Classification of Pure and Adulterated Food Samples Using ...deetoher/CASI05.pdf · of adulterants...
1
Classification of Pure and Adulterated Food Samples Using Near-Infrared Spectroscopic Data. Deirdre Toher *† , Gerard Downey * and Thomas Brendan Murphy † * Teagasc, The National Food Centre, Ashtown † Department of Statistics, Trinity College Dublin. Problem – Pure or Adulterated? The main aim of food authenticity studies is to detect when foods are not what they claim to be. Honey is defined by the E.U. as a natural product produced by the European honey bee, thus any chemical al- teration means that the product can no longer be claimed by producers to be honey. As it is a relatively expensive product to produce and naturally ex- tremely variable, honey is prone to adulteration for economic gain. Indeed cases of honey adulteration have been recorded since Roman times when concentrated grape juice was sometimes added. False claims may also be made in relation to the origin of the honey, but this study concentrates on attempting to classify samples into pure and adulterated. The adulteration was completed in the laboratory, using a number of adulterants – fructose:glucose mixtures, beet invert syrup and high fructose corn syrup – and various ratios and weight percentages of these adulterants. The spectra of these samples were then recorded over the wavelength range 1100 - 2498 nm as seen in Figure 1 below. The similarity of the pure and adulterated spectra is clearly evident, with the pure spectra almost completely masked by the adulterated spectra. 1100 1300 1500 1700 1900 2100 2300 2500 0.0 0.4 0.8 1.2 Wavelength (nm) Intensity Pure Adulterated FIGURE 1 Spectra of Pure and Adulterated Honey Dimensionality Reduction The data span 1100 to 2500 nm, with measurements taken every 2 nm. Adjacent absorption values are highly correlated, therefore dimension reduction is the first issue to be confronted. The technique chosen was that of wavelet analysis. Wavelet Analysis Wavelet analysis is used to decompose a spectrum into a series of wavelet coefficients. The coefficients can be used to reconstruct the original spectrum, so no information loss occurs. However, on exam- ining the coefficients produced by the wavelet analysis, it is evident that many are zero or close to zero. Removing such coefficents en- ables recomposition of the spectra using only fourteen coefficents. The recomposed spectra are then approximations of the original spectra [Figure 2]. Thresholding is used to select the coefficients that contain important information on the structure of the spectrum. Many thresholding tech- niques have been proposed and the choice of methods is a subjective one. The Daubechies’ wavelet is a consistently reliable type to use and is the default within wavethresh [1]. Efficient wavelet analysis methods require that the dimension of the data must be 2 m , where m is an integer. This forces 188 observations to be discarded. The central 2 9 = 512 observations were chosen – the range (1290 - 2312) nm. Methods of extending the tails of the spectra in order to have 2 10 = 1024 observations are unreliable, with resulting analysis proving problematic – often the associated variance structures are singular. 1290 1490 1690 1890 2090 2290 0.0 0.4 0.8 1.2 Wavelength (nm) Intensity Pure Adulterated FIGURE 2 Reconstructed Thresholded Spectra Wavelet Decomposition Coefficients Daub cmpct on ext. phase N=2 Translate Resolution Level 8 7 6 5 4 3 2 1 0 64 128 192 256 Wavelet Decomposition Coefficients Daub cmpct on ext. phase N=2 Translate Resolution Level 8 7 6 5 4 3 2 1 0 64 128 192 256 (a) 1290 1490 1690 1890 2090 2290 0.0 0.2 0.4 0.6 0.8 1.0 Wavelength (nm) Intensity (b) Actual Thresholded Thresholded FIGURE 3a Wavelets Decomposition – thresholded and actual/non-thresholded FIGURE 3b Actual and Reconstructed Thresholded Wavelets Figure 3a shows the structure of the wavelet analysis for one sample of pure honey, both in its full and thresholded form, while Figure 3b shows the reconstructed spectrum of the same sample after thresholding together with its actual spectrum. Classification Techniques The classification techniques used on this data set are based on Gaussian mixture models; each group is modelled using a Gaussian distribution. The covariance of each of the Gaussian models is struc- tured in a parsimonious manner using constraints. This approach offers the ability to model groups that have distinct volume, shape and ori- entation properties. Fraley and Raftery’s paper [2] describes a methodological approach to cluster analysis, with specific mention of model-based Discriminant Analysis. Their mclust [3] package was used to perform the model- based Discriminant Analysis. This allows for the possibility of the following models: TABLE 1: Parametrizations of the covariance matrix Σ k Model ID Decomposition Distribution EII Σ g = λI Spherical VII Σ g = λ g I Spherical EEI Σ g = λA Diagonal VEI Σ g = λ g A Diagonal EVI Σ g = λA g Diagonal VVI Σ g = λ g A g Diagonal EEE Σ g = λDAD T Ellipsoidal EEV Σ g = λD g AD T g Ellipsoidal VEV Σ g = λ g D g AD T g Ellipsoidal VVV Σ g = λ g D g A g D T g Ellipsoidal Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA) correspond to the EEE and VVV models respectively. The letters of the Model ID represent the volume, shape and orientation of the groups. EII VII EEI VEI EVI VVI EEE EEV VEV VVV FIGURE 4: General Shapes of Models Model Selection and Verification 400 simulations were performed, each taking a random split of the data to form a training set of 240 samples, with the remaining 238 samples used as a test set. Each training set comprised 79 pure and 161 adulterated honeys (28 using beet invert syrup, 20 using corn syrup and 113 using fructose:glucose mixtures). Test sets comprised 78 pure and 160 adulterated honeys (28 with beet invert syrup, 20 with corn syrup and 112 with fructose:glucose mixtures). Two types of classification were attempted – to classify into “pure” and “adulterated” and to classify into “pure”,“fructose:glucose”,“beet invert syrup” and “high fructose corn syrup”. Model selection techniques examined were (a) choosing the model with the best Bayesian Information Criterion value and (b) leave-one-out cross validation. Classification Results Selection method: BIC value Mean Values Predicted Adult. Pure Actual Adult. 153.93 6.07 Pure 8.44 69.56 Misclassification range Pure as adulterated: 2–22 Adulterated as pure: 0–14 Total misclassification: 4–22 Model Selected and Frequency: EEE/111; EEV/108; VEV/164; VVV/17 Mean Values Predicted BI FG CS Pure Actual BI 23.34 0.04 0.81 0.82 FG 0.02 105.99 0.00 6.00 CS 4.99 0.60 12.45 1.96 Pure 0.08 6.34 0.04 71.55 Misclassification range Pure as adulterated: 2–11 Adulterated as pure: 2–21 Total misclassification (pure/adulterated): 9–27 Incorrect type of adulteration: 2–13 Model Selected and Frequency: VEV/389; VVV/11 Selection method: Cross Validation Mean Values Predicted Adult. Pure Actual Adult. 154.07 5.95 Pure 8.13 69.88 Misclassification range Pure as adulterated: 2–22 Adulterated as pure: 0–12 Total misclassification: 7–22 Model Selected and Frequency: EEE/79; EEV/193; VEV/54; VVV/74 Mean Values Predicted BI FG CS Pure Actual BI 24.63 0.87 1.88 0.62 FG 0.00 108.08 0.07 3.85 CS 3.15 1.17 14.76 0.92 Pure 0.75 9.30 0.70 67.30 Misclassification range Pure as adulterated: 2–22 Adulterated as pure: 0–18 Total misclassification (pure/adulterated): 9–23 Incorrect type of adulteration: 2–15 Model Selected and Frequency: EEE/122; EEV/127; VEV/144; VVV/7 Conclusions and Further Work Wavelet analysis provides dramatic dimension reduction while main- taining the differences between pure and adulterated samples. The ability to identify each type of adulteration is promising. The devel- opment of a two stage technique of first classifying samples as pure or adulterated then classifying types of adulterations will be explored. Quantifying the level of adulteration also requires further exploration. These techniques are being developed towards a commerical and reg- ulatory standard, and so must be proven to perform not only on Irish honey but on samples from throughout the world. The potential for using updating algorithms will also be studied. Acknowledgements This work was funded by Teagasc under the Walsh Fellowship Scheme. References [1] G. Nason, A. Kovac (1997) and M. Maechler (1999), wavethresh: Software to perform wavelet statistics and transforms, R package version 2.2-8, (2004) [2] C. Fraley, A.E. Raftery, Model-based clustering, dis- criminant analysis, and density estimation, J. Amer. Stat. Assoc. 97, 611–631, (2002) available on- line as Technical Report no. 380, October 2000, http://www.stat.washington.edu/www/research/reports [3] C. Fraley, A.E. Raftery and R. Wehrens (R-port), mclust: Model-based cluster analysis
Transcript of Classification of Pure and Adulterated Food Samples Using ...deetoher/CASI05.pdf · of adulterants...
Classification of Pure and Adulterated Food Samples
Using Near-Infrared Spectroscopic Data.
Deirdre Toher∗†, Gerard Downey∗ and Thomas Brendan Murphy†
∗ Teagasc, The National Food Centre, Ashtown† Department of Statistics, Trinity College Dublin.
Problem – Pure or Adulterated?
The main aim of food authenticity studies is to detect when foods arenot what they claim to be. Honey is defined by the E.U. as a naturalproduct produced by the European honey bee, thus any chemical al-teration means that the product can no longer be claimed by producersto be honey.As it is a relatively expensive product to produce and naturally ex-tremely variable, honey is prone to adulteration for economic gain.Indeed cases of honey adulteration have been recorded since Romantimes when concentrated grape juice was sometimes added.False claims may also be made in relation to the origin of the honey,but this study concentrates on attempting to classify samples into pureand adulterated.The adulteration was completed in the laboratory, using a numberof adulterants – fructose:glucose mixtures, beet invert syrup and highfructose corn syrup – and various ratios and weight percentages of theseadulterants. The spectra of these samples were then recorded over thewavelength range 1100 − 2498 nm as seen in Figure 1 below. Thesimilarity of the pure and adulterated spectra is clearly evident, withthe pure spectra almost completely masked by the adulterated spectra.
1100 1300 1500 1700 1900 2100 2300 2500
0.0
0.4
0.8
1.2
Wavelength (nm)
Inte
nsity
Pure
Adulterated
FIGURE 1 Spectra of Pure and Adulterated Honey
Dimensionality Reduction
The data span 1100 to 2500 nm, with measurements taken every 2 nm.Adjacent absorption values are highly correlated, therefore dimensionreduction is the first issue to be confronted. The technique chosen wasthat of wavelet analysis.
Wavelet Analysis
Wavelet analysis is used to decompose a spectrum into a series ofwavelet coefficients. The coefficients can be used to reconstruct theoriginal spectrum, so no information loss occurs. However, on exam-ining the coefficients produced by the wavelet analysis, it is evidentthat many are zero or close to zero. Removing such coefficents en-ables recomposition of the spectra using only fourteen coefficents. Therecomposed spectra are then approximations of the original spectra[Figure 2].Thresholding is used to select the coefficients that contain importantinformation on the structure of the spectrum. Many thresholding tech-niques have been proposed and the choice of methods is a subjectiveone.The Daubechies’ wavelet is a consistently reliable type to use and isthe default within wavethresh [1]. Efficient wavelet analysis methodsrequire that the dimension of the data must be 2m, where m is aninteger. This forces 188 observations to be discarded. The central29 = 512 observations were chosen – the range (1290 − 2312) nm.Methods of extending the tails of the spectra in order to have210 = 1024 observations are unreliable, with resulting analysis provingproblematic – often the associated variance structures are singular.
1290 1490 1690 1890 2090 2290
0.0
0.4
0.8
1.2
Wavelength (nm)
Inte
nsity
Pure
Adulterated
FIGURE 2 Reconstructed Thresholded Spectra
Wavelet Decomposition Coefficients
Daub cmpct on ext. phase N=2Translate
Res
olut
ion
Leve
l
87
65
43
21
0 64 128 192 256
Wavelet Decomposition Coefficients
Daub cmpct on ext. phase N=2Translate
Res
olut
ion
Leve
l
87
65
43
21
0 64 128 192 256
(a)
1290 1490 1690 1890 2090 2290
0.0
0.2
0.4
0.6
0.8
1.0
Wavelength (nm)
Inte
nsity
(b)Actual
ThresholdedThresholded
FIGURE 3a Wavelets Decomposition– thresholded and actual/non-thresholded
FIGURE 3b Actual and Reconstructed Thresholded Wavelets
Figure 3a shows the structure of the wavelet analysis for one sampleof pure honey, both in its full and thresholded form, while Figure 3bshows the reconstructed spectrum of the same sample after thresholdingtogether with its actual spectrum.
Classification Techniques
The classification techniques used on this data set are based onGaussian mixture models; each group is modelled using a Gaussiandistribution. The covariance of each of the Gaussian models is struc-tured in a parsimonious manner using constraints. This approach offersthe ability to model groups that have distinct volume, shape and ori-entation properties.Fraley and Raftery’s paper [2] describes a methodological approachto cluster analysis, with specific mention of model-based DiscriminantAnalysis. Their mclust [3] package was used to perform the model-based Discriminant Analysis.This allows for the possibility of the following models:
TABLE 1: Parametrizations of the covariance matrix Σk
Model ID Decomposition DistributionEII Σg = λI SphericalVII Σg = λgI SphericalEEI Σg = λA DiagonalVEI Σg = λgA DiagonalEVI Σg = λAg DiagonalVVI Σg = λgAg Diagonal
EEE Σg = λDADT Ellipsoidal
EEV Σg = λDgADTg Ellipsoidal
VEV Σg = λgDgADTg Ellipsoidal
VVV Σg = λgDgAgDTg Ellipsoidal
Linear Discriminant Analysis (LDA) and Quadratic DiscriminantAnalysis (QDA) correspond to the EEE and VVV models respectively.The letters of the Model ID represent the volume, shape and orientationof the groups.
EII VII EEI VEI EVI
VVI EEE EEV VEV VVV
FIGURE 4: General Shapes of Models
Model Selection and Verification
400 simulations were performed, each taking a random split of thedata to form a training set of 240 samples, with the remaining 238samples used as a test set. Each training set comprised 79 pure and161 adulterated honeys (28 using beet invert syrup, 20 using corn syrupand 113 using fructose:glucose mixtures). Test sets comprised 78 pureand 160 adulterated honeys (28 with beet invert syrup, 20 with cornsyrup and 112 with fructose:glucose mixtures).Two types of classification were attempted – to classify into “pure”and “adulterated” and to classify into “pure”,“fructose:glucose”,“beetinvert syrup” and “high fructose corn syrup”.Model selection techniques examined were (a) choosing the model withthe best Bayesian Information Criterion value and (b) leave-one-outcross validation.
Classification Results
Selection method: BIC value
Mean Values PredictedAdult. Pure
Actual Adult. 153.93 6.07Pure 8.44 69.56
Misclassification rangePure as adulterated: 2–22 Adulterated as pure: 0–14Total misclassification: 4–22Model Selected and Frequency:EEE/111; EEV/108; VEV/164; VVV/17
Mean Values PredictedBI FG CS Pure
Actual BI 23.34 0.04 0.81 0.82FG 0.02 105.99 0.00 6.00CS 4.99 0.60 12.45 1.96
Pure 0.08 6.34 0.04 71.55
Misclassification rangePure as adulterated: 2–11 Adulterated as pure: 2–21Total misclassification (pure/adulterated): 9–27Incorrect type of adulteration: 2–13Model Selected and Frequency:VEV/389; VVV/11
Selection method: Cross Validation
Mean Values PredictedAdult. Pure
Actual Adult. 154.07 5.95Pure 8.13 69.88
Misclassification rangePure as adulterated: 2–22 Adulterated as pure: 0–12Total misclassification: 7–22Model Selected and Frequency:EEE/79; EEV/193; VEV/54; VVV/74
Mean Values PredictedBI FG CS Pure
Actual BI 24.63 0.87 1.88 0.62FG 0.00 108.08 0.07 3.85CS 3.15 1.17 14.76 0.92
Pure 0.75 9.30 0.70 67.30
Misclassification rangePure as adulterated: 2–22 Adulterated as pure: 0–18Total misclassification (pure/adulterated): 9–23Incorrect type of adulteration: 2–15Model Selected and Frequency:EEE/122; EEV/127; VEV/144; VVV/7
Conclusions and Further Work
Wavelet analysis provides dramatic dimension reduction while main-taining the differences between pure and adulterated samples. Theability to identify each type of adulteration is promising. The devel-opment of a two stage technique of first classifying samples as pure oradulterated then classifying types of adulterations will be explored.Quantifying the level of adulteration also requires further exploration.These techniques are being developed towards a commerical and reg-ulatory standard, and so must be proven to perform not only on Irishhoney but on samples from throughout the world.The potential for using updating algorithms will also be studied.
Acknowledgements
This work was funded by Teagasc under the Walsh Fellowship Scheme.
References
[1] G. Nason, A. Kovac (1997) and M. Maechler (1999),wavethresh: Software to perform wavelet statistics andtransforms, R package version 2.2-8, (2004)
[2] C. Fraley, A.E. Raftery, Model-based clustering, dis-criminant analysis, and density estimation, J. Amer.Stat. Assoc. 97, 611–631, (2002) available on-line as Technical Report no. 380, October 2000,http://www.stat.washington.edu/www/research/reports
[3] C. Fraley, A.E. Raftery and R. Wehrens (R-port), mclust:Model-based cluster analysis
