A global-local interaction and conditional consistency constrained diffusion model for SAR-guided optical image cloud removal

• A novel diffusion model framework is proposed in this study. • A global–local interaction multi-scale residual self-attention network is presented. • A new multi-condition consistency-constrained loss function is employed. Cloud cover constitutes a formidable obstacle in the field of optical remote sensing image processing, substantially impeding the extraction and utilization of surface information. Synthetic Aperture Radar (SAR) imagery, serving as a complementary informational resource, is capable of furnishing crucial auxiliary data for optical images. In recent years, diffusion-based cloud removal methodologies have made significant progress. Nevertheless, their inherent generative diversity and randomness pose challenges in meeting the realism requirements for cloud removal in optical remote sensing imagery. To address this, this paper presents a SAR-guided optical imagery cloud removal method based on global–local interaction and conditional consistency-constrained diffusion models (GLCdiffcr). Specifically, the method integrates a multi-scale residual self-attention network in the denoising module. This network captures both global and local details of SAR imagery and the captured details provide precise guidance for cloud removal. Additionally, within the reverse diffusion framework, the method directly predicts cloud-free optical images and iterates over multiple steps, reducing errors caused by generative randomness and improving consistency. Meanwhile, in order to enhance the realism of the generated images, the method employs a novel multi-condition consistency-constrained loss function, which combines pixel-level errors with structural similarity measures. Through this loss function, the gap between the generated images and real-world land cover types is further minimized. Experimental results demonstrate that the proposed method outperforms current state-of-the-art methods in both quantitative metrics and visual quality, particularly in complex regions, with higher accuracy and reliability.