BOTANICAL ORIGIN-BASED HONEY DISCRIMINATION USING VIS-NIR SPECTROSCOPY AND STATISTICAL CLUSTER ANALYSIS Diana Tsankova 1, Svetla Lekova 2

Transcription

1 Journal Journal of Chemical of Chemical Technology and and Metallurgy, 50, 5, 50, 2015, BOTANICAL ORIGIN-BASED HONEY DISCRIMINATION USING VIS-NIR SPECTROSCOPY AND STATISTICAL CLUSTER ANALYSIS Diana Tsankova 1, Svetla Lekova 2 1 University of Food Technologies 26 Maritsa blvd, 4000 Plovdiv, Bulgaria, dtsankova@yahoo.com 2 University of Chemical Technology and Metallurgy, 8 Kliment Ohridski, Sofia 1756, Bulgaria sv_lekova@uctm.edu Received 15 February 2015 Accepted 23 July 2015 ABSTRACT The aim of the article is to present results on the potential of honey discrimination (on the base of its botanical origins) by Vis-NIR spectroscopy and subsequent statistical cluster analysis. Thirty-five samples from three types of honey (acacia, linden, and honeydew) are measured by a spectrophotometer Cary100 with recorded wavelength range of nm for calibration of honey classifier. Firstly, principal components analysis (PCA) is used for reducing the number of inputs (wavelengths) and proper visualization of the experimental results. Next, the first two principal components (PCs) are combined separately with Naïve Bayes classification (NBC) and k-means clustering (KMC) to develop PC-NBC and PC- KMC models. An additional reduction of the number of used wavelengths by selecting the values at equal, predetermined intervals by a power of 2 is proposed. The PC-NBC and PC-KMC models developed for these reduced wavelengths produce almost the same performance as compared with those developed for the original number of wavelengths and thus facilitate developing a simpler and cheaper sensor for honey discrimination in practice. The high accuracy of the proposed honey classifiers is confirmed by leave-one-out cross-validation test conducted in MATLAB environment. Keywords: honey discrimination, Vis-NIR spectroscopy, principal components analysis, Bayes classification. INTRODUCTION is a natural product and nothing should be extracted or added to it. Production of natural honey is a laborious process, which is time consuming and involves a lot of costs. Therefore, honey is often subjected to falsification by adding sugar and other impurities. Furthermore, some types of honey have a higher market price than others, so in order to prevent fraud in the labeling, means for distinguishing between different types of honey should be developed. Several methods have been used for the determination of the floral origin of honey. Among them the pollen recognition and sensory analysis are the most popular ones. However, the technique of analysis of honey s pollen content is prolonged and has some limitations. The other methods are mainly based on the analysis of honey s aroma compounds, sugar profile, flavonoid pattern, non-flavonoid phenolics, organic acids, isotopic relations, protein and amino acid compositions and marker presence [1, 2]. But some of these methods are in general time-consuming, complex, and labour intensive, for routine quality control application, or require very specialised personnel to interpret the results [3, 4]. In addition, most of the analytical techniques involve some kind of sample pre-treatment. The advantages of the technique of visible (Vis) and near infrared (NIR) spectroscopy with respect to other analytical methods are the non-invasive approach, the relatively easy and quick data acquisition. Some authors [5, 6] had used this technique for qualification of adulterants in honey with good accuracy, but prediction models were developed 638

2 Diana Tsankova, Svetla Lekova without reduction of wavelengths range, which would cause computational complexity. The aim of this work is to investigate the possibility of honey discrimination (based on its botanical origin), using Vis-NIR spectroscopy in an absorbance mode. Spectroscopic data obtained undergo subsequent statistical processing including: principal components analysis (PCA) for reducing the classifiers number of inputs; Naïve Bayes classification (NBC) and k-means clustering (KMC) for cluster distinguishing. The technology of joint use of PCA and clustering techniques from classical type as NBC and KMC, not only contributes to reduction of the area of the input data, but also overcomes some computational problems such as badly scaled or close to singular matrix. In addition, a wavelengths number reduction by selecting the values at regular, predetermined intervals by a power of 2 is proposed. This will facilitate the design of a simple and inexpensive sensor for recognizing some different types of honey in practice. EXPERIMENTAL Spectrum Acquisition Thirty-five samples of three different types of honey (acacia - 14 samples; linden - 12 samples; and honeydew - 9 samples) were purchased from supermarkets (Lexie, Kaufland, Piccadilly) and from private producers. The samples were annealed at room temperature ( C). All of them were in a liquid state (without crystallization). The spectral characteristics of the honey were taken with a spectrophotometer Cary100 ranging from 350 to 900 nm at 1 nm sampling space. The method of the PCs was used to reduce dimensionality of the input data (spectral readings), then they were classified using two classifiers - a supervisor Bayes classifier and unsupervisor k-means one. Principal Components Analysis The aim of the method is to reduce the dimensionality of multivariate data (e.g., wavelengths) whilst preserving as much of the relevant information as possible. PCA is a linear transformation, that transforms the data (observations of possibly correlated variables) to a new coordinate system such that the new set of variables, the principal components, are linear functions of the original variables. PCs are uncorrelated, and the greatest variance by any projection of the data comes to lie on the first coordinate, the second greatest variance on the second coordinate, and so on. This is achieved by computing the covariance matrix for the full data set. Then, the eigenvectors and eigenvalues of the covariance matrix are computed, and sorted according to decreasing eigenvalue [7, 8]. All the principal components are orthogonal to each other. The full set of principal components is as large as the original set of variables. Usually the sum of the variances of the first few principal components exceeds 80 % of the total variance of the original data [9]. In this study, the first two PCs are used by the following two calibration methods: NBC and KMC for developing discrimination models. Naive Bayes classification algorithm The Naive Bayes classifier is fast and easy to implement. It is designed for use when features are independent of one another within each class, but it appears to work well in practice even when that independence assumption is not valid [9-12]. It classifies data in two steps - training and prediction steps. In the training step, using the training samples, the method estimates the parameters of a probability distribution, assuming that features are conditionally independent given the class. During the prediction step, for any unseen test sample, the method computes the posterior probability of that sample belonging to each class. The method then classifies the test sample according to the largest posterior probability. In this study the NBC algorithm uses normal (Gaussian) distribution. It is appropriate for features that have normal distributions in each class. The Naive Bayes classifier estimates a separate normal distribution for each class by computing the mean and standard deviation of the training data in that class. K-means clustering K-means clustering aims to partition n observations into k clusters in which each observation belongs to the cluster with the nearest mean, serving as a prototype of the cluster. Given an initial set of k means, the algorithm proceeds by alternating between two steps - assignment and update steps [13, 14]. During the assignment step each observation is assigned to the cluster whose mean yields the least within-cluster sum of squares. Since 639

3 Journal of Chemical Technology and Metallurgy, 50, 5, 2015 the sum of squares is the squared Euclidean distance, this is the nearest mean. Mathematically, this means partitioning the observations according to the Voronoi diagram generated by the means. In the update step the new means are calculated and they have to be the centroids of the observations in the new clusters. The algorithm has converged when the assignments no longer change. Since both steps optimize the withincluster sum of squares, and there exists a finite number of such partitions, the algorithm must converge to a local optimum. There is no guarantee that the global optimum is found by k-means algorithm. clusters, that with a few exceptions coincide with the three types of honey - acacia, linden and honeydew. The figure shows that three samples of linden honey are located in the cluster of acacia honey and 1 sample of honeydew - in the cluster of linden honey. Here, the determination of the type of honey is based solely on the inscription on the label by the manufacturer, i.e. on trusting the manufacturer. The first two PCs explain as high as % of variance in the spectra (93.58 % for PC-1 and 4.07 % for PC-2). The two PCs were chosen to develop PC-NBC and PC-KMC models. One-outcross-validation test was used to check the performance RESULTS AND DISCUSSION Absorbance Spectra Absorbance spectra of the three brands of honey with wavelengths ranging from 350 to 900 nm are shown in Fig. 1. The spectra present peaks at the band of nm and low absorbance in the NIR range above 760 nm. All spectra are similar in spectral shape and absorbance. Therefore, it is necessary to apply appropriate methods of multivariate analysis to distinguish the honeys. PC-NBC and PC-KMC Based Models for Full Wavelengths Spectra The spectral dimensionality was reduced to a small number (two) of principal components using PCA. The scores scatter plot of the 1st and 2nd PCs is shown in Fig. 2. It is evident that the samples form three Fig. 2. PCA of the three types of honey. Fig. 1. Absorbance spectra. Fig. 3. PC-NBC model of the three types of honey. 640

4 Diana Tsankova, Svetla Lekova Table 1. Discrimination accuracy of PC-NBC model. PC-NBC model Observed Class Predicted Class by PC-NBC dew Total dew Total Table 2. Discrimination accuracy of PC-KMC model. PC-KMC model Observed Class Predicted Class by PC-NBC dew Total dew Total of the classifiers. The prediction results of the honey s botanical origin made by the proposed classifiers, PC- NBC and PC-KMC, are shown in Fig. 3, Table 1 and Fig. 4, Table 2, respectively. Table 3 shows the good performance of the models mentioned above (PC-NBC and PC-KMC): 88.6 % and 80.0 % accuracy, respectively, for honey discrimination. Fig. 4. PC-KMC model of the three types of honey Table 3. Discrimination accuracy of PC-NBC and PC- KMC models under wavelengths reduction rates (WRRs) between 1 and 2 6. Number of wavelengths WRRs PC-NBC (%) PC-KMC (%) 551 (full) PC-NBC and PC-KMC Based Models for Reduced Wavelengths Spectra The number of the used wavelengths of spectral characteristics was reduced successively 2, 4, 8, 16, x 32, 64 times (i.e., 2, where x = 1,2,..., 6 ). Selected wavelengths were at regular intervals, starting from 350 nm. Based on these reduced wavelengths, a series of PC-NBC and PC-KMC models were established. The results of model performance under different wavelength reduction rates (WRRs) are shown in Table 3. For example, the PC-NBC and PC-KMC models developed for 8 wavelengths (WRR = 64) achieved discrimination accuracy of 88.6 % and 77.1 %, respectively, which was almost as those for all the 551 wavelengths. This means that the eight wavelengths at equal intervals are enough to develop accurate models (PC-NBC and PC-KMC) for honey discrimination and thus facilitates the design of a simple and inexpensive sensor for honey discrimination in practice. Some authors in the literature have also reported the feasibility of reducing wavelengths by WRRs. Li and Yang [15] have developed PC-SVM (support vector machine) and PC-LDA (linear discriminate analysis) based prediction models using vis-nir wavelengths, subjected to WRRs of 2, 4, 8, 16, 32, 64, 128, 256 and 512. They have found that the models, calibrated for 6 wavelengths at interval of nm, have achieved accuracy almost the same as those for full 2847 wavelengths. However, they have tested their classifiers with the same samples, which they had used for training. CONCLUSIONS Based on Vis-NIR spectroscopy for discrimination of (botanical origins of) honey, two classification mod- 641

5 Journal of Chemical Technology and Metallurgy, 50, 5, 2015 els, PC-NBC and PC-KMC, were developed for various numbers of wavelengths ranging from 350 to 900 nm. As a result of using PCA, the number of input data were reduced from 551 wavelengths to only 2 PCs. The obtained advantages are the lack of correlation between the input data and also the ability to visualize the clusters formed by different types of honey. The proposed reduction of 6 the wavelength to 2 times, deteriorates very little the performance of the classifiers (accuracy 88.6 % and 77.1 % for PC-NBC and PC-KMC, respectively), which may allow developing a simpler and cheaper sensor for honey discrimination in practice. Future work will include: selection of specific wavelengths (in unequal intervals) and statistical analysis of the significance of the selected inputs for the separation of honey according to its botanical origin; adding new samples of other types of honey; using methods of artificial intelligence to increase the classifiers accuracy, etc. Acknowledgements The paper presents research and development, supported by the Scientific Fund for Internal Competition of the University of Food Technologies, Plovdiv, under Research Project No.7/14-H. REFERENCES 1. E. Anklam, A review of the analytical methods to determine the geographical and botanical origin of honey, Food Chem., 63, 1998, I. Hermosin, R.M. Chicon, M.D. Cabezudo, Free amino acid composition and botanical origin of honey, Food Chemistry, 83, 2003, S. Bogdanov, P. Martin, authenticity, Mitt. Geb. Lebensmittelunters Hyg. 93, 2002, S. Benedetti, S. Mannino, A.G. Sabatini, G.L. Marcazzan, Electronic nose and neural network use for the classification of honey, Apidologie, 35, 2004, 1-6, INRA/DIB-AGIB/ EDP Sciences, DOI: / apido: T. Gallardo-Velazquez, G. Osorio-Revilla, M.Z.D. Loa, Y. Rivera-Espinoza, Application of FTIR-HATR spectroscopy and multivariate analysis to the quantification of adulterants in Mexican honeys, Food Research International, 42, 3, 2009, Zhu, X., S.Li, Y. Shan et al., Detection of adulterants such as sweeteners materials in honey using nearinfrared spectroscopy and chemometrics, Journal of Food Engineering, 101, 1, H. Hotelling, Analysis of a Complex of Statistical Variables into Principal Components, Journal of Educational Psychology, 24, 6-7, 1933, , I.T. Jolliffe, Principal Component Analysis, 2nd ed., Springer Series in Statistics, New York, Statistics Toolbox User s Guide, R2014a, COPY- RIGHT by The MathWorks, Inc classifier#cite_note-rennie A. McCallum, K. Nigam, A comparison of event models for Naive Bayes text classification, AAAI-98 workshop on learning for text categorization 752, J. Rennie, L. Shih, J. Teevan, D. Karger, Tackling the poor assumptions of Naive Bayes classifiers, ICML, D. MacKay, An Example Inference Task: Clustering. Information Theory, Inference and Learning Algorithms, Cambridge University Press, 2003, Chapter 20, pp Independent_component_analysis_.28ICA Y. Li, H. Yang, Discrimination Using Visible and Near-Infrared Spectroscopy, International Scholarly Research Network, ISRN Spectroscopy, v. 2012, Article ID , 4 pages, doi: /2012/