Frequency-Driven State-Space Model for Remote Sensing Pansharpening

Deep learning has significantly advanced the field of pansharpening by enabling more effective spatial-spectral fusion. However, most existing methods operate in a unified representation space, where spatial details and global structural semantics are implicitly entangled. They mainly employ shallow transformations without explicit high-frequency subband-specific enhancement, and are inadequate in modeling the long-range dependencies required to preserve low-frequency structural consistency. To address these limitations, we propose a Frequency-Driven State Space Model (FDMamba), a novel Mamba-based architecture that individually considers high-and low-frequency components and leverages wavelet transforms for multi-scale decomposition, enabling frequency-aware sequence modeling. Specifically, our approach comprises two key parts: (1) High Frequency Enhancement Module (HFEM), equipped with directionally selective convolution blocks to refine anisotropic textures and orientation-sensitive edges within the wavelet subbands; (2) Low Frequency Mamba Module (LFMM), employing the state space model to efficiently capture long-range dependencies in low-frequency components with linear complexity. To enable effective cross-frequency interaction, we further introduce a frequency-guided querying mechanism, where enhanced high-frequency features dynamically guide the token selection. Extensive experiments on three representative benchmark datasets, including WorldView-3, GaoFen-2, and QuickBird, demonstrate that our method achieves state-of-the-art performance against existing approaches. Our code will be available at https://github.com/zhangxc0105/FDMamba.