FA-Mamba: frequency attention driven Mamba for multimodal remote sensing classification

Multimodal remote sensing classification plays a vital role in areas such as resource exploration and disaster monitoring. However, noise and redundancy across different modalities often hinder effective feature fusion. Existing approaches mainly rely on Transformer and Mamba architectures. While Transformer-based models suffer from high computational cost and inefficient long-range dependency modeling, Mamba improves efficiency but still focuses primarily on spatial-temporal domain modeling, with limited exploration of frequency-domain. This restricts the discriminative capacity of the fused representations. To overcome these limitations, we propose FA-Mamba, a frequency attention driven Mamba multimodal remote sensing classification network. With a focus on computational efficiency ensured by linear complexity, FA-Mamba adaptively enhances the complementary information between modalities while effectively suppressing noise and redundancy in cross-modal data, enabling more robust feature fusion. Additionally, to effectively capture global contextual information, we design a coordinate fusion attention (CFA) mechanism. By integrating spatial cues along different directions, the method incorporates long-range correlations across modalities. This facilitates a more comprehensive understanding of complementary multimodal information. Extensive experiments conducted on three public multimodal remote sensing datasets demonstrate the effectiveness of our approach. FA-Mamba achieves superior performance compared to state-of-the-art methods, with an average overall accuracy of 95.84%. Our code is available at https://github.com/dnyeee/FA-Mamba.