An improved Pix2PixGAN sample generation method for SAR oil spill detection

Marine oil spills pose serious threats to marine ecosystems and coastal socio-economic systems, while the scarcity of annotated SAR oil-spill samples limits the training and generalization of deep learning-based detection models. To address this problem, this study proposes an improved Pix2PixGAN method for SAR oil-spill sample generation. The proposed method uses real SAR images as background priors and combines oil-spill masks with multi-channel random noise to jointly constrain background realism, target morphology, and texture diversity. A spatially adaptive normalization module, PatchGAN discriminator, least-squares adversarial loss, Dropout regularization, and segmentation-consistency constraint are integrated to enhance local texture realism and structural consistency. Experiments on the SOS dataset show that the proposed method outperforms Pix2Pix, PGGAN, and BEGAN in FID, KID, Global BIHD, ENL difference, and GMHD, indicating better consistency with real SAR samples in both feature distribution and SAR-related image-level statistical characteristics. Ablation experiments further confirm the complementary effects of spatial semantic modulation and multi-channel noise injection. In downstream U-Net segmentation validation, the model trained only with generated samples achieved performance comparable to that trained with real samples only, while combining real and generated samples produced slight numerical improvements. These results indicate that the generated samples contain learnable oil-spill structural and textural information, but the improvement should be interpreted as evidence of sample usability rather than statistically significant superiority. Cross-dataset validation on the independently constructed A Symphony-Yellow Sea dataset and additional experiments using DeepLabV3Plus, PSPNet, and SegNet further suggest that the generated samples can provide useful supplementary training cues for independent SAR oil-spill scenes when real annotated samples are limited.