Answers are grounded in this paper and cite their sources.
Hyperspectral image (HSI) classification requires capturing both global geometric structures and fine-grained local details. While graph neural networks (GNNs) offer powerful topological modeling, standard explicit graph construction suffers from two critical limitations: prohibitive resource consumption and the introduction of background noise that dilutes the spectral fidelity of the center pixel. Moreover, the smoothing nature of GNNs often leads to the loss of high-frequency local textures. To overcome these challenges, we propose the center-aware axial hierarchical sparse graph network (CHSG), a sparse graph learning framework designed to balance global topological reasoning with local detail preservation while avoiding explicit graph construction. First, the hierarchical sparse graph attention (HSGA) module employs an axial rolling strategy to achieve implicit graph topological awareness. Rather than calculating computationally intensive pairwise affinities, it functions as a directional aggregator that encodes topological continuity through roll-based neighborhood aggregation, effectively capturing anisotropic spatial structures. Second, the center-aware positional encoding (CAPE) module treats the target pixel as a spectral anchor, generating content-adaptive attention to explicitly filter out heterogeneous interference. Finally, a parallel spectral–spatial joint multiscale (SSJM) module serves as a texture compensator by aggregating multiscale spectral–spatial contexts to recover diverse local patterns smoothed out by the graph branch. Experiments on four benchmark datasets demonstrate that CHSG achieves state-of-the-art accuracy while maintaining a favorable balance between classification performance and computational cost.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: