CFM-SAR: A Class-Conditioned Flow Matching with SCM-UNet for SAR Image Generation

Monitoring and analyzing sea-ice changes is critically important for navigation through polar regions, ecosystem monitoring, and environmental risk assessment, and Synthetic Aperture Radar(SAR) imagery is one of the most reliable and useful data sources for this goal. SAR images exhibit complex backscatter patterns that are generally not interpretable without technical expertise. Consequently, accurate annotation of SAR images requires domain experts to identify the corresponding sea ice classes, resulting in labelled datasets that are both limited and often imbalanced across classes.A further challenge is that SAR image channels differ from those of conventional images, with HH, HV, and incidence angle each containing important physical information. Although current image generation methodologies have strong capabilities in RGB images and have solved several limitations, including training instability and high computational cost, some challenges in SAR image generation remain. These issues, such as limited preservation of channel-specific physical information and mode collapse, motivate us to work in generative modelling and seek to overcome them. In this study, we propose a Flow Matching framework with a class-informed prior. Instead of using a standard Gaussian distribution as the initial point for learning, the Class-Conditioned Flow Matching (CFM) model estimates the mean and variance of each image class and uses them to better approximate the data distribution from a more informative starting point. Another contribution of this paper is the proposed Swin Transformer Convolutional Hybrid (SCM-UNet) model, which generates images using a modified U-Net architecture. Experiments on the chosen dataset demonstrate that the proposed model improves visual metrics in generated images (FSD=19.23, PSNR=28.0, and SSIM=0.83) and effectively enhances the downstream task. Results show that, when trained exclusively on generated data and evaluated on real SAR images, the classification model achieved an average accuracy of 91.61% and an F1-score exceeding 91.5%.