Non-destructive detection of geographical adulteration in Chrysanthemum (“Gongju”) using hyperspectral imaging and interpretable ensemble learning under realistic conditions

The expansion of the tea beverage industry has increased demand for premium chrysanthemum raw materials. Gongju, a Geographical Indication product from She County, Anhui, China, is a primary ingredient for these beverages, but geographical origin adulteration in commercial circulation compromises its economic value and consumer interests. Existing hyperspectral studies on origin authentication mostly focus on pure sample classification, whereas robustness under practical adulteration scenarios and model interpretability remain insufficiently evaluated. In this study, hyperspectral imaging (HSI) was combined with an interpretable ensemble-learning framework for rapid and non-destructive detection of Gongju geographical adulteration. A stacking model integrating random forest, extreme gradient boosting, and multilayer perceptron was constructed, while competitive adaptive reweighted sampling (CARS) was used for wavelength selection and Shapley Additive Explanations (SHAP) provided auxiliary interpretation of wavelength contributions. The CARS-Stacking model retained 21–36 key wavelengths, reducing feature dimension by approximately 90% while maintaining test-set accuracies of 96.4%–99.2% across four origin-adulteration scenarios. External validation showed accuracies of 90.5%–100.0% under 20%–80% adulteration levels and three acquisition modes. Furthermore, SHAP highlighted wavelengths at 1483 and 1663 nm, which may be associated with O-H, C-H, and N-H absorption features. These results support HSI coupled with CARS-Stacking for Gongju geographical adulteration screening.