Answers are grounded in this paper and cite their sources.
Standard hyperspectral image classification methods, typically dependent on scalar-neuron backbones, often struggle to capture the complex spatial hierarchies and pose information of land covers. Although capsule networks offer vector-based representations to mitigate this, existing iterations suffer from limited long-range modeling, weak feature discriminability, and high latency due to iterative routing. In this work, we propose GeoCaps, a lightweight framework designed to overcome these bottlenecks. First, we introduce a ghost-graph interaction module, which efficiently fuses local spatial textures and global geometric topologies via a dual-branch architecture, replacing heavy preprocessing layers. Second, to enhance vector distinctiveness, we design the UltraCapsule backbone equipped with orthogonal capsule attention. This mechanism dynamically recalibrates feature weights along both capsule and dimension axes with linear complexity, ensuring highly discriminative representations. Finally, we propose geometric consensus routing, a noniterative strategy that aggregates vectors based on geometric centrality, significantly accelerating inference. Extensive experiments on four benchmark datasets demonstrate that GeoCaps outperforms state-of-the-art methods in accuracy while maintaining superior computational efficiency.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: