International Journal of Phytoremediation

ISSN: 1522-6514 (Print) 1549-7879 (Online) Journal homepage: http://www.tandfonline.com/loi/bijp20

Statistical optimization of the phytoremediation of

arsenic by Ludwigia octovalvis- in a pilot reed bed
using response surface methodology (RSM) versus
an artificial neural network (ANN)

Harmin Sulistiyaning Titah, Mohd Izuan Effendi Bin Halmi, Siti Rozaimah
Sheikh Abdullah, Hassimi Abu Hasan, Mushrifah Idris & Nurina Anuar

To cite this article: Harmin Sulistiyaning Titah, Mohd Izuan Effendi Bin Halmi, Siti Rozaimah
Sheikh Abdullah, Hassimi Abu Hasan, Mushrifah Idris & Nurina Anuar (2018) Statistical
optimization of the phytoremediation of arsenic by Ludwigia octovalvis- in a pilot reed bed using
response surface methodology (RSM) versus an artificial neural network (ANN), International
Journal of Phytoremediation, 20:7, 721-729

To link to this article: https://doi.org/10.1080/15226514.2017.1413337

View supplementary material

Published online: 03 May 2018.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=bijp20
INTERNATIONAL JOURNAL OF PHYTOREMEDIATION
2018, VOL. 20, NO. 7, 721–729
https://doi.org/10.1080/15226514.2017.1413337

Statistical optimization of the phytoremediation of arsenic by Ludwigia octovalvis-

in a pilot reed bed using response surface methodology (RSM) versus an artificial
neural network (ANN)
Harmin Sulistiyaning Titaha,b,d, Mohd Izuan Effendi Bin Halmie, Siti Rozaimah Sheikh Abdullaha, Hassimi Abu Hasana,
Mushrifah Idrisc, and Nurina Anuar a
a
Department of Chemical and Process Engineering, Faculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia, UKM Bangi,
Selangor, Malaysia; bDepartment of Environmental Engineering, Faculty of Civil Engineering and Planning, Institut Teknologi Sepuluh Nopember (ITS),
Keputih, Sukolilo, Surabaya, Indonesia; cTasik Chini Research Centre, Faculty of Science and Technology, Universiti Kebangsaan Malaysia, UKM Bangi,
Selangor, Malaysia; dDepartment of Civil and Structural, Faculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia, UKM Bangi,
Selangor, Malaysia; eDepartment of Land Management, Faculty of Agriculture, Universiti Putra Malaysia, Serdang, Selangor, Malaysia

ABSTRACT KEYWORDS
In this study, the removal of arsenic (As) by plant, Ludwigia octovalvis, in a pilot reed bed was optimized. A artificial neural network;
Box-Behnken design was employed including a comparative analysis of both Response Surface optimization;
Methodology (RSM) and an Artificial Neural Network (ANN) for the prediction of maximum arsenic phytoremediation; pilot
removal. The predicted optimum condition using the desirability function of both models was 39 mg kg¡1 scale; response surface
methodology
for the arsenic concentration in soil, an elapsed time of 42 days (the sampling day) and an aeration rate of
0.22 L/min, with the predicted values of arsenic removal by RSM and ANN being 72.6% and 71.4%,
respectively. The validation of the predicted optimum point showed an actual arsenic removal of 70.6%.
This was achieved with the deviation between the validation value and the predicted values being within
3.49% (RSM) and 1.87% (ANN). The performance evaluation of the RSM and ANN models showed that
ANN performs better than RSM with a higher R2 (0.97) close to 1.0 and very small Average Absolute
Deviation (AAD) (0.02) and Root Mean Square Error (RMSE) (0.004) values close to zero. Both models were
appropriate for the optimization of arsenic removal with ANN demonstrating significantly higher
predictive and fitting ability than RSM.

Introduction
Arsenic is a toxic, carcinogenic and abundant metalloid in the
environment. The cumulative arsenic production was estimated
to be 4.53 million tons worldwide due to industrial demand
(Han et al. 2003). The main causes of arsenic contamination of
drinking water, groundwater and the food chain are arsenic-pol-
luted soil, sludge and sediment. In nature, arsenic exists pre-
dominantly as arsenite (As (III)) and arsenate (As (V))
(Bhattacharya et al. 2007, Pillewan et al. 2014). Different arsenic
species show different chemical and physical characteristics and
exhibit different degrees of toxicity, mobility, and bioavailability
(Violante et al. 2010). Arsenic is very toxic to humans and is a
human carcinogen, affecting the kidneys, liver, bladder, and
skin (Jomova et al. 2011). The consumption of water polluted
with arsenic at a high concentration may cause arsenicosis and
gastrointestinal problems in humans (Pimparkar and Bhave
2010). Chronic exposure at an arsenic concentration greater
than 50 mg/L can lead to cancers, birth defects, and other
diseases (Kim et al. 2011). Thus, the permissible concentration
in water allowed by World Health Organisation (WHO) is
10 mg/L (Organization 2003). According to a United Nations
Children’s Fund (UNICEF) report in 2007, there are more than
140 million people in 70 countries affected by arsenic pollution.
Phytoremediation is a low-cost and effective technology to
remediate soils contaminated with heavy metals using certain
species of plants. It is a rising green technology that has many
advantages such as being an environmentally friendly, inex-
pensive and less destructive method compared to other con-
ventional methods (Tangahu et al. 2011, 2013, McIntyre
2003). Ludwigia octovalvis was selected for this particular
study considering its capabilities to survive in a heavy-metal-
contaminated site in Malaysia. The use of fast-growing native
plant species with a high biomass production like L. octovalvis
are highly suitable for phytoremediation of arsenic-polluted
sites (Titah et al. 2012, 2013, 2014). L. octovalvis has a sinker
root, consisting of lateral and long roots that grow just
beneath the surface of soil and the taproot (Titah et al. 2013).
The arsenic removal performance of L. octovalvis was evalu-
ated in this study using arsenic-spiked soil in a pilot reed bed
system to determine the feasibility for large-scale

CONTACT Mohd Izuan Effendi Bin Halmi m_izuaneffendi@upm.edu.my Department of Land Management, Faculty of Agriculture, Universiti Putra Malaysia,
43400 Serdang, Selangor Malaysia.
Color versions of one or more of the figures in the article can be found online at www.tandfonline.com/bijp.
Supplemental data for this article can be accessed on the publisher’s website.
© 2018 Taylor & Francis Group, LLC
722 H. S. TITAH ET AL.
phytoremediation (Titah et al. 2014, Vymazal 2005). The reed
bed system is developed to simulate the chemical, microbio-
logical and physical processes that occur naturally. A reed bed
system depends on three different factors including interac-
tion with soil microbes or rhizobacteria, the chemical and
physical attributes of the reed bed and finally the identity of
the plant itself. The performance of phytoremediation
employing a reed bed system is, however, inconsistent and
relies on several factors related to the soil and plant attributes
as well as the chemistry of arsenic in the natural environment
(Titah et al. 2013).
Optimization is performed to seek and identify the ideal
solution for certain conditions and leads to enhancing and
improving the efficiency of the process or designed system
(Myers et al. 2009). Recently, response surface methodology
(RSM) and artificial neural network (ANN) approaches have
been applied together for optimizing and modelling in many
fields including the study of the environment (Desai et al.
2008, Dutta et al. 2004). Response surface methodology
(RSM) uses a combination of specific experimental designs
to generate mathematical models of different orders, e.g., lin-
ear, quadratic, cubic, etc., to seek an optimum point from a
particular set of response variables and factors (Amini et al.
2008). In contrast, an ANN is a nonlinear computational
modelling technique that has generated increasing interest
but has only recently been applied to optimization in phytor-
emediation studies (Sanusi et al. 2016). It acts in a manner
similar to the human brain and tries to imitate human intui-
tion in reaching conclusions and making decisions when
confronted by irrelevant, noisy, complex and partial infor-
mation (Dutta et al. 2004, Basri et al. 2007). The capability
of an ANN to generalize and understand the nonlinear pro-
cess and the behaviour of any complex makes it a more use-
ful modelling tool than RSM (Bing€ol et al. 2012). The
application of ANN for optimization is found in the predic-
tion of C.I. Acid Blue 92 degradation using potential duck-
weed (Lemna Minor L) (Khataee et al. 2012).
In this study, the ANN and RSM approaches were investi-
gated and compared for their predictive, fitting and optimiza-
tion abilities for enhancing arsenic removal by L. octovalvis
using a pilot reed bed. The removal of arsenic by L. octovalvis
as a large-scale treatment method through the use of both the
RSM and ANN approaches has not been reported in the litera-
ture. Therefore, the ultimate goal of this study is to investigate
the capabilities of the RSM and ANN methods to model and
optimize large-scale arsenic removal by L. octovalvis.
Materials and method
Pilot reed bed setup and operation
L. octovalvis was propagated from plant seeds using garden-
ing soil in the greenhouse at Universiti Kebangsaan Malay-
sia. The plant seeds were grown in plastic crates with
dimensions of 37 £ 27 £ 10 cm for 3 weeks before being
transferred into polybags. The plants were 8 weeks old when
they were used in this large-scale phytoremediation. A fiber-
glass tank with dimensions of 92 £ 92 £ 60 cm and a wall
thickness of 0.5 cm was used to construct pilot reed beds. A
top layer of fine gravel (K D 10 mm) was used and a layer
of medium gravel (K D 20 mm) was placed at the bottom of
the reed bed. The thicknesses of both the fine and medium
gravel layers were 10 cm each. The As-spiked soil of 10 cm
depth was placed into the reed bed on the top of the fine
gravel. The detailed procedure of this pilot study can be
referred in Titah et al (2013, 2014). The sand used in this
study was spiked with solution of an As (V) salt, sodium
arsenate dibasic heptahydrate (Sigma Aldrich, USA). The As
(V) solution was mixed with the sand using manual method
(Titah et al. 2013). During treatment, plants were watered at
alternate days.
Arsenic concentration determination using inductively
coupled plasma-optical emission spectrometer (ICP-OES)
The wet digestion method (USEPA 1996) was used in the
determination of the total extractable arsenic concentration
in the soil (USEPA 1996; Peters and Basta 1996). The
method was carried out as follows: Two-gram samples of
the sand from the treatment system was added to 5 mL of
69% HNO3 in 100 mL Erlenmeyer flasks and incubated
overnight. The concentration of the arsenic extracted using
this method was measured using an Optima 7300DV ICP-
OES apparatus (Perkin Elmer, USA) (Lomonte et al. 2008)
with detection limits (the arsenic concentration equivalent
to three times of the standard deviation of the blank
divided into the slope of the calibration graph) of arsenic
was 0.1 ppm.
Optimization using response surface methodology (RSM)
The RSM was applied to optimize the interactive effects of
arsenic removal by L. octovalvis with the following three
variables: arsenic concentration in the soil (A), sampling
day (B), and aeration rate (C). A Box-Behnken design
(BBD) was employed as a design in this comparative study
with 17 total experiments (Al-Baldawi et al. 2014). The
three different variables were coded at three different levels
(–1, 0, and 1) representing low, medium and high levels
respectively (A: 5, 22, and 39 mg kg¡1; B: 14, 28, and
42 days; C: 0, 1, and 2 L min¡1). Design Expert (Version
6.0.1, Stat-Ease Inc., Minneapolis, MN, USA) was used for
data interpretation and regression modelling. The experi-
mental results were assessed using a number of regressions
and the significance of the regression value was examined
using an F test (Zheng and Wang 2010). An appropriate
design must be significant and lack of fit should be insignif-
icant (Noordin et al. 2004). The model was examined using
the coefficient of determination (R2) and adjusted coefficient
of determination (adjusted R2), where both correlations
should be close to 1.0 to demonstrate a good relationship
between the predicted and experimental values (Muhamad
et al. 2013; Bezerra et al. 2008). The significant term was
determined for the response used in this study. From the
regression model obtained, the regression coefficients were
used for the statistical calculations to generate response sur-
face plots (Myers et al. 2009). The observed responses from
the BBD design were then fitted to the following polynomial

equation as presented below (Equation (1)):
Y D b0 C b1 A C b2 B C b3 C C b12 AB C b13 CB C b23 CB
(1)
C b11 A2 C b22 B2 C b23 C 2 C e
where Y is the response (total arsenic removal from soil (%)
efficiency); A (arsenic concentration in soil), B (sampling
day), and C (aeration rate) represent the effect of the indepen-
dent variables; A2, B2, C2 are the square effects; AB, AC, and
BC are the interaction effects; b0 is constant; b1, b2, and b3 are
the linear coefficients; b12, b13, and b23 are the interaction
coefficients; b11, b22, and b33 are quadratic coefficients; and e is
the random error.
Optimization using artificial neural network (ANN)
An artificial neural network was employed as an alternative
to the RSM modelling tool in this comparative analysis (Ravi
Kumar Dasari et al. 2009). According to previous literature,
the ANN model is potentially more precise and accurate in
the modelling of complex nonlinear relationships than RSM
and other modelling solutions (Saha 2013). In this study, the
ANN Predictive modelling was developed using three differ-
ent parameters, i.e., the concentration of arsenic in soil, sam-
pling day and aeration rate as inputs to the model, and
arsenic removal from soil (%) as the output from the model,
respectively. ANN was applied in this study to predict the
nonlinear relationship between variables (input) and
responses (output). The experimental results obtained from
these 17 experiments were analyzed using intelligent problem
solver, STATISTICA neural network software (Version 7)
from Stat Soft Co. Ltd. ANNs can be divided into many clas-
ses such as radial basis function networks (RBF) and multi-
layer feed forward networks (MLP) with the multilayer feed
forward network (MLP) having a particularly wide applica-
tion in prediction and optimization process modelling
(Buci" nski et al. 2008; Park and Sandberg 1993). The multi-
layer feed forward network (MLP) topology consists of sev-
eral layers such as the input layer of neurons, hidden layers
and output layer. The analysis was carried out using auto-
mated neural networks comprising two different phases: a
training data set and a testing data set in which the experi-
mental data were randomly separated into 70% from the
training data set and 30% from the testing data set. As the
training proceeds, the software auto-searches for the best
network by varying the number of hidden layers and the
nodes within each layer to control the learning rate and
momentum of the algorithm. Four activation functions
(identity, logistic, tanh, and exponential) were used during
the training process for the hidden and output layers (Betiku
et al. 2014). The best layer was selected according to an eval-
uation of the highest coefficient of determination (R2) and
the lowest selection error. The best network was selected and
used in the optimization and prediction analysis. The per-
formances of the training and testing of the model were eval-
uated using the coefficient of determination (R2) (Equation
2), the average absolute deviation (AAD) (Equation 3) and
the root mean square error (RSME) (Equation 4) between
INTERNATIONAL JOURNAL OF PHYTOREMEDIATION 723
the predicted values of the network and the experimental
values. The network gave a good correlation coefficient and
determination (R2) with a value close to one. The smallest
values of AAD and RSME are thought to represent a good
network and describe the precision of the system (Basri et al.
2007, Ebrahimpour et al. 2008).
X
½.Modal Prediction/i ¡ .Experimental value/i "2
R2 D 1 ¡ X iD1¡n
½.Average experimental value/i ¡ .Experimental value/i "2
iD1¡n
(2)
(" p
# )
X ! "
AAD D j yi;exp ¡ yi;cal j =yi;exp =p £100 (3)
iD1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sX
.yi;exp ¡ yi;cal /2
RESM D (4)
n
where yi,exp and yi,cal are the experimental and calculated
responses, respectively, and p is the number of the experimen-
tal run. The value n is the number of experimental data points.
Scanning electron microscope (SEM) and energy dispersive
X-ray analysis (EDX)
The aim of this analysis was to look for the effect of arsenic on
tissues of roots, stems and leaves of L. octovalvis using an
FESEM Model Supra 55VP (Zeiss, Germany). The analysis was
carried out after the plant was exposed to arsenic for 42 days.
Sample preparation was performed according to a method
from Pathan et al. (2008).
Results and Discussion
Optimization using response surface methodology (RSM)
The Box-Behnken design, a second-order experimental
design in response surface methodology, was applied to
model the experiment for improving, developing and opti-
mizing the removal of arsenic from soil (Ferreira et al.
2007). The design was employed to find out the optimum
conditions of the arsenic concentration in soil, the sampling
day and the aeration rate for arsenic removal from soil (%)
by L. octovalvis. Adequacy checking for the model was con-
ducted in this study to determine whether the approximating
model would give inaccurate results (Pradhan and Biswas
2009). Four different high-degree polynomial models such as
linear, 2F1, quadratic and cubic models were applied and fit-
ted to the experimental results to represent the relationship
between the variables and the response (total arsenic removal
from soil) (Saqib et al. 2012). Several tests, namely the
sequential model sum of squares, model summary statistics
and lack-of-fit tests, were performed in this study to deter-
mine the adequacy of models among various models. Based
on the sequential model sum of squares, the quadratic model
was significant (p < 0.05) compared to other models. Thus,
the quadratic model was chosen as the best model. A qua-
dratic model was developed to fit the coefficients acquired
through multiple regression analysis. The quadratic model
724 H. S. TITAH ET AL.

obtained is presented in Equation (1) as below: Table 2. ANOVA analysis of the quadratic model.

Analysis of variance table [Partial sum of squares]

½Total As removal from soil" D ½ C 67:22 ¡ 4:82A C 5:28B Sum of Mean F
¡ 1:22C ¡ 3:75A2 C 10:25B2 ¡ 17:86C2 Source Squares DF Square Value Prob. > F
¡ 1:36AB ¡ 10:59AC -- 2:73BC" Model 2689.465 9 298.8294 10.87964 0.0024 significant
A 185.5701 1 185.5701 6.756148 0.0355
B 223.1328 1 223.1328 8.123713 0.0247
Table 1 shows that arsenic removal from soil ranged from C 11.9072 1 11.9072 0.433512 0.5313
34.25 to 87.07%. The maximum arsenic removal from soil A (Bhattacharya 59.33691 1 59.33691 2.16031 0.1851
et al. 2007)
(87.07%) was observed in run 12 with the experimental condi- B (Bhattacharya 442.0232 1 442.0232 16.09297 0.0051
tions of arsenic concentration in soil (A, 5 mg kg¡1), sampling et al. 2007)
day (B, 42 days) and aeration rate (C, 1 L min¡1). The mini- C (Bhattacharya 1343.298 1 1343.298 48.90614 0.0002
et al. 2007)
mum arsenic removal from soil (34.25%) was observed with AB 7.4529 1 7.4529 0.271342 0.6185
the experimental condition of arsenic concentration in soil (A, AC 448.8042 1 448.8042 16.33985 0.0049
39 mg kg¡1), sampling day (B, 28 days) and aeration rate (C, BC 29.75703 1 29.75703 1.08338 0.3326
Residual 192.268 7 27.46685
2 L min¡1). The optimum process condition must be consid- Lack of Fit 115.3645 3 38.45483 2.00016 0.2564 not significant
ered to achieve maximum arsenic removal efficiency from soil. Pure Error 76.90348 4 19.22587
The analysis of variance (ANOVA) is shown in Table 2. The Corr. Total 2881.733 16
computed model F-value is 2689.465 and the corresponding p-
value is <0.05, implying that the model is significant. There is
only a 0.24% possibility that a ‘‘Model F-value’’ could occur Table 1 shows the actual experimental values and RSM-pre-
due to noise (Francis et al. 2003). The lack of fit was not signifi- dicted values of the response. Figure 1a indicates the relation-
cant as the p-value was 0.2564. For the model predicted by ship involving the predicted output with actual experimental
Equation (1), the p-value was less than 0.05, confirming that it values for arsenic removal from soil by L. octovalvis. It can be
is significant for describing the efficiency of arsenic removal observed that the predicted values using the RSM quadratic
from soil. In this case, the individual variables in the quadratic model were in agreement with the experimental values with an
model of arsenic concentration in soil (A) and sampling day acceptable correlation among these values. Hence, the estab-
(B) are quite significant because the p-value is lower than 0.05. lished RSM model is appropriate for predicting the perfor-
The model shows an acceptable coefficient of determination mance of arsenic removal under the investigated conditions.
(R2 D 0.930) and adjusted coefficient of determination The satisfactory approximation of the chosen model was tested
(adjusted-R (Bhattacharya et al. 2007) D 0.84), indicating that using the diagnostic plots accessible in Design Expert 6.0.10
the model is adequate to represent the actual relationship software, which include the studentized residuals plotted
between the response and the significant variables (Al-Baldawi against the normal probability, predicted versus studentized
et al. 2014). The model was used to determine the quadratic residuals, run versus studentized residuals and run versus out-
effects and interaction of arsenic removal from soil in the pilot lier (Hasan et al. 2011). Figure S1a indicates that the studen-
scale operation with optimal conditions. tized residuals plotted versus the normal probability exhibited a
straight line, demonstrating a normal distribution of the experi-
Table 1. Box-Behnken design matrix for the three independent variables with the
mental data. Figure S1b, S1c, and S1d show that the predicted
observed and predicted response for arsenic removal. versus studentized residuals, run versus studentized residuals
and run versus outlier all lie within the interval §3.50, suggest-
Total As removal from soil (%)
A: Arsenic C: Aeration ing that the approximation of the model was good without
Conc. in soil B: Sampling rate Experimental RSM ANN error.
Run (mg kg¡1) day (day) (L min¡1) values Predicted Predicted

1 22 14 2 52.45 55.8288 52.37206

2 22 42 2 58.97 60.9363 58.40792 Optimization using artificial neural network
3 22 28 1 60.57 67.218 59.11453
4 39 14 1 63.08 64.9775 63.96535 An ANN offers a unique platform for dealing with the multi-
5 22 42 0 72.21 68.8313 72.46866 variate regression analysis and modelling of nonlinear experi-
6 22 28 1 71.62 67.218 72.22499
7 39 28 0 52.53 52.5988 51.99223 mental problems (Heiat 2002). The method of developing an
8 22 28 1 70.55 67.218 71.22499 ANN model is entirely different from the development tech-
9 39 28 2 34.25 28.9738 33.13092 nique using polynomial RSM. It is worth describing several dif-
10 5 28 0 35.77 41.0463 36.70356
11 5 14 1 75.19 71.88 73.36989 ferences involving RSM and ANN. RSM performs an analysis
12 5 42 1 87.07 85.1725 86.00799 using variables or factors and response. In contrast, in ANN,
13 5 28 2 59.86 59.7913 67.90245 the variables are called inputs, whereas the response is known
14 39 42 1 69.5 72.81 71.24713
15 22 28 1 67.58 67.218 67.22499 as the target (experimental response) or output (predicted
16 22 14 0 54.78 52.8138 55.29798 response) (Desai et al. 2008). In this comparative modelling
17 22 28 1 65.77 67.218 64.42321 study, multilayer feedforward neural networks (MLP) and a
ANN training set: normal. BFGS algorithm were applied to model arsenic removal by L.
ANN testing set: bold numbers. octovalvis for figuring biases and ANN weight. The MLP
 INTERNATIONAL JOURNAL OF PHYTOREMEDIATION 725

100 According to Table 1, the experimental data with boldface

numbers were from the testing data set and the rest were from
the training data set. The ANN-predicted data values for each
80
experimental datum are summarized in Table 1. The result
shows a close relationship between actual experimental values
RSM Predicted data

60 and ANN predicted values, which indicates that the ANN is

able to fit the actual experimental data accurately (Dutta et al.
R² = 0.930
2004). The close relationship between the data could be seen
40
through plotting the actual experimental data with ANN-pre-
dicted data versus the experimental run. Several neural network
20 architectures and topologies were examined using the auto-
mated neural network to determine the optimum number of
neurons in the hidden layer for estimation and prediction of
0 the responses. The best five ANN models are presented in
0 20 40 60 80 100
Experimental data
Table 3. The best topology was network no. 3, MLP 3-9-1,
which has nine neurons as the optimum. This feed-forward
a network topology (3-9-1) refers to the number of the neurons
in the inputs, the hidden layers and output layer, respectively.
100 All neurons from the hidden layer and the output layer neuron
have exponential and tan as the transfer function, respectively.
The network was chosen due to the values of R2 of both the
80
training and testing sets being close to 1.0 and also because
they showed less error than other networks (Ebrahimpour et al.
ANN Predicted data

60 2008). For the training data set, the coefficient of determination

R² = 0.973
40
20
0
0 20 40 60
Experimental data
b the ANN gave a good prediction due to the value of R2 being
Figure 1. (a) RSM predicted and (b) ANN predicted versus actual experimental
data values for arsenic removal by L. octovalvis.
network consists of the input layer, a number of hidden layers
and also an output. The inputs employed in this investigation
were the arsenic concentration in soil, the sampling day, and
the aeration rate. The target response was the arsenic removal
from soil (%). The determination of the number of hidden
layers is necessary to develop a much better ANN model and to
prevent over-fitting due to over-training. The experimental
result from the Box-Behnken Design (Table 1) was used in the
development of the ANN model. The experimental data (17
data points) were randomly trained and tested. The ANN pre-
dicted data are presented in Table 1.
(R2) and training error were 0.989215 and 0.003069, respec-
tively, whereas for the testing data set, R2 was 0.993662 and the
test error was 0.002090. This shows that the model obtained
from the ANN can be used effectively to describe the input var-
iables for arsenic removal.
The predicted values are very close to the actual values
(Figure 1b), which indicates that the nonlinear fitting effects of
the model are very good. The accuracy of the models obtained
80 100 from the ANN was examined by evaluating the values of R2,
AAD and RMSE (Nelofer et al. 2012). The results showed that
0.973, which is approximately 1.0, and the values of AAD and
RMSE being small, i.e., 0.024 and 0.0045, respectively. The
ANN showed a better fitting performance because of the higher
value of R2 and smaller values of AAD and RMSE.
Determination of the optimal point using the desirability
function method
The optimum condition predicted by both RSM and ANN was
determined by means of the desirability function method
(Bezerra et al. 2008). This method includes desires and priori-
ties for each of the variables and builds up a procedure for
determining the relationship between predicted arsenic
removal for each variable and the desirability of the responses.
When working in the pilot scale, the variables were set as

Table 3. Summary of active networks.

Index Net. name Training perf. Test perf. Training error Test error Training algorithm Hidden activation Output activation

1 MLP 3-4-1 0.834283 0.999374 0.013053 0.000184 BFGS 21 Tanh Tanh

2 MLP 3-10-1 0.930380 0.994949 0.005838 0.000131 BFGS 19 Exponential Exponential
3 MLP 3-9-1 0.989215 0.993662 0.003069 0.002090 BFGS 18 Exponential Tanh
4 MLP 3-8-1 0.943357 0.977610 0.036734 0.011855 BFGS 4 Identity Identity
5 MLP 3-7-1 0.927721 0.994551 0.035673 0.007311 BFGS 3 Exponential Logistic
726 H. S. TITAH ET AL.
follows: arsenic concentration in soil was set to maximum,
sampling day was set to maximum and the aeration rate was
set to minimum (Al-Baldawi et al. 2014).
Using the functions of numerical optimization in the Design
Expert software 6.01, four solutions were suggested for the opti-
mum condition with a desirability value greater than 0.860. It
shows a desirability of 0.922 for the maximum arsenic removal
from soil efficiencies. The maximum arsenic removal of 72.6%
was achieved at the optimized conditions of 39 mg kg¡1 arsenic
in soil (A), sampling day 42 (B) and 0.22 L min¡1 for aeration
rate (C). The Profiler in STATISTICA Version 7 (StatSoft) was
used to determine the optimum point using an Artificial Neural
Network (ANN). The ANN predicted a maximum arsenic
removal occurring at 71.4% under the optimized conditions,
similar to the predicted point by RSM.
The validation result showed that approximately 70.06%
arsenic removal was achieved for RSM and ANN. The valida-
tion values were found to be in close agreement with the pre-
dicted values, indicating the satisfactoriness of the obtained
model to optimize arsenic removal. The deviation between the
validation value and predicted values was within 3.49% and
1.87% for RSM and ANN, respectively. Thus, it can be con-
cluded that both models were appropriate in prediction and fit-
ting with ANN showing less error compared to RSM. This
implies that the comparative optimization of pilot-scale arsenic
removal by both the response surface method and an artificial
neural network was considered successful.
Performance evaluation of RSM and ANN models
The comparative study between the two models was performed
to assess the reliability and accuracy of the models (Betiku et al.
2014). The predictive ability of both models is usually decided
by R2 to support the model’s precision (Ebrahimpour et al.
2008). The R2 value must be close to 1.0 for a good model. The
correlation coefficient for RSM and ANN were 0.930 and 0.973,
respectively, with ANN showing a greater value of R2, which is
closer to 1.0 than that of RSM. However, a large value of R2
close to 1.0 does not mean that the regression model is an effec-
tive model. Other evaluation parameters such as AAD and
RMSE should be considered to validate more than one model
(Noordin et al. 2004). For a good model, the AAD and RMSE
values need to be as small as possible and should be near zero.
The higher values of RMSE and AAD indicate a higher possi-
bility of errors of prediction in both models. The AAD value
(0.4) for RSM was greater than the AAD value (0.024) for
ANN. In addition, the RMSE value (1.198) for RSM was greater
than the RMSE value (0.0045) for ANN. The built ANN model
has a greater R2, whereas the AAD and RMSE values are
smaller than the respective RSM values. Hence, the ANN has
demonstrated significantly higher accuracy and predictive abil-
ity than the RSM.
3D surface plot under the optimum conditions
The response surfaces of total arsenic removal from soil by
L. octovalvis have been established using a quadratic model
as the mathematical equation. Figure 2a–2c shows the rela-
tionships between the arsenic concentration in soil (A),
sampling day (B) and aeration rate (C). Figure 2a illustrates
the 3D surface plot of the effect of arsenic concentration in
soil (A) and sampling day (B) on arsenic removal. It can be
observed that an arsenic concentration around the range of
5–39 mg kg¡1 brought arsenic removal to a steady state.
This could possibly be due to the range chosen including
arsenic concentrations that were not significantly different.
The effect of sampling day (B) was an increase in the
response within just 14–42 days. It is believed that the later
sampling day results in higher arsenic removal (Han et al.
2003). Figure 2b shows the effect of arsenic concentration in
soil (A) and aeration rate (C) on arsenic removal from soil.
It was discovered that arsenic removal from soil was
enhanced after the aeration rate was increased. It was proven
that there was an increase in the arsenic removal from soil
when the aeration rate was increased from 0 to 1.5 L min¡1
and that it later slowly decreased after the aeration rate was
increased to 2 L min¡1; the aeration rate was still effective
for arsenic removal from soil even with the rate increased to
2 L min¡1. The actual arsenic removal from soil was 60% at
an aeration rate of 1.5 L min¡1, which is sufficient in the
pilot system to boost oxygen availability in soil for enhancing
microbial activities around the root (Titah et al. 2013). The
efficiency of phytoremediation by L. octovalvis may not
depend on only the plant itself but also on the interaction of
the plant roots with microbes. Figure 2c exhibits the effect of
sampling day (B) and aeration rate (C) on arsenic removal
from soil. It was discovered that arsenic removal was
enhanced with an increase in the aeration rate, but the effect
of sampling day was not significantly different (p > 0.05).
This proves that aeration rate is important to enhance arse-
nic removal from soil. A later sampling day and sufficient
aeration rate are very important for the maximum removal
of pollutants by phytoremediation. Because we needed the
most affordable treatment system, the aeration rate was set
at the minimum (Al-Baldawi et al. 2014).
SEM and EDX analysis
To evaluate the effect of arsenic on plant structure of L. octoval-
vis, the plant parts such as root, stem and leaves were analysed
using scanning electron microscopy (SEM) with energy disper-
sive X-ray (EDX). The SEM analysis of the control (without
arsenic) showed that the cross sections of the root could be
seen clearly after the 42-day treatment (Figure S2a). After the
plants were exposed to arsenic at a high concentration, the root
sections were less visible due to the toxic effects of arsenic
(Figure S2d). Figure S2b shows the SEM analysis of the stem
without arsenic exposure (control). It has a very clear appear-
ance without any toxic effect, whereas the plant treated with
arsenic showed damage of the stem tissues (Fig. S2e). The SEM
analysis of the leaves after being treated with arsenic showed
that the damage to the leaf tissue and the trichomes were not
visible after this treatment (Figure S2f) in comparison with the
control (Figure S2c). The results of the EDX showed that no
arsenic was detected in the roots, stems and leaves of the con-
trol plant (Figure S2a-S2c), whereas at an arsenic concentration
of 39 mg kg¡1, the presence of arsenic was shown in the root,
stem and leaves (Figure S2d-S2f).
 INTERNATIONAL JOURNAL OF PHYTOREMEDIATION 727

Figure 2. RSM Response surface for arsenic removal by L. octovalvis as a function of the variables: (a) Aeration rate (C) and arsenic concentration in soil (A), (b) Sampling
day (B) and arsenic concentration in soil (A), (c) Aeration rate (C) and Sampling day (B).

Conclusions
The comparative optimization of a pilot reed bed associated
with the efficiency of As removal from soil by L. octovalvis was
successfully performed using RSM and ANN. Both models
predicted the same optimum point: 39 mg kg¡1 for the As con-
centration in soil, a sampling day of 42 and 0.22 L min¡1 aera-
tion rate. These optimum points produced the predicted values
of As removal by RSM (72.6%) and ANN (71.4%) with the
deviation between the validation value and the predicted values
728 H. S. TITAH ET AL.
was within 3.49% and 1.87% for the RSM and ANN, respec-
tively. The result showed the built ANN model has demon-
strated significant higher predictive and fitting ability than
RSM. The cross sections of SEM for the root, stem and leaves
control could be seen clearly without any toxicity effect com-
pared to the cross section after exposure with As.
Funding
The authors would like to thank Tasik Chini Research, Universiti Kebang-
saan Malaysia (UKM) and the Ministry of Higher Education, Malaysia for
funding this research under FRGS/1/2015/SG05/UKM/01/1.
ORCID
Nurina Anuar http://orcid.org/0000-0003-2364-3634
References
Al-Baldawi IAW, Abdullah SRS, Hasan HA, Suja F, Anuar N, Mushrifah I.
2014. Optimized conditions for phytoremediation of diesel by Scirpus
grossus in horizontal subsurface flow constructed wetlands (HSFCWs)
using response surface methodology. J Environ Manage. 140:152–159.
Amini M, Younesi H, Bahramifar N, Lorestani AAZ, Ghorbani F, Daneshi
A, Sharifzadeh M. 2008. Application of response surface methodology
for optimization of lead biosorption in an aqueous solution by Aspergil-
lus niger. J Hazard Mater. 154(1):694–702.
Basri M, Rahman RN, Ebrahimpour A, Salleh AB, Gunawan ER, Rahman
MB. 2007. Comparison of estimation capabilities of response surface
methodology (RSM) with artificial neural network (ANN) in lipase-cat-
alyzed synthesis of palm-based wax ester. BMC Biotechnol. 7(1):53.
Betiku E, Omilakin OR, Ajala SO, Okeleye AA, Taiwo AE, Solomon BO.
2014. Mathematical modeling and process parameters optimization
studies by artificial neural network and response surface methodology:
a case of non-edible neem (Azadirachta indica) seed oil biodiesel syn-
thesis. Energy. 72:266–273.
Bezerra MA, Santelli RE, Oliveira EP, Villar LS, Escaleira LA. 2008.
Response surface methodology (RSM) as a tool for optimization in ana-
lytical chemistry. Talanta. 76(5):965–977.
Bhattacharya P, Welch AH, Stollenwerk KG, McLaughlin MJ, Bundschuh
J, Panaullah G. 2007. Arsenic in the environment: biology and chemis-
try. Sci Total Environ. 379(2):109–120.
Bing€ol D, Hercan M, Elevli S, Kılıç E. 2012. Comparison of the results of
response surface methodology and artificial neural network for the bio-
sorption of lead using black cumin. Bioresour Technol. 112:111–115.
Buci"
nski A, Karama"c M, Amarowicz R, Pegg RB. 2008. Modeling the tryp-
tic hydrolysis of pea proteins using an artificial neural network. LWT –
Food Sci Technol. 41(5):942–945.
Desai KM, Survase SA, Saudagar PS, Lele S, Singhal RS. 2008. Comparison
of artificial neural network (ANN) and response surface methodology
(RSM) in fermentation media optimization: Case study of fermentative
production of scleroglucan. Biochem Eng J. 41(3):266–273.
Dutta JR, Dutta PK, Banerjee R. 2004. Optimization of culture parameters
for extracellular protease production from a newly isolated Pseudomo-
nas sp. Using response surface and artificial neural network models.
Process Biochem. 39(12):2193–2198.
Ebrahimpour A, Rahman RN, Ch’ng DHE, Basri M, Salleh AB. 2008. A
modeling study by response surface methodology and artificial neural
network on culture parameters optimization for thermostable lipase
production from a newly isolated thermophilic Geobacillus sp. strain
ARM. BMC Biotechnol. 8(1):96.
Ferreira SC, Bruns R, Ferreira H, Matos G, David J, Brandao G, da Silva
EP, Portugal L, Dos Reis P, Souza A. 2007. Box-Behnken design: an
alternative for the optimization of analytical methods. Anal Chim Acta.
597(2):179–186.
Francis F, Sabu A, Nampoothiri KM, Ramachandran S, Ghosh S, Szakacs
G, Pandey A. 2003. Use of response surface methodology for optimiz-
ing process parameters for the production of a-amylase by Aspergillus
oryzae. Biochem Eng J. 15(2):107–115.
Han FX, Su Y, Monts DL, Plodinec MJ, Banin A, Triplett GE. 2003. Assess-
ment of global industrial-age anthropogenic arsenic contamination.
Naturwissenschaften. 90(9):395–401.
Hasan HA, Abdullah SRS, Kamarudin SK, Kofli NT. 2011. Response sur-
face methodology for optimization of simultaneous COD, NH4C–N
and Mn2C removal from drinking water by biological aerated filter.
Desalination. 275(1):50–61.
Heiat A. 2002. Comparison of artificial neural network and regression
models for estimating software development effort. Inform Software
Technol. 44(15):911–922.
Jomova K, Jenisova Z, Feszterova M, Baros S, Liska J, Hudecova D, Rhodes
C, Valko M. 2011. Arsenic: toxicity, oxidative stress and human disease.
J Appl Toxicol. 31(2):95–107.
Khataee A, Movafeghi A, Torbati S, Lisar SS, Zarei M. 2012. Phytoremedia-
tion potential of duckweed (Lemna minor L.) in degradation of C.I.
acid blue 92: artificial neural network modeling. Ecotoxicol Environ
Saf. 80:291–298.
Kim K-W, Chanpiwat P, Hanh HT, Phan K, Sthiannopkao S. 2011. Arse-
nic geochemistry of groundwater in southeast asia. Front Med. 5
(4):420–433.
Lomonte C, Gregory D, Baker AJ, Kolev SD. 2008. Comparative study of
hotplate wet digestion methods for the determination of mercury in
biosolids. Chemosphere. 72(10):1420–1424.
McIntyre T. 2003. Phytoremediation of heavy metals from soils. Adv Bio-
chem Eng Biotechnol. 78:97–123.
Muhamad MH, Abdullah SRS, Mohamad AB, Rahman RA, Kadhum
AAH. 2013. Application of response surface methodology (RSM) for
optimisation of COD, NH3–N and 2, 4-DCP removal from recycled
paper wastewater in a pilot-scale granular activated carbon sequencing
batch biofilm reactor (GAC-SBBR). J Environ Manage. 121:179–190.
Myers RH, Montgomery DC, Anderson-Cook CM. 2009. Response surface
methodology: process and product optimization using designed experi-
ments. 4th ed. Hoboken, NJ: John Wiley & Sons.
Nelofer R, Ramanan RN, Rahman RNZRA, Basri M, Ariff AB. 2012. Com-
parison of the estimation capabilities of response surface methodology
and artificial neural network for the optimization of recombinant
lipase production by E. coli BL21. J Ind Microbiol Biotechnol. 39
(2):243–254.
Noordin MY, Venkatesh VC, Sharif S, Elting S, Abdullah A. 2004. Applica-
tion of response surface methodology in describing the performance of
coated carbide tools when turning AISI 1045 steel. J Mater Process
Technol. 145(1):46–58.
Organization WH. 2003. Arsenic in drinking water. IWA.
Park J, Sandberg IW. 1993. Approximation and radial-basis-function net-
works. Neural Comput. 5(2):305–316.
Pathan A, Bond J, Gaskin R. 2008. Sample preparation for scanning elec-
tron microscopy of plant surfaces—horses for courses. Micron. 39
(8):1049–1061.
Peters J, Basta N. 1996. Reduction of excessive bioavailable phosphorus in
soils by using municipal and industrial wastes. J Environ Qual. 25
(6):1236–1241.
Pillewan P, Mukherjee S, Meher AK, Rayalu S, Bansiwal A. 2014. Removal
of arsenic (III) and arsenic (V) using copper exchange zeolite#a. Envi-
ron Prog Sustainable Energy. 33(4):1274–1282.
Pimparkar B, Bhave A. 2010. Arsenicosis: review of recent advances.
J Assoc Physicians India. 58:617–624.
Pradhan M, Biswas C. 2009. Modeling and analysis of process parameters
on surface roughness in EDM of AISI D2 tool steel by RSM approach.
Ravi Kumar Dasari VR, Reddy Donthireddy SR, Yugandhar Nikku M, Rao
Garapati H. 2009. Optimization of medium constituents for cephalo-
sporin C production using response surface methodology and artificial
neural networks. J Biochem Technol. 1(3):69–74.
Saha PD. 2013. Mathematical modeling of the reduction of safranin
onto chemically modified rice husks in stirred tank reactor using
response surface methodology and artificial neural network. Biorem J.
17(1):52–60.

Sanusi SNA, Halmi MIE, Abdullah SRS, Hassan HA, Hamzah FM, Idris M.
2016. Comparative process optimization of pilot-scale total petroleum
hydrocarbon (TPH) degradation by Paspalum scrobiculatum L. Hack
using response surface methodology (RSM) and artificial neural net-
works (ANNs). Ecol Eng. 97:524–534.
Saqib M, Mumtaz MW, Mahmood A, Abdullah MI. 2012. Optimized bio-
diesel production and environmental assessment of produced biodiesel.
Biotechnol Bioprocess Eng. 17(3):617–623.
Tangahu BV, Abdullah SRS, Basri H, Idris M, Anuar N, Mukhlisin M.
2013. Phytoremediation of wastewater containing lead (Pb) in pilot
reed bed using Scirpus grossus. Int J Phytorem. 15(7):663–676.
Tangahu BV, Sheikh Abdullah SR, Basri H, Idris M, Anuar N, Mukhlisin
M. 2011. A review on heavy metals (As, Pb, and Hg) uptake by plants
through phytoremediation. Int J Chem Eng. 2011.
Titah HS, Abdullah SRS, Idris M, Anuar N, Basri H, Mukhlisin M.
2012. Arsenic range finding phytotoxicity test against Ludwigia
octovalvis as first step in phytoremediation. Res J Environ Toxicol.
6(4):151.
INTERNATIONAL JOURNAL OF PHYTOREMEDIATION 729
Titah HS, Abdullah SRS, Mushrifah I, Anuar N, Basri H, Mukhlisin M.
2013. Arsenic toxicity on Ludwigia octovalvis in spiked sand. Bull Envi-
ron Contam Toxicol. 90(6):714–719.
Titah HS, Abdullah SRS, Mushrifah I, Anuar N, Basri H, Mukhlisin M.
2014. Optimization of arsenic phytoremediation by Ludwigia octovalvis
in pilot reed bed system using response surface methodology. In: Aris
A, Tengku Ismail T, Harun R, Abdullah A, Ishak M, editors. From
sources to solution. Singapore: Springer. p. 251–255.
USEPA. 1996. SW 846. Method 3050b – acid digestion of sediments,
sludge and soils, rev. 2.
Violante A, Cozzolino V, Perelomov L, Caporale A, Pigna M. 2010. Mobil-
ity and bioavailability of heavy metals and metalloids in soil environ-
ments. J Soil Sci Plant Nutr. 10(3):268–292.
Vymazal J. 2005. Horizontal sub-surface flow and hybrid constructed wet-
lands systems for wastewater treatment. Ecol Eng. 25(5):478–490.
Zheng Y, Wang A. 2010. Removal of heavy metals using polyvinyl alcohol
semi-IPN poly (acrylic acid)/tourmaline composite optimized with
response surface methodology. Chem Eng J. 162(1):186–193.