Abstract
Geographical authentication of Meconopsis integrifolia is important for the quality control and sustainable utilization of plateau medicinal plant resources, but adjacent origins often exhibit highly similar spectral and chemical profiles. In this study, a shared-private multimodal fusion (SPMF) framework was developed for origin discrimination of M. integrifolia using paired near-infrared (NIR) diffuse-reflectance spectra and ultraviolet-visible (UV-Vis) absorption spectra of the NaNO2-Al(NO3)3-NaOH chromogenic system. After spectral quality control and Isolation Forest outlier removal, sixteen preprocessing strategies combined with four classifiers were evaluated on the training set to determine the optimal NIR and UV-Vis modelling pipelines, which were then fixed for subsequent fusion experiments. Conventional low-level, mid-level, and high-level fusion strategies were compared with the proposed SPMF model under the same stratified data split. Internal five-fold cross-validation on the training set was used for model selection and stability evaluation, while an independent test set was reserved for final assessment. The SPMF model decomposed paired NIR and UV-Vis chromogenic spectra into shared latent representations and modality-specific private representations, constrained by similarity, difference and reconstruction losses to enhance cross-modal consistency, reduce redundant information and preserve modality-specific spectral characteristics. On the independent test set, SPMF correctly classified 40 of 41 samples, corresponding to an accuracy of 97.6%, with a 95% Wilson confidence interval of 87.4-99.6%. Its macro-precision, macro-Recall, macro-averaged F1 score (macro-F1), and Cohen's kappa were 0.98, 0.98, 0.98, and 0.97, respectively, and the multiclass Brier score was 0.073. The training-set five-fold cross-validation accuracy was 0.981 ± 0.004. Compared with the best traditional fusion baseline, SPMF yielded higher point estimates on the current independent test set. Under the current experimental realization, the complete SPMF model matched or exceeded the performance of the reduced ablation variants. However, the component-specific effects remain preliminary, and because all samples were collected in 2019 and no external validation across collection years, batches, independent collection campaigns or analytical instruments was performed, the reported performance should be interpreted as discrimination within the current dataset rather than as established real-world generalizability.