Leakage-controlled chemometric discrimination of 82 tea cultivars using Vis–NIR–SWIR leaf spectra
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy · 2027-03 · 368 · 129093
DOI: 10.1016/j.saa.2026.129093 ↗Kurz gefasst
Fachartikel in der Zeitschrift Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2027-03. Verfasst von Zhichao Chai und 8 weiteren Autor:innen. Beteiligte Einrichtung: Key Laboratory of Biology, Genetics and Breeding of Special Economic Animals and Plants, Ministry of Agriculture and Rural Affairs, National Center f…. Die Arbeit nennt 42 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Rapid, reagent-free spectroscopic discrimination of closely related plant materials remains challenging because performance estimates can be inflated by information leakage and inadequate validation. We evaluated Vis-NIR-SWIR leaf spectroscopic fingerprints for internal closed-set discrimination of 82 tea cultivars. The dataset comprised 1640 leaf-level mean spectra, with one leaf collected from each of 20 distinct clonally propagated plants per cultivar and ten separately saved within-leaf measurements averaged for each selected leaf. Preprocessing, model fitting, and wavelength selection were restricted to the training data. Evaluation used ten repeats of five-fold stratified cross-validation, yielding 50 held-out evaluations. With nested α selection, SNV + RidgeClassifier achieved an accuracy of 0.9937 ± 0.0040, a Macro-F1 of 0.9936 ± 0.0040, Cohen's kappa of 0.9936 ± 0.0040, and a top-3 accuracy of 0.9973 ± 0.0024. Nested Raw + RidgeClassifier showed comparable performance (Macro-F1 0.9933 ± 0.0048). Expanded nested raw and SNV PLS-LR comparisons achieved Macro-F1 values of 0.8187 and 0.9032, respectively, within the tested 5-160-component grid; the upper boundary was still frequently selected. When SNV was recomputed solely from the selected raw wavelength inputs, 50- and 100-band models retained Macro-F1 values of 0.9612 and 0.9821, respectively. Five-seed out-of-fold decision margins concentrated errors in low-confidence predictions; globally retaining 80% of 8200 prediction events left two errors among 6560 events, without providing an externally calibrated rejection threshold. These results demonstrate that leakage-controlled Vis-NIR-SWIR spectroscopic fingerprinting can support reliable internal cultivar discrimination, uncertainty-aware sample flagging, and candidate wavelength prioritization. External validation across seasons, locations, populations, and instruments remains necessary.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Key Laboratory of Biology, Genetics and Breeding of Special Economic Animals and Plants, Ministry of Agriculture and Rural Affairs, National Center for Tea Plant Improvement, Tea Research Institute, Chinese Academy of A…
- Tea Research Institute, Fujian Academy of Agricultural Sciences, Fuzhou 350013, Fujian, China. Electronic address: fengjingliu@faas.cn.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Chai, Z., Qian, X., Yu, Y., Li, X., Zhang, J., Li, X., Nie, C., Liu, F., & Sun, X. (2027). Leakage-controlled chemometric discrimination of 82 tea cultivars using Vis–NIR–SWIR leaf spectra. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 368, 129093. https://doi.org/10.1016/j.saa.2026.129093
Vancouver
Chai Z, Qian X, Yu Y, Li X, Zhang J, Li X, et al. Leakage-controlled chemometric discrimination of 82 tea cultivars using Vis–NIR–SWIR leaf spectra. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy. 2027;368:129093. doi:10.1016/j.saa.2026.129093
BibTeX
@article{Chai2027,
author = {Chai, Zhichao and Qian, Xiaona and Yu, Yongchen and Li, Xiaofeng and Zhang, Jinghe and Li, Xiwang and Nie, Chenlu and Liu, Fengjing and Sun, Xiaoling},
title = {Leakage-controlled chemometric discrimination of 82 tea cultivars using Vis–NIR–SWIR leaf spectra},
journal = {Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy},
year = {2027},
volume = {368},
pages = {129093},
doi = {10.1016/j.saa.2026.129093}
}
EUVIMED