Optical remote sensing imagery plays a crucial role in a wide range of applications, including environmental monitoring, disaster assessment, and urban planning. However, the widespread presence of clouds severely degrades image quality and limits the reliability of downstream analysis. Therefore, accurately recovering cloud-contaminated regions has become a critical problem in remote sensing. Despite recent progress, remote sensing cloud removal methods still face two practical difficulties in complex real-world scenarios: preserving large-scale structural consistency and recovering high-frequency details that are uncertain. To address this problem, we propose SCReD (Structure-Consistent Residual Distribution), a two-stage, structure-guided residual refinement framework for single-image remote sensing cloud removal. Specifically, the first stage introduces a structure-enhanced coarse reconstruction module to improve spatial consistency and provide a more reliable structural condition. The second stage performs conditional latent diffusion in the residual space, where probabilistic modeling is restricted to high-frequency residual refinement rather than full-image regeneration. In this way, SCReD explicitly separates coarse structural recovery from uncertain detail refinement, thereby helping to balance structural stability and texture restoration under severe cloud occlusion. Extensive experiments on representative cloud removal datasets show that SCReD achieves competitive quantitative and visual performance. This is especially true in more challenging real-world scenarios. Additional analyses further show that the two stages play complementary roles within the proposed task-specific framework.
Zhang et al. (Tue,) studied this question.