Oil spills identification on satellite radar data using deep learning

Marcelo Guarido, David J. Emery, Daniel O. Trad, Kristopher A. Innanen

Oil spills in oceans are a major pollutant endangering oceanic and coastal marine life, and their detection is of high environmental importance. Manually detection presents a challenging and lengthy task. We presented a deep learning model based on the U-Net structure to identify oil bodies in satellite radar images with promising results. Our model successfully classified larger oil bodies with moderate success on smaller ones. Image feature engineering, such as a four-directional cumulative sum, brought important information to the model and performed more accurate predictions. Limited by computer resources, our model was relatively simple. We used pre-trained weights from the MobileNetV2 model. Although initial results are unsatisfactory, we will continue to explore the transfer learning methodology to generate more accurate oil detection algorithms.