtering could provide a significant reduction in manual workload. The findings in this thesis are believed to pave the way for a new and more efficient way to process SAR images of marine oil spills.