Simultaneous determination of five micro-components in Chrysanthemum morifolium (Hangbaiju) using near-infrared hyperspectral imaging coupled with deep learning with wavelength selection

Abstract

Chrysanthemum morifolium (Hangbaiju) is a kind of favored flower tea due to its health benefits. Rapid and accurate determination of chemical components is important for evaluating the quality of Hangbaiju. In this study, near-infrared hyperspectral imaging was used to explore the feasibility of determining the content of buddleoside, luteolin, apigenin, quercetin, and diosmetin in fresh and dry Hangbaiju. Partial least squares regression (PLSR), support vector regression (SVR) and convolutional neural network (CNN) were used to build regression models. Competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA) and a CNN based feature selection method (CNNFS) were used to select the optimal wavelengths. The prediction performances of luteolin and quercetin in fresh and dry Hangbaiju were good using both full spectra and optimal wavelengths, illustrating the feasibility of using near-infrared hyperspectral imaging to determine luteolin and quercetin in Hangbaiju. The relatively worse results of the other three components indicated that more efforts should be made to improve the prediction performances. CNN based regression and wavelength selection showed close results to the conventional regression and wavelength selection methods, indicating that CNN based regression and wavelength selection is promising to determine chemical components of Hangbaiju and other materials.