Journal of Food Engineering 126 (2014) 107–112

Contents lists available at ScienceDirect

Journal of Food Engineering

journal homepage: www.elsevier .com/ locate / j foodeng

Preliminary study on the use of near infrared hyperspectral imagingfor quantitation and localisation of total glucosinolates in freeze-driedbroccoli

0260-8774/$ - see front matter � 2013 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.jfoodeng.2013.11.005

⇑ Corresponding author. Tel.: +353 1 805 9500; fax: +353 1 805 9550.E-mail address: gerard.downey@teagasc.ie (G. Downey).

José Miguel Hernández-Hierro a, Carlos Esquerre b, Juan Valverde c, Salvador Villacreces c, Kim Reilly d,Michael Gaffney d, M. Lourdes González-Miret a, Francisco J. Heredia a, Colm P. O’Donnell b,Gerard Downey c,⇑a Food Colour & Quality Laboratory, Department of Nutrition & Food Science, Universidad de Sevilla, Facultad de Farmacia, 41012 Sevilla, Spainb Biosystems Engineering, School of Agriculture, Food Science and Veterinary Medicine, University College Dublin, Belfield, Dublin 4, Irelandc Teagasc Food Research Centre Ashtown, Dublin 15, Dublin, Irelandd Horticulture Development Unit, Teagasc Research Centre, Kinsealy, Dublin 17, Ireland

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

Article history:Received 27 March 2013Received in revised form 5 November 2013Accepted 6 November 2013Available online 15 November 2013

Keywords:GlucosinolatesBroccoliHyperspectral imagingNear infraredVisibleChemometrics

The use of hyperspectral imaging to (a) quantify and (b) localise total glucosinolates in florets of a singlebroccoli species has been examined. Two different spectral regions (vis–NIR and NIR), a number of spec-tral pre-treatments and different mask development strategies were studied to develop the quantitativemodels. These models were then applied to freeze-dried slices of broccoli to identify regions within indi-vidual florets which were rich in glucosinolates. The procedure demonstrates potential for the quantita-tive screening and localisation of total glucosinolates in broccoli using the 950–1650 nm wavelengthrange. These compounds were mainly located in the external part of florets.

� 2013 Elsevier Ltd. All rights reserved.

1. Introduction

Glucosinolates are a class of about 120 chemicals distributed inonly 16 plant families. These compounds are well-known for theircharacteristic pungent smells and tastes which are typical of someBrassica vegetables such as cabbage, mustard, cress, cauliflower,broccoli, turnip, Brussel sprouts, radish and horseradish. Structur-ally, glucosinolates (b-thioglucoside-N-hydroxysulfates) are char-acterised by the presence of nitrogen and sulphur groups.Biosynthesis of glucosinolates is mainly carried out using glucoseand amino acids such as methionine, alanine, leucine and valine(aliphatic glucosinolates) or tryptophan and phenylalanine (aro-matic glucosinolates) (Crozier et al., 2006). Epidemiological studieshave consistently reported a reduced incidence of a number of dis-eases in subjects consuming diets rich in these compoundsalthough anti-nutritive effects of both glucosinolates and theirhydrolysis products have also been reported (Crozier et al., 2006;Jeffery and Araya, 2009; Shahindi, 1990; Verkerk et al., 2009).

Broccoli (Brassica oleracea L. var Italica) contains significantamounts of these potentially bioactive compounds (Vallejo et al.,2003; Wang et al., 2012a). This vegetable is an economicallyimportant crop in a number of countries which may act as a sourcenot only of glucosinolates but also of vitamins, minerals and otherbeneficial phytochemicals (Jeffery et al., 2003; Wang et al., 2012a).It is therefore important to both characterise the content of bioac-tive compounds in broccoli and determine in what parts of theplant these bioactive compounds are accumulated. Our previouswork has demonstrated the potential for near infrared spectros-copy to quantify total glucosinolates in freeze-dried powders withacceptable accuracy for screening purposes (Hernandez-Hierroet al., 2012). Information about the spatial distribution of the afore-mentioned compounds would also be useful but near infraredspectroscopy does not provide the capability to map the locationof constituents. Hyperspectral imaging may hold the answer to thisproblem.

Hyperspectral imaging is an emerging technique for non-destructive food analysis which provides both spatial and spectralinformation about an object. Recorded images consist of manythousands of pixels in a two-dimensional array, with each pixel

108 J.M. Hernández-Hierro et al. / Journal of Food Engineering 126 (2014) 107–112

corresponding to a specific region on the surface of the sample;each pixel in a hyperspectral image therefore contains a spectrumof the sample at that specific position. Interrogation of thesespectra makes possible the development of mathematical modelsto predict the chemical composition or functional class of a sampleat each pixel. Reflectance imaging is the most common imageacquisition mode and is usually carried out in either thevisible-near infrared (vis–NIR; 400–1000 nm) or near infrared(NIR; 1000–1700 nm) spectral regions (Gowen et al., 2007). Theuse of multivariate chemometric methods is required to handlethe large quantities of spectral data collected in each image andvery many approaches are available for the development of regres-sion models to predict constituent concentrations in a sample atpixel level.

The number of research applications of hyperspectral analysishas risen considerably in the food sector in the recent past (Burgerand Gowen, 2011; Gowen et al., 2007; Lorente et al., 2012;McGoverin et al., 2010; Sun, 2010). Hyperspectral image analysishas been used to determine moisture, total soluble solids and pHin strawberries (ElMasry et al., 2007), firmness and soluble solids inapples (Mendoza et al., 2011; Wang et al., 2012b) anthocyaninsin grape skins (Fernandes et al., 2011), chlorophyll distribution incucumber leaves (Ji-Yong et al., 2012) and maturity stage ofbananas (Rajkumar et al., 2012). Additionally, this analyticalmethod has been used to determine moisture in dehydratedprawns (Wu et al., 2012) and some quality parameters of bothlamb (Kamruzzaman et al., 2012) and pork (Barbin et al., 2012)meat. Results reported in these studies have indicated that hyper-spectral imaging is able to predict a number of food componentsand quality parameters in a wide range of biological matrices.

The aim of this study was to evaluate the potential of hyper-spectral imaging technology for the quantitative screening andlocalisation of total glucosinolates in freeze-dried broccoli. Sincepredictive models developed on freeze-dried powders by conven-tional NIR spectrometers may not be transferred directly to hyper-spectral imaging datasets, a new predictive model must begenerated using an actual hyperspectral imaging system on homo-geneous, freeze-dried broccoli powders after which it may be ap-plied to hyperspectral images of intact broccoli for localisationand quantitation of total glucosinolates. To our knowledge, this isthe first time that this analytical tool has been applied to broccolifor these purposes.

2. Material and methods

2.1. Samples and chemical analysis

Sixty-four broccoli samples were grown at Teagasc (KinsealyResearch Centre, Dublin) using different agricultural managements(see (Hernandez-Hierro et al., 2012)), cultivars (Belstar and Fiesta)and years (2009 and 2010). Samples were harvested, immediatelyfrozen then freeze-dried. Once freeze-dried, samples were milled(Blixer 4, Robot Coupe, France), vacuum-packed in polypropylenebags and stored at �20 �C prior to analysis. (Hernandez-Hierroet al., 2012). Two aliquots were taken from each sample, one forthe micellar electrokinetic capillary chromatography (MEKC) anal-ysis and the remaining aliquot for hyperspectral imaging analysisin order to develop the quantitative model on powders. Separately,whole slices of broccoli were frozen and freeze-dried in order toapply the constructed model for glucosinolate localisation studies.

Total glucosinolates were expressed as the sum of the individ-ual glucosinolates previously determined by a MEKC-DAD methodand the results obtained in this earlier work were used as referencedata for the development of quantitative models (Hernandez-Hier-ro et al., 2012). Total glucosinolate contents in the broccoli samples

ranged from 4.74 to 15.10 lmol g�1 dry matter (DM) with meanand standard deviation values of 9.53 and 1.87 lmol g�1 DMrespectively.

2.2. Hyperspectral imaging analysis

Two different spectral regions using two different devices (HSIinstrument DV Optics Ltd, Padua, Italy) were used in this study.The main components of these systems, including the illuminationsource, diffuser, moving base, optics (mirror and lens), spectro-graph, camera and computer are shown in Fig. 1. System 1 consistedof a high performance CCD Basler A312f camera (580 � 580 pixels;Balser, Ahrensburg, Germany), a spectrograph (Specim V10E)(Spectral Imaging Ltd., Oulu, Finland) attached to the camera cover-ing the spectral range between 400 and 1000 nm (spectral resolu-tion of 5 nm), a zoom lens (16 mm focal length), a halogen lamp(150 W) light source transmitted via fibre optics and a cylindricaldiffuser in front of the fibre optic line light to produce a diffuse lightsource, a moving table and a computer 2 comprised a SUI GoodrichSU320M-1.7RT InGaAs camera (320 � 240 pixels; Sensors Unlim-ited, Inc., Princeton, NJ, USA), a spectrograph (Specim ImSpectorN17E; Spectral Imaging Ltd., Oulu, Finland) covering the spectralrange between 900 and 1700 nm (spectral resolution of 7 nm), fivehalogen lamps (3 � 50 W and 2 � 20 W) alternately mounted(50 W–20 W–50 W–20 W–50 W) as source light and a cylindricaldiffuser in front of the lamps to produce a diffuse light source, amoving table and a computer system.

A two point reflectance calibration was used. A white ceramictile, the reflectance (Rw) of which was calibrated against a tile ofcertified reflectance (Ceram Research Ltd., UK), was used as a whitereference while dark current was recorded by taking a measure-ment after covering the spectrograph lens with a cap and closingthe shutter. Corrected reflectance values (R) were calculated takinginto account the relationship between sample (S), white standard(W) and dark current (D) absolute signal intensities and the knownreflectance of the ceramic tile (Rw) using the following formula:

R ¼ ½ðS� DÞ=ðW� DÞ� � RW ð1Þ

Data were recorded in units of reflectance and saved in ENVIheader format using the instrument acquisition software (SpectralScanner; DV Optics, Padua, Italy). When using System 1, only spec-tral data in the 450–900 nm region was used in data analysis dueto reduced efficiency of the light source and CCD in wavelength re-gions outside this range. In the case of System 2, the spectral rangewas attenuated to 950–1650 nm for similar operational reasons.

2.3. Data processing and analysis

A flowchart of the data processing and analysis strategy em-ployed in this study is shown in Fig. 2. Data treatment and quanti-tative model development was carried out using Matlab (R2010b;The Math Works, Inc., USA). For each hyperspectral image, regionsof interest (ROIs) of approx. 3 cm diameter were selected using aninteractive selection tool available in the acquisition software (‘ROItool’) and 500 pixels were randomly-selected within each ROI.Spectral information within each individual pixel was used in sub-sequent data analysis. Spectral data were pre-treated using thestandard normal variate (SNV) transform to diminish the effectsof light scatter. Prior to quantitative model development, principalcomponent analysis (PCA) was applied to the data (32,000 spec-tra = 64 samples � 500 pixels) in order to identify and, if necessary,eliminate spectral outliers using the T2 Hotelling value (Esquerreet al., 2012; Hotelling, 1931). Finally, quantitative calibrationswere developed by partial least squares (PLS) regression using totalglucosinolates as the dependent (Y) variable and pixel spectra asthe independent (X) variables.

Fig. 1. Imaging system line drawing: illumination source, diffuser, moving base, optics (mirror and lens), spectrograph, camera and computer.

Fig. 2. Flow chart of the data processing and analysis strategy employed in thisstudy.

J.M. Hernández-Hierro et al. / Journal of Food Engineering 126 (2014) 107–112 109

2.4. Prediction map

Slices of whole freeze-dried broccoli were scanned in the NIRzone (950–1650 nm) to apply the previously constructed modeland identify the glucosinolate allocation. Prior to the quantitativeanalysis, a thresholding rule method was applied to the broccoliimages to isolate the broccoli from other parts of image. An image

was generated using the maximum reflectance value of each pixelspectrum in a raw image. A threshold of 0.45 reflectance units wasset by analysing the corresponding histogram and drawing a tenta-tive mask image in an iterative process. Pixels with the maximumreflectance values larger than the threshold were classified as broc-coli and the remaining pixels were considered as background. SNVwas applied to minimise the effects of scattering in the mask cre-ated and then the PLS model was applied.

3. Results and discussion

3.1. Spectra and model development

Fig. 3 shows the average and standard deviation (10 timesamplified) spectra of broccoli powders over the 450–900 nm(Fig. 3a) and 950–1650 nm (Fig. 3b) ranges. Standard deviationspectra have been multiplied by a factor of 10 for display reasons.A strong feature of the sample spectra was the absorbance patternin the 650–700 nm range. This arose from pigment which dis-played a green tinge, likely to be mainly chlorophyll (Fig. 3a). Amarked absorbance pattern in the 1400–1500 nm range wasclearly present in the spectra (Fig. 3b) and this seems likely to arisefrom first overtone absorbances of OAH functional groups. Onelikely cause of such absorbances is variations in moisture contentof the sample spectra; while all samples were lyophilised, the sen-sitivity of NIR spectroscopy to water will disclose even small, inter-sample moisture content variations. Absorbance by OAH groups instructural carbohydrates (e.g. cellulose) may, of course, also beimplicated.

In the case of the first hyperspectral device (450–900 nm), using6 principal components (98.7% of the variability in the spectraldata) and the T2 Hotelling value, 1382 spectra were identifiedand eliminated as spectral outliers. The remaining 30,618 spectra,which represented about 96% of the spectral population, were re-tained and used to perform the partial least squares regression.The optimum number of PLS terms used for the calibration (PLSterms = 8) was determined by cross-validation using 10 subsets.Validation errors were combined into a single figure, the standarderror of cross-validation (SECV), which was calculated to be1.80 lmol g�1 DM in this case.

Fig. 3. Average and standard deviation (10 times amplified) spectra of broccoli powders over the 450–900 nm (a) and 950–1650 nm (b) wavelength ranges.

110 J.M. Hernández-Hierro et al. / Journal of Food Engineering 126 (2014) 107–112

When using the second hyperspectral device (950–1650 nm),using 11 principal components (71.7% of the variability in the spec-tral data) and T2 Hotelling value, 1863 spectra were identified andeliminated as spectral outliers, approximately 30% more than forthe first device reported above. The remaining 30,137 spectra,which represented about 94% of the spectral population, were se-lected and used to perform the partial least squares regression.The optimum number of PLS terms used for the calibration (PLSterms = 7) was determined by cross-validation using 10 subsets.Validation errors were combined into a single figure, the standarderror of cross-validation (SECV), which had a value of1.75 lmol g�1 DM in this case.

Better results (SECV = 1.75, 7 PLS terms) were obtained usingthe 950–1650 nm wavelength range. This spectral region containslittle information about colour variations between samples, some-thing which would have been a significant feature of the shorterwavelength range models. Subsequent analyses were carried outusing this longer wavelength range. Fig. 4 shows the loading(Fig. 4a) and beta regression coefficient (Fig. 4b) plots of the PLSmodel developed for total glucosinolate prediction. It is clear thatspectral regions around 1100 nm and 1400 nm make the mostimportant contributions to the model loadings. In loading 1, threemain features are apparent – a minimum at 1110 nm which is op-posed to a broad peak centred around 1200 nm and a reportedsharp maximum at 1420 nm. Loading 2 consists simply of a slopingbaseline (from 1000 nm to a minimum around 1300 nm) and a sin-gle broad maximum centred around 1440 nm. Loading 3 has twominima at 1110 and 1410 nm while loading 4 exhibits a maximum

Fig. 4. PLS loadings (a) and beta regression coefficients (b) of the PLS model for total gluc(Supplementary Fig. 1).

at 1115 nm and a minimum around 1170 nm. In the case of load-ings 5, 6 and 7, the main features consist of minima around 1110and 1470 nm (loading 5), maxima around 1080, 1155 and1460 nm (loading 6) and small maxima at 1110 and 1200 nm(loading 7) respectively. Absorbance around 1110 nm may reflect2nd overtone AOH stretching although a combination AS@Ostretch has been previously reported around 1020–1060 nm (Stu-art, 2004) which may be relevant given the occurrence of sul-phur-containing volatile compounds in broccoli (Jacobsson et al.,2004). Absorbances at 1200 nm have been attributed to 2nd over-tone of CAH stretch vibrations in ACH3, ACH2 or ACH structures(Maalooly and Jaillais, 2006). Definitive attribution of absorbancesaround 1420–1440 nm is difficult given that absorptions aroundthis wavelength range originate from both AOH and ACH struc-tures. There is an undoubted difference between loading 1 and 2in this regard, since the relevant peak in loading 1 is sharp, very in-tense and has a maximum of 1420 nm while that in loading 2 isbroad, less intense and is centred around 1440 nm. Given the nar-row range of moisture variation between samples since they wereall lyophilised, it seems reasonable to attribute the 1440 nm peakto water absorbance (1st harmonic AOH stretch) and the1420 nm feature to absorbances arising from ACH moieties. Giventhe widespread occurrence of such bonds in biological structures, itis not possible to be more specific about the molecular species in-volved; however, structural polysaccharides comprising the bulk ofbroccoli composition are likely to be implicated.

It is hazardous to try to interpret regression coefficients at anysignificant level of detail (Brown and Green, 2009) but a number of

osinolate prediction. For separate plots of each PLS loading, see Supplementary data

Fig. 5. Flowchart of the process for performing the image segmentation procedure and glucosinolate prediction map in broccoli.

J.M. Hernández-Hierro et al. / Journal of Food Engineering 126 (2014) 107–112 111

features of the regression vector shown in Fig. 4b merit mention.Wavelengths with the largest regression coefficients are found at1130, 1366, 1460, 1480, 1520 and 1540 nm; the latter four are con-siderably larger than the first two. Coefficients associated with the1460 and 1480 nm wavelengths are in opposition to the others.The features at 1130 and around 1460–1480 nm are somewhatconsistent with the features observed in the individual loadingsdescribed above but the emergence of large regression coefficientsat 1520 and 1540 nm and a smaller negative coefficient at 1366 nmhave no parallel in the loadings vectors. Broadly speaking,however, the regression vector indicates the importance ofspectral regions involving ACH vibrations and others which arenon-identifiable in the prediction of glucosinolate content in broccoli.

3.2. Prediction map

Fig. 5 shows a flowchart of the process for performing the imagesegmentation procedure and the glucosionolate prediction map inbroccoli. The main feature of note in this figure is that thesecompounds were almost exclusively located in the external partof the broccoli florets. Predicted glucosinolate contents in the mainstem and even in peripheral structural features of the floret werevery low (<8 lmol g�1 DM); given the predictive accuracy of themodel used for this measurement, it was not possible to indicatewhether these molecular species were absent or not from suchstructures. Although their role in plants is unclear, their potentodour and pronounced taste suggests a role in herbivore andmicrobial defence. Deposition in external plant parts, confirmedspectroscopically in this work, would be the optimal location forthese purposes.

4. Conclusion

Two different spectral regions (vis–NIR and NIR) were studiedto develop quantitative models. Better results were obtained

using the 950–1650 nm wavelength range and subsequentanalyses were therefore carried out using this spectral zone.The procedure demonstrates potential for the quantitativescreening and location of total glucosinolates in broccoli usingthe 950–1650 nm wavelength range. Nevertheless, a comprehen-sive study should be made in order to evaluate all other relevantsources of variability in the complete development of thesemodels. It is well known that in the case of agricultural productsthe range of the variability should be as large as that expectedin any future samples. Likely causes of variation in glucosino-lates in broccoli include effects of the environment, changes infarm soil, growing season, cultivar, harvest conditions, diseaseincidence, etc. Consequently, all these factors should be takeninto account in the development of future robust global models.Such a study would entail several years work but the results re-ported herein suggest the viability of obtaining useful resultsfrom such an undertaking. One noteworthy observation arisingfrom this preliminary study is that these glucosinolate com-pounds seem to be mainly located in the external part of thevegetable florets.

Acknowledgements

Authors are grateful to Jens C. Sørensen and Hilmer Sørensenfrom the Department of Basic Sciences and Environment,Chemistry and Biochemistry, Faculty of Life Sciences, Universityof Copenhagen in Denmark for their useful help on setting up theMicellar Electrokinetic Capillary Chromatography method and forsharing their expert knowledge on the glucosinolate/myrosinasesystem. J.M. Hernández-Hierro thanks the Spanish MICINN forthe Juan de la Cierva contract (JCI-2011-09201) and Universidadde Sevilla for the mobility Grant (Universidad de Sevilla ResearchPlan). Spanish MICINN Project AGL2011-30254-C02 and Junta deAndalucia PGC Project AGR 6331.

112 J.M. Hernández-Hierro et al. / Journal of Food Engineering 126 (2014) 107–112

Appendix A. Supplementary material

Supplementary data associated with this article can be found, in theonline version, at http://dx.doi.org/10.1016/j.jfoodeng.2013.11.005.

References

Barbin, D.F., ElMasry, G., Sun, D.-W., Allen, P., 2012. Predicting quality and sensoryattributes of pork using near-infrared hyperspectral imaging. Analytica ChimicaActa 719, 30–42.

Brown, C.D., Green, R.L., 2009. Critical factors limiting the interpretation ofregression factors in multivariate calibration. Trends in Analytical Chemistry28 (4), 506–514.

Burger, J., Gowen, A.A., 2011. Data handling in hyperspectral image analysis.Chemometrics and Intelligent Laboratory Systems 108 (1), 13–22.

Crozier, A., Clifford, M.N., Ashihara, H., 2006. Plant Secondary Metabolites.Occurrence, Structure and Role in the Human Diet. Blackwell Publishing,Garsington Road, Oxford, UK.

ElMasry, G., Wang, N., ElSayed, A., Ngadi, M., 2007. Hyperspectral imaging fornondestructive determination of some quality attributes for strawberry. Journalof Food Engineering 81 (1), 98–107.

Esquerre, C., Gowen, A.A., Downey, G., O’ Donnell, C.P., 2012. Wavelength selectionfor development of a near infrared imaging system for early detection of bruisedamage in mushrooms (Agaricus bisporus). Journal of Near InfraredSpectroscopy 20, 537–546.

Fernandes, A.M., Oliveira, P., Moura, J.P., Oliveira, A.A., Falco, V., Correia, M.J., Melo-Pinto, P., 2011. Determination of anthocyanin concentration in whole grapeskins using hyperspectral imaging and adaptive boosting neural networks.Journal of Food Engineering 105 (2), 216–226.

Gowen, A.A., O’Donnell, C.P., Cullen, P.J., Downey, G., Frias, J.M., 2007. Hyperspectralimaging – an emerging process analytical tool for food quality and safetycontrol. Trends in Food Science & Technology 18 (12), 590–598.

Hernandez-Hierro, J.M., Valverde, J., Villacreces, S., Reilly, K., Gaffney, M., Gonzalez-Miret, M.L., Heredia, F.J., Downey, G., 2012. Feasibility study on the use ofvisible-near-infrared spectroscopy for the screening of individual and totalglucosinolate contents in broccoli. Journal of Agricultural and Food Chemistry60 (30), 7352–7358.

Hotelling, H., 1931. The generalization of student’s ratio. Annals of MathematicalStatistics 2 (3), 360–378.

Jacobsson, J., Nielsen, T., Sjöholm, I., 2004. Influence of temperature, modifiedatmosphere packaging, and heat treatment on aroma compounds in broccoli.Journal of Agricultural and Food Chemistry 52, 1607–1614.

Jeffery, E., Araya, M., 2009. Physiological effects of broccoli consumption.Phytochemistry Reviews 8 (1), 283–298.

Jeffery, E.H., Brown, A.F., Kurilich, A.C., Keck, A.S., Matusheski, N., Klein, B.P., Juvik,J.A., 2003. Variation in content of bioactive components in broccoli. Journal ofFood Composition and Analysis 16 (3), 323–330.

Ji-Yong, S., Xiao-Bo, Z., Jie-Wen, Z., Kai-Liang, W., Zheng-Wei, C., Xiao-Wei, H., De-Tao, Z., Holmes, M., 2012. Nondestructive diagnostics of nitrogen deficiency bycucumber leaf chlorophyll distribution map based on near infraredhyperspectral imaging. Scientia Horticulturae 138, 190–197.

Kamruzzaman, M., ElMasry, G., Sun, D.-W., Allen, P., 2012. Prediction of somequality attributes of lamb meat using near-infrared hyperspectral imaging andmultivariate analysis. Analytica Chimica Acta 714, 57–67.

Lorente, D., Aleixos, N., Gómez-Sanchis, J., Cubero, S., García-Navarrete, O., Blasco, J.,2012. Recent advances and applications of hyperspectral imaging for fruit andvegetable quality assessment. Food and Bioprocess Technology 5 (4), 1121–1142.

Maalooly, J., Jaillais, B., 2006. Glucides. In: La spectroscopie infrarouge et sesapplications analytiques. Lavoisier Paris, France, pp. 175–232.

McGoverin, C.M., Weeranantanaphan, J., Downey, G., Manley, M., 2010. Theapplication of near infrared spectroscopy to the measurement of bioactivecompounds in food commodities. Journal of Near Infrared Spectroscopy 18 (2),87–111.

Mendoza, F., Lu, R., Ariana, D., Cen, H., Bailey, B., 2011. Integrated spectral and imageanalysis of hyperspectral scattering data for prediction of apple fruit firmnessand soluble solids content. Postharvest Biology and Technology 62 (2), 149–160.

Rajkumar, P., Wang, N., ElMasry, G., Raghavan, G.S.V., Gariepy, Y., 2012. Studies onbanana fruit quality and maturity stages using hyperspectral imaging. Journal ofFood Engineering 108 (1), 194–200.

Shahindi, F., 1990. Canola and Rapeseed Production, Chemistry and ProcessingTechnology. Van Nostrand Reinhold, New York, USA.

Sun, D.W., 2010. Hyperspectral Imaging for Food Quality Analysis and Control.Elsevier Science & Technology, San Diego, CA, USA.

Stuart, B.H., 2004. Infrared Spectroscopy: Fundamentals and Applications. JohnWiley & Sons, New York, USA.

Vallejo, F., Tomás-Barberán, F.A., Benavente-García, A.G., García-Viguera, C., 2003.Total and individual glucosinolate contents in inflorescences of eight broccolicultivars grown under various climatic and fertilization conditions. Journal ofthe Science of Food and Agriculture 83, 307–313.

Verkerk, R., Schreiner, M., Krumbein, A., Ciska, E., Holst, B., Rowland, I., De Schrijver,R., Hansen, M., Gerhäuser, C., Mithen, R., Dekker, M., 2009. Glucosinolates inBrassica vegetables: the influence of the food supply chain on intake,bioavailability and human health. Molecular Nutrition Food Research 53 (S2),S219.

Wang, J., Gu, H., Yu, H., Zhao, Z., Sheng, X., Zhang, X., 2012a. Genotypic variation ofglucosinolates in broccoli (Brassica oleracea var. italica) florets from China. FoodChemistry 133 (3), 735–741.

Wang, S., Huang, M., Zhu, Q., 2012b. Model fusion for prediction of apple firmnessusing hyperspectral scattering image. Computers and Electronics in Agriculture80, 1–7.

Wu, D., Shi, H., Wang, S., He, Y., Bao, Y., Liu, K., 2012. Rapid prediction of moisturecontent of dehydrated prawns using online hyperspectral imaging system.Analytica Chimica Acta 726, 57–66.