TY - JOUR
T1 - Wavelet-Driven Spectral-Spatial State-Space Modeling for Hyperspectral Image Classification
AU - Zeng, Sihao
AU - Wang, Zhen
AU - Li, Jiayuan
AU - Li, Ruixiang
AU - You, Zhuhong
N1 - Publisher Copyright:
© 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - Hyperspectral image (HSI) classification is challenged by high spectral dimensionality and complex spectral-spatial dependencies. Existing models, including convolutional neural networks (CNNs), Transformers, and Mamba, struggle to balance long-range modeling, computational efficiency, and preservation of spectral-spatial coherence. To address these limitations, we propose the wavelet-driven spectral-spatial state space network (WD4SNet), a novel and efficient architecture for HSI classification. Specifically, WD4SNet first employs a spectral-spatial fusion block (SSFB) to perform complementary integration of local spatial textures and spectral information at shallow layers, providing robust initial representations. The core of the network is a spectral-spatial state-space model (SSM), which consists of a spectral-based multiscale state space (SMSS) and a spatial-axis multiscale state space (SASS). Both modules incorporate wavelet-guided enhancement with learnable subband processing, enabling multiscale global and local spectral-spatial dependency modeling with linear complexity. To further enhance spectral-spatial synergy, a cross-axis spectral-spatial feature alignment (CASFA) module is introduced for adaptive alignment and fusion of spectral and spatial representations. Finally, a lightweight feature pyramid decoder (LFPD) reconstructs high-resolution prediction maps with minimal computational overhead. Extensive experiments on four benchmark hyperspectral datasets demonstrate that WD4SNet consistently outperforms state-of-the-art methods across multiple quantitative evaluation metrics. Furthermore, ablation studies validate the effectiveness of each module in advancing spectral-spatial fusion and representation learning. All codes and classification results will be available on https://github.com/darkseid-arch/WD4SNet
AB - Hyperspectral image (HSI) classification is challenged by high spectral dimensionality and complex spectral-spatial dependencies. Existing models, including convolutional neural networks (CNNs), Transformers, and Mamba, struggle to balance long-range modeling, computational efficiency, and preservation of spectral-spatial coherence. To address these limitations, we propose the wavelet-driven spectral-spatial state space network (WD4SNet), a novel and efficient architecture for HSI classification. Specifically, WD4SNet first employs a spectral-spatial fusion block (SSFB) to perform complementary integration of local spatial textures and spectral information at shallow layers, providing robust initial representations. The core of the network is a spectral-spatial state-space model (SSM), which consists of a spectral-based multiscale state space (SMSS) and a spatial-axis multiscale state space (SASS). Both modules incorporate wavelet-guided enhancement with learnable subband processing, enabling multiscale global and local spectral-spatial dependency modeling with linear complexity. To further enhance spectral-spatial synergy, a cross-axis spectral-spatial feature alignment (CASFA) module is introduced for adaptive alignment and fusion of spectral and spatial representations. Finally, a lightweight feature pyramid decoder (LFPD) reconstructs high-resolution prediction maps with minimal computational overhead. Extensive experiments on four benchmark hyperspectral datasets demonstrate that WD4SNet consistently outperforms state-of-the-art methods across multiple quantitative evaluation metrics. Furthermore, ablation studies validate the effectiveness of each module in advancing spectral-spatial fusion and representation learning. All codes and classification results will be available on https://github.com/darkseid-arch/WD4SNet
KW - Feature alignment
KW - hyperspectral classification
KW - pyramid decoder
KW - spectral-spatial modeling
KW - state-space model (SSM)
UR - https://www.scopus.com/pages/publications/105046131790
U2 - 10.1109/TGRS.2026.3717611
DO - 10.1109/TGRS.2026.3717611
M3 - 文章
AN - SCOPUS:105046131790
SN - 0196-2892
VL - 64
JO - IEEE Transactions on Geoscience and Remote Sensing
JF - IEEE Transactions on Geoscience and Remote Sensing
M1 - 5524921
ER -