Talanta 148 (2016) 5461

Contents lists available at ScienceDirect

Talanta
journal homepage: www.elsevier.com/locate/talanta

Prediction of banana quality indices from color features using support

vector regression
Alireza Sanaeifar a, Adel Bakhshipour a, Miguel de la Guardia b,n
a
Biosystems Engineering Department, Shiraz University, Shiraz, Iran
b
Department of Analytical Chemistry, University of Valencia, 46100 Burjassot, Spain

art ic l e i nf o a b s t r a c t

Article history: Banana undergoes signicant quality indices and color transformations during shelf-life process, which
Received 9 September 2015 in turn affect important chemical and physical characteristics for the organoleptic quality of banana. A
Received in revised form computer vision system was implemented in order to evaluate color of banana in RGB, L*a*b* and HSV
16 October 2015
color spaces, and changes in color features of banana during shelf-life were employed for the quantitative
Accepted 24 October 2015
prediction of quality indices. The radial basis function (RBF) was applied as the kernel function of support
Available online 26 October 2015
vector regression (SVR) and the color features, in different color spaces, were selected as the inputs of the
Keywords: model, being determined total soluble solids, pH, titratable acidity and rmness as the output. Experi-
Banana mental results provided an improvement in predictive accuracy as compared with those obtained by
Color features
using articial neural network (ANN).
Firmness
& 2015 Elsevier B.V. All rights reserved.
pH
Total soluble solids
Support vector regression

1. Introduction representative of the whole because of difculties of destructively

Banana is one of the most popular fruits in the world and a
main one in international trade [1]. In 2011, bananas (Musa spp.)
were grown in 10.6 million ha with an average fruit yield of
13.6 t ha
1 [2], being Cavendish the most widely banana cultivar.
Suitable temperature, humidity, time and air ow are all needed
for ripening of bananas and by using controlled ethylene gas, the
ripening treatment of banana could be articially made [3]. During
banana ripening period, the peel color changes, the avor evolves
and the pulp softens. Banana peel color changes from green to
yellow, with brown spots appearing on the yellow color at the end
of shelf-life; being the synthesis of few pigments the reason of this
change in peel color [4]. So, on-line quality control of banana
during ripening treatment is quite important to keep a rm pulp
texture, good color and avor and also to prevent from bruise [5].
This makes monitoring of quality parameters in the orchard,
package and delivery points necessary in producing acceptable
bananas for the customer [3].
Most of the fruit quality measurement methods are destructive
such as pulp to peel ratio determination and fruit rmness, which
are mainly based on rheological properties [6]. Also, these meth-
ods do not sufciently monitor the quality of banana fruits during
ripening period [5], being employed just a small set of samples as
n
Corresponding author. Fax: 34 96 3544838.
E-mail address: Miguel.Delaguardia@uv.es (M. de la Guardia).
analyzing every unit of fruit [3].
The development of instrumental non-destructive techniques
could be highly appropriate to increase the number of fruit pieces
which can be analyzed, to repeat several times the analysis on the
same sample at a given time or during its physiological evolution,
and to access to real-time information [7].
Several soft and non-destructive techniques that can be applied
for quality measurement include proton transfer reaction mass
spectrometry (PTR-MS) [8], ultrasound [9], magnetic resonance
imaging (MRI) [10], time-resolved reectance spectroscopy [11],
X-ray and computed tomography (CT) [12], laser-induced uor-
escence spectroscopy (LIFS) [13], Fourier transform infrared (FTIR)
[14], sonic technique [15], near infrared spectroscopy (NIRS) [16],
sound velocity [17], optical chlorophyll sensing system [18], ca-
pacitance technique [5] and electronic nose systems [19]. Never-
theless, most of the aforementioned analytical techniques are far
from reaching a practical application and, sometimes, are time-
consuming, expensive, difcult to be implemented and require to
be made by trained personnel [7].
Near-infrared (NIR) spectroscopy has been used to correlate
rmness and soluble solid content (SSC) in Cavendish bananas at
different stages of ripeness [3,20]. On the other hand, Baiano et al.
[7] proposed the use of hyperspectral imaging technique for pre-
diction of some physico-chemical parameters of table grapes.
Concerning banana shelf life, studies made by Wang et al. [21],
Bora et al. [22] and Mendoza et al. [23] evaluated the ripening
stage of bananas from their color but without making any

http://dx.doi.org/10.1016/j.talanta.2015.10.073
0039-9140/& 2015 Elsevier B.V. All rights reserved.
 A. Sanaeifar et al. / Talanta 148 (2016) 5461
quantitative study about their quality indices. However there is no
precedent, in our own knowledge, about the use of external color
of bananas to evaluate their quality.
At rst instance, fruits quality is evaluated by looking at their
color, gloss, size and, secondly, by texture, total soluble solids (TSS)
content and acidity. These parameters may provide important
information to the consumer in the choice of food supply. A special
emphasis should be placed on quality attribution in trading. The
ripening indices traditionally used to evaluate changes in skin
color, softening, titratable acidity, soluble solids concentration and
volatile compounds [7,24]. On the other hand, the application of
sensors; such as optical, chemical, and tactile ones, provides a high
correlation with the human senses. Different techniques have been
reported to determine various quality parameters of fruits. How-
ever, there is currently no single or combination of techniques and
computational methods available to quantify the overall quality of
foods [25].
Color can be correlated with other quality attributes such as
sensory, nutritional and visual or non-visual defects and helps to
control them directly [25,26]. Color is considered a basic physical
property of agro-food products. In fact, color plays an important
role in the evaluation of external quality in food industries and
food engineering research [7,27]. The role of color is important to
evaluate food quality, but the use of color parameters to predict
specic quality indices of foods has been less explored and re-
presents an opportunity among researchers.
Industrial food product quality was monitored and controlled
by an on-line imaging system [25, 28]. However, there are only few
studies available on modeling kinetics of greenyellow transfor-
mations in fruits and vegetables. Engineers need quantitative
models to develop and improve processes [29]. Therefore, current
investigation was conducted to study changes in color features of
banana during shelf-life and to evaluate the use of image analysis
technique as a rapid and nondestructive alternative for the accu-
rate prediction of quality indices.
On the other hand, the scientic literature shows the valuable
application of articial intelligence in food research eld [3034]
and radial basis function-based support vector regression has been
employed to process color data of bananas at different shelf-life
stages.
2. Materials and methods
2.1. Experimental material
Banana fruits (Cavendish variety) imported from the Phi-
lippines were used in this research. The banana fruits were stored
at 14 C during transportation. Then, the fruits were stored in an
airtight warehouse. Bananas' ripening was completed in 4 days. In
the rst day, fruits were stored at 20 C, and in the second day,
ethylene gas was injected. In the third day, ethylene was removed
and temperature decreased to 18 C and nally temperature was
decreased so that it achieved 11 C in the fourth day. The control of
temperature, humidity and ethylene gas concentration in the ri-
pening room are very important. Bananas were kept in the
warehouse at the humidity level of 8588% for 4 days, as this time
period needed for completing the ripening treatment of fruits. At
this site, ripeness is currently assessed visually by comparing the
peel color of banana with standardized color charts describing
various stages of ripening.
Fruit development can generally be divided into three major
steps: growth, ripening, and senescence. The period of growth
generally involves cell division and enlargement, which accounts
for the increasing size of the fruit. During ripening fruit become
soft textured, and accumulate soluble sugars, pigments and
55
volatile aroma. Senescence is the period when chemical synthe-
sizing pathways give way to destruction processes, leading to
aging and death of tissue. Additionally, it has been identied seven
stages in the ripening period of bananas which change the color
from green to yellow and brown [35].
Before entering the warehouse, bananas are in the rst ripen-
ing stage and after leaving the warehouse, they are in the fth
ripening stage. For performing the experiments, 15 bananas of the
same size, weight and ripening stage, were daily transferred from
the warehouse to the laboratory for image analysis. The bananas
were picked from the same position of each branch. Then tests
were done on the samples till the fth ripening stage. Moreover,
all the measurements were performed in the day of transferring
from the warehouse. For experiments conducted during the se-
nescence period, banana samples were selected after the end of
fth ripening stage and were studied during additional 4 days
until mealiness completion. This period represents the time in-
terval between delivering bananas from the warehouse and reach
to distribution centers and as long as are usable by the consumer.
In general, after completing the ripening period in ve stages,
senescence period starts and this period takes four days. All this
process is represented shelf-life period of banana samples.
2.2. Machine vision system
A machine vision system was designed that includes two main
parts. An image acquisition system to take images from banana
samples at any stage of shelf-life. The second part was an image
analysis algorithm written in MATLAB v8.1 (The Mathworks, Na-
tick, MA, USA) software for automatically emphasizing the fruit
surface in the images as the region of interest and extract average
color features from the skin of the banana fruits.
2.2.1. Image acquisition
The setup used to take the images from bananas consisted of
four main compartments for: (1) lighting and imaging system,
(2) computer case, (3) monitor and user interfaces and
(4) accessories.
The D65 standard light source, which is commonly used in food
image processing studies, was employed in this research to pro-
vide reproducible, uniform-steady lighting conditions [36]. A box
made of wood with the interior walls painted in mat black was
used as the illumination chamber. Two uorescent lamps (Natural
Daylight, 230 V/12 W) with a color reproduction index up to 95%
and color temperature of 6500 K were placed on left and right
sides of the box. PVC sheets were located in front of the lamps to
avoid direct lighting and shading. The banana samples were
mounted on top of a black tray with very low reectivity to be
adequately discriminable from the background. High quality
images from samples were taken by a Power shot color camera
(Canon EOS kiss  4 digital color camera, Japan, 18MP) with re-
mote capturing capability mounted at a vertical distance of 25 cm
above the samples.
Camera and uorescent lamps were turned on 1 h before image
recording to achieve a stable lightening [37]. All of the images
were taken under totally stabilized camera characteristics ac-
cording to Table 1 variables. The images were transferred auto-
matically to the computer and loaded into image processing
toolbox of MATLAB software for further operations.
2.2.2. Color feature extraction
The owchart of image analysis algorithm is illustrated in Fig. 1.
Image enhancement processes including common median ltering
was the rst operation performed on the loaded RGB images to
suppress noises.
Loaded images were segmented, which is one of the most
56 A. Sanaeifar et al. / Talanta 148 (2016) 5461

Table 1 Table 2
Camera characteristics. Performance criteria.

Variable Value Criteria Calculation

Image size 5184  3456 pixels Root mean squared error (RMSE) 1
RMSE = ni = 1 (Pi Ai )2
Zoom No zoom n
Flash mode No ash Coefcient of determination n (Pi Ai )2
Sensitivity ISO-100 R2 = 1 ( i=1
)
n Ai2
i=1
White balance Fluorescence
Operation mode Manual
Aperture Av. f/4.0
Exposure time Av. 1/30 s and CIE L*a*b* are described comprehensively in literatures
Image type JPEG [38,39].
Macro On The values describing the red, green, blue, hue, saturation, in-
Focal length 24 mm
tensity, lightness, and also a* and b* chromatic components were
Resolution 72 dpi
determined on the surface of bananas. The extracted values were
then averaged and saved for further analysis.
important steps of image analysis, to recognize the region of in-
terest automatically in the images. Based on visual testing, green 2.3. Measurement of quality indices
color component was considered as the best one to discriminate
between bananas and image background. So the green component Total soluble solids (TSS), titratable acidity (TA) and pH are
image was extracted from each RGB image. Optimal threshold of chemical indices that are used to asses quality and ripeness of
green intensities was calculated based on Otsu algorithm and banana fruit. To measure TSS, 10 g samples were picked from the
applied to the green channel image as a criterion to differentiate esh of bananas (along length of fruit) and diluted in 50 g of dis-
the banana samples from the background. tilled water in a blender for 1 min. TSS was measured using a re-
In order to extract color information from the banana surface fractometer [5]. The TA was established by titrating banana juice
and to eliminate data related to image background, the binary with 0.1 M NaOH till to obtain a pH of 8.18.3 and calculating the
image and the original color image were superimposed together result as malic acid. The pH of juice was measured by a pH meter
using logical AND operator. The result was a RGB image with color Mettler Toledo Delta 320 (Geneva, Switzerland). Firmness, which
values and zeros corresponding to the banana surface and back- is one of the parameters that best correlates with ripeness, is de-
ground respectively. ned as the maximum force in Newton required until tissue failure
The resulted image was then transformed from RGB into HSV is occurred. To measure the pulp rmness of samples, an Instron
and L*a*b* color spaces. Methods of color space conversion to HSV universal testing machine Model SANTAM ST-5, with an 8 mm

Fig. 1. Flowchart of the algorithm employed for color measurement of banana samples.
 A. Sanaeifar et al. / Talanta 148 (2016) 5461 57

evaluate the performance of the models built. This technique is

190
especially used in cases where the data set is relatively small and
180 limited. This technique picks a single observation case among all
R cases as validation data, and the other data are used to derive
170 predictive function. This operation is repeated for all observations.
G
So, all of the samples are once used as the validation data [40]. In
160 B this way the maximum number of samples are used to calibrate
RGB

150 and validate the model.

140 2.4.1. Support vector regression (SVR)

Support vector machine is one of machine learning method
130 that has recently become popular and widely used and in-
vestigated because of its ability in prediction for both, classica-
120
0 1 2 3 4 5 6 7 8 9 tion and regression [41]. The basic idea of the support vector re-
Shelf-life stages gression (SVR) is to non-linearly map the input variables into a
high dimensional feature space where they are linearly correlated
with the output variable [42].
70
The -support vector regression (SVR) was introduced as an
60 alternative to the -insensitive loss function. The purpose of the
50 SVR is to nd the function with the most deviations from the
L* actual destination vector for all obtained training information, and
40
it must be as at as possible [43,44]. SVR requires parameters
30 a*
dened by the user to set up kernel-specic parameters. Ad-
L*a*b*

20 b* ditionally, the optimal values of the legalization argument C and

size errors in sensitive area require to be ascertained. The set-
10
tings' selection controls the prediction complexity. One of the
0 major advantages of SVR lie in the algorithm which contains the
-10 0 1 2 3 4 5 6 7 8 9 resolution of the quadratic programming function leading to an
unique, optimum and comprehensive solution. According to the
-20
choice of kernel function and its parameters, the dimensionality of
-30 the feature space is determined while the complexity of the model
Shelf-life stages
is determined by an extra penalty parameter. According to a re-
view of literature, a part of support vector machine prediction has
1
been scarcely utilized in image processing studies [44,45].
Some of the most widely used kernel functions include: Linear,
0.8 Polynomial, Radial basis function (RBF), Sigmoid, etc. Different
H kernel functions were evaluated for checking the robustness of the
regression model. The RBF kernel, which is widely used in many
0.6 S
problems, eventually was applied for this study. The radial basis
HSV

V function is also a simple function and can model systems of

0.4
0.2
0
0 1 2 3 4 5 6 7
Shelf-life stages
Fig. 2. The changes of color features during shelf-life in spaces: (a) RGB, (b) L*a*b*
and (c) HSV.
cylindrical probe, was used, being the loading rate of the cross-
head 50 mm/min. The rmness was measured at 2 cm away from
the middle of the fruit centerline curvature [5]. Quality indices
were measured with three replications for each shelf-life stage.
2.4. Data analysis
For analyzing the data collected, articial neural networks and
support vector regression were developed and compared to nd
the best model for prediction the quality indices of bananas. Ex-
planations of the techniques used in this study are given below.
MATLAB v8.1 from Mathworks (Natick, MA, USA) and Unscrambler
10.3 from CAMO AS (Trondheim, Norway) were the software used
for data analysis through this study.
Leave-one-out cross validation technique was applied to
varying complexity whereas other kernels satisfy it only for certain
conditions [19]. The different SVR parameters needed for each
model were optimized with leave-one-out cross-validation for the
training set in terms of root mean squared error (RMSE). During
this process, keeping the value of equal to 0.1, for C and a broad
8 9 range of parameter settings was investigated with large steps and
after identifying a promising region, this region was searched in
detail. A new prediction model according to the support vector
regression (SVR) is being recommended in this paper in order to
predict quality indices of banana. Aiming to create a powerful SVR
model, the SVR parameters should be set up carefully. SVR looks
for to minimize the generalization error to gain generalized per-
formance rather than minimizing the training error.
2.4.2. Articial neural network (ANN)
To be able to evaluate the suggested model, an articial neural
network (ANN) model was created and the outcomes extracted
from ANN were compared with the outcomes acquired from SVR.
After several trail and errors with various types of ANN, a back-
propagation multilayer perceptron (BP-MLP) algorithm was cho-
sen as the best option for banana quality indices prediction. The
color features in different color spaces including RGB, HSV and
L*a*b* were selected as inputs while the quality indices including
rmness, TSS, TA and pH were chosen as output for ANNs. Color
features of banana samples were repeatedly fed into the network
and training was accomplished by automatically adjusting the
58 A. Sanaeifar et al. / Talanta 148 (2016) 5461

150 0.4

120
0.3
Firmness (N)

TA (%)
90
0.2
60
0.1
30

0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Shelf-life stages Shelf-life stages

5.8 25

5.6 20

TSS (% Brix)
5.4 15
PH

5.2 10

5 5

4.8 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Shelf-life stages Shelf-life stages
Fig. 3. The changes of the banana quality indices during shelf-life stages.

Table 3
Correlation between color feature values and quality indices.

R G B H S V L* a* b*

TSS .828**
.902**
.711**
.918**
.939**
.910**
.830** .943**
.389
pH
.118
.403* .050 .037
.255
.381
.484* .225
.803**
TA
.098 .628** .227 .217 .479* .608** .692**
.454* .786**
Fir
.835** .836** .778** .903** .886** .843** .758**
.894** .204

**
Correlation is signicant at the .01 level (2-tailed).
*
Correlation is signicant at the .05 level (2-tailed).

connection weights in order to reduce the error level. Also, the
MLP method is considered to be rather fast and efcient when a
small data set is being used [46]. A leave-one-out cross-validation
method was used to evaluate the performance of the nal net-
work. All data were assigned in two categories of training (80%)
and validation (20%) ones. The BP-MLP with different numbers of
hidden layers (one, two, and three) was trained and tested. The
nal results showed that two hidden layer BP-MLP ANN was en-
ough for this study. The Linear Axon was regarded as transfer
function in output layer because of the regression nature of the
study. However, the effect of error minimization algorithms,
training algorithm, epochs, and times on ANNs overall perfor-
mance were further surveyed [37].
Several actions conrmed the validity of the offered ANN and
SVR models. The root mean squared error (RMSE) served to assess
the differences between the expected and actual values. The
parameter was determined as indicated in Table 2. Where n is the
total number of test data, Pi concerns the experimental values, and
Ai the predicted values [44].
3. Results and discussion
3.1. Color features
The changes of banana color features during the shelf-life
period are shown in Fig. 2. L* was reduced during the shelf-life
period. However, the a* color feature has an upward trend over the
period which means decline in greenness of fruit skin. Generally
small changes were observed in the b* values during that period. A
gradual increase was seen in b* value till 5th day of shelf-life (end
of ripening period), which stabilized after that, addressing an in-
crease in yellow color of banana peel. It may be concluded that
progressively the color of banana peel altered from green to yellow
with ripening period and during senescence period, brown spots
on the skin of yellow bananas grow. These observations are in
consistent with the previous results in this eld [47]. The trend of
change in color spaces RGB and HSV in Fig. 2(b) and (c) is shown.
3.2. Quality indices
The changes of the banana quality indices during the ripening
 A. Sanaeifar et al. / Talanta 148 (2016) 5461 59

125 25.00

Firmness 100 20.00

75 15.00

TSS
50 10.00

25 5.00

0 0.00
0.2 0.25 0.3 0.35 0.4 12 2 -8 -18 -28
h a*

6 0.4

5.6 0.32

5.2 0.24
pH

TA
4.8 0.16

4.4 0.08

4 0
9 13 17 21 9 13 17 21
b* b*
Fig. 4. Change of quality indices vs. color features.

Table 4 period leads to decreasing the acid concentrations [51]. This result
Results of SVR for predicting the quality indices of banana. conforms with previous researches done in this eld [52,53].
Quality R2 (%) RMSE C
indices 3.3. Correlation between color features and quality indices
Calibration Validation Calibration Validation

Firmness 99.52 99.50 4.8271 4.910 1.668 100 To obtain the best set of color parameters for predicting the
TSS 99.61 99.57 .7092 .8536 1.668 100 quality of bananas, correlations between color components and
pH 99.80 99.74 .0383 .1197 1.668 100 quality indices were analyzed and the corresponding values are
TA 99.20 99.19 .0087 .0218 1.668 100
indicated in Table 3.
Signicant correlations were observed between some of color
and senescence periods are indicated in Fig. 3. components and the quality indices at 99 and 95% levels. Re-
As shelf-life period progressed, the rmness values decreased garding these results it can be stated that there are good agree-
from 102.52 to 17.40 N. The texture of banana fruits being softened ments between some of color characteristics and quality para-
by ripening. Changes of the cell wall structure by degrading en- meters of banana. So, it can be concluded that color features are
zymes and the starch degradation in fruit are the two main rea- good descriptors of banana quality indices.
sons of fruit softening [5,48].
The TSS was 7.8% Brix at stage one (green-mature) and in-
creased to 16.17% Brix at stage ve (ripe). Its value reached to 20%
4. Correlation between color feature values and quality
Brix at stage nine (full-mealiness). During the ripening treatment
indices.
the starch content of fruit is hydrolyzed into soluble sugars such as
glucose, sucrose and fructose which in turn increases the TSS
The changes of quality indices relating to the color parameters
content of fruit [49].
which provided the best correlation with each of them are shown
By the ripening process, the pH of banana decreases from 5.70
at the rst stage to 4.96 at fth stage. As the senescence period in Fig. 4.
started, there was an increasing trend in the pH value which - a*, b* and h were the best estimators of TSS, acidity parameters
nally reached to 5.60. This change was mainly due to changes of and rmness respectively. As the a* color component ranges from
fruit malic acid content that is a dominant acid in Cavendish ba- green to red, it can be concluded from Fig. 4(b) that the greener
nana varieties [50]. the banana the lower the TSS value. On the other word, unripe
As it is also shown in Fig. 3(b), The TA mean value of banana bananas have lower TSS contents than ripe ones. The higher b*
samples increased during the period of ripening and decreased values introduce the increase of the yellow color of objects. From
during the senescence period. Fig. 4(c) it can be stated that the riper bananas have an increased
Some of fruits, like banana, have the maximum amount of or- acidity (low pH) and high TA indices. The rmness reduces with
ganic acids at their complete ripeness. However, increasing of the ripening process while the banana becomes less chromatic
metabolic activities and formation of sugars during the senescence (lower hue values).
60 A. Sanaeifar et al. / Talanta 148 (2016) 5461

Firmness TSS
120 25
100

Predicted Values
Predicted Values

20
80
60 15
40
10
20
0 5
0 20 40 60 80 100 120 5 10 15 20 25
Actual Values Actual Values

pH 0.4 TA
5.8

Predicted Values
0.35
Predicted Values

5.6
5.4 0.3

5.2 0.25

5 0.2

4.8 0.15
4.8 5 5.2 5.4 5.6 5.8 0.15 0.2 0.25 0.3 0.35 0.4
Actual Values Actual Values
Fig. 5. Plots of the regression between observed and predicted banana quality indices obtained by using the original color data set and SVR_rbf model.

Table 5
Results of ANN for predicting the quality indices of banana.
Quality indices Network structure R2 (%)
Train Test
Firmness 9-4-7-1 95.71 95.01
TSS 9-10-12-1 92.39 89.99
pH 9-9-14-1 60.88 59.79
TA 9-8-14-1 57.25 54.91
4.1. SVR and ANN
Robust quality measurement tools, based on the whole color
extracted features of bananas, were built by using SVR and Neural
Networks and the performance of the different models assessed
and compared to each other for predicting the quality parameters
of banana fruits. RBF was applied as the Kernel function of SVR
method in this study. Adjusting the appropriate SVR parameter is
an important issue in order to provide well prediction perfor-
mance. C, and Kernel parameter are the optimal parameters of
SVR. C being the trade-off between the model complexity and
error tolerance. With regard to model complexity, very large C
values minimize error. is width of -intensive zone and affect the
number of support vector of the training data. is the width of
Kernel function which represents the distribution of independent
values in the training data set. If is large, there are less support
vectors which cause less complex prediction. The root mean
squared error (RMSE) and determination coefcient (R2) were
employed to evaluate the differences between the expected and
actual values of banana quality indices obtained for SVR and ANN.
Table 4 shows the optimal values of the parameters C and , as well
as R2 and RMSE values obtained for each one of the quality indices.
The parameter of SVR being 0.1. Fig. 5 shows the regression
between observed and SVR predicted data in the testing phase.
Generally, as can be seen in Table 4 and Fig. 5, SVR_rbf predicts
well the quality indices with R2 values higher than 99% and RMSE
values of the order of 5 for rmness and less than 0.85 for TSS,
0.1 for pH and 0.02 for TA. On the other hand, Table 5 shows the
prediction capability obtained by using neural network results and
RMSE
in this case R2 values varied from 57 till 95% and RMSE were 6, 1.3,
Train Test 0.2 and 0.04 for rmness, TSS, pH and TA respectively. So, data
reported in the aforementioned tables show that the SVR method
6.1222 6.1721 is more accurate and robust than that based on neural network for
1.3062 1.3614
.1975 .2129
banana quality prediction based on color features.
.0332 .0481 In the case of neural network, the best results were obtained
when the network was used to estimate the fruit rmness based
on image characteristics where the coefcient of determination
between the experimental and predicted values was 95.01%. The
corresponding value was 99.50% for SVR method. Whereas the
best result of SVR was belong to predicting the banana pH
(R2 99.74%).
Although the leave-one-out method was used for both assayed
models training it seems that the low number of input data leaded
to poorer performance of neural network than SVR, which is less
sensitive to the amount of model input data.
The capability of ANN and SVR models was also evaluated by
the comparison for classifying the banana fruit based on their
degree of shelf-life. The ANN was succeed to divide fruits into
three classes with high accuracy (R2 91.65% and RMSE 0.1962).
Among the 25 samples used to test the classier ANN, only two
were misclassied. The classication accuracy of the SVR classier
was 100% for training set and 92% for testing set. So, once again it
can be seen that the SVR model works better than the ANN one in
spite that both models had a similar classication accuracy.
5. Conclusions
Support vector regression (SVR) of color parameters provides a
useful model for prediction of the quality indices of bananas. The
use of a radial basis function (SVRrbf) permitted to estimate
quality indices, component concentrations, chemical and physical
properties from color features of banana during shelf-life. The SVR
 A. Sanaeifar et al. / Talanta 148 (2016) 5461
approache enhanced the outcomes provided by ANN and create
smaller estimation errors in comparison with ANN. Moreover, the
SVR technique can be applied in the elds related to agriculture
with lower computer time than ANN or any knowledge of the
network's architecture as in ANN techniques. From the results
reported in this study it can be concluded that SVR technique can
predict quality indices with a high estimation accuracy and short
computation time. The aforementioned ndings can be used in
developing a grading system for the banana industry based on the
rapid prediction of quality parameters of the fruit.
References
[1] P. Jaiswal, S.N. Jha, P.P. Kaur, R. Bhardwaj, A.K. Singh, V. Wadhawan, Prediction
of textural attributes using color values of banana (Musa sapientum) during
ripening, J. Food Sci. Technol. 51 (2014) 11791184.
[2] FAO, FAOSTATS, Food and Agriculture Organization of the United Nations,
Rome, Italy, 2013.
[3] C. Liew, C. Lau, Determination of quality parameters in Cavendish banana
during ripening by NIR spectroscopy, Int. Food Res. J. 19 (2012) 751758.
[4] J.F.S. Gomes, R.R. Vieira, F.R. Leta, Colorimetric indicator for classication of
bananas during ripening, Sci. Hortic. 150 (2013) 201205.
[5] M. Soltani, R. Alimardani, M. Omid, Evaluating banana ripening status from
measuring dielectric properties, J. Food Eng. 105 (2011) 625631.
[6] I. Ramma, S.B. Madhu, P. Peerthum, Post-harvest quality improvement of ba-
nana, Food Agric. Res. Counc. (2001) 187.
[7] A. Baiano, C. Terracone, G. Peri, R. Romaniello, Application of hyperspectral
imaging for prediction of physico-chemical and sensory characteristics of table
grapes, Comput. Electron. Agric. 87 (2012) 142151.
[8] D. Barbon, A. Weber, M. Vescovi, A. Tonini, A. Boschetti, S. Lannotta,
L. Fadanelli, G. Stoppa, A statistical approach for proton transfer reaction mass
spectrometry (PTR-MS) data aimed at a qualication of fruits based on VOC
emissions, Acta Hortic. (2005).
[9] J.A. Bamberger, M.S. Greenwood, Non-invasive characterization of uid food-
stuffs based on ultrasonic measurements, Food Res. Int. 37 (2004) 621625.
[10] J.E. Andaur, A.R. Guesalaga, E.E. Agosin, M.W. Guarini, P. Irarrzaval, Magnetic
resonance imaging for nondestructive analysis of wine grapes, J. Agric. Food
Chem. 52 (2004) 165170.
[11] P.E. Zerbini, M. Vanoli, M. Grassi, A. Rizzolo, M. Fibiani, G. Biscotti, A. Pifferi,
A. Torricelli, R. Cubeddu, Time-resolved reectance spectroscopy as a non-
destructive tool to assess the maturity at harvest and to model the softening of
nectarines, Acta Hortic. 682 (2005) 1459.
[12] N. McFarlane, C. Bull, R. Tillett, R. Speller, G. Royle, K. Johnson, The potential for
Compton scattered X-rays in food inspection: the effect of multiple scatter and
sample inhomogeneity, J. Agric. Eng. Res. 75 (2000) 265274.
[13] J. Wulf, M. Geyer, B. Nicola, M. Zude, Non-destructive assessment of pigments
in apple fruit and carrot by laser-induced uorescence spectroscopy (LIFS)
measured at different time-gate positions, V Int. Postharvest Symp. 682 (2004)
13871394.
[14] A. Bellincontro, I. Nicoletti, M. Valentini, A. Tomas, D. De Santis, D. Corradini,
F. Mencarelli, Integration of nondestructive techniques with destructive ana-
lyses to study postharvest water stress of winegrapes, Am. J. Enol. Vitic. 60
(2009) 5765.
[15] E. Finney, I. BenGera, D.R. Massie, An objective evaluation of changes in
rmness of ripening bananas using a sonic technique, J. Food Sci. 32 (1967)
642646.
[16] D. Prez-Marn, M.-T. Snchez, P. Paz, M.-A. Soriano, J.-E. Guerrero, A. Garrido-
Varo, Non-destructive determination of quality parameters in nectarines
during on-tree ripening and postharvest storage, Postharvest Biol. Technol. 52
(2009) 180188.
[17] P. Subedi, K.B. Walsh, Non-invasive techniques for measurement of fresh fruit
rmness, Postharvest Biol. Technol. 51 (2009) 297304.
[18] M. Li, D.C. Slaughter, J.F. Thompson, Optical chlorophyll sensing system for
banana ripening, Postharvest Biol. Technol. 12 (1997) 273283.
[19] A. Sanaeifar, S.S. Mohtasebi, M. Ghasemi-Varnamkhasti, H. Ahmadi, J. Lozano,
Development and application of a new low cost electronic nose for the ripe-
ness monitoring of banana using computational techniques (PCA, LDA, SIMCA,
and SVM), Czech J. Food Sci. 32 (2014) 538548.
[20] J. Tarkosova, J. Copikova, Determination of carbohydrate content in bananas
during ripening and storage by near infrared spectroscopy, J. Infrared Spec-
trosc. 8 (2000) 2126.
[21] Y. Wang, Y. Cui, S. Chen, P. Zhang, H. Huang, G.Q. Huang, Study on fruit quality
measurement and evaluation based on color identication, in: Proceedings of
the International Conference on Optical Instrumentation and Technology, In-
ternational Society for Optics and Photonics, 2009, pp. 75130F-75130F-75136.
[22] G.C. Bora, D. Lin, P. Bhattacharya, S.K. Bali, R. Pathak, Application of bio-image
analysis for classication of different ripening stages of banana, J. Agric. Sci. 7
(2015) p152.
[23] F. Mendoza, J. Aguilera, Application of image analysis for classication of ri-
pening bananas, J. Food Sci. 69 (2004) E471E477.
61
[24] L. Sonego, S. Lurie, Y. Zuthi, T. Kaplonov, R. Ben-Arie, I. Kosto, Factors affecting
taste scores of early season seedless table grape cv. Mystery and Prime, J.
Agric. Food Chem. 50 (2002) 544548.
[25] P.B. Pathare, U.L. Opara, F.A.-J. Al-Said, Colour measurement and analysis in
fresh and processed foods: a review, Food Bioprocess Technol. 6 (2013) 3660.
[26] F. Francis, Quality as inuenced by color, Food Qual. Preference 6 (1995)
149155.
[27] M. Abdullah, L. Guan, K. Lim, A. Karim, The applications of computer vision
system and tomographic radar imaging for assessing physical properties of
food, J. Food Eng. 61 (2004) 125135.
[28] A.C. Pereira, M.S. Reis, P.M. Saraiva, Quality control of food products using
image analysis and multivariate statistical tools, Ind. Eng. Chem. Res. 48
(2008) 988998.
[29] C. Chen, H. Ramaswamy, Color and texture change kinetics in ripening ba-
nanas, LWT Food Sci. Technol. 35 (2002) 415419.
[30] H. Zheng, L. Jiang, H. Lou, Y. Hu, X. Kong, H. Lu, Application of articial neural
network (ANN) and partial least-squares regression (PLSR) to predict the
changes of anthocyanins, ascorbic acid, Total phenols, avonoids, and anti-
oxidant activity during storage of red bayberry juice based on fractal analysis
and red, green, and blue (RGB) intensity values, J. Agric. Food Chem. 59 (2010)
592600.
[31] S. Kerdpiboon, W.L. Kerr, S. Devahastin, Neural network prediction of physical
property changes of dried carrot as a function of fractal dimension and
moisture content, Food Res. Int. 39 (2006) 11101118.
[32] H. Zheng, H. Lu, A least-squares support vector machine (LS-SVM) based on
fractal analysis and CIELab parameters for the detection of browning degree
on mango (Mangifera indica L.), Comput. Electron. Agric. 83 (2012) 4751.
[33] H. Lu, H. Zheng, Y. Hu, H. Lou, X. Kong, Bruise detection on red bayberry
(Myrica rubra Sieb. & Zucc.) using fractal analysis and support vector machine,
J. Food Eng. 104 (2011) 149153.
[34] H. Zheng, B. Jiang, H. Lu, An adaptive neural-fuzzy inference system (ANFIS)
for detection of bruises on Chinese bayberry (Myrica rubra) based on fractal
dimension and RGB intensity color, J. Food Eng. 104 (2011) 663667.
[35] S. Zomo, S. Ismail, M.S. Jahan, K. Kabir, M. Kabir, Chemical properties and shelf
life of banana (Musa sapientum L.) as inuenced by different postharvest
treatments, The Agriculturists 12 (2015) 0617.
[36] R. Quevedo, J. Aguilera, F. Pedreschi, Color of salmon llets by computer vision
and sensory panel, Food Bioprocess Technol. 3 (2010) 637643.
[37] M. Dowlati, S.S. Mohtasebi, M. Omid, S.H. Razavi, M. Jamzad, M. de la Guardia,
Freshness assessment of gilthead sea bream (Sparus aurata) by machine vision
based on gill and eye color changes, J. Food Eng. 119 (2013) 277287.
[38] K. Leon, D. Mery, F. Pedreschi, J. Leon, Color measurement in L a b units
from RGB digital images, Food Res. Int. 39 (2006) 10841091.
[39] S.J. Sangwine, R.E.N. Horne, The Colour Image Processing Handbook, Chapman
& Hall, London, 1998.
[40] R. Abbaszadeh, A. Rajabipour, H. Ahmadi, M. Mahjoob, M. Delshad, Prediction
of watermelon quality based on vibration spectrum, Postharvest Biol. Technol.
86 (2013) 291293.
[41] M. Ghasemi-Varnamkhasti, S.S. Mohtasebi, M. Siadat, H. Ahmadi, S.H. Razavi,
From simple classication methods to machine learning for the binary dis-
crimination of beers using electronic nose data, Eng. Agric. Environ. Food 8
(2015) 4451.
[42] V. Vapnik, Statistical Learning Theory, Wiley, New York, 1998.
[43] A.J. Smola, B. Schlkopf, A tutorial on support vector regression, Stat. Comput.
14 (2004) 199222.
[44] E. Omrani, B. Khoshnevisan, S. Shamshirband, H. Saboohi, N.B. Anuar, M.H.N.
M. Nasir, Potential of radial basis function-based support vector regression for
apple disease detection, Measurement 55 (2014) 512519.
[45] Z. Ramedani, M. Omid, A. Keyhani, S. Shamshirband, B. Khoshnevisan, Po-
tential of radial basis function based support vector regression for global solar
radiation prediction, Renew. Sustain. Energy Rev. 39 (2014) 10051011.
[46] A. Sanaeifar, S.S. Mohtasebi, M. Ghasemi-Varnamkhasti, M. Siadat, Application
of an electronic nose system coupled with articial neural network for clas-
sication of banana samples during shelf-life process, in: Proceedings of the
2014 International Conference on Control, Decision and Information Tech-
nologies (CoDIT), IEEE, 2014, pp. 753757.
[47] R. Mustaffa, A. Osman, S. Yusof, S. Mohamed, Physicochemical changes in
Cavendish banana (Musa cavendishii L var Montel) at different positions within
a bunch during development and maturation, J. Sci. Food Agric. 78 (1998)
201207.
[48] G.B. Seymour, J.E. Taylor, G.A. Tucker, Biochemistry of Fruit Ripening, Chap-
man & Hall, London, 1993.
[49] J. Marriott, M. Robinson, S.K. Karikari, Starch and sugar transformation during
the ripening of plantains and bananas, J. Sci. Food Agric. 32 (1981) 10211026.
[50] P. John, J. Marchal, Ripening and biochemistry of the fruit, in: S. Gowen (Ed.),
Bananas and Plantains, Chapman & Hall, London, 1995, pp. 434467.
[51] R.H. Wills, T. Lee, D. Graham, W. McGlasson, E. Hall, Postharvest: An In-
troduction to the Physiology and Handling of Fruit and Vegetables, CAB In-
ternational, Willingford, 1981.
[52] R. Mustaffa, A. Osman, S. Yusof, S. Mohamed, Physico-chemical changes in
Cavendish banana (Musa cavendishii L var Montel) at different positions
within a bunch during development and maturation, J. Sci. Food Agric. 78
(1998) 201207.
[53] A. Etienne, M. Gnard, D. Bancel, S. Benoit, C. Bugaud, A model approach re-
vealed the relationship between banana pulp acidity and composition during
growth and post harvest ripening, Sci. Hortic. 162 (2013) 125134.