Single-polarized SAR Image Preprocessing in Scope of Transfer Learning for Oil Spill Detection

This study proposes a novel preprocessing approach for improving oil spill detection from Synthetic Aperture Radar (SAR) satellite imagery using deep learning models. A transfer learning approach with the LinkNet segmentation architecture pre-trained on ImageNet is employed. The model is trained on Sentinel-1 SAR data from 2018–2023 using a designed preprocessing pipeline that converts the single-channel SAR input to a 3-channel RGB image. The proposed preprocessing involves transforming the original SAR intensity values to a normal distribution, extracting nonlinear features, and encoding them into the RGB channels. Quantitative results on a test set show the preprocessed model achieves an improvement of 0.038 in F1-score and 0.054 in Intersection over Union compared to the original dB-scale preprocessing approach. Qualitative evaluation on independent SAR scenes from the Mediterranean Sea also demonstrates the model's ability to generalize to new geographic areas after training on data from other regions. The proposed preprocessing technique shows promising performance gains for automatic oil spill segmentation from SAR imagery and potential for integration with other preprocessing methods and task-specific neural network architectures.