Skip to content
Open access

SDMCNet: Spatial–Spectral Decoupled Mamba–MoE Collaborative Network for Hyperspectral Image Classification

2026 · IEEE Transactions on Geoscience and Remote Sensing · Vol 64, pp. 5530918-5530918 · 0 citations · 44 references

Abstract

Using the whole image as input for hyperspectral image (HSI) classification, the model needs to handle broader spatial dependencies and more diverse spectral variations. In modeling nonlocal spatial relationships, scanning based on fixed windows is prone to directional bias, leading to the loss of spatial boundary details. In addition, the high dimensionality of the data makes spectral distributions complex and spectral information redundant, thereby affecting the reliable identification and effective utilization of spectral bands. To address these challenges, we propose SDMCNet, a spatial–spectral decoupled Mamba–mixture-of-expert (MoE) collaborative network for image-level HSI classification. Taking the whole image as input, SDMCNet achieves a balance between spatial structure modeling and fine-grained spectral representation. Specifically, the spatial branch progressive spatial perception feature extraction (PSPFE) integrates local and global Mamba modules to capture short- and long-range spatial contextual dependencies. A boundary constrained enhancement module (BCEM) is further introduced to suppress oversmoothing and noisy perturbations near class boundaries, thereby preserving spatial boundary details. The spectral branch anchor routed expert spectral aggregation (ARESA) builds an anchor routed expert collaboration mechanism to adaptively select complementary spectral experts to model complex dependencies among spectral bands and obtain precise spectral band combinations. The two branches collaborate through spectral channel weighting, thereby achieving synergistic spatial–spectral modeling for image-level HSI classification. Extensive experiments on the Indian Pines (IP), Pavia University (PU), Houston 2013, and HanChuan (HC) datasets demonstrate that SDMCNet consistently outperforms representative methods, validating the effectiveness and potential value of using the whole image as input for HSI classification. The code is available at https://github.com/chai-xue/SDMCNet

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.