Hyperspectral geographical origin identification of Bupleurum via spatial-spectral features based on adaptive weighted quaternion zernike moments.

Bupleurum, a vital herb in traditional Chinese medicine (TCM), possesses complex chemical compositions heavily dependent on variety and geographical origin. Misidentification often leads to unstable therapeutic efficacy and safety risks. Addressing challenges in market confusion and traceability, this study proposes a rapid, non-destructive identification framework using visible-near infrared (Vis-NIR) hyperspectral imaging. A comprehensive dataset comprising 3670 samples from 17 batches, 7 origins, and 2 varieties was established, with hyperspectral images acquired separately from the cross-sections and lateral surfaces. ROI-average spectra were extracted from both parts, while ROI spatial information was retained for AWQZM texture feature extraction. To utilize complementary information from different parts and preprocessing methods, a four-component spectral feature set was constructed from cross-section and lateral-surface preprocessed average spectra and first-derivative spectra. Adaptive Weighted Quaternion Zernike Moments (AWQZM) were introduced to extract spatial texture features from ROI images. A unified dimensionality reduction strategy and a 95% variance adaptive strategy were compared, and SVM, RF, and KNN models were optimized using Tree-structured Parzen Estimator (TPE) Bayesian optimization. The SVM model fusing SG-smoothed spectra, first-derivative spectra, and AWQZM texture features achieved the best performance, with a testing accuracy of 98.23%. These results demonstrate that fusing dual-part spectral information with spatial texture features improves Bupleurum variety and geographical origin identification, providing a feasible approach for rapid, non-destructive TCM quality traceability.