Non-destructive origin tracing and ginsenoside quantification in American ginseng using hyperspectral imaging and mixed multi-task 1DCNN

Abstract

Accurate origin tracing and ginsenoside quantification in American ginseng (Panax quinquefolius L.), are critical for its market value and therapeutic efficacy. Empirical identification and chemical analysis methods are often destructive and inefficient. This study addresses these limitations by integrating hyperspectral imaging (HSI) with a novel mixed multi-task one-dimensional convolutional neural network (MMT1DCNN) for the non-destructive and rapid evaluation of ginsenoside content and origin of American ginseng. Adaptive threshold segmentation and spectra extraction are conducted on the raw HSI images to obtain the spectral data of each American ginseng. Four common pretreatment methods are employed to clean and normalize the spectra. The successive projections algorithm is applied to select key bands from spectra, which are then fed into the MMT1DCNN to simultaneously predict the origin and ginsenoside content. A dataset containing spectral, origin, and ginsenoside data of American ginseng is constructed to validate the method's performance. Experimental results demonstrate that the proposed method achieves 95 % overall accuracy for origin classification, a correlation coefficient of 0.965, a root mean square error of 3.64 mg/g, and a residual predictive deviation of 3.54 for ginsenoside content. These findings suggest that the developed method has significant potential for non-destructive, accurate origin tracing and quantitative evaluation of American ginseng.