Hyperspectral imaging and SHAP-assisted machine learning for differentiating Rubia species and geographical-origin groups

Introduction Rubia cordifolia L. is widely used in traditional Chinese medicine, whereas closely related Rubia species are difficult to distinguish visually and authentic material varies across producing regions. Methods This study combined hyperspectral imaging with machine learning for authenticity discrimination among three Rubia species and geographical-origin classification of R. cordifolia from six regions. Conventional machine-learning and deep-learning models were compared using region-of-interest (ROI) mean spectra. Training and test data were separated by collection batch, and evaluation was repeated across ten random partitions. Results Support vector machines coupled with standard normal variate transformation (SNV-SVM) achieved the highest mean ROI-level test accuracies of 97.7% and 97.0% for species and origin classification, respectively. Successive projections algorithm, competitive adaptive reweighted sampling, and Shapley additive explanations (SHAP) were compared for wavelength reduction. Sixty wavelengths ranked using training-derived SHAP values reduced the input by 84.8% while retaining test accuracies of 94.0% and 89.7%, respectively. SHAP identified 751.72 and 979.26 nm as the most influential wavelengths. Discussion These regions may reflect overlapping X-H absorptions, water-related responses, and tissue optical effects. The results demonstrate spectral differentiation of the collected materials using batch-separated evaluation. Additional batches, external validation, and reference chemical measurements are needed to assess transferability and test the proposed spectral associations. These findings provide candidate wavelengths for further development of rapid, non-destructive screening approaches for Rubia -derived medicinal materials.