Variable selection in near-infrared spectra: Application to quantitative non-destructive determination of protein content in wheat

Abstract

With the development of information technology, online non-destructive quantitative testing technology for wheat has attracted widespread attention. Protein content is an important indicator for wheat. A near infrared (NIR) spectroscopy system was utilized for the quantitative detection of protein in wheat. Firstly, the spectra of wheat were obtained by the NIR spectroscopy system. Then, variable selection methods were used to optimize the characteristic wavelengths, including competitive adaptive reweighted sampling (CARS), variable combination population analysis (VCPA), Monte Carlo variable combination population analysis (MCVCPA), and automatic weighting variable combination population analysis (AWVCPA). AWVCPA algorithm was effective in eliminating uninformative and inference wavelength variables of NIR spectra and reduced the complexity of quantitative prediction model. The characteristic wavelengths selected by AWVCPA were highly related to the absorbance of protein. Finally, the partial least squares (PLS) was employed to construct quantitative detection models based on different variable selection method. The experimental results showed that the AWVCPA-PLS prediction model is the best among the four tested models. The R2, RMSEP, RPD and RER values of AWVCPA-PLS were 0.9753, 0.0934, 10.15 and 38.54 respectively, indicative of high prediction accuracy and robust performance. This study provides an important theoretical basis for evaluating wheat in food engineering region.