TY - JOUR
T1 - Exploring informative features in Chinese liquor (Baijiu) near-infrared hyperspectral data classification through attention mechanisms
AU - Chen, Danlei
AU - Wang, Yun
AU - Tang, Linruize
AU - Zhao, Zhengqiao
AU - Cohen, Israel
AU - Chen, Jie
AU - Lei, Tao
AU - Chen, Jingdong
N1 - Publisher Copyright:
© 2026 Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license. http://creativecommons.org/licenses/by-nc-nd/4.0/
PY - 2026/11
Y1 - 2026/11
N2 - Rapid and objective quality assessment of Chinese Liquor (Baijiu) remains challenging, as sensory evaluation is subjective and inconsistent, while Gas Chromatography–Mass Spectrometry is time-consuming and unsuitable for industrial-scale inspection. Near-infrared hyperspectral imaging (NIR-HSI) offers a promising non-destructive alternative, yet spectral redundancy and container reflections challenge model generalization. This work presents a NIR-HSI dataset of 300 images from 19 Baijiu products in quartz cuvettes and glass bottles, covering 936–1720 nm with 224 bands. A self-attention model is benchmarked against logistic regression, random forest, XGBoost, and a 1D-CNN on alcohol content (4-class, 9-class) and aroma type (5-class) classification. Attention-derived feature-ranking scores are obtained without additional inference, show high correlation with SHAP and permutation importance, and highlight wavelengths linked to O–H and C–H bond overtones within known absorption regions. Under a rank-then-retrain protocol, the cuvette data show that retaining 10% of important bands suffices for alcohol tasks and 20%–40% for aroma tasks, reducing inference time by approximately eight-fold. Moreover, rankings from one container type transfer to the other, indicating that the retained bands reflect intrinsic chemical information rather than container-specific artifacts. Reliable bottle-level voting supports product decisions, but external evaluation using independent products remains necessary.
AB - Rapid and objective quality assessment of Chinese Liquor (Baijiu) remains challenging, as sensory evaluation is subjective and inconsistent, while Gas Chromatography–Mass Spectrometry is time-consuming and unsuitable for industrial-scale inspection. Near-infrared hyperspectral imaging (NIR-HSI) offers a promising non-destructive alternative, yet spectral redundancy and container reflections challenge model generalization. This work presents a NIR-HSI dataset of 300 images from 19 Baijiu products in quartz cuvettes and glass bottles, covering 936–1720 nm with 224 bands. A self-attention model is benchmarked against logistic regression, random forest, XGBoost, and a 1D-CNN on alcohol content (4-class, 9-class) and aroma type (5-class) classification. Attention-derived feature-ranking scores are obtained without additional inference, show high correlation with SHAP and permutation importance, and highlight wavelengths linked to O–H and C–H bond overtones within known absorption regions. Under a rank-then-retrain protocol, the cuvette data show that retaining 10% of important bands suffices for alcohol tasks and 20%–40% for aroma tasks, reducing inference time by approximately eight-fold. Moreover, rankings from one container type transfer to the other, indicating that the retained bands reflect intrinsic chemical information rather than container-specific artifacts. Reliable bottle-level voting supports product decisions, but external evaluation using independent products remains necessary.
KW - Baijiu classification
KW - Feature selection
KW - Hyperspectral imaging
KW - Interpretable deep learning
KW - Near-infrared spectroscopy
KW - Self-attention mechanism
UR - https://www.scopus.com/pages/publications/105051340250
U2 - 10.1016/j.jfca.2026.109487
DO - 10.1016/j.jfca.2026.109487
M3 - 文章
AN - SCOPUS:105051340250
SN - 0889-1575
VL - 159
JO - Journal of Food Composition and Analysis
JF - Journal of Food Composition and Analysis
M1 - 109487
ER -