Non-destructive determination of ginsenosides in ginseng by combined hyperspectral and X-ray techniques based on ensemble learning.

The potential of hyperspectral imaging and X-ray techniques for the non-destructive determination of the ginsenosides Rg1 + Re and Rb1 in ginseng was investigated. The random forest (RF) models were established using spectral information extracted from hyperspectral data to predict ginsenosides content. The RF model was optimized by data pre-processing methods and feature screening methods. Multiple feature screening methods combined with partial least squares regression models were used to find hyperspectral image feature information (color information and texture information) related to ginsenosides. A significant positive correlation between density extracted from X-ray images and the ginsenosides content was found by building the univariate linear regression models. Finally, the prediction performance of the integrated learning model based on the three data blocks was better than the model constructed by single data blocks (Rg1 + Re: R = 0.8691, RMSE = 0.0439%; Rb1: R = 0.8291, RMSE = 0.0803%). The results indicate that the developed method is highly feasible for non-destructive evaluation of ginseng quality.