Lightweight Spatial–Spectral Shift Module With Multihead MambaOut for Hyperspectral Image Classification

In hyperspectral images, the high dimensionality of spectral data often leads to redundant spectral information, making it difficult to extract features. Two-dimensional CNNs fail to effec-tively extract spatial and spectral information, and deploying three-dimensional CNNs on microprocessors is challenging as these net-works consume excessive resources. If graph convolutional networks (GCN) are adopted, most networks employ superpixel segmentation for HSI classification. However, this approach tends to overlook pixel-level features and thus fails to achieve fine classification.To efficiently extract spectral features while reducing resource consumption, we proposed the spectral shift module (SPCSM). This module extracts features by circularly shifting spectral information. It emphasizes the internal correlations within the spectral and offers the advantage of fewer parameters. Based on the SPCSM, we designed the spatial–spectral shift module (S²SM). It models and analyzes correlations between spatial and spectral data, facilitating the extraction of spatial and spectral features. Additionally, it addresses the issues of redundancy in shallow information. To enable deployment on microprocessors and effectively classify hyperspectral images, we optimized the MambaOut network by integrating the S²SM with a multihead MambaOut (LS²SM-MHMambaOut). Experiments demonstrated that our network struck a favorable balance between classification accuracy and model complexity. Therefore, it is promising to be integrated into microprocessors.