*2.4. Model Building*

Models were developed for each FAs category: SFAs, MUFAs. PUFAs, ω-3 and ω-6. Therefore, the spectral data were considered as X, whereas the data obtained from the chromatographic analysis (the percentage of each FAs category for each model) were considered as Y.

X data were preprocessed before the multivariate analysis. Several methods were tested, such as standard normal variate without and with detrend (SNV and SNVd), multiplicative scatter correction (MSC), Saviztky–Golay first and second derivatives (with different polynomial orders and windows) and combinations of all of them. This step was necessary because it eliminates the irrelevant information that cannot be correctly processed [28], and it improves the regression [29]. X and Y data were mean centered in all cases before creation of the models.

To correlate the NIR spectra and the reference data (SFAs, MUFAs, PUFAs, ω-3 and ω-6 percentage), several partial least square regression (PLSR) models were developed [30]. For each developed model, two steps were followed:


As figures of merit of the models, the coefficient of determination (R2), the root mean square error (RMSE) and the bias value were calculated for the CV and the prediction. To study the distribution of the oil mixtures used in each dataset, their mean, standard deviation, minimum and maximum values were calculated and expressed in percentage.
