EFSANet: Explicit Frequency Modeling and Style-Aware Alignment for Hyperspectral–Multispectral Image Fusion

The fusion of low-resolution hyperspectral images (LR-HSIs) and high-resolution multispectral images (HR-MSIs) is an effective strategy for reconstructing high-resolution hyperspectral images (HR-HSIs), and is important for remote sensing observation, land-cover classification, and environmental monitoring. In recent years, deep learning-based hyperspectral-multispectral fusion methods have achieved remarkable progress. However, existing methods still face two key challenges. First, differences in spectral response functions and spatial resolutions introduce significant cross-modal discrepancies, which hinder accurate feature alignment and joint modeling. Second, high-frequency structural information is often weakened during deep feature extraction and interaction, leading to blurred textures, degraded edges, and reduced spatial fidelity in reconstructed images. To address these issues, this paper proposes an Explicit Frequency Modeling and Style-Aware Alignment Network (EFSANet) for hyperspectral and multispectral image fusion. The proposed method integrates frequency-domain enhancement, spatial contextual modeling, and style-adaptive correction to reduce cross-modal discrepancies and recover fine spatial details. Specifically, the Pyramid Frequency Fusion (PFF) module explicitly decomposes multi-scale frequency components and adaptively reweights them to enhance structural details. The Fine-grained Frequency-domain Collaboration (FFC) module captures long-range spatial dependencies and promotes global contextual interaction. The Style-Adaptive Refinement (SAR) module performs style-aware feature correction based on global statistics, reducing inter-modal discrepancies and improving spectral-spatial consistency. Extensive evaluations across both simulated and real-world datasets confirm the effectiveness and superiority of the proposed EFSANet over state-of-the-art methods in preserving fine spatial details and spectral consistency.