A study on the predictive model for ginsenoside content in wild ginseng based on decision tree and ensemble learning algorithms

• A robust GA-DT model for ginsenoside prediction in wild ginseng achieved R 2 of 0.9701 and RMSE of 0.1183. • GA feature selection enhanced model performance by reducing noise and improving generalization in GA-DT. • Optimal preprocessing (LSNV 8 + SAVGOL 101) minimized baseline drift and noise, enhancing predictive accuracy. This study focuses on constructing and optimizing a predictive model for ginsenoside content in wild ginseng. Decision tree (DT) and ensemble learning algorithms (Random Forest, XGBoost, LightGBM) were compared and evaluated. Near-infrared hyperspectral data (900–1701 nm) served as the input for the models, while the output was the total ginsenoside content in wild ginseng, determined by high-performance liquid chromatography (HPLC). Initially, DT and random forest (RF) models were established using the full wavelength range (452 wavelengths), but both exhibited overfitting issues. To improve model performance, genetic algorithm (GA) was employed for feature wavelength selection, and the optimized variables were used for re-modeling. Further optimization of spectral data preprocessing methods, including Savitzky-Golay (SAVGOL) smoothing, multiplicative scatter correction (MSC), and local standard normal variate (LSNV) transformation, was conducted, leading to the selection of the best combination (LSNV 8 + SAVGOL 101), which was used as the final model input. By comparing the performance of GA-DT, GA-RF, GA-XGBoost, and GA-LightGBM, GA-DT was identified as the optimal model, achieving an R 2 score of 0.9701, RMSE of 0.1183, and relative error of 0.044. Finally, the topology of the optimal model was visualized, enhancing model interpretability. This method holds promise for practical applications in the quality control of traditional Chinese medicinal herbs, including cultivated ginseng, by providing a rapid, accurate, and non-destructive approach for ginsenoside content determination. Furthermore, it offers significant industrial potential in standardizing ginseng-related products and advancing the research of herbal medicine quality evaluation. The results indicate that this method offers a scientific model optimization strategy for rapid, non-destructive detection of ginsenoside content in wild ginseng.