Convolutional neural networks for oil spill detection in SAR images

Oil pollution in the seas and oceans is critical to the biological ecosystems of marine life and, in general, to the health of our earth. One of the sources of this pollution is the bilge dumping that generates long oil spills. Satellite imagery and, in particular, SAR (Synthetic Aperture Radar) provide wide coverage of oceans and seas. The analysis of this large number of images can identify and locate these long oil spills using current machine learning methods. This information could be linked with the movements of the vessels, which makes it possible to identify and punish fraudsters. However, many oceanic and atmospheric phenomena other than oil spills may modify the sea surface and create dark zones in SAR images. These dark zones are called look-alikes and represent a real difficulty in distinguishing from oil spills, even for a human operator. Unlike many other applications, there is no dataset available to the scientific community for this type of situation. This results in limited research in this area and the inability to compare performances. In this thesis, we develop a new spill oil dataset, including ground truth masks with training and test parts. We propose a CNN-based classification system that automatically detects oil spills from SAR images, including many other look-alikes. The architecture of the model and its hyperparameters are optimized and discussed. Mean accuracy of 91.4% and a mean Intersection over Union (mIU) of 83.5% have been obtained on the test dataset.