Global or Local? A Choice for NIR Calibrations in Analyses of Forage Quality

Abstract

This paper investigates the effect of spectral data pre-treatment by using scatter correction techniques, detrending and derivatives on the standard error of NIR predictive models. It is shown that no particular spectral pre-treatment or no single derivative works best for the three constituents (protein, cellulose, organic matter digestibility) of the three forage databases which we investigated (grass-hay, tropical forages, maize whole plants). The best analytical results are obtained with SNVD, MSC or WMSC treatments. The best results are obtained with a first or second derivative with a segment and a gap of five data points. Local Regression was investigated for the prediction of forage quality. The standard errors of prediction were compared with those obtained with the best global calibration. Trial and error is the only way to fix the number of samples in the subset and the number of terms to retain in the model. Compared to the results for the traditional universal calibration method, the gain in SEP for protein, cellulose and digestibility in grass-hay, tropical forages or maize ranges between 5 and 11%.