Super-Resolving SAR Images with Diffusion Models: A Dual Evaluation of Metrics and Applications

SAR, recognized for its capability to capture images under various weather conditions and at any time, is served as a key technology in remote sensing domains. Nonetheless, the inherent limitations of imaging systems often result in SAR images that do not achieve high resolution, thereby constraining their specific applications. Drawing inspiration from the efficacy of diffusion models in generative tasks involving natural images, we have devised a model that converts Gaussian noise into a super-resolution image derived from low-resolution input, utilizing a Markov chain framework and integrating residual prediction to facilitate faster convergence. The performance of this model is assessed in comparison to leading methodologies, including Generative Adversarial Networks and Transformers, employing both traditional evaluation metrics and semantic segmentation tasks. Experimental findings derived from the MSTAR and SARBake datasets reveal that the super-resolution algorithm utilizing a diffusion model exhibits superior performance relative to alternative models, as evidenced by conventional metrics. This algorithm not only demonstrates exceptional visual quality but also proves effective in downstream applications, underscoring its considerable potential for tasks involving super-resolution of SAR images.