We evaluated near-infrared hyperspectral imaging in the range of 900–1700 nm combined with machine learning for geographic origin authentication of Curcumae Radix collected from six major Chinese production regions. Each origin was represented by 100 specimens measured both as intact root slices and as finely milled powder. Three classification algorithms—support vector machine, random forest, and k-nearest neighbor—were compared after five common spectral preprocessing methods, using stratified train–test splits and nested cross-validation to ensure reliable generalization estimates. Model robustness was further assessed through per-origin holdout testing, permutation analysis, and the introduction of Gaussian noise and illumination variation, along with bootstrap-derived confidence intervals. Support vector machine consistently achieved the highest performance, with root slices reaching an average accuracy of 95.17% and powders achieving 99.00%. Powder spectra demonstrated not only higher discriminative power but also greater resilience to measurement noise and lighting changes compared with intact-root spectra. These results indicate that morphology-aware preprocessing combined with SVM enables rapid, non-destructive, and statistically validated provenance identification of Yujin, with powder-based assays offering the optimal balance between accuracy, robustness, and operational efficiency.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号