Answers are grounded in this paper and cite their sources.
Investigates a novel method to improve classification accuracy in hyperspectral images, suggesting effective label noise management.
Supervised classification serves as a fundamental technique for hyperspectral remote sensing analysis, whereas label noise within training samples inevitably degrades classification accuracy and weakens model robustness. To alleviate the adverse effects induced by label noise, this study develops a novel Propagation Filtering with Dynamic Threshold Collaborative Detection (PF-DTCD) method. The proposed framework firstly implements propagation filtering on all training samples. Afterwards, it calculates the gray-level distance of each sample pre- and post-filtering as well as the minimal inter-class distance, which is dynamically assigned as the discrimination threshold. Samples with gray-level distances exceeding the threshold are identified as label noise and discarded to purify the training set. Extensive experiments conducted on two hyperspectral datasets with multiple mainstream classifiers verify the outstanding capability of PF-DTCD for label noise identification and elimination. Taking the KSC dataset as an example, the presented method achieves maximum improvements of 9.25% in Overall Accuracy (OA), 12.26% in Average Accuracy (AA) and 9.91% in Kappa coefficient, which remarkably promotes the comprehensive classification performance.
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: