Cereal Chem. 70:29-35   |  VIEW ARTICLE

Classification of Hard Red Wheat by Near-Infrared Diffuse Reflectance Spectroscopy.

S. R. Delwiche and K. H. Norris. Copyright 1993 by the American Association of Cereal Chemists, Inc. 

Various forms of discriminant analysis models have been developed and tested for distinguishing two classes of wheat---hard red winter and hard red spring. Near-infrared diffuse reflectance (NIR) spectroscopy was used to measure the intrinsic properties of ground samples of hard red winter and spring wheats grown during the 1987, 1988, 1989, and 1990 crop years, of which 100 samples from each of the first three years formed the calibration set for each model. Discriminant functions were developed by using the following parameters: NIR-predicted protein content (adjusted to 12% moisture), NIR-predicted hardness, NIR protein and NIR hardness, and the scores from principal component analysis (PCA) of full-range (1,100- 2,498 nm) NIR spectra. Each function was tested on 1,325 samples (excluded from training of the models) from the 1987-1989 crop years and on 678 samples from the 1990 crop year, all of known class. Model performance, expressed as the percent of misclassified samples for each year and class, was poorest for the one-parameter models, which often had misclassification rates in excess of 25%. A five-factor PCA model was the most accurate, with an average misclassification rate of 5% for 1987, 1988, and 1989 samples. However, the misclassification rate of the PCA model rose to 8% for the 1990 samples, suggesting that model accuracy is reduced when samples grown during years excluded from calibration, such as from a new year's crop, are classified. Examination of the principal component factors indicates that hardness, protein level, and the interaction of water with protein and other constituents within wheat are responsible for correct NIR-based classification.