Enhancing near-infrared spectroscopy calibration for accurate protein and gluten determination in wheat flour and intact grains using chemometric techniques

Rapid and reliable determination of protein and gluten in wheat is crucial for quality assessment and process control. Near-infrared (NIR) spectroscopy provides a nondestructive alternative to conventional chemical analysis; however, its predictive performance depends strongly on preprocessing and modeling strategy. This study evaluated how different combinations of scatter correction, derivative, and wavelength-selection methods influence NIR calibration performance for predicting protein and gluten contents in both wheat flour and intact grain samples, using Partial Least Squares (PLS) and Support Vector Machine (SVM) regression models under identical conditions. The results demonstrated that SVM achieved superior prediction accuracy for both protein and gluten contents, particularly when combined with Standard Normal Variate (SNV) preprocessing and mild smoothing. Among the best-performing models, those developed from flour-based spectra generally achieved higher coefficients of determination (R2, Coefficient of Determination, up to 0.96) than those based on grain spectra (R2 approximate to 0.88-0.90), reflecting reduced scattering and greater compositional uniformity in flour samples. The most successful combinations were SNV + SVM for protein prediction (R2 = 0.99) and smoothing + SNV + Genetic Algorithm-Partial Least Squares (GA-PLS) + SVM for gluten prediction (R2 = 0.93). Overall results revealed that combining NIR spectroscopy with optimized preprocessing and machine-learning algorithms enables rapid and precise quantification of wheat quality traits, supporting its broader application in industrial quality control and breeding programs.