Partial least squares regression (PLS1) was used to obtain the multivariate calibration models using the Unscrambler 10.2 (Camo Software, Oslo, Norway). The data set was randomly split into two subsets: the calibration set consisting of 75% of the samples and the external validation set with the remaining 25% of samples. The external validation set may be used to determine the number of latent variables (LV), and is often cited as the most realistic estimate, particularly of the prediction errors. However, it requires a large amount of samples [16,17], such as in the present study. These models were developed with the spectra transformed by taking the Savitzky–Golay second (2D) derivative using a second-order polynomial, with a window of 15 and 25 points [18]. For the extractive model, the best results were obtained by combining the standard normal variate (SNV) with first (1D) derivative transformations using a second-order polynomial, with a window of 2 points [19].