Transfer learning and single-polarized SAR imagepreprocessingfor oilspill detection

This study addresses the challenge of oil spill detection using Synthetic Aperture Radar (SAR) satellite imagery,
 employing deep learning techniques to improve accuracy and efficiency. We investigated the effectiveness of
 various neural network architectures and encoders for this task, focusing on scenarios with limited training
 data. The research problem centered on enhancing feature extraction from single-channel SAR data to improve
 oil spill detection performance.
 Our methodology involved developing a novel preprocessing pipeline that converts single-channel SAR
 data into a three-channel RGB representation. The preprocessing technique normalizes SAR intensity values
 and encodes extracted features into RGB channels.
 Through an experiment, we have shown that a combination of the LinkNet with an EfficientNet-B4 is
 superior to pairs of other well-known architectures and encoders.
 Quantitative evaluation revealed a significant improvement in F1-score of 0.064 compared to traditional
 dB-scale preprocessing methods. Qualitative assessment on independent SAR scenes from the Mediterranean
 Sea demonstrated better detection capabilities, albeit with increased sensitivity to look-alike.
 We conclude that our proposed preprocessing technique shows promise for enhancing automatic oil spill
 segmentation from SAR imagery. The study contributes to advancing oil spill detection methods, with potential
 implications for environmental monitoring and marine ecosystem protection.