Benchtop Vis-NIR spectroscopy meets machine learning for multi-task analysis in Hongmeiren citrus: Geographical origin identification and antioxidant component quantification

Abstract

Due to geographical indication advantage of Hongmeiren (HMR) citrus, economically motivated origin fraud has emerged, alongside significant differences in antioxidant components. This study employed benchtop visible and near-infrared (Vis-NIR) spectroscopy combined with machine learning models for simultaneously identifying geographical origins and quantifying antioxidant components in HMR citrus. Savitzky-Golay and area normalization pre-processing exhibited the performance improvement in classification and regression models, respectively. Subsequently, Boruta algorithm feature selection further improved the both regression and classification performance of models. Among these, feedforward neural network model achieved the highest classification accuracy (88.7 %) for geographical origin, with few misclassified samples (1–2) in each origin and an excellent AUC (0.943). It also exhibited highest regression effects in quantifying ascorbic acid (R² = 0.875), total phenolics (R² = 0.856), and total flavonoids (R² = 0.806). Thus, benchtop Vis-NIR spectroscopy offers a practical and efficient multitask analysis method for simultaneous quality assessment and geographical authenticity identification in fruits.