Application of Multidimensional Analyses to the Extraction of Discriminant Spectral Patterns from NIR Spectra

Abstract

A method for extracting NIR discriminant spectral patterns without any reference to chemical values is suggested. First, groups of samples have to be defined a priori. Second, this method involves two procedures: the application of Principal Component Analysis (PCA) on spectral data and Factorial Discriminant Analysis (FDA) on the PC scores. Discriminant spectral patterns are assessed as linear combinations of PCA eigenvectors with weights determined by FDA. This method was applied on an illustrative collection of wheat semolina conditioned at 3 levels of water concentration. Three groups were made on the basis of the 3 water levels. Two discriminant spectral patterns representative of water concentration were extracted. They showed the commonly known bands of water at 1940 and 1450 nm. The band at 1450 nm was resolved into two bands at 1410 and 1460 nm, which could be assigned to free and bound OH, respectively. The discriminant spectral patterns therefore modeled the relative proportion of free and bound water. Suitability of the method for other applications is discussed.