SGN-CR: a SAR-guided dual-branch network for cloud removal in remote sensing imagery

Optical remote sensing imagery plays a vital role in Earth observation. However, persistent cloud coverage frequently obscures critical ground information and limits downstream applications. Existing multimodal fusion methods attempt to leverage Synthetic Aperture Radar (SAR) as an auxiliary source, but often suffer from inadequate inter-modal interaction and the propagation of SAR speckle noise. In this paper, we propose a SAR-guided dual-branch network for cloud removal, which provides an effective SAR-guided structural guidance approach within a heterogeneous CNN-transformer architecture. Specifically, a SAR-guided attention modulation mechanism injects deep SAR structural priors into optical feature extraction to restore textures in thick-cloud regions. To address the trade-off between noise suppression and semantic alignment, we introduce a hierarchical collaborative feature fusion strategy, using spatially adaptive gated fusion at shallow levels to block speckle noise and cross-modal cross-attention at deep levels for precise semantic completion. In addition, a cross-axis attention mechanism captures long-range spectral dependencies while reducing computational complexity. Experimental results on the SEN12 MS-CR dataset demonstrate high quantitative and qualitative performance. The proposed method effectively preserves spectral fidelity and restores sharp geometric structures, demonstrating its effectiveness in reconstructing high-resolution multimodal remote sensing data.