One of the most difficult tasks in analytical chemis- try of wine is to identify its authenticity and geographi- cal origin. Single quality assessment parameters are not sufficient to determine whether the product conforms to its labels. To establish the authenticity and geographical origin of wines, as well as changes occurring in case of their adulteration, analytical approaches are being de- veloped that aim to determine the mineral and isotopic composition, study spectral characteristics, and identify phenolic and volatile compounds using various methods of analysis [1–2]. The identification of authenticity and origin criteria is based on obtaining a large amount of data and its processing by chemometric methods, which reveal hidden relations between the wine’s components [1–9]. The combination of modern data analysis tools

with the capabilities of chemometric methods ensures higher accuracy in identifying the geographical origin of wines. The information on the elemental composi- tion of wines can be used to both control the technologi- cal process and, in combination with chemometric data processing methods, establish the origin of wines [10, 11]. For example, wines produced in various regions of Europe differ quite markedly in the metal content [12], which makes it a good criterion for identifying their geo- graphical origin (Table 1).

Natural variability of  wine  quality  is  determined by the grapes growing conditions, such as the climate, the microelement composition of the soil, the technolo- gy of growing grapes, the period of grape harvest, etc. The mineral composition of wines can be influenced by various factors (soil, climate, relief, etc.); therefore, for

Copyright © 2019, Chernukha et al. This is an open access article distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), allowing third parties to copy and redistribute the material in any medium or format and to remix, transform, and build upon the material for any purpose, even commercially, provided the original work is properly cited and states its license.

Table 1. The metal content of wines in different countries [12]

Element                                                                                           Element content, mg/dm3

identification purposes, many researchers study those ele- ments which are least dependent on external factors in a given geographical area [3–6, 8, 9, 13, 14]. For example, some authors [13] use Sr, Mn, Mg, Li, Co, Rb, B, Cs, Zn, Al, Ba, Si, Pb, and Ca.

The content of metals in wines is  widely  diffe- rent: 10–1000 of macroelements (Ca, K, Na, and Mg), 0.1–10 mg/dm3 of minor elements (Al, Fe, Cu, Mn, Rb, Sr, and Zn), and 0.1–1000 μg/dm3 of trace elements (Ba, Cd, Co, Cr, Li, Ni, Pb, V, etc.) [12]. Therefore, the prob- lem of ascertaining the microelement “image” of grapes is of practical, as well as scientific, interest [14–18].

In cases when wines from different grape varieties have certain organoleptic similarities, for example, co- lour or astringent, sour taste, it is important to be able to identify the grape variety from the microelement com- position of the wine [19]. In fact, the task comes down to establishing the grape variety and geographical origin based on the content of microelements in a sample of un- blended wine.

The purpose of this work was to study a possibility of identifying the authenticity and geographical origin of red wines, namely Cabernet and Merlot varietal wines, based on multi-element analysis with STATISTICA Neu- ral Network.

STUDY OBJECTS AND METHODS

The study used 144 samples of Cabernet (76) and

wineries are located in different geographic zones (sub- zones) of Krasnodar Region: the South-Piedmont zone, the Black Sea zone, the Anapa subzone, and the Taman subzone. The wines were provided by the manufacturers or purchased in retail stores.

The main vineyards of Krasnodar Region are located in five cultivation areas: Temryuk (the Taman Peninsu- la, the Taman subzone), Anapa (the Anapa subzone), the Black Sea zone (Gelendzhik and Novorossiysk), Krymsk (the South-Piedmont zone), and Novokubansk. The fre- quency distribution of Cabernet and Merlot samples by zone and variety is shown in Table 2

The elemental composition of the wine samples was established by atomic emission spectroscopy with induc- tively coupled plasma using iCAP-6000 (Thermo Scien- tific). The operating conditions of the spectrometer were optimised to detect 20 elements (Li, Na, Mg, Al, K, Ca, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Rb, Sr, Cd, Ba, and Pb). The most sensitive analytical lines were used for most of the metals, with the exception of Al, V, Ca, Mg, and Sr, for which alternative lines were chosen due to spectral overlays. For some macroelements, we needed to reduce the signal intensity. When optimising the con- ditions for element detection, we studied how the ope- rating characteristics (generator power, argon flow rate) affected the analytical signal of elements in the model

Table 2. Wine sample frequencies by zone

Merlot (68) varietal dry wines produced from 2012 to                                                                                                               

2015 by the main wineries in Krasnodar Region: ZAO

Variety                 Two-entry table of frequencies by zone          

Zaporozhskoye, OOO Kuban-Vino, OAO APF Fanago- ria, OOO APK Millstream Black Sea Wines, ZAO AF

Taman subzone

Anapa subzone

South-Pied- mont zone

Black Sea zone

Total

Kavkaz,  ZAO  Abrau-Durso,  ZAO  APK  Gelendzhik,

ZAO AF Myskhako, OOO Firma Somelye, ООO AF Sauk-Dere,  and  ООО  Soyuz-Vino  (Table  2).  Theses

Cabernet    28             17             13                  18             76

Merlot          33            23             12                  0               68

Total           61            40             25                  18              144

and sample solutions. We also investigated the mutual influence of micro- and macroelements, as well as back- ground components, of the samples prepared for analysis in the model solutions containing variable amounts of the elements. The quantification of metals was carried out by diluting the wine samples, taking into account the data obtained [5, 14–18, 20, 21].

The following reference standards were used to study the test samples: GSO 7780-2000  (Li),  GSO 8062-94 (Na), GSO 7767-2000 (Mg), GSO 7854-2000 (Al), GSO (K), GSO 7772-2000 (Ca), GSO 7205-95 (Ti), GSO (Cr), GSO 8056-94 (Mn), GSO 8032-94 (Fe), GSO 7784-2000 (Co), GSO 7785-2000 (Ni), GSO 7836-2000 (Cu), GSO 8053-94 (Zn), GSO 7035-93 (Rb), GSO 7783-2000 (Sr), GSO 7874-2000 (Cd), GSO 7760-2000 (Ba), and GSO

7778- 2000 (Pb). All the reagents used in the work were of chemically pure (C.P.) grade.

The chemometric analysis was performed using STA- TISTICA Neural Networks [22].

RESULTS AND DISCUSSION

The analysis of average element contents (Tables 3, and 4) showed a significant dependence of wine compo- sition on the grape variety and place of origin. For exa-

mple, the samples from the Anapa subzone had a high content of Fe, those from the South-Piedmont zone were rich in Ba, Ti, and V, whereas the Taman wines were abundant in Na, Mg, and Rb. The Cabernet wines had significantly different contents of many elements. For example, the Cabernet samples from the South-Pied- mont zone contained the lowest concentrations of Li, Na, Al, Ca, Fe, and Sr, while the Merlot samples from the same zone had the lowest content of Al, Ca, Fe, and Li. As a rule, standard deviations did not exceed half of the average values. This suggests a small variation in the concentrations of elements, which means that an ave- rage value is a relevant characteristic of metal content in wine. The exceptions are Cu, Li, Ni, Na, Rb, and Ti; however, standard deviations exceeded the averages only in three cases: Cu (Cabernet, Anapa subzone) and Ni (Cabernet and Merlot, South-Piedmont zone).

Previously, we applied traditional statistical methods of discriminant analysis and classification trees to con- struct  probabilistic-statistical  models   that   allowed us to identify the varietal and regional origin  of  the same group of red wines using multi-element analysis data [23]. This study looked at a possibility of determi-

Table 3. Average elemental content and standard deviations (s.d.) in Cabernet samples from various geographical zones of Krasnodar Region, µg/dm3

                  Anapa       South-Piedmont   Black Sea    Taman     

Table 4. Average elemental content and standard deviations (s.d.) in Merlot samples from various geographical zones of Krasnodar Region, µg/dm3

Table 5. Average elemental contents and standard deviations (s.d.) in wine samples, µg/dm3

ning the grape variety and geographical origin using STATISTICA Neural Networks, followed by a compara- tive analysis.

To select a number of elements as predictors of neu- ral network classification models, we used a Spearman’s nonparametric correlation coefficient that characterised the correlation between the names of wine samples, the region of grape origin, and the concentrations of trace elements in the samples. In particular, the elements with the largest statistically significant correlation links be- tween the names of regions and wines (Fe, Mg, Rb, Ti, and Na) were selected as predictor variables.

In Table 5, which shows average elemental contents with standard deviations in both wine varieties from different regions, we can see some significant differenc- es in the average values – the deciding factor for buil- ding classification models with neural networks. Most distinctly these differences are visualised by means of graphs. Fig. 1, for example, shows some box plots dis- playing Mg content in the Cabernet and Merlot wines from various regions. The box plots present ranges of values of a selected variable separately for groups of ob- servations defined by the values of a categorical varia- ble. The rectangles depicted around the midpoints (or squares) represent selected ranges of variation, for exam- ple, the standard error (the ratio of the standard deviation to the square root of the sample size). The segments with their ends outside the rectangles also reflect ranges of variation (average ± 1.96 × standard error). The diagram shows that the average values of Mg content, together with variation values, differ significantly between both the regions and the grape varieties.

As in [23], we were not able to build adequate neural networks that would allow us to identify the grape vari- ety and region of origin from the concentrations of se- lected elements. Therefore, the problem was divided into two parts. First, networks were built to predict the grape variety from the concentrations of Fe, Mg, Rb, Ti, and Na. Then, based on the variety predicted (qualitative pre- dictor) and the same set of elements (quantitative predic- tors), further networks were built to determine the place of grape origin. After assessing their predictive proper- ties (productivity, number of classification errors, etc.), we selected the best network. Productivity is a percentage

of correctly classified wine samples, with 100% taken as

maximum. The higher the productivity, the more accurate the prediction. To improve predictive accuracy, the sam- ples were divided into three groups: training, control, and validation sample. The most important were the values of adequacy criteria in the test set. By combining various network options, we tried to create a network with the best predictive capabilities; therefore, at each stage of the process, the number of networks was different.

Building a neural network to establish the varietal origin of wine. The program divided 144 wine samples into three groups: training set (102), control set (21), and test set (21). The productivity of the best network (MLP 5-5-2), selected out of 50, had high values of 99.02%, 90.48%, and 95.30% in the training, control, and test sets, respectively. MLP 5-5-2 is a combination of letters and numbers that represents a topology of a multilayer perceptron. The letters stand for the type of a neural net- work, a multilayer perceptron (MLP); the first numer-

Box plots for Mg content in wines

180                Cabernet                                    Merlot

160

140

120

100

80

South-Piedmont

Black Sea

Taman

Anapa

South-Piedmont

Black Sea

Taman

60

 

 

 

 

 

 

Zone

average

*

average ± standard error average ± 1.96 × standard error

 

Fig. 1. Box plots displaying Mg content in Cabernet and Merlot wines.

Table 7. Network sensitivity analysis

 Control              6.86            3.58          5.63        3.12        2.64   

Fig. 2. Neural network to determine the variety of red wines.

al (5) refers to the number of predictor variables in the model, a sum of quantitative predictors and qualitative predictor values; the second (5) and the third (2) nume- rals refer to the numbers of hidden and output neurons, respectively.

The network topology is shown in schematic form in

Fig. 2, where we can see five entries of predictor vari- ables X ; five hidden neurons Y ; two output neurons rep-

The network sensitivity can be used to estimate a contribution of each predictor to its predictive proper- ties: in our case, a contribution of the elements to the classification model. The sensitivity values (see Table 7) indicate a decreasing sequence of Fe, Rb, Mg, Na, and Ti, which represents their contributions to the predictive properties of the network.

Building a neural network to determine the re- gional origin. The possibility of predicting the wine variety based on five microelements made it realistic to create a neural network to identify the place of grape origin using the trace elements of Fe, Mg, Rb, Ti, and Na and the varieties of Cabernet and Merlot.  In  the same way, the program divided 144 wine samples into three groups: training set (102), control set (21), and test

i                                                               j

Merlot varieties, as well as connections between them in the form of weights W , W .

set (21). The best out of 18 networks (MLP 7-9-4) had

productivity values of 100%, 80.95%, and 100% in the training, control, and test sets, respectively.

ij         jk

Table 6 shows the frequencies of correctly and in-

correctly classified wines in the sample sets. As we can see, one Merlot sample from the training and the test sets and two Merlot samples from the control set were erroneously classified as Cabernet. All Cabernet samples were correctly identified in all the sets. The total num- ber of erroneously classified samples was four out of 144 (app. 2.8%), i.e. the neural network identified 97.2% of the wine samples correctly. In [23], by comparison, the classification tree with seven terminal vertices only once misclassified a Merlot sample as a Cabernet, based on the concentration of seven microelements, i.e. 99.3% of the training sample was identified correctly.

As  can  be  seen  in  Table  8,  all  the  wine  samples

(100%) from the Black Sea zone were classified by the network correctly. The next high accuracy area was the Taman subzone with 100%, 85.71%, and 100% of cor- rectly classified samples in the training, control, and test sets, respectively.  The  lowest  accuracy  was  observed in the Anapa subzone: 100%, 71.43%, and 100%, re- spectively. The total number of misclassified  samples was four out of 144 (app. 2.8%), i.e. the neural network