Deep learning for geographical discrimination of Panax notoginseng with directly near-infrared spectra image

Herbal materials have been widely used as functional food by a certain group of people for a potentially positive effect on body health regulation. Panax notoginseng as a crude material of functional food has long medical and cultivation history for more than 400 years in China and other countries. However, the quality was fluctuated with their geographical origins and Wenshan Autonomous Prefecture was regarded as the geo-authentic location with high properties. Therefore, rapid detection method is necessary for consumer to discriminate their authentic origins. In our study, 258 near infrared spectra of root powder of P. notoginseng from five main cultivation areas were used for discrimination analysis. A deep learning strategy (residual convolutional neural network) was established with 80% spectra images. Therein, the discrimination of geographical origins of the herb was first to be reported using directly spectra images instead of data matric from these spectra. The results indicated that these samples could be correctly classified as their respective categories with 100% accuracy in training set and 91% accuracy in test set. Finally, 22 samples were accurately discriminated in 25 samples of prediction set. In general, residual convolutional neural network using direct spectra image would be a feasible strategy for geographical traceability in further discrimination research.