DIFFERENTIATION OF HARD RED WHEAT BY NEAR-INFRARED ANALYSIS OF BULK SAMPLES

Near-infrared reflectance spectroscopy (1,100 - 2,498 nm) has been use d to identify hard red winter and hard red spring wheat classes. As a followup to a previous study which involved ground wheat samples, the authors have used the same samples on a whole kernel in-bulk (80 g) ba sis. Four years of U.S. winter and spring wheats were used. A small nu mber (n = 150 samples per class) from the first three years' samples w ere used for calibration; the remaining portion (n = 1,325), plus all of the fourth year's samples (n = 778), were used to verify the models . Four types of classification algorithms were examined: multiple line ar regression (MLR), principal component analysis with Mahalanobis dis tance (PCA/MD), partial least squares (PLS) analysis, and artificial n eural networks (ANN). All four models demonstrated classification accu racies (defined as the percentage of correctly classified samples) gre ater than 88%, and most often, about 95% for samples grown during the same years as used in calibration. These accuracies were significantly better than those associated with discriminant models that were based solely on protein content, NIR-hardness, or a combination of protein and hardness. Spectrally sensed water-matrix interactions were probabl y beneficial to model accuracy; however, moisture content alone was no t deemed necessary to a model's success. When predicting the fourth ye ar, the MLR model needed a bias correction, whereas the other three mo dels performed reasonably well. The ANN model's performance was highes t, with accuracies in the range of 95-98%. At little expense to model accuracy, the number of input nodes to the ANN model could be reduced from 223 to 111, provided the full wavelength range was preserved.