A transfer learning-DCNN based oil spill detection using compact polarimetric SAR data

Abstract

Oil spills in the marine environment can cause both economic and environmental crises, which underlines the urgent need for effective detection and prevention strategies to mitigate the consequences. Synthetic Aperture Radar (SAR) is one of the commonly used sensors for oil spill detection, and the compact polarimetric (CP) SAR system, with its wide swath width and sufficient polarimetric information, is well suited for this task. This study aims to employ four deep convolutional neural network (DCNN)-based semantic binary segmentation models (i.e., U-Net, LinkNet, FPN, and PSPNet) to detect oil spills using simulated compact polarimetry in hybrid-pol mode. In the deployed methodology, we used transfer learning to improve the adaptability of the models to different sensors. We efficiently adapted the built oil spill detection models on UAVSAR to data of RADARSAT-2 while retaining the essential features and knowledge from pre-trained models by fine-tuning them. The results showed the potential of these models in oil spill detection. The PSPNet model, as the most accurate, achieved an overall accuracy (OA) of 96.00% and a kappa coefficient of 91.30% on the UAVSAR image. After fine-tuning, it yielded an OA of 98.68% and a kappa coefficient of 92.71% on the RADARSAT-2 image.