A lightweight neural network approach for identifying geographical origins and predicting nutrient contents of dried wolfberries based on hyperspectral data

A rapid and non-destructive machine learning approach was developed to identify the geographical origin and predict nutrient contents of dried wolfberries by analyzing the hyperspectral images. 1826 images of dried wolfberry samples from four different geographical origins in China were collected and used as the database for this study. The raw spectral data was pre-processed and the characteristic wavelengths between 370 and 1043 nm was extracted to establish the datasets applicable for the analysis. The novel lightweight neural network model was developed based on group convolution and average pooling. By comparing the new approach with the other two existing neural network methods, it shows that the new model performs as well as or better than the existing algorithms. This study demonstrates the feasibility of identifying the geographic origin classification and nutrient prediction based on hyperspectral data of dried wolfberries. The non-invasive method provides a reference for future research of other similar products.