Optical Remote Sensing of Oil Spills by using Machine Learning Methods in the Persian Gulf: A Multi-Class Approach

Abstract

Marine oil spills are harmful for the environment and costly for society. Coastal areas are particularly vulnerable since they provide habitats for organisms, animals and marine ecosystems. This thesis studied machine learning methods to classify thick oil in a multi-class case, using remotely sensed multi-spectral data in the Persian Gulf. The study area covers a large area between United Arab Emirates (UAE) and Iran. The dataset is extracted from 10 Sentinel-2 tiles on six spectral bands between 492 nm to 2202 nm. These images were annotated for four classes, namely thick oil, thin oil, ocean water and turbid water by using the Bonn Agreement to analyse true color composite images. A variety of machine learning methods were trained and evaluated using this dataset. Then a robustness evaluation was done by using selected machine learning methods on an independent dataset. Initially multiple machine learning methods were included; three decision trees, six K-Nearest Neighbor (KNN) models, two Artificial Neural Network (ANN) models, two Naive bayes models, and two discriminant models. Two KNN models and two ANN models were then picked for further evaluation. The results show that the fine KNN approach with two nearest neighbors had the best performance based on the computed statistical measures. However, the robustness evaluation showed that the tri-layered NN performed better. This thesis has shown that supervised machine learning with a multi-class approach can be used for oil spill monitoring using multi-spectral remote sensing data in the Persian Gulf.