FCFNet: A Frequency-Domain-Guided Cross-Modal Feature Fusion Network for Cloud Removal

Optical remote sensing images are frequently contaminated by clouds and shadows, leading to information loss that significantly degrades the performance of Earth observation tasks. Synthetic Aperture Radar (SAR), with its cloud-penetrating capability, serves as a powerful complement to optical imagery under dense cloud coverage. However, most existing SAR-optical fusion approaches mainly rely on spatial-domain operations, underutilizing frequency-domain structural cues and lacking effective strategies to resolve spatial misalignment between heterogeneous modalities. To address the limitations above-mentioned, we propose FCFNet, a Frequency-Domain Guided Cross-Modal Feature Fusion Network that jointly exploits spatial and frequency representations for cloud removal. Specifically, the FCFNet comprises three key components: a Deformable Gated Fusion (DGF) module to adaptively align SAR and optical features via learnable spatial offsets and channel-wise modulation; a Frequency-Aware Fusion (FAF) module that decomposes features into distinct frequency bands for selective integration of structural and textural information; and a Frequency-Domain Attention (FDA) mechanism that enhances high-frequency detail recovery in decoding. Additionally, we formulate a jointly optimized loss function that aligns with the network’s dual-domain design, promoting accurate reconstruction through spatial supervision and frequency-based constraints. Extensive experiments on the SEN12MS-CR dataset demonstrate that FCFNet outperforms state-of-the-art methods across various quantitative metrics, particularly under heavy cloud occlusion scenarios, validating the effectiveness of our frequency-space cooperative modeling strategy.