Automatic Detection of Oil Spills from SAR Images Using Deep Learning

Abstract

AbstractLarge tankers, ships, and pipeline cracks pour oil on sea surfaces, causing significant damage and devastation to the maritime ecosystem. From the synthetic aperture radar (SAR) images, the target scenarios are ships, sea, land surfaces, look-alikes, and oil spills. Oil spill detection and segmentation using SAR pictures is critical for leak cleaning and environmental protection. The paper presents a deep learning-based framework for the oil spill identification on the dataset generated by the European Maritime Safety Agency (EMSA). In the first step, a 23-layer Convolutional Neural Network detects oil spills by classifying the patches into less than 0.5% and more than 0.5% oil spill pixels. The second step uses a U-Net architecture to apply the semantic segmentation on the generated patches. The results evidenced that the deep learning framework and segmentation models outperform the identification of an oil spill from SAR images. The present research makes a substantial contribution to future research on oil spill detection and the processing of SAR images.KeywordsDeep learning frameworkMarine surveillanceOil spill detectionConvolutional neural networkSAR images