Reconstruction of constituent spectra for individual samples through principal component analysis of near-infrared spectra
Abstract
Compound component weights for a specific constituent were combined with near-infrared spectra of individual samples to produce a form of spectral reconstruction which highlighted the influence of individual absorbance bands in the estimation of a fitted value for the constituent. The technique is illustrated using a set of wheat flour spectra with corresponding values for protein and moisture. For moisture, all samples exhibited variation at two points in the spectrum where absorbance bands for moisture are known. There was little variation between samples in the relative responses of these bands. For protein, however, individual samples exhibited as few as two and as many as six sources of variation which were used by the model to estimate the protein content of the sample. Not all these sources of variation related to known protein bands, indicating that the model was sensitive to the presence of other constituents such as starch. “Null” points, where adjacent absorbance effects were always in balance, were identified for both moisture and protein. Variation in particle size of samples was shown to distort reconstructed spectra. A simple algorithm using “null” points for protein was shown to reduce this distortion and enabled absorption effects to be more clearly observed.