Tracing origin and cultivation practice of Lithocarpus litseifolius via multi-data fusion and machine learning approaches

Lithocarpus litseifolius (sweet tea) is a medicinal and edible plant rich in flavonoids and essential nutrients, with potential as a hepatoprotective beverages and natural sweetener. Although widely cultivated across several provinces in China, the quality and consistency of its raw material remain poorly regulated. To address this, 163 samples (n ≥ 18) from 7 main producing regions were analyzed for 22 functional compounds, 4 stable isotope ratios, and 49 multi-element to discriminate cultivation practices and geographical origins. Orthogonal partial least squares discriminant analysis (OPLS-DA) successfully generated prediction models across two cultivation regions. Integrating 8 machine-learning algorithms with multi-level data fusion identified 6 key variables-caffeine, Rb, Ce, δ¹⁵N, Sr, and 3"-O-acetylphlorizin. Five base learners built on these variables were then combined via soft-voting ensemble learning, yielding an optimal origin classifier with 100.00% accuracy. Additionally, the study delivered the first comprehensive analysis of quality variations in sweet tea and identified seven primary influenced environmental factors, offering insights into cultivation strategies and quality formation mechanisms.