SinGAN-Labeler: An Enhanced SinGAN for Generating Marine Oil Spill SAR Images with Labels

Deep learning-based SAR oil spill detection faces significant challenges due to limited labeled training data. To address this, we propose SinGAN-Labeler, an enhanced framework that generates high-quality synthetic SAR oil spill images and their labels from minimal input. The model integrates an adaptive module to automate scale parameter optimization, accelerating training, and a hybrid attention module combining spatial, channel, and global contextual mechanisms to enhance feature extraction. By leveraging multi-scale training with diverse receptive fields, the generated images retain critical structural details while ensuring diversity. Experiments demonstrate that detection models trained on synthetic data achieve performance comparable to those using real images. Notably, expanding data sets by fivefold (from 5, 10, and 15 baseline images) improves the UNet++ model’s IoU by 78.2%, 58.5%, and 22.5%, respectively. These results validate SinGAN-Labeler’s capability to mitigate data scarcity and enhance oil spill detection accuracy, particularly under extreme sample limitations.