Cloud obscuration in optical remote sensing imagery significantly degrades data quality and limits its utility for downstream analysis and practical applications. The rich multi-dimensional information inherent in multi-temporal optical images offers substantial potential for the reconstruction of cloud-covered areas. To this end, this research presents an efficient reconstruction framework that progressively integrates spectral–temporal–spatial information and employs a coarse-to-fine multi-scale optimization strategy, without requiring external cloud masks as input. The proposed method involves a progressive refining process for the identification of thick clouds and associated shadows. Spectral characteristics are first used to extract regions with anomalous luminance, while temporal information enables the differentiation of clouds and shadows from similar ground objects, accounting for the constantly changing nature of clouds. Spatial correlations between clouds and the associated shadows are also incorporated to enhance detection accuracy. For the reconstruction stage, a cross-scale cascade optimization strategy is developed to restore the identified cloud and shadow areas. At the coarse scale, a structural consistency constraint with temporal smoothness weighting is constructed to narrow down the search for compensatory patch groups to a localized region, thereby ensuring spatial continuity and improving computational efficiency. At the fine scale, patch groups are formed and restored at smaller units, enhancing the local accuracy and detail fidelity. Extensive experiments validate the superior performance of the method and demonstrate its robustness across varying image sizes and cloud coverages, even under challenging conditions where all temporal images are partially cloud-affected. The code for this research is publicly available at https://github.com/YuXiaoyu221/STSIPI-CSR.
Yu et al. (Tue,) studied this question.