Foodstuff authentication from spectral data: Toward a species-independent discrimination between fresh and frozen–thawed fish samples

Journal of Food Engineering 119 (2013) 765–775

Contents lists available at SciVerse ScienceDirect

Journal of Food Engineering journal homepage: www.elsevier.com/locate/jfoodeng

Foodstuff authentication from spectral data: Toward a species-independent discrimination between fresh and frozen–thawed fish samples Matteo Ottavian a, Luca Fasolato b, Pierantonio Facco a,⇑, Massimiliano Barolo a a b

CAPE-Lab – Computer-Aided Process Engineering Laboratory, Department of Industrial Engineering, University of Padova, via Marzolo 9, 35131 Padova PD, Italy Department of Biomedicine and Food Science, University of Padova, Viale dell’Università 16, 35020 Legnaro PD, Italy

a r t i c l e

i n f o

Article history: Received 5 April 2013 Received in revised form 4 June 2013 Accepted 8 July 2013 Available online 16 July 2013 Keywords: Near-infrared spectroscopy Foodstuff authentication PLS-DA Orthogonal PLS

a b s t r a c t The substitution of fresh fish with frozen–thawed fish is a typical fraud that can damage consumers for several reasons. In fact, not only the quality of thawed meat can be negatively affected during freezing, but also safety issues can arise, as thawed meat is more susceptible to microbial growth. Though several strategies have been proposed for fresh fish authentication, their classification ability is strongly affected by the fish species being considered. In this paper, we propose three different strategies based on latent variable modeling techniques in order to develop a multi-species classifier of the fresh/frozen–thawed status of fish samples using near-infrared spectra. Whereas the first two strategies model the information related to the species and to the fish together (either jointly or sequentially), the third strategy aims at explicitly separating them to improve the classification performance. The proposed strategies were validated over a database of more than 1200 samples of several different species, with near-infrared spectra collected with two different instruments. The overall classification accuracies ranged between 80% and 91%, according to the strategy and the instrument used. We believe that this study can contribute to the development of a species-independent approach to foodstuff classification. Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction The substitution of fresh fish with frozen–thawed fish is a typical fraud that not only damages consumers from an economical point of view, but can also cause safety issues (Pavlov, 2007). In fact, although freezing is one of the most widely used methods to extend the shelf life of seafood, it can affect the overall organoleptic properties of the product, and thawed meat is characterized by a higher susceptibility to microbial growth. Furthermore, fish authentication is important for correct product labeling (Martinez et al., 2003), as promoted by recent regulatory actions (Uddin, 2010; European Parliament Legislative Resolution, 2011). Several methods have been proposed for the identification of the fresh/frozen–thawed substitution fraud (e.g., eye lens evaluation, measurements of dielectric properties, erythrocytes lysis, hematocrit evaluation, muscles histology, enzymatic methods, etc.; Uddin, 2010). The classification ability of the majority of these systems is strongly affected by the species under investigation, the integrity of the product (whole fish or fillet) or by its shelf life (Ud⇑ Corresponding author. Tel.: +39 0498275470. E-mail address: [email protected] (P. Facco). 0260-8774/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.jfoodeng.2013.07.005

din, 2010). For example, the use of methods based on changes in dielectric properties, while being accurate on intact fish, provides poor results when applied on fillets (Duflos et al., 2002). Enzymatic assays were found to be useful in fillets, but not applicable to all species (Duflos et al., 2002). Recently, Bozzetta et al. (2012) proposed muscles histology as a simple method for the evaluation of the fresh/frozen–thawed status. Despite the good classification results obtained on a wide range of species (more than 35 different species), the method requires time for sample processing (e.g., fixation and coloration) and the use of several reagents. As an alternative to the abovementioned techniques, more rapid analytical technologies have been developed. Among them (Nott et al., 1999; Karoui et al., 2006; Vidacˇek et al., 2008; Fernández-Segovia et al., 2012; Leduc et al., 2012), near-infrared spectroscopy (NIRS) has been suggested by the promising results obtained on some species (Uddin, 2010; Sivertsen et al., 2011; Fasolato et al., 2012; Zhu et al., 2012; Kimiya et al., 2013; Ottavian et al., 2013). NIRS is a well consolidated analytical technology and plenty of applications can be found in the field of seafood authentication (Cozzolino and Murray, 2012). To the authors’ knowledge, there are currently no NIRS application to multi-species databases, i.e. so far the fresh/frozen–thawed authentication problem has been solved only analyzing single species separately.

766

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775

In this paper, we propose and compare three alternative strategies based on latent variable modeling techniques (Geladi and Kowalski, 1986; Trygg and Wold, 2002; Barker and Rayens, 2003) in order to develop a multi-species classifier of the fresh/frozen– thawed status of fish samples. While the first two strategies model the information on the species and on the fish status fresh/frozen– thawed together (either jointly or sequentially), the third strategy aims at explicitly separating them to improve the classification performance. A thorough validation of the proposed strategies is carried out using two NIR instruments exploring different spectral regions, on a total of more than 1200 samples.

Table 2 Available dataset in terms of number of samples per species, per class and per NIR instrument: validation sets V1 and V2 samples. Validation set

Species

Symbol

Number of samples

FOSS

UNITY

V1

Gilthead sea bream (Sparus aurata) (Fasolato et al., 2010b) Red mullet (Mullus barbatus) (Fasolato et al., 2010a) Sole (Solea vulgaris) (Fasolato et al., 2008)

Xsa

Fresh Frozen– thawed

27 27

U

U

Xmb

Fresh Frozen– thawed Fresh

27 27

U

U

35

U

-

Frozen– thawed Fresh Frozen– thawed

8 50 35

U

Ua

2. Materials and methods 2.1. Available dataset The number of samples available for model calibration and model validation (per species, class and instrument; see also Section 2.2) is given in Tables 1 and 2. The fresh/frozen–thawed classification models were built considering only the samples of the four species indicated in Table 1 (independently on the strategy used; Section 2.3). For model validation, instead, two datasets were considered, namely V1 and V2 (Table 2). The V1 spectra were collected at the same time of the calibration samples, whereas the V2 spectra were collected at a different time. For each species, the N spectra considered were collected into an Xsub [N  M] matrix, where M is the number of wavelengths (M = 401 for FOSS spectra and 421 for UNITY spectra) and subscript sub refers to the initial letter of the species Latin name. As an example, swordfish (Xiphias gladius L) FOSS samples were collected into Xxg [260  401]. Superscript  is used to identify the species which were used only for model validation (Table 2).

V2

2.2. NIR analysis For details about the origin, freezing, thawing and storage of the samples, the reader is referred to the references wherein they were originally presented (Fasolato et al., 2008, 2010a,b, 2012; Ottavian et al., 2012). As for the NIR analysis, the epaxial white muscles of fresh and frozen–thawed samples were minced using a Retsch Grindomix (Retsch GmbH, Hann, Germany) at 4000 rpm for 10 s. Two aliquots per sample were scanned in small ring cups in reflectance mode with two different instruments: a FOSS NIRSystem 5000 (FOSS NIRSystem Inc., Silver Spring, MD, USA) at 2 nm intervals from 1100 to 2500 nm; and a UNITY Scientific SpectraStar 2500TW (Unity Scientific, Columbia, MD, USA) at 1 nm intervals from 680 to 2500 nm. For each aliquot, a mean spectrum was

Table 1 Available dataset in terms of number of samples per species, per class and per NIR instrument: calibration samples.

a

Species

Symbol

Number of samples

Gilthead sea bream (Sparus aurata) (Fasolato et al., 2010b) Red mullet (Mullus barbatus) (Fasolato et al., 2010a)

Xsa

Fresh Frozen– thawed Fresh Frozen– thawed Fresh Frozen– thawed Fresh Frozen– thawed

Xmb

Sole (Solea vulgaris) (Fasolato et al., 2008)

Xsv

Swordfish (Xiphias gladius L) (Fasolato et al., 2012)

Xxg

FOSS

UNITY

53 53

U

U

53 53

U

U

71 17

U

-

101 74

U

Ua

Only 53 fresh and 53 frozen–thawed samples were available.

Xsv

Swordfish (Xiphias gladius L) (Fasolato et al., 2012)

Xxg

Gilthead sea bream (Sparus aurata) (Fasolato et al., 2010b) Red mullet (Mullus barbatus) (Fasolato et al., 2010a)

Xsa

Fresh Frozen– thawed

71 71

U

U

Xmb

Fresh

71

U

U

Frozen– thawed Fresh Frozen– thawed Fresh Frozen– thawed

71 71 71

U

U

– 38

U

-

Fresh Frozen– thawed Fresh Frozen– thawed

– 66

U

-

15

-

U

Swordfish (Xiphias gladius L) (Fasolato et al., 2010a) European sea bass (Dicentrarchus labrax) (Ottavian et al., 2012) Different speciesb

Xxg

Carp/tench (Cyprinus carpio/ Tinca tinca)

Xcct

Xdl

Xmix

a

Only 27 fresh and 27 frozen–thawed samples were available. Among the species included in Xmix: Sarda sarda, Pollachius virens, Scorpaena scrofa, Pangasius spp., Scomber scombrus, Gadus macrocephalus, Hippoglossus hippoglossus, etc. b

obtained by averaging multiple scans. Then, the spectrum of the sample was obtained by averaging those of the two aliquots. Reflectance (R) values were converted into absorbance (A) values through A = log(1/R). The analysis of the UNITY spectra was limited to the region between 680 and 1100 nm due to the noise characterizing the wavelength region above 1900 nm and considering that the NIR regions explored by the two instruments partially overlap. This was done in order to explore two different spectral regions with the available instruments. 2.3. Statistical analysis Principal component analysis (PCA; Jackson, 1991) was applied as an exploratory tool to reveal the internal correlation structure of the available datasets. This preliminary analysis was intended to identify the major sources of variability of the data (sample species, sample status, etc.). After the preliminary analysis, three alternative strategies were developed for sample classification, with the aim of developing a classifier for the fresh/frozen–thawed status of the samples independently from their species. A schematic of the three strategies is given in Fig. 1, while details on the modeling techniques are presented in the following subsections.

767

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775

X sa X mb

STRATEGY 1

PLS-DA

X=

Ystatus X sv X xg PLS-DA X sa

X sa X mb

STRATEGY 2

Ystatus PLS-DA

X mb

PLS-DA

Ystatus PLS-DA

Yspecies

X= X sv

X sv

Ystatus PLS-DA

X xg

X xg

Ystatus

step 2 (inner layer)

step 1 (outer layer) X sa X mb

STRATEGY 3

Orthogonal decomposition

PLS-DA X⊥

X=

+

X//

X//

Ystatus

X sv X xg

Fig. 1. Schematic of three strategies considered to build a species-independent fresh/frozen–thawed classification model. The sample status is either ‘‘fresh’’ or ‘‘frozen– thawed’’.

In the first strategy, a partial least-squares discriminant analysis (PLS-DA; Barker and Rayens, 2003) model to classify the fresh/frozen–thawed status of each sample was built considering the species altogether, i.e. calibrating the model on the matrix X obtained by stacking on the top of each other the matrices Xsa, Xmb, Xsv and Xxg of the calibration set (Table 1). In the second approach, a two-level cascade arrangement of PLS-DA models was developed: in the first level, a PLS-DA model classified the samples according to their species; in the second level, different PLS-DA models (one for each species considered in the calibration set) discriminated between fresh and frozen– thawed samples. In the third strategy, orthogonal PLS-DA (OPLS-DA; Trygg and Wold, 2002; Bylesjö et al., 2006) was used to remove the information in the spectral data which is not related to the fresh/frozen– thawed status of the samples. OPLS-DA decomposed matrix X into matrices X\ and X//, containing respectively the orthogonal (i.e. not correlated) and parallel (i.e. correlated) information concerned with the fresh/frozen–thawed status of the fish. Therefore, the information related to the species difference is removed from the spectra matrix before the fresh/frozen–thawed status classification is carried out. This enables the calibration of a species-independent fresh/frozen–thawed classifier (see Sections 3.4–5 for further details). It should be stressed that both Strategy 1 and Strategy 2 represent a multi-species fresh/frozen–thawed classifier, with the variability related to the species modeled together with the variability of the status (fresh vs. frozen–thawed). However, especially for Strategy 2, when a sample of a new species (i.e. a species not included in the calibration dataset) is analyzed, the classifica-

tion results are not reliable, since the PLS-DA model in the outer layer cannot correctly assess the sample species. Conversely, when Strategy 3 is applied, only the variability strictly related to the sample status is retained, whereas the variability related to the species is not modeled at all, facilitating a species-independent classification. In the calibration of the PLS-DA models of Strategy 1 and Strategy 2, four different spectra preprocessing techniques were considered, combining standard normal variate (SNV; Barnes et al., 1989) and first and second order derivatives (D-1 and D-2; Savitzky and Golay, 1964), namely: (i) no preprocessing at all; (ii) SNV; (iii) SNV and D-1; and (iv) SNV and D-2. For Strategy 3, since the OPLS itself can be considered a preprocessing step, no preprocessing techniques were applied. All strategies were implemented in MATLAB 7.11 (R2010b; the MathWorks Inc., Natick, MA) using the PLS Toolbox 7.0 (Eigenvector Research Inc., Manson, WA) and in-house developed codes. 2.3.1. Principal component analysis (PCA) Given a generic [N  M] X matrix, its PCA decomposition is given by

X ¼ TPTPCA þ EX

ð1Þ

with T [N  A], PPCA [M  A] and EX [N  M] being respectively the scores, loadings and residuals of the model built on A principal components (PCs), and superscript T indicating the transpose of a matrix. PCA summarizes the information stored in the X matrix by defining a low-dimensional space, whose axes (of which the A loadings PPCA are the direction cosines) represent the directions of max-

768

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775

imum variability of the original data. The scores T = [t1, t2, . . . , tA], i.e. the projections of X onto the latent space, represent the new variables.

T

X ¼ TP þ EX

ð2Þ

Y ¼ TQ T þ EY

ð3Þ

T ¼ XW

ð4Þ

where P [M  A] and Q [I  A] are the loadings relating the projections in the model space T to the data matrices X and Y, respectively. W [M  A] is the weight matrix, through which the predictor X are projected onto the latent space to give the scores T. EX [N  M] and EY [N  I] are the residual matrices, and account for the mismatch in the reconstruction of the original data. The NIPALS algorithm (Geladi and Kowalski, 1986; see Appendix A) was used to extract scores, loadings and weights. The number A of latent variables (LVs) to be retained was selected by maximizing the classification performance in terms of sensitivity, Se, and specificity, Sp (Se = % of samples of the ith class correctly classified; Sp = % of samples not belonging to the ith class correctly classified; Fawcett, 2006) in cross-validation of the calibration dataset (Wold, 1978). Furthermore, since the output of the PLS-DA model was not in the form of 0’s and 1’s, but was a real number, a threshold was chosen to define the class membership. Following a Bayesian approach (with the assumption that the predictions within each class are approximately normally distributed), the threshold value was determined in such a way as to return the best possible split among classes in cross-validation, with the least probability of false classification (Fawcett, 2006). The importance of each predictor variable m within the classification model was assessed by means of the variable importance in projection (VIP) index (Chong and Jun, 2005):

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi PA 2 a¼1 wm;a  SSY a  M VIPm ¼ SSYtotal  A

10 0 -10 -20 -30 -60

Sea bream Sole -40

-20

Red mullet Swordfish 0

20

40

60

Scores on PC1 (89.3%) Fig. 2. Scores t1 and t2 of the preliminary PCA model. Closed symbols: frozen– thawed samples. Open symbols: fresh samples.

where X\ and X// are the matrix of the information which is orthogonal to the property of interest Y, and the matrix that is parallel (i.e. correlated) to Y. Note that in the current application Y is the matrix characterizing the sample status, respectively. Several criteria have been proposed for the selection of the number of orthogonal components to remove (Trygg and Wold, 2002), namely: (i) monitoring the ratio between the norm of the orthogonal weight w\ and the norm of the PLS loadings p, (ii) cross-validation, and (iii) removing as many orthogonal components as necessary to obtain a one-component PLS model. Following Trygg and Wold (2002), in the present study a combination of these criteria was used (Section 3.4.1). 3. Results and discussion Preliminary analysis and classification results are presented in Sections 3.1–3.4 for FOSS spectra. Section 3.5 compares the three proposed strategies, whereas results for UNITY spectra are summarized in Section 3.6. Section 3.7 relates the results obtained with the proposed methods to those achieved with different methods presented in the literature. 3.1. Preliminary analysis

ð5Þ

where wm,a is the weight value for component a of variable m, SSYa is the sum of squares of explained variance for the ath component, and SSYtotal is the sum of squares explained of the matrix Y. As a rule of thumb, variables with VIP > 1 are considered to be of greater importance, whereas variables with VIP < 1 are less important. 2.3.3. Orthogonal projection to latent structures (OPLS) The orthogonal projection to latent structure (OPLS; Trygg and Wold, 2002) was originally introduced to remove the systematic variation from an input dataset X not correlated to the response Y, i.e. to remove variability in X that is not correlated (i.e., that is orthogonal) to Y. The OPLS method can be seen as either a preprocessing method or it can be made an integral part of regular PLS modeling. In fact, orthogonal components (or variables, OLV) can be obtained with a small modification of the NIPALS PLS method (Trygg and Wold, 2002), as shown in Appendix A. As mentioned earlier, the OPLS algorithm decomposes the (properly scaled) matrix X into:

X ¼ X== þ X?

20

Scores on PC2 (8.1%)

2.3.2. Partial least-squares discriminant analysis (PLS-DA) The memberships of each of the N available samples are encoded in matrix Y [N  I], where I = 2 for the classification according to the sample status, and I = 4 for the classification according to the sample species. The class attribution is expressed in the form of 0’s (the sample does not belong to the class being discriminated) and 1’s (the sample belongs to the class being discriminated). A PLS-DA model is given by:

30

ð6Þ

A 2-PC PCA model was calibrated on the matrix X obtained by stacking on the top of each other the matrices Xsa, Xmb, Xsv and Xxg of the calibration set (Table 1 and Fig. 1). No preprocessing was applied on the spectral data.1 The PCA model extracted more than 97% of the total variability (89.3% on PC1 and 8.1% on PC2). The model scores are shown in Fig. 2, which shows that the greatest source of variability of the data is given by the difference between species. In fact, samples of the same species, in fact, cluster in the same zone of the score space: sea bream samples are in the upper left region, whereas swordfish, mullet and sole samples are aligned along PC1 from the left to the right. Also the discrimination of interest (i.e. the fresh/frozen–thawed status) seems to insist mainly along PC1, with fresh and frozen–thawed samples partially overlapping. The overlap for samples of a given species can be attributed to the instrument used for spectra collection (FOSS, see Section 3.5), whereas the overlap for samples of different species is related to the collinearity between the information on the status and on the 1 Calibrating the PCA model on pretreated spectra did not change the conclusions drawn on Fig. 1. Increasing the number of preprocessing steps, in fact, only affected the distance among clusters, suggesting that pretreating the spectra does not improve the classification results.

769

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775 Table 3 Strategy 1 PLS-DA model statistics for different preprocessing techniques (calibration data). Spectra pretreatment

LVs

varY (%)

Secv (%)

Spcv (%)

No preprocessing SNV SNV and D-1 SNV and D-2

22 22 24 23

30.3 32.0 36.0 34.7

82.7 82.1 80.6 79.1

86.7 85.7 80.6 80.6

species. Another way of looking at the partial collinearity highlighted in Fig. 2 consists in the analysis of the PLS-DA models that classify the X samples according to their fresh/frozen–thawed status (Strategy 1) and to the species they belong to (Strategy 2). The angle between the first loadings of each model, in fact, was found to be 4°, i.e. the two models almost share the first latent variable (result not shown for the sake of conciseness). This means that there is a high collinearity between the major source of variability (first latent variable) of the two PLS-DA models, which are intended to capture the differences among samples due either to the status or to the species, but are built in such a way as to model the two sources of information jointly. Generally speaking, Fig. 2 suggests that it is possible to classify the samples according both to their species and to their status, but model complexity (in terms of number of retained LVs) would be high, due to the multiple sources of variability within the data. 3.2. Strategy 1 3.2.1. Model calibration Details on the classification model obtained with Strategy 1 are given in Table 3 for all the spectra preprocessing strategies explored. For each preprocessing, the number of LVs, the explained variance on the response matrix Y (varY, %), and the sensitivity and specificity (in cross-validation, Secv and Spcv) towards the fresh fish class are presented for the best classification model. Although the data in Table 3 show that the performance of the models obtained with no pretreatments and with SNV alone are similar, the former was preferred. The large number of LVs retained can be justified from the considerations drawn in the previous Section, as only a (small) fraction of the overall variability can be attributed to the fresh/frozen–thawed classification. The VIP index for the classification model (Appendix B) highlighted mainly wavelengths related to the absorbance of water and lipids (i.e. wavelengths around 1150, 1400, 1700, 1850–1900 and 2250–2400 nm). 3.2.2. Model validation The classification results for the validation datasets (Table 2) are presented in Table 4 in terms of misclassifications per class and overall accuracy (i.e., the percentage of correctly classified samples) for each species. The classification results on samples of species used in the calibration dataset were similar in validation sets V1 and V2, with an overall accuracy comparable with that shown in Table 3. An exception was represented by the Xsa (Sparus aurata) samples of the V2 dataset, whose sensitivity towards the fresh class was found to be poor (36.6%). It should be noted that similar results were obtained also with other classification strategies (see Tables 7, 9 and C2 in Appendix C, and also Fasolato et al., 2012). The fact that different rearing conditions (with respect to those of the calibration samples) may have induced a change in the proximate composition in terms of fat and water content is a possible explanation for the poor classification result obtained with Strategy 1. In fact, the fat and water contents can affect the spectral response. Unfortunately though, it was not possible to test this assumption, since proximate composition data were not available.

Table 4 Strategy 1 classification results for the validation datasets. Validation dataset

Species

V1

Xsa Xmb Xsv Xxg

V2

Xsa Xmb Xxg Xdl Xmix

Frozen–thawed samples misclassified

Classification accuracy

4 7 2 6

5 11 3 9

84.2 68.4 88.4 82.6

45 15 9 – –

5 17 9 0 22

64.8 77.5 87.3 100 66.7

Fresh samples misclassified

Although the PLS-DA model was calibrated on Xsa, Xmb, Xsv and Xxg samples, it was used to classify the Xdl and Xmix samples of validation set V2. First of all, the capability of the model to describe these samples was checked by means of the square prediction error (SPE) and Hotelling’s T2 statistics, where SPE is a measure of the representativeness of the model and T2 is a measure of the difference between the sample condition and the average spectrum. Both the SPE and the Hotelling T2 of each new sample were checked against the reference 95% confidence limits (SPElim and T 2lim ) obtained from the calibration data. It was found that both the Xdl and the Xmix samples showed an SPE value below the reference limits, but the Xdl samples had an Hotelling T2 value much higher than T 2lim . Hence, despite the model was found to be extrapolating (within the model space), the classification accuracy for sea bass samples was 100%. 3.3. Strategy 2 3.3.1. Model calibration Details on the models calibrated for Strategy 2 (Fig. 1) are given in Tables 5 and 6, respectively for the classification of the samples Table 5 Strategy 2 model statistics for the PLS-DA model discriminating among species for different preprocessing techniques (calibration data). Average sensitivities and specificities values are reported. Spectra pretreatment

LVs

varY (%)

Secv (%)

Spcv (%)

No preprocessing SNV SNV and D-1 SNV and D-2

22 22 20 16

70.0 70.7 71.0 70.9

100 99.9 100 100

100 99.8 99.8 100

Table 6 Strategy 2 models statistics for the PLS-DA discriminating among status for different preprocessing techniques (calibration data). Species

Spectra pretreatment

LVs

varY (%)

Secv (%)

Spcv (%)

Xsa

No preprocessing SNV SNV and D-1 SNV and D-2

9 11 5 5

35.3 41.3 35.4 40.4

86.8 92.5 86.8 83.0

84.9 92.5 86.8 79.2

Xmb

No preprocessing SNV SNV and D-1 SNV and D-2

13 5 8 3

41.6 39.6 43.4 39.22

88.7 81.1 86.8 86.8

90.6 83.0 92.5 86.8

Xsv

No preprocessing SNV SNV and D-1 SNV and D-2

10 8 3 4

30.6 25.5 22.3 25.9

88.2 88.2 88.2 82.4

98.6 97.2 91.5 93.0

Xxg

No preprocessing SNV SNV and D-1 SNV and D-2

15 11 8 4

38.9 34.5 37.2 28.4

89.0 89.0 86.3 79.2

91.1 89.1 89.1 78.1

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775

according to the species and according to the status (one model for each species). For the PLS-DA model discriminating among species (Table 5), average sensitivities and specificities for the four species of the calibration dataset are reported. With respect to the model discriminating among species (Table 5), spectra preprocessing only affected the number of the LVs to be retained, whereas the explained variance on Y and the sensitivities and specificities did not change across preprocessing. With the aim of minimizing the preprocessing operations, the model obtained with no preprocessing at all was selected. The VIP index of the model (Appendix B) pointed to the water and lipid contents as the major factors causing differences among the species. Note that the possibility of classifying fish (as well as meat) samples according to their species using VIS–NIR spectra has already been discussed in the literature (Cozzolino et al., 2002; Mamani-Linares et al., 2012). As for the species-tailored PLS-DA models that discriminate for the fish status, no preprocessing was used for all species except for Xsa, for which SNV was preferred.2 A comparison between Tables 6 and 3 shows that, as expected, the sensitivities and specificities towards the fresh class are higher than those obtained with Strategy 1. The VIP indexes for the PLS-DA models of Tables 5 and 6 are shown in Appendix B. 3.3.2. Model validation The accuracy of the PLS-DA model discriminating for the fish species (outer layer) was found to be almost 100% on the validation sets V1 and V2, with only four misclassifications on the Xsa, Xmb, Xsv and Xxg samples. The fish status classification results for the same validation datasets (inner layer) are presented in Table 7. The effect of having a PLS-DA model tailored on each species can be clearly appreciated from Table 7, since for the species used also in the calibration step the classification accuracy is greater with respect to that shown in Table 4 for Strategy 1, with the exception of Xmb for V2 (only one misclassification). As for the Xdl and Xmix samples, the species attribution obtained from the species PLS-DA model was necessarily erroneous, as these species were not included in the calibration dataset. The capability of the model to adequately describe the samples was poor, i.e. SPE was higher than the respective confidence limits. Nevertheless, in order to define which fresh/frozen–thawed classification model use, samples were assigned to the class towards which they showed the highest probability of attribution. For several samples, however, the class membership was not clear, since the probability of assigning the sample to a specific class was below 5% for all classes (i.e., species). The majority of the samples for which a class was clearly defined was assigned mainly to the sea bream and swordfish classes.

25 no preprocessing

Number of latent variables

770

2

Again, it was noticed that derivatives did not improve the classification accuracy, as they just affected the model structure.

SNV and D1 SNV and D2 15

10

5

0 0

5

10

15

20

25

30

Number of orthogonal latent variables Fig. 3. Effect of the number of orthogonal latent variables removed on the final PLSDA model structure (in terms of number of latent variables retained).

Table 7 Strategy 2 fish status classification results for the validation datasets. Validation dataset

Species

Fresh samples misclassified

Frozen–thawed samples misclassified

Classification accuracy

V1

Xsa Xmb Xsv Xxg

5 3 1 5

2 6 0 15

87.0 83.3 97.7 76.7

V2

Xsa Xmb Xxg Xdl Xmix

24 26 9 -

4 7 5 2 40

80.3 76.8 90.1 94.7 39.4

Table 8 Strategy 3 OPLS-DA model details (calibration data). Spectra pretreatment

OLVs

OvarX (%)

varY (%)

Secv (%)

Spcv (%)

No preprocessing No preprocessing

22 28

84.1 92.1

29.8 31.8

90.0 93.9

92.1 94.2

Table 9 Strategy 3 classification results for the validation datasets. Fresh samples misclassified

Frozen–thawed samples misclassified

Classification accuracy

Validation dataset

Species

V1

Xsa Xmb Xsv Xxg

0 3 0 1

4 7 1 2

92.6 81.5 97.7 96.5

V2

Xsa Xmb Xxg Xdl Xmix

31 15 6 – –

2 15 6 1 12

76.8 78.9 91.5 97.4 81.8

3.4. Strategy 3 3.4.1. Model calibration As mentioned in Section 2.3.3, the selection of the number of orthogonal components to remove should be derived from the comparison of the results of multiple criteria (Trygg and Wold, 2002). Fig. 3 presents the effect of the number of OLVs removed on the number of LVs retained for the PLS-DA model, as chosen by crossvalidation. Fig. 3 clearly shows the effect of the pretreating the spectra prior to the OPLS analysis: spectra preprocessing removes undesired variability and therefore reduces the number of OLV to be

SNV

20

removed in order to obtain an 1-LV classification model. The model obtained with no preprocessing on the spectra was used.3 According to Fig. 3, 22 OLVs need to be removed to obtain a model with 1 LV (note the consistency of this results with the number of LV retained 3 The variability removed by SNV and/or D1 and D2 was deemed as ‘‘uneffective’’, since the classification accuracy of the final model was similar with and without preprocessing.

771

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775 Table 10 Classification accuracy in validation for the three proposed strategies. Strategy

1 2 3

V1

V2

Total

Misclassified

Classification accuracy

Misclassified

Classification accuracy

Misclassified

Classification accuracy

47 37 18

80.2 84.4 92.4

122 117 89

77.0 77.9 83.2

169 154 107

77.8 79.9 86.0

in the PLS-DA model of Strategy 1). The final model was eventually built by removing 28 OLVs, since an improvement in the sensitivity and specificity (in cross-validation) was observed when slightly more than 22 OLVs were removed. Details on the models with 22 and 28 OLVs are given in Table 8. OvarX in Table 8 represents the percentage of the variability of X removed by the OPLS algorithm. It can be concluded that only a very small fraction of the original variability was retained (and was therefore useful) for the fresh/frozen–thawed classification, as anticipated from the preliminary analysis. A comparison between the sensitivities and specificities towards the fresh class (in cross-validation) of the three strategies (Tables 3, 5, 6 and 8) shows that the accuracy of Strategy 3 is greater than that of Strategies 1 and 2. An additional advantage of the OPLS algorithm is the possibility of analyzing separately X\ and X//. The VIP index obtained from the PLS-DA model calibrated on X\ to separate the samples according to their species (not shown) clearly resembles the one obtained for the outer PLS-DA model of Strategy 2. This is reasonable, because the PLS-DA model calibrated on X\ (no preprocessing, 17 LVs, 64.0% of explained variance on Y; Table 5) resembles the outer one calibrated for Strategy 2, i.e. the variability removed from the OPLS algorithm contains information related to the species. As a further confirmation, by calibrating a PLS-DA model on X//, it was verified that little information about the samples species was retained in X//: the model was found to explain only 3% of the variability on Y, hence having poor discriminating capability. A drawback of pretreating the spectra with the OPLS algorithm was observed in the VIP index of the OPLS-DA model. Since the portion of the original X variance retained for the fresh/frozen– thawed classification is small, all wavelengths showed VIP values very close to one. Thus, it was not possible to identify specific spectral regions responsible for the discrimination, since the information retained in X// for the classification of interest is related to the correlation among all the wavelengths of the spectra.

greatest part of the variability within the calibration data X is not related to the classification of interest (i.e. the fresh/frozen– thawed status). All the proposed strategies represent a fairly accurate (within the limitation of the FOSS spectral range) multi-species classification approach, and these results can open the route to the development of a species-independent approach to fresh/frozen–thawed fish classification. We stress the fact that all strategies were tested on samples of species not included within the calibration datasets and, although the models were found to be somewhat extrapolating (within the model plane), classification accuracies were generally good.

3.4.2. Model validation The classification results for the validation datasets are presented in Table 9. Classification results of Strategy 3 are generally good (Tables 4 and 7). As happened for Strategies 1 and 2, for the V2 dataset the Xdl samples were found to be far from the center of the model plane (large value of the Hotelling T2 statistic), and hence the results should be taken with caution.

Since several applications of fish sample classification according to the fresh/frozen–thawed status can be found in the literature, this Section compares the classification accuracies obtained using the methods proposed in this study compared to those achieved with other methods reported in the literature. Spectroscopic techniques have been applied to fish muscles, exuded juice or dry meat extract (Uddin et al., 2005; Karoui et al., 2006, 2007; Vidacˇek et al., 2008; Uddin, 2010; Sivertsen et al., 2011; Fernández-Segovia et al., 2012; Leduc et al., 2012; Fasolato et al., 2012; Zhu et al., 2012; Ottavian et al., 2013; Kimiya et al., 2013). The reported classification accuracy ranges between 85% and 96–100%, according not only to the instrument considered, but also to the classification strategy employed. The results presented in Table 10 for the FOSS instrument are generally worse than those reported by other authors, but the reason for the difference observed lays mainly in the spectral region used for the discrimination (NIR region above 1100 nm vs. VIS/NIR region below 1100 nm), as the results of the UNITY spectra seem to confirm. Additionally, many of the investigations cited above were meant to be feasibility studies, namely the results were obtained using

3.5. Comparison among the proposed strategies Table 10 provides a compact summary of the classification accuracy in validation for the three proposed strategies. Clearly, Strategy 3 outperforms Strategies 1 and 2. As reported by some authors (Trygg and Wold, 2002; Svensson et al., 2002), the introduction of the OPLS algorithm is expected to be beneficial in terms of interpretation of the results, but it should not affect the overall accuracy. Stated differently, the performance of Strategies 1 and 3 are expected to be similar. The difference in terms of classification accuracy observed in Table 10 may be related to the fact that the

3.6. Results for UNITY spectra All the results presented so far were obtained using a FOSS instrument to obtain NIR spectra. Results for UNITY spectra are reported in Appendix C. With respect to results of FOSS spectra, classification accuracies were found to be higher (82.1% for Strategy 1, 91.0% for Strategy 2 and 88.4% for Strategy 3), confirming the conclusion drawn in Section 3.5. The main difference with the spectral region considered previously is the possibility of highlighting some wavelengths that are responsible for the fresh/frozen–thawed classification, i.e. some markers of the fish status. It was found that the region above 950 nm provided the greatest contribution within the classification model of Strategy 3 (particularly the regions 970–1030 nm and 1060–1088 nm). The region 950–1100 nm was reported as informative for frozen–thawed status on pork meat (Park et al., 2001), and it is mainly related to the modification of the water content and NH2 compounds. Additionally, changes on the absorbance at 970 nm (second overtone of O–H stretching) were observed also in Psetta maxima fillets using NIR imaging (Sivertsen et al., 2011). 3.7. Comparison with other studies

772

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775

a limited number of samples (generally less than one hundred) having relatively limited variability, and only cross-validation statistics are reported. The overall classification accuracy obtained in the present study with UNITY spectra (680–1100 nm) is comparable with that of other studies (88.5% and 91.0%, respectively for Strategies 2 and 3, with values up to 97.0% for some species), but a much larger validation dataset was used in the present study. An interesting comparison can be made with the histologybased classification proposed by Bozzetta et al. (2012), both because they used part of the same dataset and because it is one of the very few examples of a multi-species approach reported in the open literature. Generally, the histology-based method achieved higher classification accuracies than those presented here, but it must be emphasized that it relied on the operator’s experience and required reagents and sample preparation. Additionally, when new (i.e. different from those used in the calibration step) species were introduced, the resulting classification accuracy was reduced.

Emanuele Teneggi and Elena Maria Cencetti (Istituto Zooprofilattico Sperimentale del Piemonte, Liguria e Valle d’Aosta, Turin, Italy) for the samples and Dr. Lorenzo Serva and Massimo Mirisola for carrying out the NIR measurements. Appendix A. OPLS algorithm (Trygg and Wold, 2002) The OPLS algorithm for a single column Y response matrix (i.e. I = 1) and an X predictor matrix (each properly scaled) is given by:

MO and MB gratefully acknowledge Fondazione CARIPARO (Project # PARO104725 – 2010) for the financial support. The authors would like to thank Valentina Tepedino (Eurofishmarket),

ðA:2Þ

t ¼ Xw

u¼

ðA:3Þ

tT Y tT t

ðA:4Þ

Yq qT q

ðA:5Þ

tT X tT t

ðA:6Þ

4. Conclusions

Acknowledgements

ðA:1Þ

YT Y

w wT w

w¼

qT ¼

This paper addressed the problem of developing a multi-species fresh/frozen–thawed classification model from NIR spectral data for fish samples through three different strategies based on latent variable modeling techniques. In the first strategy, a PLS-DA model was built by concatenating vertically the spectra of samples of different species. In the second strategy, a cascade arrangement was proposed, where first (outer layer) a PLS-DA model separated the samples according to their species, and then (inner layer) a PLSDA model tailored for each species classified the samples according to their status. In the third strategy, OPLS-DA was used to remove the variability in the data that is not related to the fresh/frozen– thawed status of the samples. Since the variability removed was found to be mainly related to the samples species, the variability retained for the classification represented a species-independent information of the fish status. The three strategies were tested on a very large database of spectra of two NIR instruments exploring different spectral regions, respectively from 680 to 1100 nm and from 1100 to 2500 nm, and using also samples of species not included in the calibration data. Strategies 2 and 3 returned the best validation classification accuracies, with values of 91% and 88.4%, and of 80% and 86%, respectively. This study demonstrated the effectiveness of NIR spectroscopy as a screening method for fresh/frozen–thawed fish control, showing the possibility of working on a multi-species database (considering also species not included in the calibration data) without tailoring the classification model on a specific species (as traditionally proposed in the literature). However, it should be noted that the applicability of the proposed strategies to samples of unknown species should be checked on a case-by-case basis and subject to further investigation, as the models were found to extrapolate the results. The interpretation of the spectral data, though not always possible, highlighted the wavelengths associated to the water absorbance as those responsible for the classification of interest, in agreement with the results reported in the literature. We believe that this study can contribute to the development of a species-independent approach to foodstuff classification.

YT X

wT ¼

pT ¼

w? ¼ p 

w? ¼



 wT p w wT w

w? wT? w?

ðA:7Þ

ðA:8Þ

t? ¼ Xw?

ðA:9Þ

tT? X tT? t?

ðA:10Þ

pT? ¼

EOPLS ¼ X  t? pT?

ðA:11Þ

Steps 1–6 represent the normal NIPALS algorithm (Geladi and Kowalski, 1986). In (A.11), EOPLS represents the residual after a portion of orthogonal variability has been removed. For additional orthogonal components w\, p\, t\, the algorithm needs to be repeated from step 3 by setting X = EOPLS .u in (A.5) represents the scores of the response matrix Y.

Appendix B. VIP index for the PLS-DA classification models The VIP index for the PLS-DA model of Strategy 1 is given in Fig. B1. Fig. 4 shows the existence of several spectral regions with a VIP index greater than 1, and thus relevant for the classification model, namely around 1150, 1400, 1700, 1850–1900 and 2250– 2400 nm. Some of these regions can be linked to the absorbance of water (1150, 1400, 1850–1900 and 2250 nm). Park et al. (2001) described the contribution of water absorption at 1153 nm in thawed beef longissimus muscles. The bands around 1400 nm (particularly 1410, 1435, 1450, and 1485 nm; first overtones of the O–H/N–H stretching modes) have been linked to an increase in free water species during thawing and to water-bonded groups (Liu and Chen, 2001). Water absorption (the combination of the O–H stretching band and the O–H bending band) has been reported around 1900 nm in several food matrices (Büning-Pfaue, 2003). Several authors (Uddin et al., 2005; Fasolato et al., 2012; Ottavian et al.,

773

5

5

4

4

3

VIP index

VIP index

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775

2

1

Sea bream Red mullet Sole Swordfish

3

2

1

0 1200

1400

1600

1800

2000

2200

0

2400

Wavelength (nm)

1200

1400

1600

1800

2000

2200

2400

Wavelength (nm)

Fig. B1. VIP index for the PLS-DA model of Strategy 1 (calibration data). The dashdotted line red line represent the threshold value (VIP = 1). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. B2. VIP index for the PLS-DA model of Strategy 2 that classifies the samples according to their species (calibration data). The dash-dotted line red line represent the threshold value (VIP = 1).

2013) highlighted wavelengths related to the absorbance of water as responsible for the fresh/frozen–thawed classification both in VIS/NIR and NIR spectra. The increase of liquids release during thawing (e.g. due to the breakdown of the cell and protein denaturation), in fact, affects the overall absorbance level due to scattering effects (Fasolato et al., 2012; Ottavian et al., 2013). As reported on intact and ground cutlets of swordfish (Fasolato et al., 2012) and in other muscle matrices (Barbin et al., 2012, 2013), a lower absorbance was detected for fresh samples compared to frozen–thawed ones. It can be concluded that the overall water content and the different behavior of water species in connection with other constituents, such as proteins, were probably responsible for the selection of the abovementioned wavelength regions.

The regions around 1700 nm and 2250–2400 nm were mainly associated to the C–H groups of components such as lipids, amino and fatty acids (Cozzolino et al., 2009). Namely, the increase of the C–H modes may be caused by the relaxation of meat lipids originated during the thawing progression (Barbin et al., 2013). Figs. B2 and B3 report the VIP indexes for the PLS-DA models of Strategy 2, respectively for the outer and the four inner layer classifiers. With respect to Fig. B3, although the results are similar for the four species, the peaks in the VIP index plot of the sea bream model seem to be slightly shifted, both for the water contribution in the region around 1400 nm and for the protein contribution in the region around 1500 nm. A possible explanation for this behavior might be related to a difference in the proximate composition

(a)

(b)

5

4

VIP index

VIP index

4 3 2 1

2

0 1200 1400 1600 1800 2000 2200 2400

1200 1400 1600 1800 2000 2200 2400

Wavelength (nm)

Wavelength (nm)

5

(d)

4

5 4

VIP index

VIP index

3

1

0

(c)

5

3 2 1

3 2 1

0

0 1200 1400 1600 1800 2000 2200 2400

1200 1400 1600 1800 2000 2200 2400

Wavelength (nm)

Wavelength (nm)

Fig. B3. VIP index for the PLS-DA models of Strategy 2 that classify the samples of a species according to their status. (a) Xsa, (b) Xmb, (c)Xsv, and (d)Xxg (calibration data). The dash-dotted line red line represent the threshold value (VIP = 1). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

774

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775

Table C1 UNITY models calibration details for the three proposed strategies. Strategy

Model

Calibration data

Spectra pretreatment

Model structure

varY (%)

Secv (%)

Spcv (%)

1

Fresh/frozen–thawed PLS-DA

X

No preprocessing

18 LV

38.8

86.8

89.3

2

Species PLS-DA Fresh/frozen–thawed PLS-DA Fresh/frozen–thawed PLS-DA Fresh/frozen–thawed PLS-DA

X Xsa Xmb Xxg

No preprocessing SNV No preprocessing No preprocessing

17 LV 11 LV 9 LV 13 LV

63.4 44.8 41.4 44.5

100 98.1 98.1 99.4

100 94.3 100 100

3

Fresh/frozen–thawed OPLS-DA

X

No preprocessing

14 OLV and 1 LV

41.7

100

98.1

Table C2 UNITY classification results for the validation datasets. Strategy

Validation dataset

Species

Fresh samples misclassified

Frozen– thawed samples misclassified

Classification accuracy

1

V1

Xsa Xmb Xxg Xsa Xmb Xxg Xcct

9 2 2 20 15 6 2

4 7 0 29 0 12 –

75.9 83.3 96.3 65.5 89.4 87.3 86.7

Xsa Xmb Xxg Xsa Xmb Xxg Xcctt

3 2 2 18 2 11 3

1 1 0 3 1 7 –

92.6 94.4 96.3 85.2 97.9 87.3 80.0

Xsa Xmb Xxg Xsa Xmb Xxg Xcctt

5 0 3 24 7 7 3

5 5 0 4 2 7 –

81.5 90.7 94.4 80.3 93.7 90.1 80.0

V2

2

V1

V2

3

V1

V2

between fresh and frozen–thawed samples. However, this is not a definitive conclusion and additional investigation is required on this issue. Appendix C. Results for UNITY spectra Results for UNITY spectra are summarized in Tables C1 and C2. References Barbin, D.F., Elmasry, G., Sun, D.-W., Allen, P., 2012. Near-infrared hyperspectral imaging for grading and classification of pork. Meat Science 90, 259–268. Barbin, D.F., Sun, D.-W., Su, C., 2013. NIR hyperspectral imaging as a non-destructive evaluation tool for the recognition of fresh and frozen–thawed porcine longissimus dorsi muscles. Innovative Food Science and Emerging Technologies. http://dx.doi.org/10.1016/j.ifset.2012.12.011. Barker, M., Rayens, W., 2003. Partial least-squares for discrimination. Journal of Chemometrics 17, 166–173. Barnes, R.J., Dhanoa, M.S., Lister, S.J., 1989. Standard normal variate transformation and de-trending of near-infrared diffuse reflectance spectra. Applied Spectroscopy 43, 727–777. Bozzetta, E., Pezzolato, M., Cencetti, E., Varello, K., Abramo, F., Mutinelli, F., Ingravalle, F., Teneggi, E., 2012. Histology as a valid and reliable tool to differenziate fresh from frozen–thawed fish. Journal of Food Protection 75, 1536–1541. Büning-Pfaue, H., 2003. Analysis of water in food by near infrared spectroscopy. Food Chemistry 82, 107–115. Bylesjö, M., Rantalainen, M., Cloarec, O., Nicholson, J.K., Holmes, E., Trygg, J., 2006. OPLS discriminant analysis: combining the strengths of PLD-DA and SIMCA classification. Journal of Chemometrics 20, 341–351. Chong, I.-G., Jun, C.H., 2005. Performance of some variable selection methods when multicollinearity is present. Chemometrics and Intelligent Laboratory Systems 78, 103–112.

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 Review 47, 207–218. Cozzolino, D., Murray, I., Scaife, J.R., 2002. Near infrared reflectance spectroscopy in the prediction of chemical characteristics of minced raw fish. Aquaculture Nutrition 8, 1–6. Cozzolino, D., Restaino, E., La Manna, A., Fernandez, E., Fassio, A., 2009. Usefulness of near infrared reflectance (NIR) spectroscopy and chemometrics to discriminate between fishmeal, meat meal and soya meal samples. Ciencia e investigación agraria 36, 209–214. Duflos, G., Le Fur, B., Mulak, V., Becel, P., Malle, P., 2002. Comparison of methods of differentiating between fresh and frozen–thawed fish or fillets. Journal of the Science of Food and Agriculture 82, 1341–1345. European Parliament Legislative Resolution of July, 6, 2011 on the Provision of Food Information to Consumers, amending Regulations (EC) No 1924/2006 and (EC) No 1925/2006 and repealing Directives 87/250/EEC, 90/496/EEC, 1999/10/EC, 2000/13/EC, 2002/67/EC, 2008/5/EC and Regulation (EC) No 608/2004 (17602/ 1/2010 – C7-0060/2011 – 2008/0028(COD). Fasolato, L., Manfrin, A., Corrain, C., Perezzani, A., Arcangeli, G., Rosteghin, M., Novelli, E., Lopparelli, L.M., Balzan, S., Mirisola, M., Serva, L., Segato, S., Bianchi, E., 2008. Assessment of quality parameters and authentication in sole (Solea vulgaris) by NIRS (near infrared reflectance spectroscopy). Industrie Alimentari 47, 355–361. Fasolato, L., Balzan, S., Valentini, K., Ferlito, J.C., Riovanto, R., Mirisola, M., Cencetti, E., Serva, L., Benozzo, F., Teneggi, E.M., Berzaghi, P., Novelli, E. 2010a. Nondestructive non touch visible-NIR transmittance spectroscopy for identification of fresh and frozen–thawed fish. In: Proceeding of the 4th Conference NIR on The Go. Fasolato, L., Cencetti, E., Riovanto, R., Mirisola, M., Novelli, E., Balzan, S., Serva, L., Ferlito, J.C., Benozzo, F., Teneggi, M.E., Berzaghi, P. 2010b. Validation of near infrared spectroscopy analysis in authentication of fresh and frozen–thawed fish products. In: Proceedings of the 14th International Conference on NIR Spectroscopy. Fasolato, L., Balzan, S., Riovanto, R., Berzaghi, P., Mirisola, M., Ferlito, J.C., Serva, L., Benozzo, F., Passera, R., Tepedino, V., Novelli, E., 2012. Comparison of visible and near-infrared reflectance spectroscopy to authenticate fresh and frozen– thawed swordfish (Xiphias gladius L). Journal of Aquatic Food Product Technology 21, 493–507. Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recognition Letters 27, 861–874. Fernández-Segovia, I., Fuentes, A., Aliño, M., Masot, R., Alcañiz, M., Barat, J.M., 2012. Detection of frozen–thawed salmon (Salmo salar) by a rapid low-cost method. Journal of Food Engineering 113, 210–216. Geladi, P., Kowalski, R., 1986. Partial least squares regression: a tutorial. Analytica Chimica Acta 185, 1–17. Jackson, J.E., 1991. A User’s Guide to Principal Components. John Wiley & Sons, New York, USA. Karoui, R., Thomas, E., Dufour, E., 2006. Utilisation of a rapid technique based on front-face fluorescence spectroscopy for differentiating between fresh and frozen–thawed fish fillets. Food Research International 39, 349–355. Karoui, R., Lefur, B., Grondin, C., Thomas, E., Demeulemester, C., De Baerdemaeker, J., Guillard, A.-S., 2007. Mid-infrared spectroscopy as a new tool for the evaluation of fish freshness. International Journal of Food Science and Technology 42, 57– 64. Kimiya, T., Sivertsen, A.H., Heia, K., 2013. VIS/NIR spectroscopy for non-destructive freshness assessment of Atlantic salmon (Salmo salar L) fillets. Journal of Food Engineering 116, 758–764. Leduc, F., Krzewinski, F., Le Fur, B., N’Guessan, A., Malle, P., Kol, O., Duflos, G., 2012. Differentiation of fresh and frozen/thawed fish, European sea bass (Dicentrarchus labrax), gilthead seabream (Sparus aurata), cod (Gadus morhua) and salmon (Salmo salar), using volatile compounds by SPME/GC/MS. Journal of the Science of Food and Agriculture 92, 2560–2568. Liu, Y., Chen, Y.-R., 2001. Two-dimensional visible/near-infrared correlation spectroscopy study of thawing behavior of frozen chicken meats without exposure to air. Meat Science 57, 299–310. Mamani-Linares, L.W., Gallo, C., Alomar, D., 2012. Identification of cattle, llama and horse meat by near infrared reflectance or transflectance spectroscopy. Meat Science 90, 378–385. Martinez, I., Aursand, M., Erikson, U., Singstad, T.E., Veliyulin, E., Van der Zwaag, C., 2003. Destructive and non-destructive analytical techniques for authentication and composition analyses of foodstuffs. Trends in Food Science and Technology 14, 489–498.

M. Ottavian et al. / Journal of Food Engineering 119 (2013) 765–775 Nott, K.P., Evans, S.D., Hall, L.D., 1999. Quantitative magnetic resonance imaging of fresh and frozen–thawed trout. Magnetic Resonance Imaging 17, 445–455. Ottavian, M., Facco, P., Fasolato, L., Novelli, E., Mirisola, M., Perini, M., Barolo, M., 2012. Use of near-infrared spectroscopy for fast fraud detection: application to the authentication of European sea bass (Dicentrarchus labrax). Journal of Agricultural and Food Chemistry 60, 639–648. Ottavian M., Fasolato, L., Serva, L., Facco, P., Barolo, M., 2013. Data fusion for food authentication: fresh/frozen–thawed discrimination in west African goatfish (Pesudupeneus prayensis) fillets, Food and Bioprocess Technology, http:// dx.doi.org/10.1007/s11947-013-1157-x. Park, B., Chen, Y.-R., Hruschka, W.R., Shackelford, S.D., Koohmaraie, M., 2001. Principal component regression of near-infrared reflectance spectra for beef tenderness prediction. Transactions of the ASAE 44, 609–615. Pavlov, A., 2007. Changes in the meat from aquaculture species during storage at low temperature and attempts for differentiation between thawed-frozen and fresh chilled meat. A review. Bulgarian Journal of Veterinary Medicine 10, 67– 75. Savitzky, A., Golay, M.J.E., 1964. Smoothing and differentiation of data by simplified least squares procedures. Analytical Chemistry 36, 1627–1639. Sivertsen, A.H., Kimiya, T., Heia, K., 2011. Automatic freshness assessment of cod (Gadus morhua) fillets by VIS/NIR spectroscopy. Journal of Food Engineering 103, 317–323.

775

Svensson, O., Kourti, T., MacGregor, J.F., 2002. An investigation of orthogonal signal correction algorithms and their characteristics. Journal of Chemometrics 16, 176–188. Trygg, J., Wold, S., 2002. Orthogonal projections to latent structures (O-PLS). Journal of Chemometrics 16, 119–128. Uddin, M., 2010. Differentiation of fresh and frozen–thawed fish. In: Leo, M., Nollet, L.M.L., Toldrá, F. (Eds.), Handbook of Seafood and Seafood Product Analysis. CRC Press, Boca Raton, USA. Uddin, M., Okazaki, E., Torza, S., Yumiko, Y., Tanaka, M., Fukuda, Y., 2005. Non destructive visible/NIR spectroscopy for differentiation of fresh and frozen– thawed fish. Journal of Food Science 70, 506–510. Vidacˇek, S., Medic´a, H., Botka-Petrakb, K., Nezˇakc, J., Petrak, T., 2008. Bioelectrical impedance analysis of frozen sea bass (Dicentrarchus labrax). Journal of Food Engineering 88, 263–271. Wold, S., 1978. Cross-validatory estimation of number of components in factor and principal components models. Technometrics 20, 397–405. Zhu, F., Zhang, D., He, Y., Liu, F., Sun, D.-W., 2012. Application of visible and near infrared hyperspectral imaging to differentiate between fresh and frozen– thawed fish fillets. Food and Bioprocess Technology. http://dx.doi.org/10.1007/ s11947-012-0825-6.