Self-Supervised Learning for Oil Spill Detection in Synthetic Aperture Radar Imagery

Labelled Synthetic Aperture Radar (SAR) imagery for oil spill detection is exorbitant and sluggish. Every annotation requires a domain expert who can distinguish between real oil spills from look-alikes, and the result is that labelled SAR datasets are small, while large volumes of unlabeled satellite imagery go unused. This study has developed a self-supervised learning pipeline using Bootstrap Your Own Latent (BYOL) to address this directly. The idea is to pre-train a ResNet18 encoder on the unlabeled portion of the SOS dataset, and fine-tune it with only a small fraction of labels. Pre-training ran for 300 epochs using BYOL without any labels. Fine-tuning used the LP-FT strategy which froze the encoder and trained only the classification head for 20 epochs, then unfreezes everything and trained end-to-end at a much smaller learning rate. Before any fine-tuning, the frozen encoder achieved 96.4% KNN accuracy on validation embeddings. That number matters because it shows BYOL actually learned something useful about SAR imagery, not just noise. With 10% of labels, the model achieves F1(oil) of 0.928. The supervised baseline from scratch gets 0.973 on F1(oil), which looks better until you account for the test set being 95.3% oil patches. On macro F1, which weights both classes equally, BYOL scores 0.552 against the supervised baseline's 0.486. On a strictly disproportionate test set, macro F1 is the reliable metric.