Answers are grounded in this paper and cite their sources.
Hyperspectral image (HSI) classification plays a crucial role in remote sensing. Recently, Mamba has been introduced into HSI classification owing to its linear computational complexity. However, existing methods still struggle to model complex nonlinear spectral responses and to distinguish complementary high-frequency and low-frequency information. To address these limitations, we propose a multiscale Kolmogorov–Arnold Mamba network (MS-KAMNet). The proposed network consists of three core modules: the multiscale spatial–spectral feature aggregation module (MS-SFAM), the multiscale spiral Kolmogorov–Arnold Network (KAN) Mamba (MS-SKM) module, and the Fast-KAN-guided frequency-domain module (FK-FDM). Specifically, the MS-SFAM uses multiscale spatial–spectral dilated convolutions and group shuffling to enhance cross-scale feature interactions. The MS-SKM module serializes HSIs via multiscale spiral scanning and integrates KAN with the Mamba architecture, enabling efficient joint modeling of spatial contextual information and complex nonlinear spectral characteristics while maintaining linear computational complexity. The FK-FDM explicitly separates low-frequency and high-frequency components through wavelet decomposition. It enhances low-frequency global structures using directional pooling and captures high-frequency local details through multiscale Fast-KAN convolutions, thereby achieving complementary spatial–frequency representation. Extensive experiments on Pavia University, OHID-1, and Wenxian demonstrate the effectiveness of MS-KAMNet, with overall accuracy values of 99.08%, 93.52%, and 95.42%, respectively.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: