Cloud Removal Method for Single-Point-of-View Optical Satellite Images in Tidal Flat Environment Using Guided Sampling of DDPM

Optical satellite imagery is severely constrained by cloud contamination, particularly in dynamic coastal environments such as tidal flats, where tidal variations are significant. In this study, we propose a diffusion model-based technique for cloud removal (CR) that operates solely on single-point-of-view satellite images without relying on multitemporal imagery or external sensor data. Specifically, a guided sampling strategy inspired by classifier guidance is employed to leverage gradient information from cloud-free reference regions, enabling image restoration that accurately reflects the underlying topography and temporal characteristics. The proposed method is evaluated against existing inpainting and repaint approaches under varying cloud cover conditions (10%–50%). In terms of quantitative metrics such as structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), correlation coefficient (CC), and root-mean-squared error (RMSE), the proposed method demonstrates excellent stability and superior reconstruction quality compared to conventional techniques. Even in cases of extensive cloud coverage, the reconstructed topography structures closely resemble those of the original imagery. Furthermore, the land–water boundary is extracted using the normalized difference water index (NDWI) by integrating UAV-derived digital elevation map (DEM) and tidal data. A comparison between the extracted boundaries and the images reconstructed using the proposed method confirms a high degree of agreement with the actual topography.