Explainable AI-driven hyperspectral imaging for non-destructive geographical discrimination of small-fruited sea buckthorn: Metabolic correlations and feature optimization

Small-fruited sea buckthorn exhibits substantial compositional variation across different geographical origins, necessitating accurate classification methods. This study established a reproducible hyperspectral imaging (HSI)-driven pipeline, in which deep learning enabled non-destructive geographical classification, and untargeted metabolomics together with SHapley Additive exPlanations (SHAP) were integrated to elucidate the metabolic basis and key spectral features underlying the classification. The convolutional neural network (CNN) trained on pre-processed HSI spectra achieved robust discrimination of samples from five geographical origins, with a validation accuracy of 93.3 %, outperforming benchmark models. The dataset comprised 90 sea buckthorn fruit samples from each of five geographical origins, with one ROI spectrum per fruit, obtained from a 700 × 700 pixel region. SHAP was applied to quantify wavelength contributions and identify globally important features, enabling a 45.9 % reduction in wavelengths without loss of predictive performance. In parallel, untargeted metabolomics analysis using UPLC-QTOF-MS under positive and negative ion modes identified origin-related differential metabolites. Correlation analysis between SHAP-selected wavelengths and metabolite abundances established spectral–metabolic associations linking characteristic wavelengths to major classes of discriminative metabolites, providing mechanistic support for the observed hyperspectral signatures and geographical discrimination of sea buckthorn. This integrated strategy supports rapid, non-destructive authentication and traceability of sea buckthorn.