**5. Identification and quantification of olive oil adulterants**

Extra virgin olive oil is adulterated with oils of low quality or price. The natural variation due to geographical origin, weather effect during growth and harvesting makes the task of detecting adulteration difficult. Analytical methods applied in the examination of chemical composition include the determination of fatty acid profiles by gas liquid chromatography (Firestone et al. 1988), high-pressure liquid chromatography (Cortesi 1993; Mariani & Fedeli 1993), pyrolysis mass spectrometry (Goodacre et al. 1993), measurement of iodine values and many other determinations. Rapid, non destructive spectroscopic methods such as Raman (Davies et al. 2000), ultraviolet (Calapaj et al. 1993), MID-IR (Lai et al. 1995; Dupuy et al. 1996; Guillen & Cabo 1999; Küpper et al. 2001) and NIR (Wesley et al. 1995; Bewig et al. 1994; Sato 1994; Wesley et al. 1996; Hourant et al. 2000) have all been applied to quantify adulterants in olive oil.

NIR spectroscopy in tandem with PCA and PLS1 regression, as studied in this laboratory, was found to be suitable for the identification and quantification of adulterants (refined olive oil, sunflower oil, maize oil and soya oil) in virgin olive oil. Binary mixtures were prepared with extra virgin olive oil and each one of the selected potential adulterants. Different amounts of refined olive oil and 3 commercial oil samples (sunflower, soya and maize) were mixed with olive oil giving four different data sets. 25 samples were prepared for each data set (binary mixture), containing additions from 5 to 95 mL of adulterant. Additionally, for each adulterant, an independent prediction set of 8 samples was prepared.

NIR spectra from the samples were obtained with a Perkin Elmer Spectrum One NTS FT-NIR spectrometer. The data were recorded in the spectral range between 10000 – 4500 cm-1, by coadding 30 scans with a resolution of 8 cm-1. Each sample was acquired five times. PCA allowed the characterization of the sample relationships (scores plans) and the recovery of their subspectral profiles (loadings) (Jolliffe 2002). For this analysis, the 6100 - 4500 cm-1 region was selected and each spectrum was SNV corrected. A calibration model was built for each adulterant and was validated with the external and independent prediction data sets.

Oil samples are distributed across PC1 axes according to the olive oil content (Figure 3a). The bands located at 4596, 4668 4704, 5880 and 6024 cm-1 are related with the samples with larger amount of olive oil (Figure 3b).

Parameters of the best calibration models built for each adulterant are shown in Table 2. The four calibration models were built in the spectroscopic region of 4536-6108 cm-1.

Several factors may reduce the predictive ability of such modeling techniques: 1) low amount of some parameters present in the olive oil; 2) NIR instrument characteristics (selectivity, specificity, signal to noise ratio, etc.) and 3) sampling distribution. At this stage seven fatty acids can be estimated using NIR, but many more could be quantified. It will be necessary to add more samples with wider ranges of parameters in order to enhance the robustness of the models. Galtier et al. (2007) have managed to quantify 14 fatty acids,

The spectral profiles extracted from infrared spectra using chemometric methods could be in many cases a substitute for the traditional analysis, for olive oil overall characterization.

Extra virgin olive oil is adulterated with oils of low quality or price. The natural variation due to geographical origin, weather effect during growth and harvesting makes the task of detecting adulteration difficult. Analytical methods applied in the examination of chemical composition include the determination of fatty acid profiles by gas liquid chromatography (Firestone et al. 1988), high-pressure liquid chromatography (Cortesi 1993; Mariani & Fedeli 1993), pyrolysis mass spectrometry (Goodacre et al. 1993), measurement of iodine values and many other determinations. Rapid, non destructive spectroscopic methods such as Raman (Davies et al. 2000), ultraviolet (Calapaj et al. 1993), MID-IR (Lai et al. 1995; Dupuy et al. 1996; Guillen & Cabo 1999; Küpper et al. 2001) and NIR (Wesley et al. 1995; Bewig et al. 1994; Sato 1994; Wesley et al. 1996; Hourant et al. 2000) have all been applied to quantify

NIR spectroscopy in tandem with PCA and PLS1 regression, as studied in this laboratory, was found to be suitable for the identification and quantification of adulterants (refined olive oil, sunflower oil, maize oil and soya oil) in virgin olive oil. Binary mixtures were prepared with extra virgin olive oil and each one of the selected potential adulterants. Different amounts of refined olive oil and 3 commercial oil samples (sunflower, soya and maize) were mixed with olive oil giving four different data sets. 25 samples were prepared for each data set (binary mixture), containing additions from 5 to 95 mL of adulterant. Additionally, for each adulterant, an independent prediction set of 8 samples was prepared. NIR spectra from the samples were obtained with a Perkin Elmer Spectrum One NTS FT-NIR spectrometer. The data were recorded in the spectral range between 10000 – 4500 cm-1, by coadding 30 scans with a resolution of 8 cm-1. Each sample was acquired five times. PCA allowed the characterization of the sample relationships (scores plans) and the recovery of their subspectral profiles (loadings) (Jolliffe 2002). For this analysis, the 6100 - 4500 cm-1 region was selected and each spectrum was SNV corrected. A calibration model was built for each

adulterant and was validated with the external and independent prediction data sets.

four calibration models were built in the spectroscopic region of 4536-6108 cm-1.

Oil samples are distributed across PC1 axes according to the olive oil content (Figure 3a). The bands located at 4596, 4668 4704, 5880 and 6024 cm-1 are related with the samples with

Parameters of the best calibration models built for each adulterant are shown in Table 2. The

**5. Identification and quantification of olive oil adulterants** 

squalene and triacilglycerols.

adulterants in olive oil.

larger amount of olive oil (Figure 3b).

(b)

Fig. 3. (a) Scores plot of the first principal component (PC1), obtained for the set of virgin olive oil (voo) adulterated with refined oil, sunflower oil, maize oil and soya oil. (b) PC1 loading profile.


Table 2. Statistical parameters obtained for the calibration models built for each adulterant.

Quality Evaluation of Olives, Olive Pomace and Olive Oil by Infrared Spectroscopy 139

pomace using NIR and PLS1 regression, a Monte Carlo Cross-Validation (MCCV) (Xu & Liang 2001) framework was used. This approach was needed for building robust calibration

The model for water content estimation was built by a preliminary spectrum pre-treatment by computing the 1st derivative, in order to minimize the baseline effect, followed by Standard Normal Deviate (SNV). A PLS1 regression model with 3 LVs (Latent Variables) was needed to obtain predictive power with a Q2 of 0.96 and a relative RMSECV of 1.1%. The **b** vector plot for the water calibration models and the relationship between measured and predicted water

(b) Fig. 4. (a) b vector plot for the water calibration and (b) relationship between measured and predicted water values using a PLS1 model with 3 Latent Variables (Reproduced with

permission from Barros et al. 2009 Springer 2009).

(a)

values using a PLS1 are presented, respectively, in Figure 4a and Figure 4b.

models for real-time industrial application.

The coefficient of determination higher than 0.99, and the low root mean squared error of prediction (RMSEP) suggest a good predictive power. PLS1 regression based calibration models were used to predict the percentage of adulterant in the independent data sets. Results presented in Table 3 suggest that NIR spectroscopy in tandem with PLS1 regression is suitable to detect and quantify adulteration with other edible oils (sunflower, soya, maize refined olive oil) in extra virgin olive oil up to 2% (w/w).


Table 3. Predicted values using the calibration model built for each adulterant.
