Chemometrics and Intelligent Laboratory Systems 111 (2012) 34–38

Contents lists available at SciVerse ScienceDirect

Chemometrics and Intelligent Laboratory Systems

j ourna l homepage: www.e lsev ie r .com/ locate /chemolab

Development of NIR calibration models to assess year-to-year variation in totalnon-structural carbohydrates in grasses using PLSR

Nisha Shetty a,⁎, René Gislum a, Anne Mette Dahl Jensen b, Birte Boelt a

a Aarhus University, Science and Technology, Department of Agroecology, Research Centre Flakkebjerg, DK-4200, Slagelse, Denmarkb University of Copenhagen, Faculty of Life Sciences, Forest and Landscape, Rolighedsvej 23, DK-1958, Frederiksberg C, Denmark

⁎ Corresponding author. Tel.: +45 89156000; fax: +E-mail address: nisha.shetty@agrsci.dk (N. Shetty).

0169-7439/$ – see front matter © 2011 Elsevier B.V. Alldoi:10.1016/j.chemolab.2011.11.004

a b s t r a c t

a r t i c l e i n f o

Article history:Received 4 July 2011Received in revised form 3 November 2011Accepted 9 November 2011Available online 20 November 2011

Keywords:NIRPLSRiPLSTNCGrassesFructan

Near-infrared (NIR) spectroscopy was used in combination with chemometrics to quantify total non-structural carbohydrates (TNC) in grass samples in order to overcome year-to-year variation. A total of1103 above-ground plant and root samples were collected from different field and pot experiments andwith various experimental designs in the period from 2001 to 2005. A calibration model was developedusing partial least squares regression (PLSR). The calibration model on a large data set spanning five yearsdemonstrated that quantification of TNC using NIR spectroscopy was possible with an acceptable low rootmean square of prediction error (RMSEP) of 1.30. However, for some years the estimated RMSEP was too op-timistic as year-to-year variation for new years was not included in the model. Interval partial least squares(iPLS) regression was applied to remove non-relevant spectral regions and in order to improve model perfor-mance, but still it was not possible to avoid year-to-year variation using iPLS, however iPLS simplified the in-terpretation of the regression model. The best option was to expand the database with samples from a newyear, to include these samples in the calibration model and to apply this on the remaining samples from thefuture year.

© 2011 Elsevier B.V. All rights reserved.

1. Introduction

Total non-structural carbohydrates (TNC) include water solublecarbohydrates (WSC) as well as starch and fructans. In cool seasongrasses, WSC consist mainly of the simple sugars: glucose, fructoseand sucrose, whereas fructans are storage carbohydrates based on along chain of fructose polymers. Fructans are considered an alterna-tive to starch for storing carbohydrates when sucrose concentrationis above a certain threshold level, and recently fructans are suggestedto provide protection against environmental stresses such as drought[1]. Fructans are gaining increased attention due to their potential forenhancing the forage quality of plants for ruminants [2] and involve-ment regrowth after defoliation [3]. The increased focus on TNC hasresulted in an increased desire to measure WSC and fructans. Al-though quantification of WSC is easier to perform than detailed struc-tural analysis of the fructans, it is still impractical to characterise alarge number of samples. Therefore there is a need for rapid methodsfor quantification of TNC. Development of an innovative method forquantification and/or screening of samples for TNC will be of interestin breeding and grass testing programmes. The ability to measurefructans in large numbers of samples rapidly will facilitate the

45 89156082.

rights reserved.

identification of grasses with higher forage quality through a higherand more stable concentration of fructan during the growing season.

Near-infrared (NIR) spectroscopy has been used for the screeningof crossbred material from cereal breeding programmes since the late1970s, and the ease and speed of the technique have made it possibleto perform thousands of quality tests in the time between harvestingof one generation and sowing of the next [4].

Despite the successful implementation of NIR within cereal breed-ing programmes, NIR has gained limited attention as a method forquality testing in grass breeding. It is not that the use of NIR for predic-tion of quality parameters in grass has not been investigated. Numer-ous studies have been performed and are thoroughly described inAndrés et al. [5], as well as in newer publications such as Feuersteinand Paul [6] describing the use of on-line NIR on a plot harvester,while the use of NIR for determination of quality parameters in dryand grinded grass samples was discussed by Gislum et al. [7]. Howev-er, most of these studies have been proof-of-concept studies made ona small number of samples (b100) not representing variation betweene.g. years, cultivars, species or locations. An exception to the smallnumber sample study was made by Shetty and Gislum [8], who devel-oped PLSRmodels using 1459 samples covering a large variation. Eventhough some of these studies have demonstrated good correlationsbetween quality parameters and NIR, they have not resulted in the de-velopment of methods that are in use in the industry.

In order to make robust models (capable of performing well whenapplied to the new data) that can handle the multivariate nature of

Table 1Mean and standard deviation of difference (SDD) of TNC for five measurements of foursamples (A, B, C and D) listed in order of TNC. In samples A, B and C five sub-samplesfrom the three samples were weighed and analysed. In sample D five sub-sampleswere measured from the same extract.

Sample Mean TNC SDD TNC

C 3.47 1.84B 15.00 3.47A 17.98 4.18D 26.57 6.07

35N. Shetty et al. / Chemometrics and Intelligent Laboratory Systems 111 (2012) 34–38

the samples and the highly collinear NIR spectroscopic data it is es-sential to apply appropriate data analytical techniques such as thechemometric methods principal component analysis (PCA) [9], par-tial least squares regression (PLSR) [10], and interval PLS (iPLS)[11]. Here, PLSR will be used in the pursuit of a robust method. Fur-thermore it will be investigated whether variable selection can im-prove the predictive ability (or robustness) of the models as well asto perceive on which vibrational bands in the spectra, informationon TNC can be found. For this purpose iPLS which searches for a spec-tral interval that is particularly informative with respect to the pa-rameter under consideration, was used. This method often leads toimprovements of prediction ability over standard full-spectrumPLSR models [11–13]. This is due to the high degree of redundancyin an NIR spectrum. Furthermore, there may be large parts of thespectra which do not correlate to the constituent of interest, andthus this part of the spectra may add noise, or inconsistencies to themodel. Making a model on only the relevant part of the spectra canthus lead to superior regression models.

PLSR was used to develop NIR calibration model to assess year-to-year variation in TNC in grasses, when the model is applied on futuresamples. Even though variation between different species could alsobe the topic, the present study focuses on the year-to-year variation.The high year-to-year variation in the present study was due to loca-tion, species and climate.

2. Materials and methods

2.1. Materials

A total of 1103 above-ground plant and root samples were collect-ed from different field and pot experiments from 2001 to 2005 (ab-breviated from 1 to 5). Plants including roots were dug up from thesoil; plants were separated from the roots at soil surface and thenroots were hand washed. The sample set consisted of red fescue (Fes-tuca rubra L.), perennial ryegrass (Lolium perenne L.), smooth stalkedmeadow grass (Poa pratensis L.) and orchard grass (Dactylis glomerataL.) samples. Plant and root samples were weighed and freeze-driedand weighed again. Afterwards the samples were ground using aCyclotec 1093 sample mill (Foss, Hillerød, Denmark) to pass througha 1 mm screen prior to determination of carbohydrate concentrationon aliquots of the samples.

2.2. Chemical analysis

TNC were extracted using a 0.1 M sodium acetate buffer andhydrolysed with 0.037 M sulphuric acid. Quantification of TNC wasdone using a coupled enzymatic assay procedure [14]. Results arepresented in percentage on a dry matter basis.

The reproducibility of the measurement of TNC (y data) was esti-mated as the standard deviation of differences (SDD). SDDwas calculat-ed on five measurements of four samples, three samples with differentTNC concentrations and one sample with the same extract (Table 1).

SDD ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi∑ di−dmð Þ2

n−1ð Þ

s

where di=difference in y between five replicatemeasurements of sam-ple i, dm=mean value of all replicate differences (∑di/n) and n=num-ber of samples.

The repeatability of the ydatawas calculated as the standarddeviation(SD) on one sample which was measured in almost each run (n=48).

SD ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi∑ yi−ymð Þ2

n−1ð Þ

s

where yi is the y-value of replicate i, ym is the average y value of all repli-cates and n is the number of replicates.

SD and SDD were used to evaluate the performance of the PLSRmodels.

2.3. NIR measurements

Reflectance spectra of dried and ground plant and root sampleswere obtained using a QFA-Flex 400 FT-NIR instrument (Q-interline,Roskilde, Denmark). The samples were packed as uniformly as possi-ble in glass vials (height 6 cm, diameter 2.6 cm) andmeasured using arotating sample device. The sample was rotated at three rounds perminute. The measuring sample window at the rotating sample devicehad a diameter of 6 mm and the analysis surface was ≈510 mm2.Spectra were collected at every 2 nm in the NIR region from 1100 to2498 nm. One spectrum was obtained for each sample as an averageof 64 sub-scans. The spectra were reported as log (1/R). Using thisprocedure, approximately 20 samples could be analysed per hour.

2.4. Multivariate data analysis

PCA was performed as an explorative data analysis in order to ob-tain a first look at the structure of the data, to identify outliers and todelineate classes. The PLSR method was used to derive calibrationmodels. PLSR models were obtained on raw and pre-processed NIRspectra. Pre-processing included Savitzky–Golay first derivative(1 d) and second derivative (2 d) [15] averaging over 5 points usinga second order polynomial and multiplicative signal correction(MSC) [16]. Root mean square error of cross-validation (RMSECV)plotted against the number of PLSR components using different pre-processing methods is used to select the optimum pre-processingmethod and optimum number of components in the PLSR model.The optimum number of PLSR components was chosen as the firstlocal minimum in the smooth declining RMSECV curve or the pointwhere this curve flattened. Random cross-validation with 10 seg-ments and 9 iterations was used.

The performance of the PLSR models were evaluated using theroot mean square error of prediction (RMSEP), standard error of per-formance (SEP) and bias. The relationship between RMSEP, SEP andbias is: RMSEP2≈SEP2+Bias2, if bias is≈0 then RMSEP2≈SEP2

[17]. Initially, the obtained model was validated using 20% of theavailable data. The test set was selected as every fifth sample aftersorting according to TNC to ensure that samples from the whole con-centration range were represented in the test set. Models were alsovalidated using one of the years as test set in a loop in order to eval-uate how year-to-year variations affected the model.

All data analyses were carried out using MATLAB version 7.9.(R2009b) (The Mathworks, Inc., Natick, MA, USA) along with thePLS toolbox 5.5.2 (Eigenvector Research, Inc., Manson, WA, USA).

3. Results and discussion

The SDD of TNC ranged from 1.84% TNC to 4.18% TNC dependingon the mean TNC concentration when sub-samples were analysedfrom the same sample (Table 1). When a repeated measurement of

Table 2Comparison of errors from the developed PLSR models using all wavelengths (1100–2498), optimal pre-processing methods and optimal number of PLSR components against2*SDD values from the chemical analysis of TNC. The numbers in parentheses represent the RMSEP, SEP and Bias obtained when all the models were run using the same 2Dpre-processing method and the same 9 PLSR components similar to 20% test set.

Test set data No. of samples in test set Pre-processing #PLSR comp. RMSEP SEP Bias 2*SDD

20% 219 2D 9 1.30 1.30 0.03Year 1 113 1D 7 1.84 (2.43) 1.84 (2.25) −0.07 (0.95)Year 2 112 1D 11 2.05 (2.96) 1.63 (1.85) −1.25 (−2.32) 3.68–8.36Year 3 442 1D 7 2.10 (2.38) 1.97 (2.20) −0.73 (−0.91)Year 4 149 2D 7 1.28 (1.29) 1.25 (1.23) 0.30 (0.42)Year 5 281 2D 10 2.36 (2.54) 1.84 (1.99) 1.48 (1.59)

RawMSC1d2d

6

5.5

5

4.5

4

3.5

RM

SE

CV

(%

TN

C)

3

2.5

2

1.5

12 4 6 8

PLS components10 12 14

Fig. 2. RMSECV plot plotted against the number of PLSR components using differentpre-processing methods.

Fig. 1. Raw spectra (A) and second derivative spectra (B) of the 1097 grass samples.

36 N. Shetty et al. / Chemometrics and Intelligent Laboratory Systems 111 (2012) 34–38

the same extract was performed SDD was 6.07 for a mean %TNC of26.57 (Table 1). The 2*SDD given in Table 2 is in accordance to thecalculations performed in Table 1.

Scatter correction of the raw NIR spectra (Fig. 1A) using 1 d, 2 d,and MSC clearly reduced the offset in the spectra. Mean 2 d spectrashow clear peaks, especially in the higher NIR wavelengths(Fig. 1B). Six samples were removed as outliers using Residual vs.Hotelling T2 plot. Fig. 2 shows lower RMSECV for 2 d compared to1 d up until five PLSR components. Whereafter there was not muchdifference between the two pre-processing methods. Raw and MSChad considerably higher RMSECV values. A steep decrease in RMSECVis a positive sign for a model as it indicates that the variation de-scribed in the first PLSR components is well focused on the Y variable.Accordingly, 2 d was chosen as final pre-processing method for cali-bration model built using 80% of the available data and nine PLSRcomponents were chosen as optimal. When the PLSR model wasbuilt using four out of five years data in calibration set (left out oneyear data was used to validate the model), 1 d and 2 d pre-processing methods still had lower RMSECV values compared toraw and MSC (data not shown). Final chosen pre-processing methodsand optimal PLSR components when each year left out in a loop isshown in Table 2. The PLSR model developed using nine PLSR compo-nents and all wavelengths (1100 to 2498 nm) had an RMSEP and SEPof 1.30% TNC and concurrently a low bias (0.03% TNC) (Table 2). PLSRmodel results for year-to-year validation are also shown in Table 2.

The iPLS plot demonstrates prediction error (RMSEP) for 20(Fig. 3A) and 40 (Fig. 3B) equidistant sub-interval models (bars)and for the full-spectrum PLSR model (line) using 20% test set valida-tion with nine PLSR components plotted together with the 2 d meanspectrum. Apparently, no interval PLSR model can compete with thefull-spectrum PLSR model using nine PLS components, even thoughPLSR interval model for interval 19 of the 20 interval model (interval37 and 38 of the 40 interval model) had RMSEP values below 2% TNC.

The predicted %TNC plotted against the measured %TNC showed ahigh linear correlation of R2=0.96 (Fig. 4).

Since it is not possible to develop a model where year-to-year var-iation is eliminated, it may be a possibility to include some samplesfrom the future year in the calibration model and apply the modelon the remaining samples from that year. The question is ‘howmany samples from the future year have to be included in the calibra-tion model’ before the RMSEP becomes stable. To answer this ques-tion samples from one year were predicted using PLSR modelsdeveloped on four of the five years and an increasing number of sam-ples from the test year. Samples from the test year were moved fromthe validation data to the calibration data one by one and the RMSEPwas calculated (Fig. 5).

Fig. 3. iPLS plot using 20% test set validation for 20 interval (A) and 40 interval (B) PLSR models (bars) and for the full-spectrum PLSR model (dotted line) using nine PLSR com-ponents plotted together with the mean 2 d spectrum. The number of PLSR components for each interval PLSR model is shown at the bottom of each bar.

37N. Shetty et al. / Chemometrics and Intelligent Laboratory Systems 111 (2012) 34–38

In order for the current model to be acceptable it must accuratelyand precisely report the TNC in samples from a current season [18].RMSEP for PLSR calibration models developed on four years and in-clusion of samples one by one from the fifth year showed a large var-iation according to which year was used for calibration model (Fig. 5).It is not surprising that RMSEP decreased when more and more sam-ples from the same year were gradually included in the calibrationmodel [19,20] For example RMSEP on samples from year 2 decreasedfrom ~3% to ~2% TNC by inclusion of ~50 samples. Year 3 behavedvery differently, and the same low RMSEP was not achieved for thisyear, even after inclusion of samples from the third year. This is prob-ably due to the larger range in TNC for this year compared to the fourother years (Fig. 5). The sharp increase/decrease in RMSEP in Fig. 5especially when adding last few samples is due to the high variationin %TNC values within each year.

In the present experiment the high number of samples and thefact that samples contain a high variation due to e.g. location, speciesand climate variation due to different years support the choice of arelatively high number of PLSR components. The complexity of thesample set could also influence the proportion of the TNC which isextracted and the particle size of the tissue after grinding. The in-creased complexity of the data also accounts for higher estimated er-rors than found in creeping bentgrass turf by Narra et al. [21], whofound standard errors of calibration of 0.7% based on a dataset of112 samples from the period 1998 to 2001. McGrath et al. [22]reported SEP of 0.98% for fructan in wheat shoots. Whereas Jafari etal. [23] found a SEP of 1.2 for WSC in perennial ryegrass based on153 samples taken in the period 1995 to 1997. Brink and Marten

Fig. 4. Predicted vs. measured plot for the full-spectrum PLSR model using 20% test setvalidation with nine PLSR components on 2 d pre-processed spectra.

[24] found standard errors of 1.17%, 1.40% and 1.59% for log 1/R, 1 dand 2 d pre-processed data respectively for TNC in alfalfa rootsbased on 72 samples. Batten et al. [25] found SEP of 1.5% and 1.4%for non-structural carbohydrates based on 61 and 34 samples in riceand wheat shoots respectively and these errors are higher than theerror obtained in the present experiment.

The error found in the current study based on 1097 samples islikely to be more realistic than the values from the small studiesdue to the inclusion of more variation. The SEP values for the modelswere compared with SDD in Table 2 showing very promising values.For practical purposes, the errors from the models are usually accept-ed if SEP≤2*SDD [17] and TNC have acceptable, low error comparedwith the chemical analysis. It should be noted that SDD varies accord-ing to concentration level (the SDD shown in Table 2 is the SDD fromTable 1 multiplied by 2).

When samples from more years are included in the calibration setone could expect better prediction ability. However results should becarefully examined when data from a new year are included in thecalibrationmodel. For example the present study showed that, the es-timated RMSEP is not quite as low when models were validated year-to-year (Table 2). Year-to-year variation is not fully covered when themodel is based on four years and applied on a fifth year. It will accord-ingly be risky to apply the current model based on five years on sam-ples from future years and believe that the RMSEP holds up.

Fig. 5. RMSEP plot using one year as test set and including samples from this one by onein the calibration set.

38 N. Shetty et al. / Chemometrics and Intelligent Laboratory Systems 111 (2012) 34–38

All six models shown in Table 2 are also run using same pre-processing method (2D) and same number of PLSR components (9)similar to 20% test set, as this makes comparison between test setseasier (results are shown in parentheses, Table 2). Similar toTable 2, the results showed high variation in RMSEP when modelswere validated year-to-year, except for year 4, which showed similarRMSEP as the 20% test set, indicating that the variation in year 4 iscovered by the other four years. The largest RMSEP of 2.96 wasfound for year 2. However this was drastically decreased to 2.05upon use of optimal pre-processing and number of PLSR components.The decrease indicates that the model is very sensitive to the pre-processing being performed. The overall worst year is year 5, indicat-ing that this year was very different compared to the four other years.

It was therefore tested whether the year-to-year variation couldbe reduced using variable selection in the data analysis. iPLS regres-sion models were developed using 10 to 50 equidistant sub-intervals, which all indicated the same wavelength areas to containmost information on Y. iPLS plots using 20% test set validation for20 and 40 equidistant sub-intervals are shown in Fig. 3. None of theiPLS models showed improved prediction however the interestingpart is the interpretation of the iPLS models. The best sub-modelsfor the 20 interval PLSR model were at interval number 19 corre-sponding to 2220–2350 nm, interval number 9 corresponding to1418–1470 nm and at interval numbers 13 and 14 corresponding to1658–1808 nm. These intervals mainly represented the functionalgroups CH, CH2 and CH3. It is interesting that the selected wave-lengths for prediction of TNC using iPLS are mostly identical withthe important wavelength selected for prediction of fructans byShetty and Gislum [8]. This underlines that TNC and fructan concen-trations are strongly correlated. It may be that the prediction of TNCis in fact based on the fructan concentration. If there is a positive cor-relation between TNC and fructans it is not a problem, however, ifnew strains are developed and WHC ratio deviates strongly fromthis linearity, the application of the TNC model may not be possible.The informative part of the spectra is, not surprisingly, dominatedby CH stretch modes and CH deformation modes [26].

iPLS models were also developed using year-to-year validation(not shown). Unfortunately, the estimate of the prediction errorusing validation according to year is not reduced using the iPLS se-lected wavelength regions (alone or together with other relevantareas). The model failed due to high spectral difference as well as dif-ference in %TNC between years (data not shown). Accordingly, vari-able selection does not reduce the year-to-year variation.

A possible strategy would be to measure the carbohydrates in thelaboratory and obtain NIR spectra in the beginning of the year. At onepoint (e.g. after 200 samples according to years 3 and 5 in Fig. 5) car-bohydrates could be predicted using NIR. A strategy from an instru-mental point of view would probably be to measure the first fewsamples with both NIR and reference method. If there was a good cor-respondence between the reference and prediction, there would beno need to update the PLS-model. If not, an update would be neces-sary, and maybe even more references.

In practice, predicted values should be followed by an outlierwarning indicating whether the predicted value is outside range orthat the sample is deviating from the calibration samples, in whichcase the value is not to be trusted and the sample should be sent foranalysis in the laboratory.

4. Conclusions

Calibration of an NIR model on a large data set spanning five yearsdemonstrated that measurement of TNC with NIR is possible with anacceptable low error. However, for some years the estimated RMSEPis too optimistic as year-to-year variation for new years was not

included in the model. It was not possible to avoid year-to-year vari-ation in the models by selecting only specific regions of the spectra forregression using iPLS. Thus the best way to overcome year-to-yearvariation is to expand the database with samples from a new year, in-clude these samples in the calibration and apply the calibration on theremaining samples from the future year. Outlier warnings will aid indeciding when the model is robust for application.

References

[1] D.P. Livingston III, D.K. Hincha, A.G. Heyer, Fructan and its relationship to abioticstress tolerance in plants, Cellular and Molecular Life Sciences 66 (2009)2007–2023.

[2] D.R. Biggs, K.R. Hancock, In vitro digestion of bacterial and plant fructans and ef-fect on ammonia accumulation in cow and sheep rumen fluids, Journal of Generaland Applied Microbiology 44 (1998) 167–171.

[3] V. Amiard, A. Morvan-Bertrand, J.P. Billard, C. Huault, M.P. Prud'homme, Fate offructose supplied to leaf sheaths after defoliation of Lolium perenne L.: assessmentby 13 C-fructose labelling, Journal of Experimental Botany 54 (2003) 1231–1243.

[4] B.G. Osborne, Applications of near infrared spectroscopy in quality screening ofearly-generation material in cereal breeding programmes, Journal of Near Infra-red Spectroscopy 14 (2006) 93–101.

[5] S. Andrés, I. Murray, A. Calleja, F.J. Giraldéz, Nutritive evaluation of forages by nearinfrared reflectance spectroscopy, Journal of Near Infrared Spectroscopy 13(2005) 301–311.

[6] U. Feuerstein, C. Paul, NIR-spectroscopy of non-dried forages as a tool in breedingfor higher quality — laboratory tests and online investigations in plot harvesters,XXVII Eucarpia symposium on improvement of fodder crops and amenity grasses.Copenhagen, Denmark, 2007.

[7] R. Gislum, T. Asp, B. Boelt, NIRS and chemometrics — exploration of grass foragequality, XXVII Eucarpia symposium on improvement of fodder crops and amenitygrasses. Copenhagen, Denmark, 2007.

[8] N. Shetty, R. Gislum, Quantification of fructan concentration in grasses using NIRspectroscopy and PLSR, Field Crops Research 120 (2011) 31–37.

[9] S. Wold, K.H. Esbensen, P. Geladi, Principal components analysis, Chemometricsand Intelligent Laboratory Systems 2 (1987) 37–52.

[10] H. Martens, T. Næs, Multivariate Calibration, Wiley, New York, 1993.[11] L. Nørgaard, A. Saudland, J. Wagner, J.P. Nielsen, L. Munck, S.B. Engelsen, Interval

partial least squares regression (iPLS): a comparative chemometric study with anexample from near infrared spectroscopy, Applied Spectroscopy 54 (3) (2000)413–419.

[12] A. Borin, R.J. Poppi, Application of mid infrared spectroscopy and iPLS for thequantification of contaminants in lubricating oil, Vibrational Spectroscopy 37(2005) 27–32.

[13] H. Winning, N. Viereck, L. Norgaard, J. Larsen, S.B. Engelsen, Quantification of thedegree of blockiness in pectins using 1H NMR spectroscopy and chemometrics,Food Hydrocolloids 21 (2007) 256–266.

[14] K. Larsson, S. Bengtsson, Determination of easily available carbohydrates in plantmaterial, Report of the National Swedish Laboratory for Agricultural Chemistry,22, 1983, pp. 1–10.

[15] A. Savitzky, M.J.E. Golay, Smoothing and differentiation of data by simplified leastsquares procedures, Analytical Chemistry 36 (1964) 1627–1639.

[16] P. Geladi, D. McDougall, H. Martens, Linearization and scatter-correction for near-infrared reflectance spectra of meat, Applied Spectroscopy 39 (1985) 491–500.

[17] K.H. Esbensen, Multivariate Data Analysis—In Practice, 4th ed. CAMO, Oslo, 2000.[18] T.M. Venås, Å. Rinnan, Determination of weight percent gain in solid wood mod-

ified with in situ cured furfuryl alcohol by near-infrared reflectance spectroscopy,Chemometrics and Intelligent Laboratory Systems 92 (2008) 125–130.

[19] J.A. Guthrie, D.J. Reid, K.B. Walsh, Assessment of internal quality attributes ofmandarin fruit. 2. NIR calibration model robustness, Australian Journal of Agricul-tural Research 56 (2005) 417–426.

[20] J.A. Guthrie, C.J. Liebenberg, K.B. Walsh, NIR model development and robustnessin prediction of melon fruit total soluble solids, Australian Journal of AgriculturalResearch 57 (2006) 1–8.

[21] S. Narra, T.W. Fermanian, J.M. Swiader, Analysis of mono- and polysaccharides increeping bentgrass turf using near infrared reflectance spectroscopy, Crop Sci-ence 45 (2005) 266–273.

[22] V.B. McGrath, A.B. Blakeney, G.D. Batten, Fructan to nitrogen ratio as an indicatorof nutrient stress in wheat crops, New Phytologist 136 (1997) 145–152.

[23] A. Jafari, V. Connolly, A. Frolich, E.J. Walsh, A note on estimation of quality param-eters in perennial ryegrass by near infrared reflectance spectroscopy, Irish Journalof Agricultural and Food Research 42 (2003) 293–299.

[24] G.E. Brink, G.C. Marten, Analysis of alfalfa root carbohydrate concentration bynear infrared reflectance spectroscopy, Crop Science 26 (1985) 159–161.

[25] G.D. Batten, A.B. Blakeney, V.B. McGrath, S. Ciavarella, Non-structural carbohy-drate: analysis by near infrared reflectance spectroscopy and its importance asan indicator of plant growth, Plant and Soil 155/156 (1993) 243–246.

[26] B.G. Osborne, T. Fearn, P.H. Hindle, Theory of near infrared spectrophotometry,Practical NIR Spectroscopy with Applications in Food and Beverage Analysis,2nd ed., Longman Scientific and Technical, Harlow, UK, 1993, pp. 13–35.