Isotope-assisted data mining techniques for authenticating rice across Chinese markets

This study aims to authenticate the geographical origin and cultivar (Japonica vs Indica) labeling of rice sold across Chinese markets using stable isotope analysis coupled with multivariate analyses and machine learning models. δ ¹³C, δ ¹⁵N, δ ¹⁸O, and δ ²H were investigated in rice using isotope ratio mass spectrometry (IRMS) and analyzed through PCA and t-SNE to visualize regional separation and cultivar differences. Different machine learning models (kNN, SVM, RF) were developed using authentic field samples and tested with market-sourced samples for origin verification. Results revealed significant isotopic discrepancies between field and market labelled samples, suggesting potential mislabeling of cultivar and origin. t-SNE demonstrated superior performance in resolving nonlinear patterns in isotope data compared to PCA. Among the different models, RF showed slightly better classification performance. Although the models achieved perfect classification accuracy on field-sourced samples, a considerable decline in performance was observed when applied to market-sourced samples, suggesting widespread origin mislabeling issues.