On-site quantification of polysaccharides in Dendrobium huoshanense using a portable NIR spectrometer: A machine learning approach with SHAP interpretation

Dendrobium huoshanense , a precious edible-medicinal plant, faces a dearth of rapid, reliable quality assessment tools, hampering efficient resource utilization and market quality supervision. To address this gap, this study developed a portable, non-destructive, and interpretable method for quantifying D. huoshanense polysaccharides (DHP) using near-infrared spectroscopy (NIRS) coupled with machine learning. A systematic optimization workflow was implemented, incorporating outlier removal, MSC-second derivative preprocessing, UVE variable selection, and comparative evaluation of four machine learning models (BP-ANN, PLSR, RFR, 1D-CNN). The proposed MSC-2nd Derivative-UVE-RFR hybrid model delivered robust prediction performance: the test set yielded an average R² of 0.863±0.072, RMSE of 1.915±0.299, and RPD of 3.058±0.753, while the optimal fold in cross-validation achieved an exceptionally high test R² of 0.938. Notably, SHAP (SHapley Additive exPlanations) analysis was integrated to enhance model interpretability, identifying key polysaccharide-associated spectral bands and clarifying their quantitative contribution mechanisms. This study provides a rapid, eco-friendly, and interpretable solution for on-site quality control of D. huoshanense , effectively mitigating batch variation in practical applications. It further facilitates sustainable utilization of this medicinal resource, reinforces market quality supervision, and lays a technical foundation for the non-destructive testing of other plant-based food and medicinal products.