Food Chemistry 145 (2014) 278–283

Contents lists available at SciVerse ScienceDirect

Food Chemistry

journal homepage: www.elsevier .com/locate / foodchem

Analytical Methods

Identification of additive components in powdered milk by NIR imagingmethods

0308-8146/$ - see front matter � 2013 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.foodchem.2013.06.116

⇑ Corresponding author. Tel./fax: +86 10 62733091.E-mail address: minsg@cau.edu.cn (S. Min).

Yue Huang a,b, Shungeng Min a,⇑, Jia Duan a, Lijun Wu a, Qianqian Li a

a College of Science, China Agricultural University, Beijing 100193, PR Chinab Third Class Tobacco Supervision Station, Beijing 101121, PR China

a r t i c l e i n f o a b s t r a c t

Article history:Received 26 May 2011Received in revised form 26 February 2013Accepted 26 June 2013Available online 4 July 2013

Keywords:Powdered milkNear-infrared imagingMelamineSemi-quantification

The express assay of excessive additives in powdered milk is of vital necessity, especially during indus-trial production. Near-infrared microscopy provides chemical information on the spatial distribution andcluster side of components in milk-based products when materials are mixed together. Distributions oftwo additive components and one banned chemical in powdered milk were simulated in this study.The distribution of inorganic additive ZnSO4 was identified using the relationship imaging mode. The dis-tribution image of lactose was obtained by assigning the wavenumber region and by using principal com-ponent analysis coupled with correlation coefficient imaging. In addition, classical least square regressionwas employed to quantify the banned additive, melamine, in the powdered milk. Lastly, the detectionlimit of melamine in powdered milk was determined using the relationship imaging mode.

� 2013 Elsevier Ltd. All rights reserved.

1. Introduction

Near infrared (NIR) spectroscopy has proven its efficiency in foodquality evaluation and pollution analysis over the past few years.NIR spectroscopy has been applied to the routine detection of com-ponents in dairy products because of its high penetrability, non-destructive behaviour, and ease of pretreatment (Wu, Feng, & He,2007). NIR imaging technology is generally noninvasive, requiresminimal sample preparation, poses less thermal damage to a sam-ple material, and can yield a response in real time compared withFTIR, Raman imaging (Zhang, Hanson, & Sekulic, 2005), or magneticresonance (MR) scanning (Alessandra, Maria, Olimpia, Massimili-ano, & Paolo, 2010). Thus, sample section and other pretreatmentsare rendered unnecessary in NIR spectroscopy. Moreover, in-situdetermination is feasible in NIR spectroscopy (Monteiroa,Ambrozin, Santos, Boffo, & Pereira-Filho 2009; Roman, Ekaterina,& Ravilya, 2010; Roman, Ekaterina, & Ravilya, 2011). On-lineimaging detection was successfully realised, e.g., acrylamide inpotato chips was predicted with R2 of 0.83 and an average predic-tion error of 266 lg/kg (Pedreschi, Segtnan, & Knutsen, 2010).

Amongst various spectroscopic analytical methods, NIRmicroscopy can determine which chemical species are present ata micro-scale level and can provide spatial information on theirdistribution within a sample. In recent years, the wide applicationof convenient detectors makes NIR imaging more promising inevaluating food quality.

The content of additive components is a crucial factor thataffects the entire quality of powdered milk. Traditional analyticalmethods such as high-performance liquid chromatography orliquid chromatography–mass spectroscopy (LC–MS) are time con-suming, destructive, and costly. Hence, spectral analyses are pre-ferred as simple and direct methods in evaluating the quality ofpowdered milk (Borin, Ferrao, Mello, Maretto, & Poppi, 2006; He,Liu, Lin, Awika, & Mustapha, 2008; Kamishikiryo-Yamashita,Oritani, Takamura, & Matoba, 1994; Wu, He, & Feng, 2008; Wuet al., 2007; He, & Sun, 2008). Qin, Xu, Zhou, and Wang (2004) per-formed a quality analysis involving the determination of fat, sugar,and protein by using the second derivative spectra of powderedmilk. They evaluated the thermal perturbation of samples by usingtwo-dimensional correlation infrared spectroscopy. The expressmethod of evaluating the quality of powdered milk based on thefingerprint of the FTIR spectra was studied by Deng, Zhou, andSun (2006). Protein, fat, and lactose in raw milk were quantifiedusing the NIR–PLS method (Roman, Ekaterina, & Ravilya, 2011;Sasic & Ozaki, 2001). Moros, Garrigues, and Guardia (2007)detected the components in infant powdered milk by using Ramanspectroscopy. Although numerous studies on detecting the compo-nents in powdered milk by using NIR spectroscopy have been per-formed, studies on hyper spectral imaging method remains rare. Infact, hyperspectral imaging at the visible and short near infrared(VIS/SNIR) region has been used to identify the target componentsin complicated materials for quite a long time already. The robust,reliable combination of chemical and digital imaging features havebeen successfully applied in diverse fields such as remote sensing(Kokaly, Asner, Ollinger, Martin, & Wessman, 2009), agricultural

Y. Huang et al. / Food Chemistry 145 (2014) 278–283 279

products (Pedreschi et al., 2010), food (Gowen, O’Donnell, Cullen,Downey, & Frias, 2007), and pharmaceuticals (Gendrin, Roggo, &Collet, 2007; Roggo, Edmond, Chalus, & Ulmschneider, 2005). Thus,NIR imaging is a valuable method in investigating the locationwhere the components being studied are distributed and in moni-toring particular banned chemicals.

The banned nitrogen-rich chemical melamine (2,4,6-triamino-1,3,5-triazine) has caused widespread food safety concerns inrecent years. Melamine abuse has been reported in products suchas milk, infant formula, pet food, and candy. The reproductive andurinary systems of animals could be damaged with prolonged in-take of melamine (Roman & Sergey, 2011). Melamine has a highnitrogen content of about 66%. The Kjeldahl method is not suitablein detecting the protein content in powdered milk because exces-sive melamine could interfere with the instrument reading value.Rapid methods of testing the presence of melamine are essential be-cause of the serious health concerns associated with melamine con-sumption and the extensive scope of affected products. Progressesin detection and quantification by using FTIR, NIR, and Raman spec-troscopy have been reported (He et al., 2008; Lin et al., 2008). Thelimit of detection (LOD) of low-concentration melamine can reachbelow 1 ppm (0.76 ± 0.11 ppm) when an appropriate nonlinearalgorithm as support vector regression was applied to spectrumanalysis (Roman & Sergey 2011). Moreover, except for a robustPLS model (R2 > 0.99, RMSECV P 0.9, and RPD P 12.), factorizationanalysis of the spectra can differentiate pure infant formula powderfrom samples of 1 ppm melamine with no misclassifications, a con-fidence level of 99.99%, and a short assay time of 2 min for detection(Mauer, Chernyshova, Hiatt, Deering, & Davis, 2009).

In the present study, the main feature of measurements throughNIR imaging includes the large amount of information collected forone sample (i.e., one spectrum could be measured over a widewavenumber range in each pixel of the image). The obtained datacan be described as a spectral hypercube. The cube has a threedimensional structure of X � Y � k, where the X and Y dimensionsrepresent the spatial information and the k dimension correspondsto the spectral pattern. Such data structures could comprise thou-sands of spectra. These raw data, normally highly correlated witheach other, would require a specific imaging mode to extract thedesired information.

In this study, four imaging methods (relationship imaging, RI;chemical imaging, CI; principal component analysis, PCA imaging;and classical least square, CLS imaging) were applied to profilingthe distribution of the two allowed additives, namely, ZnSO4 andlactose, as well as that of melamine in the counterfeit powderedmilk. The optimal imaging mode and its parameters were used toidentify the additive components in powdered milk based on theirdifferent spectral information. Moreover, the existence of mela-mine in powdered milk was profiled, and the feasibility of semi-quantification for melamine was achieved.

2. Experiment and methods

2.1. Reagents and instruments

Branded powdered milk was purchased from the market (Yilinonfat dry milk). Test objects including lactose, ZnSO4, and mela-mine were obtained from Sinopham Chemical Reagent Co., Ltd.An electronic analytical balance (Mettler Inc., Sweden), oscillator(WT2, Harbin China) and an electrical thermal vacuum drying oven(ZK82-BB, vacuum 267 kPa, Shanghai China) were used for samplepreparation. The spectra of pure materials were collected usingSpectrum One NTS (PerkinElmer Inc., US). The samples were ana-lysed in an NIR line mapping system (Spectrum Spotlight 400FT-NIR Microscope, PerkinElmer, US). The spectra were obtained

from each acquisition by using a linear MCT detector array. Spec-trum software version 5.0.1 (PerkinElmer Inc., US) was employedin interpreting the spectral data. All images were processed usingSpotlight software version 1.1.0 and Hyperview software version2.0 (PerkinElmer Inc., US).

2.2. Scanning

All materials were preserved in the dryer at 4 ± 1 �C before theexperiment. The entire experimental course, which includesweighing, loading, mixing, and scanning, was performed at ambi-ent temperature of 20 ± 1 �C. The samples were not treated withany reagent. Proper amount of powdered milk and three additiveswere manually mixed to counterfeit a binary system, which wasthen loaded into the sample container before being subjected tovisible spectroscopy. Samples in container were described as adiameter of 10 mm and a height about 2 mm, featuring a flat sur-face. The surface of each sample was scanned. One complete scan-ning procedure took about 10 min.

2.3. Imaging methods

The imaging modes used in this study include RI, PCA, and CLS.The RI mode requires a full-range spectrum of the pure material,which is used as the reference. The values of the generated imageare based on the correlation coefficient between the spectrum fromeach pixel and the reference. All of these values construct the dis-tribution of the scanned region.

PCA was performed to acquire multivariate images, and wascoupled with the correlation coefficient imaging mode to analysethe mixture containing highly correlated components. This processprovided factor scores and loading vectors, which were used tointerpret the information on covariance structures. Correlationcoefficient imaging may be supplemented to these score plots tohelp represent the relative intensities of individual score valuepairings (Geladi & Grahn, 1996).

2.4. Semi-quantification of melamine

Multivariate CLS algorithm is often applied in spectroscopicanalysis, and is usually a good choice for analysing hyperspectralimages such as those in this paper. CLS regression is based onthe assumption that the measured spectra are the sum of purecompound spectra weighted by the concentration of the com-pounds. The relative concentrations of the compounds in the sam-ple can be estimated using only the pure compound spectra basedon Beer–Lambert’s law.

In this study, the CLS imaging method was employed to identifythe banned additive melamine. The relative concentration of eachpixel within the scanned image of the mixture can be simulta-neously obtained as well. Afterwards, samples with melaminewere loaded and detected from high to low concentration to deter-mine the LOD. Each sample was scanned three times at differentareas. The RI mode was used to acquire an image of the melaminedistribution. Until no melamine was found in the scanned imageduring the three scanning times or in the extracted spectrum ofthe suspicious spot over below the identified correlation coeffi-cient, we consider that melamine was not detected by this method.

3. Results and discussion

3.1. Distribution of ZnSO4

ZnSO4 is a common inorganic nutrient additive in commercialpowdered milk. Inorganic matter has no absorbance in the

280 Y. Huang et al. / Food Chemistry 145 (2014) 278–283

near-infrared region; hence, the spectrum of ZnSO4 differs muchfrom that of powdered milk, as shown in Fig. 1. The correlationcoefficient of ZnSO4 and powdered milk is 0.40, which was ob-tained by matching their respective spectra.

The optical imaging area was about 5000 � 5000 lm. The NIRimaging area was adjusted to capture object distribution at arbi-trary positions within this optical area. The spatial resolution ofmicroscopy was 6.25 lm. The spectral region ranged from7800 cm�1 to 4000 cm�1, and the wave number resolution was32 cm�1 with an interval of 16 cm�1. Each pixel channel wasscanned 64 times. The scanned region of the visible image of themixed sample showed the total absorbance image (Fig. 2A). Fromthis image, we cannot conclude the distribution of either compo-nent. The colour band on the right side represents the intensityof absorbance. The RI mode was used to re-image the scannedregion because the contour of the component in the total absor-bance image was ambiguous. The reference spectrum was that ofpowdered milk. Clear distribution was observed in the RI image(Fig. 2B), wherein the red region represents powdered milk andthe green region represents ZnSO4.

Given the differences of the two components, the CI mode waschosen to determine their distributions. The spectra obtained fromthe different spots in the image had a major distinction in the de-fined wavenumber range of 5100–5000 cm�1. The CI image is

Fig. 1. NIR spectra of lactose, ZnSO4

Fig. 2. Scanned region of the sample which consisted of powdered milk and zinc su

shown in Fig. 2C, wherein the distinct contour of powdered milkin the green region and ZnSO4 in the red region were captured.

The inorganic additive in powdered milk, having a major spec-tral difference with the matrix, could therefore be imaged by bothRI and CI mode to visualise the distribution.

3.2. Distribution of lactose

Lactose is the primary ingredient in powdered milk. Mixed lac-tose and powdered milk are quite similar in terms of colour andparticle size. The surface of powdered milk with lactose wasscanned in reflectance mode. An area of 950 � 950 lm was imagedusing a pixel size of 25 � 25 lm. Thus, about 1560 spectra wereobtained for the original image. The wavenumber range was from7800 cm�1 to 4000 cm�1 with an interval of 16 cm�1. The wave-number resolution was 32 cm�1. Each channel was scanned 16times. The spectra from all the pixels were standardised with theSNV treatment so as to eliminate the differences in absorptionintensity of the uneven particles.

The matrix effect of the lactose and powdered milk spectrawere highly correlative, as shown in Fig. 1. Differentiating thetwo components from the total absorbance image was difficult.Thus, the spectra from the original data cube were decomposedusing the PCA algorithm. The PCA is an important approach in

, melamine and powdered milk.

lphate: (A) total absorbance image; (B) relationship image; (C) chemical image.

Fig. 3. Comparison of eigenvector spectra of the first PC and pure spectra ofpowdered milk over the full spectral region and defined region, respectively.

Y. Huang et al. / Food Chemistry 145 (2014) 278–283 281

extracting useful chemical information. It is used to describe theimportant spectral information in a reduced number of compo-nents (Baronti, Casini, Lotte, & Porcinai, 1997). Each factor is

Fig. 4. Score images of the first two PCs (upper part); correlation coefficient im

Fig. 5. CLS images and concentration histog

relatively independent, and the score image can represent themicrostructure characteristics and component distribution. Thefactor number is based on the complexity of the mixed sample.More factors would be used if the sample is composed of morecomponents. The contrariwise factor number should be controlledto avoid over dispersion of effective information (Clarke, 2004).The eigenvector spectrum of PC1 was composed of 189 points, asshown in Fig. 3 (left). By matching the spectra of the first factor’seigenvector and that of powdered milk over the full spectralregion, the first PC showed a low correlation with powdered milkwith a correlation coefficient of 0.2110. From the score image ofthe first PC, we could roughly judge the distribution of powderedmilk, as shown in the upper part of Fig. 4, whereas the distributionof PC2, representing lactose, was poorly visualised.

Two components shared close spectral shape in all wavenum-ber range. The spectra overlapped after mixing; hence, the specificband at the region from 6500 cm�1 to 5500 cm�1, where the spec-tra of the two components significantly differed, was assigned forPCA decomposition. Similarly, the eigenvectors spectra of the firsttwo PCs were selected, and were made up of 63 data points, asshown in Fig. 3 (right). The correlation coefficient of PC1 with pow-dered milk significantly improved to 86.03%, whereas the correla-tion between PC2 and its corresponding component remained low.

ages based on the loading vector spectra of the first two PCs (lower part).

rams of melamine and powdered milk.

Fig. 6. Relationship images of different concentration levels.

282 Y. Huang et al. / Food Chemistry 145 (2014) 278–283

The correlation coefficient images are shown in the lower part ofFig. 4 basing on the loading vector spectra of the first two PCs.The distribution of powdered milk can be clearly viewed fromthe left image. Lactose can be also easily identified in the rightimage.

In summary, PCA offered a contour description of the targetobject after decomposing the principal components of a spectralcorrelated mixture. Moreover, assigning the proper wavenumberrange was essential in applying PCA combined with the correlationcoefficient imaging approach when analysing components thatexert a significant matrix effect and have mutual interferences.

3.3. Semi-quantification of melamine

When treating a complicated mixture with multi-components, arelatively accurate quantification can be achieved using the CLSimaging approach. Powdered milk with melamine was selected as

the sample, and was scanned in the reflectance mode. The obtainedimage had an area of 900 � 800 lm, and had a total of 170 � 140pixels. The dispersion size for each pixel was 6.25 � 6.25 lm. Thewavenumber range was from 7500 cm�1 to 4000 cm�1 with aninterval of 16 cm�1. The wavenumber resolution was 32 cm�1. Eachchannel was scanned 64 times. The extracted spectra were pro-cessed with SNV to eliminate the differences in absorption intensityon the analyte surface.

The pure material spectra of powdered milk and melaminewere used as an input to the CLS algorithm. After conducting calcu-lation and spatial pixel reconstruction, the concentration imagesfor the two components were obtained (Fig. 5), wherein the colourband represents the relative content of the components. The redspeckles in the left image represent melamine, and the light blueregion represents the matrix. Similarly, the deep blue spots inthe right image represent melamine. The rest of the image showsa profile of yellow and red spots, which indicates a high correlation

Y. Huang et al. / Food Chemistry 145 (2014) 278–283 283

of the other components in the matrix except for melamine. Thelower half of Fig. 5 represents the concentration histograms of mel-amine and powdered milk respectively. The content of melaminewas quite low in the mixture; hence, its pixels are located in thelow concentration region (0–0.5, w/w). For components withmutual interferences and exerts a significant matrix effect in thepowdered milk, highly spectral-correlated pixels occupied mostof the concentration image, accounting for a relative content ofabove 50% in the histogram. From these results, the CLS algorithmwas concluded to be a practical approach in quantifying a targetcompound from a complicated mixture.

We attempted to reach the limit of imaging detection. As shownin Fig. 6, the red region occupies a relatively large portion in theimages of 30%, 10%, and 5% (w/w) melamine. The correlation coef-ficients of these extracted spectra and the reference were all over0.9. Variously, the red spot representing melamine was rarelyfound in the two scans when the melamine content was 1%. Thethird scan is shown in Fig. 6, and a suspicious spot of melaminewas found. The spectra from pixels within this spot had a low cor-relation of 0.607. Despite the low content of melamine and thespectral overlapping effect of powdered milk, melamine was stillconsidered to have been detected. Therefore, the detection limitof melamine was defined to be about 1% (w/w) in powdered milk,which is quite higher than that of other approaches.

4. Conclusion

The NIR micro-imaging system can be used to acquire the totalchemical information of a multi-component mixed sample and canbe used when spatial information becomes relevant for an analyt-ical application. The distribution of a target component is related toselecting a specific imaging mode and the method for data extrac-tion. In this study, the distributions of three target components in apowdered milk matrix were obtained using four different imagingmethods. Subsequently, we defined the lowest detection limit ofmelamine in powdered milk. Compared with the limit of0.05 mg/kg from the LC–MS method and 0.56 mg/kg from ELISAdetection (Lutter et al., 2011), the NIR imaging method was lesssensitive in detecting the trace amount of melamine in powderedmilk. However, in analysing powdered milk products whose mela-mine content surpasses the standard limit of melamine, the NIRimaging method may be accepted as a simple and intuitive detec-tion approach. From the overview of this work, a target componentin the matrix could be identified and detected using the NIR micro-imaging method. This method can also be applied to detect foreignmatter in a pure matrix, and to perform semi-quantification on aspecific component in the mixture. The NIR micro-imaging methodhas good visibility and proper sensitivity compared with otherqualitative detections. With NIR micro-imaging methods andchemometrics, further quantitative analysis of foreign matters ina pure material or a specific ingredient in a mixture sample canbe feasible. The simplicity of sample preparation and imaging pro-cedure would make the proposed technique promising for indus-trial applications. The possibility of real-time quality control byusing NIR imaging is also interesting for the practical implementa-tion of the proposed method.⁄NOTE: Due to one dead dot of the MCT detector, some images

in this study present transverse defect with line mapping system.Thus the dead pixels are corrected by neighboring interpolationso that they don’t affect the general analysis.

Acknowledgement

The authors gratefully acknowledge the financial supportprovided by the National Natural Science Foundation of China(No. 20575076).

References

Alessandra, C., Maria, T. A., Olimpia, M., Massimiliano, V., & Paolo, S. (2010).Seasonal chemical–physical changes of PGI Pachino cherry tomatoes detectedby magnetic resonance imaging (MRI). Food Chemistry, 122, 1253–1260.

Baronti, S., Casini, A., Lotte, F., & Porcinai, S. (1997). Principal component analysis ofvisible and near-infrared multispectral images of works of art. Chemometricsand Intelligent Laboratory System, 39, 103–114.

Borin, A., Ferrao, M. F., Mello, C., Maretto, D. A., & Poppi, R. J. (2006). Least-squaressupport vector machines and near infrared spectroscopy for quantification ofcommon adulterants in powdered milk. Analytica Chimica Acta, 579, 25–32.

Clarke, F. (2004). Extracting process-related information from pharmaceuticaldosage forms using near infrared microscopy. Vibrational Spectroscopy, 34,25–35.

Deng, Y. E., Zhou, Q., & Sun, S. Q. (2006). Analysis and discrimination of infantpowdered milk via FTIR spectroscopy. Spectroscopy and Spectral Analysis, 26(4),636–639.

Geladi, P., & Grahn, H. (1996). Multivariate image analysis. Chichester: Wiley.Gendrin, C., Roggo, Y., & Collet, C. (2007). Content uniformity of pharmaceutical

solid dosage forms by near infrared hyperspectral imaging: A feasibility study.Talanta, 73, 733–741.

Gowen, A. A., O’Donnell, C. P., Cullen, P. J., Downey, G., & Frias, J. M. (2007).Hyperspectral imaging – An emerging process analytical tool for food qualityand safety control. Trends in Food Science and Technology, 18, 590–598.

He, L., Liu, Y., Lin, M. S., Awika, J., & Mustapha, A. (2008). A new approach to measuremelamine, cyanuric acid, and melamine cyanurate using surface enhancedRaman spectroscopy coupled with gold nanosubstrates. Sensing andInstrumentation for Food Quality and Safety, 2(1), 66–71.

He, Y., & Sun, D. W. (2008). Study on infrared spectroscopy technique for fastmeasurement of protein content in milk powder based on LS-SVM. Journal ofFood Engineering, 84(1), 124–131.

Kamishikiryo-Yamashita, H., Oritani, Y., Takamura, H., & Matoba, T. (1994). Proteincontent in milk by near-infrared spectroscopy. Journal of Food Science, 59(2),313–315.

Kokaly, R. F., Asner, G. P., Ollinger, S. V., Martin, M. E., & Wessman, C. A. (2009).Characterizing canopy biochemistry from imaging spectroscopy and itsapplication to ecosystem studies. Remote Sensing of Environment, 113, 78–91.

Lin, M., He, L., Awika, J., Yang, L., Ledoux, D. R., et al. (2008). Detection of melaminein gluten, chicken feed, and processed foods using surface enhanced.

Lutter, P., Perroud, M. S., Gimenez, E. C., Meyer, L., Goldmann, T., et al. (2011).Screening and confirmatory methods for the determination of melamine incow’s milk and milk-based powdered infant formula: Validation andproficiency-tests of ELISA, HPLC–UV, GC-MS and LC–MS/MS. Food Control, 22,903–913.

Mauer, L. J., Chernyshova, A. A., Hiatt, A., Deering, A., & Davis, R. J. (2009). Melaminedetection in infant formula powder using near- and mid-infrared spectroscopyagric. Journal of Agricultural and Food Chemistry, 57, 3974–3980.

Monteiroa, M. R., Ambrozin, A., Santos, M. S., Boffo, E. F., & Pereira-Filho, E. R. (2009).Evaluation of biodiesel–diesel blends quality using 1H NMR and chemometrics.Talanta, 78, 660–664.

Moros, J., Garrigues, S., & Guardia, M. (2007). Evaluation of nutritional parameters ininfant formulas and powdered milk by Raman spectroscopy. Analytica ChimicaActa, 593, 30–38.

Pedreschi, F., Segtnan, V. H., & Knutsen, S. H. (2010). On-line monitoring of fat, drymatter and acrylamide contents in potato chips using near infrared interactanceand visual reflectance imaging. Food Chemistry, 121, 616–620.

Qin, Z., Xu, C. H., Zhou, Q., & Wang, J. (2004). Chinese Journal of Analytical Chemistry,32, 1156–1160.

Roggo, Y., Edmond, A., Chalus, P., & Ulmschneider, M. (2005). Infrared hyperspectralimaging for qualitative analysis of pharmaceutical solid forms. AnalyticaChimica Acta, 535, 79–87.

Roman, M. B., Ekaterina, L., & Ravilya, S. (2011). Neural network (ANN) approach tobiodiesel analysis: Analysis of biodiesel density, kinematic viscosity, methanoland water contents using near infrared (NIR) spectroscopy. Fuel, 90, 2007–2015.

Roman, M. B., Ravilya, S., & Ekaterina, L. (2010). Gasoline classification using nearinfrared (NIR) spectroscopy data: Comparison of multivariate techniques.Analytica Chimica Acta, 671, 27–35.

Raman spectroscopy and HPLC, (2007). Journal of Food Science, 73(8), 129–134.Roman, M. B., & Sergey, V. S. (2011). Melamine detection by mid- and near-infrared

(MIR/NIR) spectroscopy: A quick and sensitive method for dairy productsanalysis including liquid milk, infant formula, and milk powder. Talanta, 85,562–568.

Sasic, S., & Ozaki, Y. (2001). Short-wave near-infrared spectroscopy of biologicalfluids. 1. Quantitative analysis of fat, protein, and lactose in raw milk by partialleast-squares regression and band assignment. Analytical Chemistry, 73, 64–71.

Wu, D., Feng, S., & He, Y. (2007). Infrared spectroscopy technique for thenondestructive measurement of fat content in milk powder. Journal of DairyScience, 90(8), 3613–3619.

Wu, D., He, Y., & Feng, S. (2008). Short-wave near-infrared spectroscopy analysis ofmajor compounds in milk powder and wavelength assignment. AnalyticaChimica Acta, 610, 232–242.

Zhang, L., Hanson, J. M., & Sekulic, S. (2005). Multivariate data analysis for Ramanimaging of a model pharmaceutical tablet. Analytica Chimica Acta, 545, 262–278.