- Email: [email protected]
Classification of red wine based on its protected designation of origin (PDO) using Laser-induced Breakdown Spectroscopy (LIBS)
Talanta 158 (2016) 185–191
Contents lists available at ScienceDirect
Talanta journal homepage: www.elsevier.com/locate/talanta
Classification of red wine based on its protected designation of origin (PDO) using Laser-induced Breakdown Spectroscopy (LIBS) S. Moncayo a, J.D. Rosales a, R. Izquierdo-Hornillos a, J. Anzano b, J.O. Caceres a,n a b
Department of Analytical Chemistry, Faculty of Chemical Sciences, Complutense University, 28040 Madrid, Spain Laser Laboratory, Department of Analytical Chemistry, Faculty of Sciences, University of Zaragoza, 50009 Zaragoza, Spain
art ic l e i nf o
a b s t r a c t
Article history: Received 27 February 2016 Received in revised form 16 May 2016 Accepted 21 May 2016 Available online 24 May 2016
This work reports on a simple and fast classification procedure for the quality control of red wines with protected designation of origin (PDO) by means of Laser Induced Breakdown Spectroscopy (LIBS) technique combined with Neural Networks (NN) in order to increase the quality assurance and authenticity issues. A total of thirty-eight red wine samples from different PDO were analyzed to detect fake wines and to avoid unfair competition in the market. LIBS is well known for not requiring sample preparation, however, in order to increase its analytical performance a new sample preparation treatment by previous liquid-to-solid transformation of the wine using a dry collagen gel has been developed. The use of collagen pellets allowed achieving successful classification results, avoiding the limitations and difficulties of working with aqueous samples. The performance of the NN model was assessed by three validation procedures taking into account their sensitivity (internal validation), generalization ability and robustness (independent external validation). The results of the use of a spectroscopic technique coupled with a chemometric analysis (LIBS-NN) are discussed in terms of its potential use in the food industry, providing a methodology able to perform the quality control of alcoholic beverages. & 2016 Elsevier B.V. All rights reserved.
Keywords: Laser Induced Breakdown Spectroscopy Wine Neural networks Protected designation of origin LIBS PDO
1. Introduction The certification of the protected designation of origin (PDO) is one of the most important parameters to be controlled in order to protect the production and origin of agroalimentary products. Since the introduction of European regulations control [1] on this matter, many wine companies have adopted different strategies on the confirmation of wine authenticity improving the PDO controls. The type of grape, geographical origin, harvest, and vintage are parameters that determine the quality of wine products. Nevertheless, these indicators are not easily recognizable for the consumers and the PDO is considered as a single and unambiguous sign of quality that companies use to promote their brand in the market influencing the customer final decision [2]. The wine industry has seen an important growth in the last decade associated with an increase in the wine consumption. This is a strategic sector due to its importance in economic, environmental and social terms, as well as its importance representing the country’s image abroad. Since 2008, the wine industry has put an important effort in the control of counterfeiting of wines with the objective of protecting the trade-mark quality wines and to prevent their n
Corresponding author. E-mail address: [email protected] (J.O. Caceres).
http://dx.doi.org/10.1016/j.talanta.2016.05.059 0039-9140/& 2016 Elsevier B.V. All rights reserved.
illegal adulteration [3,4]. The wine adulteration consists of the addition of any substance to the natural wine, which changes its composition and may occur in many different forms. The greater part of the adulteration consists of addition of water and sugar, mixing with lower quality wines and label replacing [3]. The two main constituents of wine are water (81%) and ethanol (between 11% and 15%). Two types of flavonoids, the anthocyanins and flavanols, are key compounds for color and astringency, being responsible for the organoleptic properties and quality of wine. Other organic compounds in small amounts, such as acids, alcohols, phenols, nitrogenous compounds and inorganic substances represent the remaining 7%, making wine a complex sample and difficult to analyze [5,6]. The sensorial analyses together with chemical assays, and mineral content analysis may not be adequate for determining the PDO of wine [3]. Chromatographic techniques [7–9] require to conduct separate analysis of each component in the wine being slow and expensive process. The identification of grape variety by means of isotopic analysis [4,10], nuclear magnetic resonance (NMR) [11,12] or ADN/aRNA [13] techniques are generally used, due to their capacity of generating a fingerprint of wine providing the identification. Although these techniques produce accurate results, a large amount of sample and the use of expensive consumables are required, increasing the cost and duration of the analysis.
186
S. Moncayo et al. / Talanta 158 (2016) 185–191
This work evaluates the Laser Induced Breakdown Spectroscopy (LIBS) technique for the discrimination and the determination of geographical origin of red wines. LIBS technique is based on the interaction of a laser beam with a material target generating a plasma, the emission of the plasma contains spectroscopic information of excited atoms and ions present in the sample and reflecting its elemental composition [14]. Although there is a loss of molecular information in plasma, LIBS has provided excellent results in the identification of samples with complex matrixes [15,16]. LIBS provides spectral fingerprints characteristic of each sample based on the composition of wine The presence of mineral elements in wine is generally related to soil composition, grape variety, climate conditions, yeasts and winemaking [8]. The combination of LIBS with supervised classification methods such as NN has already shown successful results in many areas of knowledge for sample classification [17–20]. One of the major advantages of LIBS is that it does not require sample preparation, providing an economic and fast analysis, however in some cases avoiding a sample preparation goes in detrimental of the technique limiting its analytical performance. The change of the physical state of the sample transforming the liquid into solid has already been described as sample preparation [21]. Although the liquid-to-solid process produces an increase in the time analysis and an alteration of the original chemical composition, significant improvements such as the increase of the ablation rate, higher plasma temperature and electron density as well as a better laser-to-solid interaction has been observed in literature [22–24]. Moreover, avoiding the inherent drawbacks of working with liquids such as splashing and surface ripples produces lower limit of detection, better repeatability and sensitivity [25,26]. Different liquid-to-solid matrix conversion protocols have been described in the literature involving precipitation, filtering and pellets formation procedures [25,27,28]. Herein, the transformation of the liquid wine sample into gels by adding a natural collagen and its subsequent dried in an air assisted oven has been used as a new sample preparation protocol. The aim of this work was to identify the adulteration of wines collected from Spanish local markets and evaluate the capacity of LIBS coupled with NN to detect the PDO of wines with negligible compositional and spectral differences and to improve the recognition capacity of extremely similar samples that have fewer physical and spectral differences between them.
2. Materials and methods 2.1. Wine samples Thirty-eight Spanish red wines from eleven protected designation of origin, three foreign red wines and four table wines were purchased in retail stores. These wine samples were selected to cover the main Spanish wine regions (Fig. 1) including La Mancha, Ribera de Duero, Rioja, Valdepeñas, Vinos de Madrid, Cariñeña, Ribeiro, Ribera del Guadiana, Navarra and Somontano and Toro. Moreover a German, French and Italian wine was also included in the study. Most of the wines included in the study were elaborated with Tempranillo grapes, although Cabernet Sauvignon, Garnacha, Tinta de Toro and Shyrah were also considered. All samples belong to the 2011 vintage and were not affected by ageing period (young wine). Table 1 shows sample information including sample ID, commercial brand and type of grapes. 2.2. Sample preparation A methodology based on the formation of a gel of wine using a commercial collagen was applied. 50 mL of wine sample were
introduced into a beaker and 1 g of collagen gel was added and dissolved in the wine sample. 2.4 mL of this solution were allowed to stand 15 min until the formation of a gel on a square petri dish of 4 4 cm. Then, samples were introduced in a forced ventilation oven at 3572 °C during 12 h to evaporate the water, obtaining a dry solid. The final sample was completely flat with a thickness of approximately 0.35 mm. Fig. 2 shows an example of dry gel and the craters formed by single laser shots. In this process not only LIBS analysis was simplified but also pre-concentration of the sample (pre-concentration factor of 1:5) is performed allowing an improvement in the limits of detection. The gels for all samples were prepared at the same time to maintain the same conditions and avoiding the degradation and oxidation of the wine components. 2.3. LIBS set-up The LIBS technique and the methodology used in the present work together with the most significant experimental conditions have been previously described [29]. Thus, only the experimental conditions relevant to this study are presented here. LIBS measurements were obtained using a Q-switched Nd: YAG laser (Quantel, Brio model) operating at 1064 nm, with a pulse duration of 4 ns full width at half maximum (FWHM), 4 mm beam diameter and 0.6 mrad divergence. Samples were placed over an X–Y–Z manual micro-metric positionator with a 0.5 mm stage of travel at every coordinate to ensure that each laser pulse impinged on a fresh position. The laser beam was focused onto the sample surface with a 100 mm focal-distance lens, producing a spot of 100 mm in diameter. The best signal-to-background ratio was achieved at 42 mJ of pulse energy with a repetition rate of 1 Hz. The laser crater profile was measured by means of a confocal microscope after laser pulse irradiation on a fresh position. A narrow crater was created with a diameter of 450 mm and 140 mm in depth. Emission from the plasma was collected with a 4-mm aperture, and 7 mm focus fused silica collimator placed at 4 cm from the sample, and then focused into an optical fiber (1000 mm core diameter, 0.22 numerical aperture), coupled to a spectrometer. The spectrometer system was an EPP2000, StellarNet (Tampa, FL, U.S.A.) with a gated CCD detector. A grating of 300 l/ mm was selected; a spectral resolution of 0.5 nm was achieved with a 7 mm entrance slit. The wavelength range used was from 200 to 1000 nm. Therefore, 2048 data points were recorded for each sample. The detector integration time was set to 1 ms, to prevent the detection of bremsstrahlung, the detector was triggered by a 2 μs delay time between the laser pulse and the acquired plasma radiation using a digital delay generator (Stanford model DG535). The spectrometer was computer-controlled using an interface developed in Matlab. 2.4. LIBS analysis Wine samples were measured directly in air at room condition. Each LIBS spectrum was acquired from a single shot measurement. A total of 100 spectra were recorded for each wine sample by moving the sample stage about 0.25 mm to expose a fresh portion of the sample surface and avoiding areas irradiated by previous shots. Only in the case of the M1, D1, R1, V1 and VM1 samples, four data sets of 100 spectra were obtained: the first data set (training library) was used to calculate the model of the NN; whereas the last three data sets (replicate libraries) were used for validation purpose. In order to avoid data variations due to changes in the laser pulse energy, each spectrum was normalized by the intensity of one specific spectral line, i.e., K (I) 766.49 nm [30]. The spectral information of each sample was obtained in less than 2 min considering the integration time of the spectrometer and the frequency of laser pulses that was fixed to 1 Hz.
S. Moncayo et al. / Talanta 158 (2016) 185–191
187
Fig. 1. Geographical distribution of Spanish wine regions. The dashed parts represent the Protected Denomination of Origin of the Spanish analyzed wines.
Table 1. Samples used in the study. PDO
Sample ID
Commercial brand
Grape variety
Vintage
La Mancha
M1 M2 M3 M4 M5 D1 D2 D3 D4 D5 R1 R2 R3 R4 R5 V1 V2 V3 V4 V5 VM1 VM2 VM3 VM4 VM5 CR RB RG NV SM TR TW1 TW2 TW3 TW4 CH DR VC
Libertario Vereda Mayor Don Lucio Fidencio Monte Don Lucio Dehesa Valpincia Barón de Santuy Mayor de Castilla Sangre de Castilla Viña Espolón Antaño Solar Viejo Barón de Urzande Castillo de Albali Vega del Cega Calle Real Viña Albali Señorío de Los Llanos Tanis Puerta de Alcalá Vega Madroño Alma de Valdeguerra Puerta de Hierro Jesús Díaz Castillo de Aguaron Pazo 5 Viñas Diácono MonteSierra Cermeño Conde Noble Don Simon Eroski Viñas Altas Corte Alle Mura Dornfield Cimarosa
Tempranillo, Garnacha Tempranillo Tempranillo Tempranillo, Garnacha Cabernet Sauvignon Cabernet Sauvignon Tempranillo Tempranillo, Cabernet Tempranillo Tempranillo Tempranillo, Garnacha Tempranillo, Garnacha Tempranillo, Garnacha Tempranillo, Garnacha Tempranillo Tempranillo Tempranillo Tempranillo, Cabernet Tempranillo Tempranillo Tempranillo, Syrah Tempranillo, Garnacha Tempranillo Tempranillo Tempranillo, Syrah Garnacha, Syrah Tempranillo, Cabernet Tempranillo, Garnacha Tempranillo, Garnacha Tempranillo, Cabernet Tinta de Toro Mixture Mixture Mixture Mixture Sangiovese Dornfelder Cabernet Sauvignon
2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011
Ribera del Duero
Rioja
Valdepeñas
Vinos de Madrid
Cariñena Ribeiro Ribera del Guadiana Navarra Somontano Toro Table wine
Chianti Dornfelder Valle Central
188
S. Moncayo et al. / Talanta 158 (2016) 185–191
Fig. 2. (a) Sample after preparation. (b) Dry gel of wine sample with visible laser spots.
Table 2. Wavelength range considered for wine analysis. Element Wavelength range (nm) Mg Ca Na Hα K
271–291 311–340, 390–429, 513–524 581–598 640–675 760–783
2.5. Wavelength range selection In case of LIBS spectra the intensity value at each wavelength is referred to as variable. Each spectrum was composed by 2048 variables taking into account the CCD pixels. For the NN analysis a reduction in the number of variables used as inputs was done. Reducing the number of variables have shown important advantages in a classification process, enhancing the robustness of the models and being more efficient in the discrimination process, without a loss of meaningful information. Since the organic nature of the gel may affect C, H, N, O signals, the main emission lines from the mineral elements of wine, Mg, Ca, K, and Na, were selected in seven wavelengths ranges resulting into 355 variables (Table 2). It is important to note that emission lines for these elements in the selected wavelengths intervals were not observed in the pure gel. The selection of several ranges may not describe the data fully, however, it describes the PDO trend instead of the particular wine trend. The reduction of variables decreases the discrimination power of the NN model but increase the ability to generalize.
complexity [19,31]. The NN classification model is estimated by calibrating with a set of reference samples to calculate the optimum weights and biases that better represent the given classes. NN training was achieved by applying the BP algorithm based on the conjugate gradient method, one of the general-purpose second-order techniques that helps minimize the goal functions of several variables. Second order indicates that such methods use the second derivatives of the error function, whereas a first order technique, such as standard back-propagation, uses only the first derivatives. To determine when the training should be stopped, an early stopping criteria based on performance improving (error rate) of the validation set. The number of epochs was not relevant in this case. To avoid an overfitting of the NN model, the learning process was repeated until a minimum of the mean square error (MSE) of the verification data, defined in Eq. (1), was reached:
MSE =
1 N
N
∑K
( rk − yk )2
(1)
where N, yk, and rk are the number of input data, the response from each output neuron, and the observed output response, respectively. Two hidden layers were used to calculate the model considering the ratio between sensitivity and generalization ability. The output layer (classification result) is comprised of J neurons (where the J¼ number of reference samples used) for estimating the similarity between the reference sample spectra and the testing sample spectra. A detailed description of the calculation process is provided in the literature [32,33]. Once the calibration is done, the weights are frozen and the NN model is validated with a test data set (samples not included in the calibration) providing a prediction result. Although the data matrix can be considerably large, the computation time for training the NNs was always below 10 s
2.6. Artificial Neural Networks (ANN) model 2.7. Classification and validation procedures A multilayer perceptron, feedforward, supervised neural network was used in this classification process [8]. The main advantage of NN is the capacity to model nonlinear effects, which are always present in LIBS spectra. In our case the NN topology consists of several neurons arranged in three different layers (input, hidden and output). The connections are controlled by a weight that controls the output of the neuron before inputting its numerical content to a neuron in the next layer. This topology has been widely used to model systems with a similar level of
An appropriate classification system must satisfy three basic validation procedures. Firstly, it must be able to classify all wine samples included in the training (sensitivity test), for this the NN model is trained with only one wine for each PDO and the ability of the model to classify replicates of the same wine is assessed, which is considered as an internal validation of the model. Secondly, it must provide a high generalization ability (generalization test), in other words the ability of the model to classify other
S. Moncayo et al. / Talanta 158 (2016) 185–191
wines belonging to one of the PDOs but not included in the training. This is considered as an external validation of the model, since the wine was not presented to the model previously, but is related to the training spectra by its PDO. Finally, the model must be robust. In this case it should be able to classify a wine that does not belong to any PDO as unassigned and not classify it as a trained PDO, therefore, there is no relation with the training spectra being completely independent of the model and this procedure is named as independent external validation [19].
3. Results and discussions Thirty-eight wine samples and a minimum of two replicates for each of the PDO covering a spectral window from 200 to 900 nm were analyzed. Fig. 3(a) shows a typical LIBS spectrum of a wine sample and the wavelength ranges selected as inputs to the NN have been highlighted in gray color. The observed molecular emission bands are identified using the spectroscopic information tabulated in NIST Atomic Spectral Database [30]. Lines emission for Mg, Ca, K, Na, Hα, C and molecular species such as swan carbon system, among others, are mainly present in the LIBS spectra. A visual comparison of spectrum emission lines from pure gel, M1, D1, R1, V1, VM1 wines and pure gel sample assigned as I, II, III, IV, V, VI and VII, respectively are shown in Fig. 3(b). Emissions in all the spectra from analyzed wine samples correspond to the same elements showed high similarity among distinct LIBS spectral fingerprints. To consider a wine sample correctly classified, the prediction of the model must match with the actual class by an arbitrary threshold of Spectral Correlation (SC) (Eq. (2)) higher than 90% and less than 20% to the other classes, otherwise is considered incorrectly classified. A wine sample was classified as unclassified when the SC was less than 80% for all classes.
SC =
100 N
∑ δi i
(2)
where δi is the number of spectra classified correctly and N is the total number of spectra.
189
3.1. Sensitivity tests For the training of the NN model, the training library samples M1, D1, R1, V1 and VM1 were used, producing an input matrix with 355 rows (number of variables) and 500 columns (5 100 spectra of each PDO). Once the learning is done the NN model is able to classify new spectra in a class. For the first validation procedure, three replicates of the training wines were prepared and measured at the same condition, generating three new data sets used to validate the model. Although, the same wine is considered to produce the replicates, each replicated was prepared in different gels, assessing possible sample preparation differences. These data sets have not been shown in the NN model previously and therefore the spectra are different to those that were used for the training. Table 3 gives the results of the sensitivity analysis for NN prediction of wine replicates. The model is considered sensitive if the correct classification is high and the misclassification rates are low. Based on the results, it is noted that a high sensitivity is achieved for all replicates with an average of 99.2% rate in the correct classification with a negligible rate of incorrect classification and misclassification. Therefore, the model was able to classify new spectra belonging to the samples used for training despite of the intrinsic sample-to-sample variability in the preparation of each gel. 3.2. Generalization ability test The validation of the model considering replicates is not enough in some cases because they cannot be seen as independent observations. Therefore, other validation procedures must be performed, ascertaining the generalization ability of the model is an appropriate way to do it. In this second validation the same NN model as discussed in the previous section was used. To estimate the generalization ability of the model four different wine brands belonging to the PDOs used in the training were tested to obtain the prediction result. A high rate in the correct classification and low rate of incorrect classification and misclassification was sought, which means that the model was able to generalize and correctly classify other spectra than those with which it was trained. Table 4 gives the results of the generalization test. An average of 98.6% of correct classification was achieved without any wine unclassified or misclassified and thus all wine samples were correctly assigned to their respective classes showing a high generalization ability and evidence the lack of model overfitting. 3.3. Robustness test
Fig. 3. (a) LIBS spectrum of Rioja PDO with the assignation of emission lines and selected the wavelength ranges selected as inputs to the NN have been highlighted in gray color. (b) Typical spectrum of (I) dry gel (II) M1 (La Mancha: Libertario) (III) D1 (Ribera del Duero: Dehesa) (IV) R1 (Rioja: Viña Espolón) (V) V1 (Valdepeñas: Vega del Cega) (VI) VM1 (Vinos de Madrid: Puerta de Alcalá) within the spectral range of 200–1000 nm.
To assess the robustness of the NN model, thirteen wine samples were measured and introduced into the NN model evaluating the response of the model to detect samples completely external and independent of the training set. Six wines with Spanish PDO, four table wines and three foreign wines were considered. A robust model implies a high capacity to detect unknown samples “correctly as unknown”, without diminishing the prediction accuracy of the known samples. Table 5 presents the NN model classification results when completely unknown samples were introduced to the model. All wine samples were correctly classified as unknown to the model. None of the samples were classified as belonging to other classes, providing a high robustness. Only in case of TW3 a Spectral Correlation of 70% was obtained in the La Mancha class which can be related to the fact that the table wines are elaborated with mixed variety of grapes, in many cases, those grapes are collected in the La Mancha region. However, the SC is lower than the threshold fixed and therefore classified as unknown to the model. This highlights the robustness of NN in dealing with samples of unknown classes.
190
S. Moncayo et al. / Talanta 158 (2016) 185–191
Table 3. Classification results to sensitivity test (internal validation). Predicted Group Membership (Spectral Correlation, %) Sample
La Mancha
Ribera del Duero
Rioja
Valdepeñas
Vinos de Madrid
Unclassified
M1-replicate 1 M1-replicate 2 M1-replicate 3 D1-replicate 1 D1-replicate 2 D1-replicate 3 R1-replicate 1 R1-replicate 2 R1-replicate 3 V1-replicate 1 V1-replicate 2 V1-replicate 3 VM1-replicate 1 VM1-replicate 2 VM3-replicate 3
100 99 99 0 0 0 0 0 0 0 0 1 0 0 1
0 0 1 100 98 100 0 1 1 0 1 1 0 0 0
0 1 0 0 0 0 100 99 98 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 100 99 98 0 0 0
0 0 0 0 0 0 0 0 1 0 0 0 100 99 99
0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
Table 4. Classification results for the generalization ability test. Predicted Group Membership (Spectral Correlation, %) Sample ID La Mancha Ribera del Duero
Rioja Valdepeñas Vinos de Madrid
Unclassified
2M2 2M3 2M4 2M5 2D2 2D3 2D4 2D5 2R2 2R3 2R4 2R5 2V2 2V3 2V4 2V5 2VM2 2VM3 2VM4 2VM5
1 0 0 0 0 1 1 0 100 98 97 97 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 1 0 0
98 100 98 99 2 0 1 0 0 1 1 0 0 2 1 0 1 0 0 0
0 0 1 1 98 99 98 100 0 0 1 2 0 1 0 0 0 0 0 0
1 0 1 0 0 0 0 0 0 0 0 1 100 97 98 100 1 0 1 0
0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 98 99 98 100
Table 5. Classification results to the robustness test (external validation).
The results were satisfactory and the proposed methodology was used for the analysis of red wines. The wavelength ranges selected, only considering the major elements of the wine, together with NN analysis is enough to perform successful classification of wines. It is interesting to note that without performing any quantification of the elements, the wavelength ranges selected consisting the emission lines from the major elements of wines, and together with NN analysis were enough to perform successful classification of wines. The intervals selected provide a characteristic fingerprint of the sample taking into account the peak profiles that includes information from the plasma which is useful in the classification process. NN analysis, due to its capacity to model complex non-linear input-target relationships, was able to recognize the PDO pattern and classify correctly all wines tested, maximizing inter-PDO differences and minimizing intra-PDO variability. All validation procedures have produced successful results. A high sensitivity and generalization ability was achieved with an average of 99.2% and 98.6% of correct classification, respectively with a robustness of the 100%. Thus, this work offers the possibility to perform red wine classification based on LIBS measurement combined with multivariate chemometric method (NN). This LIBS/NN methodology was able to distinguish between geographically close regions. From the legal point of view, this methodology seems to be sufficiently reliable to be used in quality control procedures as a screening tool. Methodologically the developed LIBS-NN system provides a simple and fast way of identifying the PDOs without carrying out cumbersome analysis in experimental as well as mathematical point of view.
Predicted Group Membership (Spectral Correlation, %) Sample ID La Mancha Ribera del Duero
Rioja Valdepeñas Vinos de Madrid
Correctly as unknown
CR RB RG NV SM TR TW1 TW2 TW3 TW4 CH DR VC
5 8 2 0 0 2 0 32 0 0 0 0 0
✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
5 0 4 2 7 40 32 12 70 22 0 0 0
0 0 0 0 3 0 0 0 0 0 0 8 0
0 0 0 2 0 0 0 0 0 0 7 2 7
0 0 0 0 5 3 0 0 10 8 0 0 0
4. Conclusions Laser Induced Breakdown Spectroscopy (LIBS) technique combined with a supervised chemometric method Neural Network (NN) has been evaluated in a real word application for a fast and robust discrimination and classification of red wines based on their characteristic spectral fingerprints. The liquid-to-solid treatment improves the quality of the spectra and their acquisition, avoiding the strong splashing and sloshing of the liquid produced by the shock waves. The dry gel emission signals are taken by NN as a background, allowing the classification of samples without any additional spectral treatment. Despite of the difficulties in the implementation of a NN analysis compared to other standard chemometric methods, due to the diversity of functions and architectures, the results obtained in this study demonstrate that
S. Moncayo et al. / Talanta 158 (2016) 185–191
LIBS combined with NN offers an analysis highly sensitive and robust with the great generalization ability, which makes this combination a useful screening tool, given its speed, high throughput, minimal destructive, micro-analysis and ease of use. In this work, three different validation approaches (internal, generalization ability and independent external validation) have been examined to evaluate the performance of the model also considering a complete evaluation of the possible overfitting of the model. Moreover, the use of a compact spectrometer to perform the analysis improves the possibilities of building a portable LIBS prototype, reducing the dimension of the experimental setup and being able to provide an in situ classification result, facilitating the adulteration control task and improving the quality control of agroalimentary products.
Acknowledgements The authors gratefully acknowledge financial support from the Complutense University of Madrid.
References [1] Regulation (EU) No 1151/2012 of the European Parliament and of the Council of 21 November 2012 on quality schemes for agricultural products and foodstuffs, EUR-Lex -32012R1151-EN-EUR-Lex, Cited February, 2016. [2] D.M.A.M. Luykx, S.M. van Ruth, An overview of analytical methods for determining the geographical origin of food products, Food Chem. 107 (2008) 897–911. [3] K. Schlesier, C. Fauhl-Hassek, M. Forina, V. Cotea, E. Kocsi, R. Schoula, F. Jaarsveld, R. Wittkowski, Characterisation and determination of the geographical origin of wines. Part I: overview, Eur. Food Res. Technol. 230 (2009) 1–13. [4] E.I. Geana, R. Popescu, D. Costinel, O.R. Dinca, I. Stefanescu, R.E. Ionete, C. Bala, Verifying the red wines adulteration through isotopic and chromatographic investigations coupled with multivariate statistic interpretation of the data, Food Control 62 (2016) 1–9. [5] M. Aceto, O. Abollino, M.C. Bruzzoniti, E. Mentasti, C. Sarzanini, M. Malandrino, Determination of metals in wine with atomic spectroscopy (flame-AAS, GFASS and ICP-AES); a review, Food Addit. Contam. 19 (2002) 126–133. [6] F.H. Larsen, F. van den Berg, S.B. Engelsen, An exploratory chemometric study of 1H NMR spectra of table wines, J. Chemom. 20 (2006) 198–208. [7] D. Serrano-Lourido, J. Saurina, S. Hernández-Cassou, A. Checa, Classification and characterisation of Spanish red wines according to their appellation of origin based on chromatographic profiles and chemometric data analysis, Food Chem. 135 (2012) 1425–1431. [8] D.P. Makris, S. Kallithraka, A. Mamalos, Differentiation of young red wines based on cultivar and geographical origin with application of chemometrics of principal polyphenolic constituents, Talanta 70 (2006) 1143–1152. [9] L. Jaitz, K. Siegl, R. Eder, G. Rak, L. Abranko, G. Koellensperger, S. Hann, LC–MS/ MS analysis of phenols for classification of red wine according to geographic origin, grape variety and vintage, Food Chem. 122 (2010) 366–372. [10] D. Wang, Q. Zhong, G. Li, Z. Huang, Rapid method for the determination of the stable oxygen isotope ratio of water in alcoholic beverages, J. Agric. Food Chem. 63 (2015) 9357–9362. [11] M.A. Brescia, V. Caldarola, A. De Giglio, D. Benedetti, F.P. Fanizzi, A. Sacco, Characterization of the geographical origin of Italian red wines based on traditional and nuclear magnetic resonance spectrometric determinations, Anal.
191
Chim. Acta 458 (2002) 177–186. [12] E.I. Geana, R. Popescu, D. Costinel, O.R. Dinca, R.E. Ionete, I. Stefanescu, V. Artem, C. Bala, Classification of red wines using suitable markers coupled with multivariate statistic analysis, Food Chem. 192 (2016) 1015–1024. [13] L. Martins, P. Goncalves, P.G.P. De, Method and Kit for Dna Extraction from Vitis vinifera L. and for Amplification and Detection of Grapevine Varieties or Cultivars in Musts or Wines, Universidade De Tras-Os-Montes E Alto Douro; Martins Lopes, Paula Filomena; Goncalves Pereira, Maria Leonor, De Pinho Guedes Pinto, Henrique, 2011. [14] A.W. Miziolek, V. Palleschi, I. Schechter, Laser Induced Breakdown Spectroscopy (LIBS), 1 ed., Cambridge University Press, New York – USA, 2008. [15] R.J. Lasheras, C. Bello-Gálvez, E.M. Rodríguez-Celis, J. Anzano, Discrimination of organic solid materials by LIBS using methods of correlation and normalized coordinates, J. Hazard. Mater. 192 (2011) 704–713. [16] P. Lucena, A. Doña, L.M. Tobaria, J.J. Laserna, New challenges and insights in the detection and spectral identification of organic explosives by laser induced breakdown spectroscopy, Spectrochim. Acta Part B: At. Spectrosc. 66 (2011) 12–20. [17] Y. Huang, L.J. Kangas, B.A. Rasco, Applications of artificial neural networks (ANNs) in food science, Crit. Rev. Food Sci. Nutr. 47 (2007) 113–126. [18] S. Curteanu, H. Cartwright, Neural networks applied in chemistry Neural networks applied in chemistry. I. Determination of the optimal topology of multilayer perceptron neural networks, J. Chemom. 25 (2011) 527–549. [19] S. Moncayo, S. Manzoor, F. Navarro-Villoslada, J.O. Caceres, Evaluation of supervised chemometric methods for sample classification by Laser Induced Breakdown Spectroscopy, Chemom. Intell. Lab. Syst. 146 (2015) 354–364. [20] J.O. Caceres, S. Moncayo, J.D. Rosales, F.J.M. de Villena, F.C. Alvira, G.M. Bilmes, Application of Laser-Induced Breakdown Spectroscopy (LIBS) and neural networks to olive oils analysis, Appl. Spectrosc. 67 (2013) 1064–1072. [21] S.C. Jantzi, V. Motto-Ros, F. Trichard, Y. Markushin, N. Melikechi, A. De Giacomo, Sample treatment and preparation for laser-induced breakdown spectroscopy, Spectrochim. Acta Part B: At. Spectrosc. 115 (2016) 52–63. [22] H. Sobral, R. Sanginés, A. Trujillo-Vázquez, Detection of trace elements in ice and water by laser-induced breakdown spectroscopy, Spectrochim. Acta Part B: At. Spectrosc. 78 (2012) 62–66. [23] J.O. Caceres, J. Tornero López, H.H. Telle, A. González Ureña, Quantitative analysis of trace metal ions in ice using Laser-induced Breakdown Spectroscopy, Spectrochim. Acta Part B: At. Spectrosc. 56 (2001) 831–838. [24] D. Zhu, L. Wu, B. Wang, J. Chen, J. Lu, X. Ni, Determination of Ca and Mg in aqueous solution by laser-induced breakdown spectroscopy using absorbent paper substrates, Appl. Opt. 50 (2011) 5695–5699. [25] D.M. Díaz Pace, C.A. D’Angelo, D. Bertuccelli, G. Bertuccelli, Analysis of heavy metals in liquids using Laser Induced Breakdown Spectroscopy by liquid-tosolid matrix conversion, Spectrochim. Acta Part B: At. Spectrosc. 61 (2006) 929–933. [26] L. St-Onge, E. Kwong, M. Sabsabi, E.B. Vadas, Rapid analysis of liquid formulations containing sodium chloride using laser-induced breakdown spectroscopy, J. Pharm. Biomed. Anal. 36 (2004) 277–284. [27] M.A. Gondal, T. Hussain, Z.H. Yamani, M.A. Baig, Detection of heavy metals in Arabian crude oil residue using laser induced breakdown spectroscopy, Talanta 69 (2006) 1072–1078. [28] M.A. Gondal, M.N. Siddiqui, M.M. Nasr, Detection of trace metals in asphaltenes using an advanced Laser-Induced Breakdown Spectroscopy (LIBS) technique, Energy Fuels 24 (2010) 1099–1105. [29] D. Marcos-Martinez, J.A. Ayala, R.C. Izquierdo-Hornillos, F.J.M. de Villena, J. O. Caceres, Identification and discrimination of bacterial strains by laser induced breakdown spectroscopy and neural networks, Talanta 84 (2011) 730–737. [30] N. US Department of Commerce, NIST Atomic Spectra Database, 2015. [31] S. Manzoor, S. Moncayo, F. Navarro-Villoslada, J.A. Ayala, R. Izquierdo-Hornillos, F.J.M. de Villena, J.O. Caceres, Rapid identification and discrimination of bacterial strains by laser induced breakdown spectroscopy and neural networks, Talanta 121 (2014) 65–70. [32] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford University Press, USA, 1996. [33] H.B. Demuth, M.H. Beale, M.T. Hagan, Neural network toolbox for use with MATLAB: User’s guide 9th for version 6.0 (Release 2008a), Math Works 2007.
Contents lists available at ScienceDirect
Talanta journal homepage: www.elsevier.com/locate/talanta
Classification of red wine based on its protected designation of origin (PDO) using Laser-induced Breakdown Spectroscopy (LIBS) S. Moncayo a, J.D. Rosales a, R. Izquierdo-Hornillos a, J. Anzano b, J.O. Caceres a,n a b
Department of Analytical Chemistry, Faculty of Chemical Sciences, Complutense University, 28040 Madrid, Spain Laser Laboratory, Department of Analytical Chemistry, Faculty of Sciences, University of Zaragoza, 50009 Zaragoza, Spain
art ic l e i nf o
a b s t r a c t
Article history: Received 27 February 2016 Received in revised form 16 May 2016 Accepted 21 May 2016 Available online 24 May 2016
This work reports on a simple and fast classification procedure for the quality control of red wines with protected designation of origin (PDO) by means of Laser Induced Breakdown Spectroscopy (LIBS) technique combined with Neural Networks (NN) in order to increase the quality assurance and authenticity issues. A total of thirty-eight red wine samples from different PDO were analyzed to detect fake wines and to avoid unfair competition in the market. LIBS is well known for not requiring sample preparation, however, in order to increase its analytical performance a new sample preparation treatment by previous liquid-to-solid transformation of the wine using a dry collagen gel has been developed. The use of collagen pellets allowed achieving successful classification results, avoiding the limitations and difficulties of working with aqueous samples. The performance of the NN model was assessed by three validation procedures taking into account their sensitivity (internal validation), generalization ability and robustness (independent external validation). The results of the use of a spectroscopic technique coupled with a chemometric analysis (LIBS-NN) are discussed in terms of its potential use in the food industry, providing a methodology able to perform the quality control of alcoholic beverages. & 2016 Elsevier B.V. All rights reserved.
Keywords: Laser Induced Breakdown Spectroscopy Wine Neural networks Protected designation of origin LIBS PDO
1. Introduction The certification of the protected designation of origin (PDO) is one of the most important parameters to be controlled in order to protect the production and origin of agroalimentary products. Since the introduction of European regulations control [1] on this matter, many wine companies have adopted different strategies on the confirmation of wine authenticity improving the PDO controls. The type of grape, geographical origin, harvest, and vintage are parameters that determine the quality of wine products. Nevertheless, these indicators are not easily recognizable for the consumers and the PDO is considered as a single and unambiguous sign of quality that companies use to promote their brand in the market influencing the customer final decision [2]. The wine industry has seen an important growth in the last decade associated with an increase in the wine consumption. This is a strategic sector due to its importance in economic, environmental and social terms, as well as its importance representing the country’s image abroad. Since 2008, the wine industry has put an important effort in the control of counterfeiting of wines with the objective of protecting the trade-mark quality wines and to prevent their n
Corresponding author. E-mail address: [email protected] (J.O. Caceres).
http://dx.doi.org/10.1016/j.talanta.2016.05.059 0039-9140/& 2016 Elsevier B.V. All rights reserved.
illegal adulteration [3,4]. The wine adulteration consists of the addition of any substance to the natural wine, which changes its composition and may occur in many different forms. The greater part of the adulteration consists of addition of water and sugar, mixing with lower quality wines and label replacing [3]. The two main constituents of wine are water (81%) and ethanol (between 11% and 15%). Two types of flavonoids, the anthocyanins and flavanols, are key compounds for color and astringency, being responsible for the organoleptic properties and quality of wine. Other organic compounds in small amounts, such as acids, alcohols, phenols, nitrogenous compounds and inorganic substances represent the remaining 7%, making wine a complex sample and difficult to analyze [5,6]. The sensorial analyses together with chemical assays, and mineral content analysis may not be adequate for determining the PDO of wine [3]. Chromatographic techniques [7–9] require to conduct separate analysis of each component in the wine being slow and expensive process. The identification of grape variety by means of isotopic analysis [4,10], nuclear magnetic resonance (NMR) [11,12] or ADN/aRNA [13] techniques are generally used, due to their capacity of generating a fingerprint of wine providing the identification. Although these techniques produce accurate results, a large amount of sample and the use of expensive consumables are required, increasing the cost and duration of the analysis.
186
S. Moncayo et al. / Talanta 158 (2016) 185–191
This work evaluates the Laser Induced Breakdown Spectroscopy (LIBS) technique for the discrimination and the determination of geographical origin of red wines. LIBS technique is based on the interaction of a laser beam with a material target generating a plasma, the emission of the plasma contains spectroscopic information of excited atoms and ions present in the sample and reflecting its elemental composition [14]. Although there is a loss of molecular information in plasma, LIBS has provided excellent results in the identification of samples with complex matrixes [15,16]. LIBS provides spectral fingerprints characteristic of each sample based on the composition of wine The presence of mineral elements in wine is generally related to soil composition, grape variety, climate conditions, yeasts and winemaking [8]. The combination of LIBS with supervised classification methods such as NN has already shown successful results in many areas of knowledge for sample classification [17–20]. One of the major advantages of LIBS is that it does not require sample preparation, providing an economic and fast analysis, however in some cases avoiding a sample preparation goes in detrimental of the technique limiting its analytical performance. The change of the physical state of the sample transforming the liquid into solid has already been described as sample preparation [21]. Although the liquid-to-solid process produces an increase in the time analysis and an alteration of the original chemical composition, significant improvements such as the increase of the ablation rate, higher plasma temperature and electron density as well as a better laser-to-solid interaction has been observed in literature [22–24]. Moreover, avoiding the inherent drawbacks of working with liquids such as splashing and surface ripples produces lower limit of detection, better repeatability and sensitivity [25,26]. Different liquid-to-solid matrix conversion protocols have been described in the literature involving precipitation, filtering and pellets formation procedures [25,27,28]. Herein, the transformation of the liquid wine sample into gels by adding a natural collagen and its subsequent dried in an air assisted oven has been used as a new sample preparation protocol. The aim of this work was to identify the adulteration of wines collected from Spanish local markets and evaluate the capacity of LIBS coupled with NN to detect the PDO of wines with negligible compositional and spectral differences and to improve the recognition capacity of extremely similar samples that have fewer physical and spectral differences between them.
2. Materials and methods 2.1. Wine samples Thirty-eight Spanish red wines from eleven protected designation of origin, three foreign red wines and four table wines were purchased in retail stores. These wine samples were selected to cover the main Spanish wine regions (Fig. 1) including La Mancha, Ribera de Duero, Rioja, Valdepeñas, Vinos de Madrid, Cariñeña, Ribeiro, Ribera del Guadiana, Navarra and Somontano and Toro. Moreover a German, French and Italian wine was also included in the study. Most of the wines included in the study were elaborated with Tempranillo grapes, although Cabernet Sauvignon, Garnacha, Tinta de Toro and Shyrah were also considered. All samples belong to the 2011 vintage and were not affected by ageing period (young wine). Table 1 shows sample information including sample ID, commercial brand and type of grapes. 2.2. Sample preparation A methodology based on the formation of a gel of wine using a commercial collagen was applied. 50 mL of wine sample were
introduced into a beaker and 1 g of collagen gel was added and dissolved in the wine sample. 2.4 mL of this solution were allowed to stand 15 min until the formation of a gel on a square petri dish of 4 4 cm. Then, samples were introduced in a forced ventilation oven at 3572 °C during 12 h to evaporate the water, obtaining a dry solid. The final sample was completely flat with a thickness of approximately 0.35 mm. Fig. 2 shows an example of dry gel and the craters formed by single laser shots. In this process not only LIBS analysis was simplified but also pre-concentration of the sample (pre-concentration factor of 1:5) is performed allowing an improvement in the limits of detection. The gels for all samples were prepared at the same time to maintain the same conditions and avoiding the degradation and oxidation of the wine components. 2.3. LIBS set-up The LIBS technique and the methodology used in the present work together with the most significant experimental conditions have been previously described [29]. Thus, only the experimental conditions relevant to this study are presented here. LIBS measurements were obtained using a Q-switched Nd: YAG laser (Quantel, Brio model) operating at 1064 nm, with a pulse duration of 4 ns full width at half maximum (FWHM), 4 mm beam diameter and 0.6 mrad divergence. Samples were placed over an X–Y–Z manual micro-metric positionator with a 0.5 mm stage of travel at every coordinate to ensure that each laser pulse impinged on a fresh position. The laser beam was focused onto the sample surface with a 100 mm focal-distance lens, producing a spot of 100 mm in diameter. The best signal-to-background ratio was achieved at 42 mJ of pulse energy with a repetition rate of 1 Hz. The laser crater profile was measured by means of a confocal microscope after laser pulse irradiation on a fresh position. A narrow crater was created with a diameter of 450 mm and 140 mm in depth. Emission from the plasma was collected with a 4-mm aperture, and 7 mm focus fused silica collimator placed at 4 cm from the sample, and then focused into an optical fiber (1000 mm core diameter, 0.22 numerical aperture), coupled to a spectrometer. The spectrometer system was an EPP2000, StellarNet (Tampa, FL, U.S.A.) with a gated CCD detector. A grating of 300 l/ mm was selected; a spectral resolution of 0.5 nm was achieved with a 7 mm entrance slit. The wavelength range used was from 200 to 1000 nm. Therefore, 2048 data points were recorded for each sample. The detector integration time was set to 1 ms, to prevent the detection of bremsstrahlung, the detector was triggered by a 2 μs delay time between the laser pulse and the acquired plasma radiation using a digital delay generator (Stanford model DG535). The spectrometer was computer-controlled using an interface developed in Matlab. 2.4. LIBS analysis Wine samples were measured directly in air at room condition. Each LIBS spectrum was acquired from a single shot measurement. A total of 100 spectra were recorded for each wine sample by moving the sample stage about 0.25 mm to expose a fresh portion of the sample surface and avoiding areas irradiated by previous shots. Only in the case of the M1, D1, R1, V1 and VM1 samples, four data sets of 100 spectra were obtained: the first data set (training library) was used to calculate the model of the NN; whereas the last three data sets (replicate libraries) were used for validation purpose. In order to avoid data variations due to changes in the laser pulse energy, each spectrum was normalized by the intensity of one specific spectral line, i.e., K (I) 766.49 nm [30]. The spectral information of each sample was obtained in less than 2 min considering the integration time of the spectrometer and the frequency of laser pulses that was fixed to 1 Hz.
S. Moncayo et al. / Talanta 158 (2016) 185–191
187
Fig. 1. Geographical distribution of Spanish wine regions. The dashed parts represent the Protected Denomination of Origin of the Spanish analyzed wines.
Table 1. Samples used in the study. PDO
Sample ID
Commercial brand
Grape variety
Vintage
La Mancha
M1 M2 M3 M4 M5 D1 D2 D3 D4 D5 R1 R2 R3 R4 R5 V1 V2 V3 V4 V5 VM1 VM2 VM3 VM4 VM5 CR RB RG NV SM TR TW1 TW2 TW3 TW4 CH DR VC
Libertario Vereda Mayor Don Lucio Fidencio Monte Don Lucio Dehesa Valpincia Barón de Santuy Mayor de Castilla Sangre de Castilla Viña Espolón Antaño Solar Viejo Barón de Urzande Castillo de Albali Vega del Cega Calle Real Viña Albali Señorío de Los Llanos Tanis Puerta de Alcalá Vega Madroño Alma de Valdeguerra Puerta de Hierro Jesús Díaz Castillo de Aguaron Pazo 5 Viñas Diácono MonteSierra Cermeño Conde Noble Don Simon Eroski Viñas Altas Corte Alle Mura Dornfield Cimarosa
Tempranillo, Garnacha Tempranillo Tempranillo Tempranillo, Garnacha Cabernet Sauvignon Cabernet Sauvignon Tempranillo Tempranillo, Cabernet Tempranillo Tempranillo Tempranillo, Garnacha Tempranillo, Garnacha Tempranillo, Garnacha Tempranillo, Garnacha Tempranillo Tempranillo Tempranillo Tempranillo, Cabernet Tempranillo Tempranillo Tempranillo, Syrah Tempranillo, Garnacha Tempranillo Tempranillo Tempranillo, Syrah Garnacha, Syrah Tempranillo, Cabernet Tempranillo, Garnacha Tempranillo, Garnacha Tempranillo, Cabernet Tinta de Toro Mixture Mixture Mixture Mixture Sangiovese Dornfelder Cabernet Sauvignon
2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011 2.011
Ribera del Duero
Rioja
Valdepeñas
Vinos de Madrid
Cariñena Ribeiro Ribera del Guadiana Navarra Somontano Toro Table wine
Chianti Dornfelder Valle Central
188
S. Moncayo et al. / Talanta 158 (2016) 185–191
Fig. 2. (a) Sample after preparation. (b) Dry gel of wine sample with visible laser spots.
Table 2. Wavelength range considered for wine analysis. Element Wavelength range (nm) Mg Ca Na Hα K
271–291 311–340, 390–429, 513–524 581–598 640–675 760–783
2.5. Wavelength range selection In case of LIBS spectra the intensity value at each wavelength is referred to as variable. Each spectrum was composed by 2048 variables taking into account the CCD pixels. For the NN analysis a reduction in the number of variables used as inputs was done. Reducing the number of variables have shown important advantages in a classification process, enhancing the robustness of the models and being more efficient in the discrimination process, without a loss of meaningful information. Since the organic nature of the gel may affect C, H, N, O signals, the main emission lines from the mineral elements of wine, Mg, Ca, K, and Na, were selected in seven wavelengths ranges resulting into 355 variables (Table 2). It is important to note that emission lines for these elements in the selected wavelengths intervals were not observed in the pure gel. The selection of several ranges may not describe the data fully, however, it describes the PDO trend instead of the particular wine trend. The reduction of variables decreases the discrimination power of the NN model but increase the ability to generalize.
complexity [19,31]. The NN classification model is estimated by calibrating with a set of reference samples to calculate the optimum weights and biases that better represent the given classes. NN training was achieved by applying the BP algorithm based on the conjugate gradient method, one of the general-purpose second-order techniques that helps minimize the goal functions of several variables. Second order indicates that such methods use the second derivatives of the error function, whereas a first order technique, such as standard back-propagation, uses only the first derivatives. To determine when the training should be stopped, an early stopping criteria based on performance improving (error rate) of the validation set. The number of epochs was not relevant in this case. To avoid an overfitting of the NN model, the learning process was repeated until a minimum of the mean square error (MSE) of the verification data, defined in Eq. (1), was reached:
MSE =
1 N
N
∑K
( rk − yk )2
(1)
where N, yk, and rk are the number of input data, the response from each output neuron, and the observed output response, respectively. Two hidden layers were used to calculate the model considering the ratio between sensitivity and generalization ability. The output layer (classification result) is comprised of J neurons (where the J¼ number of reference samples used) for estimating the similarity between the reference sample spectra and the testing sample spectra. A detailed description of the calculation process is provided in the literature [32,33]. Once the calibration is done, the weights are frozen and the NN model is validated with a test data set (samples not included in the calibration) providing a prediction result. Although the data matrix can be considerably large, the computation time for training the NNs was always below 10 s
2.6. Artificial Neural Networks (ANN) model 2.7. Classification and validation procedures A multilayer perceptron, feedforward, supervised neural network was used in this classification process [8]. The main advantage of NN is the capacity to model nonlinear effects, which are always present in LIBS spectra. In our case the NN topology consists of several neurons arranged in three different layers (input, hidden and output). The connections are controlled by a weight that controls the output of the neuron before inputting its numerical content to a neuron in the next layer. This topology has been widely used to model systems with a similar level of
An appropriate classification system must satisfy three basic validation procedures. Firstly, it must be able to classify all wine samples included in the training (sensitivity test), for this the NN model is trained with only one wine for each PDO and the ability of the model to classify replicates of the same wine is assessed, which is considered as an internal validation of the model. Secondly, it must provide a high generalization ability (generalization test), in other words the ability of the model to classify other
S. Moncayo et al. / Talanta 158 (2016) 185–191
wines belonging to one of the PDOs but not included in the training. This is considered as an external validation of the model, since the wine was not presented to the model previously, but is related to the training spectra by its PDO. Finally, the model must be robust. In this case it should be able to classify a wine that does not belong to any PDO as unassigned and not classify it as a trained PDO, therefore, there is no relation with the training spectra being completely independent of the model and this procedure is named as independent external validation [19].
3. Results and discussions Thirty-eight wine samples and a minimum of two replicates for each of the PDO covering a spectral window from 200 to 900 nm were analyzed. Fig. 3(a) shows a typical LIBS spectrum of a wine sample and the wavelength ranges selected as inputs to the NN have been highlighted in gray color. The observed molecular emission bands are identified using the spectroscopic information tabulated in NIST Atomic Spectral Database [30]. Lines emission for Mg, Ca, K, Na, Hα, C and molecular species such as swan carbon system, among others, are mainly present in the LIBS spectra. A visual comparison of spectrum emission lines from pure gel, M1, D1, R1, V1, VM1 wines and pure gel sample assigned as I, II, III, IV, V, VI and VII, respectively are shown in Fig. 3(b). Emissions in all the spectra from analyzed wine samples correspond to the same elements showed high similarity among distinct LIBS spectral fingerprints. To consider a wine sample correctly classified, the prediction of the model must match with the actual class by an arbitrary threshold of Spectral Correlation (SC) (Eq. (2)) higher than 90% and less than 20% to the other classes, otherwise is considered incorrectly classified. A wine sample was classified as unclassified when the SC was less than 80% for all classes.
SC =
100 N
∑ δi i
(2)
where δi is the number of spectra classified correctly and N is the total number of spectra.
189
3.1. Sensitivity tests For the training of the NN model, the training library samples M1, D1, R1, V1 and VM1 were used, producing an input matrix with 355 rows (number of variables) and 500 columns (5 100 spectra of each PDO). Once the learning is done the NN model is able to classify new spectra in a class. For the first validation procedure, three replicates of the training wines were prepared and measured at the same condition, generating three new data sets used to validate the model. Although, the same wine is considered to produce the replicates, each replicated was prepared in different gels, assessing possible sample preparation differences. These data sets have not been shown in the NN model previously and therefore the spectra are different to those that were used for the training. Table 3 gives the results of the sensitivity analysis for NN prediction of wine replicates. The model is considered sensitive if the correct classification is high and the misclassification rates are low. Based on the results, it is noted that a high sensitivity is achieved for all replicates with an average of 99.2% rate in the correct classification with a negligible rate of incorrect classification and misclassification. Therefore, the model was able to classify new spectra belonging to the samples used for training despite of the intrinsic sample-to-sample variability in the preparation of each gel. 3.2. Generalization ability test The validation of the model considering replicates is not enough in some cases because they cannot be seen as independent observations. Therefore, other validation procedures must be performed, ascertaining the generalization ability of the model is an appropriate way to do it. In this second validation the same NN model as discussed in the previous section was used. To estimate the generalization ability of the model four different wine brands belonging to the PDOs used in the training were tested to obtain the prediction result. A high rate in the correct classification and low rate of incorrect classification and misclassification was sought, which means that the model was able to generalize and correctly classify other spectra than those with which it was trained. Table 4 gives the results of the generalization test. An average of 98.6% of correct classification was achieved without any wine unclassified or misclassified and thus all wine samples were correctly assigned to their respective classes showing a high generalization ability and evidence the lack of model overfitting. 3.3. Robustness test
Fig. 3. (a) LIBS spectrum of Rioja PDO with the assignation of emission lines and selected the wavelength ranges selected as inputs to the NN have been highlighted in gray color. (b) Typical spectrum of (I) dry gel (II) M1 (La Mancha: Libertario) (III) D1 (Ribera del Duero: Dehesa) (IV) R1 (Rioja: Viña Espolón) (V) V1 (Valdepeñas: Vega del Cega) (VI) VM1 (Vinos de Madrid: Puerta de Alcalá) within the spectral range of 200–1000 nm.
To assess the robustness of the NN model, thirteen wine samples were measured and introduced into the NN model evaluating the response of the model to detect samples completely external and independent of the training set. Six wines with Spanish PDO, four table wines and three foreign wines were considered. A robust model implies a high capacity to detect unknown samples “correctly as unknown”, without diminishing the prediction accuracy of the known samples. Table 5 presents the NN model classification results when completely unknown samples were introduced to the model. All wine samples were correctly classified as unknown to the model. None of the samples were classified as belonging to other classes, providing a high robustness. Only in case of TW3 a Spectral Correlation of 70% was obtained in the La Mancha class which can be related to the fact that the table wines are elaborated with mixed variety of grapes, in many cases, those grapes are collected in the La Mancha region. However, the SC is lower than the threshold fixed and therefore classified as unknown to the model. This highlights the robustness of NN in dealing with samples of unknown classes.
190
S. Moncayo et al. / Talanta 158 (2016) 185–191
Table 3. Classification results to sensitivity test (internal validation). Predicted Group Membership (Spectral Correlation, %) Sample
La Mancha
Ribera del Duero
Rioja
Valdepeñas
Vinos de Madrid
Unclassified
M1-replicate 1 M1-replicate 2 M1-replicate 3 D1-replicate 1 D1-replicate 2 D1-replicate 3 R1-replicate 1 R1-replicate 2 R1-replicate 3 V1-replicate 1 V1-replicate 2 V1-replicate 3 VM1-replicate 1 VM1-replicate 2 VM3-replicate 3
100 99 99 0 0 0 0 0 0 0 0 1 0 0 1
0 0 1 100 98 100 0 1 1 0 1 1 0 0 0
0 1 0 0 0 0 100 99 98 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 100 99 98 0 0 0
0 0 0 0 0 0 0 0 1 0 0 0 100 99 99
0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
Table 4. Classification results for the generalization ability test. Predicted Group Membership (Spectral Correlation, %) Sample ID La Mancha Ribera del Duero
Rioja Valdepeñas Vinos de Madrid
Unclassified
2M2 2M3 2M4 2M5 2D2 2D3 2D4 2D5 2R2 2R3 2R4 2R5 2V2 2V3 2V4 2V5 2VM2 2VM3 2VM4 2VM5
1 0 0 0 0 1 1 0 100 98 97 97 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 1 0 0
98 100 98 99 2 0 1 0 0 1 1 0 0 2 1 0 1 0 0 0
0 0 1 1 98 99 98 100 0 0 1 2 0 1 0 0 0 0 0 0
1 0 1 0 0 0 0 0 0 0 0 1 100 97 98 100 1 0 1 0
0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 98 99 98 100
Table 5. Classification results to the robustness test (external validation).
The results were satisfactory and the proposed methodology was used for the analysis of red wines. The wavelength ranges selected, only considering the major elements of the wine, together with NN analysis is enough to perform successful classification of wines. It is interesting to note that without performing any quantification of the elements, the wavelength ranges selected consisting the emission lines from the major elements of wines, and together with NN analysis were enough to perform successful classification of wines. The intervals selected provide a characteristic fingerprint of the sample taking into account the peak profiles that includes information from the plasma which is useful in the classification process. NN analysis, due to its capacity to model complex non-linear input-target relationships, was able to recognize the PDO pattern and classify correctly all wines tested, maximizing inter-PDO differences and minimizing intra-PDO variability. All validation procedures have produced successful results. A high sensitivity and generalization ability was achieved with an average of 99.2% and 98.6% of correct classification, respectively with a robustness of the 100%. Thus, this work offers the possibility to perform red wine classification based on LIBS measurement combined with multivariate chemometric method (NN). This LIBS/NN methodology was able to distinguish between geographically close regions. From the legal point of view, this methodology seems to be sufficiently reliable to be used in quality control procedures as a screening tool. Methodologically the developed LIBS-NN system provides a simple and fast way of identifying the PDOs without carrying out cumbersome analysis in experimental as well as mathematical point of view.
Predicted Group Membership (Spectral Correlation, %) Sample ID La Mancha Ribera del Duero
Rioja Valdepeñas Vinos de Madrid
Correctly as unknown
CR RB RG NV SM TR TW1 TW2 TW3 TW4 CH DR VC
5 8 2 0 0 2 0 32 0 0 0 0 0
✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
5 0 4 2 7 40 32 12 70 22 0 0 0
0 0 0 0 3 0 0 0 0 0 0 8 0
0 0 0 2 0 0 0 0 0 0 7 2 7
0 0 0 0 5 3 0 0 10 8 0 0 0
4. Conclusions Laser Induced Breakdown Spectroscopy (LIBS) technique combined with a supervised chemometric method Neural Network (NN) has been evaluated in a real word application for a fast and robust discrimination and classification of red wines based on their characteristic spectral fingerprints. The liquid-to-solid treatment improves the quality of the spectra and their acquisition, avoiding the strong splashing and sloshing of the liquid produced by the shock waves. The dry gel emission signals are taken by NN as a background, allowing the classification of samples without any additional spectral treatment. Despite of the difficulties in the implementation of a NN analysis compared to other standard chemometric methods, due to the diversity of functions and architectures, the results obtained in this study demonstrate that
S. Moncayo et al. / Talanta 158 (2016) 185–191
LIBS combined with NN offers an analysis highly sensitive and robust with the great generalization ability, which makes this combination a useful screening tool, given its speed, high throughput, minimal destructive, micro-analysis and ease of use. In this work, three different validation approaches (internal, generalization ability and independent external validation) have been examined to evaluate the performance of the model also considering a complete evaluation of the possible overfitting of the model. Moreover, the use of a compact spectrometer to perform the analysis improves the possibilities of building a portable LIBS prototype, reducing the dimension of the experimental setup and being able to provide an in situ classification result, facilitating the adulteration control task and improving the quality control of agroalimentary products.
Acknowledgements The authors gratefully acknowledge financial support from the Complutense University of Madrid.
References [1] Regulation (EU) No 1151/2012 of the European Parliament and of the Council of 21 November 2012 on quality schemes for agricultural products and foodstuffs, EUR-Lex -32012R1151-EN-EUR-Lex, Cited February, 2016. [2] D.M.A.M. Luykx, S.M. van Ruth, An overview of analytical methods for determining the geographical origin of food products, Food Chem. 107 (2008) 897–911. [3] K. Schlesier, C. Fauhl-Hassek, M. Forina, V. Cotea, E. Kocsi, R. Schoula, F. Jaarsveld, R. Wittkowski, Characterisation and determination of the geographical origin of wines. Part I: overview, Eur. Food Res. Technol. 230 (2009) 1–13. [4] E.I. Geana, R. Popescu, D. Costinel, O.R. Dinca, I. Stefanescu, R.E. Ionete, C. Bala, Verifying the red wines adulteration through isotopic and chromatographic investigations coupled with multivariate statistic interpretation of the data, Food Control 62 (2016) 1–9. [5] M. Aceto, O. Abollino, M.C. Bruzzoniti, E. Mentasti, C. Sarzanini, M. Malandrino, Determination of metals in wine with atomic spectroscopy (flame-AAS, GFASS and ICP-AES); a review, Food Addit. Contam. 19 (2002) 126–133. [6] F.H. Larsen, F. van den Berg, S.B. Engelsen, An exploratory chemometric study of 1H NMR spectra of table wines, J. Chemom. 20 (2006) 198–208. [7] D. Serrano-Lourido, J. Saurina, S. Hernández-Cassou, A. Checa, Classification and characterisation of Spanish red wines according to their appellation of origin based on chromatographic profiles and chemometric data analysis, Food Chem. 135 (2012) 1425–1431. [8] D.P. Makris, S. Kallithraka, A. Mamalos, Differentiation of young red wines based on cultivar and geographical origin with application of chemometrics of principal polyphenolic constituents, Talanta 70 (2006) 1143–1152. [9] L. Jaitz, K. Siegl, R. Eder, G. Rak, L. Abranko, G. Koellensperger, S. Hann, LC–MS/ MS analysis of phenols for classification of red wine according to geographic origin, grape variety and vintage, Food Chem. 122 (2010) 366–372. [10] D. Wang, Q. Zhong, G. Li, Z. Huang, Rapid method for the determination of the stable oxygen isotope ratio of water in alcoholic beverages, J. Agric. Food Chem. 63 (2015) 9357–9362. [11] M.A. Brescia, V. Caldarola, A. De Giglio, D. Benedetti, F.P. Fanizzi, A. Sacco, Characterization of the geographical origin of Italian red wines based on traditional and nuclear magnetic resonance spectrometric determinations, Anal.
191
Chim. Acta 458 (2002) 177–186. [12] E.I. Geana, R. Popescu, D. Costinel, O.R. Dinca, R.E. Ionete, I. Stefanescu, V. Artem, C. Bala, Classification of red wines using suitable markers coupled with multivariate statistic analysis, Food Chem. 192 (2016) 1015–1024. [13] L. Martins, P. Goncalves, P.G.P. De, Method and Kit for Dna Extraction from Vitis vinifera L. and for Amplification and Detection of Grapevine Varieties or Cultivars in Musts or Wines, Universidade De Tras-Os-Montes E Alto Douro; Martins Lopes, Paula Filomena; Goncalves Pereira, Maria Leonor, De Pinho Guedes Pinto, Henrique, 2011. [14] A.W. Miziolek, V. Palleschi, I. Schechter, Laser Induced Breakdown Spectroscopy (LIBS), 1 ed., Cambridge University Press, New York – USA, 2008. [15] R.J. Lasheras, C. Bello-Gálvez, E.M. Rodríguez-Celis, J. Anzano, Discrimination of organic solid materials by LIBS using methods of correlation and normalized coordinates, J. Hazard. Mater. 192 (2011) 704–713. [16] P. Lucena, A. Doña, L.M. Tobaria, J.J. Laserna, New challenges and insights in the detection and spectral identification of organic explosives by laser induced breakdown spectroscopy, Spectrochim. Acta Part B: At. Spectrosc. 66 (2011) 12–20. [17] Y. Huang, L.J. Kangas, B.A. Rasco, Applications of artificial neural networks (ANNs) in food science, Crit. Rev. Food Sci. Nutr. 47 (2007) 113–126. [18] S. Curteanu, H. Cartwright, Neural networks applied in chemistry Neural networks applied in chemistry. I. Determination of the optimal topology of multilayer perceptron neural networks, J. Chemom. 25 (2011) 527–549. [19] S. Moncayo, S. Manzoor, F. Navarro-Villoslada, J.O. Caceres, Evaluation of supervised chemometric methods for sample classification by Laser Induced Breakdown Spectroscopy, Chemom. Intell. Lab. Syst. 146 (2015) 354–364. [20] J.O. Caceres, S. Moncayo, J.D. Rosales, F.J.M. de Villena, F.C. Alvira, G.M. Bilmes, Application of Laser-Induced Breakdown Spectroscopy (LIBS) and neural networks to olive oils analysis, Appl. Spectrosc. 67 (2013) 1064–1072. [21] S.C. Jantzi, V. Motto-Ros, F. Trichard, Y. Markushin, N. Melikechi, A. De Giacomo, Sample treatment and preparation for laser-induced breakdown spectroscopy, Spectrochim. Acta Part B: At. Spectrosc. 115 (2016) 52–63. [22] H. Sobral, R. Sanginés, A. Trujillo-Vázquez, Detection of trace elements in ice and water by laser-induced breakdown spectroscopy, Spectrochim. Acta Part B: At. Spectrosc. 78 (2012) 62–66. [23] J.O. Caceres, J. Tornero López, H.H. Telle, A. González Ureña, Quantitative analysis of trace metal ions in ice using Laser-induced Breakdown Spectroscopy, Spectrochim. Acta Part B: At. Spectrosc. 56 (2001) 831–838. [24] D. Zhu, L. Wu, B. Wang, J. Chen, J. Lu, X. Ni, Determination of Ca and Mg in aqueous solution by laser-induced breakdown spectroscopy using absorbent paper substrates, Appl. Opt. 50 (2011) 5695–5699. [25] D.M. Díaz Pace, C.A. D’Angelo, D. Bertuccelli, G. Bertuccelli, Analysis of heavy metals in liquids using Laser Induced Breakdown Spectroscopy by liquid-tosolid matrix conversion, Spectrochim. Acta Part B: At. Spectrosc. 61 (2006) 929–933. [26] L. St-Onge, E. Kwong, M. Sabsabi, E.B. Vadas, Rapid analysis of liquid formulations containing sodium chloride using laser-induced breakdown spectroscopy, J. Pharm. Biomed. Anal. 36 (2004) 277–284. [27] M.A. Gondal, T. Hussain, Z.H. Yamani, M.A. Baig, Detection of heavy metals in Arabian crude oil residue using laser induced breakdown spectroscopy, Talanta 69 (2006) 1072–1078. [28] M.A. Gondal, M.N. Siddiqui, M.M. Nasr, Detection of trace metals in asphaltenes using an advanced Laser-Induced Breakdown Spectroscopy (LIBS) technique, Energy Fuels 24 (2010) 1099–1105. [29] D. Marcos-Martinez, J.A. Ayala, R.C. Izquierdo-Hornillos, F.J.M. de Villena, J. O. Caceres, Identification and discrimination of bacterial strains by laser induced breakdown spectroscopy and neural networks, Talanta 84 (2011) 730–737. [30] N. US Department of Commerce, NIST Atomic Spectra Database, 2015. [31] S. Manzoor, S. Moncayo, F. Navarro-Villoslada, J.A. Ayala, R. Izquierdo-Hornillos, F.J.M. de Villena, J.O. Caceres, Rapid identification and discrimination of bacterial strains by laser induced breakdown spectroscopy and neural networks, Talanta 121 (2014) 65–70. [32] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford University Press, USA, 1996. [33] H.B. Demuth, M.H. Beale, M.T. Hagan, Neural network toolbox for use with MATLAB: User’s guide 9th for version 6.0 (Release 2008a), Math Works 2007.