Oil Spill Classification from Multi-Spectral Satellite Images: Exploring Different Machine Learning Techniques

This work describes the potential of oil spill classification from optical satellite images, as investigated by applying different machine learning techniques to a dataset of more than 300 oil spill candidates, which have been detected from multi-spectral satellite sensors during the years 2008 and 2009, over the entire area of the Mediterranean Sea. A set of geometrical and grey level features from Synthetic Aperture Radar (SAR) literature has been extracted from the regions of interest in order to characterize possible oil spills and feed the classification system. Results obtained by applying different machine learning classifiers to the dataset, and the achieved performance are discussed. In particular, as a first approach to oil spill classification, simple statistical classifiers and neural networks were used. Then, a more interpretable fuzzy rule-based classifier was employed, and performance evaluation was refined by exploiting Receiver Operating Characteristic (ROC) analysis. Finally, since oil spill dataset collection happens incrementally, a suitable technique for online classification was proposed, encompassing at the same time cost-oriented classification, in order to allow for a dynamic change of the misclassification costs. This latter goal has been achieved by building an ensemble of cost-oriented, incremental and decremental support vector machines, exploiting the concept of the ROC convex hull.