74 Comparison of calibrations based on partial least squares (PLS), and multiple linear regression (MLR) for near-infrared prediction of composition and functionality in grains.

Phil Williams. Canadian Grain Commission, Grain Research Laboratory, 1404-303, Main Street, Winnipeg, Manitoba, Canada R3C 3G8.

Modern computers and software enable very rapid development and evaluation of calibrations for prediction of composition and functionality in cereal technology. Data are presented to illustrate differences in the pattern of results obtained by the two methods, having been applied to a large number of calibrations for both types of application. Excellent results could be obtained by either method, but neither method was consistently superior to the other in accuracy or precision. While the MLR method selects discrete wavelength points about which calibrations are developed, the PLS approach is based on all wavelengths scanned. The “weights” developed during development of PLS calibrations indicate areas of the spectrum which have contributed positively or negatively, and to a greater or lesser extent than others. These “peaks” of variance often coincided with wavelengths selected by MLR. The implications of this observation are discussed.