Answers are grounded in this paper and cite their sources.
Few-sample hyperspectral image classification faces a fundamental representation bottleneck under limited supervision and computational resources. Reliable recognition requires subtle spectral differences to be interpreted jointly with local spatial structures. However, spectral filters learned freely from scarce labels may overfit sample-specific variations, while resolution reduction in lightweight networks may erase boundaries and small regions that are critical for distinguishing spectrally similar categories. These effects can accumulate during feature extraction, degrading both spectral reliability and spatial discriminability. To address this challenge, we propose SS-MCN, a lightweight network that couples anchored spectral adaptation with aliasing-aware spectral–spatial compression. SS-MCN introduces a basis–residual spectral filter bank in which a class-agnostic analytical basis provides a stable transformation anchor, while a scaled learnable residual enables task-specific adjustment. Unlike physical spectral priors, prototype spectra, or endmember dictionaries, the proposed anchor is not material-dependent; its role is to regularize the parameterization of spectral learning rather than encode category semantics. The resulting spectral responses are fused with multi-scale spatial features through a content-adaptive gate before resolution reduction, allowing spectral evidence and local structures to jointly shape the compressed representation. A fixed low-pass operation is then applied before downsampling to reduce aliasing-induced degradation. Experiments on three benchmark datasets show that SS-MCN improves overall accuracy by 0.98, 1.47, and 1.15% points on UP, WHU-Hi-HongHu, and WHU-Hi-HanChuan, respectively, compared with the best-performing competing methods. These results demonstrate the effectiveness of anchored spectral adaptation and aliasing-aware feature compression for lightweight few-sample hyperspectral image classification.
No takes yet. Share an insight, caveat, or question.
Dang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: