Optical remote sensing images are often contaminated by clouds and shadows, resulting in missing data, which greatly hinders consistent Earth observation missions. Cloud and shadow removal is one of the most important tasks in optical remote sensing image processing. Due to the characteristics of active imaging that enable synthetic aperture radar (SAR) to penetrate cloud cover and other climatic conditions, SAR data are extensively utilized to guide optical remote sensing image cloud and shadow removal. Nevertheless, SAR data are highly corrupted by speckle noise, which generates artifact pollution to spectral features extracted from optical images and makes SAR-optical fusion ill-posed to generate cloud and shadow removal results while retaining high spectral fidelity and reasonable spatial structures. To overcome the aforementioned drawbacks, this paper presents a novel hierarchical spectral and structure-preserving fusion network (HS2P), which can recover cloud and shadow regions in optical remote sensing imagery based on the hierarchical fusion of optical and SAR remote sensing imagery. In HS2P, we present a deep hierarchical architecture with stacked residual groups (ResGroups), which progressively constrains the reconstruction. To pursue the adaptive selection of more informative features for fusion and reduce attention to the features with artifacts brought by clouds and shadows in optical data or speckle noise in SAR data, residual blocks with a channel attention mechanism (RBCA) are recommended. Additionally, a novel collaborative optimization loss function is proposed to preserve spectral features while enhancing structural details. Extensive experiments on the publicly open dataset (i.e., SEN12MS-CR) demonstrate that the proposed method can robustly recover diverse ground information in optical remote sensing imagery with various cloud types. Compared with the state-of-the-art cloud and shadow removal methods, our HS2P achieves significant improvements in terms of quantitative and qualitative results. The source code is publicly available at https://github.com/weifanyi515/HS2P.
