1 Prediction of beef chemical composition by NI Hyperspectral imaging 3 4 5 6 7 8 9 10 11 1 13 14 15 16 17 18 19 0 1 3 Gamal ElMasry a ; Da-Wen Sun a and Paul Allen b a Food efrigeration and Computerised Food Technology (FCFT), School of Biosystems Engineering, University College Dublin, Belfield, Dublin 4, Ireland. b Ashtown Food esearch Centre, Teagasc, Dublin 15, Ireland. Corresponding author. E-mail: gamal.elmasry@ucd.ie Abstract Developing a non-destructive method of food safety and quality monitoring became an essential request from meat product industry. Hyperspectral imaging technique provides extraordinary advantages over the traditional imaging and spectroscopy techniques in food quality evaluation due to the spatial and spectral information it offers. In this study, a pushbroom hyperspectral imaging system in the reflectance mode was developed in the near infrared range (900-1700 nm) for a rapid determination of the major chemical compositions of beef. Beef samples were collected from different breeds and then scanned by the system followed by traditional assessment of their chemical composition by using the ordinary wet-chemical methods. The extracted spectral data and the measured quality parameters were modelled by partial least squares (PLS) regression for predicting such parameters. The results indicated that hyperspectral imaging technique had a great potential for rapid and non-destructive assessment of meat quality traits. Keywords: Hyperspectral imaging, beef, meat, multivariate analysis, quality, PLS. 4 5 6 7 8 9 30 31 3 33 1. Introduction Hyperspectral imaging, also known as chemical or spectroscopic imaging, is an emerging technology that integrates conventional imaging and spectroscopy to attain both spatial and spectral information from an object -which otherwise cannot be achieved with either conventional imaging or spectroscopic techniques (ElMasry & Sun, 010). Therefore, the man aim of this study was to develop a NI hyperspectral imaging system (900-1700 nm) for objective prediction of the major chemical composition (protein, water and fat) of intact fresh beef. Indeed, the avoidance of using extraneous, harsh, toxic, corrosive and/or expensive chemicals during food compositional analyses would be accompanied not only with cost
34 35 36 37 38 39 40 41 4 43 44 45 46 47 48 49 50 51 5 53 54 55 56 57 58 59 60 61 6 63 64 65 66 67 68 69 savings but also with the avoidance of disposal problems resulting in several environmental benefits. Moreover, the direct and simultaneous measurement of several constituents without tedious sample preparation adds another merit of this expeditious technique to improve the production capabilities of a process line and manufacturing. Although there are a number of NIS analyzers are commercially available (Prieto et al., 009) these instruments are usually confined to point analysis where spatial information is neglected. If the sample itself was inhomogeneous, the average value might be misleading to represent the bulk sample. For detailed food analysis, concentration gradients of certain chemical constituents are often more attractive than average concentrations (Wold et al., 006). Similarly, digital imaging technique provides only spatial information of the object being analyzed without any consideration to its spectral fingerprints leading to neglecting chemical background of such object. On the other side, hyperspectral imaging systems stand with a superior capability of acquiring hundreds of spectra with high spectral and spatial resolution for providing spatial and spectral information of each pixel over certain wavelength range. This capability results in building chemical images to show the spatial distribution of the major chemical constituents of the tested object. Building such chemical images requires transferring the multivariate prediction models to each pixel of the hyperspectral image (Burger and Geladi, 006). These chemical images are normally created to visualize quantitative spatial distribution of food sample components and their relative concentrations. Since the quantitative assessment of chemical attributes is quite important, the distribution of such attributes is extremely significant to fully characterizing food products. In this sense, hyperspectral imaging technique has been used successfully proposed and undergone various degrees of development and application in chemical composition distribution on some meat products (Wold, et al., 006; ElMasry and Wold, 008; Ottestad, et al., 009). ecently, this technique have been exploited for non-destructive prediction and visualization of different meat quality attributes such as water holding capacity, ph, colour and tenderness in lamb, pork and beef (ElMasry et al., 011; Kamruzzaman et al., 01; Barbin et al., 01). This study aimed to investigate the development of hyperspectral imaging system for the non-contact assessment of the major constituents (water, fat and protein contents) in fresh beef and provide their distribution over the beef samples.. Materials and Methods.1. Beef Samples and spectral data extraction A total of 7 bulls from three different breeds (Holstein-Friesian, Jersey Holstein-Friesian and Norwegian ed Holstein-Friesian) were slaughtered at a commercial slaughterhouse
70 71 7 73 74 75 76 77 78 79 80 81 8 83 84 85 86 87 88 89 90 91 9 93 94 95 96 97 98 99 100 101 10 (Meadow Meats, athdowney, Co. Laois, Ireland). At 4 h post-mortem, three muscles (M. longissimus dorsi (LD), M. semitendinosus (ST) and M. psoas major (PM)) were dissected from each carcass and then sliced to 1-inch thick slices by a mechanical slicer. The slices were labelled and vacuum packed and stored at 4 C until the next day when quality parameters were measured. Moisture and fat contents were analysed using the Smart Trac (CEM Corporation, North Carolina, USA), and protein content was measured using a LECO FP-48 Nitrogen Determinator (LECO Instruments Ltd., UK). After chemical measurements, the samples were minced and imaged again in the hyperspectral system. The configuration of the developed hyperspectral imaging system is explained in details in ElMasry et al. (011). In practise, beef sample was placed on the conveying stage to be scanned line by line using 10 ms exposure time to build a hyperspectral image ( 0 ) which is then corrected against dark and white references. The average spectrum of each sample was extracted by segmentation to locate the lean parts of the beef sample as the main region of interest (OI). Only one average spectrum was used to represent each sample and the same routine was repeated for all hyperspectral images of beef samples. Predictive partial least squares (PLS) regression models were developed for each constituent so that these attributes can be predicted in the future directly from the measured spectra. The accuracy and the predictive capabilities of the model were evaluated based on coefficient of determination in calibration ( determination in cross-validation ( and the root mean square error by cross-validation (MSECV). 3. esults and Discussion 3.1. Spectral features C ), coefficient of CV ), the root mean square error of calibration (MSEC) The hyperspectral image described as I(x,y,λ) can be viewed either as a separate spatial sub-image I(x,y) at each wavelength (λ) as shown in Figure 1a, or as a spectrum I(λ) at any pixel (x,y) as shown in Figure 1b. The peak at 974 nm was apparent in the beef sample due to the water absorption bands related to O H stretching second overtones, whereas a second peak at 111 nm was due to the fat absorption related to C H stretching second overtone. The performance of the developed PLS regression models for predicting the major chemical composition of beef samples in calibration and leave-one-out cross-validation was reasonably good.
103 104 105 106 107 108 109 110 111 11 113 114 115 116 117 118 119 10 11 1 13 14 15 16 17 FIGUE 1 (a) Assortment of some NI sub-images at wavelengths indicated, (b) Spectral 3.. Prediction of chemical constituents features of the examined muscles. Table 1 shows the main statistical parameters resulted from each model in cross validation and prediction samples. The results indicated that the PLS models were very efficient in predicting these chemical constituents with a determination coefficient ( cv ) of 0.91, 0.89 and 0.91 using 9, 8, 10 factors for fat, moisture and protein respectively. When the model used in predicting these constituents in the samples of a prediction set, the model gave good prediction ability with determination coefficient ( p ) of 0.89, 0.86 and 0.75 for water, fat and protein, respectively. However, the predictability of protein content might be considered uncertain as the value of P (0.75) is rather small and the number of latent factors was higher compared with the other two constituents. The lack of a strong predictability for protein in meat suing spectral analyses is not new, and several authors have reported similar results (Alomar 003). It was very advantageous to select significant variables during a multivariate regression in order to improve the predictive ability of the model and to augment the processing speed. These feature-related wavelengths identified from the PLS s weighted regression coefficients of each constituent indicated that eight wavelengths (934, 1048, 1108, 1155, 1185, 11, 165, 1379 nm) were defined as the most relevant wavelengths in predicting water contents. Similarly, seven wavelengths (934, 978, 1078, 1138, 115, 189, 1413 nm) were selected for fat and ten wavelengths (94, 937, 1018, 1048, 1108, 1141, 118, 11, 1615, and 1665 nm) were selected for efficient prediction of protein. The results shown in Table 1 revealed that the predictability of these models is still good indicating the robustness
18 19 130 131 13 133 134 135 of the developed models. Water, fat and protein were predicted with determination coefficient ( p ) of 0.89, 0.84 and 0.86 with standard error of prediction (SEP) of 0.46, 0.65 and 0.9%, respectively. More importantly, the protein model was much better compared with that developed with the full spectral range. TABLE 1 Performance of the developed PLS model in predicting water, fat and protein contents in beef samples using the full spectral range and the selected feature-related wavelengths. No of Cross-validation Prediction Constituent wavelength s LFs cv SECV(%) p SEP(%) Water Fat Protein (1) 37 9 0.91 0.48 0.89 0.47 () 8 6 0.89 0.51 0.89 0.46 (1) 37 8 0.89 0.65 0.86 0.6 () 7 6 0.88 0.66 0.84 0.65 (1) 37 10 0.91 0.7 0.75 0.39 () 10 9 0.88 0.31 0.86 0.9 136 137 138 139 140 141 14 143 144 145 146 147 148 149 150 151 15 (1) Models developed using the full spectral range (37 wavelengths) () Models developed using the selected feature-related wavelengths (8, 7 and 10 wavelengths for water, fat and protein content, respectively). Acknowledgements Funding by the Irish Department of Agriculture, Fisheries and Food through Food Institutional esearch Measure (FIM) strategic research initiative is acknowledged. 4. eferences ElMasry, G., Sun, D.-W., Allen, P. (011). Non-destructive determination of water-holding capacity in fresh beef by using NI hyperspectral imaging. Food esearch International, 44 (9): 64-633. Prieto, N., oehe,., Lavín, P., Batten, G. & Andrés, S. (009). Application of near infrared reflectance spectroscopy to predict meat and meat products quality: A review. Meat Science, 83 (), 175-186. Wold, J. P., Johansen, Ib-., Haugholt, K, H., Tschudi, J., Thielemann, J., Segtnan, V, H.,Narum, B. & Wold, E. (006). Non-destructive tranreflectance near infrared
153 154 155 156 157 158 159 160 161 16 163 164 165 166 167 168 169 170 171 17 173 174 175 imaging for representative on-line sampling of dried salted coalfish (bacalao). Journal of Near Infrared Spectroscopy, 14, 59-66. Burger, J. & Geladi, P. (006). Hyperspectral NI imaging for calibration and prediction, a comparison between image and spectrometer data for studying organic and biological samples. The Analyst, 131, 115-1160. ElMasry, G., and Wold, J. P. (008). High-Speed Assessment of Fat and Water Content Distribution in Fish Fillets Using Online Imaging Spectroscopy. Journal of Agricultural and Food Chemistry, 56(17), 767-7677. Ottestad, S., Høy, M., Stevik, A., and Wold, J.P. (009). Prediction of ice fraction and fat content in super-chilled salmon by non-contact interactance near infrared imaging. Journal of Near-Infrared Spectroscopy, 17(), 77-87. Kamruzzaman, M., ElMasry, G., Sun, D-W, Allen, P. (01). Prediction of some quality attributes of lamb meat using NI hyperspectral imaging and multivariate analysis. Analytica Chimica Acta, 714 (10): 57-67. Barbin D., ElMasry G., Sun, D.-W. & Allen P. (01). Predicting quality and sensory attributes using near-infrared hyperspectral imaging. Analytica Chimica Acta, 714 (10): 57-67. Alomar, D., Gallo, C., Castañeda, M. & Fuchslocher,. (003). Chemical and discriminant analysis of bovine meat by near infrared reflectance spectroscopy (NIS). Meat Science, 63, 441-450. ElMasry, G. & Sun, D-W. (010). Meat Quality Assessment Using a Hyperspectral Imaging System. In: Hyperspectral Imaging for Food Quality Analysis and Control (Edited by Da-Wen Sun), PP. 73-94, Academic Press / Elsevier, San Diego, California, USA.