Oil spill segmentation of noisy SAR images using domain adaptation based UNet

Oil spills dampen the water bodies, causing serious threats to the marine environment. In this paper, oil spill detection and segmentation in satellite imagery in the presence of speckle noise is proposed. Most of the algorithms have not given concern to noise in the images due to which segmentation results may not generate perfect results. One of the best deep learning networks, UNet is notable for its application in segmentation tasks, but when applied for SAR images corrupted with speckle noise, it gives improper results with low values of accuracy and metric scores. Taking into consideration of this we have modified the established UNet architecture with the domain adaptation block of the gradient reversal layer. This new UNet model, applied Sentinel-1 SAR images raw dataset with speckle noise and had produced given promising results with a validation accuracy approach of 95% and F1 score is 95.86% which indicates that it has the ability to differentiate and partition oil spills locations within SAR images. Proposed model robustness is observed by the values of dice similarity score and IoU obtained for all the 5 different classes. Classification of the classes in the the scenarios like disaster management, marine pollution control, and various applications, this method provides a good segmentation results.