SAR-Conditioned Consistency Model for Effective Cloud Removal in Remote Sensing Images

Cloud contamination, especially thick cloud cover, severely limits the usability of optical remote sensing imagery by obscuring surface information. Due to the strong penetrability of microwave signals, Synthetic Aperture Radar (SAR) has emerged as an effective source for thick cloud removal. While SAR-assisted deep learning methods, such as CNNs and GANs, have made notable progress, the quality of generated imagery still requires improvement. Diffusion models, which offer strong potential for enhancing generation fidelity, could address this limitation but suffer from slow sampling speeds that constrain practical use and underscore the need for greater efficiency. To simultaneously enhance both reconstruction quality and sampling efficiency, this paper proposes a fast-sampling SAR-conditioned consistency model based on consistency distillation, named CM-CR, which adopts a teacher–student architecture to divide the reconstruction process into a rapid coarse prediction stage and a detailed refinement stage, significantly reducing per-scene processing time while maintaining high reconstruction fidelity. Specifically, a SAR-Conditioned Score-Based Diffusion Model (SCSBD) is first developed as the teacher network for learning a SAR-conditioned optical image generation model. Consistency distillation is then used to derive the student network SAR-conditioned consistency model (SCCM), which enables a rapid coarse prediction through single-step sampling. Finally, a Progressive Denoising via Multistep Resampling (PDMSR) strategy is introduced to iteratively refine the single-step output, producing fine-grained reconstructions. Comparative experiments conducted on the widely used cloud removal benchmark dataset SEN12MS-CR demonstrate that the proposed CM-CR method achieves state-of-the-art (SOTA) performance across all image quality metrics. Notably, although its design uses approximately 80 times more parameters compared with a standard Denoising Diffusion Probabilistic Model (DDPM), it delivers up to a 40-fold acceleration at inference.