Portable VIS–NIR spectroscopy integrated with machine learning for robust authentication and geographical origin classification of rice

This is the authors' abstract. We don't add key points for this paper.

In this study, a handheld spectroscopic–Machine Learning (ML) framework was developed to discriminate the geographical origin of three Iranian rice varieties (Hashemi, Khazar, and Shiroodi) at two spatial scales: provincial (Mazandaran vs. Gilan) and urban (Sari, Amol, and Rasht). Vis-NIR spectral data were collected under two illumination sources (LED and Bulb), and multiple classification models—including Ridge, Logistic Regression, Multilayer Perceptron (MLP), and Support Vector Machine (SVM)—were evaluated in combination with feature selection methods (LARS and LARS-Lasso) and two data-splitting strategies (Stratified and Kennard-Stone). At the provincial scale, Hashemi and Khazar achieved perfect test accuracy (100%) in their optimal configurations. Shiroodi exhibited comparatively lower provincial separability, with test accuracy below 90%. In the urban scale (three-class problem), overall variability increased, and performance became more sensitive to model configuration and illumination source. Nevertheless, optimized models achieved up to 100% test accuracy for Khazar and above 94% for Shiroodi, while Hashemi reached a maximum of 88.9%. Confusion matrix analysis revealed that most misclassifications occurred between Sari and Amol (within Mazandaran Province), whereas Rasht samples were generally classified with higher stability. Overall, the results demonstrate that spectroscopic signatures, combined with appropriate ML strategies, enable highly accurate classification of rice's geographical origin among the regions.