Automated Classification of Oil Spill Events in Satellite Imagery Using Deep Learning and Spectral Decomposition

Abstract

Oil spills represent a major environmental challenge, the devastating effects of which on marine ecosystems require early detection and classification to mitigate their consequences. In this study, an innovative methodology based on deep learning and spectral decomposition techniques is proposed for the automatic classification of spill events in multispectral satellite images. A convolutional neural network architecture (CNN) is implemented combined with dimensional reduction techniques by spectral decomposition (PCA and SVD), optimizing the recognition of spectral patterns characteristic of hydrocarbons. The results show a high accuracy (>95%) in the classification of spills against other surface anomalies such as algae or solar reflections, validating the usefulness of this approach for automated environmental monitoring.