Milling-Free Prediction of Breadmaking Quality in Hard Winter Wheat Breeding Lines Using NIR Spectroscopy of Intact Kernels

Abstract

Wheat breeding programs generate thousands of lines annually, yet baking quality evaluation remains constrained by grain milling and large sample requirements, limiting early generation screening. This study evaluated near-infrared (NIR) spectroscopy combined with machine learning for predicting and classifying bread loaf volume from wheat flour and intact-kernel spectra. Models were evaluated across three independent data splits, with model selection performed exclusively within the calibration data. Among the regression approaches, artificial neural network regression (ANNR) showed the best overall performance, with held-out R2 Pred values of 0.61 ± 0.04 for flour and 0.60 ± 0.02 for kernels. Three-class partial least squares discriminant analysis (PLS-DA) showed moderate discrimination across the full loaf volume range, whereas classification of contrasting low- and high-loaf volume samples achieved accuracy of 0.881 ± 0.052 and 0.897 ± 0.020 for flour and kernels, respectively. Overall, intact-kernel NIR provided predictive and classification performance comparable to that of flour spectra without requiring milling, supporting its potential as a rapid, nondestructive screening tool for wheat breeding materials.