Clustering of variables to analyze spectral data

A cluster analysis of variables around latent variables is presented and applied in order to identify groups among near‐infrared (NIR) spectral variables. By organizing multivariate data into a small number of clusters, each of them being represented by a component, this approach makes it possible to reduce the dimensionality of the problem. For the NIR data considered herein, it turned out that the groups of spectral variables are associated with various spectral regions. This feature can be helpful for the interpretation of the outcomes. For a predictive perspective the groups of variables can be used as blocks in multiblock partial least squares models. Alternatively the latent variables associated with the various clusters can be used as predictors. The cluster analysis procedure together with how its outcomes can be used for prediction purposes are illustrated on the basis of sensory and NIR data on green peas. Copyright © 2005 John Wiley & Sons, Ltd.