COMPARISON OF ROBUST MODELING TECHNIQUES ON NIR SPECTRA USED TO ESTIMATE GRAPE QUALITY

Multivariate calibration models are of critical importance for spectroscopic applications and so great efforts is placed in robust modelling. During the 2006 growing season, over 480 samples of wine grape berries Cabernet cultivar were scanned with a portable, hand-held NIR spectrometer and subsequently processed to determine °Brix using a digital refractometer. This study compares the performance of modelling techniques: partial least squares regression (PLSR) with or without spectral wavelength selection and external parameter ortogonalization (EPO) based on virtual standards. Models are performed on raw spectra, normalized spectra and spectra obtained after multiplicative scatter correction. PLSR with variable selection applied on normalized spectra was found as the best performing in terms of fitting (r²=0.72, RPD=1.64), with the minimum standard error of calibration and prediction on cross-validation 0.53 and 0.61 °Brix respectively.