Fusion of Oil Spill Detection Results from Thresholding and Deep Learning Using Landsat Data over the North Sea

Abstract

The contamination of marine and coastal environment by oil pollution has a considerable impact on the surrounding ecosystems. It is therefore imperative that oil spills are identified at the earliest possible stage in order that the relevant monitoring frameworks can be put in place and appropriate response measures initiated. The timely and accurate detection of oil is of great benefit in the prevention of pollution and the facilitation of clean-up operations, which serve to minimise the negative impact on the environment and identify the source of the pollution. In particular, the use of Synthetic Aperture Radar (SAR) has been established as an effective method for monitoring large marine areas for many years.

This study presents the fusion results of two complementary approaches for automatic oil spill detection in optical satellite imagery using Landsat data. The first approach is a traditional thresholding analysis, while the second employs a convolutional neural network (CNN) in the type of a U-Net architecture. The proposed fusion approach is evaluated on two datasets: a larger set of 48 Landsat-8 images to analyse general detection performance, and a subset of 15 images for which manually labelled binary oil masks are available enabling quantitative validation. The aim of this study is to improve detection accuracy and reduce false positive detections under the assumption that the two methods produce different types of errors.

The results demonstrate that the combination of both methods thresholding and deep learning partially optimizes detection accuracy by reducing false positive detections, although some false positives remains and certain oil spills are reduced in size or they are lost.