Thick Cloud Removal Using Cloud Mask-Guided SAR-Optical Fusion Network

Thick cloud cover in optical satellite imagery often leads to significant loss of surface information. Synthetic Aperture Radar (SAR) provides a complementary data source to assist in reconstructing cloud-covered regions. Recently, deep learning-based SAR-optical fusion approaches have been widely explored for cloud removal, but two key challenges remain: (1) the difficulty of cross-modal feature alignment due to inherent modality differences, and (2) coherent SAR noise interferes with the reconstruction quality. Specifically, global interactions across modalities are often underutilized due to the computational burden of full-scene modeling in recent research. Meanwhile, although cloud masks can localize contaminated regions and enable spatially selective attention, their potential for region aware noise suppression has been largely ignored. To address these limitations, we propose PSF-CR, a novel cloud-mask guided, global perception-enhanced SAR-optical fusion network for cloud removal. PSF-CR incorporates a gated attention module guided by the cloud mask to constrain SAR noise and enhance fusion in contaminated regions. To further emphasize on key regions and enhance fine-grained texture recovery while preserving global consistency, we incorporate the cloud mask features into a weighted L1 loss and introduce a composite spatial-frequency loss for joint supervision. In addition, PSF-CR incorporates two frequency-domain attention modules that leverage the Fourier Transform to reduce computational complexity: a global cross-attention module for cross-modal alignment, and a global self-attention module for full-scene reconstruction. Extensive experiments on two public multimodal remote sensing datasets demonstrate the competitive performance of our method against state-of-the-art approaches.