- Email: [email protected]
Potential of multispectral imaging combined with chemometric methods for rapid detection of sucrose adulteration in tomato paste
Accepted Manuscript Potential of multispectral imaging combined with chemometric methods for rapid detection of sucrose adulteration in tomato paste
Changhong Liu, Guang Hao, Min Su, Ying Chen, Lei Zheng PII:
S0260-8774(17)30324-2
DOI:
10.1016/j.jfoodeng.2017.07.026
Reference:
JFOE 8968
To appear in:
Journal of Food Engineering
Received Date:
09 April 2017
Revised Date:
21 July 2017
Accepted Date:
24 July 2017
Please cite this article as: Changhong Liu, Guang Hao, Min Su, Ying Chen, Lei Zheng, Potential of multispectral imaging combined with chemometric methods for rapid detection of sucrose adulteration in tomato paste, Journal of Food Engineering (2017), doi: 10.1016/j.jfoodeng. 2017.07.026
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 1
Potential of multispectral imaging combined with chemometric methods for
2
rapid detection of sucrose adulteration in tomato paste
3
Running title: Rapid detection of sucrose adulteration in tomato paste
4 5
Changhong Liu a,b,‡, Guang Hao a,‡, Min Su c, Ying Chen d,*, Lei Zheng a,b,*
6 7 8 9 10
a School of Food Science and Engineering, Hefei University of Technology, Hefei 230009, China b Key Laboratory for Agricultural Products Processing of Anhui Province, Hefei 230009, China
11
c Xinjiang Entry-Exit Inspection and Quarantine Bureau, Urumqi, 830063, China
12
d Agro-product Safety Research Centre, Chinese Academy of Inspection and
13
Quarantine, Beijing, 100123, China
14 15 16
‡ These authors contribute equally to this work.
17
* Corresponding authors: Lei Zheng, Tel: +86-551-62919398.
18 19
E-mail:
[email protected],
[email protected] (Y Chen).
[email protected]
(L
Zheng);
ACCEPTED MANUSCRIPT 21
Abstract: Confirmation of the authenticity of tomato paste is an increasing
22
challenger for food processors and regulatory authorities. This study focuses on the
23
rapid qualitative and quantitative detection of sucrose adulteration in tomato paste
24
using multispectral imaging (405-970 nm) combined with chemometric methods.
25
Partial least squares (PLS), least squares-support vector machines (LS-SVM), and
26
back propagation neural network (BPNN) were used to develop quantitative
27
models. Compared with PLS and BPNN, LS-SVM considerably improved the
28
prediction performance with coefficient of determination in prediction ( RP2 ) of
29
0.936 and 0.966, the root-mean-square error of prediction (RMSEP) of 0.521% and
30
0.445%, and residual predictive deviation (RPD) of 5.014 and 5.865 for batch 1 and
31
batch 2 of tomato paste, respectively. Besides, multispectral imaging was feasible to
32
detect sucrose adulteration in tomato paste at very low proportions (1%) with no
33
misclassification using chemometric methods. It was concluded that multispectral
34
imaging has an excellent potential for rapid determination of sucrose adulteration in
35
tomato paste.
36
Keywords: Multispectral imaging; Tomato paste; Sucrose; Rapid detection;
37
Chemometric methods
ACCEPTED MANUSCRIPT 39
1. Introduction
40
Tomato fruit (Solanum lycopersicum) is the second most important vegetable
41
crop worldwide which is consumed fresh as well as in the form of processed
42
products, such as tomato juice, tomato puree, tomato sauce, ketchup, and tomato
43
paste. Tomato paste is the most commonly consumed processed tomato product,
44
which is concentrated from fresh tomatoes. There is an increasing demand of
45
tomato paste consumption due to the beneficial health effects (Rizwan et al., 2011;
46
Burton‐Freeman et al., 2012). Therefore, tomato paste is an essential food in
47
people’s daily life in European and American countries, and becomes more and
48
more widely accepted in many other countries such as China.
49
Tomato paste is usually stored and used as an intermediate product with water
50
and other ingredients to be reconstituted into final products, such as ketchups and
51
tomato sauces. Since tomato paste is the main ingredient in the final products,
52
maintaining the quality of the paste is crucial for the tomato processing industry
53
(Zhang et al., 2014). In order to obtain higher commercial profits, some
54
unscrupulous traders add illegal additives to tomato paste. The viscosity and shear
55
thinning behavior are essential characteristics of tomato pastes (Bayod et al., 2008;
56
Berta et al., 2016). Addition of sucrose can improve the viscosity of tomato paste.
57
The adulteration of tomato paste products is not only detrimental to the interests of
58
consumers but also adversely affect the development of the industry. In order to
59
ensure the quality and safety of tomato paste products, it is necessary to develop a
60
rapid and effective quality evaluation method for identifying adulterated tomato
61
paste.
62
Hyperspectral imaging is an emerging non-destructive technology that
63
integrates conventional imaging and spectroscopy to attain both spatial and spectral
ACCEPTED MANUSCRIPT 64
information from an object simultaneously (Dale et al., 2013; Cheng and Sun,
65
2014). Recently, hyperspectral imaging technology has been used to detect
66
adulteration in beef (Kamruzzaman et al., 2015), lamb meat (Kamruzzaman et al.,
67
2013), prawn (Wu et al., 2013), and milk powders (Fu et al., 2014; Huang et al.,
68
2016; Lim et al., 2016). However, the rich information in hyperspectral imaging
69
results in difficulties in data processing, which makes it hard for industrial online
70
applications.
71
Recently, a simplified multispectral imaging has been increasingly applied as a
72
powerful analytical tool for non-destructive and rapid determination of food quality
73
and safety, including constituent analysis (Dissing et al., 2011; Liu et al., 2015),
74
quality evaluation (Lleó et al., 2009; Løkke et al., 2013), and so on (Feng and Sun,
75
2012; Qin et al., 2013; Pu et al., 2015). Previous studies also showed that
76
multispectral imaging is especially suitable for adulteration detection, such as
77
detection of beef and pork adulteration (Ropodi et al., 2015; Ropodi et al., 2016),
78
identification of water-injected beef samples (Liu et al., 2016), and detection of
79
dicyandiamide in infant formula powder (Liu et al., 2017). Due to the advantages of
80
the multispectral imaging technology, the objective of this study was to investigate
81
the feasibility of using this technique for the detection and quantification of sucrose
82
adulteration in tomato paste samples. The specific goals include to: (1) evaluate the
83
potential use of multispectral imaging to detect sucrose adulteration in tomato paste;
84
(2) compare the detection and prediction performance of different chemometric
85
methods, including partial least squares (PLS), least squares-support vector
86
machines (LS-SVM), and back propagation neural network (BPNN) models to
87
detect sucrose adulteration in tomato paste; and (3) identify the lowest proportion of
88
sucrose in tomato paste that can be safely detected.
ACCEPTED MANUSCRIPT 89
2. Materials and methods
90
2.1. Samples preparation
91
Two batches of pure concentrated tomato paste were provided by the Technique
92
Center of Xinjiang Entry-Exit Inspection and Quarantine Bureau (Urumchi City,
93
China), including the soluble solids content was 30 oBrix (batch 1) and 36 oBrix
94
(batch 2). Analytical grade sucrose used for adulteration was purchased from
95
Sinopharm Chemical Reagent Co., Ltd (Shanghai, China). Sucrose produced mainly
96
from sugarcane was purchased from Carrefour supermarket (Hefei City, China).
97
The concentrations of sugars (sucrose, fructose, and glucose) from the tomato paste
98
prior to the experiment were measured by using high performance liquid
99
chromatography (HPLC). The results showed that the concentration of fructose and
100
glucose were 6.60±0.05 and 5.15±0.07 g/100g in batch 1 and 8.75±0.07 and
101
7.60±0.28 g/100g in batch 2 of tomato paste, respectively, and the sucrose was not
102
detected. Analytical grade sucrose or sucrose from sugarcane was mixed into the
103
tomato paste to make mixtures at different proportion levels (w/w) of 1%, 2%, 3%,
104
4%, 5%, 6%, 7%, 8% and 9%, respectively. In addition, samples of pure tomato
105
paste were also prepared for experiment. In order to make sure the sucrose particles
106
were uniformly distributed in the tomato paste, the crystal sucrose was ground into
107
powder. Then, powdered sucrose was added to tomato paste and stirred for about 10
108
min with a glass stick so that the sucrose was completely mixed with tomato paste.
109
The mixture samples were placed in Petri dishes and levelled across the top using a
110
card to smooth the sample surface and remove excess adulteration tomato paste, and
111
then snapshots were taken using multispectral imaging system for every proportion
112
level of adulteration, each Petri dish was considered as a replicate in the experiment
113
(15 × 9 samples in calibration set and 5 × 9 samples in prediction set, respectively).
ACCEPTED MANUSCRIPT 114
In addition, when determining the detection limit of sucrose proportions in tomato
115
paste, new samples were generated and each Petri dish was also considered as a
116
replicate in the experiment (30 × 2 samples in calibration set and 20 × 2 samples in
117
prediction set, respectively).
118
2.2. Multispectral imaging system and image acquisition
119
The multispectral images of tomato paste samples were captured using the
120
VideometerLab equipment (Videometer A/S, Hørsholm, Denmark) which acquired
121
multispectral images at 19 different wavelengths. The detailed information of the
122
measured wavelengths used in the study were 405, 435, 450, 470, 505, 525, 570,
123
590, 630, 645, 660, 700, 780, 850, 870, 890, 910, 940, and 970 nm. Fig. 1 shows the
124
principal setup of the rapid detection and screening platform. The ready-to-use
125
system integrates an integrating sphere, light emitting diodes (LEDs), a
126
monochrome gray scale CCD camera, and computer technology with advanced
127
digital image analysis and statistics. Although the CCD camera itself has a factor of
128
10 times lower sensitivity in NIR than in the optimal green wavelength, the
129
instrument has balanced out this sensitivity with the intensity of illumination, such
130
that 10 times more NIR light than green light is typically sent. And in this way the
131
same signal to noise ratio over all wavelengths can be obtained. The advantage of
132
this apparatus is that it not only uses the information of visible and short-wave NIR
133
spectral regions, but also uses the spatial information of each pixel.
134
The system was first calibrated radiometrically using both a diffuse white and a
135
dark target and then geometrically calibrated with a geometric target to ensure pixel
136
correspondence for all spectral bands. Following this, a light setup based on the type
137
of object to be recorded called “auto light”. In auto light, it is always the brightest
138
sections in the image that dictate the final result. For each proportion of
ACCEPTED MANUSCRIPT 139
adulteration, a different auto light procedure was employed. A more detailed
140
description of this instrument can be found in the reference (Panagou et al., 2014).
141
Image segmentation was performed using the VideometerLab software version
142
2.12.23. To remove the image background, all items except the tomato paste were
143
removed by a Canonical Discriminant Analysis (CDA) and segmented using a
144
simple threshold. The image of tomato paste sample without the background could
145
be transformed to spectra based on a mean calculation. Thus, each image
146
contributed with a single spectrum for the model calibration.
147
2.3. Data analysis
148
Different chemometric methods, including principal component analysis
149
(PCA), PLS, partial least squares discriminant analysis (PLSDA), LS-SVM and
150
BPNN were used for qualitative and quantitative analysis of the spectral data. All of
151
these chemometrics analysis and statistics were performed using the commercial
152
software Matlab 2009 (The Mathworks Inc., Natick, MA, USA), and Origin 8.5.
153
Principal component analysis (PCA), a common unsupervised method of
154
identification, was firstly applied to the initial exploration of data visualization
155
frameworks and the identification of abnormal value. In this study, PCA was used
156
to distinguish the pure tomato paste from the sucrose-blended tomato paste. With
157
the analysis of spectral data, PCA can provide very important information about the
158
potential capability of differentiating the samples.
159
PLS, as a frequently-used multivariate statistical data analysis method, is based
160
on full wavelength that participates in multivariate calibration model. It compressed
161
a large number of variables into a few much smaller number of latent variables
162
(LVs) that were linear combinations of the spectral data (X) and used these factors
163
to ascertain for the analyte’s proportion (Y), explaining much of the covariance of X
ACCEPTED MANUSCRIPT 164
and Y. The number of LVs, as a critical parameter in the algorithm, was determined
165
by minimizing the value of the prediction residual error sum of square (PRESS).
166
The principal components in the whole reflectance spectral matrix X and the
167
proportion level of adulteration matrix Y were separated (Panagou, et al., 2011;
168
Zheng et al., 2016), which are represented as follow:
X =TPT +E Y =UQT +F
169
170 171
(1) (2)
where T and P, U and Q are the vectors of X and Y PLS scores and loadings respectively, and E, F are the X and Y residuals.
172
PLSDA is a partial least squares application for the optimum separation of
173
classes, and each sample is assigned a dummy variable 1 or 2 as a reference value. It
174
is an arbitrary number indicating whether the sample belongs to a particular group
175
or not. The criteria for the selected cutoff were similar to those reported by Xie et al.
176
(2007).
177
LS-SVM proposed by Cortes and Vapnik (1995) was used as a learning
178
algorithm for classification and regression tasks. It possesses the advantage of not
179
only good generalization performance as support vector machines (SVM), but also
180
simpler structure and shorter optimization time. Assume a set of training set is given
181
as {xk , yk }N K 1 , with the input xk R N and yk R . The following regression model
182
is constructed by nonlinear mapping function , which maps the input data to a
183
higher dimensional feature space. And the ( xk )T ( x j ) = K(xi, xj) (i, j = 1, 2, …,
184
N ) is defined as the kernel function. In this study, a radial basis function (RBF)
185
kernel with Gaussian function K(x, xk) was selected as the kernel function
186 187
K ( x, xk ) exp( || x xk ||2 / 2 )
(3)
The LS-SVM model can be simply expressed as:
ACCEPTED MANUSCRIPT N
y ( x) ak K ( x, xk ) b
188
(4)
k 1
189
Where αk is Lagrange multipliers that is the parameters in the optimization
190
problem called support value, xk is the input vector, b is the bias term. RBF kernel
191
function was used to reduce the computational complexity of the training procedure
192
and give a good performance under the general smoothness assumptions. Two
193
crucial parameters (γ, σ2) were needed for LS-SVM and higher performance of the
194
model could be obtained by adjusting the value of parameters. The details of LS-
195
SVM algorithm could be found in the previous reported researches (Devos et al.,
196
2009).
197
BPNN could solve complex problems more accurately than linear techniques,
198
which was successfully used in many fields especially for pattern recognition. A
199
three-layer structure BPNN that was an input layer, a hidden layer and an output
200
layer was selected. Whole spectral regions including 19 wavelengths reflection were
201
used to be the BPNN input layer. The obtained eigenvectors were processed by the
202
neural network and the output of network expressed the resemblance that an object
203
corresponds with a training pattern. Leave-one-out cross-validation procedure was
204
used during the calibration step. Several network architectures were tested by
205
varying the number of neurons in the hidden layer with different initial weights. The
206
optimal parameters such as hidden nodes, the goal error and iteration times were
207
determined by the least prediction error. The theory of BPNN has been described in
208
detail (Dai and MacBeth, 1997).
209
2.4. Evaluation of model performance
210
The quality of the models was evaluated by the root mean square error of
211
calibration (RMSEC), root mean square error of prediction (RMSEP), bias and the
ACCEPTED MANUSCRIPT 212
coefficient of determination (R2) in calibration ( RC2 ) and prediction ( RP2 ).
213
Meanwhile, the residual predictive deviation (RPD) was used to verify the accuracy
214
of the calibration models developed. The higher the value of the RPD the greater the
215
probability of the model to predict the chemical composition in samples set
216
accurately. An RPD value range between 2.4 and 3.0 is considered poor and the
217
models could be applied only for very rough screening, while an RPD value greater
218
than 3.0 could be considered fair and recommended for screening purposes and
219
good for quality control, respectively (Sinelli et al., 2008). Generally, an optimum
220
model should have high RC2 , RP2 , and RPD values, and low RMSEC and RMSEP
221
values.
222
3. Results and discussion
223
3.1. Spectral features of tomato paste
224
Average reflectance spectra in the wavelength range of 405-970 nm of the
225
examined pure tomato paste, and sucrose and tomato paste mixtures with different
226
proportions of sucrose are shown in Fig. 2. For the low sucrose proportions under
227
investigation, the mean spectra of sucrose and tomato paste mixtures were very
228
similar to the spectra of pure tomato paste across the whole tested wavelength
229
region. Furthermore, the reflectance values of the samples decreased with the
230
increase of sucrose proportions. A main absorption band observed between 940 and
231
970 nm related to O-H third stretching overtone was assigned to water molecule
232
(Wu et al., 2008). This observation suggests that the spectral evaluation may allow
233
effective determination of sucrose proportions in tomato paste. However, it is
234
difficult to determine the sucrose proportion directly from the observed spectra.
235
Especially when a lot of samples were measured simultaneously, their spectral
236
curves showed many overlaps, making the quantitative determination even more
ACCEPTED MANUSCRIPT 237
complicated. In order to solve this problem, multivariate data analysis method was
238
introduced to analyze the multiple variables of spectral data.
239
3.2. Quantification of sucrose proportion in tomato paste
240
The main purpose of the investigation was to quantify the level of adulteration
241
in tomato paste. Quantitative calibration models were generated using PLS, LS-
242
SVM and BPNN mathematical algorithms. Through comparison of the value of
243
PRESS, the number of latent variables for PLS model both were determined as 5 for
244
predicting sucrose proportions in both batches of tomato paste. Based on the
245
spectral information, the performance of these three models for predicting sucrose
246
proportions in tomato paste was shown in Table 1. Compared with PLS and BPNN
247
models, the LS-SVM model had the best predictive accuracy with RC2 of 0.990, RP2
248
of 0.936, RMSEC of 0.263%, bias of 0.325%, and RMSEP of 0.521% for batch 1 of
249
tomato paste. Furthermore, the largest RPD of 5.014 showed that the LS-SVM
250
model was considered to be adequate for sucrose proportions prediction. Although
251
the RC2 and RMSEC in BPNN model were similar to those in LS-SVM model, the
252
RMSEP was higher and the RP2 and RPD were lower indicating that the BPNN
253
model was slightly less good for sucrose proportions prediction. Regarding the PLS,
254
the RMSEP was higher and RPD was lower than LS-SVM and BPNN models in
255
prediction set indicating that the model was less good. While the RMSEP was
256
slightly larger than RMSEC with lower bias in PLS indicated that the PLS is a
257
potential better-behaved model for practical application.
258
Similar results were also obtained for batch 2 of tomato paste. In general, the
259
LS-SVM model gave better prediction abilities with RP2 of 0.966 than PLS and
260
BPNN models. The corresponding values for RMSEP and RPD were found to be
261
0.445% and 5.865, respectively. Therefore, LS-SVM was considered as the best
ACCEPTED MANUSCRIPT 262
way for establishing the quantitative model of sucrose proportion in tomato paste
263
and had the best generalization capability. This was due to the fact that LS-SVM
264
can learn in high-dimensional characteristic space with fewer training samples. It is
265
worth mentioning that PLS, although used extensively in chemometrics, did not
266
perform always as well as LS-SVM and BPNN, at least for this particular case. The
267
obtained prediction results confirmed the suitability of multispectral imaging for
268
determination of sucrose proportion in tomato paste in a rapid and non-destructive
269
manner.
270
Furthermore, the results from the ReliefF method showed that the 630, 645,
271
700, 780 and 850 nm for batch 1 and 630, 645, 850, 890 and 940 nm for batch 2 of
272
tomato paste were more important than other wavelengths for sucrose detection.
273
The results were consistent with previously reported result, that the combination of
274
NIR wavelengths (λ= 730, 830, 909-915, 960, and 965 nm) within C-H and O-H
275
bands can reliably quantify sucrose (Omar et al., 2012).
276
3.3. PCA data analysis
277
In the study, PCA was performed initially to visualize any variation among pure
278
tomato paste, tomato paste containing 1% and 9% sucrose (the minimum and
279
maximum proportions in the current study, respectively) in principal component
280
(PC) space. Fig. 3 shows the two-dimensional score plot of the first two principal
281
components (PCs) from the spectral reflectance values extracted from the
282
multispectral images of the samples. The first two PCs, which account for the most
283
spectral variations above 97% (97.96% and 98.87% in batch 1 and batch 2 of
284
tomato paste, respectively), were used to make differentiation clear. The results
285
showed that the PCA method can clearly identify the pure tomato paste, as well as
286
1% and 9% sucrose adulteration in tomato paste.
ACCEPTED MANUSCRIPT 287
3.4. Detection limit of sucrose proportions in tomato paste
288
Currently although there are no international standards for acceptable levels of
289
sucrose in tomato paste yet, tomato paste mixed with sucrose seriously endangering
290
the interests of consumers and market order. Therefore, identification of adulterated
291
tomato paste is very important. Especially, the ability to differentiate between
292
unadulterated tomato paste and those containing the 1% sucrose (the lowest
293
proportion in the current study) is crucial. Spectra of pure tomato paste and tomato
294
paste containing 1% sucrose were collected using multispectral imaging system and
295
analyzed using PLSDA, LS-SVM and BPNN models. All three models were able to
296
distinguish between adulterated (1% sucrose) and unadulterated tomato paste. The
297
high classification accuracies of 100% were obtained in prediction set for both
298
batches of tomato paste using PLSDA, LS-SVM or BPNN model. The results show
299
that the detection limit of sucrose in tomato paste was 1% using multispectral
300
imaging combined with chemometric methods.
301
4. Conclusions
302
The detection of adulteration is attracting much attention for tomato paste
303
manufacturer, researchers and consumers. This research was carried out to
304
investigate the prediction performance of the multispectral imaging for rapid
305
detection of sucrose adulteration in tomato paste. Better prediction performance and
306
generalization ability for quantification of sucrose proportions in both batches of
307
tomato paste were achieved using LS-SVM. Besides, multispectral imaging was
308
able to detect sucrose adulteration in tomato paste at the 1% level. The study
309
confirmed that multispectral imaging technology in tandem with chemometric
310
method has the potential to be a rapid and non-destructive method to accurately
311
identification of the adulteration proportions in tomato paste. Further study is
ACCEPTED MANUSCRIPT 312
needed to improve the adaptability of the LS-SVM model or to develop new
313
mathematical model for the practical applications to detect adulteration in tomato
314
paste.
315 316
Acknowledgements
317
This work was supported by the National Key Research and Development Plan
318
of China (2016YFD0401104), the National Natural Science Foundation of China
319
(31401544), the Key Science & Technology Specific Projects of Anhui Province
320
(15czz03117), the Funds for Huangshan Professorship of Hefei University of
321
Technology (407-037019), and the Fundamental Research Funds for the Central
322
Universities (JZ2016HGTB0712).
323 324
References
325
Bayod, E., Willers, E.P., Tornberg, E., 2008. Rheological and structural
326
characterization of tomato paste and its influence on the quality of ketchup.
327
LWT-Food Sci. Technol. 41, 1289-1300.
328
Berta, M., Wiklund, J., Kotzé, R., Stading, M., 2016. Correlation between in-line
329
measurements of tomato ketchup shear viscosity and extensional viscosity. J.
330
Food Eng. 173, 8-14.
331
Borin, A., Ferrão, M.F., Mello, C., Maretto, D.A., Poppi, R.J., 2006. Least-squares
332
support vector machines and near infrared spectroscopy for quantification of
333
common adulterants in powdered milk. Anal. Chim. Acta 579, 25-32.
334
Burton‐Freeman, B., Talbot, J., Park, E., Krishnankutty, S., Edirisinghe, I., 2012.
335
Protective activity of processed tomato products on postprandial oxidation and
336
inflammation: a clinical trial in healthy weight men and women. Mol. Nutr.
ACCEPTED MANUSCRIPT 337
Food Res. 56, 622-631.
338
Cheng J.H., Sun D.W., 2014. Hyperspectral imaging as an effective tool for quality
339
analysis and control of fish and other seafoods: current research and potential
340
applications. Trends Food Sci. Tech. 37, 78-91.
341
Cortes, C., Vapnik, V., 1995. Support vector network. Mach. Learn. 20, 273-297.
342
Dai, H., MacBeth, C., 1997. Effects of learning parameters on learning procedure
343
and performance of a BPNN. Neural Networks 10, 1505-1521.
344
Dale, L.M., Thewis, A., Boudry, C., Rotar, I., Dardenne, P., Vincent, B., Fernández,
345
P.J.A., 2013. Hyperspectral imaging applications in agriculture and agro-food
346
product quality and safety control: A review. Appl. Spectrosc. Rev. 48, 142-
347
159.
348
Devos, O., Ruckebusch, C., Durand, A., Duponchel, L., Huvenne, J.P., 2009.
349
Support vector machines (SVM) in near infrared (NIR) spectroscopy: focus on
350
parameters optimization and model interpretation. Chemometr. Intell. Lab. 96,
351
27-33.
352
Dissing, B.S., Nielsen, M.E., Ersbøll, B.K., Frosch, S., 2011. Multispectral imaging
353
for determination of astaxanthin concentration in salmonids. PLoS ONE 6,
354
e19032.
355
Feng, Y.Z., Sun, D.W., 2012. Application of hyperspectral imaging in food safety
356
inspection and control: a review. Crit. Rev. Food Sci. Nutr. 52, 1039-1058.
357
Fu, X., Kim, M.S., Chao, K., Qin, J., Lim, J., Lee, H., Garrido-Varo, A., Pérez-
358
Marín, D., Ying, Y., 2014. Detection of melamine in milk powders based on
359
NIR hyperspectral imaging and spectral similarity analyses. J. Food Eng. 124,
360
97-104.
361
Huang, M., Kim, M.S., Delwiche, S.R., Chao, K., Qin, J., Mo, C., Zhu, Q., 2016.
ACCEPTED MANUSCRIPT 362
Quantitative analysis of melamine in milk powders using near-infrared
363
hyperspectral imaging and band ratio. J. Food Eng. 181, 10-19.
364
Kamruzzaman, M., Makino, Y., Oshita, S., 2015. Rapid and non-destructive
365
detection of chicken adulteration in minced beef using visible near-infrared
366
hyperspectral imaging and machine learning. J. Food Eng. 170, 8-15.
367
Kamruzzaman, M., Sun, D.W., ElMasry, G., Allen, P., 2013. Fast detection and
368
visualization of minced lamb meat adulteration using NIR hyperspectral
369
imaging and multivariate image analysis. Talanta 103, 130-136.
370
Lim, J., Kim, G., Mo, C., Kim, M., Chao, K., Qin, J., Fu, X., Baek, I., Cho, B.K.,
371
2016. Detection of melamine in milk powders using near-infrared
372
hyperspectral imaging combined with regression coefficient of partial least
373
square regression model. Talanta 151, 183-191.
374
Liu, C., Liu, W., Chen, W., Yang, J., Zheng, L., 2015. Feasibility in multispectral
375
imaging for predicting the content of bioactive compounds in intact tomato
376
fruit. Food Chem. 173, 482-488.
377
Liu, C., Liu, W., Yang, J., Chen, Y., Zheng, L., 2017. Non-destructive detection of
378
dicyandiamide in infant formula powder using multi-spectral imaging coupled
379
with chemometrics. J. Sci. Food Agr. 97, 2094-2099.
380
Liu, J., Cao, Y., Pan, W., Ma, F., Liu, C., Chen, W., Yang, J., Zheng, L., 2016.
381
Rapid and non-destructive identification of water-injected beef samples using
382
multispectral imaging analysis. Food Chem. 190, 938-943.
383
Lleó, L., Barreiro, P., Ruiz-Altisent, M., Herrero, A., 2009. Multispectral images of
384
peach related to firmness and maturity at harvest. J. Food Eng. 93, 229-235.
385
Løkke, M.M., Seefeldt, H.F., Skov, T., Edelenbos, M., 2013. Color and textural
386
quality of packaged wild rocket measured by multispectral imaging.
ACCEPTED MANUSCRIPT 387
Postharvest Biol. Technol. 75, 86-95.
388
Omar, A.F., Atan, H., MatJafri, M.Z., 2012. Peak response identification through
389
near-infrared spectroscopy analysis on aqueous sucrose, glucose, and fructose
390
solution. Spectrosc. Lett. 45, 190-201.
391
Panagou, E.Z., Mohareb, F.R., Argyri, A.A., Bessant, C.M., Nychas, G.J.E., 2011.
392
A comparison of artificial neural networks and partial least squares modelling
393
for the rapid detection of the microbial spoilage of beef fillets based on Fourier
394
transform infrared spectral fingerprints. Food Microbiol. 28, 782-790.
395
Panagou, E.Z., Papadopoulou, O., Carstensen, J.M., Nychas, G.J.E., 2014. Potential
396
of
multispectral
imaging
technology
for
rapid
and
non-destructive
397
determination of the microbiological quality of beef filets during aerobic
398
storage. Int. J. Food Microbiol. 174, 1-11.
399
Pu, H., Kamruzzaman, M., Sun, D.W., 2015. Selection of feature wavelengths for
400
developing multispectral imaging systems for quality, safety and authenticity
401
of muscle foods-a review. Trends Food Sci. Tech. 45, 86-104.
402
Qin, J., Chao, K., Kim, M.S., Lu, R., Burks, T.F., 2013. Hyperspectral and
403
multispectral imaging for evaluating food safety and quality. J. Food Eng. 118,
404
157-171.
405
Rizwan, M., Rodriguez-Blanco, I., Harbottle, A., Birch-Machin, M.A., Watson,
406
R.E.B., Rhodes, L.E., 2011. Tomato paste rich in lycopene protects against
407
cutaneous photodamage in humans in vivo: a randomized controlled trial. Brit.
408
J. Dermatol. 164, 154-162.
409
Ropodi, A.I., Panagou, E.Z., Nychas, G.J.E., 2016. Multispectral imaging (MSI): A
410
promising method for the detection of minced beef adulteration with
411
horsemeat. Food Control 73, 57-63.
ACCEPTED MANUSCRIPT 412
Ropodi, A.I., Pavlidis, D.E., Mohareb, F., Panagou, E.Z., Nychas, G.J., 2015.
413
Multispectral image analysis approach to detect adulteration of beef and pork
414
in raw meats. Food Res. Int. 67, 12-18.
415
Sinelli, N., Spinardi, A., Di Egidio, V., Mignani, I., Casiraghi, E., 2008. Evaluation
416
of quality and nutraceutical content of blueberries (Vaccinium corymbosum
417
L.) by near and mid-infrared spectroscopy. Postharvest Biol. Technol. 50, 31-
418
36.
419
Wu, D., He, Y., Feng, S., 2008. Short-wave near-infrared spectroscopy analysis of
420
major compounds in milk powder and wavelength assignment. Analytica
421
Chimica Acta 610, 232-242.
422
Wu, D., Shi, H., He, Y., Yu, X., Bao, Y., 2013. Potential of hyperspectral imaging
423
and multivariate analysis for rapid and non-invasive detection of gelatin
424
adulteration in prawn. J. Food Eng. 119, 680-686.
425
Xie, L., Ying, Y., Ying, T., Yu, H., Fu, X., 2007. Discrimination of transgenic
426
tomatoes based on visible/near-infrared spectra. Anal. Chim. Acta 584, 379-
427
384.
428
Zhang, L., Schultz, M.A., Cash, R., Barrett, D.M., McCarthy, M.J., 2014.
429
Determination of quality parameters of tomato paste using guided microwave
430
spectroscopy. Food Control 40, 214-223.
431
Zheng, J., Song, Z., Ge, Z., 2016. Probabilistic learning of partial least squares
432
regression model: Theory and industrial applications. Chemometr. Intell.
433
Lab. 158, 80-90.
ACCEPTED MANUSCRIPT
435 436
Fig. 1. Principal setup of the multispectral imaging system. An integrating sphere
437
with a matte white coating ensures optimal lighting conditions. The light emitting
438
diodes (LEDs) located in the rim of the sphere ensures narrowband illumination.
439
The image acquisition is performed by a monochrome grayscale CCD camera
440
mounted in the top of the sphere.
441
ACCEPTED MANUSCRIPT
443 444
Fig. 2. Average reflection spectra of pure tomato paste and different proportions of
445
sucrose in batch 1 (a) and batch 2 (b) of tomato paste, respectively.
ACCEPTED MANUSCRIPT
447 448
Fig. 3. Two-dimensional score plot of the first two principal components conducted
449
on the spectral data in batch 1 (a) and batch (b) of tomato paste. The classification
450
was clear among the three groups of the pure tomato paste (○), 1% sucrose in
451
tomato paste (*) and 9% sucrose in tomato paste (+).
ACCEPTED MANUSCRIPT 453
Table 1
454
Performance of PLS, LS-SVM and BPNN models for predicting sucrose
455
proportions (%) in tomato paste. Batch Batch 1
Batch 2
Sources of sucrose
Chemometrics
RC2
RMSEC
RP2
Bias
RMSEP
RPD
Analytical
PLS
0.949
0.569
0.927
0.262
0.630
4.140
grade
LS-SVM
0.990
0.263
0.936
0.325
0.521
5.014
BPNN
0.990
0.259
0.919
0.322
0.620
4.210
PLS
0.894
0.800
0.878
0.454
0.703
3.715
LS-SVM
0.994
0.205
0.966
0.066
0.445
5.865
BPNN
0.958
0.511
0.910
0.462
0.548
4.767
Sugarcane
2 C
2 P
456
R , coefficient of determination in calibration; R , coefficient of determination in prediction;
457
RMSEC, root mean square error of calibration; RMSEP, root mean square error of prediction;
458
RPD, residual predictive deviation.
ACCEPTED MANUSCRIPT Research highlights
1. Multispectral imaging could determine sucrose proportion in tomato paste accurately. 2. Multispectral imaging was able to detect low proportion sucrose in tomato paste. 3. Multispectral imaging successfully used as a rapid screening technique for sucrose.
Changhong Liu, Guang Hao, Min Su, Ying Chen, Lei Zheng PII:
S0260-8774(17)30324-2
DOI:
10.1016/j.jfoodeng.2017.07.026
Reference:
JFOE 8968
To appear in:
Journal of Food Engineering
Received Date:
09 April 2017
Revised Date:
21 July 2017
Accepted Date:
24 July 2017
Please cite this article as: Changhong Liu, Guang Hao, Min Su, Ying Chen, Lei Zheng, Potential of multispectral imaging combined with chemometric methods for rapid detection of sucrose adulteration in tomato paste, Journal of Food Engineering (2017), doi: 10.1016/j.jfoodeng. 2017.07.026
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 1
Potential of multispectral imaging combined with chemometric methods for
2
rapid detection of sucrose adulteration in tomato paste
3
Running title: Rapid detection of sucrose adulteration in tomato paste
4 5
Changhong Liu a,b,‡, Guang Hao a,‡, Min Su c, Ying Chen d,*, Lei Zheng a,b,*
6 7 8 9 10
a School of Food Science and Engineering, Hefei University of Technology, Hefei 230009, China b Key Laboratory for Agricultural Products Processing of Anhui Province, Hefei 230009, China
11
c Xinjiang Entry-Exit Inspection and Quarantine Bureau, Urumqi, 830063, China
12
d Agro-product Safety Research Centre, Chinese Academy of Inspection and
13
Quarantine, Beijing, 100123, China
14 15 16
‡ These authors contribute equally to this work.
17
* Corresponding authors: Lei Zheng, Tel: +86-551-62919398.
18 19
E-mail:
[email protected],
[email protected] (Y Chen).
[email protected]
(L
Zheng);
ACCEPTED MANUSCRIPT 21
Abstract: Confirmation of the authenticity of tomato paste is an increasing
22
challenger for food processors and regulatory authorities. This study focuses on the
23
rapid qualitative and quantitative detection of sucrose adulteration in tomato paste
24
using multispectral imaging (405-970 nm) combined with chemometric methods.
25
Partial least squares (PLS), least squares-support vector machines (LS-SVM), and
26
back propagation neural network (BPNN) were used to develop quantitative
27
models. Compared with PLS and BPNN, LS-SVM considerably improved the
28
prediction performance with coefficient of determination in prediction ( RP2 ) of
29
0.936 and 0.966, the root-mean-square error of prediction (RMSEP) of 0.521% and
30
0.445%, and residual predictive deviation (RPD) of 5.014 and 5.865 for batch 1 and
31
batch 2 of tomato paste, respectively. Besides, multispectral imaging was feasible to
32
detect sucrose adulteration in tomato paste at very low proportions (1%) with no
33
misclassification using chemometric methods. It was concluded that multispectral
34
imaging has an excellent potential for rapid determination of sucrose adulteration in
35
tomato paste.
36
Keywords: Multispectral imaging; Tomato paste; Sucrose; Rapid detection;
37
Chemometric methods
ACCEPTED MANUSCRIPT 39
1. Introduction
40
Tomato fruit (Solanum lycopersicum) is the second most important vegetable
41
crop worldwide which is consumed fresh as well as in the form of processed
42
products, such as tomato juice, tomato puree, tomato sauce, ketchup, and tomato
43
paste. Tomato paste is the most commonly consumed processed tomato product,
44
which is concentrated from fresh tomatoes. There is an increasing demand of
45
tomato paste consumption due to the beneficial health effects (Rizwan et al., 2011;
46
Burton‐Freeman et al., 2012). Therefore, tomato paste is an essential food in
47
people’s daily life in European and American countries, and becomes more and
48
more widely accepted in many other countries such as China.
49
Tomato paste is usually stored and used as an intermediate product with water
50
and other ingredients to be reconstituted into final products, such as ketchups and
51
tomato sauces. Since tomato paste is the main ingredient in the final products,
52
maintaining the quality of the paste is crucial for the tomato processing industry
53
(Zhang et al., 2014). In order to obtain higher commercial profits, some
54
unscrupulous traders add illegal additives to tomato paste. The viscosity and shear
55
thinning behavior are essential characteristics of tomato pastes (Bayod et al., 2008;
56
Berta et al., 2016). Addition of sucrose can improve the viscosity of tomato paste.
57
The adulteration of tomato paste products is not only detrimental to the interests of
58
consumers but also adversely affect the development of the industry. In order to
59
ensure the quality and safety of tomato paste products, it is necessary to develop a
60
rapid and effective quality evaluation method for identifying adulterated tomato
61
paste.
62
Hyperspectral imaging is an emerging non-destructive technology that
63
integrates conventional imaging and spectroscopy to attain both spatial and spectral
ACCEPTED MANUSCRIPT 64
information from an object simultaneously (Dale et al., 2013; Cheng and Sun,
65
2014). Recently, hyperspectral imaging technology has been used to detect
66
adulteration in beef (Kamruzzaman et al., 2015), lamb meat (Kamruzzaman et al.,
67
2013), prawn (Wu et al., 2013), and milk powders (Fu et al., 2014; Huang et al.,
68
2016; Lim et al., 2016). However, the rich information in hyperspectral imaging
69
results in difficulties in data processing, which makes it hard for industrial online
70
applications.
71
Recently, a simplified multispectral imaging has been increasingly applied as a
72
powerful analytical tool for non-destructive and rapid determination of food quality
73
and safety, including constituent analysis (Dissing et al., 2011; Liu et al., 2015),
74
quality evaluation (Lleó et al., 2009; Løkke et al., 2013), and so on (Feng and Sun,
75
2012; Qin et al., 2013; Pu et al., 2015). Previous studies also showed that
76
multispectral imaging is especially suitable for adulteration detection, such as
77
detection of beef and pork adulteration (Ropodi et al., 2015; Ropodi et al., 2016),
78
identification of water-injected beef samples (Liu et al., 2016), and detection of
79
dicyandiamide in infant formula powder (Liu et al., 2017). Due to the advantages of
80
the multispectral imaging technology, the objective of this study was to investigate
81
the feasibility of using this technique for the detection and quantification of sucrose
82
adulteration in tomato paste samples. The specific goals include to: (1) evaluate the
83
potential use of multispectral imaging to detect sucrose adulteration in tomato paste;
84
(2) compare the detection and prediction performance of different chemometric
85
methods, including partial least squares (PLS), least squares-support vector
86
machines (LS-SVM), and back propagation neural network (BPNN) models to
87
detect sucrose adulteration in tomato paste; and (3) identify the lowest proportion of
88
sucrose in tomato paste that can be safely detected.
ACCEPTED MANUSCRIPT 89
2. Materials and methods
90
2.1. Samples preparation
91
Two batches of pure concentrated tomato paste were provided by the Technique
92
Center of Xinjiang Entry-Exit Inspection and Quarantine Bureau (Urumchi City,
93
China), including the soluble solids content was 30 oBrix (batch 1) and 36 oBrix
94
(batch 2). Analytical grade sucrose used for adulteration was purchased from
95
Sinopharm Chemical Reagent Co., Ltd (Shanghai, China). Sucrose produced mainly
96
from sugarcane was purchased from Carrefour supermarket (Hefei City, China).
97
The concentrations of sugars (sucrose, fructose, and glucose) from the tomato paste
98
prior to the experiment were measured by using high performance liquid
99
chromatography (HPLC). The results showed that the concentration of fructose and
100
glucose were 6.60±0.05 and 5.15±0.07 g/100g in batch 1 and 8.75±0.07 and
101
7.60±0.28 g/100g in batch 2 of tomato paste, respectively, and the sucrose was not
102
detected. Analytical grade sucrose or sucrose from sugarcane was mixed into the
103
tomato paste to make mixtures at different proportion levels (w/w) of 1%, 2%, 3%,
104
4%, 5%, 6%, 7%, 8% and 9%, respectively. In addition, samples of pure tomato
105
paste were also prepared for experiment. In order to make sure the sucrose particles
106
were uniformly distributed in the tomato paste, the crystal sucrose was ground into
107
powder. Then, powdered sucrose was added to tomato paste and stirred for about 10
108
min with a glass stick so that the sucrose was completely mixed with tomato paste.
109
The mixture samples were placed in Petri dishes and levelled across the top using a
110
card to smooth the sample surface and remove excess adulteration tomato paste, and
111
then snapshots were taken using multispectral imaging system for every proportion
112
level of adulteration, each Petri dish was considered as a replicate in the experiment
113
(15 × 9 samples in calibration set and 5 × 9 samples in prediction set, respectively).
ACCEPTED MANUSCRIPT 114
In addition, when determining the detection limit of sucrose proportions in tomato
115
paste, new samples were generated and each Petri dish was also considered as a
116
replicate in the experiment (30 × 2 samples in calibration set and 20 × 2 samples in
117
prediction set, respectively).
118
2.2. Multispectral imaging system and image acquisition
119
The multispectral images of tomato paste samples were captured using the
120
VideometerLab equipment (Videometer A/S, Hørsholm, Denmark) which acquired
121
multispectral images at 19 different wavelengths. The detailed information of the
122
measured wavelengths used in the study were 405, 435, 450, 470, 505, 525, 570,
123
590, 630, 645, 660, 700, 780, 850, 870, 890, 910, 940, and 970 nm. Fig. 1 shows the
124
principal setup of the rapid detection and screening platform. The ready-to-use
125
system integrates an integrating sphere, light emitting diodes (LEDs), a
126
monochrome gray scale CCD camera, and computer technology with advanced
127
digital image analysis and statistics. Although the CCD camera itself has a factor of
128
10 times lower sensitivity in NIR than in the optimal green wavelength, the
129
instrument has balanced out this sensitivity with the intensity of illumination, such
130
that 10 times more NIR light than green light is typically sent. And in this way the
131
same signal to noise ratio over all wavelengths can be obtained. The advantage of
132
this apparatus is that it not only uses the information of visible and short-wave NIR
133
spectral regions, but also uses the spatial information of each pixel.
134
The system was first calibrated radiometrically using both a diffuse white and a
135
dark target and then geometrically calibrated with a geometric target to ensure pixel
136
correspondence for all spectral bands. Following this, a light setup based on the type
137
of object to be recorded called “auto light”. In auto light, it is always the brightest
138
sections in the image that dictate the final result. For each proportion of
ACCEPTED MANUSCRIPT 139
adulteration, a different auto light procedure was employed. A more detailed
140
description of this instrument can be found in the reference (Panagou et al., 2014).
141
Image segmentation was performed using the VideometerLab software version
142
2.12.23. To remove the image background, all items except the tomato paste were
143
removed by a Canonical Discriminant Analysis (CDA) and segmented using a
144
simple threshold. The image of tomato paste sample without the background could
145
be transformed to spectra based on a mean calculation. Thus, each image
146
contributed with a single spectrum for the model calibration.
147
2.3. Data analysis
148
Different chemometric methods, including principal component analysis
149
(PCA), PLS, partial least squares discriminant analysis (PLSDA), LS-SVM and
150
BPNN were used for qualitative and quantitative analysis of the spectral data. All of
151
these chemometrics analysis and statistics were performed using the commercial
152
software Matlab 2009 (The Mathworks Inc., Natick, MA, USA), and Origin 8.5.
153
Principal component analysis (PCA), a common unsupervised method of
154
identification, was firstly applied to the initial exploration of data visualization
155
frameworks and the identification of abnormal value. In this study, PCA was used
156
to distinguish the pure tomato paste from the sucrose-blended tomato paste. With
157
the analysis of spectral data, PCA can provide very important information about the
158
potential capability of differentiating the samples.
159
PLS, as a frequently-used multivariate statistical data analysis method, is based
160
on full wavelength that participates in multivariate calibration model. It compressed
161
a large number of variables into a few much smaller number of latent variables
162
(LVs) that were linear combinations of the spectral data (X) and used these factors
163
to ascertain for the analyte’s proportion (Y), explaining much of the covariance of X
ACCEPTED MANUSCRIPT 164
and Y. The number of LVs, as a critical parameter in the algorithm, was determined
165
by minimizing the value of the prediction residual error sum of square (PRESS).
166
The principal components in the whole reflectance spectral matrix X and the
167
proportion level of adulteration matrix Y were separated (Panagou, et al., 2011;
168
Zheng et al., 2016), which are represented as follow:
X =TPT +E Y =UQT +F
169
170 171
(1) (2)
where T and P, U and Q are the vectors of X and Y PLS scores and loadings respectively, and E, F are the X and Y residuals.
172
PLSDA is a partial least squares application for the optimum separation of
173
classes, and each sample is assigned a dummy variable 1 or 2 as a reference value. It
174
is an arbitrary number indicating whether the sample belongs to a particular group
175
or not. The criteria for the selected cutoff were similar to those reported by Xie et al.
176
(2007).
177
LS-SVM proposed by Cortes and Vapnik (1995) was used as a learning
178
algorithm for classification and regression tasks. It possesses the advantage of not
179
only good generalization performance as support vector machines (SVM), but also
180
simpler structure and shorter optimization time. Assume a set of training set is given
181
as {xk , yk }N K 1 , with the input xk R N and yk R . The following regression model
182
is constructed by nonlinear mapping function , which maps the input data to a
183
higher dimensional feature space. And the ( xk )T ( x j ) = K(xi, xj) (i, j = 1, 2, …,
184
N ) is defined as the kernel function. In this study, a radial basis function (RBF)
185
kernel with Gaussian function K(x, xk) was selected as the kernel function
186 187
K ( x, xk ) exp( || x xk ||2 / 2 )
(3)
The LS-SVM model can be simply expressed as:
ACCEPTED MANUSCRIPT N
y ( x) ak K ( x, xk ) b
188
(4)
k 1
189
Where αk is Lagrange multipliers that is the parameters in the optimization
190
problem called support value, xk is the input vector, b is the bias term. RBF kernel
191
function was used to reduce the computational complexity of the training procedure
192
and give a good performance under the general smoothness assumptions. Two
193
crucial parameters (γ, σ2) were needed for LS-SVM and higher performance of the
194
model could be obtained by adjusting the value of parameters. The details of LS-
195
SVM algorithm could be found in the previous reported researches (Devos et al.,
196
2009).
197
BPNN could solve complex problems more accurately than linear techniques,
198
which was successfully used in many fields especially for pattern recognition. A
199
three-layer structure BPNN that was an input layer, a hidden layer and an output
200
layer was selected. Whole spectral regions including 19 wavelengths reflection were
201
used to be the BPNN input layer. The obtained eigenvectors were processed by the
202
neural network and the output of network expressed the resemblance that an object
203
corresponds with a training pattern. Leave-one-out cross-validation procedure was
204
used during the calibration step. Several network architectures were tested by
205
varying the number of neurons in the hidden layer with different initial weights. The
206
optimal parameters such as hidden nodes, the goal error and iteration times were
207
determined by the least prediction error. The theory of BPNN has been described in
208
detail (Dai and MacBeth, 1997).
209
2.4. Evaluation of model performance
210
The quality of the models was evaluated by the root mean square error of
211
calibration (RMSEC), root mean square error of prediction (RMSEP), bias and the
ACCEPTED MANUSCRIPT 212
coefficient of determination (R2) in calibration ( RC2 ) and prediction ( RP2 ).
213
Meanwhile, the residual predictive deviation (RPD) was used to verify the accuracy
214
of the calibration models developed. The higher the value of the RPD the greater the
215
probability of the model to predict the chemical composition in samples set
216
accurately. An RPD value range between 2.4 and 3.0 is considered poor and the
217
models could be applied only for very rough screening, while an RPD value greater
218
than 3.0 could be considered fair and recommended for screening purposes and
219
good for quality control, respectively (Sinelli et al., 2008). Generally, an optimum
220
model should have high RC2 , RP2 , and RPD values, and low RMSEC and RMSEP
221
values.
222
3. Results and discussion
223
3.1. Spectral features of tomato paste
224
Average reflectance spectra in the wavelength range of 405-970 nm of the
225
examined pure tomato paste, and sucrose and tomato paste mixtures with different
226
proportions of sucrose are shown in Fig. 2. For the low sucrose proportions under
227
investigation, the mean spectra of sucrose and tomato paste mixtures were very
228
similar to the spectra of pure tomato paste across the whole tested wavelength
229
region. Furthermore, the reflectance values of the samples decreased with the
230
increase of sucrose proportions. A main absorption band observed between 940 and
231
970 nm related to O-H third stretching overtone was assigned to water molecule
232
(Wu et al., 2008). This observation suggests that the spectral evaluation may allow
233
effective determination of sucrose proportions in tomato paste. However, it is
234
difficult to determine the sucrose proportion directly from the observed spectra.
235
Especially when a lot of samples were measured simultaneously, their spectral
236
curves showed many overlaps, making the quantitative determination even more
ACCEPTED MANUSCRIPT 237
complicated. In order to solve this problem, multivariate data analysis method was
238
introduced to analyze the multiple variables of spectral data.
239
3.2. Quantification of sucrose proportion in tomato paste
240
The main purpose of the investigation was to quantify the level of adulteration
241
in tomato paste. Quantitative calibration models were generated using PLS, LS-
242
SVM and BPNN mathematical algorithms. Through comparison of the value of
243
PRESS, the number of latent variables for PLS model both were determined as 5 for
244
predicting sucrose proportions in both batches of tomato paste. Based on the
245
spectral information, the performance of these three models for predicting sucrose
246
proportions in tomato paste was shown in Table 1. Compared with PLS and BPNN
247
models, the LS-SVM model had the best predictive accuracy with RC2 of 0.990, RP2
248
of 0.936, RMSEC of 0.263%, bias of 0.325%, and RMSEP of 0.521% for batch 1 of
249
tomato paste. Furthermore, the largest RPD of 5.014 showed that the LS-SVM
250
model was considered to be adequate for sucrose proportions prediction. Although
251
the RC2 and RMSEC in BPNN model were similar to those in LS-SVM model, the
252
RMSEP was higher and the RP2 and RPD were lower indicating that the BPNN
253
model was slightly less good for sucrose proportions prediction. Regarding the PLS,
254
the RMSEP was higher and RPD was lower than LS-SVM and BPNN models in
255
prediction set indicating that the model was less good. While the RMSEP was
256
slightly larger than RMSEC with lower bias in PLS indicated that the PLS is a
257
potential better-behaved model for practical application.
258
Similar results were also obtained for batch 2 of tomato paste. In general, the
259
LS-SVM model gave better prediction abilities with RP2 of 0.966 than PLS and
260
BPNN models. The corresponding values for RMSEP and RPD were found to be
261
0.445% and 5.865, respectively. Therefore, LS-SVM was considered as the best
ACCEPTED MANUSCRIPT 262
way for establishing the quantitative model of sucrose proportion in tomato paste
263
and had the best generalization capability. This was due to the fact that LS-SVM
264
can learn in high-dimensional characteristic space with fewer training samples. It is
265
worth mentioning that PLS, although used extensively in chemometrics, did not
266
perform always as well as LS-SVM and BPNN, at least for this particular case. The
267
obtained prediction results confirmed the suitability of multispectral imaging for
268
determination of sucrose proportion in tomato paste in a rapid and non-destructive
269
manner.
270
Furthermore, the results from the ReliefF method showed that the 630, 645,
271
700, 780 and 850 nm for batch 1 and 630, 645, 850, 890 and 940 nm for batch 2 of
272
tomato paste were more important than other wavelengths for sucrose detection.
273
The results were consistent with previously reported result, that the combination of
274
NIR wavelengths (λ= 730, 830, 909-915, 960, and 965 nm) within C-H and O-H
275
bands can reliably quantify sucrose (Omar et al., 2012).
276
3.3. PCA data analysis
277
In the study, PCA was performed initially to visualize any variation among pure
278
tomato paste, tomato paste containing 1% and 9% sucrose (the minimum and
279
maximum proportions in the current study, respectively) in principal component
280
(PC) space. Fig. 3 shows the two-dimensional score plot of the first two principal
281
components (PCs) from the spectral reflectance values extracted from the
282
multispectral images of the samples. The first two PCs, which account for the most
283
spectral variations above 97% (97.96% and 98.87% in batch 1 and batch 2 of
284
tomato paste, respectively), were used to make differentiation clear. The results
285
showed that the PCA method can clearly identify the pure tomato paste, as well as
286
1% and 9% sucrose adulteration in tomato paste.
ACCEPTED MANUSCRIPT 287
3.4. Detection limit of sucrose proportions in tomato paste
288
Currently although there are no international standards for acceptable levels of
289
sucrose in tomato paste yet, tomato paste mixed with sucrose seriously endangering
290
the interests of consumers and market order. Therefore, identification of adulterated
291
tomato paste is very important. Especially, the ability to differentiate between
292
unadulterated tomato paste and those containing the 1% sucrose (the lowest
293
proportion in the current study) is crucial. Spectra of pure tomato paste and tomato
294
paste containing 1% sucrose were collected using multispectral imaging system and
295
analyzed using PLSDA, LS-SVM and BPNN models. All three models were able to
296
distinguish between adulterated (1% sucrose) and unadulterated tomato paste. The
297
high classification accuracies of 100% were obtained in prediction set for both
298
batches of tomato paste using PLSDA, LS-SVM or BPNN model. The results show
299
that the detection limit of sucrose in tomato paste was 1% using multispectral
300
imaging combined with chemometric methods.
301
4. Conclusions
302
The detection of adulteration is attracting much attention for tomato paste
303
manufacturer, researchers and consumers. This research was carried out to
304
investigate the prediction performance of the multispectral imaging for rapid
305
detection of sucrose adulteration in tomato paste. Better prediction performance and
306
generalization ability for quantification of sucrose proportions in both batches of
307
tomato paste were achieved using LS-SVM. Besides, multispectral imaging was
308
able to detect sucrose adulteration in tomato paste at the 1% level. The study
309
confirmed that multispectral imaging technology in tandem with chemometric
310
method has the potential to be a rapid and non-destructive method to accurately
311
identification of the adulteration proportions in tomato paste. Further study is
ACCEPTED MANUSCRIPT 312
needed to improve the adaptability of the LS-SVM model or to develop new
313
mathematical model for the practical applications to detect adulteration in tomato
314
paste.
315 316
Acknowledgements
317
This work was supported by the National Key Research and Development Plan
318
of China (2016YFD0401104), the National Natural Science Foundation of China
319
(31401544), the Key Science & Technology Specific Projects of Anhui Province
320
(15czz03117), the Funds for Huangshan Professorship of Hefei University of
321
Technology (407-037019), and the Fundamental Research Funds for the Central
322
Universities (JZ2016HGTB0712).
323 324
References
325
Bayod, E., Willers, E.P., Tornberg, E., 2008. Rheological and structural
326
characterization of tomato paste and its influence on the quality of ketchup.
327
LWT-Food Sci. Technol. 41, 1289-1300.
328
Berta, M., Wiklund, J., Kotzé, R., Stading, M., 2016. Correlation between in-line
329
measurements of tomato ketchup shear viscosity and extensional viscosity. J.
330
Food Eng. 173, 8-14.
331
Borin, A., Ferrão, M.F., Mello, C., Maretto, D.A., Poppi, R.J., 2006. Least-squares
332
support vector machines and near infrared spectroscopy for quantification of
333
common adulterants in powdered milk. Anal. Chim. Acta 579, 25-32.
334
Burton‐Freeman, B., Talbot, J., Park, E., Krishnankutty, S., Edirisinghe, I., 2012.
335
Protective activity of processed tomato products on postprandial oxidation and
336
inflammation: a clinical trial in healthy weight men and women. Mol. Nutr.
ACCEPTED MANUSCRIPT 337
Food Res. 56, 622-631.
338
Cheng J.H., Sun D.W., 2014. Hyperspectral imaging as an effective tool for quality
339
analysis and control of fish and other seafoods: current research and potential
340
applications. Trends Food Sci. Tech. 37, 78-91.
341
Cortes, C., Vapnik, V., 1995. Support vector network. Mach. Learn. 20, 273-297.
342
Dai, H., MacBeth, C., 1997. Effects of learning parameters on learning procedure
343
and performance of a BPNN. Neural Networks 10, 1505-1521.
344
Dale, L.M., Thewis, A., Boudry, C., Rotar, I., Dardenne, P., Vincent, B., Fernández,
345
P.J.A., 2013. Hyperspectral imaging applications in agriculture and agro-food
346
product quality and safety control: A review. Appl. Spectrosc. Rev. 48, 142-
347
159.
348
Devos, O., Ruckebusch, C., Durand, A., Duponchel, L., Huvenne, J.P., 2009.
349
Support vector machines (SVM) in near infrared (NIR) spectroscopy: focus on
350
parameters optimization and model interpretation. Chemometr. Intell. Lab. 96,
351
27-33.
352
Dissing, B.S., Nielsen, M.E., Ersbøll, B.K., Frosch, S., 2011. Multispectral imaging
353
for determination of astaxanthin concentration in salmonids. PLoS ONE 6,
354
e19032.
355
Feng, Y.Z., Sun, D.W., 2012. Application of hyperspectral imaging in food safety
356
inspection and control: a review. Crit. Rev. Food Sci. Nutr. 52, 1039-1058.
357
Fu, X., Kim, M.S., Chao, K., Qin, J., Lim, J., Lee, H., Garrido-Varo, A., Pérez-
358
Marín, D., Ying, Y., 2014. Detection of melamine in milk powders based on
359
NIR hyperspectral imaging and spectral similarity analyses. J. Food Eng. 124,
360
97-104.
361
Huang, M., Kim, M.S., Delwiche, S.R., Chao, K., Qin, J., Mo, C., Zhu, Q., 2016.
ACCEPTED MANUSCRIPT 362
Quantitative analysis of melamine in milk powders using near-infrared
363
hyperspectral imaging and band ratio. J. Food Eng. 181, 10-19.
364
Kamruzzaman, M., Makino, Y., Oshita, S., 2015. Rapid and non-destructive
365
detection of chicken adulteration in minced beef using visible near-infrared
366
hyperspectral imaging and machine learning. J. Food Eng. 170, 8-15.
367
Kamruzzaman, M., Sun, D.W., ElMasry, G., Allen, P., 2013. Fast detection and
368
visualization of minced lamb meat adulteration using NIR hyperspectral
369
imaging and multivariate image analysis. Talanta 103, 130-136.
370
Lim, J., Kim, G., Mo, C., Kim, M., Chao, K., Qin, J., Fu, X., Baek, I., Cho, B.K.,
371
2016. Detection of melamine in milk powders using near-infrared
372
hyperspectral imaging combined with regression coefficient of partial least
373
square regression model. Talanta 151, 183-191.
374
Liu, C., Liu, W., Chen, W., Yang, J., Zheng, L., 2015. Feasibility in multispectral
375
imaging for predicting the content of bioactive compounds in intact tomato
376
fruit. Food Chem. 173, 482-488.
377
Liu, C., Liu, W., Yang, J., Chen, Y., Zheng, L., 2017. Non-destructive detection of
378
dicyandiamide in infant formula powder using multi-spectral imaging coupled
379
with chemometrics. J. Sci. Food Agr. 97, 2094-2099.
380
Liu, J., Cao, Y., Pan, W., Ma, F., Liu, C., Chen, W., Yang, J., Zheng, L., 2016.
381
Rapid and non-destructive identification of water-injected beef samples using
382
multispectral imaging analysis. Food Chem. 190, 938-943.
383
Lleó, L., Barreiro, P., Ruiz-Altisent, M., Herrero, A., 2009. Multispectral images of
384
peach related to firmness and maturity at harvest. J. Food Eng. 93, 229-235.
385
Løkke, M.M., Seefeldt, H.F., Skov, T., Edelenbos, M., 2013. Color and textural
386
quality of packaged wild rocket measured by multispectral imaging.
ACCEPTED MANUSCRIPT 387
Postharvest Biol. Technol. 75, 86-95.
388
Omar, A.F., Atan, H., MatJafri, M.Z., 2012. Peak response identification through
389
near-infrared spectroscopy analysis on aqueous sucrose, glucose, and fructose
390
solution. Spectrosc. Lett. 45, 190-201.
391
Panagou, E.Z., Mohareb, F.R., Argyri, A.A., Bessant, C.M., Nychas, G.J.E., 2011.
392
A comparison of artificial neural networks and partial least squares modelling
393
for the rapid detection of the microbial spoilage of beef fillets based on Fourier
394
transform infrared spectral fingerprints. Food Microbiol. 28, 782-790.
395
Panagou, E.Z., Papadopoulou, O., Carstensen, J.M., Nychas, G.J.E., 2014. Potential
396
of
multispectral
imaging
technology
for
rapid
and
non-destructive
397
determination of the microbiological quality of beef filets during aerobic
398
storage. Int. J. Food Microbiol. 174, 1-11.
399
Pu, H., Kamruzzaman, M., Sun, D.W., 2015. Selection of feature wavelengths for
400
developing multispectral imaging systems for quality, safety and authenticity
401
of muscle foods-a review. Trends Food Sci. Tech. 45, 86-104.
402
Qin, J., Chao, K., Kim, M.S., Lu, R., Burks, T.F., 2013. Hyperspectral and
403
multispectral imaging for evaluating food safety and quality. J. Food Eng. 118,
404
157-171.
405
Rizwan, M., Rodriguez-Blanco, I., Harbottle, A., Birch-Machin, M.A., Watson,
406
R.E.B., Rhodes, L.E., 2011. Tomato paste rich in lycopene protects against
407
cutaneous photodamage in humans in vivo: a randomized controlled trial. Brit.
408
J. Dermatol. 164, 154-162.
409
Ropodi, A.I., Panagou, E.Z., Nychas, G.J.E., 2016. Multispectral imaging (MSI): A
410
promising method for the detection of minced beef adulteration with
411
horsemeat. Food Control 73, 57-63.
ACCEPTED MANUSCRIPT 412
Ropodi, A.I., Pavlidis, D.E., Mohareb, F., Panagou, E.Z., Nychas, G.J., 2015.
413
Multispectral image analysis approach to detect adulteration of beef and pork
414
in raw meats. Food Res. Int. 67, 12-18.
415
Sinelli, N., Spinardi, A., Di Egidio, V., Mignani, I., Casiraghi, E., 2008. Evaluation
416
of quality and nutraceutical content of blueberries (Vaccinium corymbosum
417
L.) by near and mid-infrared spectroscopy. Postharvest Biol. Technol. 50, 31-
418
36.
419
Wu, D., He, Y., Feng, S., 2008. Short-wave near-infrared spectroscopy analysis of
420
major compounds in milk powder and wavelength assignment. Analytica
421
Chimica Acta 610, 232-242.
422
Wu, D., Shi, H., He, Y., Yu, X., Bao, Y., 2013. Potential of hyperspectral imaging
423
and multivariate analysis for rapid and non-invasive detection of gelatin
424
adulteration in prawn. J. Food Eng. 119, 680-686.
425
Xie, L., Ying, Y., Ying, T., Yu, H., Fu, X., 2007. Discrimination of transgenic
426
tomatoes based on visible/near-infrared spectra. Anal. Chim. Acta 584, 379-
427
384.
428
Zhang, L., Schultz, M.A., Cash, R., Barrett, D.M., McCarthy, M.J., 2014.
429
Determination of quality parameters of tomato paste using guided microwave
430
spectroscopy. Food Control 40, 214-223.
431
Zheng, J., Song, Z., Ge, Z., 2016. Probabilistic learning of partial least squares
432
regression model: Theory and industrial applications. Chemometr. Intell.
433
Lab. 158, 80-90.
ACCEPTED MANUSCRIPT
435 436
Fig. 1. Principal setup of the multispectral imaging system. An integrating sphere
437
with a matte white coating ensures optimal lighting conditions. The light emitting
438
diodes (LEDs) located in the rim of the sphere ensures narrowband illumination.
439
The image acquisition is performed by a monochrome grayscale CCD camera
440
mounted in the top of the sphere.
441
ACCEPTED MANUSCRIPT
443 444
Fig. 2. Average reflection spectra of pure tomato paste and different proportions of
445
sucrose in batch 1 (a) and batch 2 (b) of tomato paste, respectively.
ACCEPTED MANUSCRIPT
447 448
Fig. 3. Two-dimensional score plot of the first two principal components conducted
449
on the spectral data in batch 1 (a) and batch (b) of tomato paste. The classification
450
was clear among the three groups of the pure tomato paste (○), 1% sucrose in
451
tomato paste (*) and 9% sucrose in tomato paste (+).
ACCEPTED MANUSCRIPT 453
Table 1
454
Performance of PLS, LS-SVM and BPNN models for predicting sucrose
455
proportions (%) in tomato paste. Batch Batch 1
Batch 2
Sources of sucrose
Chemometrics
RC2
RMSEC
RP2
Bias
RMSEP
RPD
Analytical
PLS
0.949
0.569
0.927
0.262
0.630
4.140
grade
LS-SVM
0.990
0.263
0.936
0.325
0.521
5.014
BPNN
0.990
0.259
0.919
0.322
0.620
4.210
PLS
0.894
0.800
0.878
0.454
0.703
3.715
LS-SVM
0.994
0.205
0.966
0.066
0.445
5.865
BPNN
0.958
0.511
0.910
0.462
0.548
4.767
Sugarcane
2 C
2 P
456
R , coefficient of determination in calibration; R , coefficient of determination in prediction;
457
RMSEC, root mean square error of calibration; RMSEP, root mean square error of prediction;
458
RPD, residual predictive deviation.
ACCEPTED MANUSCRIPT Research highlights
1. Multispectral imaging could determine sucrose proportion in tomato paste accurately. 2. Multispectral imaging was able to detect low proportion sucrose in tomato paste. 3. Multispectral imaging successfully used as a rapid screening technique for sucrose.