Acta Scientiarum Polonorum Technologia Alimentaria

Background. Tea, a globally popular beverage, is frequently adulterated with various additives for profit. However, research on these additives – particularly their common combinations in fraudulent practices – remains limited.
Materials and methods. This study uses mid-infrared (MIR) spectroscopy and machine learning techniques to identify such additives in tea. We trained several models – AdaBoost, random forest, K-nearest neighbours (KNN), support vector classification, Gaussian naive Bayes, and decision tree – to detect adulterants such as white sugar, talcum powder, sulphur, paraffin wax, food colouring, and flavourings, as well as their combina­tions. To improve accuracy and efficiency, we employed the Successive Projections Algorithm (SPA) and Competitive Adaptive Reweighted Sampling (CARS) for feature selection, prioritising features that exhibited the strongest correlations with the additives.
Results. Compared to models relying solely on raw spectra, those incorporating SPA and CARS consistently achieved high accuracy and reduced detection time by at least 90%. The selected wavenumbers also serve as biomarkers for specific additives. The SPA-KNN model performed exceptionally well, with an accuracy of 0.956, macro-precision of 0.964, macro-recall of 0.956, and a macro-F1 score of 0.956, all with a detection time of just 0.9 seconds.
Conclusion. These results highlight the effectiveness of combining MIR spectroscopy with advanced clas­sifiers and feature selection algorithms, offering a rapid and precise method for detecting tea additives. This approach benefits producers, distributors, and consumers alike.