Food Chemistry 144 (2014) 80–86

Contents lists available at SciVerse ScienceDirect

Food Chemistry
journal homepage: www.elsevier.com/locate/foodchem

Monitoring methods and predictive models for water status in Jonathan

apples
Lucia Carmen Trincă a,⇑, Adina-Mirela Căpraru a, Dragosß Arotăritßei b, Irina Volf c, Ciprian Chirutßă a
a
‘‘Ion Ionescu de la Brad’’ University of Agricultural Sciences and Veterinary Medicine, Faculty of Horticulture, Str. Aleea M. Sadoveanu, No. 3, 700490 Iasi, Romania
b
‘‘Grigore T. Popa’’ University of Medicine and Pharmacy, Faculty of Medical Bioengineering, Iasi, Str. Universitătßii, No. 16, 700115 Iasi, Romania
c
‘‘Gheorghe Asachi’’ Technical University of Iasi, Faculty of Chemical Engineering and Environmental Protection, 73 Prof. dr. Doc. Dimitrie Mangeron Street, 700050 Iasi, Romania

a r t i c l e i n f o a b s t r a c t

Article history: Evaluation of water status in Jonathan apples was performed for 20 days. Loss moisture content (LMC)
Available online 6 June 2013 was carried out through slow drying of wholes apples and the moisture content (MC) was carried out
through oven drying and lyophilisation for apple samples (chunks, crushed and juice).
Keywords: We approached a non-destructive method to evaluate LMC and MC of apples using image processing
Water status and multilayer neural networks (NN) predictor. We proposed a new simple algorithm that selects the tex-
Moisture content ture descriptors based on initial set heuristically chosen. Both structure and weights of NN are optimised
Jonathan apples
by a genetic algorithm with variable length genotype that led to a high precision of the predictive model
Imagistic analysis
Neural network
(R2 = 0.9534).
In our opinion, the developing of this non-destructive method for the assessment of LMC and MC (and
of other chemical parameters) seems to be very promising in online inspection of food quality.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction normal process of drying in the oven, or with IR radiation (El-Sayd

Water is the main component of the most foods. Depending on
the availability and biological activity, water’s status includes free
water and bound water forms (Goñi, Muñoz, Ruiz-Cabello, Escrib-
ano, & Merodio, 2007; Rahman, 2006). Currently, the scientific lit-
erature presents characterization of water forms from a level of
general concept (Schmidt, 2007) with limitations determined by
the specialized field approached (biochemically, physiologically
or technologically). In food, generally free water is associated with
the loss of the moisture content (LMC) while bound water is deter-
mined as moisture content (MC).
In food industry, LMC provides information for the rate of the
metabolic and physiologic processes (perspiration, respiration)
that continue in vegetal products even after harvesting. Also,
LMC is correlated with the weight loss, the changes of layout/shape
or the variations of the biochemical composition for dried/dehy-
drated products (Jan & Rab, 2012) and it’s easily determined as
the weight loss (of the food product) by slow drying (Veraverbeke,
Verboven, Van Oostveldt, & Nicolaï, 2003).
Food freshness can be associated with MC. To determine food
moisture content, most standard methods rely either on the
⇑ Corresponding author. Tel.: +40 745320134.
E-mail addresses: lctrinca@yahoo.com (L.C. Trincă), amcapraru@yahoo.com
(A.-M. Căpraru), dragos_aro@yahoo.com (D. Arotăritßei), inavolf@yahoo.com (I. Volf),
chiruta_c@yahoo.com (C. Chirutßă).
& Makawy, 2010). Classical methods are time-consuming and do
not ensure high-precision determination because of the potential
errors induced by the too many factors involved in the drying pro-
cess (Bradley, 2010; Fontana, 2007). The modern methods are fast
but require calibration based on the reference method (Isengard,
Merkh, Schreib, Labitzke, & Dubois, 2010). These drawbacks ex-
plain the general need to look for modern and innovative methods
either for rapid evaluation or precise prediction of food quality.
One of the most suitable non-destructive techniques relies on im-
age processing (Gonzalez & Woods, 2008). Imagistic analysis
parameters can be used as inputs in neural network (NN) predic-
tive models (Haykin, 1999) in order to reduce the importance of
experimental errors in the analytic process.
In our approach, we proposed a non-destructive method to
evaluate LMC from apples using image processing and multilayer
neural networks predictor. The NN predictor was also used for pre-
diction of MC obtained for various samples and methods. We ap-
proached a new simple algorithm that selects the texture
descriptors based on initial set, heuristically chosen. Both structure
and weights of NN are optimised by a genetic algorithm with var-
iable length genotype.
In our opinion, the developing of this non-destructive method
for the assessment of MC and LMC (and of other chemical param-
eters) seems to be very promising in online inspection of food
quality.

0308-8146/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.foodchem.2013.05.131
 L.C. Trincă et al. / Food Chemistry 144 (2014) 80–86
2. Material and methods
Apples are very popular fruits in Europe. There are many variety
of apple fruits having different colours (green, red, yellow, or
mixed red-yellow) and different changes in aspect and structure
during storage.
There are many approaches that deal with the subject of drying
fruits, most of them being focused on pieces of fruits (Fernández,
Castillero, & Aguilera, 2005; Seiiedlou, Ghasemzadeh, Hamdami,
Talati, & Moghaddam, 2010). Our approach is focused on the mod-
ification of the moisture content in apples during storage at ordin-
ary temperature.
At early November, 100 healthy-looking Jonathan apples, that
weren’t exposed to preservation treatments, were selected ran-
domly from the same source, with roughly the same shape and
weight (91–103 g), as being the gross sample of analysis. Chemical
and imagistic tests have been conducted every 5 days: lots of 15
apples were investigated in every testing day for MC evaluation.
Also, 15 other apples were analysed all experimental period for
LMC determination. All apples were stored during experimental
period at 20 °C room temperature and 70% relative humidity.
A balance type RADWAG AS 220/C/2 with accuracy of 104 g
was used for weighing, a SLW 115 ECO drying oven was used for
drying apple samples at temperature of 90 °C and a Freeze Dryer
ALPHA type 1–4 LD was used for lyophilisation.
For image capturing an AF-S DX Zoom-Nikon ED 18–70 mm, f/
3.5–4.5 GIF was used.
2.1. Methods for chemical analysis
Analyses were performed for a period of 20 days. For LMC eval-
uation each entire apple was weighed and photographed.
Also, MC analysis had involved the weighing and photographing
of each lot of 15 apples, in the first step. For MC evaluation, in the
second step, chemical analyses were performed on the samples
immediately obtained by processing the apple as chunks, crushed
and juice. The apple has been peeled and then has been shredded
with a special device to get apple chunks (0.4  2 cm). Also, peeled
apple was crushed in order to get crushed apple or pressed to get
apple juice.
2.1.1. Loss moisture content determination
At every 5 days, apples were weighed regularly. By weighing
the difference was attributed to the loss of moisture content
(LMC, %), in accordance with the Eq. (1):
LMC; % ¼ 100  ðmi  mf Þ=mi ð1Þ
where mi = initial mass of sample (g), mf = final mass of sample (g).
2.1.2. Oven drying
5 g of fresh sample (each apple was processed to get chunks,
crushed or juice) were dried at 90 °C until a constant mass
6102 g was reached. Moisture content (MC, %) was determined
in accordance with the Eq. (2):
MC; % ¼ 100  ðm  m1 Þ=m2 ð2Þ
where m = mass of the sample before drying (g), m1 = mass of the
sample after drying (g), m2 = mass of the analysed sample (g).
2.1.3. Freeze drying (lyophilisation)
Lyophilisation is dehydration process at cold: water freezes fas-
ter than other ingredients and is removed in the form of ice with-
out changing the food structure.
Lyophilisation was carried out in a type 1–4 LD plus, Freeze-
Dryer ALPHA.
81
5 g of fresh apple sample (chunks, crushed or juice) were sub-
jected to freeze drying at temperature of 50 to 60 °C, and pres-
sure ranging from 0.02 to 0.03 mbar until constant mass. Water
content was determined using the Eq. (2).
2.2. Methods for imagistic analysis
Non-destructive methods are usually based on image process-
ing techniques (Gonzalez & Woods, 2008; Petrou & Sevilla, 2006;
Yam & Papadakis, 2004). The images are acquired under special
conditions (illumination, angle between axis of camera lens and
the lighting source – 45°, in order to capture the diffused reflection,
colour temperature of the fluorescent lamp and so on.) in accor-
dance with the image’s acquisition standards (Francis, 2005)
Analysis of texture can offer useful information about quality of
fruits (Gonzalez & Woods, 2008). At apples, the wrinkles usually
denote long time storage: more wrinkles involve a larger storage
period and a higher loss of the moisture content. Also, it looks like
wrinkle’s texture parameters are related with the variation of LMC
(Vesali, Gharibkhani, & Komarizadeh, 2011)
Fourier descriptors for 2-D digital images seem to be useful for
characterization of LMC. However, usually way to use Fourier
descriptors are related to shape shrinking due to drying process.
For rectangular pieces of apples, the change in shape due to drying
process is significant. The evolution of shape in time can be mea-
sured if the shape is coded in a vector form by Fourier descriptors
(Fernández et al., 2005). In our case, these descriptors doesn’t offer
useful information related to LMC due faintly the fact that Jonathan
apple variety provides in time a very small change in their shape.
The experiment achievement until the moment that the shape of
apple significantly changed was not of interest to us, because in
this case, the apple cannot be accepted by consumers.
Unlike the data presented in the literature (Fernández et al.,
2005; Seiiedlou et al., 2010) for which investigations have been
conducted on apple’s cuticle, our experiments focused on the im-
age acquired for the entire apple.
Tests have been conducted every five days: lots of 15 apples
were investigated in every testing day for MC evaluation. Also,
15 other apples were analysed all experimental period for LMC
determination. For LMC evaluation each entire apple was photo-
graphed. Also, MC analysis had involved the photographing of each
lot of 15 apples, in the first step.
For image capturing an AF-S DX Zoom-Nikon ED 18–70 mm, f/
3.5–4.5 GIF was used. The pictures were saved in RGB colour space
with 1936  1926 pixels size. In order to ensure the same calibra-
tion of dimension in all pictures, a box was designed
(16  16  32 cm) to provide a fixed distance between photo cam-
era and apples.
The photos were processed and specific texture parameters
were determined (see discussion at Section 3.2.) in order to build
NN predictive models for LMC and MC of apples.
2.3. Predictive methods
2.3.1. Neural networks predictive methods
Neural networks (NN) proven to be universal approximators
(Haykin, 1999). In 1989, Cybenko demonstrated, using a Kolmogo-
rov’s older result, that multilayer feed-forward network with a sin-
gle hidden layer, which contains finite number of hidden neurons,
is a universal approximator (Cybenko, 1989). The demonstration
has been made for sigmoid activation function. In the most com-
mon sense, the universal approximator can approximate any
non-linear function with a desired precision, if its architecture is
large enough. We will use a multilayer neural network (multilayer
perceptron, MLP) as predictor for LMC and MC. The structure and
weights of MLP are simultaneous adapted using a genetic
82 L.C. Trincă et al. / Food Chemistry 144 (2014) 80–86
Fig. 1. The coded chromosome for layer 1.
algorithm with variable length genotype (Munir-ul, Chowdhury, &
Yun Li, 1996).
2.3.2. Genetic algorithm with variable length genotyp
Genetic algorithms have been proved to be useful tools for a
large class of optimization of real-world problems. We propose
to use genetic algorithm with variable length genotype (GAVLG)
with known sequence of operation: selection (based on tourna-
ment selection), crossover, and mutation (Munir-ul et al., 1996).
The fitness function (measure of performance) is chosen to be
ffitness = 1  R2, and the objective of algorithm is the minimization
of ffitness (maximisation of R2).
The NN is coded by three chromosomes, corresponding to
three levels of weights. Each chromosome is composed by weights
from current layer and each weight is coded in binary 16 bits
length. Each weight is scaled in [20.0, +20.0] range. The length
of chromosome is variable, according to NN architecture. For prac-
tical calculus, the number of neurons in each layer is set to be in
the range [2, 64]. Let’s denote with wkji the weight from layer k,
k = {1, 2, 3} that connect the weight from layer k and node j to
a node in the layer k-1 and node i. By convention, the input layer
has k = 0. The chromosome is code as in Fig. 1 where each wkji
weight is coded as a string of 16 sequences of ‘0’ and ‘1’ bits
(Fig. 1). The encode/decode procedure maps two-way binary-float
weight.
The crossover is made between parents with variable length
genotype. We propose to use for crossover a modified cut-and-
splice operator from messy genetic algorithm proposed by Mu-
nir-ul et al. (1996).
The cut operator splits the chromosomes at a random chosen
position with cutting probability Pc and the splice operator concat-
enates two strings in a randomly chosen order with probability Ps
(Fig. 2).
It is clear that the offspring can contain a residual gene (a gene
that has a length lower than 16 bits and it cannot be codified as a
weight (this is a residual gene) and this gene will be deleted.
3. Results and discussion
3.1. Methods for chemical analysis
Dynamics of loss moisture content (LMC) for the set of 15 ap-
ples is presented in Fig. 3. One may observe from Fig. 3, a linear in-
crease of loss moisture content, which ranged between 1.19–1.56%
(after 5 days), 2.05–3.44% (after 10 days), 2.83–4.42% (after
15 days) and 4.23–6.42% (after 20 days). We consider LMC, as
being ‘‘free-water form’’ of the apple which results by perspiration,
a normal physiological process. Loss of water occurs through lenti-
cels, stomata, calyx, peduncle or injured tissues and decreases tis-
sues turgidity which it become soft and with reduced juice.
Harvested fruits are deprived of the possibility of recovering the
water lost through perspiration, what causes irreversible organo-
leptic and biochemical changes (Jan & Rab, 2012).
Vesali et al. (2011) reported for LMC, (%) determined
experimentally on Golden Delicious apples, stored at 20 °C, varia-
tions between 3.2–10.4% (after 8 days), 9.1–23.7% (after 16 days)
and 15.5–31.4% (after 24 days). Jan and Rab (2012), reported for
Golden Delicious harvested at commercial maturity and stored at
5 ± 1 °C with 60–70% relative humidity, a 1.06% weight loss after
30 days and a 4.53% weight loss after 150 days.
These significant differences can be explained by the peculiari-
ties of these two varieties: Golden Delicious apples, kept in regular
temperature, allow release of free water in greater proportion
through the many free pore spaces (Menges & Ertekin, 2006; Rah-
man, 2006). Unlike Golden apples, those of Jonathan variety can be
stored at 20 °C, for a long time, without significant loss of water
content, probably due to the naturally thicker waxy layer (Vera-
verbeke, Lammertyn, Saevels, & Nicolaï, 2001), characteristic of
this cultivar.
In order to see the influence of the way of preparing the sam-
ples, MC determination considered different type of apple samples
(chunks, crushed, juice). Also, for the same types of samples (apple
chunks) were used two different methods (oven drying and lyo-
philisation) in order to appreciate the resulting differences for
moisture content determination.
As it can be observed in Table 1, moisture content decreased
over time, for all types of apple samples, with variation between
1.15% and 2.46%. We consider MC, as being the ‘‘main bound –
water form’’ of the apple. The reduction of bound water content
was caused by water’s loss, mainly in the form of free water, if
we consider the bigger variation of loss moisture content (1.42–
5.36%). It seems that intrinsic relationships between respiration
and perspiration rata affects water loss and moisture content (Er-
turk, Akbuclak, & Ozer, 2003; Jan & Rab, 2012).
One may see from Fig. 4 that dynamics of the moisture content
presents variations due to the different ways of samples prepara-
tion. For example, compared to chunks apple, when prepares
crushed or juice apples, significant changes of the water status
may happen through destruction of cellular compartments. The
parenchyma impairment (Hills & Remigereau, 1997) increases
the general content of water that can be determined, in this case,
as MC. Literature presents data about inter-conversion of water
forms when vegetable food substrates are subjected to drying by
different techniques (Derossi, Severini, & Cassi, 2011; Okos et al.,
2007).
Also, when two methods of moisture content determination
were investigated different results were obtained. For sure, the
lowest values of MC registered in the case of lyophilisation
(Fig. 4) can be explained considering that freeze drying in high vac-
uum reduces at minimum the chemical alterations implied in er-
rors occurrence in oven drying method.
Correlation coefficient (DeVore & Berk, 2007) was calculated
through 2012 The MathWorks, Inc. MATLAB program (http://
www.mathworks.com). Small values of the correlation coefficient
(R2) for the entire experimental period showed insignificant
correlations between variations of the two forms of water
(LMC/MC or MC/MC for all types of samples/methods combina-
tions). These results can be explained if considering the meta-
bolic processes (e.g. respiration) of the main biochemical
compounds that lead to the formation of endogenous water
(Nguyen, Dresselaers, & Verboven, 2006) and low molecular
weight compounds (slightly volatile), which modifies the water
content/status (Nguyen et al., 2006; Veraverbeke et al., 2003)
and affects the weight of apple.
Significant differences of the dynamics and the inconclusive sta-
tistical results of the chemical analysis of LMC and MC led us to the
necessity to establish other different correlations between chemi-
cal and imagistic parameters, through predictive models.
3.2. Imaging analysis method
It is normal to suppose LMC is somewhat related to volume of
fruit. A factor that uses an approximation for apple volume is pro-
posed to use as descriptor. In 3-D diffusion approached, in order to
simplify the problem the objects are usually approximated by a
regular 3-D form: plate, parallelepiped, cube, and sphere.
 L.C. Trincă et al. / Food Chemistry 144 (2014) 80–86 83

Fig. 2. Cut-and-splice operator.

6.5 Apple1
Apple2
6
Apple3
5.5
Apple4
5 Apple5
LMC, (%)

4.5 Apple6
Apple7
4
Apple8
3.5 Apple9
3 Apple10

2.5 Apple11
Apple12
2
Apple13
1.5 Apple14

1 Apple15
Day 5 Day 10 Day 15 Day 20 Average
Period of time, (days)

Fig. 3. Dynamics of loss moisture content (LMC) for the set of 15 apples.

Table 1
Loss moisture content (LMC,%) and moisture content (MC,%) of apples samples during the experiment.

Method Sample Value Day 0 Day 5 Day 10 Day 15 Day 20

Loss moisture content (LMC), %
Slow drying Whole apples Minimum 0 1.199 2.051 2.833 4.234
Maximum 0 1.565 3.442 4.420 6.427
Average ± standard deviation 0 1.423 ± 0.148 2.878 ± 0.546 3.671 ± 0.668 5.363 ± 0.944

Moisture content (MC), %

Oven drying Chunksapples Minimum 78.867 78.925 76.865 71.048 73.166
Maximum 79.251 80.578 80.655 80.580 81.208
Average ± standard deviation 79.100 ± 0.161 79.801 ± 0.587 79.543 ± 1.699 77.690 ± 3.885 77.193 ± 2.848
Crushed apples Minimum 79.908 71.998 77.693 77.075 76.105
Maximum 80.198 86.264 80.274 80.133 81.653
Average ± standard deviation 80.134 ± 0.111 79.802 ± 5.106 79.384 ± 1.035 78.925 ± 1.612 77.666 ± 2.379
Juice apples Minimum 78.998 77.998 79.543 76.556 77.349
Maximum 83.415 83.372 80.655 82.965 80.871
Average ± standard deviation 80.633 ± 1.725 80.424 ± 1.982 80.262 ± 0.484 80.209 ± 2.324 79.482 ± 1.366
Lyophilisation Chunksapples Minimum 78.251 75.432 77.020 77.202 75.457
Maximum 84.588 83.854 79.888 79.216 78.340
Average ± standard deviation 79.999 ± 1.836 78.340 ± 3.248 78.025 ± 1.211 77.923 ± 0.960 77.528 ± 1.185

Jonathan apple’s weight loss explains the LMC increasing,
although a very significant change doesn’t appeared in their
dimension.
During the experiment, images of 90 wholes apple were cap-
tured and the pictures were saved in RGB colour space with
1936  1926 pixels size (Fig. 5). From each apple four equal rectan-
gular areas that not include stem area (pedicel area) were denoted,
so, will result finally 90  4 = 360 samples.
The apples are approximated by spheres and the radius of this
sphere is used as form descriptor (FD) for prediction of LMC. The
84 L.C. Trincă et al. / Food Chemistry 144 (2014) 80–86

82 X
L
m¼ zi pðzi Þ ð3Þ
81 i¼1
1
MC (%)

80 S¼1 ð4Þ
1 þ r2
79
X
L

78
l3 ¼ ðzi  IÞ3 pðzi Þ ð5Þ
i¼1
77 X
L
0 5 10 15 20 U¼ p2 ðzi Þ ð6Þ
Period of time (days) i¼1

Chunks Crushed X
L

Juice Lyophilizaon e ¼  pðzi Þlog2 pðzi Þ ð7Þ

i¼1
Fig. 4. Dynamics of moisture content (MC) for apple samples during experiment.
In the Eqs. (3)–(7), zi is the random variable denoting intensity
level, p(zi), i = 0, 1, 2,. . .., L  1, is the corresponding histogram for L
number of discrete intensity levels and r is the variance.
Imaging analysis parameters (m, S, l3, U, e,) and FD were deter-
mined for all Jonathan apple lots considered and further used to
build neural network predictive models.

3.3. Predictive methods

3.3.1. Neural network predictive models

Fig. 5. Sample of an apple in RGB format.
red–green–blue (RGB) image in jpg format is segmented and the
contour is detected. We selected the green (G) matrix from RGB
image due to fact that G colour is less sensitive to noise. The con-
tour of the apple is approximated by a circle using least mean
square (LMS) method (Fig. 6).
It is obvious from Fig. 6 that for this variety of apples, the con-
tour can be more or less from a circle with a greater error than for
Golden Delicious apples (Vesali et al., 2011).
We used a statistical approach to specify the texture of surface
(Gonzalez & Woods, 2008). The selected parameters from 360
pieces of samples are: mean intensity (m), smoothness (S), third
momentum (l3), uniformity (U) and entropy (e):
A multilayer perceptron is proposed to be used to model the
LMC and MC. The NN architecture is a feed-forward one, with sig-
moid activation function of nodes in hidden layer/layers. In order
to have a greater flexibility of input and output vectors we used
linear activation function in the nodes that belong to these layers.
The NN is a fitting one type model, linear function allows to avoid
saturation for these nodes and the range vector elements of input/
output layer to be larger than [1, 1].
Our experimental trials showed that a single hidden layer is not
enough for a good prediction for this nonlinear relation between
texture descriptors and LMC (MC). So, we choose a NN with two
hidden layers (Fig. 7).
The basic inputs in NN predictor are vectors V = {m, S, l3, U, e,
FD} where FD is radius of the circle approximated as apple’s con-
tour. We evaluate the fitting performance using the R2 value.
R2 (coefficient of determination) is a statistical measure for
model performance between input data and predicted outcome.
A close value to 1 signifies a good model while a zero close value
suggests that the model is not suitable for current data set.
The data set was split in two disjoint groups: the group used for
learning stage (75 apples–300 samples) and a group used for test
stage (15 apples–60 samples).

Fig. 6. The G image of apple, the contour of apple (cyan) and the approximation by
a circle (blue). Fig. 7. The multilayer perceptron used for NN model predictor.
 L.C. Trincă et al. / Food Chemistry 144 (2014) 80–86 85

Table 2
Performance of NN using inputs V = {m, S, l3, U, e, FD}, R2 values.

Model Method
Neural network Slow drying Oven drying Lyophilisation
LMC, % (whole apple) MC, % (chunks apple) MC, % (crushed apple) MC, % (juice apple) MC, % (chunks apple)
Learning stage 0.8941 0.9915 0.9915 0.9915 0.9915
Testing stage 0.6381 0.8335 0.8335 0.8335 0.8335

Table 3
Performance of NN using inputs V = {m, S, 3, U, e, FD, Hk}, R2 values.

Haralick descriptors Method

Slow drying Oven drying Lyophilisation
LMC, % (whole apple) MC, % (chunks apple) MC, % (crushed apple) MC, % (juice apple) MC, % (chunks apple)
H1 0.7188 0.9499 0.9499 0.9499 0.9499
H2 0.9499 0.9499 0.9499 0.9499 0.9499
H3 0.9499 0.9499 0.9499 0.9499 0.9499
H4 0.9499 0.9499 0.9499 0.9499 0.9499
H5 0.9499 0.9499 0.9499 0.9499 0.9499
H6 0.9499 0.9499 0.9499 0.9499 0.9499
H7 0.9499 0.9499 0.9499 0.9499 0.9499
H8 0.9534 0.9534 0.9534 0.9534 0.9534
H10 0.9534 0.9534 0.9534 0.9534 0.9534
H11 0.9534 0.9534 0.9534 0.9534 0.9534
H12 0.9534 0.9534 0.9534 0.9534 0.9534
H13 0.9534 0.9534 0.9534 0.9534 0.9534

Table 4
Performance of NN using inputs V = {m, S, 3, U, e, FD, H8}, R2 values.

Model Method
Neural network Slow drying Oven drying Lyophilisation
LMC, % (whole apple) MC, % (chunks apple) MC, % (crushed apple) MC, % (juice apple) MC,% (chunks apple)
Learning stage 0.9989 0.9989 0.9989 0.9989 0.9989
Testing stage 0.9534 0.9534 0.9534 0.9534 0.9534

The predicted outcome of NN can be one of the variables: LMC
(whole apples) or MC (chunks/crushed/ juice/lyophilised apple
samples).
3.3.2. Genetic algorithm with variable length genotype
The results for learning stage and testing case using V = {m, S,
l3, U, e, FD} are shown in Table 2.
We propose a novel algorithm to improve the results. We use
the 13th texture descriptors proposed by Haralick, Shanmugam,
and Dinstein (1973), denoted by Hk, k = 1 .... 13. We eliminated H9
(entropy) from the list because this term is already taken in V.
Main steps of GAVLG algorithm for prediction of LMC/MC are:
Step 1. Set the basic list Vinput = V and additional list Hk, k2Lh,
Lh = {1, ... ,8,10, ... 13}.
Step 2. Run learn and test stage for NN using Vinput list. Obtain R2
for LMC.
Step 3. For all, k2Lh run learn and test stage for NN using
Vinput = V[Hk.
Step 4. The descriptors Hk that produced a greater value than in
the algorithm. If no improvement is made, take into account
the list Vinput  Hp and stop the algorithm. If the computational
effort is acceptable, go to Step 3.
Step 6. Clear Hp from list Lh.
The algorithm GAVLG runs separately for each population of
chromosomes. We have three populations of chromosomes and
we start with 50 individuals, with lengths between 4 and 10
weights. These values have been chosen by empirical assertions,
based on programmer’s experience. After 40 generations, an opti-
mal architecture has been found: 15 nodes in the first hidden layer
and 10 nodes in the second hidden layer.
We can see from Table 3 that if we append the H8 descriptor,
the precision of fitting increase from 0.6381 (that is an unaccept-
able value for a good model) to 0.9534, a good value for the model.
H8 refers to the sum of entropy:
X
2L
H8 ¼  pxþy ðzi Þ log pxþy ðzi Þ ð8Þ
i¼2

Step 2 are candidates for selected descriptor. If two or more X

L X
L

candidates have the same values, select the candidate with a

smaller computational effort. Even this factor is not relevant,
select one of them randomly. The new candidate Hp is append
to list Vinput, that is Vinput = Vinput[Hp.
Step 5. Stop if the desired precision is acquired. If the computa-
tional effort is too high, make a trade-off between precision and
performance and take into account the list Vinput  Hp and stop
pxþy ðzi Þ ¼ pðzij Þ ð9Þ
i¼1 j¼1iþj¼k
The experiments showed that more descriptors from Lh are ap-
pended the greatest is R2 value (Table 4). In the same pattern if we
add together H10, H11, H12, and H13 (Difference variance, Difference
entropy, and Information Measure of Correlation – both values) we
can reach a higher precision of the model, R2  0.9817.
86 L.C. Trincă et al. / Food Chemistry 144 (2014) 80–86
In accordance with step 5 of GAVLG algorithm the computa-
tional effort in this case is substantially increased, so, the optimal
Vinput can be considered V = {m, S, l3, U, e, FD, H8}.
4. Conclusions
Evaluation of water status in Jonathan apples was performed for
twenty days. Loss moisture content (LMC) was carried out through
slow drying of wholes apples and the moisture content (MC) was
carried out through oven drying and lyophilisation for apple sam-
ples (chunks, crushed and juice).
We approached a non-destructive method to evaluate LMC and
MC of apples using image processing and multilayer neural net-
works (NN) predictor.
We proposed a new simple algorithm that selects the texture
descriptors based on initial set heuristically chosen. Both structure
and weights of NN are optimised by a genetic algorithm with var-
iable length genotype that led to a high precision of the predictive
model (R2 = 0.9534).
In our opinion, the developing of this non-destructive method
for the assessment of LMC and MC (and of other chemical param-
eters) seems to be very promising in online inspection of food
quality.
References
Bradley, R. L. Jr., (2010). Moisture and total solids analysis. In S. S. Nielsen (Ed.), Food
analysis (pp. 85–104). New York: Springer Science, Business Media, LLC.
Cybenko, G. (1989). Approximations by superpositions of a sigmoidal function.
Mathematics of Control, Signals, and Systems, 2(4), 303–314.
Derossi, A., Severini, C., & Cassi, D. (2011). Mass transfer mechanisms during
dehydration of vegetable food: Traditional and innovative approaches. In M. El-
Amin (Ed.), Advanced topics in mass transfer (pp. 305–354). Rijeko: InTech –
Open Access Publisher.
Devore L.J. Berk N.K. Modern mathematical statistics with applications 2007,
Belmont: Duxbury.
El-Sayd, N. I., & Makawy, M. M. (2010). Comparison of methods for determination of
moisture in food. Research Journal of Agriculture and Biological Sciences, 6(6),
906–911.
Erturk, U., Akbuclak, B., & Ozer, M. H. (2003). Quality changes of some apple
cultivars stored in normal atmosphere for long period. Acta Horticulturae, 599,
665–672.
Fernández, L., Castillero, C., & Aguilera, J. M. (2005). An application of image analysis
to dehydration of apple discs. Journal of Food Engineering, 67(1–2), 185–193.
Fontana, A. J. Jr., (2007). Measurement of water activity, moisture sorption
isotherms and moisture content of foods. In G. V. Barbosa-Cánovas, A. J.
Fontana, Jr., S. J. Schmidt, & T. P. Labuza (Eds.), Water activity in foods:
Fundamentals and applications (pp. 155–173). Ames, Iowa: IFT Press, Blackwell
Publishing.
Francis, F. J. (2005). Colorimetric Property of Foods. In M. A. Rao, S. S. H. Rizvi, & A. K.
Datta (Eds.), Engineering Property of Foods (pp. 703–732). Boca-Raton, Florida:
CRC Press, Taylor & Francis Group.
Goñi, O., Muñoz, M., Ruiz-Cabello, J., Escribano, M. I., & Merodio, C. (2007). Changes
in water status of cherimoya fruit during ripening. Postharvest Biology and
Technology, 45, 147–150.
Gonzalez, R. C., & Woods, R. E. (2008). Digital image processing (3th ed.). Upper
Saddle River, New Jersey: Prentice Hall.
Haralick, R. M., Shanmugam, K., & Dinstein, I. (1973). Textural features for image
classification. IEEE Transaction on Systems, Man and Cybernetics, SMC-3(6),
610–621.
Haykin, S. (1999). Neural networks: A comprehensive foundation (2th ed.). Upper
Saddle River, New Jersey: Prentice Hall.
Hills, B. P., & Remigereau, B. (1997). NMR studies of changes in subcellular water
compartmentation in parenchyma apple tissue during drying and freezing.
International Journal of Food Science and Technology, 32, 51–61.
Isengard, H.-D., Merkh, G., Schreib, K., Labitzke, I., & Dubois, C. (2010). The influence
of the reference method on the results of the secondary method via calibration.
Food Chemistry, 122(2), 429–435.
Jan, I., & Rab, A. (2012). Influence of storage duration on physico-chemical changes
of fruits of apples cultivars. The Journal of Animal and Plant Sciences, 22(3),
708–714.
The MathWorks, Inc. MATLAB program. URL <http://www.mathworks.com/>.
Accessed 5.10.12.
Menges, H. O., & Ertekin, C. (2006). Mathematical modeling of thin layer drying of
Golden apples. Journal of Food Engineering, 77(1), 119–125.
Munir-ul, M., Chowdhury, M., & Yun Li (1996). Messy genetic algorithm based new
learning method for structurally optimised neurofuzzy controllers. Proceedings
of The IEEE International Conference on Industrial Technology (ICIT ‘96). 274–278.
Nguyen, T. A., Dresselaers, T., & Verboven, P. (2006). Finite element modelling and
MRI validation of 3D transient water profiles in pears during postharvest
storage. Journal of the Science of Food and Agriculture, 86, 745–756.
Okos, M. R., Osvaldo, C., Campanella, O., Narsimhan, G., Singh, R. K., & Weitnauer, A.
(2007). Food dehydration. In D. R. Heldman & D. B. Lund (Eds.), Handbook of food
engineering (pp. 601–743). Boca Raton: CRC Press, Taylor and Francis Group.
Petrou, M., & Sevilla, P. G. (2006). Image processing: Dealing with texture. West
Sussex: John Wiley & Sons, Ltd..
Rahman, M. S. (2006). State diagram of foods: Its potential use in food processing
and product stability. Trends in Food Science and Technology, 17, 129–141.
Schmidt, S. J. (2007). Water mobility in foods. In G. V. Barbosa-Cánovas, A. J.
Fontana, Jr., S. J. Schmidt, & T. P. Labuza (Eds.), Water activity in foods:
Fundamentals and applications (pp. 47–108). Ames, Iowa: IFT Press & Blackwell
Publishing.
Seiiedlou, S., Ghasemzadeh, H. R., Hamdami, N., Talati, F., & Moghaddam, M. (2010).
Convective drying of apple: Mathematical modeling and determination of some
quality parameters. International Journal of Agriculture and Biology, 12(2),
171–178.
Veraverbeke, E., Lammertyn, J., Saevels, S., & Nicolaï, B. (2001). Changes in chemical
wax composition of three different apple (Malus domestica Borkh.) cultivars
during storage. Postharvest Biology and Technology, 23(3), 197–208.
Veraverbeke, E., Verboven, P., Van Oostveldt, P., & Nicolaï, B. (2003). Prediction of
moisture loss across the cuticle of apple (Malus sylvestris subsp mitis (Wallr.)
during storage: Part 2. Model simulations and practical applications. Postharvest
Biology and Technology, 30(1), 89–97.
Vesali, F., Gharibkhani, M., & Komarizadeh, M. H. (2011). An approach to estimate
moisture content of apple with image processing method. Australian Journal of
Crop Science, 5(2), 111–115.
Yam, K. L., & Papadakis, S. E. (2004). A simple digital imaging method for measuring
and analyzing color of food surfaces. Journal of Food Engineering, 61(1), 137–142.