Answers are grounded in this paper and cite their sources.
Demonstrates improved hyperspectral classification in noisy conditions, indicating robust performance enhancements.
Existing noisy-label hyperspectral image classification (HSIC) methods usually address clean sample selection and representation regularization as separate problems, although the reliability of observed labels varies substantially across samples in hyperspectral data. This issue is amplified by mixed pixels, boundary ambiguity, spectral overlap, and limited labeled samples, which make hard clean samples difficult to distinguish from mislabeled ones. We therefore propose PDAM, a sample-reliability-guided training framework for noisy-label HSIC. The method first estimates feature-space class consistency by comparing each sample with the prototype of its observed class and converting this consistency into a reliability probability with a Gaussian mixture model. To reduce conservative false negatives, matched high-confidence selection is further used to recover hard but correctly labeled samples. The resulting reliability estimate then determines how strongly the observed label is trusted through target refinement and how strongly the input is perturbed through reliability-guided masking. Finally, masked reconstruction provides label-independent structural regularization so that uncertain samples can still contribute to spectral–spatial representation learning. Under the evaluated synthetic symmetric noise settings on the University of Pavia (UP), Salinas Valley (SV), and Kennedy Space Center (KSC) datasets, PDAM achieves the best OA and Kappa in most reported comparisons and improves robustness under both moderate and severe noise. At 30% noise, PDAM reaches 97.30% OA on UP, 98.13% OA on SV, and 95.37% OA on KSC. Ablation studies further support the necessity of reliability estimation, hard clean sample recovery, and reliability-guided supervision and regularization within this unified training mechanism.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: