Beyond Prediction: Extracting Relevant Information from near Infrared Spectra

Abstract

With the increase of near infrared (NIR) applications, numerous chemometric methods have been developed. Among the mathematical treatments available, principal comoponent analysis (PCA) is certainly the most well-known when considering highly correlated data. In the field of near infrared spectroscopy, it allows the study of spectra without deleting wavelengths and without making any preliminary assumptions on the data. One advantage of PCA lies in the graphical displays obtained and, more precisely, on the similarity maps and spectral patterns. While the maps reveal clusters of the samples, the spectral patterns make a spectral interpretation possible. The present paper reviews our contribution to the development and application of PCA to NIR spectroscopy. It shows that PCA is the core of various mathematical treatments such as principal component regression (PCR), factorial discriminant analysis (FDA) and canonical correlation analysis (CCA). One advantage of using PCA in the prediction techniques lies in the use of all the wavelengths in the predictive model. The extraction of relevant and comprehensive wavelengths can be guided by CCA which allows the description of the samples by taking both mid- and near infrared data into account. Besides a comprehensive presentation of the mathematical treatements, examples are given.