CLASSIFICATION OF WHEAT BY VISIBLE AND NEAR-INFRARED REFLECTANCE FROMSINGLE KERNELS

Identification of wheat class is a necessary component of the official inspection of U.S. wheat, owing to differences in functionality and h ence in trade value. Because of the numerous cultivars for the several U.S. wheat classes, segregation by cultivar is generally impractical during postharvest handling. Cultivars of differing wheat classes are sometimes inadvertently mixed, resulting in classification of the lot to a mixed category, thus lowering its value. Single-kernel near-infra red reflectance scans from two spectral regions (551-750 nm for distin ctions based on color, 1,120-2,476 nm for distinctions based on intrin sic properties) were collected on 10 randomly drawn kernels from each of 318 unique samples obtained from commercial sources. Partial least squares and multiple linear regression analyses were used to develop b inary decision models for various combinations of two wheat classes, c hoosing from five classes: hard white (HWH), hard red spring (HRS), ha rd red winter (HRW), soft red winter (SRW), and soft white (SWH). Two- class model accuracy, defined as the proportion of correctly identifie d kernels of a known wheat class, was greatest (99%) when red and whit e classes such as HRW vs. HWH were compared. Accuracies declined to ty pically 78-91% when the two classes were of similar color (e.g., HRW v s. SRW, HWH vs. SWH). Using a cascade of binary comparisons similar to two-class models, a five-class model structure was developed. Five-cl ass model accuracy ranged from 65% for SRW wheat to 92% for SWH.