DisentangledMamba: Decoupling Spectral-Spatial Dependencies for Hyperspectral Image Classification

Hyperspectral image classification faces a fundamental challenge in that existing deep learning methods process spectral and spatial information jointly, inevitably entangling two physically distinct types of dependencies and degrading the discriminative capacity of each dimension. This article proposes the Disentangled Mamba Network (DisentangledMamba), a selective state space model (SSM) based network that addresses this limitation through explicit spectral-spatial decoupling. The architecture incorporates four contributions. First, a multiscale input projection applies four parallel convolutional branches with distinct receptive fields to derive complementary spectral and spatial structural representations before sequence modeling. Second, the feature stream is partitioned into spectral and spatial components, each directed to a dedicated SSM branch with independent parameters: the spectral branch employs a depthwise convolutional kernel of size k = 7 for interband correlation modeling, while the spatial branch adopts k = 4 for local neighborhood structure. Third, a cross-linear interaction module injects bounded residual signals between branches after independent processing, implementing a sequential strategy of decoupling followed by selective recoupling. Fourth, a content-based adaptive weight module assigns per-sample branch weights through Sigmoid normalization, permitting flexible spectral-spatial integration. To support stable training across six stacked SSM blocks, numerical constraints are applied at multiple levels—including clamped discretization matrices, bounded hidden states, restricted skip-connection gains, and scaled residual outputs—collectively preventing state divergence without external gradient clipping. Experiments on the Houston2013 and XiongAn datasets demonstrate that DisentangledMamba achieves the highest overall accuracy, average accuracy, and Kappa coefficients among fourteen competing methods spanning convolutional neural networks-based, attention-based, Transformer-based, and SSM-based categories, while maintaining a compact model size of 12.91 MB.