Comparative Analysis of Models for Identifying and Tracing Rice Flour Adulteration Using Raman Spectroscopy

To address the issue of rice flour adulteration, where lower-cost rice flour is mixed with higher-grade varieties to reduce costs, this work proposes a rapid identification method using Raman spectroscopy. Rice varieties from Heilongjiang Province were selected for adulteration experiments, in which Longqingdao 8 (LQD) was mixed with Sanjiang 6, Longyang 16, Suijing 18, Longdao 18, and Daohuaxiang 2 in varying proportions. Six machine learning models were employed for classification, with four different preprocessing methods. The models' performance was evaluated using the receiver operating characteristic (ROC) curves, and key characteristic bands for each rice variety were identified. For Sanjiang 6, the optimal preprocessing method was standard normal transformation, and the best-performing model was the artificial neural network, which achieved an area under the curve (AUC) of 96.2% and an accuracy of 94.8%. For Daohuaxiang 2, smoothing forest and random forest yielded the best results, with an AUC of 92.4% and an accuracy of 94.8%. Similarly, for Longdao 18, standard normal transformation and artificial neural network provided the highest accuracy (99.6%) with an AUC of 99.3%. Longyang 16 also showed optimal results with standard normal transformation and artificial neural network, achieving an AUC of 93.7% and an accuracy of 96.6%. Finally, for Suijing 18, multivariate scattering correction and random forest were the most effective, with an AUC of 99.3% and an accuracy of 99.6%. This comparative analysis of traceability models demonstrates a promising approach to identifying rice flour adulteration. The identification of compounds influencing different rice varieties further enhances the traceability of rice types, providing a robust reference for future studies.