Variable Selection Based Near Infrared Spectroscopic Quantitative Analysis on Wheat Crude Protein Content

Abstract:

Wheat is one of the main cereal grain which was produced not only in China but also at abroad. The aim of this paper is to study the feasibility of several selected variables from short wavelength near infrared spectroscopy to quantify crude protein of the whole wheat grain. In total, 52 whole wheat grain samples were collected, including 39 samples for calibration and 13 samples for validation. On the one hand, the crude protein of these samples were analyzed by using the Chinese standard of Kjeldahl method; on the other hand, those were scanned to obtain near infrared spectra with the wavelength range of 900~1700nm by using a wheat analysis system developed by Chinese Academy of Agricultural Mechanization Sciences (CAAMS). Both of spectroscopic pretreatment method and sensitive variables were optimized, then the model was built by partial least squares regression method. The results showed the combination of multiple scattering correction and wavelet transform performed better. Competitive adaptive reweighted sampling method showed an efficient variable selection, which picking 12 variables and taking 2% of the full range spectral variables, including 1028, 1158, 1199, 1367, 1407, 1445, 1478, 1494, 1550, 1584, 1661, 1686nm. Based on the optimized pretreatment method and selected variables, the model showed that the prediction determination coefficient and prediction root mean square error were 0.961 and 0.369, respectively. Competitive adaptive reweighted sampling variable selection based short wavelength near infrared spectroscopic technique showed a potential for crude protein quantitative analysis on whole wheat grain.