gg a

ngzhentr

Hyperspectral imaging

erspcesnceroscartMS

regression (MLR) were compared. The optimal RS-MLR model was considered to be the best for determin-ing the moisture content of salted pork, with a R2 of 0.917 and RMSEP of 1.48%. Visualisation of moisture

perationdersying orelatess. In

ture content (MC) is an important chemical parameter affecting

ucts, MC is closely related to the sensory, textural and microbiolog-ical quality attributes. Several studies have showed that there is ahigh relationship between texture parameters with moisture con-tents in dry-cured meat (Ruiz-Ramirez, Arnau, Serra, & Gou, 2005;Serra, Ruiz-Ramirez, Arnau, & Gou, 2005). Moreover, moisture is

due to its fast measurement, simple sample preparation and& Andres, 2009;1). NIR spmples rela

the vibration behaviour of molecular bonds such as CAHand NAH (Ghosh & Jayas, 2009), caused by their interactioelectromagnetic radiation absorbed at wavelengths betweand 2500 nm. The capability of using NIR spectroscopy for mois-ture analysis is mainly due to the high absorbance of the NIR radi-ation by water (Osborne, 2006). Many studies have conrmed theability of NIR spectroscopy to predict moisture content of rawmeat(Barlocco, Vadell, Ballesteros, Galietta, & Cozzolino, 2006; Kestens,Charoud-Got, Bau, Bernreuther, & Emteborg, 2008), as well as porkproducts such as fermented sausages (Collell, Gou, Picouet,Arnau, & Comaposada, 2010; Gaitan-Jurado, Ortiz-Somovilla,

Corresponding author at: College of Light Industry and Food Sciences, SouthChina University of Technology, Guangzhou 510641, PR China. Tel.: +353 17167342; fax: +353 1 7167493.

E-mail address: dawen.sun@ucd.ie (D.-W. Sun).

Food Chemistry 152 (2014) 197204

Contents lists availab

Food Che

lseURLs: http://www.ucd.ie/refrig, http://www.ucd.ie/sun (D.-W. Sun).processing and storage of agri-food products (Cui, Xu, & Sun,2004; Delgado & Sun, 2003; Sun, 1999; Sun & Byrne, 1998; Sun& Woods, 1993, 1994a, 1994b, 1994c, 1997). For salted meat prod-

non-invasiveness (Prieto, Roehe, Lavin, Batten,Weeranantanaphan, Downey, Allen, & Sun, 201provides complex structural information of sa0308-8146/$ - see front matter 2013 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.foodchem.2013.11.107ectrumted to, OAHn withen 700products with high quality, it is important to maintain strictcontrols during the salting process. Therefore, an in-depth studyof rapid detection method for quality evaluation of salted meatproducts is a necessity for process improvement and quality con-trol technology development. Decient controls during the saltingprocess could result in nutrition loss and texture problems. Mois-

ture analyser (Sleagun & Popa, 2009). However, these analyticalmethods are time-consuming and difcult to perform on-line ina production setting. Furthermore, these methods are not suitablefor monitoring the moisture contents continuously or non-destruc-tively. Near infrared (NIR) spectroscopy has been widely used forthe estimation of quality attributes in meat and meat productsMoisture contentPorcine meatSalting processSpectroscopic transformation

1. Introduction

Salting treatment is a common ohigh quality meat products, which rehigher long-term stability, by delagrowth. The quality of salted meat issition and the manufacturing procec

distribution in each pixel of the hyperspectral image using the prediction model display moisture evolu-tion and migration in pork slices.

2013 Elsevier Ltd. All rights reserved.

n in the production ofmeat products of muchr preventing microbiald to both their compo-order to produce meat

also a key factor for inhibiting the growth of foodborne pathogensand spoilage bacteria (Mathlouthi, 2001), inuencing the shelf lifeof salted meat products. Therefore, it is necessary and useful torapidly and accurately monitor moisture contents in order to con-trol the salting process and the quality of salted products.

Common methods for moisture analysis include oven-drying(AOAC., 1997) and microwave drying methods, and infrared mois-Keywords:Non-destructive

c c

lengths identied using the regression coefcient, two calibration models of PLSR and multiple linearFeasibility of using hyperspectral imaginof porcine meat during salting process

Dan Liu a, Da-Wen Sun a,b,, Jiahuan Qu a, Xin-An ZenaCollege of Light Industry and Food Sciences, South China University of Technology, Guab Food Refrigeration and Computerised Food Technology, Agriculture and Food Science C

a r t i c l e i n f o

Article history:Received 12 August 2013Received in revised form 21 October 2013Accepted 19 November 2013Available online 27 November 2013

a b s t r a c t

The feasibility of using hyp(MC) during the salting protance spectra (RS), absorbagate the inuence of spectThe best full-wavelength ptance spectra (R2 = 0.969, R

journal homepage: www.eto predict moisture content

, Hongbin Pu a, Ji Ma a

ou 510641, PR Chinae, University College Dublin, National University of Ireland, Beleld, Dublin 4, Ireland

ectral imaging technique (10002500 nm) for predicting moisture contents of porcine meat was assessed. Different spectral proles including reec-spectra (AS) and KubelkaMunk spectra (KMS) were examined to investi-opic transformations on predicting moisture content of salted pork slice.ial least squares regression (PLSR) models were acquired based on reec-EC = 0.921%; R2 = 0.941, RMSEP = 1.23%). On the basis of the optimal wave-

le at ScienceDirect

mistry

vier .com/locate / foodchem

also required. In the meat industry, during the meat saltingprocess, it is important to guarantee a controlled concentration

istrand homogeneous distribution of moisture in nal salted meatproducts, therefore knowledge about spatial distribution of mois-ture content within the specimen is required. However, it is noteasy to obtain moisture content at different spots within the spec-imen, by either conventional methods or spectroscopy.

Recently, a new technique referred to as hyperspectral imaging(HSI) has been developed (Sun, 2010). By integrating both spectro-scopic and imaging techniques (Du & Sun, 2005; Zheng, Sun, &Zheng, 2006a, 2006b) in one system, hyperspectral imaging cangenerate a spatial map of spectral variation, displaying spatial dis-tribution of inherent chemical and physical properties of the speci-men (Lorente, Aleixos, & Gomez-Sanchis, 2012). Applications of thistechnique in the meat industry have recently been reviewed(Elmasry, Barbin, Sun, & Allen, 2012), revealing that HSI has thepotential of predicting different attributes of meat quality quicklyand accurately such as beef (ElMasry, Sun, & Allen, 2011; Wuet al., 2012), pork (Barbin, ElMasry, Sun, & Allen, 2013; Tao, Peng,Li, Chao, & Dhakal, 2012), lamb (Kamruzzama, ElMasry, Sun, & Allen,2011, 2012), ham (Iqbal, Sun, &Allen, 2013), andsh (He,Wu, & Sun,2012; Zhu, Zhang, He, Liu, & Sun, 2012). Particularly for moisturecontent evaluation, Barbin et al. (2013) applied a pushbroomhyper-spectral imaging system to determine the moisture of intact andminced pork. Wu et al. (2012) applied the hyperspectral imagingtechnique to determine themoisture content of beef slices at differ-ent periods of the dehydration process. Recently, HSI has also beenused to determine the moisture content in salmon llet (He et al.,2012). Besides, several other authors have also explored the feasibil-ity of the HSI technique in moisture prediction of strawberries(ElMasry, Wang, ElSayed, & Ngadi, 2007), banana (Rajkumar, Wang,Elmasry, Raghavan, & Gariepy, 2012) and mushroom (Taghizadeh,Gowen, & ODonnell, 2009). To the best of our knowledge, noresearch endeavourers have been reported yet for determining themoisture distribution in porcine meat during the salting processusing hyperspectral imaging. Therefore, it is of our interest to imple-ment the HSI technique to analyse the moisture evolution andmigration in meat products during the salting process.

Therefore, the specic objectives of the current study were to(1) establish a satisfactory approach to extract spectral data fromhyperspectral images of salted pork samples acquired in the NIRrange (10002500 nm); (2) compare three spectral parameters,i.e., reectance spectra (RS), absorbance spectra (AS) and Kub-elkaMunk spectra (KMS), to improve the robustness of predictionmodels; (3) build robust PLSR calibration models between theobtained spectral information and the reference MC values; (4)identify the most signicant wavelengths linked to MC predic-tions; (5) build new quantitative models with the selected impor-tant wavelengths based on different spectral parameters, and (6)apply the prediction models to hypercubes to obtain distributionmaps depicting the variation of MC within salted meat samples.

2. Materials and methods

2.1. Sample preparation and measurement of moisture content

Pork meat samples (longissimus dorsi) were obtained from localEspana-Espana, Perez-Aparicio, & De Pedro-Sanz, 2008). However,although traditional NIR spectroscopy can determine the majorcomposition of meat products rapidly and non-invasively, it cannotdetect compositional gradients within the sample. In reality, thereare many cases where spatial evolution of quality parameters is

198 D. Liu et al. / Food Chemmarkets at 0.5 days post-mortem. For salting treatment, 30% NaCl(w/w) was employed and samples were treated at room tempera-ture (25 C). Ninety meat slices (7 mm in thickness) were cut. Inorder to obtain representative samples at different salting periods,a salting treatment was conducted on these slices for different timeperiods (0, 5, 15, 30, 60, and 180 min). After treatment, the surfacemoisture was wiped by paper towels before image acquisition. Byimplementing the above method, a contrast was highlightedbetween pork meat samples with different moisture contents toachieve better predictions. Reference moisture content of meatwas determined using the AOAC oven dryingmethod, and themois-ture content was calculated based on the mass loss after drying.Each sample was analysed in duplicate and the average values ofmoisture for each sample were used in subsequent analyses. Therewerewide variations in themoisture content for the examined porkslices. The overall measuredmoisture contents varied from 55.2% to74.9% with a mean of 63.4% and a standard deviation of 5.22%.

2.2. Hyperspectral imaging system

A typical pushbroom hyperspectral imaging system (10002500 nm) was employed. It consisted of a line-scan spectrograph(Specim V25E, Spectral Imaging Ltd., Oulu, Finland) covering thespectral range of 10002500 nm, a high performance 320 256CCD camera (XC403, Xenics Infrared Solutions, Leuven, Belgium),camera lens (OLES30, Xenics Infrared Solutions, Leuven, Belgium)for the spectral range of 10002500 nm, two halogen lamps form-ing the illumination unit (3900-ER, Illumination Technologies Inc.,New York, USA), a conveyer belt operated by a stepper motor(IRCP0076-1COMB, Isuzu Optics Corp., Taiwan, China), data acqui-sition software (Spectral Image software, Isuzu Optics Corp., Tai-wan, China) and a computer. The spectrograph had a xed-sizeinternal slit (30 lm) to dene a eld of view (FOV) for the spatialline (horizontal pixel direction) and collected spectral images inthe reectance mode in the wavelength range of 9162534 nmwith a spectral increment of about 6.32 nm between the contigu-ous bands, thus producing a total of 256 bands. The speed of theconveyer belt was adjusted to 22 mm s1 to synchronise with theimage acquisition by the Spectral Image software, which controlledthe exposure time, motor speed, binning mode, wavelength rangeand image acquisition.

2.3. Image processing and extraction of spectral data

For each salting period, pork samples were placed on the con-veyer belt and then moved to the eld of view of camera and werescanned line by line. Finally, 90 hyperspectral images (15 pork sam-ples 6 periods) were obtained. The acquired images were storedin a raw format before being processed. Each acquired image calledhypercube contains a stack of two-dimensional images at differentwavelengths and can be described as I (x, y, k). A visual inspection ofthe acquired hyperspectral images revealed a high level of noise atboth ends of the spectral range, thus being not useful for the spectraldata extraction. Therefore, the range spanning from 916 to 1000 nmand from 2452 to 2534 nm of the images were removed, leading tothe images being resized to the spectral range of 10052445 nmwith a total of 228 bands. To eliminate the differences in cameraquantum and physical conguration of imaging systems, the origi-nal hyperspectral images (R0) were corrected into the reectancemode (Rc) based on white reference imagesW for a standard Teonwhite tile (100% reectance) and black reference images B for darkcurrent (0% reectance). The formula applied was as follows:

Rc R0 BW B 100% 1

After image acquisition and reectance calibration, the region of

y 152 (2014) 197204interests (ROIs) can be easily identied based on segmentationwith a simple thresholding, due to the distinctive spectral differ-ences between meat sample and background spectrum. The

ture content and spectral data extracted from porcine meat at

istrdifferent salting periodswere established using partial least squaresregression (PLSR). PLSR is a robust and reliable method for con-structing empirical predictive models when the experimental fac-tors are numerous and highly collinear (He et al., 2012; Lin, Chen,& He, 2012). The spectra utilised here included two categories, i.e.,full-wavelength spectra and simplied spectra selected by coef-cient regression. Therefore, two categories of PLSR calibrationmod-els were built. MLR is another simple calibration algorithm, whichbuilds a relationship between the response variable and observedspectral variables in the form of a linear equation. However, thisfails when variables are more than the number of samples and iseasy to be affected by the colinearity between the variables (Wuet al., 2012). Therefore, a MLR model was established based on theselected optimal wavelength instead of full wavelength.

In general, in order to obtain efcient and reliable models, it isimportant to select the optimal number of latent variables to avoidover-tting and under-tting (ElMasry et al., 2011; Liu, He,Wang, &Sun, 2011). The optimal number of latent variable (LVs) for estab-lishing the PLSR calibration model was determined by using theminimum value of predicted residual error sum of squares (PRESS),which shows the squared sum of deviation between predicted andreference values of quality parameters. The quality of calibrationelimination of the background was performed by a masking oper-ation that converted the background pixels to zero reectance. Foreach of the calibrated images, a mask was created by thresholdinga gray image that was produced by subtracting the image at band150 (of low reectance value) from the image at band 25 (of highreectance value). The resulted image was then segmented by asimple thresholding of 0.24. Finally, the isolated sample portionwas treated as the main region of interest (ROI) to be used forextracting spectral and spatial information. The same mannerwas also conducted for spectral extraction of target object (onlylean part without fat layer), by isolating only the lean portion ofthe sample from the fat and background. Background segmenta-tion and extraction of reectance spectra from the hyperspectralimages were carried out using the software ENVI 4.8 (ITT VisualInformation Solutions, Boulder, CO, USA). To compare the abilityof different spectral proles for the quantitative determination ofMC in salted pork slices, three spectral parameters, namely reec-tance spectra (RS), absorbance spectra (AS) and KubelkaMunkspectra (KMS) were considered in this study. The spectroscopictransformations were realised by the following equations:

A log10R 2

KM 1 R2

2R3

where A is absorbance, R is reectance and KM is KubelkaMunk.

2.4. Multivariate data analysis

Before the analysis of the samples, the whole dataset (90 sam-ples) was manually divided into two groups; the calibration setand the prediction set. To ensure that both groups covered the sim-ilar range of reference values, a calibration set was formed by 66samples with 11 samples for each salting period and a predictionset was formed by 24 samples with 4 samples for each salting per-iod. The large spectral data containing hidden information requiresa reliable method to process and extract features of the spectra.Therefore, it is important to build a calibration model for quantita-tive or qualitative analysis. The quantitative models betweenmois-

D. Liu et al. / Food Chemmodels were evaluated by coefcient of determination in calibra-tion, cross-validation as well as prediction (R2c , R

2cv and R

2p , respec-

tively) and root mean square error by calibration, cross-validation2.5. Optimal wavelength selection strategy

The acquired high dimensional hyperspectral images suffer fromthe problem of multicollinearity during a multivariate analysis.Some congruent wavelengths are related to similar constituents,and consequently contain much of the same information (Sun,2010). Improvements are needed to reduce the large number of cor-related variables to a new subset that can explain the maximumvariance. Many research endeavours have directed to nd a vitalfew wavelengths that are the most inuential on the quality evalu-ation of products, and to eliminate wavelengths having irrelevantinformation (Liu, Sun, & Zeng, 2013). However, there is no standardmethod to select the signicant wavelengths from the whole spec-trum. A number of approaches have been proposed for wavelengthselection in various spectral analyses, such as stepwise regression(Feng & Sun, 2012), articial neural network (ElMasry, Wang, &Vigneault, 2009), principal component analysis (Kamruzzamanet al., 2011), independent component analysis (Shao, Bao, & He,2011) and others (ElMasry et al., 2011; Yang, Chao, & Chen, 2005).

In this study, weighted PLS regression coefcients also calledb-coefcients from the PLSR models, were used to select the mostinuential wavelengths on MC prediction. The principle of themethod is based on the calculation of the weighted regressioncoefcients that correspond to the model with full spectra. Thewavelengths with the highest absolute values of regression coef-cients are selected as the optimal wavelengths. Using only theselected wavelengths, new optimised PLSR models and MLRmodels were then developed with three spectral calibration sets,AS, RMS and RS. All steps described for spectral analysis were car-ried out in multivariate analysis software (Unscrambler version9.7, CAMO, Trondheim, Norway).

2.6. Visualisation of MC

Recognising moisture distribution in salted pork slice is usefulto understand the migration of moisture content in pork slice dur-ing salting treatment. Chemical imaging is a methodology to visu-alise this phenomenon. Hyperspectral imaging is a usefultechnique to visualise the spatial evolution of moisture content,by creating concentration images or maps of moisture content.The best optimised models were used to visualise and map eachpixel of the hyperspectral image in the form of chemical imageto predict moisture distribution in the tested muscles. The result-ing chemical image or prediction map is displayed with a linearcolour scale, where the moisture content linearly correspondedto the colour scale. In this way, by checking the colour variationin the developed map, one can easily assess the predicted distribu-tion of moisture content within a pork slice as well as the moistureevolution at different salting periods. All steps involved in visual-isation purposes were implemented with a programme developedusing Matlab (Version 2009a, The Mathworks Inc., Mass, USA). Themain steps of the experimental procedure are presented in Fig. 1.

3. Results and discussion

3.1. Spectral features of salted meatand by prediction (RMSEC, RMSECV, RMSEP). A good model shouldhave low values of RMSEC, RMSECV and RMSEP, as well as high val-ues of R2c , R

2cv, R

2p , and small difference between RMSEC, RMSECV and

RMSEP (Barbin et al., 2013; ElMasry et al., 2011; Nicola et al., 2007).

y 152 (2014) 197204 199After checking the spectral prole, it was found that the imagesat wavelength before 1005 nm and after 2445 nm revealed heavysalt and pepper noises, due to the low signal-to-noise ratio in these

expe

200 D. Liu et al. / Food Chemistrtwo spectral regions. Therefore, only spectra within the range of10052445 nm were selected for analysis. The average reectancespectra of porcine meat samples at different periods of salting areshown in Fig. 2. In general, the most prominent features that inu-enced the near-infrared spectra in the pork samples include theCAH stretching overtone associated to fat, OAH stretching over-tone related to water and NAH from amides and amines in organiccompounds related to the protein (Barlocco et al., 2006). The dom-inating bands observed at 1070, 1445 and 1940 nmwere due to thecombination of symmetric and antisymmetric OAH stretchingmodes (rst overtone) (Collell et al., 2010; Iqbal et al., 2013). Ataround 1255 nm there is another peak, representing the absorption

Fig. 1. Main steps ofband of CAH stretching second overtone (ElMasry et al., 2011),which is related to fat content of the sample (Fig. 2). It was noticedthat the spectral reectance curves of pork samples originatedfrom different salting periods exhibited a similar pattern acrossthe whole tested wavelength region. Fresh meat (salted 0 min) dif-fered to a large extent from the other ve salted samples in themagnitude of reectance throughout the whole spectrum. More-over, signicant variations in spectra amongst samples from thesame pork slice were also found. Variations observed in the spec-tral reectance amongst pork muscles could be related to the bigdifference in moisture content between fresh sample and the other

0

0.2

0.4

0.6

900 1400 1900 2400Wavelength, nm

Ref

lect

ance

0 min5 min15 min30 min60 min180 min

Fig. 2. Mean near infrared spectra extracted from the salted meat in thehyperspectral wavelength range of 10002500 nm.salted samples. Reectance values of samples decreased with anincrease in sample MC (absorbance increased). The only differenceamongst pork spectra is the magnitude of reectance, which makesthe determination of moisture content a difcult procedure. Tosolve this problem, it was necessary to conduct the chemometricanalysis for the data mining.

3.2. Prediction of moisture contents using full spectral range

The prediction of moisture content was performed using PLSRmodels based on three different spectral prole (absorbance, KMand reectance) in the full wavelength range (228 wavelengths).

rimental procedure.

y 152 (2014) 197204These established models were denoted as AS-PLSR, KMS-PLSRand RS-PLSR, models, respectively. Table 1 shows the main statis-tics used to evaluate the performance of the developed calibration,cross-validation and prediction models for predicting moisturecontent of the examined pork samples.

A critical step in building a robust PLSR model was to choosethe correct number of latent variables (LVs), to avoid an overt-ting or undertting model (Lavine, 2003). The optimum numberof latent variables was estimated at the lowest value of predictedresidual errors sum of squares (PRESS), as shown in Fig. 3. Thenumber of latent variables for predicting moisture content withAS, KMS and RS was twelve, eleven and nine, respectively. Tovisualise graphically the performance of the PLSR models, themeasured values obtained from the laboratory measurementsand its predicted values for calibration and prediction sets wereplotted and are displayed in Fig. 3. It was demonstrated thatthe RS-PLSR model had better performance than those of theother two ones, with the highest correlation coefcients of0.969 and 0.941 as well as the lowest RMSEs of 0.921 and1.23% for calibration and prediction, respectively. When theabsorbance and KM spectra were considered, decreases in theprediction accuracy (R2p) values was observed especially for KMspectra, indicating the reectance spectra were more suitablefor determining the moisture content of salted pork slice thanthe other two spectral proles. This result was inconsistent withthat reported by other authors (Feng & Sun, 2012), which demon-strated that the full wavelength PLSR model based on absorbancespectra give better performance than reectance and KM spectra

l par

Cali

R2c

istrTable 1Performance of PLSR (full and simplied) and MLR models based on different spectra

Spectral prole Model type No wavelength LVs

D. Liu et al. / Food Chemto determine total viable count (TVC) in chicken. This discrepancymight be elucidated by the dissimilar samples in terms of theirphysicochemical properties. The differences in quality attributes,uneven physical structure of the meat samples, as well as unxedscatter of the surface, led to variations in the recorded spectralproles, thus exerting different inuences on the subsequentaccuracy of calibration modelling.

AS PLSR 228 12 0.96PLSR 9 6 0.94MLR 9 0.94

KMS PLSR 228 11 0.96PLSR 14 7 0.86MLR 14 0.91

RS PLSR 228 9 0.96PLSR 7 7 0.93MLR 7 0.94

Fig. 3. Prediction of moisture content using full wavelength PLSR models for calibration athe optimum number of LVs for (a) AS, (b) KMS and (c) RS. Measured versus predictedameters.

bration Cross-validation Prediction

RMSEC R2cv RMSECV R2p

RMSEP

y 152 (2014) 197204 2013.3. Multivariate statistical analysis based on optimal wavelengths

In practice, improvements are needed to address the issue oflarge dimensionality of hyperspectral data, which limits imple-mentations of HSI for on-line systems. Therefore, it is importantto conduct wavelength selection and to further design an opti-mised multispectral imaging system with a limited number of

2 1.02 0.934 1.36 0.920 1.442 1.26 0.923 1.47 0.912 1.513 1.24 0.913 1.54 0.910 1.495 0.980 0.878 1.85 0.795 2.307 1.90 0.781 2.48 0.813 2.203 1.54 0.826 2.18 0.862 1.899 0.921 0.948 1.21 0.941 1.234 1.34 0.919 1.51 0.914 1.499 1.18 0.927 1.41 0.917 1.48

nd prediction sets based on different spectral parameters. PRESS plot for identifyingvalue of moisture content by (d) AS, (e) KMS and (f) RS.

wavebands to meet the needs of real-time acquisition and pro-cessing. In this study, b-coefcients resulting from PLSR modelswere employed to allocate important wavelengths aiming toestablish simplied PLSR models. As a result, nine (1178, 1299,1363, 1389, 1643, 1788, 1941, 2213, 2439 nm), fourteen (1178,1299, 1363, 1401, 1643, 1807, 1921, 1941, 2036, 2118, 2193,2276, 2382, 2389 nm), and six (1089, 1127, 1166, 1293, 1719,1928, 2238 nm) wavelengths were identied as important wave-lengths for AS, KMS and RS, respectively (Fig. 4). These optimalwavelengths provided potential indication of changes in chemicalcomponents, namely, water, fat and amides and amino acids,which can account for quality change of porcine meat during salt-ing treatment. Once the optimal wavelengths were selected, thesewavelengths were then used as effective wavelengths to replacethe full range spectra, for the determination of moisture content

in pork slices. PLSR and MLR were carried out, respectively basedon the reduced spectral data, and the results are shown in Table 1.It can be seen that the accuracy of optimal PLSR and MLR modelsbased on AS or RS was slightly lower compared to the modelsdeveloped using the whole spectral range (228 wavelengths). Ingeneral, the accuracy of a regression model is regarded as excel-lent when its R2 is higher than 0.9 (Guy, Prache, Thomas,Bauchart, & Andueza, 2011). Both optimal AS and RS regressionmodels can therefore be concluded as good models (AS: PLSRR2p = 0.912, RMSEP = 1.51%; MLR R

2p = 0.910, RMSEP = 1.49%; RS:

PLSR R2p = 0.914, RMSEP = 1.49%; MLR R2p = 0.917, RMSEP = 1.48%).

On the other hand, the optimised PLSR and MLR models basedon KM spectra resulted in a R2 of less than 0.9, indicating thatthey were less robust calibration models. Based on the optimisedmodel performance in terms of R2 and RMSEs, the RS-MLR model

-30

-20

-10

0

10

20

30

1000 1300 1600 1900 2200 2500Wavelength, nm

Reg

ress

ion

coef

ficie

nt 1643

1299

1389 2213

1941

1178

2439

(a)

17881363

-4

-2

0

2

4

1000 1300 1600 1900 2200 2500

ave

Reg

ress

ion

coef

ficie

nt

1363

12991401

1941

1643

1178

2193

(b)

1807

23821921

2036

2118 2276

2389

ave

17

202 D. Liu et al. / Food Chemistry 152 (2014) 197204W

-80

-40

0

40

80

120

1000 1300 1600W

Regr

essio

n co

effic

ient

1127

1166

12931089

(c)Fig. 4. Selection of important wavelengths (variables) corresponding to large values ospectral parameters: (a) AS; (b) KMS; (c) RS.length, nm

1900 2200 2500length, nm

19

1928

2238f regression coefcients (regardless of sign) of PLSR calibration models for tested

was found as the best calibration model for the prediction ofmoisture content during salting process and therefore was usedfor subsequent visualisation purposes.

3.4. Visualisation of moisture distribution

The nal RS-MLR model established using optimal wave-lengths was transferred to each pixel of the image, to predictthe moisture content in all spots of the pork sample. A predictionimage (called distribution map) was created after multiplying themodels regression coefcients by the spectrum of each pixel inthe image, to visualise the moisture content within the muscles.In the resultant distribution map, pixels that had similar spectralfeatures gave the same predicted value of moisture content,which were then visualised in a similar colour in the image. Dif-ferent moisture contents from high to low were shown in differ-ent colours from red to green (colour bar at the right in Fig. 5).Although it is difcult if not impossible to gure out the differ-ence in MC from point to point by the naked eye, the difference

future research will be conducted to shed more insight into therapid evaluation of informative parameters, such as water phase

supported by China Postdoctoral Science Foundation(2013M530366) and Fundamental Research Funds for the Central

D. Liu et al. / Food Chemistrin moisture contents from location to location within the samesample could be easily discerned from the nal distribution maps.It can be seen that the unsalted slice (0 min) contains a signicantamount of moisture as predicted by our model. It is also worthnoting the non-uniform distribution of water content within apork slice illustrated by a mixture of different colours. The evolu-tions and distributions of moisture content in pork slices at differ-ent salting periods exhibited a general trend of decrease in overallmoisture content from red to yellow, which is a clear indicationof moisture migration during the salting treatment of pork slices(Fig. 5).

In brief, the concentration maps can provide very signicantinformation not available from either conventional imaging orspectroscopy alone. By combining both spatial and spectral fea-tures in one single system, NIR hyperspectral imaging has theadvantage of being effective in predicting and visualising moistureevolution in pork at different salting periods and facilitates theexploration of quality formation mechanism of salted meat prod-ucts. This technique could be possibly benecial to the meat pro-cessing industry by performing quality evaluation and inspection,as a process monitoring means at the early stage of production.This technique offers a novel objective and quantitative methodfor the quality evaluation of meat products and can lead to ahealthy development of the meat processing industry.Fig. 5. Visualisation of moisture content of a single meat at different saltingperiods.Universities (Project No. 2014ZM0027). Mr. Yao-Ze Feng is grate-fully acknowledged for kind assistance and suggestions for thehyperspectral data mining.

References

AOAC (1997). Ofcial methods of analysis of AOAC international (16th ed., 3rdrevision). Washington, D.C.: Association of Ofcial Analytical Chemists. Method950.46Forced Air oven drying method; Method 985.14moisture in meat andpoultry products, rapid microwave drying method.

Barbin, D. F., ElMasry, G., Sun, D.-W., & Allen, P. (2013). Non-destructivedetermination of chemical composition in intact and minced pork using near-infrared hyperspectral imaging. Food Chemistry, 128, 11621171.

Barlocco, N., Vadell, A., Ballesteros, F., Galietta, G., & Cozzolino, D. (2006). Predictingintramuscular fat, moisture and WarnerBratzler shear force in pork muscleusing near infrared reectance spectroscopy. Animal Science, 82, 111116.

Collell, C., Gou, P., Picouet, P., Arnau, J., & Comaposada, J. (2010). Feasibility of near-infrared spectroscopy to predict aw and moisture and NaCl contents offermented pork sausages. Meat Science, 85, 325330.

Cui, Z. W., Xu, S. Y., & Sun, D.-W. (2004). Effect of microwave-vacuum drying on thecarotenoids retention of carrot slices and chlorophyll retention of chinese chiveleaves. Drying Technology, 22(3), 563575.

Delgado, A. E., & Sun, D.-W. (2002). Desorption isotherms for cooked and cured beefand pork. Journal of Food Engineering, 51(2), 163170.

Du, C. J., & Sun, D.-W. (2005). Comparison of three methods for classication ofpizza topping using different colour space transformations. Journal of FoodEngineering, 68(3), 277287.

Elmasry, G., Barbin, D. F., Sun, D.-W., & Allen, P. (2012). Meat Quality evaluation byhyperspectral imaging technique: An overview. Critical Reviews in Food Scienceand Nutrition, 52(8), 689711.

ElMasry, G., Sun, D.-W., & Allen, P. (2011). Non-destructive determination of water-holding capacity in fresh beef by using NIR hyperspectral imaging. FoodResearch International, 44(9), 26242633.salt content (WPS) and water activity aw, in relation to the safetyand stability of investigated meat products during the saltingprocess using HSI.

Acknowledgments

The authors are grateful to the Guangdong Province Govern-ment (China) for support through the programme of Leading Tal-ent of Guangdong Province (Da-Wen Sun). This research was also4. Conclusions

The results indicated that moisture contents in porcine meatsduring the salting process could be determined accurately by HSImeasurements. The predicted values of moisture obtained withboth full wavelengths and optimal wavelengths, based on reec-tance or absorbance spectra, produced a good performance. How-ever, PLSR and MLR models with optimal wavelengths were lessrobust in case of KM spectra. The optimal RS-MLR model turnedout to be the best with high correlation coefcients of 0.949 and0.917 as well as low root mean squared errors of 1.18% and1.48% for calibration and prediction, respectively. By transferringthe RS-MLR model to each pixel of the image, prediction mapswere generated to visualise the distribution and migration of MCon the meat surface. This feasibility study demonstrated that HSIis a useful non-destructive tool for monitoring salting process inthe meat industry. The determination of moisture content usingHSI technique for industrial uses is of signicant importance, how-ever, only MC is not sufcient to predict the shelf-life of meat prod-ucts, which should be completed with the measurement ofthermodynamic activity of water in food. A further step to revealthe behaviour of water in foods should be meaningful. Therefore,

y 152 (2014) 197204 203ElMasry, G., Wang, N., ElSayed, A., & Ngadi, M. (2007). Hyperspectral imaging fornondestructive determination of some quality attributes for strawberry. Journalof Food Engineering, 81(1), 98107.

ElMasry, G., Wang, N., & Vigneault, C. (2009). Detecting chilling injury in reddelicious apple using hyperspectral imaging and neural networks. PostharvestBiology and Technology, 52, 18.

Feng, Y.-Z., & Sun, D.-W. (2012). Determination of total viable count (TVC) inchicken breast llets by near-infrared hyperspectral imaging and spectroscopictransforms. Talanta. http://dx.doi.org/10.1016/j.talanta.2012.11.042.

Gaitan-Jurado, A. J., Ortiz-Somovilla, V., Espana-Espana, F., Perez-Aparicio, J., & DePedro-Sanz, E. J. (2008). Quantitative analysis of pork dry-cured sausages toquality control by NIR spectroscopy. Meat Science, 78, 391399.

Ghosh, P. K., & Jayas, D. S. (2009). Use of spectroscopic data for automation in foodprocessing industry. Sensing & Instrumentation for Food Quality, 3, 311.

Guy, F., Prache, S., Thomas, A., Bauchart, D., & Andueza, D. (2011). Prediction of lambmeat fatty acid composition using near-infrared reectance spectroscopy(NIRS). Food Chemistry, 127, 12801286.

He, H.-J., Wu, D., & Sun, D.-W. (2012). Non-destructive and rapid analysis ofmoisture distribution in farmed Atlantic salmon (Salmo salar) llets usingvisible and near-infrared hyperspectral imaging. Innovative Food Science andEmerging Technologies. http://dx.doi.org/10.1016/j.ifset.2013.02.009.

Iqbal, A., Sun, D.-W., & Allen, P. (2013). Prediction of moisture, colour and pH incooked, pre-sliced turkey hams by NIR hyperspectral imaging system. Journal ofFood Engineering, 11, 742751.

Ruiz-Ramirez, J., Arnau, J., Serra, X., & Gou, P. (2005). Relationship between watercontent, NaCl content, pH and texture parameters in dry-cured muscles. MeatScience, 70, 579587.

Serra, X., Ruiz-Ramirez, J., Arnau, J., & Gou, P. (2005). Texture parameters of dry-cured ham m. biceps femoris samples dried at different levels as a function ofwater activity and water content. Meat Science, 69, 249254.

Shao, Y., Bao, Y., & He, Y. (2011). Visible/near-infrared spectra for linear andnonlinear calibrations, a case to predict soluble solids contents and pH value inPeach. Food and Bioprocess Technology, 4, 13761383.

Sleagun, G., & Popa, M. (2009). Determination of the moisture content in infrareddried apples. International Symposium Euro-aliment. Galati, Romania, 910October.

Sun, D.-W. (1999). Comparison and selection of EMC ERH isotherm equations forrice. Journal of Stored Products Research, 35(3), 249264.

Sun, D.-W. (2010). Hyperspectral imaging for food quality analysis and control. SanDiego, CA: Academic Press, Elsevier.

Sun, D.-W., & Byrne, C. (1998). Selection of EMC/ERH isotherm equations forrapeseed. Journal of Agricultural Engineering Research, 69(4), 307315.

Sun, D.-W., & Woods, J. L. (1993). The moisture-content relative-humidityequilibrium relationship of wheat - a review. Drying Technology, 11(7),15231551.

Sun, D.-W., & Woods, J. L. (1994a). The selection of sorption isotherm equations for

204 D. Liu et al. / Food Chemistry 152 (2014) 197204Kamruzzaman, M., ElMasry, G., Sun, D.-W., & Allen, P. (2011). Prediction of somequality attributes of lamb meat using NIR hyperspectral imaging andmultivariate analysis. Analytica Chimica Acta. http://dx.doi.org/10.1016/j.aca.2011.11.037.

Kamruzzaman, M., ElMasry, G., Sun, D.-W., & Allen, P. (2012). Non-destructiveprediction and visualization of chemical composition in lamb meat using NIRhyperspectral imaging and multivariate regression. Innovative Food Science andEmerging Technologies. http://dx.doi.org/10.1016/j.ifset.2012.06.003.

Kestens, V., Charoud-Got, J., Bau, A., Bernreuther, A., & Emteborg, H. (2008). Onlinemeasurement of water content in candidate reference materials by acousto-optical tuneable lter near-infrared spectrometry (AOTF-NIR) using pork meatcalibrants controlled. Food Chemistry, 106, 13591365.

Lavine, B. (2003). A user-friendly guide to multivariate calibration andclassication. Journal of Chemometrics, 10, 571572.

Lin, P., Chen, Y., & He, Y. (2012). Identication of geographical origin of olive oilusing visible and near-infrared spectroscopy technique combined withchemometrics. Food and Bioprocess Technology, 5, 235242.

Liu, D., Sun, D.-W., & Zeng, X. A. (2013). Recent advances in wavelength selectiontechniques for hyperspectral image processing in the food industry. Food andBioprocess Technology. http://dx.doi.org/10.1007/s11947-013-1193-6.

Liu, F., He, Y., Wang, L., & Sun, G. M. (2011). Detection of organic acids and pH offruit vinegars using near-infrared spectroscopy and multivariate calibration.Food and Bioprocess Technology, 4, 13311340.

Lorente, D., Aleixos, N., & Gomez-Sanchis, J. (2012). Recent advances andapplications of hyperspectral imaging for fruit and vegetable qualityassessment. Food and Bioprocess Technology, 5, 11211142.

Mathlouthi, M. (2001). Water content, water activity, water structure and thestability of food stuffs. Food Control, 12, 409417.

Nicola, B. M., Beullens, K., Bobelyn, E., Peirs, A., Wouter, S., Theron, K. I., et al. (2007).Nondestructive measurement of fruit and vegetable quality by means of NIRspectroscopy: A review. Postharvest Biology and Technology, 46(2), 99118.

Osborne, B. G. (2006). Near-infrared spectroscopy in food analysis. In R. A. Meyers(Ed.), Encyclopedia of Analytical Chemistry (pp. 114). Chichester, British: JohnWiley & Sons Ltd..

Prieto, N., Roehe, R., Lavin, P., Batten, G., & Andres, S. (2009). Application of nearinfrared reectance spectroscopy to predict meat and meat products quality: Areview. Meat Science, 83(2), 175186.

Rajkumar, P., Wang, N., Elmasry, G., Raghavan, G. S. V., & Gariepy, Y. (2012). Studieson banana fruit quality and maturity stages using hyperspectral imaging.Journal of Food Engineering, 108(1), 194200.wheat-based on the tting of available data. Journal of Stored Products Research,30(1), 2743.

Sun, D.-W., & Woods, J. L. (1994b). Low-temperature moisture transfercharacteristics of wheat in thin-layers. Transactions of the ASAE, 37(6),19191926.

Sun, D.-W., & Woods, J. L. (1994c). Low-temperature moisture transfercharacteristics of barley - thin-layer models and equilibrium isotherms.Journal of Agricultural Engineering Research, 59(4), 273283.

Sun, D.-W., & Woods, J. L. (1997). Simulation of the heat and moisture transferprocess during drying in deep grain beds. Drying Technology, 15(10),24792508.

Taghizadeh, M., Gowen, A., & ODonnell, C. P. (2009). Prediction of white buttonmushroom (Agaricus bisporus) moisture content using hyperspectral imaging.Sensonry and Instrumentation for Food Quality, 3, 216219.

Tao, F., Peng, Y., Li, Y., Chao, K., & Dhakal, S. (2012). Simultaneous determination oftenderness and Escherichia coli contamination of pork using hyperspectralscattering technique. Meat Science, 90(3), 851857.

Weeranantanaphan, J., Downey, G., Allen, P., & Sun, D.-W. (2011). A review of nearinfrared spectroscopy in muscle food analysis: 20052010. Journal of NearInfrared Spectroscopy, 19, 61104.

Wu, D., Wang, S., Wang, N., Nie, P., He, Y., Sun, D.-W., et al. (2012). Application oftime series hyperspectral imaging (TS-HSI) for determining water distributionwithin beef and spectral kinetic analysis during dehydration. Food andBioprocess Technology. http://dx.doi.org/10.1007/s11947-012-0928-0.

Yang, C. C., Chao, K., & Chen, Y. R. (2005). Development of multispectral imageprocessing algorithms for identication of wholesome, septicemic, andinammatory process chickens. Journal of food Engineering, 69, 225234.

Zheng, C. X., Sun, D.-W., & Zheng, L. Y. (2006a). Recent applications of image texturefor evaluation of food qualities - a review. Trends in Food Science & Technology,17(3), 113128.

Zheng, C. X., Sun, D.-W., & Zheng, L. Y. (2006b). Recent developments andapplications of image features for food quality evaluation and inspection - areview. Trends in Food Science & Technology, 17(12), 642655.

Zhu, F., Zhang, D., He, Y., Liu, F., & Sun, D.-W. (2012). Application of visible and nearinfrared hyperspectral imaging to differentiate between fresh and frozen-thawed sh llets. Food and Bioprocess Technology. http://dx.doi.org/10.1007/s11947-012-0825-6.

Feasibility of using hyperspectral imaging to predict moisture content of porcine meat during salting process1 Introduction2 Materials and methods2.1 Sample preparation and measurement of moisture content2.2 Hyperspectral imaging system2.3 Image processing and extraction of spectral data2.4 Multivariate data analysis2.5 Optimal wavelength selection strategy2.6 Visualisation of MC

3 Results and discussion3.1 Spectral features of salted meat3.2 Prediction of moisture contents using full spectral range3.3 Multivariate statistical analysis based on optimal wavelengths3.4 Visualisation of moisture distribution

4 ConclusionsAcknowledgmentsReferences