From visible imagery to synthetic short-wave infrared for maritime ship detection

Deep learning-based detection systems in maritime surveillance must generalize beyond the training distribution to unseen vessels, weather conditions, and less frequent incidents. However, capturing sufficient real visible and short-wave infrared (SWIR) imagery of such rare events is impractical in operational settings. We address this by generating synthetic SWIR data to augment limited field observations and extend model performance across operational conditions. Our pipeline first renders synthetic visible imagery and then translates it to synthetic SWIR using a generative adversarial network translator (CycleGAN-Turbo) augmented with an SSIM (structural similarity index) term. We assess the synthetic SWIR using established image-quality measures, as well as by evaluating its impact on the performance of a state-of-the-art ship detector. We find that the GAN-based translator can retain visible-like detail, whereas adding SSIM to the loss improves structural and spatial-frequency alignment with real SWIR. This leads to reconstructions that better preserve discriminative features for detection. Overall, synthetic SWIR improves detector robustness in low-data regimes and enhances generalization under domain shift.