SSP-Mamba: Spatial–spectral pyramid Mamba for hyperspectral image classification

Hyperspectral image classification has always received wide attention in the field of remote sensing. However, due to the multidimensional and redundant information of hyperspectral data, many hyperspectral imaging (HSI) classification methods are unable to balance performance and computational efficiency, which brings great challenges to HSI classification. A recently proposed Mamba based on selective state space model (S6) can achieve remote sequence modeling while still maintaining high computational efficiency. Meanwhile, the pyramid hierarchical structure enables flexible feature extraction of data and reduces the resource consumption of the model. In view of this, this article combines S6 with the pyramid structure to form a pyramid Mamba structure, and proposes a spatial–spectral pyramid Mamba (SSP-Mamba) for HSI classification. SSP-Mamba performs feature extraction through two channels: space and spectrum. Each channel includes a one-dimensional spectral (spatial) sequence generation module and a pyramid Mamba feature extraction module. The features extracted from the two channels are first fused and enhanced, then input into the linear classification layer for classification. In addition, in order to fully capture spatial contextual information, a feature enhancement module (FEM) is introduced before generating the spatial sequence, effectively reducing the loss of local spatial information. In this model, the pyramid Mamba structure demonstrates efficient feature extraction capabilities, while the dual-channel structure fully utilizes spectral and spatial information features. A large number of experiments conducted on four widely used public HSI datasets have shown that the proposed SSP-Mamba achieves better classification performance, opening up a new window for HSI classification.