Abstract
This research implemented a miniaturized near-infrared spectroscopy (NIRS) system integrated with machine learning approaches for the quantitative evaluation of dry gluten content (DGC), wet gluten content (WGC), and the gluten index (GI) in wheat flour in a noninvasive manner. Five different algorithms were employed to mine the relationship between the full-range spectra (900-1700 nm) and three parameters, with support vector regression (SVR) demonstrating the best prediction performance for all gluten parameters (R(P) = 0.9370-0.9430, RMSEP = 0.3450-0.4043%, and RPD = 3.1348-3.4998). Through a comparative evaluation of five wavelength selection techniques, 25-30 optimal wavelengths were identified, enabling the development of optimized SVR models. The improved whale optimization algorithm iWOA-based SVR (iWOA-SVR) model exhibited the strongest predictive capability among the five optimal wavelengths-based models, achieving comparable accuracy to the full-range spectra SVR for all gluten parameters (R(P) = 0.9190-0.9385, RMSEP = 0.3927-0.5743%, and RPD = 3.0424-3.2509). The model's robustness was confirmed through external validation and statistical analyses (p > 0.05 for F-test and t-test). The results highlight the effectiveness of micro-NIRS combined with iWOA-SVR for the nondestructive gluten quality assessment of wheat flour, providing a more valuable reference for expanding the use of NIRS technology and developing portable specialized NIRS equipment for industrial-level applications in the future.