Constructing Rapid and Undamaged Detection Models for Main Physicochemical Indexes of Brewing Sorghum Based on Near Infrared Spectrum

To satisfy the demands of rapid determination of amylose, amylopectin, protein, fat, and tannin contents in brewing sorghums, in this paper, 17 spectral data preprocessing methods and 4 wavelength band selection algorithms were used to establish the near infrared spectral analysis models for these indexes. The results showed that the best spectral preprocessing methods for each index were 1st der (1st)+multiplicative scatter correction (MSC)+Z-score standardization (ZS), vector normalization (VN)+mean centering (MC), standard normal variate transformation (SNV)+ZS, MSC, SNV+ZS, respectively. The best wavelength band selection algorithm for predicting amylose, amylopectin, protein, and tannin contents was monte-carlo uninformative variable elimination, and that of fat was competitive adaptive reweighted sampling. The R² in the optimal models for these 5 indexes of whole grain sorghums were 0.9560, 0.8765, 0.9069, 0.8658, 0.8841, and the RMSECV values were 1.3222, 2.3477, 0.3549, 0.2164, 0.1077, respectively. The validation results of external independent samples showed that the models had a high prediction accuracy. The NIR analysis model established in this study could provide a technical reference for the rapid detection of sorghums in the brewing industry.