Skip to main content
ARS Home » Research » Publications at this Location » Publication #94519
Title: THE DEVELOPMENT OF NEAR INFRARED WHEAT QUALITY MODELS BY LOCALLY WEIGHTED REGRESSIONS

Author
item BARTON II, FRANKLIN
item SHENK, J.S. - PENNSYLVANIA STATE UNIV
item WESTERHAUS, M.O. - PENNSYLVANIA STATE UNIV
item FUNK, D.B. - USDA-GRAIN INSPECTION

Submitted to: Near Infrared Spectroscopy Journal
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/21/2000
Publication Date: N/A
Citation: N/A

Interpretive Summary: The number of calibrations that must be maintained for a regulatory agency to determine protein in wheat is quite large. This requires a big effort in time and personnel to keep all instruments properly calibrated. The technique of Locally Weighted Regressions replaces all the chemometric models with one standard library from which an individual model for each sample is determined. This procedure can help in two ways. First, the workload will be greatly diminished for the regulatory agency. What required a team of scientists and technicians for extended periods of time will permit those assets to be used on other projects. Second, the method will help to integrate the analysis across national borders and permit all nations to grade and market wheat from the same perspective. Currently, we are conducting tests of these procedures with the European Union, Canada, and Australia.

Technical Abstract: Large near infrared (NIR) data sets are necessary for complex products or those of diverse origins. Such is the case for wheat in the international marketplace. Global calibrations have always been less precise than those for individual instruments or for more regionalized sample collections. Three possibilities exist to cope with broad populations. First, one can accept the larger errors in precision and/or develop specific calibrations for classes of commodities within certain areas. Second, neural networks can be employed to develop calibrations and take into account the nonlinearities inherent in very large data sets. Third, Locally Weighted Regressions (LWR) can be used to develop single sample calibrations which handle nonlinearity by producing a "local" linear model in reduced space and potentially are more easily interpretable than neural networks. Results will be reported which show LWR to be as good as specific calibrations by PLS for sub-classes of products and precise enough for use for regulatory purposes. Results were obtained from a set of wheats comprising five different classes and a total of 2207 samples. The use of LWR over normal learning sets has other advantages, such as easier calibration update, easier transferability, and the possibility to include authentication and classification as part of the model.