Near-Infrared Spectroscopy and Machine Learning for Geographic-Origin Screening of Dendrobium crepidatum Lindl. et Paxt.

Dendrobium crepidatum Lindl. et Paxt. is a medicinal Dendrobium species whose quality and market value may vary with geographic origin, making rapid origin traceability important for batch management, market supervision, and application promotion. This study used near-infrared spectroscopy (NIRS) combined with multivariate analysis and machine learning to discriminate the origin of D. crepidatum. Fifty batches of stem samples from Yunnan, Guangxi, and Guizhou, China, were analyzed after Savitzky-Golay smoothing, standard normal variate transformation, and first-derivative preprocessing. Principal component analysis (PCA) showed origin-related spectral variation, and a three-class partial least squares-discriminant analysis (PLS-DA) model achieved a mean cross-validated accuracy of 70.2% with a significant permutation-test result (p = 0.0020). Six machine learning algorithms, including KNN, CART, RF, NB, LDA, and ANN, were further compared using repeated nested cross-validation. KNN performed best, with an accuracy of 0.811 ± 0.029 and a macro F1-score of 0.813 ± 0.029, followed by RF (0.804 ± 0.038 and 0.805 ± 0.037, respectively). Key spectral variables were mainly located at 4231–4235 and 5523–5624 cm−1 corresponding mainly to C-H-dominated overtone or combination absorptions with possible C-O/O-H-related contributions from carbohydrates, polysaccharides, phenolics, flavonoids, and other organic constituents. These results demonstrate the feasibility of NIRS combined with machine learning for preliminary origin traceability of D. crepidatum and provide spectral clues for future investigation of origin-related chemical variation and quality discrimination.