TensorMamba: a tensorized state space model for hyperspectral image classification

In recent years, hyperspectral image classification (HSIC) technology has been rapidly evolving from traditional machine learning methods to deep learning methods. Among these methods, Transformers, with their powerful long-range modelling capabilities, have demonstrated outstanding performance in HSIC tasks. However, the quadratic complexity of its self-attention mechanism results in high computational costs, limiting its practical applications. As an emerging architecture, Mamba achieves efficient long-range modelling with linear complexity, making it a strong competitor to Transformer. However, existing Mamba-based methods generally extract the spatial and spectral features of HSI samples independently, overlooking their important spatial–spectral coupling characteristics and thereby limiting classification performance. We propose a tensor-operation neural network based on a state-space model, termed TensorMamba, to efficiently extract spatial–spectral coupling features. Specifically, an HSI sample is spatially divided into small patches, and the resulting patch sequence is fed into the state-space model (SSM). Within the SSM, Tucker Decomposition Inverse (TDI) operations are performed along the three modes of each tensor patch to extract spatial–spectral coupling features. This enhances the representation-learning capability of the model while effectively reducing the number of model parameters. The experimental results validate the favourable classification performance and lightweight advantages of TensorMamba. https://github.com/zhwt-xidian/TensorMamba