CentralGraphFormer: Center-Guided Spectral–Spatial Graph Transformer for Hyperspectral Image Classification

Hyperspectral image (HSI) classification requires effective exploitation of spectral–spatial information under high spectral redundancy, spatial heterogeneity, and limited labeled samples. Existing CNN- and Transformer-based methods often treat surrounding pixels uniformly and rely on dense attention or shallow fusion strategies, which are poorly aligned with the inherently center-focused supervision of patch-based classification and incur high computational cost. To address these issues, we propose CentralGraphFormer, a center-guided spectral–spatial representation framework that explicitly aligns feature learning and global reasoning with the central-pixel decision objective. The framework integrates three stages: (i) a lightweight spectral shift separable network for efficient spectral continuity modeling and redundancy suppression, (ii) a spatial relevance screening mechanism that adaptively selects semantically consistent neighbors via stochastic gating and biologically inspired smoothing, and (iii) a Central Graph Attention Module (CGAM) that performs sparse, center-anchored global reasoning using graph-aware keys and a single center query, avoiding dense self-attention. Extensive experiments on four benchmark HSI datasets demonstrate that CentralGraphFormer consistently outperforms state-of-the-art methods while maintaining favorable computational efficiency. Ablation studies further confirm the effectiveness and complementarity of the proposed center-guided spectral, spatial, and global modeling components.