|Wesley, Ian - BRI, AUSTRALIA,LTD.|
|Osborne, Brian - BRI, AUSTRALIA,LTD.|
|Anderssen, Robert - CSIRO, AUSTRALIA|
Submitted to: Cereal Chemistry
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: December 1, 2002
Publication Date: June 1, 2003
Citation: WESLEY, I.J., OSBORNE, B.O., ANDERSSEN, R.S., DELWICHE, S.R., GRAYBOSCH, R.A. CHEMOMETRIC LOCALIZATION APPROACH TO NIR MEASUREMENT OF APPARENT AMYLOSE CONTENT OF GROUND WHEAT. CEREAL CHEMISTRY. 2003. Interpretive Summary: In recent years, work has been underway in wheat breeding programs to introduce, by conventional breeding practices, new lines of wheat that are waxy. The term, "waxy," is used to describe starchy substances that are extremely low in concentration of a macromolecule that is normally abundant in starch. Waxy varieties, which are high in amylopectin, but very low in amylose, have unique processing characteristics when combined with other ingredients in the formation of breads, crackers, and noodles. Further, waxy varieties have the potential for use in niche products in industrial uses of wheat starch. One of the challenges for waxy wheat breeding programs is to find a reliable analytical method that can rapidly screen thousands of samples for the waxy trait. Our previous research efforts have demonstrated the feasibility of using near infrared (NIR) spectroscopy to classify, as well as quantify the waxy condition in wheat. The present study has reexamined the mathematical basis of NIR regression models that predict wheat amylose content. An additional mathematical technique was applied and compared to the conventional structure of an NIR-based multivariate regression model. The newly applied technique, called locally weighted regression, attempts to tailor-make a regression equation to each 'unknown' sample by using samples of known constitution that are spectrally similar to the unknown. Results of this study have demonstrated some advantages of locally weighted regression, especially on sample populations whose distributions of the constituent under investigation (amylose content in the present study) are bimodal. This study on multivariate regression refinement for NIR analysis of agricultural products is for the benefit of the statistician or mathematical model developer, who, at the request of the plant breeder is seeking to develop rapid early-generation screening techniques for crop quality improvement.
Technical Abstract: The development of new wheat varieties that target specific end-uses, such as low or zero amylose contents of partially-waxy and waxy wheats, has become a modern focus of wheat breeding. But, for efficient and cost-effective breeding, inexpensive and high-throughput quality testing procedures, such as near infrared (NIR) spectroscopy, are required. The genetic nature of a set of wheat lines, which included waxy to non-waxy varieties, results in a bimodal distribution of amylose contents that presents some special challenges for the formulation of stable NIR calibrations for this property. The obvious and intuitive solution lies in the use of some form of localization procedure and this sample set provided the opportunity to explore three localization algorithms in comparison with the default partial least squares. Localization with respect to the waxy (zero amylose) varieties resulted in a MPLS calibration with a standard error of prediction of 0.16%. The key problem with the measurement of amylose is the laboratory reference method, which in reality only measures the apparent amylose content of the wheat. As a direct consequence, the measurements of amylose have such a large error that traditional calibration-and-prediction procedures generate unacceptable results. In order to resolve this difficulty, a statistically-based resampling strategy is proposed as a method of identifying samples where there is a large error in the reference measurement. The results presented in this paper establish unambiguously that, after the application of an appropriate resampling strategy, there are advantages in performing a suitable localization in order to achieve a reliable NIR calibration-and-prediction. It resolves the issue of how to utilize the bimodal distribution of apparent amylose values. The best results are obtained when the localization is performed simultaneously with respect to the sample property under investigation and the NIR spectra.