DualPathmamba: a dual-path state space model for hyperspectral image classification

Current hyperspectral image classification methods face two major challenges: convolutional neural networks struggle to effectively capture long-range dependencies in the spectral dimension, while Transformer architectures suffer from inefficiency due to quadratic computational complexity when processing hyperspectral data with hundreds of bands. This paper proposes DualPathMamba, a state space model-based method featuring: (1) a dual-path architecture that simultaneously processes spectral sequences through SSM operations and spatial features through convolutional operations, achieving complementary feature fusion through element-wise addition; (2) linear computational complexity O(L) achieved through state space modelling with four core operations – time modulation, state transition, input mapping, and output mapping – compared to quadratic complexity O(L2) in attention-based methods; (3) efficient processing of high-dimensional spectral data while maintaining approximately 2.0 M parameters. Experiments demonstrate that DualPathMamba achieves overall accuracies of 93.85%, 98.85%, and 94.52% on the Indian Pines, Pavia University, and Kennedy Space Center benchmark datasets, respectively, significantly outperforming existing methods while demonstrating superior computational efficiency and robustness.