Application of supervised chemometric techniques and synchronized excitation-emission spectrofluorometric analysis for the verification of Maltese extra virgin olive oils

Abstract

The authentication of virgin olive oil samples usually requires the use of sophisticated and very expensive analytical techniques. In this study, the potential of fluorescence spectroscopy for the authentication and discrimination of Maltese extra virgin olive oils was carried out using synchronized excitation-emission spectroscopy. Samples were collected from various producers around the Maltese islands. Synchronous excitation emission spectra were collected in the region of 240–750 nm with wavelength intervals of 10, 30, 60, 80 120 and 185 nm and subjected to several supervised chemometric procedures. Partial least square regression, linear discriminate analysis, and artificial neural network were used to define the origin of the Maltese olive oil against olive oils derived from other neighboring countries in the Mediterranean region. After subjecting the spectroscopic data to different pre-treatments and variable selection procedures results obtained evidenced a higher classification accuracy. This accuracy and predictability were highly dependent on the wave interval used and on the chemometric method used, however it was found that in general spectra obtained using δ 10 nm were deemed the most appropriate, with PLS, ANN and LDA reaching 100% accuracy and predictability in discriminating Maltese extra virgin olive oils when using derivatized spectral transformations.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13197-022-05371-x.

Introduction

Cultivation of the olive tree (Olea europea L.) is strongly associated with the shores of the Mediterranean basin. Cultivation started in the third millennium B.C. and since then it influenced the Mediterranean regions and its populations (Loukas and Krimbas 1983). This historical legacy and the uniqueness of the olive tree and olive oil called for the establishment of production protection systems, such as the designation of origin PDO and protected geographical indication of origin (PGI) (EEC Regulation No. 2082/92 and later No. 1151/2012). These systems were designated to protect the typical features and authenticity of food products and to discourage competition from similar replacement products. According to the EEC regulation, (Reg. 1151/2012) for a product to be classified under PDO status, the entire production cycle, from raw material to finished product including processing and packaging, must be carried out in one given territory. Due to the amalgamation of different factors, including the starting raw materials, environmental physiognomies, location, and skillfulness of the producer, such a product is unique and not reproducible elsewhere, bearing distinctive and exclusive characteristics such as defined chemical composition and distinct organoleptic parameters.

The determination of the origin and the authenticity of olive oils have been studied extensively in the past few years using extremely varied physicochemical techniques in conjunction with chemometric studies. These studies can be classified into two main categories (Dupuy et al. 2005). In one category the samples are previously subjected to a chemical treatment to determine and quantify the individual chemical constituents such as fatty acids, triacylglycerols (Bucci et al. 2002; Ollivier et al. 2003), sterols (Sakouhi et al. 2008), phenolic compounds (Talhaoui et al. 2016; Rigane et al. 2011), and inorganic multi-elemental composition (Lia et al. 2020). The other category is based on sample preservation, where a small sample is analysed without extensive preliminary treatment, including ¹H and ¹³C NMR analysis (Rosa et al. 2016). Fourier transform infrared spectroscopy (FTIR) (De Luca et al. 2012). Near infra-red (Galtier et al. 2007) and Synchronous excitation-emission fluorescence spectroscopy (SEEFS) (Dupuy et al. 2005; Guime et al. 2004). Several different studies have been carried out throughout the years which assessed the uniqueness of the Maltese EVOOs and the trees from which they were derived from, these included genetic analysis (Mazzitelli et al. 2015) and secondary metabolites (Gatt et al. 2021).

Fluorescence spectroscopy is an emergent analytical technique, which presents good sensitivity, with minimal sample preparation. In the case of spectrofluorometric analysis, molecules exhibiting a fluorescent nature are analysed through the simultaneous scanning of the excitation and emission wavelengths resulting in an excitation-emission matrix, known also as a total luminescence spectrum or fluorescence landscape. Total luminescence spectra are usually presented as a three-dimensional plot, with the fluorescence intensity plotted in function of the excitation and the emission wavelengths, or as a two-dimensional contour map, in which one axis represents the emission and another the excitation wavelength, and the contours are plotted by linking points of equal fluorescence intensity (Ndou and Isiah 1991). Fluorescence spectroscopy has been used for the analysis of minor constituents found in olive oil including chlorophylls, tocopherols, and phenolic compounds (Kyriakidis and Skarkalis 2000; Zandomeneghi and Zandomeneghi 2005). In comparison with other absorption methods (NMR, FTIR and Raman), fluorescence spectroscopy is 100–1000 times more sensitive enabling the measurement of concentrations down to parts per billion levels (Zandomeneghi et al. 2006). This is mainly attributed to the detection of fluorescence against a low background, increasing the overall signal to noise ratio.

An alternative to the multicomponent fluorescent systems, synchronous fluorescence techniques proposed by Lloyd (1971), have been successfully employed in the characterization and discrimination of edible oils (Sikorska et al. 2005). Rather than using the whole excitation-emission matrix (EEM), adjusting both excitation and emission monochromators to scan in the same instant with a constant wavelength interval between excitation and emission yields a synchronised excitation-emission spectrum (SEEFs). This allows narrowing of the spectral bands and facilitates data handling. Although the SEEF spectra contain less information than the EEM, they are potentially more informative than single excitation and emission spectra. The selection of an appropriate offset between excitation and emission wavelengths allows the study of singled out fluorophores, in so doing increasing the sensitivity and selectivity. Coupling this technique to chemometrics enables the distinction of commercially available samples of virgin olive oils, pure olive oils, and olive pomace oils (Poulli et al. 2005) and determines the overall quality and possible adulteration of the olive oil (Kyriakidis and Skarkalis 2000; Zaroual et al. 2020; Venturini et al. 2021). Recent applications of fluorescent spectroscopy in conjunction with advanced statistical analysis have proven useful in the discrimination of oils deriving from different cultivars and producing regions (Al Riza et al. 2021; Kontzedaki et al. 2020).

The aim of this study was to employ the use of SEEF spectra coupled with chemometrics to differentiate the Maltese EVOO’s from other EVOO’s derived from other countries within the Mediterranean region, thus developing a quick, easy, and cost-saving verification of the origin of EVOOs from the Maltese islands, paving the path for the application of protected designation of origin. In this study, SEEF collected in the region of 240–750 nm at different wave intervals of 10, 30, 60, 80 120 and 185 nm data were analysed both by a discriminant analysis using partial least squares (PLS) and linear discriminate analysis (LDA) and a modelling chemometric tool using artificial neural networks (ANN). Moreover, the study aims to identify a clear understanding of which signal pretreatment could be better for authentication purposes using different chemometric methods. Furthermore, the effect of the different spectral transformations on the final classification outcomes was also investigated in this study.

Materials and methods

Extra virgin olive samples

For this study, a total of 65 extra virgin olive oil samples were collected from the Maltese islands over four harvest seasons from 2013–2016 and from other neighboring Mediterranean countries (Supplementary Material 1). The samples were all taken from different oil producers to cover a representative sample of the Maltese islands in terms of pedological and microclimatic conditions, whilst also accounting for manufacturing techniques and the different presses employed. Foreign olive oils obtained were bought with a protected designation of origin to ensure traceability of the product. All the samples were stored at 4 °C in the absence of light prior to the analysis. The samples were preheated to 35 °C in a water bath for 1 h prior to use and then mixed to ensure homogeneity.

SEEF spectral acquisition and pretreatment

A three-dimensional (3D) matrix excitation-emission matrix (EEM) was obtained for each sample using a Jasco FP-8300 fluorescence spectrophotometer, with both the excitation and the emission bandwidths set at 5 nm for a measurement range between 240 to 750 nm. The acquisition interval and the integration time were maintained at 0.5 nm and 10 ms respectively, with a scan speed of 5000 nm.min⁻¹. The oil samples were examined by means of right-angle geometry. SEEFS were acquired by simultaneous scanning of the excitation and the emission monochromators, with a constant distance, δλ, of 10, 30, 60, 80, 120 and 185 nm. All analyses were carried out in duplicates, and the results reported as mean values. Fluorescence intensities were plotted as a function of the excitation wavelength. The spectra were exported as an ASCII file using the instrumental software Spectrum (PerkinElmer Inc., Waltham, MA) and imported directly into The Unscrambler X 10.3 (CAMO Software Oslo, Norway) for all subsequent mathematical data processing. The spectra obtained were subjected to different spectroscopic signal processing techniques, which were evaluated and compared. The spectra were normalised, a transformation that put all spectra on the same scale, thus eliminating the fluctuations in intensities between spectra arising from slightly different sample concentrations. Both peak normalisation and area normalisation were carried out separately on the baseline corrected spectrum. Normalisation was followed by detrending and deresolving procedures. The detrend transformation removes the effects of nonlinear trends, showing only the absolute changes in values across spectra by removing the least-squares line of best fit from the data, thus focusing only on fluctuations between data. Deresolve is a noise-reducing transformation that operates by artificially lowering the resolution of the spectra. Other treatments applied to the baseline corrected spectrum include multiplicative and orthogonal scatter corrections (MSC and OSC), and standard normal variate (SNV). MSC was corrected for scaling effects by performing a regression of a spectrum against a reference spectrum obtained from an average of spectra in the training set, thereby correcting the spectrum using the slope of the fit was obtained from the regression. OSC removes variance from the factors that is not related to the response, by finding directions in X that describe large variances while being orthogonal to Y and subtracting them from the data. The SNV transformation works similarly to MSC, however it standardises each spectrum using data from the spectrum itself rather than data averaged from all the spectra. Several derivatising procedures (1^st and 2^nd derivatives, Savitzky-Golay) were also carried out. The 1^st derivative removes baseline effects while the 2^nd derivative also removes the slope of the spectrum by measuring the change in slope, thereby sharpening spectral features. The Savitzky-Golay derivative fits a low-degree polynomial to adjacent points in a spectrum, thereby smoothing the spectrum while minimally affecting the signal-to-noise ratio.

Chemometric analysis, unsupervised and supervised methods

Principal component analysis (PCA) was carried out on the data, to identify possible outliers or any possible clustering of the Maltese samples present within the data set. The whole dataset was then split into two sets: the training and test sets (the former to build the model, the latter to validate it). To preserve the diversity in the training and test sets and to account for the fact that different pre-treatments had to be tested a unique sample splitting scheme was used. The following method was adopted to cover maximum variation in the two sets and at the same time being able to compare the outcomes after the different pre-treatments. The Maltese and the non-Maltese samples were grouped in an ascending way so that the first 30 samples would represent Maltese EVOO’s whilst the rest (35 samples) correspond to non –Maltese EVOO’s. A Venetian blinds cross-validation which selects every s^th sample from the data by making data splits such that all samples are left out exactly once (s = 5) was used. This sampling method excluded 20% of the observations so that they would be retained as the testing set. The leave one out cross-validation method was done on the training set consisting of 80% of the data. Double cross validation method was thus carried out because the internal cross validation method carried out by the software tends to be an overestimation. The fact that the model was double cross validated, prevented overfitting of the model and the selected samples for prediction were selected in a way to provide equal representation for both geographical origins. The supervised chemometric treatment was performed using PLS. To classify the geographical origin with regards to Maltese and non-Maltese samples, a dummy variable system was used as previously employed by Dupuy et al. (2015). The Maltese samples were assigned a dummy variable of 1, while the non-Maltese samples were assigned a value of 0. The predicted origins of samples were classified as follows: scores < 0 were denoted 0; scores > 1 were denoted 1. The cut-off point was set at 0.5 whereby samples with a predicted value of > 0.5 were labelled as foreign, while the remaining samples were labelled as local. The predictability of the model was assessed using excluded rows validation (ERV), with the exclusion of one third of the samples from each class to assess for model overfitting. The training set was furthermore subjected to leave one out cross-validation (LOOCV), after which PLS was repeated on the excluded samples.

Variable selection

Once the optimum number of factors was determined, the data points which had a VIP (variable importance in projection) > 0.8 were then used to develop subsequent PLS models. The VIP score is a measure of a variable’s importance in the PLS model. It represents the contribution of a variable to the PLS model and is determined through a weighted sum of the squared correlations between the model components and the original variable. A value less than 0.8 is typically considered to be a small VIP, and thus, a candidate for deletion from the model (Wold et al. 1998). The VIP of a predictor is a value that expresses the contribution of the individual variable in the definition of the latent vector model. It is defined according to the formula:

$$VI{P}_j=\sqrt{N_{\text{var}s}\frac{{\sum }_{k=1}^F{(b_k^2{t}_k^{T}t)\left(\frac{W_{\mathit{jk}}}{W_k}\right)}^2}{{\sum }_{k=1}^F(b_k^2{t}_k^{T}t)}}$$ 1

where t^k is the vector of sample scores along the k^th latent variable, b_k is the coefficient of the k^th PLS inner relationship, N_vars is the number of experimental variables and w_jk and W_k are the weight of the j^th variable for the k^th latent variable and the weight vector for the kth latent variable, respectively. A VIP score is a measure of a variable’s importance in modelling both X and Y. Furthermore, stepwise linear canonical discriminant analysis (SLC-DA), as implemented within JMP 10, was also used, to reduce the number of variables used in PLS-DA models. A stepwise analysis allows for the manual selection of variables used to build the linear model up to a maximum number of entries (n–1), where n is the number of samples in the sample set. The model containing the most discriminant variables was selected based on a low F-ratio and a high p-value. To further reduce the variables an inspection of the canonical coefficients of the selected variables was carried out. The selected variables obtained in SLC-DA were arranged in ascending order in terms of their scoring coefficients. A smaller set of variables was selected which consisted of 20 variables which corresponded to 10 of the most positive and 10 most negative standardized scoring coefficients.

Artificial neural networks

To determine the suitability of the whole SEEF spectra for discrimination of EVOOs of Maltese origin, an artificial neural network (ANN) analysis was carried out. The main advantage of a neural network model is that it can efficiently model different response surfaces due to its nonlinearity, allowing a better fit to the data given enough hidden nodes and layers, providing an accurate prediction for many kinds of data. Unlike other modeling and discriminate methods, the main disadvantage of a neural network model is that the results are not easily interpretable, due to the presence of several intermediate hidden layers. In this experiment, 25 iterations were carried out using a TanH activation function as the standard neuron activation function. In the case of ANN, three different cross-validation techniques were employed to prevent model overfitting; the k-fold (CV-10), hold back (33.3%), and excluded rows (Venetian blinds). Thirty-three percent of the samples were held back from the model during holdback validation, which operates by randomly splitting the dataset into training and validation sets. Thirty-three percent of the data was thus ‘held back’ to form the validation set. Excluded rows holdback uses those rows that were excluded by the Venetian blinds method as the validation set. K-fold validation divides the dataset into ‘k’ number of subsets where each subset contains a fraction ‘1/k’ of the data. Each of these sets is used to validate the model thereby fitting ‘k’ number of models. The best fitting model is presented as the final output. In this study, K-fold validation was carried out using 5 k-folds.

Model performance

The root mean square error RMSE for the model was calculated as shown by the equation below, to further assess the accuracy of the model, where y_pred corresponds to the value between 0 and 1 generated by the model, whilst y_ref corresponds to the dummy variable to which the oil was assigned.

$$\text{RMSE}=\sqrt{\frac{\sum _{i=1}^n{\left(y_{\text{pred}}-y_{\text{ref}}\right)}^2}{n}}$$ 2

The optimum model for each transformation was chosen after an assessment of the PLS model parameters. The classification accuracy of the LOOCV and ERV models, explained X and Y variance, RMSE and number of factors were used to evaluate the performance of the chemometric models. The accuracy of training testing and prediction by PLS was determined as the numerical coordinate systems were round up to the nearest integer of either zero or one. A negative value was assigned a value of zero, whereas a value greater than one was assigned a value of one. The numerical output obtained was compared to the previously assigned value based on the known origin. In the case of LDA, the output obtained classified the sample as either zero or one, so there was no need for any manipulation. The % accuracy was determined by the following equation.

$$\%Accuracy=100-\left(\left(\frac{K_m}{K_t}\right)\times 100\right)$$ 3

where K_m is the number of samples that are misclassified, K_t is the total number of samples used in training, testing, or prediction model.

Results

Identification of synchronous fluorescence spectra

Examination of the total synchronous fluorescence spectra of EVOOs as shown in Fig. 1 showed that spectral shape and intensity is dependent on the wavelength interval (δ nm) used.

Fig. 1.

The full EEM obtained for extra virgin olive oil, and the corresponding SEEFS measured at δ 10, 30, 60, 80, 120, 185 nm using right angle geometry were obtained. The black line corresponds to EVOOs derived from the Maltese islands; the red line corresponds to EVOOs derived from foreign countries

Chemometric analysis of synchronous fluorescence spectra

After splitting the data according to the procedure described above, chemometric classification models were built and tested on all the SEEF δΔ nm and their corresponding spectral pretreatment using a PLS regression algorithm using JMP 10. Table 1 shows the number of latent variables extracted, the predicted root means square error and the % variation explained in terms of X and Y for some of the different spectral pretreatment methods for each SEEF.

Table 1.

PLS-DA performance for the whole SEEF spectrum recorded at different δ nm under different pretreatments

	PLS
	LV	%X	%Y	PRESS	%A	%P	LV	%X	%Y	PRESS	%A	%P
	δ 10 nm						δ 30 nm
Raw	3	64.29	74.02	0.77	98.21	100	4	84.44	58.58	0.87	98.21	100
Normalized	5	75.96	84.55	0.68	96.43	81.82	4	82.82	62.78	0.82	92.86	100
Q Norm	4	63.79	88.34	0.66	98.21	90.91	1	43.91	24.03	0.98	89.29	100
Baseline	4	86.13	68.35	0.78	91.07	90.91	2	41.37	41.73	0.83	89.29	90.91
Detrend	3	65.35	68.98	0.8	94.64	90.91	15	99.22	89.37	0.8	94.64	100
Deresolve	3	64.99	58.53	0.77	85.71	72.73	5	90.72	64.36	0.82	98.21	90.91
SNV	5	88.09	76.52	0.77	94.64	72.73	5	93.65	60.96	0.82	91.07	100
MSC	2	53.27	48.48	0.82	85.71	81.82	7	99.07	66.07	0.9	92.86	90.91
OSC	3	57.81	75.63	0.71	98.21	90.91	3	66	55.4	0.88	92.86	100
Savitzky Golay	3	43.99	81.32	0.71	98.21	100	3	60.89	72.08	0.83	100	100
1st Derivative	3	52.97	71.89	0.78	96.43	90.91	3	60.89	72.08	0.83	100	100
2nd Derivative	4	47.97	85.69	0.68	100	100	3	52.27	74.31	0.87	100	100
	δ 60 nm						δ 80 nm
Raw	1	36.28	20.37	0.98	75	63.64	5	87.79	69.98	0.87	98.21	100
Normalized	3	74.48	51.72	0.86	87.5	90.91	5	83.76	70.24	0.76	94.64	100
Q Norm	1	48.63	18.86	0.98	82.14	90.91	3	66.78	63.77	0.78	89.29	81.82
Baseline	1	42.65	16.43	1.02	75	63.64	6	94.68	75.53	0.79	100	100
Detrend	2	60.19	39.81	0.97	75	63.64	5	98.16	60.49	0.82	98.21	100
Deresolve	1	37.34	19.75	0.98	87.5	90.91	6	93.27	69.88	0.8	96.43	100
SNV	2	66.93	27.09	0.98	80.36	81.82	5	93.1	68.3	0.74	94.64	100
MSC	3	88.12	30.95	0.96	76.79	81.82	8	99.13	76.68	0.74	96.43	100
OSC	1	38.09	20.94	0.98	73.21	63.64	4	78.3	71.08	0.78	100	100
Savitzky Golay	3	49.8	68.16	0.84	96.43	90.91	3	64.11	67.31	0.85	96.43	90.91
1st Derivative	3	49.8	68.16	0.84	96.43	90.91	3	64.11	67.31	0.85	96.43	90.91
2nd Derivative	1	20.44	39.81	0.96	78.57	72.73	3	58.39	69.36	0.9	96.43	100
	δ 120 nm						δ 180 nm
Raw	4	83.26	65.69	0.79	96.43	100	1	55.19	19.24	0.97	69.64	63.63
Normalized	4	83.25	65.69	0.79	92.86	81.82	1	49.92	21.61	0.97	73.21	72.73
Q Norm	5	80.31	72.42	0.78	96.43	81.82	1	37.9	27.4	0.95	78.57	81.82
Baseline	3	83.6	46.05	0.95	83.93	72.73	1	44.98	19.59	0.94	76.79	81.82
Detrend	7	95.85	79.68	0.77	94.64	90.91	2	80.42	31.25	0.96	87.5	81.82
Deresolve	4	84.24	64.25	0.77	98.21	90.91	6	88.74	58.71	0.87	87.5	100
SNV	3	77.51	63.95	0.76	91.07	72.73	1	55.93	21.52	0.94	73.21	72.73
MSC	14	99.62	92.03	0.78	94.64	90.91	1	58.48	25.66	0.91	83.93	81.82
OSC	4	75.04	73.35	0.87	92.86	81.82	1	50.56	26.94	0.94	80.36	90.91
Savitzky Golay	1	30.43	35.56	1.02	82.14	90.91	1	40	26.77	0.96	73.21	63.64
1st Derivative	3	70.29	63.59	0.79	91.07	81.82	2	53.19	45.68	0.93	91.07	100
2nd Derivative	9	80.93	93.33	0.82	92.86	72.73	1	37.41	25.13	1.01	76.79	63.64

Figure 2 shows that analysis of the VIP > 0.8 identified relevant features in the spectra, particularly peaks which corresponded to previously identified compounds.

Fig. 2.

Selected variables having a VIP score > 0.8 (red bars) for the different spectra pretreatments obtained for the different SEEF obtained at δ10, 30, 60, 80, 120, 185 nm (black lines)

Following VIP variable selection, a SLC-DA was performed on the SEEF data from all the pre-treatment methods in order to extract only a small amount of highly discriminate variables which would enable an easier and faster discrimination between the origins of EVOOs. The main advantage of using SLC-DA over the convention LDA is the ability to perform a feature selection. Only those variables which helped to improve classification performance were used whereas variables without discriminant information were discarded. Table 2 shows the results obtained from the PLS using the data set composed of variables which had a VIP score > 0.8 and were selected during the SLC-DA analysis.

Table 2.

Comparison of PLS-DA and LDA models for the SEEFs recorded at different δ nm under different pretreatments after variable selection

	PLS-DA		LDA		PLS-DA		LDA
	%A	%P	%A	%P	%A	%P	%A	%P
	δ 10 nm				δ 30 nm
Normalized	98.21	100	100	93.3	96.43	90.91	95	73.3
Q Norm	98.21	90.91	97.5	66.7	78.57	81.82	77.5	53.3
Baseline	98.21	90.91	100	93.3	96.43	100	90	80
Detrend	92.86	90.91	97.5	100	85.71	100	97.5	80
Deresolve	100	100	100	86.7	94.64	90.91	97.5	80
SNV	100	100	100	100	98.21	100	100	93.3
MSC	100	100	100	100	96.43	100	95	80
OSC	96.43	81.82	95	73.3	76.79	72.73	87.5	73.3
Savitzky Golay	96.43	100	100	100	98.21	100	97.5	93.3
1st Derivative	100	100	100	100	96.43	90.91	97.5	86.7
2nd Derivative	100	100	100	100	96.43	90.91	97.5	86.7
	δ 60 nm				δ 80 nm
Normalized	96.43	90.91	97.5	73.3	91.07	100	95	66.7
Q Norm	96.43	90.91	90	66.7	89.29	100	85	80
Baseline	83.93	90.91	97.5	86.7	98.21	100	92.5	80
Detrend	98.21	100	90	86.7	89.29	90.91	95	66.7
Deresolve	98.21	90.91	92.5	66.7	98.21	100	97.5	86.7
SNV	94.64	90.91	97.5	46.7	92.86	100	92.5	73.3
MSC	96.43	90.91	95	86.7	89.29	90.91	87.5	86.7
OSC	91.07	72.73	92.5	73.3	98.21	90.91	95	60
Savitzky Golay	96.43	90.91	95	80	98.21	100	95	66.7
1st Derivative	94.64	90.91	95	60	92.86	81.82	95	86.7
2nd Derivative	98.21	90.91	97.5	86.7	94.64	100	97.5	80
	δ 120 nm				δ 180 nm
Normalized	94.64	81.82	92.5	73.3	98.21	90.91	95	80
Q Norm	89.29	81.82	87.5	80	87.5	72.73	90	80
Baseline	98.21	100	92.5	86.7	91.07	90.91	92.5	73.3
Detrend	75	63.64	92.5	66.7	87.5	90.91	97.5	13.3
Deresolve	98.21	100	92.5	86.7	98.21	100	95	73.3
SNV	92.86	100	82.5	66.7	91.07	90.91	85	73.3
MSC	96.43	100	90	73.3	96.43	100	95	80
OSC	92.86	100	85	73.3	91.07	90.91	87.5	73.3
Savitzky Golay	94.64	90.91	92.5	86.7	94.64	100	85	80
1st Derivative	98.21	90.91	92.5	93.3	87.5	90.91	97.5	66.7
2nd Derivative	91.07	90.91	92.5	86.7	96.43	100	97.5	86.7

Table 2 also provides a comparison between LDA and PLS-DA model performance. A similar overall trend was observed, however for most of the spectral pre-treatments a slightly lower performance was observed for the LDA models.

Figure 3 compares the results obtained using LDA and PLS score projections. The two chemometric methods tend to corroborate the authenticity and verification of Maltese EVOOs.

Fig. 3.

a PLS score plot, and b LDA plot obtained using the Savitzky Golay derived SEEF obtained at 10 nm after variable selection (● = Maltese, ■ = non-Maltese origin)

Given that artificial neural networks can relate large amounts of the input and output parameters the application of AANs was computed on either SEEF spectrum without any form of variable selection methods. The models obtained for all the different SEEF spectra and their corresponding transformations are presented in Table 3.

Table 3.

Results summarizing the ANN model performance with no variable selection using three different cross-validation methods for the different SEEFs pretreatments obtained δ10, 30, 60, 80, 120, 185 nm

	33% Holdback		10-CV		Excluded Row		33% Holdback		10-CV		Excluded Row
	% A	%P	% A	%P	% A	%P	% A	%P	% A	%P	% A	%P
	δ10 nm						δ30 nm
Normalized	92.73	90.91	94.55	90.91	92.73	90.91	94.55	90.91	96.36	90.91	94.55	90.91
Q Norm	94.55	81.82	98.18	90.91	94.55	90.91	87.27	72.73	92.73	81.82	96.36	90.91
Baseline	100	100	100	100	100	100	96.36	100	90.91	90.91	98.18	100
Detrend	100	100	100	100	100	100	94.55	100	94.55	90.91	94.55	90.91
Deresolve	94.55	100	100	100	96.36	100	100	100	100	100	96.36	100
SNV	98.18	100	100	100	100	100	92.73	81.82	96.36	90.91	94.55	90.91
MSC	100	100	98.18	90.91	100	100	100	100	98.18	100	92.73	90.91
OSC	92.73	81.82	98.18	90.91	98.18	100	89.09	81.82	94.55	81.82	96.36	90.91
Savitzky Golay	100	100	100	100	100	100	94.55	90.91	94.55	90.91	96.36	90.91
1st Derivative	100	100	100	100	100	100	96.36	100	96.36	100	92.73	90.91
2nd Derivative	100	100	100	100	100	100	94.55	90.91	96.36	100	94.55	90.91
	δ60 nm						δ80 nm
Normalized	100	100	98.18	100	92.73	100	94.55	100	100	100	100	100
Q Norm	94.55	81.82	96.36	90.91	98.18	90.91	90.91	90.91	98.18	100	100	100
Baseline	98.18	90.91	98.18	100	98.18	100	96.36	100	100	100	100	100
Detrend	100	100	100	100	100	100	100	100	100	100	100	100
Deresolve	100	100	100	100	96.36	100	100	100	100	100	100	100
SNV	89.09	72.73	92.73	100	100	100	98.18	100	100	100	100	100
MSC	98.18	100	100	100	87.27	90.91	92.73	100	100	100	100	100
OSC	96.36	90.91	98.18	90.91	100	100	98.18	100	100	100	100	100
Savitzky Golay	100	100	100	100	98.18	100	98.18	100	100	100	100	100
1st Derivative	100	100	100	100	100	100	100	100	100	100	100	100
2nd Derivative	96.36	100	100	100	98.18	100	100	100	100	100	100	100
	δ120 nm						δ185 nm
Normalized	96.36	100	100	100	100	100	87.27	100	100	100	100	100
Q Norm	92.73	90.91	94.55	72.73	96.36	90.91	89.09	90.91	96.36	100	98.18	100
Baseline	100	100	100	100	100	100	85.45	100	100	100	100	100
Detrend	98.18	100	100	100	100	100	89.09	90.91	100	100	96.36	81.82
Deresolve	94.55	90.91	100	100	96.36	90.91	100	100	100	100	100	100
SNV	96.36	90.91	100	100	100	100	94.55	100	100	100	100	100
MSC	96.36	90.91	98.18	100	100	100	85.45	100	100	100	98.18	100
OSC	94.55	90.91	94.55	90.91	96.36	90.91	92.73	90.91	96.36	90.91	100	100
Savitzky Golay	87.27	81.82	100	100	96.36	90.91	90.91	100	100	100	100	100
1st Derivative	94.55	100	100	100	98.18	100	89.09	100	100	100	100	100
2nd Derivative	90.91	90.91	100	100	96.36	100	100	100	100	100	100	100

Discussion

The spectra of the virgin olive oil recorded at δ10 nm show two major bands with their maxima at around 310 and 680 nm as shown in Fig. 1. These peaks were previously identified by Sikorska et al. (2004, 2005, 2008) to be attributed to the presence of phenolic/tocopherol and the chlorophyll related compounds. These conclusions were further supported by the same authors in 2008 by comparison with respective standards. Appearance of new bands or splitting of existing bands is typically observed with increasing wavelength interval. Emission bands are present in the excitation region below 310 nm, 310–350 nm, 350–380 nm, and above 550 nm in spectra of virgin olive oils (Papoti and Tsimidou 2009). Similar spectral characteristics for virgin olive oil were reported by Poulli et al. (2005). At δ120 nm and δ185 nm the bathochromic shift of lower wavelength is further accentuated as three peaks could be observed at 440, 445 and 470 nm which at δ185 nm these three peaks shift to 510, 540, 555 nm respectively. Furthermore, at δ185 nm peaks observed at higher wavelengths split further and become more distinguishable as two maxima were observed at 645 nm and 700 nm. The distinct peaks observed at different δ nm cannot be attributed to one single compound. One of the pioneering papers published by Kyriakidis and Skarkalis (2000) showed that this intense peak was attributed to different forms of tocopherols present within olive oil and their corresponding oxidised derivatives.

From the results obtained it was shown that whilst the δ10 nm showed only emission bands at 310 nm and 670 nm, a bathochromic shift of the 310 nm peak was observed as the δ nm was increased. At δ30 nm two major peaks were observed, one at 335 nm and another one at 380 nm as highlighted through the use of quantile normalization and 2nd order derivatization which corresponds to same peak which was observed for the SEEFS obtained at 10 nm. This peak was attributed to the presence of tocopherol whilst the peak centered at 650 nm was attributed to the presence of chlorophyll compounds. Moving to δ60 nm three major peaks were observed, in the region 385–440 nm whilst at δ80nnm only two peaks were observed in the region 405 nm and 440 nm. However as highlighted by the different normalization processes including quantile normalization and SNV, two peaks were observed emitting at high wavelengths, 675 nm and 685 nm respectively rather than at a single peak as previously observed at δ10 and δ30 nm.

The identification of the emission bands relies mainly on comparison to the spectra of chemically pure fluorescent components. The short wavelength band in total fluorescence spectra, which covers the excitation region of 270–330 nm and 295–360 nm corresponding to the emission band at 280–310 nm in the synchronous fluorescence spectra, was assigned to tocopherols and phenols. This assignment has been confirmed by several observations (Sikorska et al. 2004, 2005, 2008; Papoti and Tsimidou 2009; Kyriakidis and Skarkalis 2000). On analysis of the different spectral pretreatment methods on the different SEEFS it was shown that EVOOs of Maltese origin had more intense peaks observed in the 395–445 nm region at δ60 nm and 645–680 nm as observed at δ120 and δ185 nm. These observations suggest that the EVOOs derived from the Maltese islands have a higher phenolic/ tocopherol and chlorophyll content. These results corroborate the results obtained using high performance liquid chromatography whereby it was shown that EVOOs of Maltese origin had a significantly higher concentration of secoiridoids (Lia et al. 2019).

During application of PLS-DA on the entire SEEF spectra as shown in Table 1, models obtained using δ30 and δ60 nm had the lowest performance under both cross-validation methods for most of the spectral pretreatments. On the other hand, it was found that SEEF obtained at δ80nm had the highest performance. In comparison, the study carried out by Dupuy et al. (2005) showed that SEEFs obtained at δ30nm using normalization as a sole spectral transformation was enough for the discrimination of virgin olive oil samples from five French registered designation of origins. Furthermore, in this study it was observed that the PLS-DA models obtained differed not only with SEEF δ nm but each SEEF δ nm had different model performance depending on the spectral pretreatment used. In the case of δ10 nm, it was observed that the model reaches optimal performance (100% accuracy and precision) using the 2nd order derived spectra. At δ80 nm it was found that Savitzky Golay, 1st and 2nd order derivatization were equally effective in predicting the geographical origin of EVOOs, as the models obtained using these functions obtained 100% accuracy and precision. For the δ120 nm it was observed that spectra pretreated using baseline and orthogonal signal correction functions reached 100% accuracy and precision were obtained. On the other hand, SEEFs obtained at δ185 nm reached optimal performance after deresolve and 1st order derivatization. Recent studies carried out by Kontzedaki et al. (2021) showed that the application of PLS-DA to fluorescence spectra obtained from olive oils obtained from various Greece regions had a 81.69% rate of correct classification. Studies published by Al Riza et al. (2021) showed that the application of fluorescence spectra in combination multivariate statistics were not only able to discriminate between different cultivars but also between geographical regions. Furthermore, studies carried out by Zaroual et al. (2020) showed a clear discrimination between EVOOs according to their geographic origin (96.72%) and variety (95.12%).

Variable selection carried out through the inspection of the VIP > 0.8 shown in Fig. 2 revealed that in the case of SEEFs obtained at δ10 nm using baseline, detrending, deresolve, MSC and SNV pretreatments, the selected variables were mainly concentrated in the 300–400 nm and 600–700 nm region corresponding to phenolic/tocopherol and pigment fluorescence. Variables selected in the SEEF obtained at δ30 nm showed that most of the spectral pretreatments were selected throughout the entire spectrum rather than being focused on fluorescent peaks. However, spectra pretreated using Savitzky Golay, 1st and 2nd order derivatisation, almost completely exclude variables in the 400–500 nm range where no fluorescent compounds were found. In the case of SEEFs obtained at δ60 nm and δ80 nm applying only detrended and baselined corrected spectra showed a more specific variable selection focused around 350–450 nm and 650–700 nm that correspond to major fluorescent peaks, the other spectral pretreatments had a rather non-specific variable selection. At δ120 nm and δ185 nm analysis of VIP’s revealed that even though the major fluorescent peak obtained was centered around 400–500 nm, in the majority of the spectral pretreatments, variables were also selected between the 600–700 nm which correspond to shouldering peaks of chlorophyll pigments and their corresponding derivatives.

Variable selection procedures showed a marked increase in the performance throughout all the spectral pre-treatments for all the SEEFs as shown in Table 2. The increase in performance is given in terms of % accuracy and % precision. LDA models using variables which had the highest and the lowest scoring coefficients obtained from the SLC-DA shown in Table 3 revealed that for SEEF obtained at δ10 nm the majority of the discriminate variables were found in the 650–750 nm range which is concordant to the results obtained using PLS. The removal of variables using the combination of two techniques greatly improves the modelling power via the removal of collinear and redundant variables. A similar overall trend was observed, however for most of the spectral pre-treatments a slightly lower performance was observed for the LDA models. From the results obtained it was further confirmed that the SEEF spectra obtained at δ10 nm had the best discriminatory power. SEEF spectra at δ10 nm after pre-treatment with SNV, MSC, Savitzky Golay, 1st and 2nd order derivation obtained the best model performance (100% accuracy and precision) when compared to all the SEEF spectra and their corresponding pre-treatment. These results further confirm that the SEEF obtained at δ10 nm have a higher discriminatory power when compared to the other SEEFs, as SEEF spectra obtained at δ10 nm after deresolve, SNV, MSC, 1^st and 2^nd order derivatisation enabled the correct classification of the samples used in both the training and in the validation test set. It was shown that variables selected in the range 300–450 nm and 510–650 nm had a higher modelling power for the Maltese EVOOs, whilst those obtained in the range 280–350 nm and 465-490 nm had a higher modelling power for modelling the non-Maltese EVOO group. Figure 3 showed that variables selected in the range 345–380 nm and 525–555 nm had a very high discriminatory power for SEEF spectrum obtained at δ10 nm after Savitzky Golay derivatisation. These ranges correspond to the tocopherol fluorescent peaks and shouldering peaks of chlorophyll pigments.

ANN models shown in Table 3 further corroborate the trends obtained using PLS on the entire spectrum without any form of feature selection. It was shown that SEEF obtained at δ10 nm, δ80 nm, δ120 nm and δ185nm had a higher ANN performance throughout all the cross-validation methods used, reaching 100% accuracy and 100% precision for most spectral pretreatments. From the results obtained it was shown that the performance of the ANN model is dependent not only on the SEEF and the data pre-processing used but also on the cross-validation method employed. In general, excluded row validation and fivefold cross-validation (CV-10) yielded models with a higher performance. This was attributed to the fact that whilst CV-10 and excluded row validation there is some form of control on the training and validation set, the 33% holdback is purely random, thus it increases the chance that the model is built using an uneven or skewed number of samples belonging to different classes. This results in a model which is inheritably biased towards one class and fails to identify samples of the other class when it comes to the validation data set, causing a lower % predictability. Nonetheless an overall higher performance was observed throughout all the SEEF spectra studied for the different pretreatments, independent of the cross-validation method used. This observation was attributed to the higher flexibility and modelling power of ANNs when compared to PLS. Like the PLS models SEEF spectra pretreated using detrending function and Savitzky Golay, 1st and 2nd order derived spectra had the highest ANN model performance throughout all the cross-validation methods used. The recent application of fluorescence spectroscopy conjunction with artificial neural networks and other machine learning algorithms have been successfully employed by Venturini et al. (2021), for the classification of EVOO, virgin olive oil, and lampante olive oil with an accuracy of 100%.

Conclusions

In conclusion, it was shown that SEEF spectra in conjunction with a number of chemometric methods, provided a cheap, fast and reliable way for the determination of the geographical origin of EVOOs, especially when it comes to discrimination of Maltese EVOOs from non-Maltese EVOOs. From the preliminary assessment using only unsupervised PCA models, only very few spectral pretreatments for different SEEF spectra managed to identify significant clustering, and such a method was deemed to be unsatisfactory when it comes to discrimination of geographical origin. Application of supervised methods of classification namely PLS-DA, ANN, and LDA were shown to be highly effective in classifying local and non-local EVOOs samples. The use of the variable selection methods significantly increased the effectiveness of PLS-DA models. Application of ANN, and LDA models were also shown to offer similar classification rates to PLS-DA models and thus corroborate the results obtained. Results showed that different SEEF spectra can greatly affect the discrimination of EVOOs. It was shown that independent of the chemometric technique used, SEEF spectra obtained at δ10 nm showed a higher model performance. It was shown that the most discriminate variables were those attributed to the different concentration of phenolic, tocopherol and chlorophyll compounds.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

SEEF

Synchronized excitation-emission spectrofluorometery

EVOO

Extra virgin olive oil

PLS

Partial least square analysis

LDA

Linear discriminate analysis

ANN

Artificial neural network

Author contributions

Frederick Lia, data acquisition, research paper conceptualization, methodology, software, validation, formal analysis, investigation, data curation, writing—original draft preparation, writing—review and editing Marion Zammit Mangion, data acquisition, methodology, software, validation, writing—review and editing, conceptualization, writing—original draft preparation, Claude Farrugia., conceptualization, software, writing—original draft preparation, formal analysis supervision and project administration.

Funding

This research was funded by the Malta Government Scholarships Post-Graduate Scheme for 2014 (MGSS-PG 2014).

Declarations

Conflicts of interest

All authors have read and agreed to the published version of the manuscript. The authors whose names are listed immediately below certify that they have NO affiliations with or involvement in any organization or entity with any financial interest (such as honoraria; educational grants; participation in speakers’ bureaus; membership, employment, consultancies, stock ownership, or other equity interest; and expert testimony or patent-licensing arrangements), or non-financial interest (such as personal or professional relationships, affiliations, knowledge or beliefs) in the subject matter or materials discussed in this manuscript.

Consent for publication

The authors, give their consent for the publication of identifiable details, which can include photograph(s) and/or videos and/or case history and/or details within the text to be published in the Journal of Food Science and Technology. The authors confirm that they have seen and been given the opportunity to read both the Material and the Article to be published by Springer.

Availability of data and material

Due to the sensitive nature of the data raw data will provided on request directly from the corresponding author.