Answers are grounded in this paper and cite their sources.
Although optical remote sensing imagery has been extensively used in various Earth observation tasks, its practical utility is frequently compromised by cloud cover and adverse weather, which lead to missing data and disrupt the temporal continuity of observations. In contrast, Synthetic Aperture Radar (SAR) is an active remote sensing technology that functions independently of solar illumination and atmospheric conditions. Its ability to penetrate clouds and fog ensures the provision of reliable all-weather imagery. Motivated by the complementary characteristics of optical and SAR observations, we present OS-MambaDiff, a Mamba-based diffusion framework to fuse SAR information and restore cloud-contaminated optical images. This method can effectively fuse information from SAR and optical images without the need for a cloud mask, achieving high-precision cloud-free reconstruction. To this end, we design a Mamba-based Cross Modal Fusion Module (MCMF), which introduces Mamba state space modeling and its 2D extension SS2D to enable efficient feature fusion between heterogeneous modalities. Furthermore, a Cloud Loss based on differentiated constraint is designed. This loss function guides the model to learn cloud-free features while suppressing overfitting to cloud-contaminated condition images, thereby preventing the model from simply copying the conditional image output. Additionally, a Sparsity-Aware Gated Fusion (SAGF) mechanism is constructed to adaptively inject fused SAR and optical features during the diffusion denoising process, while sparse regularization is employed to achieve stability and selectivity in feature modulation. Extensive experiments on the newly released cloud removal benchmark dataset SMILE-CR and the large-scale benchmark cloud removal dataset SEN12MS-CR show that OS-MambaDiff outperforms existing optical synthetic aperture radar (Optical–SAR) cloud-removal techniques.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: