Geographical origin traceability of green tea based on terahertz time-domain spectroscopy coupled with an optimized...

Rapid and objective geographical origin traceability is important for safeguarding the market value and Geographical Indication (GI) status of green tea. In this study, Xinyang Maojian green tea was used as a representative case to explore the feasibility of terahertz time-domain spectroscopy (THz-TDS) combined with machine learning for origin discrimination. Different spectral preprocessing strategies, four feature selection algorithms, and multiple classifiers were systematically compared. The results showed that Raw + SVM achieved the best performance among the full-frequency models, while RFE-SVM was the most suitable feature selection strategy, selecting 110 informative frequencies. On an independent test set based on tea-sample-level partitioning, the final Raw + RFE-SVM model achieved an Accuracy of 0.8833 and a Macro F1-score of 0.8819. SHAP analysis indicated that several low-frequency bands, such as 0.2167 THz, contributed to model classification and may serve as potential spectral indicators. Overall, this study suggests that THz-TDS combined with RFE-based feature selection and SVM classification has potential for supporting the geographical origin traceability of Xinyang Maojian tea.