Most of the observed variation between n.i.r. spectra in any sample collection is due to particle size variation; in addition, absorbance readings are highly inter-correlated. Application of multiple linear regression techniques to such spectral data produces a multiplicity of solutions which vary in the wavelengths selected and in predictive accuracy. Evaluation of all or most of these calibrations requires a significant amount of time. In the present study, alternative data treatment methods are described which reduce the effect of particle size variations, select only those wavelengths which contain significant information, and overcome the problem of inter-correlation by means of principal component analysis. These methods were applied to the analysis of dried silage for crude protein (CP) and in vitro dry matter digestibility (IVDMD); accuracy of the novel data treatments (best standard error of prediction of CP and IVDMD equal to 0.63 and 2.7, respectively) was better than that generally achieved by more conventional statistical techniques.
