Comparison of CCA and PLS to explore and model NIR data

Partial Least Squares (PLS) regression is the most widely used technique for developing NIR calibrations. PLS uses several factors to reach the optimum models which can be helpful in a physical interpretation of the sources of correlation between x and y variables. However, it suffers from later factors not arising in the order of the explained variance. Canonical Correlation Analysis (CCA) overcomes this problem by selecting the latent variables as the directions of maximum x-y correlation. Calibration of moisture, crude protein, dry gluten and resistance of dough to deformation of wheat flour samples from NIR spectra is here studied using PLS-1, PLS-2, CCA-1 and CCA-2. The calibration set contains 429 samples while 215 extra independent samples are used for the validation set. It is shown that a 2-D CCA-2 calibration model gathers the highest explained variance between the models studied. When particular calibration models of each property are compared, CCA requires regularization to avoid instability of the regression coefficients. A regularization term that tends to reduce the regression coefficients and the Durbin-Watson test or the Test for Runs to select the regularization parameter have been used. Both statistical tests led to similar values of the regularization parameter and the resulting regression coefficients and RMSEP of the CCA-1 models became similar to those of the PLS-1 models.

Financial support from the Spanish Ministry of “Economia y Competitividad” (Project CTM2012-39183) and from the “Comissionat per a Universitats i Recerca del Departament d’Economia i Coneixement de la Generalitat de Catalunya” (2014 SGR 1132) is acknowledged.