The misuse of Baijiu origin labeling poses significant risks to both producers and consumers. In this study, the geographic origin of Sichuan-style Nongxiangxing Baijiu was investigated using attenuated total reflectance Fourier transform infrared (ATR-FTIR) spectroscopy combined with machine learning techniques. A rapid, non-destructive, and cost-effective analytical strategy was developed by integrating spectral preprocessing, the Synthetic Minority Oversampling Technique (SMOTE), and feature selection. Under specific model configurations, SMOTE combined with k-Nearest Neighbors (KNN) and Support Vector Machines (SVM) achieved 100% classification performance on the validation set, indicating strong discriminative capability under controlled experimental conditions. Furthermore, the introduction of a bionic feature selection strategy improved model generalization performance. The KNN model constructed using 14 selected spectral markers achieved an accuracy of 91.67% on an independent test set, representing a 41.67% improvement compared with the model without feature selection. Shapley additive explanations identified wavenumbers at 2360 cm⁻¹, 2343 cm⁻¹, and 2336 cm⁻¹ as the most influential spectral features contributing to classification. Overall, this study demonstrates the potential of ATR-FTIR spectroscopy integrated with intelligent data-driven strategies for Baijiu origin tracing and provides a transferable framework for non-targeted spectral classification tasks. • A non-destructive method was developed to trace the origin of baijiu. • Category imbalance will affect the training and identification of classifiers. • A bionic feature selection strategy effectively enhanced model generalizability. • Independent test set validation highlights the practical application of the model. • SHAP analysis revealed critical wavelengths contributing to classification decisions.
Zhou et al. (2026) studied this question.