Development of machine learning models using multi-source data for geographical traceability and content prediction of Eucommia ulmoides leaves

Rapid and scientific quality evaluation is a hot topic in the research of food and medicinal plants. With the increasing popularity of derivative products from Eucommia ulmoides leaves, quality and safety have attracted public attention. The present study utilized multi-source data and traditional machine learning to conduct geographical traceability and content prediction research on Eucommia ulmoides leaves. Explored the impact of different preprocessing methods and low-level data fusion strategy on the performance of classification and regression models. The classification analysis results indicated that the partial least squares discriminant analysis (PLS-DA) established by low-level fusion of two infrared spectroscopy techniques based on first derivative (FD) preprocessing was most suitable for geographical traceability of Eucommia ulmoides leaves, with an accuracy rate of up to 100 %. Through regression analysis, it was found that the preprocessing methods and data blocks applicable to the four chemical components were inconsistent. The optimal partial least squares regression (PLSR) model based on aucubin (AU), geniposidic acid (GPA), and chlorogenic acid (CA) had a residual predictive deviation (RPD) value higher than 2.0, achieving satisfactory predictive performance. However, the PLSR model based on quercetin (QU) had poor performance (RPD = 1.541) and needed further improvement. Overall, the present study proposed a strategy that can effectively evaluate the quality of Eucommia ulmoides leaves, while also providing new ideas for the quality evaluation of food and medicinal plants.