Application of Principal Component Analysis on NIR Spectral Collection after Elimination of Interference by a Least-Squares Procedure

Abstract

In NIR spectroscopy, multidimensional analyses such as Principal Component Analysis (PCA) may be applied to examine the similarity between spectra of natural products. However, such an approach is often limited by the effect of spectral interference due to water or particle size distribution of the samples. In the present work, the advantage of the elimination of such spectral interference before performing PCA was investigated. Unwanted component spectra were eliminated by a least-squares procedure. They were first orthogonalized and normalized by the Gram-Schmidt orthogonalization method. The subtraction coefficients were then assessed, similarly to principal component (PC) scores, by projection of the NIR spectra on the orthogonalized component spectra, and PCA was performed on the corrected spectra. This method was applied on an illustrative collection of wheat semolina conditioned in three levels of water content. Water was the component to be eliminated and had been previously modeled by two spectral patterns. These spectral patterns were used as the unwanted component spectra. PCA was applied independently before and after spectral correction of the collection of spectra and graphs obtained by the two procedures were compared. The squared correlation coefficient of the 3 first PC scores with water content was 0.979 before correction, with the 3 groups of water content appearing clearly on PCA graphs. After correction, the corresponding squared correlation coefficient for the 7 first PC scores was 0.016. PCA graphs obtained with corrected spectra also showed that the water effect was completely eliminated. At this moment, samples were separated according to their technological nature. The procedure developed may be useful in pattern recognition study and for automatic clustering of NIR spectra. It may also be applied in fields other than NIR spectroscopy.