- Email: [email protected]
Classification of edible oils and modeling of their physico-chemical properties by chemometric methods using mid-IR spectroscopy
Spectrochimica Acta Part A 100 (2013) 109–114
Contents lists available at SciVerse ScienceDirect
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy journal homepage: www.elsevier.com/locate/saa
Classification of edible oils and modeling of their physico-chemical properties by chemometric methods using mid-IR spectroscopy Aderval S. Luna a,∗ , Arnaldo P. da Silva a , Joan Ferré b , Ricard Boqué b a b
Rio de Janeiro State University, Department of Analytical Chemistry, Rua São Francisco Xavier, 524, Maracanã, Rio de Janeiro CEP 20550-013, Brazil Universitat Rovira i Virgili, Department of Analytical Chemistry and Organic Chemistry, C/Marcel.lí Domingo s/n–43007, Tarragona, Spain
a r t i c l e
i n f o
Article history: Received 30 November 2011 Received in revised form 4 May 2012 Accepted 25 June 2012 Keywords: Edible oils Fourier transform mid-infrared spectroscopy First order calibration models Refraction index Relative density
a b s t r a c t This research work describes two studies for the classification and characterization of edible oils and its quality parameters through Fourier transform mid infrared spectroscopy (FT-mid-IR) together with chemometric methods. The discrimination of canola, sunflower, corn and soybean oils was investigated using SVM-DA, SIMCA and PLS-DA. Using FT-mid-IR, DPLS was able to classify 100% of the samples from the validation set, but SIMCA and SVM-DA were not. The quality parameters: refraction index and relative density of edible oils were obtained from reference methods. Prediction models for FT-mid-IR spectra were calculated for these quality parameters using partial least squares (PLS) and support vector machines (SVM). Several preprocessing alternatives (first derivative, multiplicative scatter correction, mean centering, and standard normal variate) were investigated. The best result for the refraction index was achieved with SVM as well as for the relative density except when the preprocessing combination of mean centering and first derivative was used. For both of quality parameters, the best results obtained for the figures of merit expressed by the root mean square error of cross validation (RMSECV) and prediction (RMSEP) were equal to 0.0001. © 2012 Elsevier B.V. All rights reserved.
1. Introduction Public interest in food quality and methods of production has increased significantly in the recent decades, due in part to changes in eating habits, consumer behavior, and the increased industrialization and globalization of food supply chains. Demand for high levels of quality and safety in food production obviously requires high standards in quality assurance and process control; satisfying this demand in turn requires appropriate analytical tools for food analysis both during and after production. Desirable features of such tools include speed, ease-of-use, minimal or no sample preparation, and the avoidance of sample destruction. These features are characteristic of a range of spectroscopic methods including the mid-infrared (MIR) and near infrared (NIR) [1]. The laboratory routine measurements of sample properties in edible oils are based on well-established standard methods, following procedures determined by international organizations like the American oil chemists’ society (AOCS) and national as Instituto Adolpho Lutz (Brazil). Although these methods are widely accepted, in many cases they exhibit several disadvantages.
∗ Corresponding author. Tel.: +55 21 23340563; fax: +55 21 23340159. E-mail address: [email protected] (A.S. Luna). 1386-1425/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.saa.2012.06.034
While fast and reliable methods are needed to monitor the quality of the various streams in a factory, most of the standard test methods cannot be applied on-line. An alternative to these limitations is the use of spectroscopic techniques in on-line or off-line mode in conjunction with multivariate calibration schemes [2,3]. Olive oil adulteration adulteration with less expensive edible oils is a major issue for the olive oil industry. Obeidat et al. used Fourier transform infrared (FTIR) spectroscopy to classify different edible oils including the virgin olive oil [4]. Tegou et al. used the same analytical technique to determine extra virgin olive oil adulteration with lower priced vegetable oils (sunflower oil, soyabean oil, sesame oil, and corn oil) [5]. Edible oils classification has also been performed by chemometric treatment of chromatographic profiles [6,7], Fourier transform Raman spectra [8] and 13 C NMR spectra [9]. Araújo et al. proposed an electroanalytical methodology for classification of edible vegetable oils with respect to type (canola, sunflower, corn and soybean) and conservation state (expired and non-expired shelf life) [10]. Baeten and Aparicio used Raman spectroscopy to discriminate virgin olive oil from other edible oils [11] and Materny et al. used visible Raman spectroscopy for the discrimination of olive oils from different vegetable oils and the detection of adulteration [12]. Many older tests involved determination of physical properties such as refractive index, melting point, and viscosity. In this context, Araújo et al. proposed an analytical method by using near
110
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
infrared spectroscopy (NIR) measurements for determination of acidity, refractive index and viscosity in four types of edible vegetable oils (corn, soya, canola and sunflower) [2] and de la Guardia et al. developed a chemometric method for the determination of acidity and peroxide index in edible oils of different types and origins by using the same analytical technique [13]. Therefore, the combination of mid-IR spectroscopy and chemometric techniques furnish a powerful tool to monitor a large variety of processes. Its interest is increasing for quality control purposes, because it is a fast technique, there is no (or little) need of sample pre-treatment and it offers good results to discriminate and determine physico-chemical properties of edible oils. On the other hands, it is necessary a specialist on chemometrics in the group of research or in the industry to manipulate the calculations or to interpret the results. Nowadays, there are friendly commercial packages that can be used for solve this kind of problem. In this work, we offer a different methodology to evaluate the quality parameters and classification of edible oils. Therefore, the aim of our study was the simultaneous evaluation of edible oil quality parameters, such as refraction index and relative density values, by means of a fast and clean methodology based on FT-mid-IR spectroscopy combined with chemometric analysis (classification of edible oils by soft independent modeling of class analogies (SIMCA), support vector machines-discriminant analysis (SVM-DA), partial least squares-discriminant analysis (DPLS) and quantification of the quality parameters by partial least squares (PLS) regression and support vector machines (SVM) regression. 2. Chemometric theory 2.1. Preprocessing methods In this study some pre-treatments were used in various steps of the analysis being presented in a summarized below. 2.1.1. Mean center (MC) The MC preprocessing is performed by calculating the average data vector or spectrum of all n rows in a data set and subtracting it point by point from each row vector in the data set [14]. 2.1.2. Multiplicative scatter correction (MSC) Multiplicative signal correction is a processing step that attempts to account for scaling and offset effects. This correction is achieved by regressing a measured spectrum against a reference spectrum and then correcting the measured spectrum using the slope and intercept of this fit [15]. Its purpose is to correct, with a mathematical function, the noisy spectral dispersion and reflection. The basis of the MSC is the fact that the wavelength of light scattering is different from the wavelength of light absorption. 2.1.3. Standard normal variate (SNV) The standard normal variate normalization is a weighted normalization. SNV calculates the standard deviation of all the pooled variables for a given spectrum. The entire spectrum is then normalized by this value, thus giving the spectrum a unit standard deviation [15]. It was developed to reduce deviation of the baseline signal and scattering in NIR spectroscopy [16]. It is applied to each spectrum individually, which is centered in the middle and then standardized by standard deviation, having a zero mean and variance) [17]. 2.1.4. Orthogonal signal correction (OSC) Orthogonal signal correction (OSC) was used to remove systematic variation from the response matrix that is unrelated, or
orthogonal, to the property matrix. Therefore, one can be certain that important information regarding the analyte is retained [16]. 2.1.5. Savitsky–Golay Savitsky–Golay method is a derivative and smoothing method in which a polynomial (usually of first or second degree) is fitted by least squares regression, to the points in a spectral window. The window size and the degree of the polynomial are critical, because when a high number of points are chosen some relevant information can be removed and if a low number of point are chosen, the noise can still kept in the spectral data [14]. 2.2. Chemometric methods 2.2.1. Principal component analysis (PCA) Principal component analysis is a very popular tool for data compression and information extraction. PCA finds combinations of variables that describe principal trends in the data. If X is a data matrix with n rows and m columns, and with each variable being a column and each sample a row, PCA decomposes X as the sum of r ti · pTi products, where r is the rank of X and E is the matrix of residuals: X = t1 pT1 + t2 pT2 + · · · + tk pTk + · · · + tr pTr + E
(1)
Here r is less than or equal to the smaller dimension of X. The ti · pTi pairs are ordered by the amount of variance captured, ti is the scores vector and pi is the loadings vector [15]. The first principal component, PC1, covers the maximum information direction and is orthogonal (that is, explains complementary information) to PC2. This PC2 is orthogonal and explains more information than PC3, and so on. Since noise and measurement error are modeled by the latest PCs, the graphical representation of the scores on the first PCs shows the most significant variability among samples while minimizing the effect of noise [14]. 2.2.2. Soft Independent modeling of class analogies (SIMCA) SIMCA uses PCA to model the shape and position of the samples that constitute a class. A multidimensional box is build for each class and the classification of future samples is carry out by determining within which box, if any, the sample lies [18]. The best classification performance for SIMCA is obtained when there is a largest distance between the classes [14]. 2.2.3. Support vector machines-discriminant analysis (SVM-DA) SVM is a classification technique. In the application of SVM, the input patterns are transformed to a feature vector of high dimensionality, whose goal is to separate the features linearly in space [19]. When SVM is used for discriminant analysis, its function is to identify the samples in the training set and the similarity of each item has, compared to class that is informed and tries to separate the samples of the validation set of classes that the analyst suggests. SVM-DA is a technique that is used for binary classification but supports classification of multiple classes [14]. 2.2.4. Partial least squares–discriminant analysis (DPLS) In DPLS, a PLS regression model is calculated that relates the independent variables (e.g. spectra) to an integer y that designates the class of the sample. For example, the number one (1) is used to indicate that the training sample belongs to the class of interest, and a zero (0) indicates that the sample belongs to a different class. Classification of an unknown sample is derived from the value predicted by the PLS model, yˆ . This value is a real number, not an integer, which should be ideally close to the values used to codify the class (here either 0 or 1). A cut-off value between 0 and 1 is established so that a sample is assigned to class 1 if the prediction is larger than the cut-off value, or assigned to class 0 otherwise. The
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
simplest approach is to use an arbitrary cut-off value, such as 0.5, although other cut-off values can be used [20]. 2.2.5. Partial least squares regression The PLS is a multivariate regression method widely used in spectroscopic analysis. PLS aims at building a linear regression model by relating the spectral data (X) with a physico-chemical variable (y) [21]. Like PCA, PLS also calculates latent variables (LV). Unlike the PCA, the PLS decomposition of X matrix of dimension (n × m), is guided, during regression, by the variation in Y dimension array (n × z). That is, the covariance between X and Y is maximized, so that the variation in X correlates directly with Y. The purpose of PLS is to build a linear model to enable the prediction of y (the chemical/physical variable) from x (a measured spectrum). The linear model between the matrix Y containing the centered reference data and the matrix X containing the centered spectra data can be described by Y = Xb + E
(2)
where b is a vector which contains the regression coefficients that are determined during the calibration and E is the matrix of residuals. In order to obtain a good estimation of b, the PLS model needs to be calibrated on samples which span the variation in Y well and in general are representative of the future samples. Depending of the complexity of the future samples, this may require a huge number of samples. 2.2.6. Support vector machine regression SVM regression performs calibration and application of SVM models. These are non-linear models which can be used for regression or classification problems. The model consists of a number of support vectors (essentially samples selected from the calibration set) and non-linear model coefficients which define the non-linear mapping of variables in the input X-block to allow prediction of either the continuous Y-block variable for regression problems [22]. 2.3. Figures of merit Researchers of multivariate calibration typically use different strategies for determining the prediction error for a model. In this work, three figures of merit for estimating errors in y are discussed. They are (a) the root mean square error of calibration (RMSEC), (b) the root mean square error of prediction (RMSEP), and (c) the root mean square error of cross-validation (RMSECV) [23]. The RMSEC describes the degree of agreement between the estimated values for the calibration samples and the accepted true values for the calibration samples used to obtain the model parameters in Eq. (3), according to
RMSEC =
1 2 (yi − yˆ i ) n−m−1 n
1/2 (3)
i=1
If the intercept b0 is omitted from the calibration model, then the number of degree of freedom for RMSEC is n − m. If the data has been mean-centered, the degrees of freedom remain n − m − 1 [23]. To obtain the RMSEP, validation samples prepared and measured independently are used. The number of validation samples, p, should be large, so that the estimated prediction error accurately reflects all sources of variability in the calibration model. The RMSEP is computed by
RMSEP =
1 2 (yi − yˆ i ) p p
i=1
1/2 (4)
111
The cross-validation approach can also be used to estimate the predictive ability of a calibration model. One method of crossvalidation is leave-one-out cross-validation (LOOCV). LOOCV is performed by estimating n calibration models, where each of the n calibration samples is left out one at a time in turn. The resulting calibration models are then used to estimate the sample left out, which acts as an independent validation sample and provides an independent prediction of each yi value, yˆ (i) , where the notation i indicates that the ith sample was left out during model estimation. This process of leaving a sample out is repeated until all of the calibration samples have been left out. The predictions yˆ (i) can be used in Eq. (3) to estimate the RMSECV [23]. The coefficient of determination of prediction could be used to examine the predictive value of a regression model. This parameter is similar to the multiple coefficient of determination, and it is expressed by Eq. (5). It is denoted Rp2 and is defined, for np new observations, as follows:
np Rp2 = 1 −
i=1 np i=1
(yi − yˆ i )
2
(yi − yˆ i )
2
(5)
In this relation, yi indicates the actual value of the dependent variable for the new individual i (i = 1, . . ., np ). yˆ i is the predicted value for this individual given by the regression model, y¯ is the arithmetic mean of n observations of the dependent variable in the sample which was used to establish the model. 3. Experimental 3.1. Samples A first group of 103 edible oil samples were used for building the chemometric models. Twenty-three of these samples are canola oil samples; 25 of them are sunflower oils; 28 are corn oil and the rest of the samples (27) are soybean oil. It provided a representative set of edible oils consumed in Rio de Janeiro city and was used to test the robustness of the chemometric models as a function of the variability due to the oil nature and different origin. All the samples were purchased in different local supermarkets at Rio de Janeiro city, Brazil. 3.2. Measurement of physico-chemical properties of edible vegetable oils The oil samples were kept at room temperature and protected from light in a cabinet during the period elapsed since the purchase until analysis. The measurement of the relative density was carried out at a temperature of 20 ◦ C, using the model 5000 Densimeter from Anton Paar (Austria). The measurement of refractive index was carried out at a temperature of 40 ◦ C, using the model J357 Automatic Refractometer from Rudolf Research (USA). The procedure adopted in this work to measure the physico-chemical properties of edible oils was based on the official methods [24]. The official methods accepted a standard deviation equal to 0.0002 for refractive index and did not mention for relative density [24]. 3.3. Spectra collecting The absorption spectra in the mid-infrared region were obtained with the Fourier transform infrared spectrometer model 100 of PerkinElmer. The measurements were taken in KBr cells of 0.59 mm optical path and absorbance measurements were made in the range 4000–450 cm−1 , with nominal resolution of 4 cm−1 , resulting in a spectrum with 3551 points. The final spectra were the average of 20 scans. Before the conversion of spectral data to ASCII format, each
112
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
Fig. 1. Fourier transform mid-infrared spectrum of the edible oil samples.
spectrum was baseline-aligned using the software (Spectrum 6 ES, Perkin-Elmer, CT, USA)
the stretching of C O (typical of ester) bond are responsible to the absorption bands at 1240 and 1170 cm−1 . The absorption band at 720 cm−1 is attributable to the (CH2 )n group for n ≥ 4 [25].
3.4. Data analysis 4.2. Principal component analysis of FT-mid-IR spectral data Chemometric analysis, including classification and quantification, was carried out using the Solo software, version 6.3 (Eigenvector Research Inc., USA). The FT-mid-IR data spectra could be imported by the software directly.
In this study, PCA was used to examine data sets for expected or unexpected clusters, including the presence of outliers. The PCA plot of all samples did not show the need for removal of any sample (Fig. 2).
4. Results and discussion 4.1. FT-MIR spectra analysis results
4.3. Prediction of the quality parameters using FT-mid-IR spectra data
Fig. 1 shows FT-MIR spectra of canola, sunflower, soybean and corn oils. FT-MIR spectra of edible oils are very similar. The most intense absorption bands are clearly observed at 2930, 2850, 1750 around 3020, 1170 and 1470 cm−1 together with small absorption bands in the range between 1380 and 720 cm−1 . Absorption bands at 1750 cm−1 arise from the C O stretching vibrations, while those at 3020, 2930 and 2850 cm−1 are attributable to the CH stretching vibrations of CH3 , and CH2 . Absorption bands at 1470 and 1380 are attributable to the bending vibration of C H as the same time
The set of samples of edible oils was divided into two groups: calibration and validation sets. We used the algorithm of Kennard–Stone (KS) [26] for the selection of samples for each type of oil. The calibration set was built with 15 canola oil, 15 sunflower oil, 18 corn oil and 19 soybean oil samples whereas the validation set was built with 8 canola oil, 10 sunflower oil, 10 corn oil and 8 soybean oil samples. After selection of the oil samples different types of pre-processing were applied to the spectra before building the regression models: mean center (MC), multiplicative scatter
Fig. 2. PCA of the matrix of edible oil samples showing PC1 × PC2.
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
113
Table 1 Figures of merit for refractive index obtained after the use of PLS and SVM regression models using FT-mid-IR spectra data. Figures of merit
LV RMSEC RMSEP 2 RPred
PLS/SVM (types of preprocessing) MC
MC + MSC
MC + SNV
MC + OSC
MC + 1st. Der
3/– 0.0002/0.0001 0.0002/0.0001 0.9323/0.9769
6/– 0.0001/0.0001 0.0001/0.0001 0.9730/0.9790
3/– 0.0001/0.0001 0.0001/0.0001 0.9730/0.9792
2/– 0.0001/0.0001 0.0001/0.0001 0.9804/0.9797
6/– 0.0001/0.0001 0.0001/0.0001 0.9622/0.9411
Obs. LV = latente variable. Table 2 Figures of merit for relative density obtained after the use of PLS and SVM regression models using FT-mid-IR spectra data. Figures of merit
LV RMSECV RMSEP 2 RPred
PLS/SVM (Types of preprocessing) MC
MC + MSC
MC + SNV
MC + OSC
MC + 1st. Der
5/– 0.0002/0.0001 0.0002/0.0001 0.9613/0.9808
6/– 0.0001/0.0001 0.0002/0.0001 0.9811/0.9916
2/– 0.0001/0.0001 0.0002/0.0001 0.9810/0.9910
4/– 0.0001/0.0001 0.0001/0.0001 0.9911/0.9910
5/– 0.0002/0.0003 0.0002/0.0003 0.9698/0.9568
Obs. LV = latente variable. Table 3 Percentage of edible oils samples correctly classified obtained after the application of SVM-DA, DPLS and SIMCA techniques, respectively. Types of preprocessing
SVM-DA/DPLS/SIMCA Canola
Sunflower
Corn
Soybean
Training set MC MC + MSC MC + SNV MC + 1st. Der
100/100/100 100/100/100 100/100/100 100/100/100
100/94/100 93/100/100 100/100/100 93/100/100
100/100/100 94/100/100 100/100/100 100/100/100
100/100/100 100/100/100 100/100/100 100/100/100
Validation set MC MC + MSC MC + SNV MC + 1st. Der
100/100/75 100/100/100 100/100/100 88/100/88
90/100/80 90/100/100 100/100/100 90/100/100
100/90/100 100/100/100 90/100/100 100/100/100
100/100/100 100/100/88 100/100/88 100/100/100
correction (MSC), standard normal variate (SNV), orthogonal signal correction (OSC), derivate (1st. Der), alone or in combination. Tables 1 and 2 show the figures of merit obtained after the application of PLS and SVM regression methods. For the refractive index, good figures of merit were obtained with the R2 of prediction, greater than 0.98 and 0.97, respectively when PLS and SVM were used after pre-processing the data by mean centering followed by orthogonal signal correction. The RMSEP values were equal or slightly lower those obtained by official methods. For the relative density, good figures of merit were obtained with the coefficient of determination of prediction, both greater than 0.99, when PLS and SVM were used after pre-processing the data by mean centering followed by orthogonal signal correction. The RMSEP were also equal but slightly lower than those obtained by official methods.
4.4. Classification of edible oils using SVM-DA, PLS-DA and SIMCA using FT-mid-IR spectra data The same procedure described above (Section 4.3) was used for the classification of edible oils samples. Table 3 shows the percentage of edible oils samples correctly classified the SVM-DA, PLS-DA and SIMCA techniques. The methods PLS-DA and SVM-DA were able to provide a 100% accuracy in the classification of edible oils for the training and validation sets. The first method above mentioned gave the same performance for all combinations of preprocessing types whereas the second only for the combination of mean center and orthogonal signal correction.
5. Conclusion The following conclusions can be drawn from this study: (a) The task of edible oil (canola, corn, soybean and sunflower oils samples) classification by type can be successfully solved by PLS-DA and SVM-DA techniques based on FT-MIR spectroscopic data using the combination of preprocessing mean center with orthogonal signal correction. The classification error per class was 0% for both techniques. (b) For the refractive index, good figures of merit were obtained with the coefficient of determination of prediction, greater than 0.98 and 0.97, respectively when PLS and SVM were used after pre-processing of mean center combined with orthogonal signal correction for FT-MIR spectra data. (c) For the relative density, good figures of merit were obtained with the coefficient of determination of prediction, both greater than 0.99, when PLS and SVM were used after pre-processing of mean center combined with orthogonal signal correction for FT-MIR spectra data. (d) For both parameters, the RMSEP values were equal or slightly lower those obtained by official methods. Therefore, these methodology can be applied to classify edible oils (canola, sunflower, corn and soybean) with successful and can be also used to estimate the refractive index and relative density. Acknowledgements ASL thanks for FAPERJ for financial support, and CNPq and UERJ – Programa Prociência for Research Grants.
114
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
References [1] R. Karoui, G. Downey, C. Blecker, Chemical Reviews 110 (2010) 6144–6168. [2] A.F.C. Pereira, M.J.C. Pontes, F.F.G. Neto, S.R.B. Santos, R.K.H. Galvão, M.C.U. Araujo, Food Research International 41 (2008) 341–348. [3] N. Pasadakis, S. Sourligas, Ch. Fotteinopoulos, Fuel 85 (2006) 1131–1137. [4] S.M. Obeidat, M.S. Khanfar, W.M. Obeidat, Australian Journal of Basic and Applied Sciences 3 (2009) 2048–2053. [5] N. Vlachos, Y. Skopelitis, M. Psaroudaki, V. Konstantinidou, A. Chatzilazarou, E. Tegou, Analytica Chimica Acta 573–574 (2006) 459–465. [6] O. Eddib, G. Nickless, Analyst 112 (1987) 391–395. [7] L. Kryger, Talanta 28 (1981) 871–887. [8] V. Baeten, M. Meurens, M.T. Morales, R.J. Aparicio, Journal of Agricultural and Food Chemistry 44 (1996) 2225–2230. [9] A.D. Shaw, A. di Camillo, G. Vlahov, A. Jones, G. Bianchi, J. Rowland, D.B. Kell, Analytica Chimica Acta 348 (1997) 357–374. [10] F.F. Gambarra-Neto, G. Marino, M.C.U. Araújo, R.K.H. Galvão, M.J.C. Pontes, E.P. de Medeiros, R.S. Lima, Talanta 77 (2009) 1660–1666. [11] V. Baeten, R. Aparicio, Biotechnology, Agronomy, Society and Environment 4 (2000) 196–203. [12] R.M. El-Abassy, P. Donfack, A. Materny, Journal of Raman Specroscopy 40 (2009) 1284–1289. [13] S. Armenta, S. Garrigues, M. de la Guardia, Analytica Chimica Acta 596 (2007) 330–337.
[14] A.S. Luna, A.P. da Silva, J.S.A. Pinho, J. Ferré, R. Boqué, Spectrochim. Acta: Mol. Biomol. Spectrosc. (2012), http://dx.doi.org/10.1016/j.saa.2012.02.085. [15] K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics A Practical Guide, John Wiley & Sons, Inc., NY, 1998. [16] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Journal of Near Infrared Spectroscopy 1 (1993) 185–186. [17] A. Moghimi, M.H. Aghkhani, A. Sazgarnia, M. Sarmad, Biosystems Engineering 106 (2010) 295–302. [18] J. Sjöblom, O. Svensson, M. Josefson, H. Kullberg, S. Wold, Chemometrics and Intelligent Laboratory Systems 44 (1998) 229–244. [19] F.L. Baronoski, MSc dissertation, Pontifícia Universidade Católica do Paraná, Paraná, Brazil, 2005. [20] N.F. Pérez, J. Ferré, R. Boqué, Chemometrics and Intelligent Laboratory Systems 95 (2009) 122–128. [21] R.W. Gerlach, B.R. Kowalski, H.O.A. Wold, Analytica Chimica Acta 112 (1979) 417–421. [22] PLS Toolbox, version 6.5, Eigenvector Research, Inc., WA, USA. [23] P. Gemperline (Ed.), Practical Guide to Chemometrics, 2nd ed., CRC Press, Boca Raton, USA, 2006. [24] O. Zenebon, N.S. Pascuet, P. Tiglea (Eds.), Métodos Físico-Químicos para Análise de Alimentos, 4th ed., Instituto Adolfo Lutz, São Paulo, 2008. [25] F. Settle (Ed.), Handbook of Instrumental Techniques for Analytical Chemistry, Prentice-Hall, Inc., NJ, USA, 1997. [26] F. Sales, M.P. Callao, F.X. Ruis, Chemometrics and Intelligent Laboratory Systems 38 (1997) 63–73.
Contents lists available at SciVerse ScienceDirect
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy journal homepage: www.elsevier.com/locate/saa
Classification of edible oils and modeling of their physico-chemical properties by chemometric methods using mid-IR spectroscopy Aderval S. Luna a,∗ , Arnaldo P. da Silva a , Joan Ferré b , Ricard Boqué b a b
Rio de Janeiro State University, Department of Analytical Chemistry, Rua São Francisco Xavier, 524, Maracanã, Rio de Janeiro CEP 20550-013, Brazil Universitat Rovira i Virgili, Department of Analytical Chemistry and Organic Chemistry, C/Marcel.lí Domingo s/n–43007, Tarragona, Spain
a r t i c l e
i n f o
Article history: Received 30 November 2011 Received in revised form 4 May 2012 Accepted 25 June 2012 Keywords: Edible oils Fourier transform mid-infrared spectroscopy First order calibration models Refraction index Relative density
a b s t r a c t This research work describes two studies for the classification and characterization of edible oils and its quality parameters through Fourier transform mid infrared spectroscopy (FT-mid-IR) together with chemometric methods. The discrimination of canola, sunflower, corn and soybean oils was investigated using SVM-DA, SIMCA and PLS-DA. Using FT-mid-IR, DPLS was able to classify 100% of the samples from the validation set, but SIMCA and SVM-DA were not. The quality parameters: refraction index and relative density of edible oils were obtained from reference methods. Prediction models for FT-mid-IR spectra were calculated for these quality parameters using partial least squares (PLS) and support vector machines (SVM). Several preprocessing alternatives (first derivative, multiplicative scatter correction, mean centering, and standard normal variate) were investigated. The best result for the refraction index was achieved with SVM as well as for the relative density except when the preprocessing combination of mean centering and first derivative was used. For both of quality parameters, the best results obtained for the figures of merit expressed by the root mean square error of cross validation (RMSECV) and prediction (RMSEP) were equal to 0.0001. © 2012 Elsevier B.V. All rights reserved.
1. Introduction Public interest in food quality and methods of production has increased significantly in the recent decades, due in part to changes in eating habits, consumer behavior, and the increased industrialization and globalization of food supply chains. Demand for high levels of quality and safety in food production obviously requires high standards in quality assurance and process control; satisfying this demand in turn requires appropriate analytical tools for food analysis both during and after production. Desirable features of such tools include speed, ease-of-use, minimal or no sample preparation, and the avoidance of sample destruction. These features are characteristic of a range of spectroscopic methods including the mid-infrared (MIR) and near infrared (NIR) [1]. The laboratory routine measurements of sample properties in edible oils are based on well-established standard methods, following procedures determined by international organizations like the American oil chemists’ society (AOCS) and national as Instituto Adolpho Lutz (Brazil). Although these methods are widely accepted, in many cases they exhibit several disadvantages.
∗ Corresponding author. Tel.: +55 21 23340563; fax: +55 21 23340159. E-mail address: [email protected] (A.S. Luna). 1386-1425/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.saa.2012.06.034
While fast and reliable methods are needed to monitor the quality of the various streams in a factory, most of the standard test methods cannot be applied on-line. An alternative to these limitations is the use of spectroscopic techniques in on-line or off-line mode in conjunction with multivariate calibration schemes [2,3]. Olive oil adulteration adulteration with less expensive edible oils is a major issue for the olive oil industry. Obeidat et al. used Fourier transform infrared (FTIR) spectroscopy to classify different edible oils including the virgin olive oil [4]. Tegou et al. used the same analytical technique to determine extra virgin olive oil adulteration with lower priced vegetable oils (sunflower oil, soyabean oil, sesame oil, and corn oil) [5]. Edible oils classification has also been performed by chemometric treatment of chromatographic profiles [6,7], Fourier transform Raman spectra [8] and 13 C NMR spectra [9]. Araújo et al. proposed an electroanalytical methodology for classification of edible vegetable oils with respect to type (canola, sunflower, corn and soybean) and conservation state (expired and non-expired shelf life) [10]. Baeten and Aparicio used Raman spectroscopy to discriminate virgin olive oil from other edible oils [11] and Materny et al. used visible Raman spectroscopy for the discrimination of olive oils from different vegetable oils and the detection of adulteration [12]. Many older tests involved determination of physical properties such as refractive index, melting point, and viscosity. In this context, Araújo et al. proposed an analytical method by using near
110
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
infrared spectroscopy (NIR) measurements for determination of acidity, refractive index and viscosity in four types of edible vegetable oils (corn, soya, canola and sunflower) [2] and de la Guardia et al. developed a chemometric method for the determination of acidity and peroxide index in edible oils of different types and origins by using the same analytical technique [13]. Therefore, the combination of mid-IR spectroscopy and chemometric techniques furnish a powerful tool to monitor a large variety of processes. Its interest is increasing for quality control purposes, because it is a fast technique, there is no (or little) need of sample pre-treatment and it offers good results to discriminate and determine physico-chemical properties of edible oils. On the other hands, it is necessary a specialist on chemometrics in the group of research or in the industry to manipulate the calculations or to interpret the results. Nowadays, there are friendly commercial packages that can be used for solve this kind of problem. In this work, we offer a different methodology to evaluate the quality parameters and classification of edible oils. Therefore, the aim of our study was the simultaneous evaluation of edible oil quality parameters, such as refraction index and relative density values, by means of a fast and clean methodology based on FT-mid-IR spectroscopy combined with chemometric analysis (classification of edible oils by soft independent modeling of class analogies (SIMCA), support vector machines-discriminant analysis (SVM-DA), partial least squares-discriminant analysis (DPLS) and quantification of the quality parameters by partial least squares (PLS) regression and support vector machines (SVM) regression. 2. Chemometric theory 2.1. Preprocessing methods In this study some pre-treatments were used in various steps of the analysis being presented in a summarized below. 2.1.1. Mean center (MC) The MC preprocessing is performed by calculating the average data vector or spectrum of all n rows in a data set and subtracting it point by point from each row vector in the data set [14]. 2.1.2. Multiplicative scatter correction (MSC) Multiplicative signal correction is a processing step that attempts to account for scaling and offset effects. This correction is achieved by regressing a measured spectrum against a reference spectrum and then correcting the measured spectrum using the slope and intercept of this fit [15]. Its purpose is to correct, with a mathematical function, the noisy spectral dispersion and reflection. The basis of the MSC is the fact that the wavelength of light scattering is different from the wavelength of light absorption. 2.1.3. Standard normal variate (SNV) The standard normal variate normalization is a weighted normalization. SNV calculates the standard deviation of all the pooled variables for a given spectrum. The entire spectrum is then normalized by this value, thus giving the spectrum a unit standard deviation [15]. It was developed to reduce deviation of the baseline signal and scattering in NIR spectroscopy [16]. It is applied to each spectrum individually, which is centered in the middle and then standardized by standard deviation, having a zero mean and variance) [17]. 2.1.4. Orthogonal signal correction (OSC) Orthogonal signal correction (OSC) was used to remove systematic variation from the response matrix that is unrelated, or
orthogonal, to the property matrix. Therefore, one can be certain that important information regarding the analyte is retained [16]. 2.1.5. Savitsky–Golay Savitsky–Golay method is a derivative and smoothing method in which a polynomial (usually of first or second degree) is fitted by least squares regression, to the points in a spectral window. The window size and the degree of the polynomial are critical, because when a high number of points are chosen some relevant information can be removed and if a low number of point are chosen, the noise can still kept in the spectral data [14]. 2.2. Chemometric methods 2.2.1. Principal component analysis (PCA) Principal component analysis is a very popular tool for data compression and information extraction. PCA finds combinations of variables that describe principal trends in the data. If X is a data matrix with n rows and m columns, and with each variable being a column and each sample a row, PCA decomposes X as the sum of r ti · pTi products, where r is the rank of X and E is the matrix of residuals: X = t1 pT1 + t2 pT2 + · · · + tk pTk + · · · + tr pTr + E
(1)
Here r is less than or equal to the smaller dimension of X. The ti · pTi pairs are ordered by the amount of variance captured, ti is the scores vector and pi is the loadings vector [15]. The first principal component, PC1, covers the maximum information direction and is orthogonal (that is, explains complementary information) to PC2. This PC2 is orthogonal and explains more information than PC3, and so on. Since noise and measurement error are modeled by the latest PCs, the graphical representation of the scores on the first PCs shows the most significant variability among samples while minimizing the effect of noise [14]. 2.2.2. Soft Independent modeling of class analogies (SIMCA) SIMCA uses PCA to model the shape and position of the samples that constitute a class. A multidimensional box is build for each class and the classification of future samples is carry out by determining within which box, if any, the sample lies [18]. The best classification performance for SIMCA is obtained when there is a largest distance between the classes [14]. 2.2.3. Support vector machines-discriminant analysis (SVM-DA) SVM is a classification technique. In the application of SVM, the input patterns are transformed to a feature vector of high dimensionality, whose goal is to separate the features linearly in space [19]. When SVM is used for discriminant analysis, its function is to identify the samples in the training set and the similarity of each item has, compared to class that is informed and tries to separate the samples of the validation set of classes that the analyst suggests. SVM-DA is a technique that is used for binary classification but supports classification of multiple classes [14]. 2.2.4. Partial least squares–discriminant analysis (DPLS) In DPLS, a PLS regression model is calculated that relates the independent variables (e.g. spectra) to an integer y that designates the class of the sample. For example, the number one (1) is used to indicate that the training sample belongs to the class of interest, and a zero (0) indicates that the sample belongs to a different class. Classification of an unknown sample is derived from the value predicted by the PLS model, yˆ . This value is a real number, not an integer, which should be ideally close to the values used to codify the class (here either 0 or 1). A cut-off value between 0 and 1 is established so that a sample is assigned to class 1 if the prediction is larger than the cut-off value, or assigned to class 0 otherwise. The
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
simplest approach is to use an arbitrary cut-off value, such as 0.5, although other cut-off values can be used [20]. 2.2.5. Partial least squares regression The PLS is a multivariate regression method widely used in spectroscopic analysis. PLS aims at building a linear regression model by relating the spectral data (X) with a physico-chemical variable (y) [21]. Like PCA, PLS also calculates latent variables (LV). Unlike the PCA, the PLS decomposition of X matrix of dimension (n × m), is guided, during regression, by the variation in Y dimension array (n × z). That is, the covariance between X and Y is maximized, so that the variation in X correlates directly with Y. The purpose of PLS is to build a linear model to enable the prediction of y (the chemical/physical variable) from x (a measured spectrum). The linear model between the matrix Y containing the centered reference data and the matrix X containing the centered spectra data can be described by Y = Xb + E
(2)
where b is a vector which contains the regression coefficients that are determined during the calibration and E is the matrix of residuals. In order to obtain a good estimation of b, the PLS model needs to be calibrated on samples which span the variation in Y well and in general are representative of the future samples. Depending of the complexity of the future samples, this may require a huge number of samples. 2.2.6. Support vector machine regression SVM regression performs calibration and application of SVM models. These are non-linear models which can be used for regression or classification problems. The model consists of a number of support vectors (essentially samples selected from the calibration set) and non-linear model coefficients which define the non-linear mapping of variables in the input X-block to allow prediction of either the continuous Y-block variable for regression problems [22]. 2.3. Figures of merit Researchers of multivariate calibration typically use different strategies for determining the prediction error for a model. In this work, three figures of merit for estimating errors in y are discussed. They are (a) the root mean square error of calibration (RMSEC), (b) the root mean square error of prediction (RMSEP), and (c) the root mean square error of cross-validation (RMSECV) [23]. The RMSEC describes the degree of agreement between the estimated values for the calibration samples and the accepted true values for the calibration samples used to obtain the model parameters in Eq. (3), according to
RMSEC =
1 2 (yi − yˆ i ) n−m−1 n
1/2 (3)
i=1
If the intercept b0 is omitted from the calibration model, then the number of degree of freedom for RMSEC is n − m. If the data has been mean-centered, the degrees of freedom remain n − m − 1 [23]. To obtain the RMSEP, validation samples prepared and measured independently are used. The number of validation samples, p, should be large, so that the estimated prediction error accurately reflects all sources of variability in the calibration model. The RMSEP is computed by
RMSEP =
1 2 (yi − yˆ i ) p p
i=1
1/2 (4)
111
The cross-validation approach can also be used to estimate the predictive ability of a calibration model. One method of crossvalidation is leave-one-out cross-validation (LOOCV). LOOCV is performed by estimating n calibration models, where each of the n calibration samples is left out one at a time in turn. The resulting calibration models are then used to estimate the sample left out, which acts as an independent validation sample and provides an independent prediction of each yi value, yˆ (i) , where the notation i indicates that the ith sample was left out during model estimation. This process of leaving a sample out is repeated until all of the calibration samples have been left out. The predictions yˆ (i) can be used in Eq. (3) to estimate the RMSECV [23]. The coefficient of determination of prediction could be used to examine the predictive value of a regression model. This parameter is similar to the multiple coefficient of determination, and it is expressed by Eq. (5). It is denoted Rp2 and is defined, for np new observations, as follows:
np Rp2 = 1 −
i=1 np i=1
(yi − yˆ i )
2
(yi − yˆ i )
2
(5)
In this relation, yi indicates the actual value of the dependent variable for the new individual i (i = 1, . . ., np ). yˆ i is the predicted value for this individual given by the regression model, y¯ is the arithmetic mean of n observations of the dependent variable in the sample which was used to establish the model. 3. Experimental 3.1. Samples A first group of 103 edible oil samples were used for building the chemometric models. Twenty-three of these samples are canola oil samples; 25 of them are sunflower oils; 28 are corn oil and the rest of the samples (27) are soybean oil. It provided a representative set of edible oils consumed in Rio de Janeiro city and was used to test the robustness of the chemometric models as a function of the variability due to the oil nature and different origin. All the samples were purchased in different local supermarkets at Rio de Janeiro city, Brazil. 3.2. Measurement of physico-chemical properties of edible vegetable oils The oil samples were kept at room temperature and protected from light in a cabinet during the period elapsed since the purchase until analysis. The measurement of the relative density was carried out at a temperature of 20 ◦ C, using the model 5000 Densimeter from Anton Paar (Austria). The measurement of refractive index was carried out at a temperature of 40 ◦ C, using the model J357 Automatic Refractometer from Rudolf Research (USA). The procedure adopted in this work to measure the physico-chemical properties of edible oils was based on the official methods [24]. The official methods accepted a standard deviation equal to 0.0002 for refractive index and did not mention for relative density [24]. 3.3. Spectra collecting The absorption spectra in the mid-infrared region were obtained with the Fourier transform infrared spectrometer model 100 of PerkinElmer. The measurements were taken in KBr cells of 0.59 mm optical path and absorbance measurements were made in the range 4000–450 cm−1 , with nominal resolution of 4 cm−1 , resulting in a spectrum with 3551 points. The final spectra were the average of 20 scans. Before the conversion of spectral data to ASCII format, each
112
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
Fig. 1. Fourier transform mid-infrared spectrum of the edible oil samples.
spectrum was baseline-aligned using the software (Spectrum 6 ES, Perkin-Elmer, CT, USA)
the stretching of C O (typical of ester) bond are responsible to the absorption bands at 1240 and 1170 cm−1 . The absorption band at 720 cm−1 is attributable to the (CH2 )n group for n ≥ 4 [25].
3.4. Data analysis 4.2. Principal component analysis of FT-mid-IR spectral data Chemometric analysis, including classification and quantification, was carried out using the Solo software, version 6.3 (Eigenvector Research Inc., USA). The FT-mid-IR data spectra could be imported by the software directly.
In this study, PCA was used to examine data sets for expected or unexpected clusters, including the presence of outliers. The PCA plot of all samples did not show the need for removal of any sample (Fig. 2).
4. Results and discussion 4.1. FT-MIR spectra analysis results
4.3. Prediction of the quality parameters using FT-mid-IR spectra data
Fig. 1 shows FT-MIR spectra of canola, sunflower, soybean and corn oils. FT-MIR spectra of edible oils are very similar. The most intense absorption bands are clearly observed at 2930, 2850, 1750 around 3020, 1170 and 1470 cm−1 together with small absorption bands in the range between 1380 and 720 cm−1 . Absorption bands at 1750 cm−1 arise from the C O stretching vibrations, while those at 3020, 2930 and 2850 cm−1 are attributable to the CH stretching vibrations of CH3 , and CH2 . Absorption bands at 1470 and 1380 are attributable to the bending vibration of C H as the same time
The set of samples of edible oils was divided into two groups: calibration and validation sets. We used the algorithm of Kennard–Stone (KS) [26] for the selection of samples for each type of oil. The calibration set was built with 15 canola oil, 15 sunflower oil, 18 corn oil and 19 soybean oil samples whereas the validation set was built with 8 canola oil, 10 sunflower oil, 10 corn oil and 8 soybean oil samples. After selection of the oil samples different types of pre-processing were applied to the spectra before building the regression models: mean center (MC), multiplicative scatter
Fig. 2. PCA of the matrix of edible oil samples showing PC1 × PC2.
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
113
Table 1 Figures of merit for refractive index obtained after the use of PLS and SVM regression models using FT-mid-IR spectra data. Figures of merit
LV RMSEC RMSEP 2 RPred
PLS/SVM (types of preprocessing) MC
MC + MSC
MC + SNV
MC + OSC
MC + 1st. Der
3/– 0.0002/0.0001 0.0002/0.0001 0.9323/0.9769
6/– 0.0001/0.0001 0.0001/0.0001 0.9730/0.9790
3/– 0.0001/0.0001 0.0001/0.0001 0.9730/0.9792
2/– 0.0001/0.0001 0.0001/0.0001 0.9804/0.9797
6/– 0.0001/0.0001 0.0001/0.0001 0.9622/0.9411
Obs. LV = latente variable. Table 2 Figures of merit for relative density obtained after the use of PLS and SVM regression models using FT-mid-IR spectra data. Figures of merit
LV RMSECV RMSEP 2 RPred
PLS/SVM (Types of preprocessing) MC
MC + MSC
MC + SNV
MC + OSC
MC + 1st. Der
5/– 0.0002/0.0001 0.0002/0.0001 0.9613/0.9808
6/– 0.0001/0.0001 0.0002/0.0001 0.9811/0.9916
2/– 0.0001/0.0001 0.0002/0.0001 0.9810/0.9910
4/– 0.0001/0.0001 0.0001/0.0001 0.9911/0.9910
5/– 0.0002/0.0003 0.0002/0.0003 0.9698/0.9568
Obs. LV = latente variable. Table 3 Percentage of edible oils samples correctly classified obtained after the application of SVM-DA, DPLS and SIMCA techniques, respectively. Types of preprocessing
SVM-DA/DPLS/SIMCA Canola
Sunflower
Corn
Soybean
Training set MC MC + MSC MC + SNV MC + 1st. Der
100/100/100 100/100/100 100/100/100 100/100/100
100/94/100 93/100/100 100/100/100 93/100/100
100/100/100 94/100/100 100/100/100 100/100/100
100/100/100 100/100/100 100/100/100 100/100/100
Validation set MC MC + MSC MC + SNV MC + 1st. Der
100/100/75 100/100/100 100/100/100 88/100/88
90/100/80 90/100/100 100/100/100 90/100/100
100/90/100 100/100/100 90/100/100 100/100/100
100/100/100 100/100/88 100/100/88 100/100/100
correction (MSC), standard normal variate (SNV), orthogonal signal correction (OSC), derivate (1st. Der), alone or in combination. Tables 1 and 2 show the figures of merit obtained after the application of PLS and SVM regression methods. For the refractive index, good figures of merit were obtained with the R2 of prediction, greater than 0.98 and 0.97, respectively when PLS and SVM were used after pre-processing the data by mean centering followed by orthogonal signal correction. The RMSEP values were equal or slightly lower those obtained by official methods. For the relative density, good figures of merit were obtained with the coefficient of determination of prediction, both greater than 0.99, when PLS and SVM were used after pre-processing the data by mean centering followed by orthogonal signal correction. The RMSEP were also equal but slightly lower than those obtained by official methods.
4.4. Classification of edible oils using SVM-DA, PLS-DA and SIMCA using FT-mid-IR spectra data The same procedure described above (Section 4.3) was used for the classification of edible oils samples. Table 3 shows the percentage of edible oils samples correctly classified the SVM-DA, PLS-DA and SIMCA techniques. The methods PLS-DA and SVM-DA were able to provide a 100% accuracy in the classification of edible oils for the training and validation sets. The first method above mentioned gave the same performance for all combinations of preprocessing types whereas the second only for the combination of mean center and orthogonal signal correction.
5. Conclusion The following conclusions can be drawn from this study: (a) The task of edible oil (canola, corn, soybean and sunflower oils samples) classification by type can be successfully solved by PLS-DA and SVM-DA techniques based on FT-MIR spectroscopic data using the combination of preprocessing mean center with orthogonal signal correction. The classification error per class was 0% for both techniques. (b) For the refractive index, good figures of merit were obtained with the coefficient of determination of prediction, greater than 0.98 and 0.97, respectively when PLS and SVM were used after pre-processing of mean center combined with orthogonal signal correction for FT-MIR spectra data. (c) For the relative density, good figures of merit were obtained with the coefficient of determination of prediction, both greater than 0.99, when PLS and SVM were used after pre-processing of mean center combined with orthogonal signal correction for FT-MIR spectra data. (d) For both parameters, the RMSEP values were equal or slightly lower those obtained by official methods. Therefore, these methodology can be applied to classify edible oils (canola, sunflower, corn and soybean) with successful and can be also used to estimate the refractive index and relative density. Acknowledgements ASL thanks for FAPERJ for financial support, and CNPq and UERJ – Programa Prociência for Research Grants.
114
A.S. Luna et al. / Spectrochimica Acta Part A 100 (2013) 109–114
References [1] R. Karoui, G. Downey, C. Blecker, Chemical Reviews 110 (2010) 6144–6168. [2] A.F.C. Pereira, M.J.C. Pontes, F.F.G. Neto, S.R.B. Santos, R.K.H. Galvão, M.C.U. Araujo, Food Research International 41 (2008) 341–348. [3] N. Pasadakis, S. Sourligas, Ch. Fotteinopoulos, Fuel 85 (2006) 1131–1137. [4] S.M. Obeidat, M.S. Khanfar, W.M. Obeidat, Australian Journal of Basic and Applied Sciences 3 (2009) 2048–2053. [5] N. Vlachos, Y. Skopelitis, M. Psaroudaki, V. Konstantinidou, A. Chatzilazarou, E. Tegou, Analytica Chimica Acta 573–574 (2006) 459–465. [6] O. Eddib, G. Nickless, Analyst 112 (1987) 391–395. [7] L. Kryger, Talanta 28 (1981) 871–887. [8] V. Baeten, M. Meurens, M.T. Morales, R.J. Aparicio, Journal of Agricultural and Food Chemistry 44 (1996) 2225–2230. [9] A.D. Shaw, A. di Camillo, G. Vlahov, A. Jones, G. Bianchi, J. Rowland, D.B. Kell, Analytica Chimica Acta 348 (1997) 357–374. [10] F.F. Gambarra-Neto, G. Marino, M.C.U. Araújo, R.K.H. Galvão, M.J.C. Pontes, E.P. de Medeiros, R.S. Lima, Talanta 77 (2009) 1660–1666. [11] V. Baeten, R. Aparicio, Biotechnology, Agronomy, Society and Environment 4 (2000) 196–203. [12] R.M. El-Abassy, P. Donfack, A. Materny, Journal of Raman Specroscopy 40 (2009) 1284–1289. [13] S. Armenta, S. Garrigues, M. de la Guardia, Analytica Chimica Acta 596 (2007) 330–337.
[14] A.S. Luna, A.P. da Silva, J.S.A. Pinho, J. Ferré, R. Boqué, Spectrochim. Acta: Mol. Biomol. Spectrosc. (2012), http://dx.doi.org/10.1016/j.saa.2012.02.085. [15] K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics A Practical Guide, John Wiley & Sons, Inc., NY, 1998. [16] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Journal of Near Infrared Spectroscopy 1 (1993) 185–186. [17] A. Moghimi, M.H. Aghkhani, A. Sazgarnia, M. Sarmad, Biosystems Engineering 106 (2010) 295–302. [18] J. Sjöblom, O. Svensson, M. Josefson, H. Kullberg, S. Wold, Chemometrics and Intelligent Laboratory Systems 44 (1998) 229–244. [19] F.L. Baronoski, MSc dissertation, Pontifícia Universidade Católica do Paraná, Paraná, Brazil, 2005. [20] N.F. Pérez, J. Ferré, R. Boqué, Chemometrics and Intelligent Laboratory Systems 95 (2009) 122–128. [21] R.W. Gerlach, B.R. Kowalski, H.O.A. Wold, Analytica Chimica Acta 112 (1979) 417–421. [22] PLS Toolbox, version 6.5, Eigenvector Research, Inc., WA, USA. [23] P. Gemperline (Ed.), Practical Guide to Chemometrics, 2nd ed., CRC Press, Boca Raton, USA, 2006. [24] O. Zenebon, N.S. Pascuet, P. Tiglea (Eds.), Métodos Físico-Químicos para Análise de Alimentos, 4th ed., Instituto Adolfo Lutz, São Paulo, 2008. [25] F. Settle (Ed.), Handbook of Instrumental Techniques for Analytical Chemistry, Prentice-Hall, Inc., NJ, USA, 1997. [26] F. Sales, M.P. Callao, F.X. Ruis, Chemometrics and Intelligent Laboratory Systems 38 (1997) 63–73.