authenticity detection of black rice by near-infrared...

Research ArticleAuthenticity Detection of Black Rice by Near-InfraredSpectroscopy and Support Vector Data Description

Hui Chen12 Chao Tan 1 and Zan Lin13

1Key Lab of Process Analysis and Control of Sichuan Universities Yibin University Yibin Sichuan 644000 China2Hospital Yibin University Yibin Sichuan 644000 China3The First Affiliated Hospital Chongqing Medical University Chongqing 400016 China

Correspondence should be addressed to Chao Tan chaotan1112163com

Received 26 March 2018 Revised 3 June 2018 Accepted 11 June 2018 Published 9 July 2018

Academic Editor Richard G Brereton

Copyright copy 2018 Hui Chen et alThis is an open access article distributed under the Creative CommonsAttribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Black rice is an important rice species in Southeast Asia It is a common phenomenon to pass low-priced black rice off as high-priced ones for economic benefit especially in some remote townsThere is increasing need for the development of fast easy-to-useand low-cost analytical methods for authenticity detection The feasibility to utilize near-infrared (NIR) spectroscopy and supportvector data description (SVDD) for such a goal is explored Principal component analysis (PCA) is used for exploratory analysis andfeature extraction Another two data descriptionmethods ie k-nearest neighbor data description (KNNDD) andGAUSSmethodare used as the reference A total of 142 samples from three brands were collected for spectral analysis Each time the samples of abrand serve as the target class whereas other samples serve as the outlier class Based on both the first two principal components(PCs) and original variables three types of data descriptions were constructed On average the optimized SVDD model achievesacceptable performance ie a specificity of 100 and a sensitivity of 942 on the independent test set with tight boundary Itindicates that SVDD combined with NIR is feasible and effective for authenticity detection of black rice

1 Introduction

Black rice is an economically important special rice speciesand has been consumed for a long time in Southeast Asiaincluding China [1ndash3] Many researches have showed thatblack rice has considerably strong free-radical scavenging andantioxidation effects as well as other biological effects of itsextracts such as antimutagenic and anticarcinogenic [4 5]Black rice quality in terms of nutrition is also valuable forits protein content and the balance of essential amino acidsIn fact black rice is also a mixture of various carbohydratesThere exist varying amounts of nutrient in different kinds ofblack rice because of genetic and environmental factors Inmarket there exist many brands of black rice The qualityand price of them vary greatly and renowned brands havehigher price However illegal tradesman often passes low-priced black rice off as high-priced ones for economic benefitespecially in some remote towns

How to discriminate different types of black rice isinteresting Up to now it is mainly dependent on human

senses More objective and novel methods are maybe basedon complex instruments such as high performance liquidchromatography or mass spectroscopy (MS) [6] In recentyears molecular spectroscopy has drawn more attention andproved to be a powerful tool for authenticity detection [7ndash9] In particular near-infrared (NIR) spectroscopy becomesthe most widely used technique in various fields includingcigarettes [10] food [11] textile [12] medicine [13] anddrug [14] It is capable of rapidly obtaining a vectormatrixsignal of a complex sample and therefore provides the chanceof executing a in-depth qualitative or quantitative analysisDetection of food authenticity is a important task in foodanalysis and aims to answer the question on which class aparticular sample belongs to by its spectral signal Oftenit can be realized by comparing spectra of a specimen tobe identified with spectra of ldquoknownrdquo or ldquostandardrdquo Asfor NIR spectroscopy however spectral signals for complexfood systems are characterized by peak overlapping andpoor resolution So an appropriate chemometric model isindispensable for a NIR-based application

HindawiInternational Journal of Analytical ChemistryVolume 2018 Article ID 8032831 8 pageshttpsdoiorg10115520188032831

2 International Journal of Analytical Chemistry

For the perspective of modeling chemometrics involvingqualitative tasks can be divided into two categories clas-sification and one-class classification ie data description[15] Classification is vey often considered as a synonym ofdiscriminant analysismethods since they assign a new sampleto one of a set of predefined classes The correspondingclassifier is trained on a training set Data description differsin one essential aspect from the conventional classificationsince it is assumed that only information on a single classis available Data description problems are common in thereal world where positive objects are widely available butnegative ones are maybe hard expensive or even impossibleto gather [16] In the literature three main approaches can bedistinguished the density estimation the boundarymethodsand the reconstructionmethods [17] General demand of anyauthentication problem is that a genuine class ie a targetclass must be known [18 19]The target class is always uniquefor a specific authentication problem Any other objects orclasses of objects that are not members of the target classare considered as outliers This also means that just samplesof the target class can be utilized and that no informationon the other classes is present For data description theboundary surrounding the target class has to be estimatedfrom available data such that it accepts as much of the targetsamples as possible and minimizes the error of acceptingoutlier Up to now much effort has been expended to developclassification algorithms and the concept of data descriptionis also of interest and noticeable [20ndash23] especially in thecases where it is impossible to meaningfully define all ofthe classes and obtain fully representative samples In foodauthenticity the interest is focused on a single target class soas to verify compliance of samples with the features of thatclass and a data description approach should be adopted tobuild an enclosed boundary around the target class

The present work focuses on exploring the feasibilityto utilize near-infrared (NIR) spectroscopy and supportvector data description (SVDD) for authenticity diction ofblack rice Principal component analysis (PCA) is used forexploratory analysis and feature extraction Another two datadescriptionmethods ie k-nearest neighbor data description(KNNDD) and GAUSS method are used as the referenceA total of 142 samples from three brands were collected forspectral analysis All spectra were preprocessed beforehandby standard normal transformation (SNV) Each time thesamples of a brand serve as the target class whereas othersamples serve as the outlier class Based on both the firsttwo principal components (PCs) and original variables threetypes of data descriptions were constructed On average theoptimized SVDD model achieves acceptable performanceie a specificity of 100 and a sensitivity of 942 onthe independent test set with tight boundary The effect oftraining set size and the parameter of kernel width have alsobeen discussed It indicates that SVDD combined with NIR isfeasible and effective for authenticity detection of black rice

2 Theory and Methods

Many methods have been developed to solve the one-classor data description problem and they can be divided into

three main categories density boundary and reconstruc-tion methods Here three algorithms ie support vectordata description Nearest Neighbor Method and GaussianMethod are introduced and used for experiments amongwhich the first two are boundary methods and the last onebelongs to density method

21 Support Vector Data Description (SVDD) SVDD is anovel algorithm for one-class classification problems whichhas been proposed by Tax [15] inspired by the idea of thesupport vector machines It focuses on finding a minimumhypersphere around the target class The hypersphere can beused to decide whether new objects are targets or outliersSuch a sphere is characterized only by center a and radius 119877When seeking sphere it needs to minimize the volume of thesphere by minimizing 1198772 and demand that the sphere coversas many training samples as possible Given the training setx119894 119894 = 1 2 119873 the task in SVDD is to minimize errorfunction

min 119871 (119877 a 120577119894) = 1198772 + 119862sum119894

120577119894 (1)

st 1003817100381710038171003817x119894 minus a1003817100381710038171003817 le 1198772 + 120577119894 120577119894 ge 0 foralli (2)

where a and 119877 are the center and the radius of the hyper-sphere respectively 119862 is the penalty factor which regulatesthe hyperspherical volume and error ie the number oftarget objects rejected 120577119894 is a slack variable for allowableerror limitation Almost all objects are within the sphereThisoptimization problem can be solved by Lagrange multipliermethod [24]

Because the target class is not spherically distributed inmost cases some traditional decision rules may not workwell To make a more effective and flexible decision theoriginal data can be implicitly transformed to a higher dimen-sion by the so-called kernel function119870(x119894 x119895) Several kernelfunctions including linear polynomial Gaussian radial basisfunction (RBF) are available [25 26] In this work the RBFkernel the most commonly used kernel in machine learningwas used The form of RBF kernel is

119870(x119894 x119895) = exp(minus10038171003817100381710038171003817x119894 minus x1198951003817100381710038171003817100381721205902 ) (3)

where 120590 is a key parameter for controlling the boundarytightness

22 Nearest Neighbor Method Themost straightforward andsimplest method to obtain a one-class model is to estimatethe density of the training set Unfortunately it often requiresa large number of samples to avoid the curse of dimension-ality Instead of estimating whole probability densities anindication of the resemblance can also be acquired by com-paring distances Nearest neighbor method can be derivedfrom a local density estimation [27] It avoids the explicitdensity estimation by only using distances to the first nearestneighbor In the process of density estimation a cell often anhypersphere in d-dimension space is centered around the test

International Journal of Analytical Chemistry 3

object z The cell volume is grown until it contains 119896 objectsfrom the training set The local density can be estimated by

119901119873119873 (z) = 119896119873119881119896 1003817100381710038171003817z minus 119873119873119905119903119896 (z)1003817100381710038171003817 (4)

where119873119873119905119903119896 (z) and 119881119896 are the 119896 nearest neighbors of z in thetraining set and the volume of the cell containing this objectLater we will use KNNDD to denote this method

For an unknown test object z the distance from it to itsnearest neighbor in the training set NN119905119903(z) is compared withthe distance from NN119905119903(z) to its nearest neighbor The testobject z can be accepted when its local density is larger orequal to the density of the nearest neighbor It seems to bevery useful for distributions characterized by fast decayingprobabilities Obviously themethod can easily be generalizedto a larger number of neighbors k That is instead of takingthe first nearest neighbor into account the 119896th neighborshould be considered

23 Gaussian Method When a proper probability model isassumed and the sample size is sufficient density method isadvantageous for one-class problem With the optimizationof the threshold a minimum volume can be automaticallyfound for the given probability density When only a littleamount of samples is available the simplest model is theunimodal GaussianNormal distribution It fits a probabilitydensity model as follows

119901119873 (x) 1(2120587)1198892 |Σ|05 exp minus12 (x minus 120583)119879 Σminus1 (x minus 120583) (5)

where 120583 is the mean and Σ is the covariance matrix Bothshould be estimated from the training set For 119889 dimensionaldata the number of the parameters is

119889 + 12119889 (119889 minus 1) (6)

The method imposes a strict unimodal and convex densitymodel on the data The main computational effort is maybethe inversion of the covariance matrix In case of badlyscaled data or data with singular directions it is difficult tocalculate the inverse of Σ and it can be approximated by thepseudoinverse Σ+ = Σ119879(ΣΣ119879)minus1or by introducing regulari-zation (adding a small constant 120582 to the diagonal ie Σ1015840 =Σ + 120582I) In the last case the user needs to supply a parameter120582 This is also the only magic parameter that requires a userto provide

Finally a threshold on the probability density needs tobe set for distinguishing between target and outlier dataAccepting 95 of the objects requires a threshold on theMahanalobis distance

(x minus 120583)119879 Σminus1 (x minus 120583) (7)

of

120579119873 = (1205942119889)minus1 (095) (8)

where (1205942119889)minus1is the inverse 1205942119889 with 119889 degrees of freedomThis method is expected to work effectively only if the datais unimodal and convex To obtain a more flexible densitymethod it can be extended to a mixture of Gaussians Laterwe will use GAUSS to denote this method

3 Experimental

31 Sample Preparation A total of 142 samplesbag of blackrice of three brands were purchased from local supermarketsin China They were from different supplier and let us markthem as A B and C brands These samples were collectedfrom three batches of A two batches of B and three batchesof C but different packages For A or C forty-eight bags ofricewere sampled sixteen bags for each batch For B forty-sixbags of rice were sampled twenty-three bags for each batchIn total the number of samples belonging to A B and C are48 46 and 48 respectively The time it takes to collect thesample is about six months The samples of each brand couldserve as the target class whereas other samples acted as theoutlier class All samples were stored in the laboratory kept at25∘C for more than 7 days in order to achieve a temperaturebalance To reduce the effect of environment the NIR spectraof all samples were recorded on the same day

32 Spectral Measurement and Preprocessing Spectra of dif-ferent samples collected on an Antaris II FT-NIR spectrom-eter (Thermo Scientific CoUSA) were equipped with anintegrating sphere module a rotating sample cup and aInGaAs detector aswell as a tungsten lamp as the light sourceThe sample was poured into a standard sample cup with a50 mm diameter and the height was controlled on about 30mm for preventing light leak An internal gold reference wasused for automatic background collection A specific samplecup spinner accessory for the integrating sphere samplingmodule that allows multipoint reflection measurements ofheterogeneous solids such as powders granules and pelletswas used for obtaining NIR spectra of high quality Inthis way the final spectrum is the average of the spectracollected at different locations which can reduce the effectof heterogeneity of solids to some extent

The NIR spectrum was measured in the region of10000ndash4000 cmminus1 with 32 scans at a resolution of 3856 cmminus1Each spectrum contains 1557 data points The experimentaltemperature and the related humidity were controlled around25∘C and 60 respectively Preprocessing of spectra is oftenof great importance if reasonable results need to be obtainedwhether it is concerned with qualitative or quantitative tasksSeveral methods of preprocessing were attempted In com-parison with other preprocessing methods standard normaltransformation (SNV) achieved a satisfactory performancewithout the need of a reference spectrum and user decisionfor the computation So all spectra were preprocessed bySNVThe spectral measurement was controlled by the Resultsoftware [28] DD toolbox was used for one-class classifier


5000 6000 7000 8000 9000 100004000Wavenumber (＝Ｇminus1

)

05

1

15Lo

g(1R)

(a)


)

minus1

0

1

Log(1R)

(b)

Figure 1 Original near-infrared (NIR) spectra (a) and all the preprocessed spectra (b) by standard normal transformation (SNV)

modeling [15] All calculation was made on MATLAB 2015bfor Windows

4 Results and Discussions

41 NIR Spectral Analysis Figure 1 shows the NIR spectraand all the preprocessed spectra of black rice samples bySNV Seen from Figure 1 the spectra of three types of blackrice share very similar absorbance patterns in the range of4000-11000 cmminus1 They can hardly be distinguished just bynaked eyes General features of a NIR spectrum of solidsamples include a multiplicative response to changes inparticle size SNV treatment autoscales each spectrum basedon calculating the mean and standard deviation between thedensities It is also clear in Figure 1 that by preprocessingsome additive and multiplicative effects have been removed

It is well known that major components of black rice arecomplex molecules from the polymerization of monomerssuch as amino acids or carbohydrates Each monomerexhibits specific chemical groups such as carboxylic andamine functions in amino acids As each chemical groupmay absorb the infrared region light it appears useful toclearly identify the characteristic NIR bands of these groupsBecause NIR spectrum corresponds to molecular responsesof the overtone and combination bands for each funda-mental absorption band there exists several overtones withdecreasing intensity corresponding to the increasingmultipleor transition number All the bands can form a myriad ofcombination bands with intensities increasing as frequencydecreases NIR band intensities are much weaker than theircorresponding mid-infrared fundamentals by a factor of 10-100 In Figure 1 two strong bands at 5175 cmminus1 and 6930cmminus1 result from the absorbance of water among which thepeaks around 5175 cmminus1 are the combination of asymmetricstretching and bending vibration of H2O The band of 8200-8600 cmminus1 can be attributed to the second overtones of C-Hstretching in various groups The wider bands in 6100ndash7000cmminus1 are mainly caused by the overlapping of the firstovertones of O-H and N-H stretching The two peaks at4266cmminus1 and 4335cmminus1 which can be attributed to C-Hstretching and C-H deformation are very stable and carrymuch useful information However accurate assignmentsof each peak were maybe difficult due to low resolutionand baseline shift therefore it is necessary to resort tochemometricmethods to extract the useful information fromspectra for identification purposes

Furthermore one of the most interesting applications ofNIR technique in the food analysis is total quality evalua-tion as it can provide fingerprint information of a sampleDifferent brands of black rice mean different balancesratiosof diverse chemical constituents and physicochemical prop-erties rather than simple amount of each constituent NIRspectra contain rich information on chemical constituentsand physicochemical properties Although the quality ofblack rice is generally assessed by sensory evaluation its tasteis actually a function of chemical constituents such as proteinmoisture amylose fatty acid and minerals Therefore anoverall evaluation is preferred based on NIR spectroscopy

Principal component analysis (PCA) the most wide-spread multivariate tool was used for an exploratory analysisand dimensional reduction Unlike other applications themain goal of the present work using PCA was to map theoriginal data into its principal component score space (iethe first two) based on which the subsequent modelingwas carried out So all samples were considered as a wholefor PCA and mean-centering pretreatment By computationthe first two PCs explain 794 and 184 of the totalvariances respectively and they may contain most of theuseful information in the original spectra Because of thiswe decided to use the first two components as the input ofsubsequent data description methods

42 Authenticity Detection by Data Description Given adataset in general the selection of a representative trainingset upon which training the predictionmodel is performed isvery important For this purpose in our work the Kennardand Stone (KS) algorithm [29] was first used to rank allsamples of each class in the dataset under considerationthereby producing three sequences (A B and C) The KSalgorithm consists of two main steps taking the pair ofsamples between which the Euclidean distance of x-vectors(predictor) is largest and then sequentially selecting a sampleto maximize the Euclidean distances between x-vectors ofalready selected samples and the remaining samples Thisprocess is repeated until all samples are picked out Theformer samples are more representative than the latter oneWhen A class served as the target class only the first thirtysamples in A sequence were used as the training set forconstructing data description The remaining samples in Asequence and all samples in B and C sequences were usedas the test set (the same partition of the sample set for thecases using B or C as the target class) Based on the first


KnnddGaussSvdd

minus14 minus1 minus08 minus06 minus04 minus02 0 02minus12PC1

01

02

03

04

05

06

07

08

09

PC2

Figure 2 Data description boundary of class A on the first two-principal-component space based on the training set

knnddSvddGauss

minus1 minus05 050 1 15 2minus15PC1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

PC2

Figure 3 Application of the data description models of class A onthe test set

two PCs of the training set three types of data descriptionsmentioned above ie SVDD KNNDD and GAUSS wereconstructed SVDD used the Gaussian kernel Figure 2 givesthe optimized data description boundary of class A basedon the training set It seems that the boundary of SVDDis tightest All the descriptions differ from conventionalclassification because they always obtain a closed boundaryaround one of the target classes Unlike densitymethods suchas GAUSS SVDD does not require a strict representativesampling of the target class a sampling containing extremeobjects is also acceptable This can be found explicitly in theerror definition of SVDD whichminimizes the volume of thedescription plus the sum of slack variables for objects outside

knnddSvddGauss

minus15 minus1 minus05 0 05 1 15 2minus2PC1

minus08

minus06

minus04

minus02

0

02

04

06

08

PC2

Figure 4 Application of the data description models of class B onthe test set

the description A conventional classifier on the contrarydistinguishes between twomultiple classes without focusingon any of the classes and aims to minimize the probability ofoverall error It is expected to perform very poorly when justthe target class is available or the dataset is relatively smallFood validation or authenticity detection is often the case

Figure 3 shows the application of the data descriptionmodels of class A on the test set Only one target samplewas identified as outlier by SVDD Even if the KNNDDand GAUSS correctly identified all samples the false pos-itive would increase when more test samples were used inthe future Similarly Classes B and C were considered asthe target class and three corresponding data descriptionswere constructed Figure 4 shows the application of thedata description models of class B on the test sets Nowall the models correctly identified the target samples andthe corresponding outliers but the SVDD use the tightestboundary maybe implying better generalization ability BothKNNDD and GAUSS produce looser borders It should benoted that each time the so-called ldquofakerdquo black rice is actuallysimulated by the samples from nontarget class

The character of the SVDD heavily depends on the widthparameter of the Gaussian kernel which is very crucial asit can provide different prediction performance and leads tooverfitting problem Several previous studies have reportedhow to optimize SVDD [30] The penalty term is samplerejection rate ie the approximate proportion of samplesmisclassified in a training set The other tunable parameteris kernel width A large width can lead to a less complicatedboundary and a relatively large width (compared to themaximum distance between samples in training set) couldlead to a rigid hypersphere In this work based on the averagenearest neighbor distance in the dataset one can distinguishthree types of cases very small very large and intermediatevalues By changing the value the description ranges from


minus1 0 1 2minus2PC1

minus1

minus05

0

05

1PC

2

(a) minus045


minus1

minus05

0

05

1

PC2

(b) minus065

minus1

minus05

0

05

1

PC2


(c) minus085


minus1

minus05

0

05

1

PC2

(d) minus15

Figure 5 The influence of the kernel parameter on the boundary of support vector data description (SVDD) using class A as the target class(classification error on the target class is set as 01)

Parzen density estimation via a mixture of Gaussian to therigid hypersphere can be observed in Figure 5 which showsthe influence of the kernel parameter on the boundary ofSVDD when using class A as the target class The boundaryof SVDD seems to be sensitive to the kernel parameter Withthe increase of the width the boundary undergoes a complexchange it gradually achieves the optimum and then getsworse For different cases the number of support vectors isalso different In order to facilitate the comparison Figure 6gives a similar ensemble plot of the influence of the kernelparameter on the boundary Based on the shape and thecompactness of the edge of the description the optimal widthparameter is 085 for this case A Such a boundary containsall the target samples among which six samples are just onthe edge and the shape is also simple Also one importantadvantage of SVDD over some traditional methods is thatthe classifier does not require that the data follow a normaldistribution However there exist some alternative proce-dures for optimizing kernel width such as cross-validationbootstrapping and the consistency evaluation of the classifierusing only the error of the nontarget class [15 30 31]

Taking the first case as an example (A as the targetclass) instead of the PCs the original spectral variableswere used as the independent variables for constructing datadescriptions On the independent test set all these modelsincluding SVDDKNNDD andGAUSS achieved a specificityof 100 (the ration of outliers that were rejected) whilethe corresponding sensitivity ie the ratio of the targetclass that was accepted is 100 for GUASS and 944 forboth SVDD and KNNDD despite whether PCs or originalvariables are used It indicates that using PCs or originalvariables does not make substantial difference Howeverusing all features is likely to result in overfitting while usingPCs will likely reduce overfitting Also using PCs makes thecomputation to be faster and to bemore convenient for visualpurposes When B or C is the target class the correspondingspecificity and sensitivity have also been summarized inTable 1 On average the SVDD achieves best prediction withthe specificity of 100 and the sensitivity of 942

On the whole the data description especially SVDDachieved an acceptable sensitivity and specificity for the so-called small-sample problem Such a procedure is maybe


Table 1 Summary of the performance of different models

Target class GAUSS KNNDD SVDDSPE SEN SPE SEN SPE SEN

A 100 100 100 944 100 944B 968 875 968 938 100 938C 978 889 989 889 100 944Average 982 921 985 923 100 942Note SPE and SEN denote the specificity and sensitivity respectively

minus02minus04minus08 minus06 0 02minus12 minus1minus14PC1

01

02

03

04

05

06

07

08

09

PC2

0508085

092

Figure 6 Ensemble of the influence of the kernel parameter on theboundary of support vector data description (SVDD) on the sameplot

potential tool for authenticity detection of various foodsincluding black rice

5 Conclusions

The work reveals that NIR spectroscopy combined withsupport vector data description is feasible and advantageousto implement authenticity detection of black rice It canserve as an alternative to laborious time-consuming wetchemical methods and sensory analysis of human Howeverfor obtaining more reliable results more samples need to becollected which remains our next work

Data Availability

The spectra data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (21375118 J1310041) Scientific ResearchFoundation of Sichuan Provincial Education Department ofChina (17TD0048) Scientific Research Foundation of YibinUniversity (2017ZD05) the Applied Basic Research Programsof Science and Technology Department of Sichuan Provinceof China (2018JY0504) and Opening Fund of Key Lab ofProcess Analysis and Control of Sichuan Universities ofChina (2015006 2016002)

References

[1] K Tananuwong and W Tewaruth ldquoExtraction and applicationof antioxidants from black glutinous ricerdquo LWT- Food Scienceand Technology vol 43 no 3 pp 476ndash481 2010

[2] C Hu J Zawistowski W Ling and D D Kitts ldquoBlack rice(Oryza sativa L indica) pigmented fraction suppresses bothreactive oxygen species and nitric oxide in chemical and bio-logical model systemsrdquo Journal of Agricultural and Food Chem-istry vol 51 no 18 pp 5271ndash5277 2003

[3] C J Liu H M Wang C L Liu L Wang and X M Mengldquouncertainty evaluation Of toatal phenol determination onblack rice by spectrophotometryrdquo China Food Additives vol 2pp 172ndash175 2017

[4] J W Hyun and H S Chung ldquoCyanidin and malvidin fromOryza sativa cv heugjinjubyeo mediate cytotoxicity againsthumanmonocytic leukemia cells by arrest of G(2)M phase andinduction of apoptosisrdquo Journal of Agricultural and Food Chem-istry vol 52 no 8 pp 2213ndash2217 2004

[5] Z F LiangW JWang X YWang andY Fang ldquoDeterminationof trace elements content of the same ogill of black rice andordinary ricerdquo Chemical Engineering vol 11 pp 27ndash31 2015

[6] B Zhang Z Q Rong Y Shi J GWu andCH Shi ldquoPredictionof the amino acid composition in brown rice using differentsample status by near-infrared reflectance spectroscopyrdquo FoodChemistry vol 127 no 1 pp 275ndash281 2011

[7] D Cozzolino ldquoNear Infrared Spectroscopy and Food Authen-ticityrdquo Advances in Food Traceability Techniques and Technolo-gies Improving QualityThroughout the Food Chain pp 119ndash1362016

[8] E Domingo A A Tirelli C A Nunes M C Guerreiro and SM Pinto ldquoMelamine detection in milk using vibrational spec-troscopy and chemometrics analysis a reviewrdquo Food ResearchInternational vol 60 pp 131ndash139 2014


[9] H Chen C Tan Z Lin and T Wu ldquoDetection of melamineadulteration in milk by near-infrared spectroscopy and one-class partial least squaresrdquo Spectrochimica Acta Part A Molec-ular and Biomolecular Spectroscopy vol 173 pp 832ndash836 2017

[10] C TanM Li andXQin ldquoStudy of the feasibility of distinguish-ing cigarettes of different brands using an Adaboost algorithmand near-infrared spectroscopyrdquo Analytical and BioanalyticalChemistry vol 389 no 2 pp 667ndash674 2007

[11] J Zhao H Lin Q Chen X Huang Z Sun and F ZhouldquoIdentification of eggrsquos freshness using NIR and support vectordata descriptionrdquo Journal of Food Engineering vol 98 no 4 pp408ndash414 2010

[12] C Ruckebusch F Orhan A Durand T Boubellouta and JP Huvenne ldquoQuantitative analysis of cotton-polyester textileblends from near-infrared spectrardquo Applied Spectroscopy vol60 no 5 pp 539ndash544 2006

[13] H Chen Z Lin HWu LWang TWu and C Tan ldquoDiagnosisof colorectal cancer by near-infrared optical fiber spectroscopyand random forestrdquo Spectrochimica Acta Part A Molecular andBiomolecular Spectroscopy vol 135 pp 185ndash191 2015

[14] K Degardin A Guillemain N V Guerreiro and Y RoggoldquoNear infrared spectroscopy for counterfeit detection using alarge database of pharmaceutical tabletsrdquo Journal of Pharmaceu-tical and Biomedical Analysis vol 128 pp 89ndash97 2016

[15] D M Tax One-class classification Delft University of Technol-ogy Delft The Netherlands 2001

[16] B Krawczyk andMWozniak ldquoDiversitymeasures for one-classclassifier ensemblesrdquoNeurocomputing vol 126 pp 36ndash44 2014

[17] O Mazhelis ldquoOne-Class Classifiers A Review and Analysis ofSuitability in the Context of Mobile-Masquerader DetectionrdquoAdvances in end-user data-mining techniques vol 30 pp 39ndash472006

[18] POliveri ldquoClass-modelling in food analytical chemistryDevel-opment sampling optimisation and validation issues - Atutorialrdquo Analytica Chimica Acta vol 982 pp 9ndash19 2017

[19] P Oliveri and G Downey ldquoMultivariate class modeling forthe verification of food-authenticity claimsrdquo TrAC - Trends inAnalytical Chemistry vol 35 pp 74ndash86 2012

[20] M Forina C Armanino R Leardi and G Drava ldquoA class-modelling technique based on potential functionsrdquo Journal ofChemometrics vol 5 no 5 pp 435ndash453 1991

[21] O Y Rodionova P Oliveri and A L Pomerantsev ldquoRigorousand compliant approaches to one-class classificationrdquo Chemo-metrics and Intelligent Laboratory Systems vol 159 pp 89ndash962016

[22] L Xu S Yan C Cai and X Yu ldquoOne-class partial least squares(OCPLS) classifierrdquo Chemometrics and Intelligent LaboratorySystems vol 126 pp 1ndash5 2013

[23] R G Brereton ldquoOne-class classifiersrdquo Journal of Chemometricsvol 25 no 5 pp 225ndash246 2011

[24] Z S Pan B Chen ZMMiao andG Q Ni ldquoOverview of studyon one-class classifierrdquoActa Electronica Sinica vol 37 pp 2496ndash2503 2009

[25] F S Uslu H Binol M Ilarslan and A Bal ldquoImproving SVDDclassification performance on hyperspectral images via correla-tion based ensemble techniquerdquo Optics and Lasers in Engineer-ing vol 89 pp 169ndash177 2016

[26] L Duan M Xie T Bai and J Wang ldquoA new support vectordata description method for machinery fault diagnosis withunbalanced datasetsrdquo Expert Systems with Applications vol 64pp 239ndash246 2016

[27] D Tax and R Duin ldquoData description in subspacesrdquo in Pro-ceedings of the 15th International Conference on Pattern Recog-nition pp 672ndash675 Barcelona Spain 2002

[28] Thermo Scientific Result Integration Software user Guide[29] R W Kennard and L A Stone ldquoComputer aided design of

experimentsrdquo Technometrics vol 11 no 1 pp 137ndash148 1969[30] S Kittiwachana D L S Ferreira G R Lloyd et al ldquoOne class

classifiers for process monitoring illustrated by the applicationto online HPLC of a continuous processrdquo Journal of Chemomet-rics vol 24 no 3-4 pp 96ndash110 2010

[31] L S Zhong and C R Hou ldquoFault monitoring of industrialprocess based on independent component and support vectordescription (IC-SVDDrdquo Computers and Applied Chemistry vol34 pp 285ndash290 2017

TribologyAdvances in

Hindawiwwwhindawicom Volume 2018


International Journal ofInternational Journal ofPhotoenergy


Journal of

Chemistry


Advances inPhysical Chemistry

Hindawiwwwhindawicom

Analytical Methods in Chemistry

Journal of

Volume 2018

Bioinorganic Chemistry and ApplicationsHindawiwwwhindawicom Volume 2018

SpectroscopyInternational Journal of


Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Medicinal ChemistryInternational Journal of


NanotechnologyHindawiwwwhindawicom Volume 2018

Journal of

Applied ChemistryJournal of



Biochemistry Research International


Enzyme Research


Journal of

SpectroscopyAnalytical ChemistryInternational Journal of


MaterialsJournal of



BioMed Research International Electrochemistry

International Journal of


Na

nom

ate

ria

ls


Journal ofNanomaterials

Submit your manuscripts atwwwhindawicom


For the perspective of modeling chemometrics involvingqualitative tasks can be divided into two categories clas-sification and one-class classification ie data description[15] Classification is vey often considered as a synonym ofdiscriminant analysismethods since they assign a new sampleto one of a set of predefined classes The correspondingclassifier is trained on a training set Data description differsin one essential aspect from the conventional classificationsince it is assumed that only information on a single classis available Data description problems are common in thereal world where positive objects are widely available butnegative ones are maybe hard expensive or even impossibleto gather [16] In the literature three main approaches can bedistinguished the density estimation the boundarymethodsand the reconstructionmethods [17] General demand of anyauthentication problem is that a genuine class ie a targetclass must be known [18 19]The target class is always uniquefor a specific authentication problem Any other objects orclasses of objects that are not members of the target classare considered as outliers This also means that just samplesof the target class can be utilized and that no informationon the other classes is present For data description theboundary surrounding the target class has to be estimatedfrom available data such that it accepts as much of the targetsamples as possible and minimizes the error of acceptingoutlier Up to now much effort has been expended to developclassification algorithms and the concept of data descriptionis also of interest and noticeable [20ndash23] especially in thecases where it is impossible to meaningfully define all ofthe classes and obtain fully representative samples In foodauthenticity the interest is focused on a single target class soas to verify compliance of samples with the features of thatclass and a data description approach should be adopted tobuild an enclosed boundary around the target class

The present work focuses on exploring the feasibilityto utilize near-infrared (NIR) spectroscopy and supportvector data description (SVDD) for authenticity diction ofblack rice Principal component analysis (PCA) is used forexploratory analysis and feature extraction Another two datadescriptionmethods ie k-nearest neighbor data description(KNNDD) and GAUSS method are used as the referenceA total of 142 samples from three brands were collected forspectral analysis All spectra were preprocessed beforehandby standard normal transformation (SNV) Each time thesamples of a brand serve as the target class whereas othersamples serve as the outlier class Based on both the firsttwo principal components (PCs) and original variables threetypes of data descriptions were constructed On average theoptimized SVDD model achieves acceptable performanceie a specificity of 100 and a sensitivity of 942 onthe independent test set with tight boundary The effect oftraining set size and the parameter of kernel width have alsobeen discussed It indicates that SVDD combined with NIR isfeasible and effective for authenticity detection of black rice

2 Theory and Methods

Many methods have been developed to solve the one-classor data description problem and they can be divided into

three main categories density boundary and reconstruc-tion methods Here three algorithms ie support vectordata description Nearest Neighbor Method and GaussianMethod are introduced and used for experiments amongwhich the first two are boundary methods and the last onebelongs to density method

21 Support Vector Data Description (SVDD) SVDD is anovel algorithm for one-class classification problems whichhas been proposed by Tax [15] inspired by the idea of thesupport vector machines It focuses on finding a minimumhypersphere around the target class The hypersphere can beused to decide whether new objects are targets or outliersSuch a sphere is characterized only by center a and radius 119877When seeking sphere it needs to minimize the volume of thesphere by minimizing 1198772 and demand that the sphere coversas many training samples as possible Given the training setx119894 119894 = 1 2 119873 the task in SVDD is to minimize errorfunction

min 119871 (119877 a 120577119894) = 1198772 + 119862sum119894

120577119894 (1)

st 1003817100381710038171003817x119894 minus a1003817100381710038171003817 le 1198772 + 120577119894 120577119894 ge 0 foralli (2)

where a and 119877 are the center and the radius of the hyper-sphere respectively 119862 is the penalty factor which regulatesthe hyperspherical volume and error ie the number oftarget objects rejected 120577119894 is a slack variable for allowableerror limitation Almost all objects are within the sphereThisoptimization problem can be solved by Lagrange multipliermethod [24]

Because the target class is not spherically distributed inmost cases some traditional decision rules may not workwell To make a more effective and flexible decision theoriginal data can be implicitly transformed to a higher dimen-sion by the so-called kernel function119870(x119894 x119895) Several kernelfunctions including linear polynomial Gaussian radial basisfunction (RBF) are available [25 26] In this work the RBFkernel the most commonly used kernel in machine learningwas used The form of RBF kernel is

119870(x119894 x119895) = exp(minus10038171003817100381710038171003817x119894 minus x1198951003817100381710038171003817100381721205902 ) (3)

where 120590 is a key parameter for controlling the boundarytightness

22 Nearest Neighbor Method Themost straightforward andsimplest method to obtain a one-class model is to estimatethe density of the training set Unfortunately it often requiresa large number of samples to avoid the curse of dimension-ality Instead of estimating whole probability densities anindication of the resemblance can also be acquired by com-paring distances Nearest neighbor method can be derivedfrom a local density estimation [27] It avoids the explicitdensity estimation by only using distances to the first nearestneighbor In the process of density estimation a cell often anhypersphere in d-dimension space is centered around the test



119901119873119873 (z) = 119896119873119881119896 1003817100381710038171003817z minus 119873119873119905119903119896 (z)1003817100381710038171003817 (4)






119889 + 12119889 (119889 minus 1) (6)




of

120579119873 = (1205942119889)minus1 (095) (8)


3 Experimental






)

05

1

15Lo

g(1R)

(a)


)

minus1

0

1

Log(1R)

(b)










KnnddGaussSvdd


01

02

03

04

05

06

07

08

09

PC2


knnddSvddGauss


minus08

minus06

minus04

minus02

0

02

04

06

08

1

PC2



knnddSvddGauss


minus08

minus06

minus04

minus02

0

02

04

06

08

PC2







minus1

minus05

0

05

1PC

2

(a) minus045


minus1

minus05

0

05

1

PC2

(b) minus065

minus1

minus05

0

05

1

PC2


(c) minus085


minus1

minus05

0

05

1

PC2

(d) minus15










01

02

03

04

05

06

07

08

09

PC2

0508085

092



5 Conclusions


Data Availability




Acknowledgments


References






































Journal of

Chemistry





Journal of

Volume 2018






Volume 2018




Journal of






Enzyme Research


Journal of



MaterialsJournal of






Na

nom

ate

ria

ls






119901119873119873 (z) = 119896119873119881119896 1003817100381710038171003817z minus 119873119873119905119903119896 (z)1003817100381710038171003817 (4)






119889 + 12119889 (119889 minus 1) (6)




of

120579119873 = (1205942119889)minus1 (095) (8)


3 Experimental






)

05

1

15Lo

g(1R)

(a)


)

minus1

0

1

Log(1R)

(b)










KnnddGaussSvdd


01

02

03

04

05

06

07

08

09

PC2


knnddSvddGauss


minus08

minus06

minus04

minus02

0

02

04

06

08

1

PC2



knnddSvddGauss


minus08

minus06

minus04

minus02

0

02

04

06

08

PC2







minus1

minus05

0

05

1PC

2

(a) minus045


minus1

minus05

0

05

1

PC2

(b) minus065

minus1

minus05

0

05

1

PC2


(c) minus085


minus1

minus05

0

05

1

PC2

(d) minus15










01

02

03

04

05

06

07

08

09

PC2

0508085

092



5 Conclusions


Data Availability




Acknowledgments


References






































Journal of

Chemistry





Journal of

Volume 2018






Volume 2018




Journal of






Enzyme Research


Journal of



MaterialsJournal of






Na

nom

ate

ria

ls






)

05

1

15Lo

g(1R)

(a)


)

minus1

0

1

Log(1R)

(b)










KnnddGaussSvdd


01

02

03

04

05

06

07

08

09

PC2


knnddSvddGauss


minus08

minus06

minus04

minus02

0

02

04

06

08

1

PC2



knnddSvddGauss


minus08

minus06

minus04

minus02

0

02

04

06

08

PC2







minus1

minus05

0

05

1PC

2

(a) minus045


minus1

minus05

0

05

1

PC2

(b) minus065

minus1

minus05

0

05

1

PC2


(c) minus085


minus1

minus05

0

05

1

PC2

(d) minus15










01

02

03

04

05

06

07

08

09

PC2

0508085

092



5 Conclusions


Data Availability




Acknowledgments


References






































Journal of

Chemistry





Journal of

Volume 2018






Volume 2018




Journal of






Enzyme Research


Journal of



MaterialsJournal of






Na

nom

ate

ria

ls





KnnddGaussSvdd


01

02

03

04

05

06

07

08

09

PC2


knnddSvddGauss


minus08

minus06

minus04

minus02

0

02

04

06

08

1

PC2



knnddSvddGauss


minus08

minus06

minus04

minus02

0

02

04

06

08

PC2







minus1

minus05

0

05

1PC

2

(a) minus045


minus1

minus05

0

05

1

PC2

(b) minus065

minus1

minus05

0

05

1

PC2


(c) minus085


minus1

minus05

0

05

1

PC2

(d) minus15










01

02

03

04

05

06

07

08

09

PC2

0508085

092



5 Conclusions


Data Availability




Acknowledgments


References






































Journal of

Chemistry





Journal of

Volume 2018






Volume 2018




Journal of






Enzyme Research


Journal of



MaterialsJournal of






Na

nom

ate

ria

ls






minus1

minus05

0

05

1PC

2

(a) minus045


minus1

minus05

0

05

1

PC2

(b) minus065

minus1

minus05

0

05

1

PC2


(c) minus085


minus1

minus05

0

05

1

PC2

(d) minus15










01

02

03

04

05

06

07

08

09

PC2

0508085

092



5 Conclusions


Data Availability




Acknowledgments


References






































Journal of

Chemistry





Journal of

Volume 2018






Volume 2018




Journal of






Enzyme Research


Journal of



MaterialsJournal of






Na

nom

ate

ria

ls









01

02

03

04

05

06

07

08

09

PC2

0508085

092



5 Conclusions


Data Availability




Acknowledgments


References






































Journal of

Chemistry





Journal of

Volume 2018






Volume 2018




Journal of






Enzyme Research


Journal of



MaterialsJournal of






Na

nom

ate

ria

ls

































Journal of

Chemistry





Journal of

Volume 2018






Volume 2018




Journal of






Enzyme Research


Journal of



MaterialsJournal of






Na

nom

ate

ria

ls




authenticity detection of black rice by near-infrared...

Documents