Hierarchical principal component analysis (PCA) and projection to latent structure (PLS) technique on spectroscopic data as a data pretreatment for calibration

Spectroscopic data consists of several hundred to some thousand variables, wherein most of the variables are: autocorrelated. When PCA and PLS techniques are used for the interpretation of these kinds of data, the loading plots are usually complex due to the covariation in the spectrum, and therefore difficult to correlate to the corresponding score plot. One of the standard methods used to decrease the influence of light scatter or shifts of the spectra is the multiplicative scatter correction technique. Another technique is the hierarchical multiblock segmentation technique, where new variables are created from the original data by blocking the spectra into sub spectra, and then projecting the sub spectra by PCA. These new variables are then used in the coming PCA or PLS calculations. These techniques reduce the random and non-wanted signals from e.g. light scatter, but still conserve all systematic information in the signals, but the greatest advantage is that the technique gives an easier interpretation of the correlation between scores and the loadings. Two examples are presented: the attenuated total reflection (ATR) and NIR, which show the advantages as well as the implementation of the method. Copyright (C) 2001 John Wiley & Sons, Ltd.