Simultaneous prediction of Panax ginseng origin and ginsenoside Re content using hyperspectral imaging and a multi-task one-dimensional convolutional neural network

Abstract

Origin tracing and saponin content determination are key factors in ginseng quality control, but traditional methods require independent modeling and encounter difficulties in small-sample scenarios. Statistical analysis revealed that ginsenoside Re content exhibits origin-specific associations. In this study, a multi-task one-dimensional convolutional neural network integrating channel attention modules was constructed, combined with hyperspectral technology and an uncertainty-driven dynamic weighting strategy, to achieve simultaneous origin classification and content prediction. The relationship between wavelength, origin, and saponins was analyzed through visualization of attention weights. The model achieved a classification accuracy of 92.59%, F1 score of 0.8969, precision of 0.9125, and recall of 0.8857. The regression task yielded an R² of 0.9074, RMSE of 0.0060, and RPD of 3.2864. Results demonstrate that compared to classical machine learning models such as random forest and support vector machine, as well as single-task 1DCNN, the proposed model exhibits efficient and accurate prediction of ginseng origin and ginsenoside Re content. High-weight channels correspond to wavelengths highly relevant to the prediction tasks. This method provides a novel approach for ginseng origin tracing and quality control, demonstrating the significant potential of multi-task prediction and establishing a foundation for the standardization and industrialization of traditional Chinese medicine.