Comparison of multivariate calibration methods for prediction of feeding value by near infrared reflectance spectroscopy.

Abstract

The spectrum of absorbance of near infrared spectroscopy (NIR) measurements contains indirect, non-specific information about the feeding value of the feed sample and can be used to predict this value. A linear calibration model was estimated from experimental data and a model was used to predict unknown in vitro values with measured spectra in future samples. Multicolinearity in the NIR measurements occurs frequently. Inclusion of all absorbances in a calibration model leads to complications and large prediction errors. To overcome multicollinearity several methods were proposed. The methods are described and applied to data in which in vitro digestibility of organic matter of maize for cattle was predicted by means of absorbances measured at 351 wavelengths. Comparison of methods showed that for these data Partial Least Squares was the best method. Multiplicative scatter correction of the spectra prior to estimation gave better predictions for all methods. (Abstract retrieved from CAB Abstracts by CABI’s permission)