Statistical analysis of near infra-red reflectance data
Near Infra-red (NIR) reflectance spectroscopy is an instrumental technique to analyse the chemical composition (eg. nitrogen content) of organic materials. As an approach it is rapid, accurate and hence cost effective. Composition is currently determined by calibration equations which relate traditional wet chemical measurements to NIR spectral measurements for the same sample. This thesis examines statistical methods of estimating composition from the NIR spectra and suggests new methods. The relative merits of each of the methods is described. Variation within the spectra is also affected by physical attributes like particle size. Several transformations are examined for their ability to reduce non-chemical differences. These include some transformations previously found to be useful and also some new approaches. Methods of calibration are then investigated. Of the 'standard* methods, stepwise multiple linear regression, principal component regression, latent root regression and partial least squares are discussed. In addition, some new methods are considered. Firstly, three new calibration models are derived which, like some of the 'standard' methods, use information from the whole spectra. Next is a slightly different approach, whereby only the information from informative areas of the spectra, called 'windows', are used for calibration. Finally, hierarchical models for combining information from different sample sets in a flexible way are considered and adapted for NIR data types.