Near-infrared hyperspectral imaging for determination of protein content in barley samples using convolutional neural network

The protein content is an essential quality parameter that is measured by the breeding programs and food industries to maintain high-quality standards for barley. The traditional methods for determining protein content are destructive, time-consuming, and require the use of analytical reagents and chemical solvents. Near-infrared (NIR) spectroscopy is a rapid, non-destructive technology used to predict protein content by collecting spectral information. Alternatively, near-infrared hyperspectral imaging (NIR-HSI) integrates both spatial and spectral information of the samples. In this study, spatially resolved spectral information provided by the NIR-HSI was applied to predict the protein content of the barley samples. A total of 972 barley samples (889 hulled and 83 naked barley samples) were collected, and their reference protein values were measured using an elemental analyzer. The protein content ranged from 7.4 to 14.2%, with higher values for naked barley compared to hulled barley samples. The spatial information obtained from hyperspectral imaging system was used to extract the multiple mean spectra from each sample. The spectra were pre-treated with different spectral preprocessing techniques, which were then used as inputs of convolutional neural network (CNN) and conventional predictive models. The CNN model performed better on the unprocessed spectra (herein called raw spectra) than on the preprocessed spectral data. Additionally, the end-to-end CNN model trained using multiple mean spectra extracted from each sample outperformed the conventional models. The CNN model achieved a coefficient of determination (R2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{R}}^{2}$$\end{document}) of 0.9962, root mean square error (RMSE) of 0.0823, and residual prediction deviation (RPD) of 16.15. Finally, prediction maps were used to visualize the predicted protein content of the test barley samples. The overall results support the conclusion that the CNN model established using multiple mean spectra extracted from NIR hyperspectral images of barley samples can be used to accurately predict the protein content.