Combined Maximum R2 and Partial Least Squares Method for Wavelengths Selection and Analysis of Spectroscopic Data

Abstract

The selection of wavelengths in multivariate analysis is of utmost importance in order to build a strong and robust predictive model. The aim of this research was to investigate the feasibility of an automated selection of sets of relevant wavelengths in Visible/Near Infra-Red (VIS/NIR) spectroscopy by combining Maximum R² (MAXR) method with Partial Least Squares (PLS) regression (MAXR-PLS) to build a PLS predictive model. The data used to test this method was derived from the determination of albumen pH and Haugh Unit (HU) as tools for testing the egg quality. For this purpose, 360 eggs were stored during 16 days under a temperature of 18^oC and a relative humidity of 55%. For each egg, the VIS/NIR transmission spectra and the two most widely used methods for the assessment of egg quality namely the HU and the albumen pH were performed. A PLS model was built using the full spectra and compared with the models built by selected wavelengths using MAXR-PLS method. Using the mentioned method, the correlation coefficients between the measured and predicted values were up to 95% and the Root Mean Square Error for Cross-validation (RMSECV) were 0.05 and 5.05 for pH and HU, respectively. In addition, this method reduces the complexity of the models by reducing the Latent Variables (LV). Despite the complexity of the spectral data, the Maximum R² method leads to a robust predictive model that uses the informative wavelengths.