**2. Instrumentation**

48 Olive Oil – Constituents, Quality, Health Properties and Bioconversions

A light beam illuminating an olive oil sample gives rise to reflected, transmitted, and scattered intensities. Optical absorption spectroscopy, as shown in Figure 2, makes use of a broadband UV-VIS-NIR source of intensity I0 to illuminate the olive oil sample. Then, the transmitted light intensity I, as a function of the illumination wavelength is measured. The change in light intensity, providing the transmittance T, is determined by the molar absorptivity the concentration of absorbing species C, and the optical path L, via the Lambert-Beer relationship, expressed by Equations 1 and 2. T is frequently expressed logarithmically as in Equation 3, to give the so called optical absorbance A, which results

> () C <sup>I</sup> <sup>I</sup> <sup>0</sup> L

> > *I I CL* <sup>0</sup> exp

exp( ) *<sup>I</sup> <sup>T</sup> C L*

<sup>0</sup> log *<sup>I</sup> A CL <sup>I</sup>*

This chapter focuses on extra virgin olive oil quality evaluation achieved by means of UV-VIS-NIR absorption spectroscopy. The composition of olive oil is about 98% triglycerides




In practice, the entire UV-VIS-NIR absorption spectrum can be considered an optical signature, a sort of univocal *fingerprint* of the olive oil. The spectroscopic data can be suitably processed for obtaining a correlation to quality indicators, to the geographic origin

such as anthocyanins present distinctive absorption bands in the VIS.

of production, to product authenticity as well as to adulteration detection.

0

*I* 

(1)

(2)

(3)

linearly dependent on concentration.

Fig. 2. Optical absorption spectroscopy: the working principle

and approximately 2% non glycerid constituents.

Márquez, 2003; Cerretani et al., 2005).

(Osborne et al., 1993; Ozaki et al., 2007).

Absorption spectroscopy in the UV-VIS-NIR range is one of the most popular measuring methods of conventional analytic chemistry (Mellon, 1950; Bauman, 1962). The most relevant advantages offered are:


The conventional instrument for absorption spectroscopy is the double-beam spectrophotometer, the working principle of which is depicted in Figure 3. Since quartzbased optical fibers are transparent in the UV-VIS-NIR range, they are used to equip conventional spectrophotometers by flexible means. Indeed, optical fibers offer the unique possibility of localized probing, a particularly attractive feature for online measurements, which can be carried out in real time without any sample drawing. Moreover, the recent availability of bright LEDs and miniaturized spectrometers further enhances the intrinsic optical and mechanical characteristics of optical fibers and makes it possible to implement compact and moderate-cost instruments.

Fig. 3. Working principle of a double-beam spectrophotometer

The conventional spectrophotometer, implemented by means of optical fiber technology, is depicted in Figure 4. In this case, optical fibers are used for both illumination and detection,

Fig. 4. Fiber optic setup for absorption spectroscopy

Optical Absorption Spectroscopy for Quality Assessment of Extra Virgin Olive Oil 51


R2 R = 1 <sup>2</sup> << 1

Fig. 6. Steps toward a multicomponent analysis by spectroscopy and chemometrics

non-destructive method for quality assessment of extra virgin olive oil.


starting variables;

Given the nature and complexity of the spectroscopic and analytical data sets involved, many multivariate chemometric techniques have been proposed (Ingle et al., 1988; Mark et al., 2007). The challenge of every multivariate data processing method is to provide excellent classification performance even when few training data are available. Indeed, smart data processing by means of chemometrics makes UV-VIS-NIR optical spectroscopy a rapid and

One of the most popular techniques for explorative analysis, data dimensionality reduction, and clustering is the Principal Component Analysis (PCA) that provides the coordinates for identifying the samples in a 2D or 3D map (Jackson, 2003). It linearly combines the spectroscopic data characterizing each oil sample to produce new variables. The coefficients giving the weight of each variable in the linear combination are called *loadings*. The new


variables are called Principal Components (PCs), and have the following properties:

along a straight line with a narrow confidence limit.

further steps of calibration and validation.

**Wavelength (nm)**

measured spectra

**Transmittance (%)**

extraction of sample cluster coordinates

any aberrant data. Ideally, the scatter plot should contain data points distributed evenly

1) data dimensionality reduction and clustering according to

2) calibration and predictive model building for evaluation of quality

peculiar features

analytical data

selected reference samples

3) validation

indicators

and a single cuvette is used. The reference spectrum is measured prior to sample analysis. While the optical fiber strand for illumination can be a single optical fiber or a bundle, the detection is necessarily carried out by means of a bifurcated bundle, or by a coupling device, so as to split the detected light intensity into two spectrometers, for the UV-VIS and the NIR spectroscopic range, respectively.

#### **3. A touch of chemometrics**

The chemical information given by an absorption spectrum is contained in the positions and intensities of the absorption bands. Whereas the band positions give information about the appearance and the structure of certain chemical compounds in a mixture, the intensities of the bands are related to the yield of these compounds. Since olive oil contains numerous compounds, the UV-VIS-NIR absorption spectrum shows broad peaks resulting from the convolution of the many overlapping bands, as summarized in Figure 5 (Osborne, 2000).

Fig. 5. Absorption spectrum – convolution of absorption spectra of many compounds in a mixture

For qualitative and quantitative analysis, the spectroscopic fingerprinting must be calibrated against reference analytical data from a database of samples representing the best variability in the population. Figure 6 summarizes the steps to follow for achieving a multicomponent analysis from absorption spectroscopy.


and a single cuvette is used. The reference spectrum is measured prior to sample analysis. While the optical fiber strand for illumination can be a single optical fiber or a bundle, the detection is necessarily carried out by means of a bifurcated bundle, or by a coupling device, so as to split the detected light intensity into two spectrometers, for the UV-VIS and the NIR

The chemical information given by an absorption spectrum is contained in the positions and intensities of the absorption bands. Whereas the band positions give information about the appearance and the structure of certain chemical compounds in a mixture, the intensities of the bands are related to the yield of these compounds. Since olive oil contains numerous compounds, the UV-VIS-NIR absorption spectrum shows broad peaks resulting from the convolution of the many overlapping bands, as summarized in Figure 5 (Osborne, 2000).

Fig. 5. Absorption spectrum – convolution of absorption spectra of many compounds in a

For qualitative and quantitative analysis, the spectroscopic fingerprinting must be calibrated against reference analytical data from a database of samples representing the best variability in the population. Figure 6 summarizes the steps to follow for achieving a multicomponent




spectroscopic range, respectively.

**3. A touch of chemometrics** 

mixture

analysis from absorption spectroscopy.

which the spectrum of a test sample may be matched.

more likely to be correlated with spectroscopic data.

any aberrant data. Ideally, the scatter plot should contain data points distributed evenly along a straight line with a narrow confidence limit.


Fig. 6. Steps toward a multicomponent analysis by spectroscopy and chemometrics

Given the nature and complexity of the spectroscopic and analytical data sets involved, many multivariate chemometric techniques have been proposed (Ingle et al., 1988; Mark et al., 2007). The challenge of every multivariate data processing method is to provide excellent classification performance even when few training data are available. Indeed, smart data processing by means of chemometrics makes UV-VIS-NIR optical spectroscopy a rapid and non-destructive method for quality assessment of extra virgin olive oil.

One of the most popular techniques for explorative analysis, data dimensionality reduction, and clustering is the Principal Component Analysis (PCA) that provides the coordinates for identifying the samples in a 2D or 3D map (Jackson, 2003). It linearly combines the spectroscopic data characterizing each oil sample to produce new variables. The coefficients giving the weight of each variable in the linear combination are called *loadings*. The new variables are called Principal Components (PCs), and have the following properties:


Optical Absorption Spectroscopy for Quality Assessment of Extra Virgin Olive Oil 53

evaluated on the calibration set. The closeness of RMSEC and RMSEP provides an estimation of robustness of the predictive model. R2 is, instead, the squared correlation coefficients between predicted and reference values, for the calibration set; thus the fit is as

The extra virgin olive oil is a blend of *cultivars* – different varieties of fruit species – which determine the blend's organoleptic properties. In addition to a distinctive taste, each oil blend has a distinctive color and turbidity. Color is mainly determined by the pigment content of the olives, and by the stage of ripening when they are harvested, whereas turbidity is mainly related to the mill type. Moreover, some commercial oils are filtered before bottling, while other oils are bottled unfiltered to provide a natural appearance. Anyway, both color and turbidity provide a means for oil assessment, and standard



These color and turbidity measurement methods are popular because of their generality, simplicity, and applicability. In the case of color, since the definition of color is independent of the substance, the method can be applied to any liquid. In the case of turbidity, the independence of the test material is attained by reference to an ISO standard turbid material. However, both methods view the characteristics separately, i.e., the color method never considers the liquid's turbidity, while the turbidity method never considers the color. Usually, in order to avoid mutual interferences, the olive oil is filtered prior to color measurements, while turbidity measurements are performed at a color-independent wavelength – typically at 830 nm. In order to evaluate the olive oil sample as it is, a new technique combining color and turbidity measurements was proposed, which was called scattered colorimetry or spectral nephelometry, since it extended the color and turbidity standards by adding light sources and

Scattered colorimetry makes use of four white light sources, which span the 450-630 nm spectroscopic range, and a miniaturized optical fiber spectrometer as detector. The sources, positioned at different angles with respect to the detector, are sequentially switched on to measure, in addition to the transmitted spectrum, the scattered spectra at the given angles. The transmitted spectrum mainly provides information regarding color, which is also dependent on the turbidity, while the scattered spectra mainly provide information on turbidity, which is also dependent on the color. Measurements on a vial filled by distilled water are carried out prior to sample analysis, so as to obtain reference signals. Then, the spectroscopic data are processed by means of chemometrics, thus providing few coordinates

nephelometric turbidity units (Figure 7, top-right) (ISO 7027, 1999).

observation angles, as shown in Figure 7-bottom (Mignani et al., 2003).

that summarize the combined effects of color and turbidity.

better as this value is closer to 1 (R2 1).

**4. Scattered colorimetry – Looking at olive oil as it is** 

techniques are used for their independent measurement.

(Billmeyer et al., 1981; Hunt, 1987).


This means that high order PCs have little weight in characterizing the oil samples, and can be disregarded with little loss of information. The loading plots are useful to interpret the score map: they show what variables are important for a given PC. A variable with 0 loading has no importance, a variable with high (positive or negative) loading is important. If a PC has a positive or negative loading at a given wavelength, a sample having high absorbance at that wavelength will tend to have a positive or negative score along that PC.

PCA can also be used for prediction of categorical variables, that is classification, at least in those cases in which they are clearly distinguishable. However, when PCA does not provide satisfactory classification, a more dedicated multivariate data processing tool must be used, such as the Linear Discriminant Analysis (LDA). LDA is a powerful tool that provides both a reduction in dimensionality and automatic object classification. Like PCA, LDA projects a high-dimensional pattern onto a subspace of smaller dimension, but the axes for projection are chosen using a different criterion (Vandeginste et al., 1998).

From the point of view of discriminant analysis, PCA has a drawback: because it weighs the variables in terms of their variance, the features with good discriminating power but limited variance are disregarded. On the contrary, LDA is a tool that is specifically suited for identification, and looks for those variables that show a large spread among different clusters (inter-class variance), but limited variance within each cluster (intra-class variance). Given an N-class problem, the LDA extracts from the data matrix N-1 Discriminating Functions (DFs), which correspond to Principal Components in the PCA, but show a better resolution with regard to poorly-separated clusters. LDA is a *supervised* method. It needs a training set of already classified objects to estimate inter-class and intra-class variances. If all starting variables obey the Gaussian distribution, LDA also provides an easy way to classify unknown patterns. In this case, the points of each cluster in the DF space are distributed following the (N-1)-dimensional Gaussian density function. Thus, classification is achieved simply by evaluating the coordinates of a pattern in the DF space, and then by seeing which of the N density function is higher at that point.

One of the most popular techniques for the prediction of quantitative variables, such as the concentration of an analyte in a multicomponent mixture, is the Partial Least Squares regression (PLS) (Wold et al., 2001). This method is used when the predictor matrix has many collinear variables and the usual Multiple Linear Regression (MLR) cannot be applied. PLS looks for a limited number of PLS factors (PF) which are linear combinations of the original predictors. These new variables are mutually orthogonal (thus uncorrelated) and have the maximum possible covariance with the target variable, among all possible combinations of the original predictors. The idea is that each PF should be linked to a different source of data variance, with the first PF being the most linked to the target variable. The estimation of the optimal number of factors needed to fit the data is a critical issue of PLS. The optimal number of factors is usually assessed by testing each PLS model on the validation set and by minimizing the RMSEP (Root Mean Square Error of Prediction). There are two other fundamental parameters for assessing the goodness of the fit: the RMSEC (Root Mean Square Error of Calibration) and the determination coefficient (R2), respectively. RMSEC is, like RMSEP, an estimation of the expected prediction error, but is evaluated on the calibration set. The closeness of RMSEC and RMSEP provides an estimation of robustness of the predictive model. R2 is, instead, the squared correlation coefficients between predicted and reference values, for the calibration set; thus the fit is as better as this value is closer to 1 (R2 1).
