Oil Spill Discrimination of SAR Satellite Images Using Deep Learning Based Semantic Segmentation

Abstract

This study examined the effective network architecture to discriminate oil spills from look-alikes using deep learning-based semantic segmentation. Data were collected from three different SAR satellites, ENVISAT, ALOSPOLSAR, and TERRASAR. To ensure better accuracy, different characteristics of SAR satellites were analyzed based on spatial and temporal resolution. Semantic segmentation helps in finer inference by making dense predictions from labels of each class. Before entering into the segmentation phase, image preprocessing is done for effective segmentation results. The image Preprocessing phase includes log normalization, hybrid median filter, and scaling. Oil spill discrimination is done by Semantic segmentation using CNN and deeplabv3plus with pretrained RESNET18 on three different satellite images with the high-level and low-level dataset. It is observed that a high-level dataset of TERRASAR images outperforms rather than another SAR dataset. The comparison technique reveals that the proposed DeeplabV3plus segmentation has resulted in a great accuracy of 99% and is considered as the effective segmentation technique for oil spill discrimination.