Relational Graph-Guided Selective State-Space Model for Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification remains challenging under extreme label scarcity and spatially disjoint supervision, where models must generalize across heterogeneous materials while producing spatially consistent maps at scale. This article proposes RGMamba (RGM), a relational graph-guided selective state-space model (SSSM) for HSI classification in scarce-label, disjoint settings. RGM constructs physically grounded relational neighborhoods by jointly enforcing spatial proximity and spectral affinity, thereby promoting materially coherent context. Neighborhood-derived distributional statistics are fused with spectral signatures to enrich per-pixel representations while preserving pixel-to-map correspondence. These graph-derived descriptors are directly interpretable in terms of spatial–spectral neighborhood consistency and spectral homogeneity; however, the subsequent dual-branch encoder and SSSM model are learned transformations and do not share the same level of transparency. A pixel-aligned dual-branch encoder integrates local spatial aggregation and bandwise spectral mixing without patch pooling, and an SSSM captures long-range dependencies with linear-time sequence complexity. While the SSSM itself scales linearly with sequence length, the overall framework incurs additional computational overhead from relational neighborhood construction, graph-statistics extraction, and pixel-aligned token processing. Experiments on six benchmark datasets show that RGM consistently outperforms representative Transformer and state-space baselines under scarce supervision. Under a stringent spatially disjoint protocol with 0.05% labeled samples, RGM achieves over 96% overall accuracy.