Using near infrared reflectance spectroscopy to estimate the nutritive value of senescing annual ryegrass (Lolium rigidum): a comparison of calibration methods

Abstract

Summary. The suitability of near infrared reflectance (NIR) spectroscopy for predicting the concentration of several quality traits in samples of annual ryegrass (Lolium rigidum Gaud.) herbage was assessed in 2 separate experiments. In the first experiment, NIR calibration equations were developed for 6 traits (water-soluble carbohydrates, dry matter digestibility, neutral detergent solubles, neutral detergent solubles digestibility, neutral detergent fibre digestibility and nitrogen) using 4 calibration methods. No significant differences were found in the accuracy of NIR equations developed using either stepwise multiple linear regression (SMLR) or partial least squares regression (PLS) techniques when the equations were used to predict the concentration of constituents in those samples not used during the calibration process. The process of removing samples identified by the computer as spectral outliers was found to improve those statistics that related NIR data to the reference data of the samples used during calibration development (i.e. improved the goodness of fit of the regressions). However, when the resulting equations were used on all of the samples there was no improvement in the accuracy of the prediction of composition, and the estimates were less accurate for 2 of the equations. In the second experiment, plant part-specific equations (leaf blade, stem and leaf sheath) were developed. The specific equations were found to be no more accurate than those developed using a subset of all samples when used to analyse samples of the same plant part. However, using equations developed on either stem or leaf sheath samples to predict the composition of leaf blade samples led to inaccurate estimates of composition, illustrating the potential for error when NIR calibration equations are used on dissimilar samples. The similarity of the NIR estimates of decline in nutritive value and those obtained using reference analyses was illustrated by plotting the actual and predicted decline in nutritive value. The results of the experiments in this paper illustrate the need to monitor the accuracy of any NIR prediction of nutritive value. Striving for very low standard errors of calibration either by eliminating outliers or by limiting the plant tissues used during calibration did not lead to more accurate predictions of the composition of samples other than those used during the calibration process.