Geographical origin discrimination of Fructus Aurantii by visible near-infrared imaging with machine learning and explainable analysis

Fructus Aurantii (FA), the dried processed product of Citrus aurantium L. , is widely used in food and functional food applications, and its quality is notably affected by geographical origin. In this study, a visible near-infrared (Vis/NIR) imaging platform was used to acquire reflectance images of FA samples from four production regions. An optimized and interpretable modeling framework was developed for FA origin discrimination by integrating spectral preprocessing, dimensionality reduction, machine learning classification, reinforcement-learning-inspired band selection, SOA-based XGBoost optimization, and quantitative SHAP analysis. The SOA-XGB model achieved an accuracy of 0.9213 ± 0.0148 in the four-class origin discrimination task and maintained stable performance on the external validation set. The band selection results showed that a limited number of informative wavelengths could achieve performance close to that of the full-band model. Quantitative SHAP analysis further indicated that wavelengths mainly distributed around 570–730 nm and 850–980 nm contributed notably to model decisions. These results demonstrate that Vis/NIR imaging combined with wavelength optimization, parameter optimization, and quantitative interpretability analysis provides a rapid, non-destructive, and reliable approach for FA geographical origin discrimination.