SEMANTIC SEGMENTATION OF OIL SPILLS IN SAR IMAGES

The large increase in the number of satellite images available in recent years allow the assessment of the most diverse issues, such as monitoring pollution, deforestation and land use. However, the interpretation of these data is a challenging problem at scale. Obtaining useful representations from such images requires a rich understanding of the nature of their information. This work explores the above problem by designing an automated framework to extract semantic maps from synthetic aperture radar images to track the presence of oil spills at sea. Conceiving it as a supervised machine learning problem, image processing techniques and a convolutional neural network are implemented and evaluated. A set of publicly available data and structures is used for this purpose and the results reveal a considerable improvement in relation to the previous techniques. In relation to the original U-Net for the class of greatest interest (OS), it was observed an improvement of 28.5% in the IoU metric and the convergence process was about 13 times faster.