Preliminary study on the use of near infrared hyperspectral imaging for quantitation and localisation of total glucosinolates in freeze-dried broccoli

Preliminary study on the use of near infrared hyperspectral imaging for quantitation and localisation of total glucosinolates in freeze - dried broccoli José Miguel Hernández - Hierro, Carlos Esquerre, Juan Valverde, Salvador Villacreces, Kim Reilly, Mich ael Gaffney, M. Lourdes González - Miret, Francisco J. Heredia, Colm P. O’Donnell, Gerard Downey This article is provided by the author(s) and Teagasc T - Stór in accordance with publish er policies. Please cite the published version. The correct citation is available in the T - Stór record for this article. This item is made available to you under the Creative Commons Attribution - Non commercial - No Derivatives 3 .0 License. NOTICE: This is the author’s version of a work that was accepted for publication in Journal of Food Engineering . Changes resulting f rom the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A defi nitive version was subsequently published in Journal of Food Engineering, Volume 126, April 2014, Pages 107 - 112. DOI 10.1016/j.jfoodeng.2013.11.005. Preliminary study on the use of near infrared hyperspectral imaging for quantitation and localisation of total glucosinolates in freeze - dried brocc oli José Miguel Hernández - Hierro 1 , Carlos Esquerre 2 , Juan Valverde 3 , Salvador Villacreces 3 , Kim Reilly 4 , Michael Gaffney 4 , M. Lourdes González - Miret 1 , Francisco J. Heredia 1 , Colm P. O’Donnell 2 , Gerard Downey 3,* 1 Food Colour & Quality Laboratory, Departmen t of Nutrition & Food Science, Universidad de Sevilla, Facultad de Farmacia, 41012 Sevilla, Spain. 2 Biosystems Engineering, School of Agriculture, Food Science and Veterinary Medicine , University College Dublin, Belfield, Dublin 4, Ireland. 3 Teagasc Food R esearch Centre Ashtown , Dublin 15, Dublin, Ireland. 4 Horticulture Development Unit, Teagasc Research Centre, Kinsealy, Dublin 17, Ireland. * Corresponding author: Gerard Downey Phone: (+353 1) 805 9500 Fax: (+353 1) 805 9550 E - mail: g erard. d owney@teagasc. ie Abstract 1 T he use of hyperspectral imaging to (a) quantify and (b) loca lise total glucosinolates in 2 florets of a single broccoli species has been examined . Two different spectral re gions 3 ( v is - NIR and NIR) , a number of spectral pre - treatments and differe nt mask development 4 strategies were studied to develop the quantitative models. These models were then 5 applied to freeze - dried slices of broccoli to identify regions within individual florets 6 which were rich in glucosinolates. The procedure demonstrates po tential for the 7 quantitative screening and loca lisa tion of total glucosinolates in broccoli using the 950 - 8 1650 nm wavelength range. T hese compounds were mainly located in the external part 9 of florets. 10 Keywords: Glucosinolates, broccoli, hyperspectral imagi ng , near infrared, visible, 11 chemometrics. 12 1. Introduction 13 Glucosinolates are a class of about 120 chemicals distributed in only 16 plant families. 14 The se compounds are well - known for their characteristic pungent smells and tastes 15 which are typical of some Brassica vegetables such as cabbage, mustard, cress, 16 cauliflower, broccoli, turnip, B russel sprouts, radish and horseradish. Structurally, 17 glucosinolates (β - thioglucoside - N - hydroxysulfates) are characterised by the presence of 18 nitrogen and sulphur groups . Biosynthesis of glucosinolates is mainly carried out using 19 glucose and amino acid s such as methionine, alanine, leucine and v aline (aliphatic 20 glucosinolates) or tryptophan and phenylalanine (aromatic glucosinolates) (Crozier et 21 al., 2006) . Epidemiological studies have consistently reported a reduc ed incidence of a 22 n umber of diseases in subjects consuming diets rich in these compounds although anti - 23 nutritive effects of both glucosinolates and their hydrolysis products have also been 24 reported (Crozier et al., 2006 ; Jeffery and Araya, 2009; Shahindi, 1990; Verkerk et al., 25 2009) . Broccoli ( Brassica oleracea L. var Italica ) contains significant amounts of these 26 potential ly bioactive compounds (Vallejo et al., 2003; Wang et al., 2012a) . This 27 vegetable is an economically important crop in a number of countries which may act as 28 a source not only of glucosinolates but also of vitamins, minerals and other beneficial 29 phytochemicals (Jeffery et al., 2003; Wang et al., 2012a) . It is therefore important to 30 both characterise the content of bioactive compounds in broccoli and determine in what 31 parts of the plant the se bioactive compounds are accumulated. Our previous work has 32 demonstrate d the potential for near infrared spectroscopy to quantify total 33 glucosinolates in freeze - dried powders with acceptable accuracy for screening purposes 34 (Hernandez - Hierro et al., 2012) . Information about the spatial distribution of the 35 aforementioned compounds would also be useful but near infrared spectroscopy do es 36 not provi de the capability to map the location of constituents . Hyperspectral imaging 37 may hold the answer to this problem . 38 Hyperspectral imaging is an emerging technique for non - destructive food analysis 39 which provides both spatial and spectral information about an object . Recorde d images 40 consist of many thousands of pixels in a two - dimensional array, with each pixel 41 corresponding to a specific region on the surface of the sample ; e ach pixel in a 42 hyperspectral image therefore contains a spectrum of the sample at tha t specific 43 position . Interrogation of these spectra makes possible the development of mathematical 44 models to predict the chemical composition or functional class of a sample at each 45 pixel. Reflectance imaging is the most common image acquisition mode and i s usually 46 carried out in either the visible - near infrared (vis - NIR ; 400 - 1000 nm) or near infrared 47 (NIR ; 1000 - 1700 nm) spectral regions (Gowen et al., 2007) . The use of multivariate 48 chemometric methods is required to handl e the large quantities of spectral data collected 49 in each image and very many approaches are available for the development of 50 regression models to predict constitu ent concentrations in a sample at pixel level. 51 The number of research applications of hyperspectral analysis has risen considerably in 52 the food sector in the recent past (Burger and Gowen, 2011; Gowen et al., 2007; 53 Lorente et al., 2012; McGoverin et al., 2010; Sun, 2010) . H yperspectral image analysis 54 has been used to determine moisture, total soluble solid s and pH in strawberries 55 (ElMasry et al., 2007) , firmness and soluble solids in apples (Mendoza et al., 2011; 56 Wang et al., 2012b) anthocyanins in grape skins (Fernandes et al., 2011) , chlorophyll 57 distribution in cucumber lea ves (Ji - Yong et al., 2012) and maturity stag e of bananas 58 (Rajkumar et al., 2012) . Additionally , this analytical method has been used to determine 59 moisture in dehydrated prawns (Wu et al., 2012) and some qua lity parameters of both 60 lamb (Kamruzzaman et al., 2012) and pork (Barbin et al., 2012) meat . R esults reported 61 in these studies have indicated that hyperspectral imaging is able to predict a number of 62 food components and q uality parameters in a wide range of biological matrices. 63 The aim of this study was to evaluate the potential of hyperspectral imaging technology 64 for the quantitative screening and loca lisa tion of total glucosinolates in freeze - dried 65 broccoli. Since predic tive models developed on freeze - dried powders by conventional 66 NIR spectrometers may not be transferred directly to hyperspectral imaging datasets, a 67 new predictive model must be generated using a n actual hyperspectral imaging system 68 on homogeneous, freeze - dried broccoli powders after which it may be applied to 69 hyperspectral images of intact broccoli for localisation and quantitation of total 70 glucosinolates. To our knowledge, this is the first time that this analytical tool has been 71 applied to broccoli for t hese purposes. 72 2. Material and methods 73 2.1. Samples and chemical analysis 74 Sixty - four broccoli samples were grown at Teagasc ( Kinsealy Research Centre , Dublin) 75 using different agricultural managements (see (Hernandez - Hierro et al., 2012) ) , 76 cultivars (Belstar and Fiesta) and years (2009 and 2010) . Samples were harvested , 77 imme diately frozen then freeze - dried . Once freeze - dried, samples were milled (Blixer 4, 78 Robot C oupe, France), vacuum - packed in polypropylene bags and stored at −20°C prior 79 to analysis. (Hernandez - Hierro et al., 2012) . Two aliquots were taken from each sample, 80 one for the micellar electrokinetic capillary chromatography (MEKC) an alysis and the 81 remaining aliquot for hyperspectral imaging analysis in order to develop the quantitative 82 model on powders . Separately, whole slices of broccoli were frozen and freeze - dried in 83 order to apply the constructed model for glucosi nolate loca lisa t ion studies. 84 Total glucosinolates were expressed as the sum of the individual glucosinolates 85 previously determined by a MEKC - DAD method and the results obtained in this earlier 86 work were used as reference data for the development of quantitative models 87 (Hernandez - Hierro et al., 2012) . T otal glucosinolate contents in the broc coli samples 88 ranged from 4.74 to 15.10  mol g - 1 dry matter (DM) with m ean value and standard 89 deviation values of 9.53 and 1.87  mol g - 1 DM respectively. 90 2. 2 . Hyperspectral imaging analysis 91 Two different spectral regions using two different devices (HSI ins trument DV Optics 92 Ltd, Padua, Italy) were used in this study . The main components of these systems, 93 including the illumination source, diffuser, moving base, optics (mirror and lens) , 94 spectrograph , camera and computer are shown in Fig . 1 . System 1 consiste d of a high 95 performance CCD Basler A312f camera (580x580 pixels ; Balser, Ahrensburg, 96 Germany) , a spectrograph (Specim V10E) (Spectral Imaging Ltd., Oulu, Finland) 97 attached to the camera covering the spectral range between 400 and 1000 nm (spectral 98 resoluti on of 5 nm) , a zoom lens (16 mm focal length) , a halogen lamp (150W) light 99 source transmitted via fibre optics and a cylindrical diffuser in front of the fibr e optic 100 line light to produce a diffuse light source , a moving table and a computer system. 101 Syste m 2 comprised a SUI Goodrich SU320M - 1.7RT InGaAs camera (320 × 240 102 pixel s ; Sensors Unlimited, Inc . , Princeton, NJ, USA) , a spectrograph ( Specim 103 ImSpector N17E ; Spectral Imaging Ltd., Oulu, Finland) covering the spectral range 104 between 900 and 1700 nm (spect ral resolution of 7 nm) , five halogen lamp s ( 3 x 50W 105 and 2 x 20W ) alternately mounted (50 W - 20 W - 50 W - 20 W - 50 W ) as source light and a 106 cylindrical diffuser in front of the lamps to produce a diffuse light source, a moving 107 table and a computer system. 108 A two poin t reflectance calibration was used. A white ceramic tile , the reflectance (R w ) 109 of which was calibrated against a tile of certified reflectance (Ceram Research Ltd, 110 UK) , was used as a white reference while dark current was recorded by taking a 111 measurement a fter covering the spectrograph lens with a cup and closing the shutter. 112 C orrected reflectance value s ( R ) w ere calculated taking into account the relation ship 113 between sample ( S ) , white standard ( W ) and dark current ( D ) absolute signal intensities 114 and the kn own reflectance of the ceramic tile (R w ) using the following formula: 115 R= [ (S - D)/(W - D) ]·R W (1) 116 Data were recorded in units of reflectance and saved in ENVI header format using the 117 instrument acquisition software ( Spectral Scanner ; DV Optics, Padua, Italy) . When 118 using S ystem 1 , only spectral data in the 450 – 90 0 nm regions were used in data 119 analysis due to reduced efficiency of the light source and CCD in wavelength regions 120 outside this range . In the case of System 2 , the spectral range was attenuated to 95 0 - 121 1650 nm for similar operational reasons . 122 2.3. Data processing and analysis 123 A flowchart of the data processing and analysis strategy employed in this study is 124 shown in Fig . 2 . Data treatment and quantitative model development was carried out 125 using M atla b ( R2010b ; The Math Works, Inc. USA). For each hyperspectral image, 126 regions of interest (ROI s ) of approx. 3 cm diameter were selected using an interactive 127 selection tool available in the acquisition software (‘ROI tool’) and 500 pixels were 128 randomly - select ed within each ROI. S pectral information within each individual pixel 129 was used in subsequent data analysis. Spectral data were pre - treated using the s tandard 130 n ormal v ariate (SNV) transform to diminish the effects of light scatter. Prior to 131 quantitative mod el development , principal component analysis (PCA) was applied to 132 the data (32000 spectra = 64 samples x 500 pixels) in order to identify and, if necessary, 133 eliminate spectral outliers using the T 2 Hotelling value (Esquerre et al., 2012; Hotelling, 134 1931) . Finally , quantitative calibrations were developed by partial least squares (PLS) 135 regression using total glucosinol ates as the dependent (Y) variable and pixel spectra as 136 the independent (X) variables . 137 2.4 Prediction map 138 Slices of whole freeze - dried broccoli were scanned in the NIR zone (950 - 1650 nm) to 139 apply the previously constructed model and identify the glucosinolate allocation. Prior 140 to the quantitative analysis, a thresholding rule method was applied to the broccoli 141 images to isolate the broccoli from other parts of image. An image was generated using 142 the maximum reflectance value of each pixel spectrum in a raw image. A threshold of 143 0.45 reflectance units was set analysing the corresponding histogram and drawing a 144 tentative mask image in an iterative process . Pixels with the maximum reflectance 145 values larger than the threshold were classified as broccoli and the remaining pixels 146 were considered as background. SNV was applied to minimise the effects of scattering 147 in the mask created and then the PLS model was applied. 148 3. Results and discussion 149 3.1. Spectra and m odel development 150 Fig. 3 shows the average and standard deviation (10 times amplified) spectra of broccoli 151 powders over the 450 - 900 nm ( Fig. 3 a ) and 950 - 1650 nm ( Fig. 3 b ) ranges. Standard 152 deviation spectra have been multiplied by a factor of 10 for display reasons. A strong 153 feature of the sample spectra was the absorbance pattern in the 650 - 700 nm range. This 154 arose from pigment which displayed a green tinge , likely to be mainly chlorophyll ( Fig. 155 3 a ) . A mark ed absorbance pattern in the 1400 - 1500 nm range was clearly present in the 156 spectra ( Fig. 3 b ) and this seems likely to arise from first overtone absorbance s of O - H 157 functional group s. One likely cause of such absorbances is variations in moisture 158 content of the sample spectra ; w hile all samples were lyophilised, the sensitivity of NIR 159 spectroscopy to water will disclose even small, inter - sample moisture content 160 Jeffery, E.H., Brown, A.F., Kurilich, A.C., Keck, A.S., Matusheski, N., Klein, B.P., 301 Juvik, J.A., (2003). Variation in content of bioactive components in broccoli. Journal of 302 Food Composition and Analy sis 16(3), 323 - 330. 303 Ji - Yong, S., Xiao - Bo, Z., Jie - Wen, Z., Kai - Liang, W., Zheng - Wei, C., Xiao - Wei, H., 304 De - Tao, Z., Holmes, M., (2012). Nondestructive diagnostics of nitrogen deficiency by 305 cucumber leaf chlorophyll distribution map based on near infrared hy perspectral 306 imaging. Scientia Horticulturae 138, 190 - 197. 307 Kamruzzaman, M., ElMasry, G., Sun, D. - W., Allen, P., (2012). Prediction of some 308 quality attributes of lamb meat using near - infrared hyperspectral imaging and 309 multivariate analysis. Analytica Chimica Acta 714, 57 - 67. 310 Lorente, D., Aleixos, N., Gómez - Sanchis, J., Cubero, S., García - Navarrete, O., Blasco, 311 J., (2012). Recent Advances and Applications of Hyperspectral Imaging for Fruit and 312 Vegetable Quality Assessment. Food and Bioprocess Technology 5(4), 1121 - 1142. 313 Maalooly, J . , Jaillais, B. (2006). Glucides. In: La spectroscopie i nfrarouge et ses 314 a pplications analytiques. (pp 175 - 232). Lavoisier Paris, France. 315 McGoverin, C.M., Weeranantanaphan, J., Downey, G., Manley, M., (2010). The 316 application of near i nfrared spectroscopy to the measurement of bioactive compounds in 317 food commodities. Journal of near Infrared Spectroscopy 18(2), 87 - 111. 318 Mendoza, F., Lu, R., Ariana, D., Cen, H., Bailey, B., (2011). Integrated spectral and 319 image analysis of hyperspectral s cattering data for prediction of apple fruit firmness and 320 soluble solids content. Postharvest Biology and Technology 62(2), 149 - 160. 321 Rajkumar, P., Wang, N., E lM asry, G., Raghavan, G.S.V., Gariepy, Y., (2012). Studies 322 on banana fruit quality and maturity st ages using hyperspectral imaging. Journal of 323 Food Engineering 108(1), 194 - 200. 324 Shahindi, F., (1990). Canola and rapeseed production, chemistry and processing 325 technology . Van Nostrand Reinhold, New York, USA. 326 Sun, D.W., (2010). Hyperspectral Imaging for Foo d Quality Analysis and Control . 327 Elsevier Science & Technology, San Diego, CA, USA. 328 Stuart, B.H. (2004). Infrared Spectroscopy: Fundamentals and Applications. John 329 Wiley & Sons, New York, USA (2004). 330 Vallejo, F., Tomás - Barberán, F.A., Benavente - García, A.G. , García - Viguera, C., 331 (2003). Total and individual glucosinolate contents in inflorescences of eight broccoli 332 cultivars grown under various climatic and fertilization conditions. Journal of the 333 Science of Food and Agriculture 83, 307 – 313. 334 Verkerk, R., Schr einer, M., Krumbein, A., Ciska, E., Holst, B., Rowland, I., De 335 Schrijver, R., Hansen, M., Gerhäuser, C., Mithen, R., Dekker, M., (2009). 336 Glucosinolates in Brassica vegetables: The influence of the food supply chain on intake, 337 bioavailability and human heal th. Molecular Nutrition Food Research 53(S2), S219 - 338 S219. 339 Wang, J., Gu, H., Yu, H., Zhao, Z., Sheng, X., Zhang, X., (2012a). Genotypic variation 340 of glucosinolates in broccoli ( Brassica oleracea var. italica ) florets from China. Food 341 Chemistry 133(3), 735 - 74 1. 342 Wang, S., Huang, M., Zhu, Q., (2012b). Model fusion for prediction of apple firmness 343 using hyperspectral scattering image. Computers and Electronics in Agriculture 80, 1 - 7. 344 Wu, D., Shi, H., Wang, S., He, Y., Bao, Y., Liu, K., (2012). Rapid prediction of 345 moisture content of dehydrated prawns using online hyperspectral imaging system. 346 Analytica Chimica Acta 726, 57 - 66. 347 348 349 Figure captions 350 Fig . 1. Imaging system line drawing: illumination source, diffuser, moving base, optics 351 (mirror and lens) , spectrograph , camera and computer. 352 Fig. 2 . Flow chart of the data processing and analysis strategy employed in this study. 353 Fig . 3 . Average and standard deviation (10 times amplified) spectra of broccoli powders 354 over the 450 - 900 nm (a) and 950 - 1650 nm (b) wavelength ran ges. 355 Fig 4 . PLS l oading s (a) and beta regression coefficients (b) of the PLS model for total 356 glucosinolate prediction . For separate plots of each PLS loading , see 357 Supplementary data ( Supplementary Fig. 1. ). 358 Fig 5 . Flowchart of the process for performing th e image segmentation procedure and 359 glucosinolate prediction map in broccoli. 360 361 Fig. 1 Fig. 2 Modelling strategy Read image Select ROI Select 500 pixels randomly SNV spectral pre - treatment PCA 32000 spectra matrix (64 samples x 500 pixels ) Spectral matrix Reference values vector ( total glucosinolates ) + PLS 1 model Eliminate spectral outliers using the T 2 Hotelling value Repeat for each sample (n=64) Fig. 3 Fig . 4 Fig. 5 Maximum reflectance values R> 0.45 broccoli R< 0.45 background 950 nm 1650 nm 950 nm 1650 nm Threshold 0.45 reflectance units (1) Masking (2) SNV pre - treatment (3) PLS model regression coefficients Mask Original hypercube