Hyperspectral imaging with multivariate analysis for technological parameters prediction and classification of muscle foods: A review

    Hyperspectral imaging with multivariate analysis for technological parameters prediction and classification of muscle foods: A review Jun-Hu Cheng, Bart Nicolai, Da-Wen Sun PII: DOI: Reference:

S0309-1740(16)30325-4 doi:10.1016/j.meatsci.2016.09.017 MESC 7105

To appear in:

Meat Science

Received date: Revised date: Accepted date:

27 April 2016 26 September 2016 29 September 2016

Please cite this article as: Cheng, J.-H., Nicolai, B. & Sun, D.-W., Hyperspectral imaging with multivariate analysis for technological parameters prediction and classification of muscle foods: A review, Meat Science (2016), doi:10.1016/j.meatsci.2016.09.017

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT Hyperspectral Imaging with Multivariate Analysis for Technological Parameters Prediction and Classification of Muscle Foods: A Review

RI P

T

Jun-Hu Chenga, b, c, Bart Nicolai c, Da-Wen Suna, b, d*

School of Food Science and Engineering, South China University of Technology, Guangzhou 510641, China

b

Academy of Contemporary Food Engineering (ACFE), South China University of Technology, Guangzhou Higher

SC

a

NU

Education Mega Center, Guangzhou 510006, China

MeBioS, Department of Biosystems, KU Leuven, Willem de Croylaan 42, 3001 Heverlee, Belgium

d

Food Refrigeration and Computerized Food Technology (FRCFT), Agriculture and Food Science Centre,

MA

c

ED

University College Dublin, National University of Ireland, Belfield, Dublin 4, Ireland

PT

Abstract: Muscle foods are very important for a well-balanced daily diet. Due to their perishability and vulnerability, there is a need for quality and safety evaluation of such foods. Hyperspectral

AC CE

imaging (HSI) coupled with multivariate analysis is becoming increasingly popular for the non-destructive, non-invasive, and rapid determination of important quality attributes and the classification of muscle foods. This paper reviews recent advances of application of HSI for predicting some significant muscle foods parameters, including color, tenderness, firmness, springiness, water-holding capacity, drip loss and pH. In addition, algorithms for the rapid classification of muscle foods are also reported and discussed. It will be shown that this technology has great potential to replace traditional analytical methods for predicting various quality parameters and classifying muscle foods. Keywords: muscle food, HSI, technological parameter, chemometrics, classification * Corresponding author. Tel: +353-1-7167342, Fax: +353-1-7167493, E-mail: [email protected], Website: www.ucd.ie/refrig; www.ucd.ie/sun. 1

ACCEPTED MANUSCRIPT Contents 1. Introduction ........................................................................................................................................ 3

RI P

T

2. Applications ....................................................................................................................................... 6 2.1. Prediction of Technological Parameters .................................................................................. 6

SC

2.1.1. Tenderness ...................................................................................................................... 6 2.1.2. Color Attributes .............................................................................................................. 9

NU

2.1.3. Drip loss (DL) .............................................................................................................. 12

MA

2.1.4. Water-holding capacity (WHC) ................................................................................... 13 2.1.5. Firmness ....................................................................................................................... 14

ED

2.1.6. Springiness ................................................................................................................... 15

PT

2.1.7. pH................................................................................................................................. 16 2.2. Classification.......................................................................................................................... 18

AC CE

2.2.1 Grading ......................................................................................................................... 19 2.2.2 Muscles discrimination ................................................................................................. 20 2.2.3 Differentiation of fresh and frozen meat products ........................................................ 21 3. Conclusions and Future Trends ........................................................................................................ 23 Acknowledgements .............................................................................................................................. 25 References ............................................................................................................................................ 25

2

ACCEPTED MANUSCRIPT

1. Introduction

RI P

T

Muscle foods, such as pork, beef, lamb, chicken, fish and others play an important role in providing us with valuable proteins and other nutrients (Jiménez-Colmenero et al., 2001; Ryan et al.,

SC

2011), but they are very perishable and vulnerable. They undergo microbial spoilage and chemical degradation during postmortem storage, which will cause quality deterioration and loss of freshness

NU

(Borras et al., 2015; Cheng et al., 2015b). Changes of some significant quality parameters during

MA

spoilage, including color features, textural properties (firmness/hardness, tenderness, cohesiveness, springiness, gumminess, and chewiness), water-holding capacity (WHC), drip loss (DL), and pH

ED

affect quality and safety considerably (Cheng et al., 2013). Thus, there is a need for reliable and accurate determination of these parameters. In addition, in order to avoid adulteration and fraudulent

PT

practices for quality inspection purposes, effective and precise authentication and classification

AC CE

methods are also required (Kamruzzaman et al., 2015). However, the currently used analytical methods and techniques for the measurement and classification of muscle foods, such as mechanical and instrumental analysis, chromatography techniques, and manual sorting, are destructive, time-consuming, and tedious with many chemicals, producing toxic waste and causing environmental pollution (Alander et al., 2013). In recent several years, hyperspectral imaging (HSI) technique, also known as spectral imaging or chemical imaging, has shown to be an innovative platform for quality and safety assessment of food. HSI technique combines traditional optical spectroscopy and computer vision into one system to generate the spatial and spectral information of the tested sample simultaneously. More specifically, a traditional spectroscopic instrument provides a single spectrum I(λ), while an imaging system 3

ACCEPTED MANUSCRIPT normally generates 2-D image information I(x, y). Accordingly, a 3-D hypercube I(x, y, λ) is obtained as shown in Fig. 1. It can be understood as a separated spatial image I(x, y) at each individual

RI P

T

wavelength (λ), or a spectrum I(λ) at each single pixel (x, y) (Sun, 2010). Moreover, each pixel in the obtained spectral image contains a spectral profile with its corresponding spatial position, which is

SC

very meaningful to acquire the visualization distribution of the changes of quality information (Gowen et al., 2007). Generally speaking, a typical HSI system normally includes light source,

NU

wavelength dispersion device (spectrograph), area detector, and a computer control system. Light

MA

source is used to generate the light for illumining the testing sample. Spectral dispersion device is to distribute wideband light into different wavelengths. Area detector, such as the commonly used

ED

charge-coupled device (CCD), is capable of controlling and quantifying the intensity of the obtained

PT

light by means of transferring incident photons into electrons. Computer control system is used to acquire and calibrate the hyperspectral images and to regulate the exposure time, motor speed,

AC CE

combining mode, and wavelength range. Furthermore, the hyperspectral image acquisition methods are referred to point scanning, line scanning, and area scanning. The used image sensing models are listed as reflectance, transmittance and interactance that are of difference in the positions of optical detector and the light source (Cheng & Sun, 2014). Because HSI is a rapid, non-invasive, non-destructive, chemical-free, and environment friendly tool, it has been successfully used for food quality detection and assessment under laboratory conditions or in research scenarios (Liu et al., 2013; He et al., 2015; Mathiassen et al., 2011) and shows great potential for replacing traditional analytical methods for on-line industrial applications (Cheng et al., 2013; Cozzolino & Murray, 2012; Dowlati et al., 2012). Due to the hypercube structure with thousands of spectra (spectral signature) distributed over the measured area (spatial signature), 4

ACCEPTED MANUSCRIPT the obtained spectral and spatial information needs to be processed statistically. Chemometrics analysis is very useful to analyze the data hypercube. The great advantage of chemometrics is its

RI P

T

ability to reduce the complexity of large data sets, to build classification and prediction models, and to enhance the accuracy and robustness of models based on spectral data analysis (Lohumi et al.,

SC

2015). Spectral pretreatment techniques mainly including multiplicative scatter correction (MSC), standard normal variate (SNV), smoothing, baseline removal, and first and second derivatives, are

NU

used to reduce and correct possible interferences related to scattering, baseline drift, path-length

MA

variation, and overlapping bands (Borràs et al., 2015; Rinnan et al., 2009; Tavallaie et al., 2011). Variable selection techniques, such as principal component analysis (PCA), regression coefficient

ED

analysis (RC, also called β-coefficient), the successive projections algorithm (SPA), uninformative

PT

variable elimination (UVE), and genetic algorithms (GA), are commonly used to select the most informative spectral regions/the optimal wavelengths for simplifying the modeling purposes and

AC CE

constructing on-line multispectral imaging detection system (Andersen & Bro, 2010; Blanco & Villarroya, 2002; Dai et al., 2015b; Liu et al., 2014). Frequently used modeling methods for quantitative analysis include multiple linear regression (MLR), partial least squares regression (PLSR), artificial neural network (ANN), and least squares-support vector machine (LS-SVM) (Cheng & Sun, 2015; Dai et al., 2014; Kleinbaum et al., 2013). The obtained model is generally evaluated by a series of statistical parameters such as the determination coefficients of calibration (R2C), cross-validation (R2CV) and prediction (R2P); the corresponding root mean square error estimated by calibration (RMSEC), cross-validation (RMSECV) and prediction (RMSEP); as well as the overall indicator of residual predictive deviation (RPD). Generally speaking, an admirable and good model should have higher values of R2C, R2CV, R2P and RPD, and lower values of RMSEC, 5

ACCEPTED MANUSCRIPT RMSECV and RMSEP as well as a small difference between them. Up to now, no review is available which addresses the use of HSI in combination with multivariate

RI P

T

analysis to measure technological parameters of muscle food. Therefore, this review aims to report and discuss recent developments of HSI for the rapid and non-destructive prediction of technological

SC

parameters of muscle foods and their classification in the five recent years.

NU

2. Applications

MA

Fig. 2 illustrates the flowchart of using HSI method in tandem with multivariate data analysis for the determination of quality attributes of muscle foods. The published investigations for the

ED

prediction of technological parameters in muscle foods are highlighted and summarized in Table 1

PT

and detailed as follows.

AC CE

2.1. Prediction of Technological Parameters

2.1.1. Tenderness

Tenderness is a significant technological parameter used for evaluating the eating quality of muscle foods from a consumer’s perception point of view. Consumers differentiate the level of tenderness through sensory evaluation at consumption and then make re-purchase decisions. Due to the complex chemical and biochemical changes during postmortem storage, the degree of proteolysis and the differing collagen content, which causes variation in tenderness among muscle food samples (Cheng et al., 2014c). The currently applied instrumental method such as Warner-Bratzler shear force (WBSF) is more objective and accurate for quantitative measurement of tenderness compared with sensory evaluation method. However, shear force measurement is invasive and inefficient and 6

ACCEPTED MANUSCRIPT usually causes larger deviation in tenderness measurement. HSI has been used for estimating meat tenderness. Barbin et al. (2013b) built PLSR prediction models between the reference tenderness

RI P

T

measured using the instrumental slice shear force (SSF) and spectral information, image feature information, and the combinations extracted from the pork meat hyperspectral images. The resulting

SC

PLSR model using spectra and wavelet feature parameters showed the strongest performance with an R2CV of 0.75 and R2P of 0.74 for the measured tenderness. However, due to the relatively low

NU

reliability and accuracy of the established PLSR models, it is perhaps difficult to develop this HSI

MA

technology for further on-line detection of pork tenderness. Consequently, further investigations are needed to improve the model predictive robustness. Another type of HSI system in the NIR spectral

ED

range of 900-1700 nm was used for the rapid and non-invasive measurement of beef tenderness (ElMasry et al., 2012). The PLSR model established using the spectral signatures extracted from the

PT

hyperspectral images of beef samples showed an acceptable result with an R2CV of 0.83 and

AC CE

RMSECV of 40.75 N. Fifteen of the most important wavelengths were identified by the weighted regression coefficient analysis (RC) method based on PLSR analysis and were selected to reduce the high dimensionality of the hyperspectral data. The simplified PLSR model using these wavelengths showed satisfactory prediction performance (R2CV = 0.77 and RMSECV = 47.75 N). Similarly, the tenderness of lamb meat as measured by the traditional instrumental (Warner-Bratzler shear force, WBSF) method and sensory evaluation by trained panelists was predicted by means of HSI (Kamruzzaman et al., 2013). The established PLSR models showed a reasonable prediction accuracy (RCV = 0.84 and RMSECV = 5.71 N for WBSF; RCV = 0.69 and RMSECV = 1.23 N for sensory tenderness). It is obvious to observe that HSI with PLSR analysis is more suitable for estimating the instrumental tenderness than that of sensory tenderness of lamb meat (RCV: 0.84 > 0.69). In order to 7

ACCEPTED MANUSCRIPT construct a multispectral imaging system for rapid detection of WBSF-related tenderness values, eleven optimal wavelengths were identified by SPA. The simplified SPA-PLSR model also resulted

RI P

T

in satisfactory results (RCV = 0.84 and RMSECV = 5.83 N). However, the performance for predicting tenderness of the mentioned muscle samples (Barbin et al., 2013b; ElMasry et al., 2012;

SC

Kamruzzaman et al., 2013) is not very good and is needed to be enhanced for further industrial applications. Therefore, some other researchers (Wu et al., 2012; Tao et al., 2012; Tao & Peng, 2014)

NU

used some different function relationships to investigate the feasibility of HSI for predicting

MA

tenderness. For example, the MLR calibration models established using Lorentzian distribution function parameters combined with the optimal wavelengths selected by a stepwise discrimination

ED

method showed it was possible to predict beef tenderness (RCV = 0.91 and RMSECV = 9.93 N; RP =

PT

0.95 and RMSEP = 7.95 N) (Wu et al., 2012). Tao et al. (2012) used three parameters of the Lorentzian distribution function to estimate the scattering from hyperspectral images for determining

AC CE

pork tenderness. The results indicated that good prediction results were obtained using MLR for pork tenderness (R = 0.83, 0.86, 0.86, and 0.93). In order to advance the robustness and reliability of their prediction models, these authors (Tao & Peng, 2014) developed a HSI system (400-1000 nm) and used a modified Gompertz function with four parameters for pork meat tenderness prediction. They established an MLR model for tenderness using these parameters and obtained a good prediction performance (RC = 0.99 and RMSEC = 2.80 N; RCV = 0.95 and RMSECV = 5.70 N). Based on these investigations above-mentioned, it can be demonstrated that HSI in combination with chemometrics analysis is useful and has the potential for the rapid and non-destructive estimation of pork, beef and lamb tenderness. Tenderness is also an important quality attribute of fish fillet. Therefore, prediction of tenderness 8

ACCEPTED MANUSCRIPT in fresh farmed salmon fillets using Vis/NIR HSI (400-1700 nm) was investigated (He et al., 2014). Linear PLSR and nonlinear LS-SVM were used to correlate spectral data of salmon samples in two

RI P

T

wavelength ranges (spectral range I, 400-1000 nm; spectral range II, 900-1700 nm) with reference tenderness values estimated by the traditional instrumental WBSF approach. Regardless of which

SC

spectral range used, the prediction results were better when using LS-SVM compared to PLSR (I: RP = 0.90 > 0.89 and RMSEP = 1.10 < 1.16; II: RP = 0.88 > 0.86 and RMSEP = 1.20 < 1.32).

NU

Furthermore, it appeared that spectral range I was superior for modeling than range II. Afterwards,

MA

four wavelengths including 555 nm, 605 nm, 705 nm, and 930 nm were selected from spectral range I using SPA. The optimized SPA-LS-SVM showed the best reliability with a RP of 0.91 and RMSEP

ED

of 1.09. In addition, the authors extracted three kinds of textural features from the hyperspectral

PT

images of fish only at the four selected wavelengths using an image information extraction algorithm named gray level co-occurrence matrix (GLCM). A new LS-SVM model was constructed using the

AC CE

data fusion of spectra and texture information and showed an acceptable performance (RP = 0.89 and RMSEP = 1.17) (He et al., 2014). The results indicated that combining image textural features and spectral information was more suitable for building an LS-SVM model for tenderness prediction of salmon fillet.

2.1.2. Color Attributes

Color is one of the most important sensory properties of meat. It has been generally recognized that color is associated with the physical, chemical, biochemical, microbial, and sensory evolution of food quality during growth, maturation, postharvest handling and processing, packaging and storage as well as trading and marketing (Vestergaard et al., 2000; Wu and Sun, 2013a). Particularly, the 9

ACCEPTED MANUSCRIPT main contribution and application of color features is to play an important role in grading and assessment of external quality for meat freshness, ripeness, and attractiveness, and further for meat

RI P

T

safety control (Grunert et al., 2004). Some consumers use color features to form expectations about meat quality at the point of purchase, based on their own experience and informational cues available

SC

in the shopping environment (Grunert et al., 2004). Therefore, color to some extent influences the purchasing decisions of consumers.

NU

Traditional sensory evaluation and instrumental methods for measuring color are not suitable for

MA

large numbers of samples and on-line applications (McCaig, 2002). A HSI system in the wavelength range of 400-1000 nm in combination with a Lorentzian distribution function to describe scattering

ED

was used to measure the color attributes (L*, a*, and b*) of fresh beef with an RCV of 0.96, 0.96, and

PT

0.97, respectively (Wu et al., 2012). Another HSI system operating in the NIR region of 900-1700 nm was also used for the quantitative determination of color information of beef muscle (ElMasry et

AC CE

al., 2012). The established PLSR model showed a reasonably good performance for predicting L* and b* with an R2CV of 0.88 and 0.81 and a corresponding RMSECV of 1.21 and 0.58, respectively. However, a* was not well predicted, maybe due to the narrow range of a* values measured by the traditional method. Six and five critical wavelengths were identified using the weighted RC method from PLSR analysis for predicting L* and b*. The PLSR models based on the critical wavelengths also produced comparable results with the original ones (RCV = 0.88 and 0.80; RMSECV = 1.24 and 0.60). This type of imaging system was also used to predict color features in cooked, pre-sliced turkey hams (Iqbal et al., 2013). Surprisingly, only a* was successfully predicted based on PLSR analysis with an R2CV of 0.72 and RMSECV of 0.37. The prediction results of L* and b* were not satisfactory. This was possibly related to the surface appearance of the meat slices or associated with 10

ACCEPTED MANUSCRIPT the processing parameters of the turkey ham, resulting in the interference of measurement of L* and b*. Likewise, nine optimal wavelengths were selected for predicting a* to result in an R2CV of 0.74

RI P

T

and RMSECV of 0.35. However, the obtained results (Iqbal et al., 2013) showed that HSI was not efficient to predict color information of cooked and pre-sliced turkey hams due to the low accuracy.

SC

More endeavors and attempts are needed to develop the HSI technology for further detecting the color variations during meat processing and handling conditions such as cooking. In another work, a

NU

long wave NIR-HSI system used for determination of color features in salmon fillet was investigated

MA

(Wu et al., 2012). PLSR models based on the full wavelength range were developed to predict L* with an RCV of 0.87, RMSECV of 2.26, and RP of 0.86, and RMSEP of 2.42, and to predict b* with

ED

an RCV of 0.83, RMSECV of 2.14, RP of 0.80, and RMSEP of 2.06. The performance of a PLSR

PT

model for predicting a* was poor with lower values of RCV (0.73) and RP (0.74). In addition, four, six, and ten key wavelengths were identified by the SPA method for the prediction of L*, a*, and b*. The

AC CE

reduced PLSR models showed results that were comparable to those obtained with the original models for L*, a*, and b* prediction in salmon fillets (Wu et al., 2012). Similarly, in another work, the authors (Cheng et al., 2014b) compared the prediction performance of HSI with two different wavelength regions (400-1000 nm and 1000-2500 nm) for the measurement of color components in grass carp fillet. Both with PLSR and LS-SVM satisfactory results for L* and a* were obtained using two spectral ranges; b* was only poorly predicted. More importantly, based on a statistical analysis, it was proved that the HSI system using the wavelength region of 400-1000 nm performed much better than the other system (1000-2500 nm) based on the higher values of R2P and lower values of RMSEP. It can be concluded that the HSI system with a wavelength range of 400-1000 nm was more appropriate and trustworthy for the prediction of color attributes (L* and a*), which was in 11

ACCEPTED MANUSCRIPT disagreement with the above study addressing the NIR wavelength range reported by Wu et al. (2012). These authors found that the HSI in the spectral range of 900-1700 nm successfully predicted

RI P

T

the color components of salmon fillets. The above differences in prediction results may be due to the different fish species and the different compositions of fish muscles used in the studies that to some

SC

extent influenced the color properties (Cheng et al., 2014b).

NU

2.1.3. Drip loss (DL)

MA

Drip loss (DL) usually influences the juiciness, flavor, and texture of food (Huff-Lonergan & Lonergan, 2005). DL indicates the loss of water soluble nutritional compounds such as water-soluble

ED

and sarcoplasmic proteins, which provides a nutritious medium for microbial activity (Duun & Rustad, 2008). Unfortunately, the current DL measurement methods are normally tedious,

PT

contaminative, and time-consuming for large amounts of samples. Therefore, the feasibility of HSI in

AC CE

the NIR region of 900-1700 nm for predicting DL of lamb meat originating from four different breeds was investigated (Kamruzzaman et al., 2012a). A PLSR model was established using reference measured DL values and the NIR spectral data extracted from the tested muscles, and relatively poor results were obtained with an R2CV of 0.77, RMSECV of 0.28%, and RPD of 2.11, respectively. However, these calibration models (Kamruzzaman et al., 2012a) showed lower prediction accuracy and they may be further enhanced by including more tested samples in the calibration process to increase the variability of the dataset and improve the predictive effectiveness. Another study used two HSI systems with spectral range I (400-1000 nm) and spectral range II (900-1700 nm) for non-invasively and rapidly measuring DL value in salmon fillets (He et al., 2014). The PLSR calibration model using spectral range I performed better than that of using spectral range 12

ACCEPTED MANUSCRIPT II for DL prediction, with a higher value of RCV (0.81 vs. 0.69) and a lower value of RMSECV (0.07 vs. 0.09). Moreover, by analyzing regression coefficients of the PLSR model, eleven informative

RI P

T

wavelengths that were important for the DL prediction were selected for model optimization purposes. Compared to Model I using the full spectra, similar results were obtained with an RCV of

SC

0.83 and RMSECV of 0.07 (He et al., 2014). Similarly, in another study (Cheng et al., 2016), five key wavelengths were chosen using a combination of GA and SPA for predicting the DL in

NU

frozen-thawed grass carp fillet. The established MLR model had a better performance than the

MA

LS-SVM model with an increase of 0.62 and 0.03 in RPD and R2P, and a decrease of 0.22% in RMSEP. The results confirmed that it is feasible to identify wavelength features using variable

PT

systems in future applications.

ED

selection methods and chemometrics analysis for developing on-line multispectral imaging detection

AC CE

2.1.4. Water-holding capacity (WHC)

Determination of the WHC value in muscle food is an effective and useful way for describing changes of meat quality (Olsson et al., 2003). A HSI system was developed for predicting WHC in fresh beef from different breeds based on spectral features of hyperspectral images (ElMasry et al., 2011). The full spectrum PLSR model showed a very good performance in both calibration and cross-validated sets with an R2C of 0.92, RMSEC of 0.21%, R2CV of 0.89, and RMSECV of 0.26%. Based on the analysis of the regression coefficients of the PLSR model, a simplified PLSR model using six informative wavelengths also generated a superior performance (R2C = 0.89 and RMSEC = 0.25%; R2CV = 0.87 and RMSECV = 0.28%). In another work, both PLSR and LS-SVM were applied to measure WHC related indicators including percentage fat loss (PFL), percentage water 13

ACCEPTED MANUSCRIPT remained (PWR), percentage water loss (PWL), and percentage liquid loss (PLL) of salmon fillet muscle with two HSI systems in the spectral range of 400-1000 nm and 900-1700 nm (Wu & Sun,

RI P

T

2013b). The results showed that the models established by using the first spectral range performed better with perfect results of RC and RP higher than 0.90. Also, based on the selected wavelengths

SC

analysis, compared with LS-SVM, PLSR models had better performances for WHC prediction in salmon fillet flesh, with satisfactory results of RP of 0.82, 0.94, 0.97, and 0.94 for PFL, PLL, PWR,

NU

and PWL, respectively. Similarly, PLSR and LS-SVM were applied to monitor the WHC of red

MA

meats including pork, beef, and lamb muscle using a Vis/NIR HSI system (400-1000 nm) (Kamruzzaman et al., 2016). The LS-SVM model was superior to the PLSR model when the full

ED

spectral data was used, with R2C, R2P, and RPD values equal to 0.97, 0.94, and 4.58, and RMSEC

PT

and RMSEP values of 0.36% and 0.50%, which was in disagreement with the above mentioned study (Wu & Sun, 2013b). In addition, eight feature wavelengths were chosen by the RC method for

AC CE

designing a convenient multispectral system for red meat industrial application. The simplified RC-LS-SVM model also produced excellent results with an R2P and RPD of 0.93 and 4.09 as well as an RMSEP of 0.56% (Kamruzzaman et al., 2016). It can be observed that HSI in combination with multivariate analysis showed superior results for DL prediction in muscle foods and it can be demonstrated that the HSI system may replace the laborious, time-consuming, and destructive methods as a rapid and non-destructive tool for detecting the WHC in meat muscle samples.

2.1.5. Firmness

Firmness is one of the most important quality traits of muscle foods. It is also a vital attribute for mechanical processing of meat (Casas et al., 2006). Because the current firmness measurement 14

ACCEPTED MANUSCRIPT methods such as the Warner-Bratzler test, Kramer shear compression, and texture profile analysis are destructive and invasive, thus, a Vis/NIR HSI system in combination with chemometrics analysis

RI P

T

was used to measure the firmness of grass carp fillet treated by frozen at -20 °C for 24 h and then thawed at 4 °C for 5 d (Cheng et al., 2014a). The LS-SVM model resulted in a better prediction

SC

performance with the increase of 0.02, 0.02, and 0.03 for R2C, R2CV, and R2P, and the decrease of 0.69, 0.82, and 0.84 for RMSEC, RMSECV, and RMSEP compared with the PLSR model (R2C =

NU

0.91, R2CV = 0.91, and R2P = 0.90; RMSEC = 2.16 N, RMSECV = 2.16 N, and RMSEP = 2.19 N).

MA

HSI thus appears to be capable of measuring the firmness in fish fillets during frozen storage. In addition, three variable selection methods of RC, SPA, and GA were used to recognize the most

ED

significant feature wavelengths for reflecting the variations of firmness. A simplified GA-LS-SVM

PT

model using seven selected wavelengths showed the best results with an R2P value of 0.94 and an RMSEP of 1.23 N (Cheng et al., 2014a). It means that the new developed model using the optimal

AC CE

wavelengths showed a comparable prediction accuracy with that of the corresponding original models. Therefore, seven key wavelengths chosen by GA instead of the whole wavelengths are sufficient to predict the firmness for quality evaluation and control of fish fillets influenced by frozen storage.

2.1.6. Springiness

Springiness plays a great role in the eating quality of common meat. Generally, springiness indicates the mechanical capability of muscle to resist external forces. Undesirable springiness of meat would lead to quality degradation and loss of freshness (Bourne, 2002). Application of Vis/NIR HSI was attempted to predict the springiness of fresh chicken meat (Xiong et al., 2014). A PLSR 15

ACCEPTED MANUSCRIPT model resulted in an acceptable predictive performance with a RC and RMSEC of 0.91 and 0.14, and RP and RMSEP of 0.89 and 0.14. A nonlinear ANN model produced relatively good results with an

RI P

T

RC of 0.84, RMSEC of 0.17, RP of 0.77, and RMSEP of 0.19, respectively. It can be inferred that both PLSR and ANN were suitable for predicting springiness. Linear factors (PLSR analysis) were

SC

better correlated to the spectral variables than nonlinear factors (LS-SVM analysis). In addition, the new optimized PLSR model established using ten critical wavelengths chosen by SPA showed better

NU

predictive results with a higher RP of 0.84 and a lower RMSEP of 0.16 (Xiong et al., 2014). The

ED

and non-invasive manner using HSI.

MA

results pointed out the great potential for the measurement of springiness of chicken meat in a rapid

2.1.7. pH

PT

pH is a critical technological attribute that affects microbial growth. It also has a great influence

AC CE

on color, flavor, WHC, water activity, and shelf-life of meat (Iqbal et al., 2013). Determination of pH is often conducted by a portable pH meter or a surface electrode, which is destructive, inefficient, and not suitable for large scale industrial applications. A pushbroom HSI system in the NIR region (900-1700 nm) was implemented to detect the pH of lamb meat from three different part of muscles (Kamruzzaman et al., 2012b). A full cross-validated PLSR model was developed using the reference pH of the tested muscles and NIR reflectance spectral data. The corresponding prediction results were obtained with an R2CV of 0.65 and RMSECV of 0.09. However, it can be observed that the statistical data showed a relatively poor performance for pH prediction, which was possibly attributed to the narrow range of pH that was used for constructing the prediction models. Better results were obtained in another study using the same spectral imaging technique (ElMasry et al., 16

ACCEPTED MANUSCRIPT 2012). Satisfactory results were obtained with the PLSR model based on the full spectra range with an R2C of 0.83 with RMSEC of 0.05 and an R2CV of 0.73 with RMSECV of 0.06. For further

RI P

T

improving the prediction accuracy and reliability, Barbin et al. (2012b) used a similar technique to measure the pH of pork samples from the longissimus dorsi muscle. In this study, first and second

SC

derivative, SNV, and MSC were applied for pre-processing the raw spectral data. The established PLSR model with SNV showed the best performance for pH predictions (R2CV = 0.88 and RMSECV

NU

= 0.10; R2P = 0.87 and RMSEP = 0.10). In addition, independent sets of feature-related wavelengths

MA

(947 nm, 1057 nm, 1161 nm, 1308 nm, and 1680 nm) were selected and very satisfactory statistical results were obtained using five selected wavelengths with an R2P of 0.90 and RMSEP of 0.09

ED

(Barbin et al., 2012b). Another HSI system (900-1700 nm) was used to determine the pH of cooked, pre-sliced turkey hams (Iqbal et al., 2013). A PLSR model based on the entire spectrum resulted in

PT

good prediction results (R2C = 0.86, RMSEC = 0.01, R2CV = 0.77, and RMSECV = 0.02). In order to

AC CE

increase the prediction capability of the calibration models and to simplify the established models, eight feature wavelengths were selected, and a good prediction performance was also achieved with an R2C of 0.88, R2CV of 0.81 with RMSEC of 0.02 and RMSECV of 0.02 (Iqbal et al., 2013), which indicated a reasonable accuracy and reliability of the used PLSR model. In another work, the pH of farmed salmon fillets was predicted with a HSI system (He et al., 2014). A PLSR model was constructed using the full wavelength region of 400-1000 nm and good correlations with an R2C of 0.80 and RMSEC of 0.04, and an R2CV of 0.76 and RMSECV of 0.05 as well as RPD of 2.00 were generated. Also, ten important wavelengths having great impact on pH prediction were identified for further optimization. The optimized PLSR model resulted in an R2C of 0.80, RMSEC of 0.04, R2CV of 0.77, and RMSECV of 0.05, respectively. It can be noticed that the performance of the simplified 17

ACCEPTED MANUSCRIPT model based on the most important wavelengths was comparable to that of the model using the full wavelength range, although the number of wavelengths was reduced from 121 to 10 wavelengths, or

RI P

T

a reduction of 92%. The new model was more robust and practical for predicting the pH. Another HSI system (900-1700 nm) was used to determine the pH of Atlantic salmon fillets during cold

SC

storage for 0, 3, 6, 9 and 12 d at 1 oC (Xu et al., 2016). The PLSR model displayed less satisfactory prediction results with an RCV and RP of 0.75 and 0.80, and the absolute differences between

NU

RMSECV and RMSEP of 0.01 for pH. It is obvious that, compared with the developed HSI in the

MA

spectral range of 400-1000 nm, the used wavelength region in 900-1700 nm showed poor results for prediction of pH. It can be demonstrated that different spectral range of HSI showed great influence

ED

on model accuracy and reliability.

PT

2.2. Classification

AC CE

Classification is one of the most important aspects for the evaluation of meat quality, because different level of meat shows different price and different quality. It is very necessary to clearly classify the meat according to the grading level. However, consumers normally use human visual decision-making process to purchase the meat. This kind of behavior is subjective and misjudgment. Classification algorithms, also called classifiers, can offer a simple yes/no answer and also can estimate the probability of which an object belongs to each of the candidate classes (Ireland & Møller, 2000). Commonly used classifiers include principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA), soft independent modeling of class analogy (SIMCA), LS-SVM, artificial neural networks (ANN), and probabilistic neural networks (PNN) (Salimi-Khorshidi et al., 2014). Table 2 illustrates the comparisons of common used classification 18

ACCEPTED MANUSCRIPT algorithms based on unsupervised and supervised pattern. In combination with these classifiers, HSI has a great potential to be used instead of the human visual decision-making process with automatic

RI P

T

procedures and has been increasingly developed for the grading, adulteration, and differentiation of fresh and frozen meat products.

SC

2.2.1 Grading

NU

The reliability of HSI coupled with classifiers analysis was investigated to identify and grade three

MA

different kinds of pork, beef, and lamb muscle (Longissimus dorsi) (Kamruzzaman et al., 2012a). The corresponding hyperspectral images were acquired and their spectra were extracted and used for

ED

modeling by PCA and PLS-DA. The spectral data were pretreated by the second derivative method, and six critical wavelengths were identified. A PLS-DA model based on these selected wavelengths

PT

yielded an overall correct classification rate (CCR) of 98.67% for three types of meat samples in the

AC CE

validation set. Likewise, hyperspectral images of longissimus dorsi pork muscle with three grades of PSE, RFN, and DFD were obtained in the NIR range (900-1700 nm) (Barbin et al., 2012a). Six significant wavelengths were identified in the second derivative spectra, and explained the obvious changes of the spectra, which indicated the variations among pork classes. PCA in combination with the analysis of these selected wavelengths was used for classification, and very satisfactory results with a CCR of 100%, 100%, and 92.31% were acquired for PSE, DFD, and RFN samples, respectively. This study clearly demonstrated that HSI in combination with classifiers has a potential for the rapid classification of diverse pork meat grades.

19

ACCEPTED MANUSCRIPT 2.2.2 Muscles discrimination

T

Muscles discrimination or categorization is a very important aspect in muscle food quality

RI P

evaluation. A NIR-HSI system was used to evaluate the discrimination of three types of semitendinosus (ST), longissimus dorsi (LD), and psoas major (PM) cut from the lamb samples

SC

(Kamruzzaman et al., 2011). PCA was used for dimensionality reduction and wavelength selection.

NU

Based on the analysis of the eigenvector plot of PCA, six feature variables were chosen and then used for classification purposes. Excellent results with an overall accuracy of 100% were obtained

MA

for discriminating three types of lamb muscles. It also means that this HSI technology has the potential and great capacity for detecting the adulteration of pork muscles from the same species.

ED

Likewise, four different lamb muscle cuts from longissimus dorsi, psoas major, semimembranosus,

PT

and semitendinosus were also successfully classified with a CCR of 96.67% using HSI in the

AC CE

wavelength range 380-1028 nm with a linear least mean squares (LMS) classifier (Sanz et al., 2015). Another HSI system (922-1739 nm) in combination with multivariate data analysis was used to categorize two different tenderness beef muscle from the same longissimus dorsi portion (Cluff et al., 2013). The slice shear force (SSF) values were then used to categorize the beef steak samples into two tenderness categories (i.e., tender and tough). Samples that had a SSF ı 206 N were considered as tough and samples that had a SSF < 206 N were considered as tender. The developed PCA model showed classification results with CCR of 83.30% and 75% for tough and tender samples (Cluff et al., 2013). However, the obtained results were a little poor with higher errors and it is difficult to develop this kind of imaging system for very accurate classification of tough and tender beef samples. In another work, wavelet analysis was used to extract information from NIR hyperspectral images for the categorization of lamb muscles using PCA and LS-SVM. A 20

ACCEPTED MANUSCRIPT classification model based on PCA coupled with wavelet analysis showed a higher CCR of 96.15%

RI P

adulteration of muscle foods in a rapid and non-destructive way.

T

(Pu et al., 2014b). Based on these CCR results, HSI with classifiers is capable of detection of

SC

2.2.3 Differentiation of fresh and frozen meat products

HSI coupled with classification algorithms has also been used to differentiate fresh and

NU

frozen-thawed muscle foods. The potential of a pushbroom HSI method in the NIR range of

MA

900-1700 nm was explored for the identification of fresh and frozen-thawed pork longissimus dorsi muscles (Barbin et al., 2013a). An overall CCR of 100% was found for an unknown set of pork meat

ED

samples using PLS-DA. All in all this showed that developing a rapid and reliable spectral imaging system for classification and certification of fresh and frozen meat samples for the benefit of the

PT

retailers and the consumers is feasible. The potential of Vis/NIR HSI in the spectral region of

AC CE

400-1000 nm to classify fresh and frozen-thawed pork meat by combining spectral and image information was also explored (Pu et al., 2015). Six key wavelengths were chosen using UVE and SPA methods. The image texture information of the principal component hyperspectral images were extracted using the gray level-gradient co-occurrence matrix (GLGCM) algorithm. A PNN model based on both optimal wavelengths and GLGCM image features resulted in the highest CCR of 93.14% in calibration and 90.91% in cross validation. In order to enhance the classification accuracy, the same authors (Pu et al., 2015) used wavelet analysis for extracting more image textural features of Vis/NIR hyperspectral images to distinguish fresh from frozen-thawed pork. Five important wavelengths were chosen by RC based on PLSR analysis. Fourth-order Daubechies wavelets were used as the wavelet mother function for wavelet textural extraction of the feature images depending 21

ACCEPTED MANUSCRIPT on the selected feature variables with the wavelet decomposition levels from 1 to 4. Four textural features were calculated in the horizontal, vertical, and diagonal orientations at each level. An

RI P

T

LS-SVM model was established using 48 textural features extracted from each feature image for the distinction of the fresh and frozen pork meat samples and resulted in an improved CCR of 98.48%

SC

and 93.18% in the training set and the testing set, respectively (Pu et al., 2015). The above results demonstrated the potential of HSI for the classification of fresh and frozen-thawed pork meat

NU

samples. It has also been explored for monitoring the handling of aquatic products. A pushbroom

MA

HSI system in the spectral region of 380-1030 nm was used to discriminate fresh, fast freezing (-70 o

C), and slow freezing (-20 oC) fish samples (Zhu et al., 2012). PCA was applied to reduce the

ED

spectral dimension, and the first three PCs explaining more than 98% of the variance of the full

PT

spectral information were extracted. Thirty-six image textural feature variables were obtained by using GLCM on the three PC images. The LS-SVM classification model that incorporated both

AC CE

spectral and textural information was most successful, with a CCR of 97.22%, compared to when individual spectral information was used or image textural features. In another work, the potential of a Vis/NIR HSI system (400-1000 nm) was considered to differentiate fresh, cold-stored, and frozen-thawed shelled shrimp (Qu et al., 2015). Eight characteristic wavelengths were extracted by UVE and SPA methods from the whole variables. GLCM was used to extract the textural information from the first three PCs images explaining over 99% of variances. Two classifiers, SIMCA and random forest (RF), had a good performance with a CCR of 88.89% and 91.11% in the prediction set based on the data fusion of textural and optimal spectral information. The statistical results also confirmed the feasibility of HSI for the classification of fresh and frozen shrimps. A similar study attempted to rapidly detect prawn freshness using Vis/NIR HIS (Dai et al., 2015a). Three kinds of 22

ACCEPTED MANUSCRIPT classification models, LS-SVM, BP-ANN, and adaptive boosting (AdaBoost), were used for classification using first and second derivative spectra based on both the full spectrum as well as

RI P

T

selected critical wavelengths by SPA. Through comparative analysis, the simplified SPA-LS-SVM model showed the best classification performance with an average CCR of 98.33% and 95% for

SC

fresh and frozen samples. Also, Cheng et al. (2015a) compared four classifiers (SIMCA, PLS-DA, LS-SVM, and PNN) and four spectral pretreatment techniques (MSC, SNV, first and second

NU

derivative) to discriminate fresh from cold-stored (4 oC for 7 d) and frozen-thawed (-20 oC and -40 oC

MA

C for 30 d) grass carp fillets. Compared with raw models established based on the full wavelength range, SIMCA, LS-SVM, and PNN coupled with first derivative spectral pretreatment exhibited the

ED

best classification capability with a maximum CCR of 94.29%. Based on the optimal wavelengths

PT

analysis selected by SPA, the new simplified SPA-LS-SVM model combined with first derivatives also showed comparable prediction precision with the CCR of 91.43%. In addition, Ivorra et al.

AC CE

(2013) reported a rapid method for detecting fresh and expired packaged salmon using HSI based on PLS-DA. A fairly good CCR of 82.7% was acquired in cross-validation based on the factually commercial salmon samples. In another study, the prediction ability of Vis/NIR and LW-NIR HSI for the evaluation of freshness of rainbow trout discretely conserved in ice for 1, 3, 5, and 7 d were compared (Khojastehnazhand et al., 2014). Both PCA and PLS-DA were used for discrimination. An overall CCR of 100% and 75% for two spectral ranges was achieved. Summarizing, hyperspectral imaging in combination with classifiers is feasible to used for classifying different types of muscle foods.

3. Conclusions and Future Trends

23

ACCEPTED MANUSCRIPT Rapid, objective, and non-destructive measurement and prediction of technological parameters and classification is important for quality and safety assessment of muscle foods. The currently applied

RI P

T

traditional analytical methods and techniques are, however, invasive, laborious, time-consuming, destructive, and not suitable for real-time and on-line applications. Consequently, HSI has been

SC

evaluated as a rapid, chemical-free, non-contact, objective, and sensitive analytical tool for quality detection and safety control of muscle foods. This paper reviews the recent advances of HSI for

NU

predicting some major quality technological parameters of muscle food including color components

MA

(L*, a*, and b*), textural properties (tenderness, firmness, and springiness), WHC, DL, and pH based on chemometrics analysis using SNV, MSC, and derivatives for preprocessing the raw spectral data,

ED

and PLSR, LS-SVM, MLR, and ANN for establishing the prediction models, as well as PCA, SPA,

PT

GA, and RC for selecting the optimal variables. In addition, the potential of this technology for the discrimination and classification of muscle foods in combination with classifiers such as SIMCA,

AC CE

PLS-DA, SVM, and PNN has also been reported. These investigations have confirmed that HSI in tandem with multivariate analysis shows great potential for prediction of technological parameters and classification of muscle foods. However, there are still some limitations and restrictions to be surmounted for HSI before moving this technology from the laboratories to the industry real-time and on-line inspections. Spectral imaging is considered as a secondary analytical tool, and accurate and reliable reference methods are required. More importantly, hyperspectral images contain much redundant data that pose considerable challenges for data mining, and hardware running speed of HSI system needs to be enhanced to satisfy the rapid acquisition and analysis of the enormous hyperspectral data with 3-5 s for one sample. A multispectral imaging system only depending on several informative wavelengths will be more appropriate to satisfy the speed requirement of quality 24

ACCEPTED MANUSCRIPT inspection with low cost and high efficiency. Investigation of the most suitable model calibration strategy with high accuracy, efficiency, and reliability for further application is also needed to be

RI P

T

considered.

SC

Acknowledgements

This research was supported by the International S&T Cooperation Program of China

NU

(2015DFA71150), the Natural Science Foundation of Guangdong Province (2014A030313244), the

MA

Key Projects of Administration of Ocean and Fisheries of Guangdong Province (A201401C04), the Collaborative Innovation Major Special Projects of Guangzhou City (201508020097), and the

ED

International S&T Cooperation Projects of Guangdong Province (2013B051000010). The authors were also grateful to the Guangdong Province Government (China) for its support through the

PT

program of “Leading Talent of Guangdong Province (Da-Wen Sun)” and the financial support

AC CE

provided by China Scholarship Council (CSC) for supporting Jun-Hu Cheng’s PhD study at KU Leuven in Belgium.

References

Alander JT, Bochko V, Martinkauppi B, Saranwong S, Mantere T. (2013). A review of optical non-destructive visual and near-infrared methods for food quality and safety. International Journal of Spectroscopy 13, 1-36. Andersen CM, Bro R. (2010). Variable selection in regression-a tutorial. Journal of Chemometrics 24(11-12), 728-737. Barbin DF, Elmasry G, Sun DW, Allen P. (2012a). Near-infrared hyperspectral imaging for grading and classification of pork. Meat Science 90(1), 259-268. Barbin DF, ElMasry G, Sun DW, Allen P. (2012b). Predicting quality and sensory attributes of pork using near-infrared hyperspectral imaging. Analytica Chimica Acta 719, 30-42. 25

ACCEPTED MANUSCRIPT Barbin DF, Sun DW, Su C. (2013a). NIR hyperspectral imaging as non-destructive evaluation tool for the recognition of fresh and frozen–thawed porcine longissimus dorsi muscles. Innovative Food Science & Emerging Technologies 18, 226-236.

RI P

T

Barbin DF, Valous NA, Sun DW. (2013b). Tenderness prediction in porcine longissimus dorsi muscles using instrumental measurements along with NIR hyperspectral and computer vision imagery. Innovative Food Science & Emerging Technologies 20, 335-342.

SC

Blanco M, Villarroya I. (2002). NIR spectroscopy: a rapid-response analytical tool. TrAC Trends in Analytical Chemistry 21(4), 240-250.

NU

Borràs E, Ferré J, Boqué R, Mestres M, Aceña L, Busto O. (2015). Data fusion methodologies for food and beverage authentication and quality assessment–A review. Analytica Chimica Acta 891, 1-14.

MA

Bourne M. (2002). Food texture and viscosity: Concept and measurement: Academic Press. ISBN 0-12-119062-5. Canizales-Rodríguez DF, Ocaño-Higuera VM, Marquez-Rios E, Graciano-Verdugo AZ, Cárdenas-López JL,

ED

Yepiz-Gómez MS, Castillo-Yáñez FJ. (2013). Biochemical, physical, chemical, and microbiological assessment of blue shrimp (Litopenaeus stylirostris) stored in ice. Journal of Aquatic Food Product

PT

Technology 24(3), 259-269.

Casas C, Martinez O, Guillen M, Pin C, Salmeron J. (2006). Textural properties of raw Atlantic salmon (Salmo

AC CE

salar) at three points along the fillet, determined by different methods. Food Control 17(7), 511-515. Castro P, Millán R, Penedo JC, Sanjuán E, Santana A, Caballero MJ. (2012). Effect of storage conditions on total volatile base nitrogen determinations in fish muscle extracts. Journal of Aquatic Food Product Technology 21(5), 519-523.

Cheng JH, Dai Q, Sun DW, Zeng XA, Liu D, Pu HB. (2013). Applications of non-destructive spectroscopic techniques for fish quality and safety evaluation and inspection. Trends in Food Science & Technology 34(1), 18-31. Cheng, JH, Sun, DW. (2014). Hyperspectral imaging as an effective tool for quality and safety analysis and control of fish and other seafoods: current research and potential applications. Trends in Food Science & Technology, 37(2), 78-91. Cheng JH, Qu JH, Sun DW, Zeng XA. (2014a). Visible/near-infrared hyperspectral imaging prediction of textural firmness of grass carp (Ctenopharyngodon idella) as affected by frozen storage. Food Research International 56, 190-198. Cheng JH, Sun DW, Pu HB, Chen X, Liu Y, Zhang H, Li JL. (2015a). Integration of classifiers analysis and 26

ACCEPTED MANUSCRIPT hyperspectral imaging for rapid discrimination of fresh from cold-stored and frozen-thawed fish fillets. Journal of Food Engineering 161, 33-39. Cheng JH, Sun DW, Pu H. (2016). Combining the genetic algorithm and successive projection algorithm for the

RI P

T

selection of feature wavelengths to evaluate exudative characteristics in frozen-thawed fish muscle. Food Chemistry 19, 855-863.

Cheng JH, Sun DW, Pu H, Zeng XA. (2014b). Comparison of visible and long-wave near-infrared hyperspectral

SC

imaging for colour measurement of grass carp (Ctenopharyngodon idella). Food and Bioprocess Technology 7(11), 3109-3120.

NU

Cheng JH, Sun DW, Zeng XA, Liu D. (2015b). Recent advances in methods and techniques for freshness quality determination and evaluation of fish and fish fillets: A review. Critical Reviews in Food Science and

MA

Nutrition 55(7), 1012-1025.

Cheng JH, Sun DW. (2015). Recent applications of spectroscopic and hyperspectral imaging techniques with

ED

chemometric analysis for rapid inspection of microbial spoilage in muscle foods. Comprehensive Reviews in Food Science and Food Safety 14(4), 478-490.

PT

Cheng JH, Sun DW, Han Z, Zeng XA. (2014c). Texture and structure measurements and analyses for evaluation of fish and fillet freshness quality: a review. Comprehensive Reviews in Food Science and Food Safety 13(1),

AC CE

52-61.

Cluff K, Naganathan GK, Subbiah J, Samal A, Calkins CR. (2013). Optical scattering with hyperspectral imaging to classify longissimus dorsi muscle based on beef tenderness using multivariate modeling. Meat Science 95(1), 42-50.

Cluff K, Naganathan GK, Subbiah J, Lu R, Calkins CR, Samal A. (2008). Optical scattering in beef steak to predict tenderness using hyperspectral imaging in the VIS-NIR region. Sensing and Instrumentation for Food Quality and Safety, 2(3), 189-196. Cozzolino D, Murray I. (2012). A review on the application of infrared technologies to determine and monitor composition and other quality characteristics in raw fish, fish products, and seafood. Applied Spectroscopy Reviews 47(3), 207-218. Dai Q, Cheng JH, Sun DW, Pu H, Zeng XA, Xiong Z. (2015a). Potential of visible/near-infrared hyperspectral imaging for rapid detection of freshness in unfrozen and frozen prawns. Journal of Food Engineering 149, 97-104. Dai Q, Cheng JH, Sun DW, Zeng XA. (2015b). Advances in feature selection methods for hyperspectral image 27

ACCEPTED MANUSCRIPT processing in food industry applications: a review. Critical reviews in Food Science and Nutrition 55(10), 1368-1382. Dai Q, Cheng JH, Sun DW, Zhu Z, Pu H. (2016). Prediction of total volatile basic nitrogen contents using wavelet

RI P

T

features from visible/near-infrared hyperspectral images of prawn (Metapenaeus ensis). Food Chemistry 197, 257-265.

Dai Q, Sun DW, Xiong Z, Cheng JH, Zeng XA. (2014). Recent advances in data mining techniques and their

SC

applications in hyperspectral image processing for the food industry. Comprehensive Reviews in Food Science and Food Safety 13(5), 891-905.

NU

Dowlati M, de la Guardia M, Mohtasebi SS. (2012). Application of machine-vision techniques to fish-quality assessment. TrAC Trends in Analytical Chemistry 40, 168-179.

MA

Duun AS, Rustad T. (2008). Quality of superchilled vacuum packed Atlantic salmon (Salmo salar) fillets stored at-1.4 and-3.6 oC. Food Chemistry 106(1), 122-131.

ED

ElMasry G, Sun DW, Allen P. (2011). Non-destructive determination of water-holding capacity in fresh beef by using NIR hyperspectral imaging. Food Research International 44(9), 2624-2633.

PT

ElMasry G, Sun DW, Allen P. (2012). Near-infrared hyperspectral imaging for predicting colour, pH and tenderness of fresh beef. Journal of Food Engineering 110(1), 127-140.

AC CE

Gowen A, O'Donnell C, Cullen P, Downey G, Frias J. (2007). Hyperspectral imaging–an emerging process analytical tool for food quality and safety control. Trends in Food Science & Technology 18(12), 590-598. Grunert KG, Bredahl L, Brunsø K. (2004). Consumer perception of meat quality and implications for product development in the meat sector-a review. Meat Science, 66(2), 259-272. He HJ, Wu D, Sun DW. (2014a). Potential of hyperspectral imaging combined with chemometric analysis for assessing and visualising tenderness distribution in raw farmed salmon fillets. Journal of Food Engineering 126, 156-164. He HJ, Wu D, Sun DW. (2015). Non-destructive spectroscopic and imaging techniques for quality evaluation and assessment of fish and fish products. Critical reviews in Food Science and Nutrition 55(6), 864-886. He HJ, Wu D, Sun DW. (2014b). Rapid and non-destructive determination of drip loss and pH distribution in farmed Atlantic salmon (Salmo salar) fillets using visible and near-infrared (Vis-NIR) hyperspectral imaging. Food Chemistry 156, 394-401. Huang Q, Chen Q, Li H, Huang G, Ouyang Q, Zhao J. (2015). Non-destructively sensing pork’s freshness indicator using near infrared multispectral imaging technique. Journal of Food Engineering 154, 69-75. 28

ACCEPTED MANUSCRIPT Huff-Lonergan E, Lonergan SM. (2005). Mechanisms of water-holding capacity of meat: The role of postmortem biochemical and structural changes. Meat Science 71(1), 194-204. Iqbal A, Sun DW, Allen P. (2013). Prediction of moisture, color and pH in cooked, pre-sliced turkey hams by NIR

RI P

T

hyperspectral imaging system. Journal of Food Engineering 117(1), 42-51.

Ireland JD, Møller A. (2000). Review of international food classification and description. Journal of Food Composition and Analysis 13(4), 529-538.

SC

Ivorra E, Girón J, Sánchez AJ, Verdú S, Barat JM, Grau R. (2013). Detection of expired vacuum-packed smoked salmon based on PLS-DA method using hyperspectral images. Journal of Food Engineering 117(3),

NU

342-349.

foods. Meat Science, 59(1), 5-13.

MA

Jiménez-Colmenero F, Carballo J, Cofrades S. (2001). Healthier meat and meat products: their role as functional

Kamruzzaman M, Barbin D, ElMasry G, Sun DW, Allen P. (2012a). Potential of hyperspectral imaging and pattern

Technologies 16, 316-325.

ED

recognition for categorization and authentication of red meat. Innovative Food Science & Emerging

PT

Kamruzzaman M, ElMasry G, Sun DW, Allen P. (2011). Application of NIR hyperspectral imaging for discrimination of lamb muscles. Journal of Food Engineering 104(3), 332-340.

AC CE

Kamruzzaman M, ElMasry G, Sun DW, Allen P. (2012b). Prediction of some quality attributes of lamb meat using near-infrared hyperspectral imaging and multivariate analysis. Analytica Chimica Acta 714, 57-67. Kamruzzaman M, Elmasry G, Sun DW, Allen P. (2013). Non-destructive assessment of instrumental and sensory tenderness of lamb meat using NIR hyperspectral imaging. Food Chemistry 141(1), 389-396. Kamruzzaman M, Makino Y, Oshita S. (2015). Non-invasive analytical technology for the detection of contamination, adulteration, and authenticity of meat, poultry, and fish: A review. Analytica Chimica Acta 853, 19-29. Kamruzzaman M, Makino Y, Oshita S. (2016). Hyperspectral imaging for real-time monitoring of water holding capacity in red meat. LWT-Food Science and Technology 66, 685-691. Khojastehnazhand M, Khoshtaghaza MH, Mojaradi B, Rezaei M, Goodarzi M, Saeys W. (2014). Comparison of visible–near infrared and short wave infrared hyperspectral imaging for the evaluation of rainbow trout freshness. Food Research International 56, 25-34. Kleinbaum D, Kupper L, Nizam A, Rosenberg E. (2013). Applied regression analysis and other multivariable methods: Cengage Learning. 29

ACCEPTED MANUSCRIPT Liu D, Sun DW, Zeng XA. (2014). Recent advances in wavelength selection techniques for hyperspectral image processing in the food industry. Food and Bioprocess Technology 7(2), 307-323. Liu D, Zeng XA, Sun DW. (2013). NIR spectroscopy and imaging techniques for evaluation of fish quality-a

RI P

T

review. Applied Spectroscopy Reviews 48(8), 609-628.

Lohumi S, Lee S, Lee H, Cho BK. (2015). A review of vibrational spectroscopic techniques for the detection of food authenticity and adulteration. Trends in Food Science & Technology 46(1), 85-98.

SC

Mathiassen JR, Misimi E, Bondø M, Veliyulin E, Østvik SO. (2011). Trends in application of imaging technologies to inspection of fish and fish products. Trends in Food Science & Technology 22(6), 257-275.

NU

McCaig T. (2002). Extending the use of visible/near-infrared reflectance spectrophotometers to measure colour of food and agricultural products. Food Research International 35(8), 731-736.

MA

Olsson GB, Ofstad R, Lødemel JB, Olsen RL. (2003). Changes in water-holding capacity of halibut muscle during cold storage. LWT-Food Science and Technology 36(8), 771-778.

ED

Pu H, Sun DW, Ma J, Liu D, Cheng JH. (2014a). Using wavelet textural features of visible and near infrared hyperspectral image to differentiate between fresh and frozen–thawed pork. Food and Bioprocess

PT

Technology 7(11), 3088-3099.

Pu H, Sun DW, Ma J, Cheng JH. (2015). Classification of fresh and frozen-thawed pork muscles using visible and

AC CE

near infrared hyperspectral imaging and textural analysis. Meat Science 99, 81-88. Pu H, Xie A, Sun DW, Kamruzzaman M, Ma J. (2014b). Application of wavelet analysis to spectral data for categorization of lamb muscles. Food and Bioprocess Technology 8(1), 1-16. Qiao J, Wang N, Ngadi MO, Gunenc A, Monroy M, Gariepy C, Prasher SO. (2007). Prediction of drip-loss, pH, and color for pork using a hyperspectral imaging technique. Meat Science, 76(1), 1-8. Qu JH, Cheng JH, Sun DW, Pu H, Wang QJ, Ma J. (2015). Discrimination of shelled shrimp (Metapenaeus ensis) among fresh, frozen-thawed and cold-stored by hyperspectral imaging technique. LWT - Food Science and Technology 62(1), 202-209. Rinnan Å, van den Berg F, Engelsen SB. (2009). Review of the most common pre-processing techniques for near-infrared spectra. TrAC Trends in Analytical Chemistry 28(10), 1201-1222. Ryan, JT, Ross, RP, Bolton, D, Fitzgerald, GF, Stanton, C. (2011). Bioactive peptides from muscle sources: meat and fish. Nutrients, 3(9), 765-791. Salimi-Khorshidi G, Douaud G, Beckmann CF, Glasser MF, Griffanti L, Smith SM. (2014). Automatic denoising of functional MRI data: combining independent component analysis and hierarchical fusion of classifiers. 30

ACCEPTED MANUSCRIPT Neuroimage 90, 449-468. Sanz J A, Fernandes A M, Barrenechea E, Silva S, Santos V, Gonçalves N, Paternain D, Jurio A, Melo-Pinto P. (2016). Lamb muscle discrimination using hyperspectral imaging: comparison of various machine learning

RI P

T

algorithms. Journal of Food Engineering 174, 92-100.

Sun DW. (2010). Hyperspectral imaging for food quality analysis and control: Elsevier. Tao F, Peng Y. (2014). A method for non-destructive prediction of pork meat quality and safety attributes by

SC

hyperspectral imaging technique. Journal of Food Engineering 126, 98-106.

Tao F, Peng Y, Li Y, Chao K, Dhakal S. (2012). Simultaneous determination of tenderness and Escherichia coli

NU

contamination of pork using hyperspectral scattering technique. Meat Science 90(3), 851-857. Tavallaie R, Talebpour Z, Azad J, Soudi MR. (2011). Simultaneous determination of pyruvate and acetate levels in

Food Chemistry 124(3), 1124-1130.

MA

xanthan biopolymer by infrared spectroscopy: effect of spectral pre-processing for solid-state analysis.

ED

Wang X, Zhao M, Ju R, Song Q, Hua D, Wang C, Chen T. (2013). Visualizing quantitatively the freshness of intact fresh pork using acousto-optical tunable filter-based visible/near-infrared spectral imagery. Computers and

PT

Electronics in Agriculture 99, 41-53.

Wu D, Sun DW. (2013a). Colour measurements by computer vision for food quality control–A review. Trends in

AC CE

Food Science & Technology 29(1), 5-20. Wu D, Sun DW, He Y. (2012). Application of long-wave near infrared hyperspectral imaging for measurement of color distribution in salmon fillet. Innovative Food Science & Emerging Technologies 16, 361-372. Wu D, Sun DW. (2013b). Application of visible and near infrared hyperspectral imaging for non-invasively measuring distribution of water-holding capacity in salmon flesh. Talanta 116, 266-276. Wu J, Peng Y, Li Y, Wang W, Chen J, Dhakal S. (2012). Prediction of beef quality attributes using VIS/NIR hyperspectral scattering imaging technique. Journal of Food Engineering 109(2), 267-273. Vestergaard M, Oksbjerg N, Henckel P. (2000). Influence of feeding intensity, grazing and finishing feeding on muscle fibre characteristics and meat color of semitendinosus, longissimus dorsi and supraspinatus muscles of young bulls. Meat Science, 54(2), 177-185. Xiong Z, Sun DW, Dai Q, Han Z, Zeng XA, Wang L. (2014). Application of visible hyperspectral imaging for prediction of springiness of fresh chicken meat. Food Analytical Methods 8(2), 380-391. Xiong Z, Sun DW, Pu H, Xie A, Han Z, Luo M. (2015). Non-destructive prediction of thiobarbituricacid reactive substances (TBARS) value for freshness evaluation of chicken meat using hyperspectral imaging. Food 31

ACCEPTED MANUSCRIPT Chemistry 179, 175-181. Xu JL, Riccioli C, Sun DW. (2016). Efficient integration of particle analysis in hyperspectral imaging for rapid assessment of oxidative degradation in salmon fillet. Journal of Food Engineering 169, 259-271.

RI P

T

Zhu F, Zhang D, He Y, Liu F, Sun DW. (2012). Application of visible and near infrared hyperspectral imaging to differentiate between fresh and frozen–thawed fish fillets. Food and Bioprocess Technology 6(10),

AC CE

PT

ED

MA

NU

SC

2931-2937.

32

ACCEPTED MANUSCRIPT Figure caption: Fig. 1. Hypercube information of hyperspectral image with a three-dimensional dataset I(x, y, λ), which can be

T

understood as a separated spatial image I(x, y) at each individual wavelength (λ), or a spectrum I(λ) at each single

RI P

pixel (x, y).

SC

Fig. 2. Flowchart of HSI system coupled with chemometrics analysis for prediction of technological parameters and classification of muscle foods. It mainly includes image analysis (image acquisition, segmentation,

AC CE

PT

ED

MA

selection methods, and model regression methods).

NU

calibration, and distribution) and multivariate data analysis (spectral pre-processing methods, variable

33

MA

NU

SC

RI P

T

ACCEPTED MANUSCRIPT

AC CE

PT

ED

Fig. 1. Hypercube information of hyperspectral image

34

PT

ED

MA

NU

SC

RI P

T

ACCEPTED MANUSCRIPT

AC CE

Fig. 2. Flowchart of HSI system coupled with chemometrics analysis for prediction of technological parameters and classification of muscle foods

35

Parameter

Tenderness

Tenderness

Tenderness

Tenderness

Tenderness

Tenderness

Tenderness

Tenderness

Tenderness

Tenderness

L*

L*

a*

b*

L*

b*

L*

Muscle food

Pork

Beef

Beef

Lamb

Lamb

Pork

Pork

Salmon fillet

Salmon fillet

Salmon fillet

Pork

Beef

Beef

Beef

Beef

Beef

Beef

PT

AC CE

947, 1078, 1151, 1215, 376

900-1700

900-1700

400-1000

400-1000

400-1000

434, 494, 561, 637, 669, 703

555, 605, 705, 930

900-1700

400-1000

400-1100

400-1100

1615, 1655

36

ED LS-SVM

PLSR

PLSR

PLSR

MLR

MLR

MLR

FNN

0.949

0.930

0.840

0.840

0.670

0.950

0.740

R2 or R

0.880

0.810

0.880

0.970

0.960

0.960

0.860

He et al. (2014a)

He et al. (2014a)

He et al. (2014a)

Tao and Peng (2014)

Tao et al. (2012)

Kamruzzaman et al. (2013)

Kamruzzaman et al. (2013)

Cluff et al. (2008)

Wu et al. (2012)

Barbin et al. (2013b)

Reference

T

ElMasry et al. (2012)

ElMasry et al. (2012)

ElMasry et al. (2012)

Wu et al. (2012)

Wu et al. (2012)

Wu et al. (2012)

Qiao et al. (2007)

RI P

SC 0.905

0.884

0.902

NU

LS-SVM

LS-SVM

MLR

MLR

PLSR

PLSR

MLR

MLR

PLSR

Model

MA

934, 964, 1017, 1081, 1144, 1215, 1265, 1341, 1455,

900-1700

496-1036

400-1100

900-1700

Spectral range (nm)

Table 1. Application of HSI with chemometrics analysis for detection of technological parameters of muscle foods

ACCEPTED MANUSCRIPT

b*

a*

a*

L*

b*

a*

L*

b*

a*

L*

a*

b*

L*

a*

b*

L*

a*

b*

DL

Beef

Turkey hams

Turkey hams

Salmon fillet

Salmon fillet

Salmon fillet

Salmon fillet

Salmon fillet

Salmon fillet

Grass carp fillet

Grass carp fillet

Grass carp fillet

Grass carp fillet

Grass carp fillet

Grass carp fillet

Grass carp fillet

Grass carp fillet

Grass carp fillet

Lamb

900-1700

465, 510, 585, 615, 740, 860,

465, 585, 660, 720, 950 950

466, 525, 590, 620, 715, 850, 955

400-1000

400-1000

400-1000

400-1000

400-1000

400-1000

964, 1081, 1105, 1158, 1295, 1406

1530

ED

37

LS-SVM

LS-SVM

LS-SVM

LS-SVM

LS-SVM

PLSR

0.735

0.798

0.866

0.730

0.798

0.864

0.740

0.720

0.800

0.770

0.731

0.891

0.912

0.722

0.905

0.916

Cheng et al. (2014b)

Cheng et al. (2014b)

Cheng et al. (2014b)

Wu et al. (2012)

Wu et al. (2012)

Wu et al. (2012)

Wu et al. (2012)

Wu et al. (2012)

Wu et al. (2012)

Iqbal et al. (2013)

Iqbal et al. (2013)

ElMasry et al. (2012)

T

Kamruzzaman et al. (2012b)

Cheng et al. (2014b)

Cheng et al. (2014b)

Cheng et al. (2014b)

Cheng et al. (2014b)

Cheng et al. (2014b)

Cheng et al. (2014b)

RI P

SC 0.702

0.887

0.906

NU

LS-SVM

PLSR

PLSR

PLSR

PLSR

PLSR

PLSR

PLSR

PLSR

PLSR

PLSR

PLSR

PLSR

MA

964, 1024, 1084, 1105, 1131, 1161, 1212, 1295, 1440,

PT

AC CE

1295, 1362, 1440, 1527

900-1700

900-1700

900-1700

914, 931, 991, 1115, 1164, 1218, 1282, 1362, 1638

900-1700

934, 1074, 1138, 1399, 1665

ACCEPTED MANUSCRIPT

DL

DL

DL

DL

DL

WHC

WHC

Firmness

Firmness

Firmness

Firmness

Springiness

Springiness

Springiness

pH

pH

pH

pH

pH

Salmon fillet

Salmon fillet

Grass carp fillet

Grass carp fillet

Pork

Beef

Beef

Grass carp fillet

Grass carp fillet

Grass carp fillet

Grass carp fillet

Chicken

Chicken

Chicken

Pork

Pork

Pork

Pork

Lamb

900-1700

494, 571,637, 669, 703, 978

400-1000

400-1000

900-1700

38

416, 458, 581, 637, 696, 722, 740, 754, 773, 973

400-1000

400-1000

450, 530, 550, 616, 720, 955, 980

450, 530, 550, 616, 720, 955, 980

400-1000

400-1000

ED

PT

940, 997, 1144, 1214, 1342, 1443

900-1700

459, 618, 655, 685, 755, 953

432, 500, 588, 660, 980

AC CE

432, 500, 588, 660, 980

415, 445, 500, 590, 605, 675, 760, 825, 880, 955, 990

400-1000

PLSR

FNN

LS-SVM

PLSR

PLSR

PLSR

ANN

PLSR

0.925

0.932

0.899

0.870

0.890

0.770

0.900

0.926

0.834

0.808

0.650

0.550

0.714

0.794

0.900

0.840

Xiong et al. (2014).

Xiong et al. (2014).

Cheng et al. (2014a)

Cheng et al. (2014a)

Cheng et al. (2014a)

Cheng et al. (2014a)

ElMasry et al. (2011)

ElMasry et al. (2011)

Qiao et al. (2007)

Cheng and Sun. (2016)

Cheng and Sun. (2016)

He et al. (2014b)

He et al. (2014b)

T

Kamruzzaman et al. (2012b)

Qiao et al. (2007)

Pu et al. (2014a)

Liu et al. (2014)

Barbin et al. (2012b)

Xiong et al. (2014).

RI P

SC 0.770

0.890

0.941

NU

LS-SVM

PLSR

MA

LS-SVM

PLSR

PLSR

PLSR

FNN

LS-SVM

MLR

PLSR

PLSR

ACCEPTED MANUSCRIPT

Beef

pH

900-1700

39

ED

PT

AC CE

T

ElMasry et al. (2012)

RI P

SC

0.730

NU

MA

PLSR

ACCEPTED MANUSCRIPT

ACCEPTED MANUSCRIPT

Table 2. A comparison of classification algorithms Common algorithm

Advantages

Disadvantages

PCA, KNN, PNN,

Easy to understand and implement

Sensitive to noise

K-means clustering

Free from parameter estimations

LDA, PLS-DA,

Effective

LS-SVM, SIMCA

Suitable for analyzing complex problems

RI P

Unsupervised

T

Classification

AC CE

PT

ED

MA

NU

SC

Supervised

40

Less effective Need a training set Heavy calculation load

ACCEPTED MANUSCRIPT Highlights Technological parameters of muscle foods using HSI system are predicted.

l

Classification of muscle foods using HSI system are presented.

l

Multivariate analysis for supporting the application of HSI are described.

l

Future trends of HSI technique are proposed.

AC CE

PT

ED

MA

NU

SC

RI P

T

l

41