Near-Infrared Analysis of Whole Kernel Barley: Comparison of Three Spectrometers

Near-Infrared Analysis of Whole Kernel Barley: Comparison ofThree Spectrometers

MIRYEONG SOHN,* DAVID S. HIMMELSBACH, FRANKLIN E. BARTON II,CARL A. GRIFFEY, WYNSE BROOKS, and KEVIN B. HICKSRichard B. Russell Agricultural Research Center, ARS, USDA , P.O. Box 5677, Athens, Georgia 30605 (M.S., D.S.H., F.E.B.); Dept. of Crop &

Soil Env. Sciences, Virginia Polytechnic Institute and State University, Blacksburg, Virginia 24061 (C.A.G., W.B.); and Eastern Regional

Research Center, ARS, USDA, 600 E. Mermaid Lane, Wyndmoor, Pennsylvania 19038 (K.B.H.)

This study was conducted to develop calibration models for determining

quality parameters of whole kernel barley using a rapid and nondestruc-

tive near-infrared (NIR) spectroscopic method. Two hundred and five

samples of whole barley grains of three winter-habit types (hulled, malt,

and hull-less) produced over three growing seasons and from various

locations in the United States were used in this study. Among these

samples, 137 were used for calibration and 68 for validation. Three NIR

instruments with different resolutions, one Fourier transform instrument

(4 cm�1 resolution), and two dispersive instruments (8 nm and 10 nm

bandpass) were utilized to develop calibration models for six components

(moisture, starch, b-glucan, protein, oil, and ash) and the results were

compared. Partial least squares regression was used to build models, and

various methods for preprocessing of spectral data were used to find the

best model. Our results reveal that the coefficient of determination for

calibration models (NIR predicted versus reference values) ranged from

0.96 for moisture to 0.79 for b-glucan. The level of precision of the model

developed for each component was sufficient for screening or classification

of whole kernel barley, except for b-glucan. The higher resolution Fourier

transform instrument gave better results than the lower resolution

instrument for starch and b-glucan analysis. The starch model was most

improved by the increased resolution. There was no advantage of using a

higher resolution instrument over a lower resolution instrument for other

components. Most of the components were best predicted using first-

derivative processing, except for b-glucan, where second-derivative

processing was more informative and precise.

Index Headings: Whole kernel barley; Fourier transform near-infrared

spectroscopy; FT-NIR spectroscopy; Near-infrared spectroscopy; NIR

spectroscopy; Resolution; Fuel ethanol; Partial least squares; PLS

regression; Starch.

INTRODUCTION

There has been growing interest in using barley as afeedstock for fuel ethanol production in the United States.‘Winter’ barley has a shorter growing season than other wintersmall grains on the East Coast, allowing harvest early enoughfor sustainable production of three crops in two years with arotation of winter barley, followed by soybeans, and then corn.However, due to traditional hulled barley’s physical andchemical properties—an abrasive hull, low starch content andhigh viscosity of barley mash—it has not been considered as afuel ethanol feedstock in the U.S. Recently, new varieties ofhull-less barley that have a loosely attached hull, which resultsin removal of the hull easily during harvest and grain cleaning,have been developed and the high starch and protein content ofthese varieties makes them potentially useful for fuel ethanolproduction.1 Barley is composed of starch, protein, moisture,oil, b-glucan, and ash as well as fibrous arabinoxylans and

cellulosic polysaccharides. Starch content is directly related toethanol yield. b-glucan is related to viscosity during fermen-tation, requiring expensive enzymes for break down. Proteinand lipid are factors for quality evaluation of co-products.Therefore, determination of these components prior to barleyuse for ethanol production is important to estimate ethanolyield, to know the dose of enzymes required, or to evaluate co-product quality.

Near infrared spectroscopy (NIRS) has been used for grainanalysis since about 1972 and it is now a well-establishedmethod of analysis for grain quality.2–6 It is a rapid andnondestructive analysis, requiring no sample preparation,allowing short analysis time.

In the previous study,7 we investigated the potential of NIRSfor assessment of barley quality (moisture, starch, b-glucan,protein, lipid, and ash) using ground material and all of thecomponents have been measured successfully with anacceptable accuracy for screening or classification of barley,except for b-glucan. Both the Fourier transform (FT) systemand the dispersive system showed feasibility to be used forbarley analysis.

Being able to predict the component composition of barleyas whole kernel, avoiding a grinding and any wet chemistry, isthe most important use for NIRS as a rapid assessment tool ofbarley for ethanol production. Previous attempts have beenmade to use NIRS as a method for screening of barley qualityin breeding programs.8,9 Chemometric prediction of barleycomponents from NIRS data has proven useful for classifica-tion of barley into low, medium, and high quality groups10 andfor explorative classification in genetic breeding.11

In this study, NIR calibration models were developed forpredicting six quality parameters of starch, b-glucan, moisture,protein, oil, and ash in whole kernel barley. Three NIRinstruments with different resolutions were employed to obtainthe best calibration models and to determine whether spectralresolution had a bearing on prediction results of whole kernelquality parameters.

EXPERIMENTAL

Barley Samples. The source of the whole kernel barleysamples was the same as that of the ground barley in ourprevious study.7 These samples were provided by researchersat Virginia Polytechnic Institute and State University andchemical analyses were conducted at the Eastern RegionalResearch Center (USDA-ARS) in Wyndmoor, Pennsylvania. Atotal of 205 whole kernel barley samples were used in thisstudy and all the equations developed are based on wholebarley kernel.

Reference Analysis. Reference values of the six grain

Received 31 October 2007; accepted 17 January 2008.* Author to whom correspondence should be sent. E-mail: [email protected].

Volume 62, Number 4, 2008 APPLIED SPECTROSCOPY 4270003-7028/08/6204-0427$2.00/0

� 2008 Society for Applied Spectroscopy

compositional quality parameters (moisture, starch, b-glucan,protein, oil, and ash) compared to corresponding data for wholekernel barley are the same as those obtained from the previousstudy,7 except for moisture content. The moisture content forwhole kernel barley was determined by drying 10 grams at 1308C for 20 hours.12

Near-Infrared Spectroscopy. Three NIR spectrometerswere utilized in this study, a Bruker Vector22/N (BrukerOptics, Billerica, MA), a NIRSystems 6500 (FOSS NIRSys-tems, Inc., Laurel, MD), and a Foss Rapid Content Analyzer-XDS (FOSS North America, Eden Prairie, MN). The first oneis a Fourier transform system and the next two instruments aredispersive systems. For the Vector22/N, whole kernel sampleswere placed in a cylindrical rotating cup (89 mm i.d.) andspectral data were collected in a diffuse reflection mode over arange of 10 000 to 4000 cm�1 (3112 data points) using BrukerOPUS software (ver. 3.0.19). For the NIRSystems 6500 andXDS, samples were placed in a square transport cell (3 3 14in.) and spectral data were collected in a diffuse reflectionmode over a range of 400 to 2498 nm with 2 nm intervals(1050 data points). WinISI software (ver. 2.01, InfrasoftInternational Inc., Port Matilda, PA) and ISIscan software(ver. 2.83, Foss NIRSystems, Inc.) were, respectively, used for

collecting data. The resolutions of the instruments are 4 cm�1

for Vector22/N, 10 nm bandpass for NIRSystems 6500, and 8nm bandpass for XDS. Each of the samples were scanned threetimes (by repacking) and the triplicate spectra were averaged.

Data Processing and Chemometrics. Chemometric pro-cessing of the data was conducted using Matlab software (ver.7.3, Mathworks, Inc., Natick, MA) with PLS_Toolbox (ver.4.0, Eigenvector Research, Inc., Manson, WA). The FT datawere converted to Unscrambler (ver. 9.0, CAMO, Norway)format and then imported into Matlab. The NIRSystems 6500data were imported into Matlab through JCAMP andUnscrambler format, in that order. The XDS data wereimported into Matlab through NSAS and Unscrambler format,in that order. The 205 samples were sorted by reference datafor each constituent, and every third sample was extracted foran external validation set (n ¼ 68), giving six calibration setsand six validation sets for six constituents. Calibration modelswere developed using the remaining 137 samples with partialleast squares (PLS) regression. Full leave-one-out cross-validation was performed during model development todetermine the optimal number of PLS factors. Savitzky–Golayfirst- or second-derivative processing followed by multiplica-tive scatter correction (MSC) was utilized as preprocessing ofthe spectral data. A third-order polynomial with 7 points (or 17points) convolution intervals was used for derivative process-ing. Four preprocessing regimes were used in this study:D(1,3,7)þMSC, D(1,3,17)þMSC, D(2,3,7)þMSC, andD(2,3,17)þMSC, where the three numbers in parenthesisrepresent derivative order, polynomial order, and convolutionintervals, respectively. All data were mean centered beforeanalysis. Outlier detection was based on student residuals,where samples with overþ2 or less than�2 were removed for99% confidence. Performance of each PLS model was reportedas the multiple coefficient of determination (R2) and root meansquare error of cross-validation (RMSECV). Predictionperformance was reported as the root mean square error ofprediction (RMSEP) and a ratio of deviation to performance(RPD), which is the ratio of the standard deviation of thereference data to the RMSEP, which provides a method-independent standardization of the RMSEP.13 Correlationmethods with RPD values ranging from 2.4 to 3.0 areconsidered adequate for rough screening purposes, values of3.1 to 4.9 are adequate for standard screening purposes, andvalues of 5.0 to 6.4 are adequate for quality control.13

RESULTS AND DISCUSSION

In the previous study,7 using the NIR region (1100–2498nm) gave better results than using the entire visible–NIR region(400–2498 nm) in the development of calibration models forground barley. Therefore, in this study, the dispersive datawere truncated to remove the visible region (400–1098 nm)and only the NIR region (1100–2498 nm) was used for wholekernel barley analysis. The FT data were not truncated with arange of 10 000 to 4000 cm�1.

Figure 1a shows FT-NIR spectra of 205 whole kernel barleysamples. The spectra were treated with multiplicative scattercorrection to remove scatter effects due to inhomogeneoussamples. Unlike the spectra of ground barley,7 the spectra ofwhole kernel barley were not consistent even after MSCtreatment, giving large variations between samples at mostwavelengths. A difference in the absorption pattern wasobserved, particularly for the wavelengths between 7000 and

FIG. 1. FT-NIR spectra of kernel barley. (a) Multiplicative scatter correctiontreated spectra of 205 barley samples; (b) average spectrum of 163 hull-lessspectra, 26 hulled spectra, and 16 malt spectra.

428 Volume 62, Number 4, 2008

4000 cm�1. It is caused by the barley type and could be bettervisualized by averaging the spectra of the various types (Fig.1b). Three spectra show the differences among hull-less barley(thick solid line), malt barley (thin solid line), and hulled barley(dotted line). Each spectrum is the average of 163, 16, and 26spectra, respectively. Clearly there are differences betweencovered barley (hulled and malt) and hull-less barley. The hull-less type appears from its spectrum to have a lower watercontent (near 5100 cm�1) and lower ligno-cellulosic content(4500–4000 cm�1) than the covered type.

Partial least squares (PLS1) regression was performed topredict the components of interest from intact barley kernels (n¼ 137) using the three devices, high-resolution FT (Vector22/N), mid-resolution (XDS), and low-resolution (NIRSystems6500) instruments, respectively. Calibration models developedwere tested by the validation set (n ¼ 68), which was notinvolved in calibration development. Table I shows thereference data for the kernel barley samples used in this study.Range, mean, and standard deviation of the moisture content

obtained for kernel barley were 9.64–18.45%, 13.03%, and1.52, respectively. The fact that moisture values for groundbarley (range: 9.03–13.37%, mean: 10.59%)7 are lower thanthose for whole kernel barley even though they are the samesample set is due to the heat generated during the grindingprocess, which drives off some of the water.

Tables II through IV summarize the PLS regressions andvalidation results of the models for six whole graincomponents. The model for each component shown in thetables is the best result from four models developed usingdifferent pretreatments. The selection of the best model wasbased on the lowest RMSEP, the highest R2 value, and thelowest number of PLS factors.

Moisture content was best predicted using the first derivativewith either 7 or 17 points in all three instruments. The R2 andRMSEP values of the models obtained using three factorsranged from 0.940 to 0.963 and 0.39% to 0.40%, respectively.There was no difference in accuracy of the moisture modelsobtained from three instruments. The RPD value of 3.7 to 3.8indicates that the NIR method could be used for standardscreening of moisture in whole kernel barley. The whole kernelmodel resulted in a higher prediction error than the groundmodel (RMSEP¼ 0.3), but a better RPD was obtained. This iscaused by the higher standard deviation of the reference data(SD ¼ 1.52) compared to the ground barley (SD ¼ 0.81).

For starch, the best equation was obtained from the use ofthe high-resolution FT instrument, which gave an R2 of 0.915and a RMSEP of 1.39% using five factors. The slight decreasein resolution afforded by the mid-resolution instrument (XDS)resulted in the same RMSEP value (1.39%), but more factors (7factors) were required to achieve this result and a lower R2

(0.816) was obtained. Using the low-resolution instrument(NIRSystems 6500) decreased the R2 to 0.777 and increased

TABLE I. Chemical composition of kernel barley samples (%, dryweight basis).a

Components

Calibration set Validation set

N Range Mean SD N Range Mean SD

Moisture 137 9.64–18.45 13.03 1.52 68 9.76–17.83 13.05 1.49Starch 137 49.83–68.19 60.92 3.64 68 51.86–67.53 61.02 3.49b-glucan 137 2.33–5.76 4.32 0.66 68 2.66–5.66 4.34 0.63Protein 137 6.81–12.22 9.10 1.11 68 7.26–12.09 9.10 1.04Oil 95 1.10–2.96 2.33 0.37 46 1.63–2.96 2.36 0.38Ash 137 1.30–2.55 1.88 0.28 68 1.30–2.45 1.86 0.28

a Source, Ref. 7. N¼ number of samples; SD¼ standard deviation

TABLE II. PLS models for kernel barley developed using the FT-NIR instrument.a

Components Pretreatmentsb

Calibration Validation

N PLS factors R2 RMSECV (%) N RMSEP (%) RPD

Moisture Der(1,3,17)þMSCþMC 138 3 0.963 0.30 68 0.40 3.73Starch Der(1,3,7)þMSCþMC 138 5 0.915 1.76 68 1.39 2.51b-glucan Der(2,3,17)þMSCþMC 138 5 0.796 0.52 68 0.43 1.47Protein Der(1,3,17)þMSCþMC 138 7 0.948 0.35 68 0.39 2.67Oil Der(1,3,17)þMSCþMC 95 6 0.864 0.21 47 0.14 2.71Ash Der(1,3,7)þMSCþMC 138 5 0.906 0.12 68 0.12 2.33

a N¼number of samples; R2¼ coefficient of determination for calibration; RMSECV¼ root mean square error of cross-validation; RMSEP¼ root mean square errorof prediction; RPD ¼ SD/RMSEP.

b Der(1,3,7)þMSCþMC¼ first derivative with third polynomial and 7 points followed by multiplicative scatter correction and mean centering.

TABLE III. PLS models for kernel barley developed using the XDS instrument.a







APPLIED SPECTROSCOPY 429

the RMSEP to 1.52%. Consequently, the high-resolution FTinstrument gave the best result in predicting starch with theleast error, the highest R2, and the lowest number of factors,indicating that the starch prediction could be improved byhigher resolution.

Protein was best predicted using first-derivative processingfor all three instruments. Models developed using seven factorsresulted in good coefficients of determination (. 0.90). Themodels from the XDS and NIRSystems 6500 gave the sameresult for R2 and similar results for RMSEP (0.34% to 0.35%).The FT data improved the R2 to 0.948 but did not decrease theRMSEP. This indicates that there is no advantage in the use ofa high-resolution instrument over a dispersive system.Estimation of oil content resulted in an RMSEP of 0.13% to0.14% for all three instuments. For ash, all instruments gaveequivalent RMSEP values of 0.12%, even though theVector22/N model was slightly better than the 6500 andXDS models in R2. Most of the components were bestpredicted using the first derivative, whereas b-glucan was bestpredicted using the second derivative for all three instruments.The b-glucan models from the Vector22/N and NIRSystems6500 gave similar results for RMSEP (0.43% to 0.44%) but abetter R2 was obtained from the Vector22/N. The XDS modelhad the lowest R2 and the highest RMSEP, as well as thehighest number of factors. However, comparison of theaccuracy between the models from the two dispersiveinstruments was not easy to evaluate because of their lowvalues for R2 (0.474 to 0.505). As discussed in the previousstudy,7 poor results for the b-glucan could be caused by thenarrow range of the reference values for the commercial barleysamples. The scatter plots of NIR predicted versus measuredvalues of the six components in whole kernel barley by the bestfit models are presented in Fig. 2.

Typically, calibration models obtained with non-milledsamples are not as good as those obtained with milled samplesbecause of spectral variability caused by sample inhomogene-ity. However, there are some reports that non-milled samplesgave similar or even better results than milled samples.14,15 Inthis study, the results for the whole kernel barley werecomparable to those for the ground barley7 for most of thecomponents. The whole kernel models resulted in slightlyhigher values for RMSEP than the ground models, but thedifference was insignificant. The R2 was higher and the numberof factors used was lower for the whole kernel models. Theonly exception was the protein model, which was more precisewith the ground material. When compared to other studies onwhole kernel barley,8,10,16 the accuracy of the modelsdeveloped in this study was somewhat improved for most of

the components. However, the protein results obtained byEdney et al.17 (SECV ¼ 0.31%) were a little better than thoseobtained in the current study (SEP ¼ 0.34%).

On the basis of the RPD values, it was concluded that theNIR calibration models developed in this study could be usedfor screening for moisture, starch, protein, and oil of wholekernel barley. For ash and b-glucan, the accuracy of the modelsdid not meet the requirment for screening but the equationscould still be used to classify barley into low and high groups.Apparently, resolution is of more importance when the samplesare intact kernels than when the samples are ground. The low-resolution instrument was accurate enough to predict starch andb-glucan on ground barley, whereas the high-resolution FTinstrument was required for whole kernel barley. For othercomponents analyzed, there was no predictive advantage inusing a higher-resolution FT instrument over a lower resolutioninstrument. Thus, either a high-resolution, a mid-resolution, ora low-resolution instrument would be suitable for themeasurement. The FT instrument is more flexible than thedispersive instruments. The FT system could be set up forlower resolution, taking advantage of the improved signal-to-noise ratio to predict protein, moisture, oil, and ash, and set tohigh resolution to predict b-glucan and starch, or alternativelythe higher resolution data could just be smoothed more. In mostcases where estimates of concentrations of all grain compo-nents is desired, use of a single NIR system would be mostpractical.

CONCLUSION

The current research supports the possibility of using NIRmethods for determining quality parameters of whole kernelbarley that could be used as a feedstock for fuel ethanolproduction and other end uses. NIR calibration modelsdeveloped using three instruments with different resolutionsrevealed that the higher resolution FT instrument (Vector 22/N)offered better results over the lower resolution dispersiveinstruments (XDS and NIRSystems 6500) for starch and b-glucan analysis. However, there was no advantage in using theFT instrument for other components. First-derivative process-ing was more useful than second-derivative processing for theanalysis of most of the components. The level of precision ofthe model for each component was sufficient to be used for arapid screening or classification of whole kernel barley, givingRPD values ranging from 2.5 to 3.8, except for the b-glucanand ash models. The NIR calibration models should be usefulto estimate ethanol yield or to evaluate co-product quality from

TABLE IV. PLS models for kernel barley developed using the NIRSystems 6500 instrument.a








the barley prior to ethanol production. Further research will be

conducted on calibration transfer between NIR instruments.

ACKNOWLEDGMENTS

The authors would like to thank Michael Kurantz, Eastern Regional

Research Center, USDA-ARS, Wyndmoor, PA, for the wet chemical analysis

of the barley samples. Thanks are also given to Ms. Judy Davis, Richard B.

Russell Research Center, USDA-ARS, Athens, GA, for the NIR data collection.

1. K. B. Hicks, R. A. Flores, F. Taylor, A. J. McAloon, R. A. Moreau, D. B.

Johnston, G. E. Senske, W. S. Brooks, and C. A. Griffey, The 15th Annual

EPAC Ethanol Conference (Cody, WY, June 12–14, 2005).

2. K. H. Norris and J. R. Hart, Meas. Control Sci. Ind. Pap. Int. Symp. 4, 19

(1965).

3. I. Ben Gera and K. H. Norris, Isr. J. Agric. Res. 18, 125 (1968).

4. D. D. Archibald, D. B. Funk, and F. E. Barton II, Proc.-Int. Soc. Opt. Eng.

3543, 141 (1998).

5. F. E. Barton II, J. S. Shenk, M. O. Westerhaus, and D. B. Funk, J. Near

Infrared Spectrosc. 8, 201 (2000).

6. M. Sohn, D. S. Himmelsbach, and F. E. Barton II, Cereal Chem. 81, 429

(2004).

7. M. Sohn, D. S. Himmelsbach, F. E. Barton II, C. A. Griffey, W. Brooks,

and K. B. Hicks, Appl. Spectrosc. 61, 1178 (2007).

8. J. H. Helm, L. Oatway, and P. Juskiw, 18th North American Barley

Researchers Workshop (Alberta, CA, July 17–20, 2005), http://www1.

agric.gov.ab.ca/$department/deptdocs.nsf/all/fcd10228.

FIG. 2. Near-infrared predicted versus measured values of moisture, starch, b-glucan, protein, oil, and ash content in whole kernel barley on the best PLS modelsfrom Tables II through IV. Black circles are for the calibration set (n ¼ 137) and white circles are for the validation set (n ¼ 68).

APPLIED SPECTROSCOPY 431

9. B. G. Osborne, J. Near Infrared Spectrosc. 14, 93 (2006).10. L. A. Oatway and J. H. Helm, Barley Newsletter (1998).11. S. Jacobsen, I. Søndergaard, D. Møller, T. Desler, and L. Munck, J. Cereal

Sci. 42, 281 (2005).12. ASAE S352.2 (Am. Soc. Agric. Engrs., St Joseph, Michigan, 1988).13. P. C. Williams, ‘‘Implementation of near-infrared technology’’, in Near-

infrared technology in the agricultural and food industries, P. C. Williamsand K. H. Norris, Eds. (Am. Assoc. Cereal Chem., St. Paul, MN, 2001),2nd ed., p. 165.

14. D. C. Perez-Marin, A. Garrido-Varo, J. E. Guerrero-Ginel, and A. Gomez-Cabrera, Anim. Feed Sci. Technol. 116, 333 (2004).

15. M. J. De la Haba, A. Garrido-Varo, J. E. Guerrero-Ginel, and D. C. Perez-Marin, J. Agric. Food Chem. 54, 7703 (2006).

16. C. K. Black and J. F. Panozzo, Proceedings of the 10th InternationalConference for Near Infrared Spectroscopy, R. K. Cho and A. M. C.Davies, Eds. (NIR Publications, Chichester, UK, 2002), p. 339.

17. M. J. Edney, J. E. Morgan, P. C. Williams, and L. D. Campbell, J. NearInfrared Spectrosc. 2, 33 (1994).


Documents

Near-Infrared Analysis of Whole Kernel Barley: Comparison of Three Spectrometers