Answers are grounded in this paper and cite their sources.
Patch-based hyperspectral image classification (HSIC) uses a spatial cube to predict the label of its center pixel, yet the descriptor passed to the classifier is often a whole-patch summary. This mismatch is mild in homogeneous neighborhoods, but in mixed or boundary-rich regions, uniform summarization can overemphasize peripheral responses and weaken center-relevant evidence. Existing Mamba-based HSIC networks usually address this issue indirectly by modifying the encoder, the token order, or the scan route. In this article, we focus instead on the postencoding aggregation step. The proposed CTMamba keeps a compact three-stage visual state-space backbone and adds four short residual refinement branches before classification: a center context adapter, a center pyramid adapter, a boundary contrast adapter, and a raw spectral statistics adapter. These branches share a common fusion dimension and are added to the pooled backbone descriptor through a single residual update. On Pavia University, Houston 2013, WHU-Hi-HanChuan, and WHU-Hi-HongHu, CTMamba achieves 98.35%, 97.49%, 94.48%, and 95.19% overall accuracy under a 30/10 per-class training protocol averaged over ten seeds, outperforming thirteen recent transformer- and Mamba-based baselines. Ablation, replacement, scan-route, region-robustness, patch-size, and threshold-sensitivity studies show that the gains mainly come from decision-stage refinement and remain stable under reasonable design changes.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: