[Geographical origin discrimination of Magnoliae Officinalis Cortex based on hyperspectral imaging technology]

As a commonly used bark medicinal material in clinical practice, Magnoliae Officinalis Cortex(MOC) is widely distributed across different production areas, with noticeable quality differences among origins. However, it is difficult to determine its origin based solely on macroscopic characteristics. In this study, MOC samples(stem bark collected above 1 m from trees aged 15-25 years, shade-dried) from 26 production areas in 8 provinces were selected as research objects. Hyperspectral data of the outer surface, cross-section, and inner surface were collected, and six preprocessing algorithms were applied for spectral denoising. Partial least squares discriminant analysis, support vector machine, random forest, and extreme gradient boosting were then employed to construct provincial-scale origin discrimination models of MOC. Prediction accuracy was used as the evaluation index to screen the optimal model, and classification performance was assessed using confusion matrices. The results showed that, among models established with single-surface hyperspectral data, the optimal model was built with full-band inner-surface data preprocessed by first derivative and modeled with random forest. Among dual-surface combinations, the optimal model was established with full-band "inner surface + cross-section" data preprocessed by first derivative and modeled with random forest. The prediction accuracies of these two models were comparable, 95.68%(inner surface) and 95.99%(inner surface + cross-section), slightly lower than that of the three-surface combination model( "inner surface + cross-section + outer surface", 97.22%). To eliminate redundant hyperspectral information and improve modeling efficiency, competitive adaptive reweighted sampling, successive projection algorithm, and uninformative variable elimination were applied to extract characteristic wavelengths from the inner surface dataset and the three-surface combination dataset. The results indicated that the prediction accuracy of the inner-surface characteristic wavelength model was 94.14%, slightly lower than that of the full-band model, while that of the three-surface combination characteristic wavelength model reached 97.53%, comparable to the full-band model. This study provides a rapid and effective discrimination model for determining the origin of MOC, as well as experimental evidence for constructing models based on characteristic wavelengths.