Neural Comput & Applic

DOI 10.1007/s00521-016-2404-7

ORIGINAL ARTICLE

Application of artificial intelligence (AI) techniques in water

quality index prediction: a case study in tropical region, Malaysia
Mohammed Hameed1 • Saadi Shartooh Sharqi2 • Zaher Mundher Yaseen1 •

Haitham Abdulmohsin Afan1 • Aini Hussain3 • Ahmed Elshafie4

Received: 26 February 2016 / Accepted: 30 May 2016

 The Natural Computing Applications Forum 2016

Abstract The management of river water quality is one performance evaluation and the sensitivity analysis of the
the most significant environmental challenges. Water variables have been conducted. The results achieved are
quality index (WQI) describes several water quality vari- positively promising with high performance accuracy
ables at a certain aquatic environment and time. Classi- belonging to RBFNN model for both scenarios. Further-
cally, WQI is commonly computed using the traditional more, the proposed approach offers an effective alternative
methods which involved lengthy computation, consume to compute and predict WQI, to the fact that WQI manual
timing and occasionally associated with accidental errors calculation methods involved lengthy computations,
during subindex calculation. Thus, providing an accurate transformations, use of various subindex formulae for each
prediction model for WQI is highly required. Recently, the value of the constituent water quality variables, and con-
artificial neural networks (ANNs) have been examined for suming time.
similar prediction applications and exhibited a remarkable
ability to capture the nonlinearity pattern between predic- Keywords Artificial neural networks  Water quality
tors and predictand. In the current research, two different index  Tropical environment  RBFNN  BPNN  Water
ANN algorithms, namely radial basis function neural net- quality variables
work (RBFNN) and back propagation neural networks
models, have been applied to examine and mimic the
relationship of WQI with the water quality variables in a 1 Introduction
tropical environment (Malaysia). The input variables cat-
egorized into two different architectures and have been Last two decades, water river pollution considered as one
inspected. In addition, comprehensive analysis for the of the global concern to the communities which is required
full attention from the environmental researchers [1].
However, water river quality is one of the main charac-
& Zaher Mundher Yaseen
[email protected]
teristics that needs full attention from environmental
scholars. The quality of the water becomes a growing
1
Civil and Structural Engineering Department, Faculty of concern throughout the developing world. The process of
Engineering and Built Environment, Universiti Kebangsaan abstraction water for domestic use, agricultural production,
Malaysia (UKM), 43600 Bangi, Selangor Darul Ehsan,
Malaysia
mining industrial production, power generation, and for-
2
estry practices can lead to deterioration in water quality
Department of Civil Engineering, Al Anbar University,
Al Anbar, Iraq
and quantity that impact not only the aquatic ecosystem
3
(i.e., the assemblage of organisms living and interacting
Department of Electrical, Electronics and Systems
together within an aquatic environment), but also the
Engineering, Faculty of Engineering, Universiti Kebangsaan
Malaysia (UKM), 43600 Bangi, Selangor Darul Ehsan, availability of safe water for human consumption. Thus,
Malaysia assessment of the quality of surface waters is important in
4
Civil Engineering Department, Faculty of Engineering, hydro-environmental management and it is very significant
University of Malaya, 50603 Kuala Lumpur, Malaysia in monitoring the concentration of pollutants in rivers. The

123
 Neural Comput & Applic

evaluation of the river water quality is a phrase to describe method that used for calculating the WQI in Malaysia
the chemical, biological, and physical characteristic of involves lengthy calculations, transformations, consum-
water. ing time, and effort. Therefore, recommending an alter-
Over the past few decades, increasing in number of native technique which is direct and quicker with high
industrial areas blooms throughout the country resulting to accuracy of computing the WQI is obligatory.
enormous amount of anthropogenic activities being intro- Most recently, artificial intelligence methods such as
duce to the environment which eventually effect the water artificial neural networks (ANNs) have been successfully
bodies. Nowadays, with the advanced science and tech- employed to solve many prediction problems related to
nology, the increase in human population, industries, engineering [6–13]. ANNs modeling have the ability to
agriculture activities, and urban development have causes deal with unknown knowledge in order to learn the com-
the riverbanks to widen. Nonetheless, unstoppable human plex model functions from examples, i.e., by training the
activities such as agricultural using pesticides, domestic input/output data sets. The greatest advantage that ANNs
sewage, factories effluents, and even soil erosion due to characterized over other modeling approaches is their
improper development have led to pollution in the water ability to model the complex pattern, nonlinear processes
bodies [2]. However, water pollution mainly occurs due to without any advance knowledge form of the relationship
the overloading of waste in the water system. The con- between input and output variables. The last two decades,
tamination of streams, lakes, underground water, bays, or applications of ANN have been reported in many studies of
oceans by substances can be harmful to living things, not water resources and engineering. ANN’s techniques have
only human, but the wildlife and plants. Pollutants that been utilized intensively in the field of water resource for
have been discharges into water body via point source and pattern matching, optimization [14], data compression,
nonpoint source are difficult to identify due to indefinite forecasting, and predictions [15–17]. Several researchers
origin of the pollutants. However, factors such as effluent have been used ANN in water resources application; for
from industrial area, sewage disposals, and land clearings instance, storm water prediction [18], waste water model-
greatly influence the water quality. In fact, the water pol- ing [19], heavy metal prediction [20], sediment transport
lution affects not only human health, but also the entire modeling [21], stream-flow forecasting [22, 23], water
environment, especially aquatic lives that live inside it level forecasting [24].
[3, 4]. The use of soft computing techniques has been suc-
For the time being, Malaysia is struggling to develop cessfully applied in modeling freshwater quality [25–29]
the infrastructure of the country to meet the vision of and in seawater [30–32]. Muttil and Chau [33] investigated
2020. Unluckily, the evolution that had been carried the past, present, and prospect of integrating AI into water
through the infrastructure contributes a bad impact to the modeling. The researchers indicated that the limitations of
environment, particularly in regard to water quality. water quality data and the high cost of water quality
Water quality index considered as the basic function for monitoring often pose a serious problem for process-based
the environmental assessment of surface water in rela- modeling methods. Hence, artificial intelligence gives the
tion to pollution load categorization and designation of optimal solution, as they are computationally very fast and
classes and beneficial uses as provided by Interim require many fewer input parameters and input conditions
National Water Quality Standards (INWQS) in Malaysia than deterministic models. The earliest research has been
[5]. Malaysia has 189 river basins nationwide; in par- conducted in water quality modeling by Lek and Guégan
ticular, Peninsular Malaysia is characterized by highly [34]. Zou et al. [35] proposed neural network embedded
diverse ecosystem and support extensive artisanal fish- Monte Carlo (NNMC) approach to model the uncertainty
eries. In addition, rivers are considered the main source in water quality applications. The results indicated that
of drinking water for the zone. Rainfall is the main NNMC approach has potential for efficient uncertainty
source of water for all kinds of surface water (i.e., analysis of water quality modeling. Kralisch et al. [36] used
channels, rivers, impoundment, and reservoirs). Those an ANN technique for the optimization of watershed
forms contribute 96 % of water supply sources. management to maintain a reasonable balance between
According to Department of Environment (DOE), out of water quality demand and consequent restrictions for the
the 189 rivers 120 rivers are controlled and monitored; farming industry. In 2004, Maier et al. [37] predicted the
around 44.5 % clean rivers, 48.4 % slightly polluted, and optimal alum doses and treated water quality variables.
7.1 % are polluted. Forty years ago, the DOE suggested Authors found that the proposed method is capable to
an adoption of WQI to assess and rank the level of water produce accurate prediction model. Diamantopoulou et al.
rivers pollution. Thereafter, the DOE approved an [38] used the application ANN to predict stream-flow water
approach called ‘‘Opinion Poll WQI (OP-WQI)’’ for quality parameters. The results show a satisfactory pre-
calculating the index of water quality of local rivers. The diction modeling of nitrates (NO3), specific electrical

123
Neural Comput & Applic

conductivity (EC), and dissolved oxygen parameters. WQI ¼ 0:22SLDO þ 0:19SLBOD þ 0:16SLCOD
Elhatip and Kömür [39] studied the water quality param- þ 0:16SLSS þ 0:15SLAN þ 0:12SLPH : ð1Þ
eters in reservoir. The results illustrated the capability of
ANN to predict the dissolved oxygen of recharge and Apparently, the DO carries the maximum weight value,
discharge of reservoir bank. Singh et al. [40] predicted the whereas pH carries the minimum weight value. Finally, the
biochemical oxygen demand (BOD) and dissolved oxygen evaluation of WQI formula consists of the subindexes,
(DO) in river banks using feed forward neural network which are calculated according to the best-fit relations
algorithm (FFNN). The researchers concluded that FFNN given in Table 1. The main advantage of using AI tech-
algorithm performed good results in predicting BOD and nique over manual calculation to assess WQI is that the
DO water quality variables in rivers. An assessment for the manual calculation consumes a lot of time and converting
effecting of physiochemical on the treatment performance the raw data to subpollutant index is needed.
of contaminated wetland sediment in replicate integrated
constructed wetland was conducted by Dong et al. [41]. An
evaluation for wetland surface water quality utilizing fuzzy 3 Materials and methods
neural network was accomplished by Wang et al. [42].
Total dissolved solid (TDS) prediction in river runoff using 3.1 Study area and data collection
two ANN algorithms, namely multi-layer perceptron
(MLP) and radial basis function (RBF), computed by Two rivers have been used in this research as a study area
Niroobakhsh [43]. Authors indicated that MLP and RBF Langat River and Klang River, Peninsular Malaysia. Langat
are able to simulate water quality variables in river with River has a total catchment area approximately 1815 km2.
more than 90 % accuracy. The total length of the river around 141 km, and it is located
The aim of this research is to construct two different 40 km east Kuala Lumpur. There are two reservoirs located
ANN algorithms, namely back propagation neural network on this river Langat Reservoir and the Semenyih Reservoir.
(BPNN) and radial basis function neural network Climatology, the annual temperature, humidity, and rainfall
(RBFNN), models to produce an efficient predictive model are considerably high. This fact has a dominant influence on
of WQI. In addition, to explore whether it is possible to the hydrological and geomorphological process of the river
predict WQI with rolling out the BOD variable as it environment. Lately, Langat River water quality has
requires a high controlled laboratory conditions and needs become considerably poor due to rapid urbanization along
time that makes BOD test could be costly, furthermore, to the river. The rapid growth of urban land use has been at the
establish a neural network model that applicable to predict expense of agricultural land and reclamation of peat and
WQI status of stream-flow river basin in tropical zone, and swampy areas. Whereas, Klang River has a total length
thus to provide an authoritative alternative to compute about 120 km with estimated area catchment around
WQI instead of manual calculation. In the following sec- 1290 km2. Klang River is a significant river in the west
tion, an introduction to the manual calculation for the WQI coast of Peninsular Malaysia, as it encompasses many cities
is provided. In Sect. 3, materials and methods are pre- in that area. Klang River basin is considered the most
sented. In Sect. 4, the application and discussion are densely populated area of the country. The watershed has
described. Finally, the conclusions from this research are experienced the highest economic growth in the country as
presented in Sect. 5. it included about 35 % of which is developed for residen-
tial, commercial, industrial, and institutional use. It is worth
mentioning that the discarding of unprocessed sewage from
2 Manual calculation of WQI for Malaysia treatment plants and livestock farms contributes to the
further deterioration of the Klang River’s water quality,
Thirty years ago, an opinion poll formula adopted to cal- adding to its solid waste and suspended-sediment loading.
culate WQI for Malaysia. The formula was proposed by the Geographically, Fig. 1 illustrates the two river location.
Department of Environment (DOE), Malaysia. A panel of The selected data for this research represent the monthly
professionals in water quality was consulted on the choice measurement of water quality variables and the evaluated
of the parameters and the weightage of each parameter. water quality index. The water quality variables are dis-
The six parameters that have been chosen by DOE are solved oxygen (DO), biochemical oxygen demand (BOD),
dissolved oxygen (DO), biochemical oxygen demand chemical oxygen demand (COD), ammoniacal nitrogen
(BOD), chemical oxygen demand (COD), suspended solid (NH3-N), suspended solid (SS), and PH used to evaluate
(SS), ammoniacal nitrogen (AN), and pH value (pH). The the WQI, for Malaysia. The data obtained from several
WQI approved by the DOE is calculated based on the monitoring stations located on Klang Basin and Langat
mentioned six parameters, the formula as follow: Basin and their branches that pass Selangor and Kuala

123
 Neural Comput & Applic

Table 1 Subindex calculation

Parameter Valuea Subindex equation
formulae for the local WQI [44]
DO (%saturation)b XB8 SLDO ¼ 0
8 \ x \ 92 SLDO ¼ 0:395 þ 0:03X 2  0:0002X 3
x C 92 SLDO ¼ 0
BOD XB5 SLBOD ¼ 100:4  4:2X
x[5 SLBOD ¼ ð108e0:055X Þ  0:1X
COD X B 20 SLCOD ¼ 99:1  1:33X
x [ 20 SLBOD ¼ ð103e0:0157X Þ  0:04X
NH3-N X B 0.3 SLAN ¼ 100:5  105X
0.3 \ x \ 4 SLAN ¼ ð94e0:573X Þ  5ðX  2Þ
xC4 SLAN ¼ 0
SS X B 100 SLSS ¼ ð97:5e0:00676X Þ  0:05X
100 \ x \ 1000 SLSS ¼ ð71e0:0016X Þ  0:015X
x C 1000 SLSS ¼ 0
PH X \ 5.5 SLPH ¼ 17:2  17:2X  6:67X 2
5.5 B x \ 7 SLPH ¼ 242 þ 95:5X  2:76X 2
7 B x \ 8.75 SLPH ¼ 181 þ 82:4X  6:05X 2
x C 8.75 SLPH ¼ 536 þ 77X  2:76X 2

For DO, X refers to DO percentage saturation and for pH it refers to the pH value
a
X is the concentration of the indicated parameter in mg/L, except for pH and DO
b
DO (%saturation) = [DO (mg/L) * 12.795] - 0.05

Lumpur states. The data set consisted of 5233 samples for weight to shift the activation function. xj and y1 indicate the
the time period from January 2001 to October 2010. These input/out variables, respectively. They can be expressed by
data divided into two set: The first set used for training the the following equations:
model that involved the data for the time period from 2001 " ! #
X
K X J
to 2008 and the rest of data used for testing the model. y1 ¼ f 1 wlk f 2 wkj xj þ bk þ b1 ð2Þ
K¼1 J¼1
3.2 Artificial neural network modeling 2
f 2ðpÞ ¼ 1 ð3Þ
ð1 þ e2p Þ
3.2.1 Back propagation neural network (BPNN)
where f1(.) and f2(.) represent the linear and tansigmoidal
ANN is inspired by simulating the function of a human brain, activation functions of the algorithm. In general, the choice
owing to its ability to model high complex relationships or to of the transfer function in the hidden layer and output layer is
find partners in data; ANN can be widely used to represent a very significant task because it can effect essentially on the
nonlinear mapping between input and output vectors. BPNN target of the model. Bayesian Regulation algorithm has been
is the most broadly neural network algorithm has been employed to train the ANN modeling because it updates the
employed for prediction of water resources and management weight and bias values according to Levenberg–Marquardt
problems [45]. Figure 2 illustrates the typical architecture of optimization and save time, in addition, its ability to reduce
BPNN: input layer, hidden layer, and output layer. The input the combination of mean-squared errors that caused by using
layer indicates the correspondence water quality variables huge number of data.
that connected to the hidden layer. This connection repre-
sented by specific weights to determine the relationship with 3.2.2 Radial basis function (RBF)
is donated by Wkj. The main duty of the hidden layer nodes is
to solve the complexity of the problem being modeled. The ANNs are flexible processing approaches that have the
hidden layer nodes comprise the activation function that efficiency to mimic the relationship between the inputs and
handles nonlinearly transforming of the inputs parameters output data sets, in addition to pattern recognition capability.
into an alternative space where the training samples are lin- Radial basis function algorithm is the remarkable efficient
early separable [46]. The connection between the hidden layer approach used in various engineering applications
and the output layer (WQI targeted) exhibited by Wlk. bk and [21–23, 47–49]. Initially, RBF was introduced by Lowe and
b1 factors denote the bias of the corresponding hidden/output Broomhead [50]. The architecture of RBF neural network is
layer neurons, respectively. They imply a percentage of similar to the BPNN structure, except it has a radial basis

123
Neural Comput & Applic

Fig. 1 Location of the two rivers, Langat and Klang River Basins, Malaysia

activation function in the hidden layer that to characterize xl2

uðx; lÞ ¼ e 2d2 ð4Þ
the partitions of the input space. Here, RBF implies an
approximation between the input/output phases with linear where l is the center of the Gaussian function (mean value
combination of the radial basis function. The main advan- of x) and d is the distance (radius) from the center of u(x,
tage of RBF is response by (decreasing or increasing) l), which gives a measure of the spread of the Gaussian
monotonically with distance from a central point. Figure 3 curve. During the training processes of RBF, the center l
illustrates the architecture of the RBF neural network algo- and the spread d are determined. Figure 4 reveals the
rithm. What is necessary to highlight here, the middle layer Gaussian radial basis function manners, with large value of
connects by linear function with output layer. In addition, the d produce less sensitivity of the used function to the sup-
radial basis functions u1, u2,… uN are known as hidden plied data in the input phase. Radial basis functions number
functions while [{ui (x)}Ni = 1] is called the hidden space. inside the hidden layer subjected to the complexity of the
Gaussian function used in this study as a radial basis func- problem nor on the size of the data set, where it can appear
tion, and the following formula can use to express the one- in the utilization of multi-layer perceptron algorithm
dimensional representation of Gaussian function: [51–53].

123
 Neural Comput & Applic

Fig. 2 Three layer of feed

forward artificial neural network
back propagation configuration

Fig. 3 Structure of radial basis

function algorithm

3.2.3 Model configuration development NH3-N, SS, and pH) used as a predictors for WQI, as
clearly shown in Figs. 2 and 4. At the initial stage, input of
In this research, water quality index has been predicted for the experiment, input variables, and output variable are
tropical environment, Langat River and Klang River, normalized linearly in the range of 0.1 and 0.9. The nor-
Malaysia. Water quality parameters (i.e., DO, BOD, COD, malization is done according to the following formula:

123
Neural Comput & Applic

Fig. 4 Radial bases function with different stages of spread value. a Normal spread, b small spread, c large spread

X  Xmin WQIa  WQIp

Xnew ¼ 0:8 þ 0:1 ð5Þ %RE ¼  100 ð9Þ
Xmax  Xmin WQIa

where Xnew is the normalized value of an original param- where WQIa is the observed value of water quality index,
eter, X is the original data point, Xmin and Xmax are the WQIp is the predicted value of water quality index and
minimum and maximum values in the data set, respec- N number of samples. Furthermore, sensitivity data
tively. The normalization is done to ensure that the mini- analysis has been investigated. Sensitivity analysis indi-
mum value in the data set is normalized to 0.1, and the cator is a method to extract the cause and effect rela-
maximum value is normalized to 0.9. This pattern of nor- tionship between input/output phases. The principal
malizing the data is chosen because it tends to provide a concept of this approach is that the inputs to the network
better outcome on the water quality application. In addi- are shifted slightly and the corresponding change in the
tion, the variables are usually measured in different units, output is reported either as a percentage or a raw
by normalizing the variables and recasting them in difference.
dimensionless units, the arbitrary effect of similarity
between objects is removed [54].
4 Application and discussion
3.2.4 Model performance criteria
The current research employed two different ANN algo-
In order to evaluate the prediction modeling accuracy, four rithms BPNN and RBFNN. Initially, the number of neurons
criteria were used for comparative evaluation of the per- for the input layer fixed six (number of the water quality
formance of the modeling. A multi-criteria approach was parameters; DO, BOD, COD, NH3-N, SS, and pH, for the
adopted for assessing the models developed, in which first scenario) and five neurons for the second scenario with
model performance was evaluated using several statistical rolling out the BOD parameter. On the other hand, one
error and goodness-of-fit measures, including the root- neuron in the output layer presents the WQI. The monthly
mean-squared error (RMSE), coefficient of determination data of water quality parameters (2001–2010) were nor-
(R2), Nash–Sutcliffe coefficient (NE), and relative error malized within the range of 0.1–0.9 in order to accelerate
(RE%). The performances criteria formulas are presented the training procedure and achieve minimum (MSE). Then,
as: the optimal data division scheme was conducted using
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi NeuroSolutions 7 software, 80 % for training the network
1 Xn
RMSE ¼ ðWQIa  WQIP Þ2 ð6Þ and 20 % for testing the network, based on trial-and-error
N t¼1
procedure. However, one of the most critical features of
Pn    
2 i¼1 WQIa  WQIa  WQIp  WQIp BPNN approach is determining the number of neurons in
R ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn  2 Pn  2 the hidden layer. An insufficient number of neurons are
i¼1 WQIa  WQIa i¼1 WQIp  WQIp utilized, and the model will be unable to capture the
ð7Þ complexity of the relationship between input and output
Pn  2 phases. Thus, in this research, a trial-and-error procedure
t¼1 WQIa  WQIp used to determine the optimal hidden layer neurons. The
NE ¼ 1  Pn  2 ð8Þ
t¼1 WQIa  WQIa
trial-and-error procedure started with two hidden neurons

123
 Neural Comput & Applic

initially, and the number of hidden neurons was increased performed error distribution between -20 and ?20, while,
up to 10 with a step size of 2 in each trial. The optimal there are more than 85 % of the testing data set limited the
architecture was determined based on three performance distribution error of the RE between -8 and ?7. This is
indices R2, RMSE, and NE (see Table 2). In the RBFNN return back to the Gaussian radial basis function, which is
approach, different spread values were utilized to achieve capable to mimic the pattern phenomena of the water
the best performance indices. quality parameters that influenced the WQI.
In the systematic analysis, the influence of each water
4.1 First scenario: six parameters water quality quality parameter was investigated using the sensitivity
modeling the WQI analysis procedure. Figure 9 presented the sensitivity anal-
ysis for the best ANN algorithm, which is RBFNN modeling.
In this scenario, prediction model was built based on the According to Fig. 9a, it can be observed that the DO is the
six water quality variables (i.e., DO, BOD, COD, NH3-N, most effective parameter correlated with the predicted WQI,
SS, and pH). The best performances evaluation criteria for around 28 %. The observation based on this percentage
the first scenario using BPNN algorithm have been pre- reveals that the sensitivity analysis of this parameter was
sented in Table 2. According to Table 2, it was determined harmonized with manual calculated WQI. Furthermore, the
that six neurons in the input layer (number of water quality nature of Langat and Klang rivers is characterized by pol-
parameters), eight neurons in the hidden layer, and one luted environment, which indicated that the dissolved oxy-
neuron in the output layer (WQI as a target) performed the gen is the most effected parameter in the aquatic system of
best R2, RMSE, and NE (0.7472, 0.0699 and 0.7701, the rivers. The followed parameters percentages are BOD
respectively), whereas according to Table 3, the first sce- 21 %, COD 18 %, SS 15 %, NH3-N 9 %, and pH 9 %.
nario using RBFNN model showed the best performances Accordingly, the results modeling of sensitivity analysis
evaluation criteria of R2, RMSE, and NE (0.9872, 0.0157, briefed the significant of each variable to WQI.
and 0.9871, respectively) with spread value 0.8. In order to
validate the functionality of the ANN models, Figs. 5 and 6
indicate the scatterplots between the observed and the 4.2 Second scenario: five parameters water quality
predicted value of WQI for the testing data set modeling the WQI
(2009–2010), for the both scenarios. The BPNN model
performed poorly comparing with RBFNN model (see The main objective of this scenario is to reduce the water
Figs. 5, 6). This observation is clearly explaining the quality parameters that needed to predict WQI without
capability of the radial basis function neural network in much loss of information; this is for the reason that BOD
capturing the nonlinearity between the water quality pre- parameter requires a high controlled laboratory conditions,
dictors and the predicted (WQI). needs time and costly tested parameter. Therefore, the five
To have more assessment of the performances of the parameters (DO, COD, NH3-N, SS, and pH) used as pre-
both models, the RE computed for the both models. The dictors for the WQI. The accuracy of the models examined
(%RE) distribution as computed by the formula 9 is illus- using the formal quantitative measures of accuracy
trated in Figs. 7 and 8 for BPNN and RBFNN models, including R2, RMSE, and NE. Table 2 presents the per-
respectively. As it can be seen, BPNN model did not per- formance criteria for the second scenario using BPNN
form very well for the first scenario (see Fig. 7a). The model. The BPNN model performed the best results of R2,
percentage of the error residual exceeds the 45 % for more RMSE, and NE (0.7007, 0.0867, and 0.7164, respectively)
than 14 % of the testing data set and limited between -35 with five neurons input layer, six neurons hidden layer, and
and ?15 for more than 65 % of testing data set. The one neuron in the output layer, while RBFNN model per-
RBFNN model performed very well for the first scenario formance criteria for the second scenario is tabulated in
(see Fig. 7b). Less than 5 % of the testing data set Table 3. The best evaluation criteria of R2, RMSE, and NE

Table 2 BPNN model

Models First scenario Second scenario
performances assessment for
2
both scenarios Architecture R RMSE NE Architecture R2 RMSE NE

M1 6-2-1 0.6853 0.0983 0.7156 5-2-1 0.7001 0.0981 0.7151

M2 6-4-1 0.7409 0.0734 0.7631 5-4-1 0.6976 0.0928 0.7149
M3 6-6-1 0.7309 0.0801 0.7592 5-6-1 0.7007 0.0867 0.7164
M4 6-8-1 0.7472 0.0699 0.7701 5-8-1 0.6593 0.1018 0.6828
M5 6-10-1 0.7267 0.0783 0.7434 5-10-1 0.5109 0.1314 0.5372

123
Neural Comput & Applic

Table 3 RBFNN model

Models First scenario Second scenario
performances assessment for
2
both scenarios Spread values R RMSE NE Spread values R2 RMSE NE

M1 0.2 0.9552 0.0294 0.9547 0.2 0.9228 0.0405 0.9140

M2 0.4 0.9242 0.0399 0.9165 0.4 0.9228 0.0405 0.9140
M3 0.6 0.9852 0.0168 0.9852 0.6 0.9807 0.0194 0.9803
M4 0.8 0.9872 0.0157 0.9871 0.8 0.9705 0.0247 0.9678
M5 1.0 0.9820 0.0186 0.9819 1.0 0.8999 0.0452 0.8926

Fig. 5 First scenario scatterplot (a) 120

between the observed and the y = 0.7032x + 19.599
(b) 120
y = 0.9917x + 0.5061
predicted value of WQI. 100 R² = 0.7472 R² = 0.9872
100
PREDICTED WQI

PREDICTED WQI
a BPNN model using six
variables of water quality, 80 80
b RBFNN model using six
60 60
variables of water quality
40 40

20 20

0 0
0 50 100 150 0 50 100 150
OBSERVED WQI OBSERVED WQI

Fig. 6 Second scenario (b)

(a) 120 120
scatterplot between the y = 0.6449x + 24.292 y = 0.9826x + 0.875
observed and the predicted 100 R² = 0.7007 100 R² = 0.9807
PREDICTED WQI

PREDICTED WQI
value of WQI. a BPNN model
using five variables of water 80 80
quality, b RBFNN model using 60 60
five variables of water quality
40 40

20 20

0 0
0 50 100 150 0 50 100 150
OBSERVED WQI OBSERVED WQI

are 0.9807, 0.0194, and 0.9803, respectively, with spread to make the water quality prediction without losses much
value 0.6. information and still have an acceptable performance.
Figure 6a shows the correlation of coefficient (R2), To further analyze the effectiveness of the modeling, the
which provides information for linear dependence between RE is demonstrated in Fig. 8a, d (BPNN and RBFNN
observed and predicted WQI using BPNN model. Similar models, respectively). The divergence between the actual
to the first scenario, the scatter plot performed less accu- and the predicted testing data exceeds (?65 and -65 %)
racy comparatively with RBFNN model. However, Fig. 6b 10 % of the data, while the RE reaches a level of -40 %
indicates that RBFNN model falls close to the ideal line for for more than 38 % of the testing data. On the other hand,
the scatterplot. Eventually, the result indicated that the RBFNN model accomplished the RE percentage with very
RBFNN model is more reliable and suitable in prediction well divergence between ?5 and -5 % for more than
WQI than BPNN model because it could capture the pat- 87 % of the testing data set. Finally, for a better under-
tern behavior between the predictors and the predictand. standing of the above observations, further analysis was
Furthermore, it is obvious that RBFNN algorithm has accomplished to study the sensitivity analysis of each
predicted WQI accurately and with a satisfactory accept- variable. The sensitivity analysis is the influence of each
able performance even without using BOD parameter. variable has been investigated. Indeed, this indicator
Nevertheless, this study clearly shows that it is possible to showed up to what extend each input variables correlated
reduce the number of the environmental parameters needed with the predicted (target output). The percentage of the

123
 Neural Comput & Applic

(a) 45

25

Relative error % 5

-15 1 101 201 301 401 501 601 701

-35

-55
2 years teasting phase

(b) 30

20
Relative error %

10

0
1 101 201 301 401 501 601 701
-10

-20

-30
2 years teasting phase

Fig. 7 First scenario relative error distribution of the water quality index modeling, WQI. a BPNN algorithm using six variables of water
quality, b RBFNN algorithm using six variables of water quality

(a) 80
60
40
Relative error %

20
0
-20 1 101 201 301 401 501 601 701

-40
-60
-80
2 years teasting phase
(b) 20
15
10
Relative error %

5
0
-5 1 101 201 301 401 501 601 701
-10
-15
-20
2 years teasting phase

Fig. 8 Second scenario relative error distribution of the water quality index modeling, WQI. a BPNN algorithm using five variables of water
quality, b RBFNN algorithm using five variables of water quality

association of the BOD almost divided equally to the other predicting water quality index. Based on this observation,
water quality parameters. DO influenced the WQI predic- the integration of soft computing with modeling water
tion by 35 %, COD 23 %, SS 17 %, NH3-N 13 %, and pH quality is a very useful application that could enhance the
12 % (see Fig. 9b). environmental engineering field in term of online imple-
In the light of the application and analysis, machine mentation monitoring and accurate prediction. However,
learning approaches produce a versatile modeling in better and more proficient use of the database

123
Neural Comput & Applic

(a) 1. Generally, ANN employed effectively in predicting

Water Quality Index
water quality index in tropical environment, Langat
pH, 8.81, 9%
DO, 27.85325, and Klang river basins, Malaysia.
NH3-N, 8.69, 28% 2. The current research emphasizes that ANN modeling,
9%
in particular RBFNN model constitutes an effective
tool for assessment of river water quality that simpli-
fies the computing of WQI.
SS, 14.6854,
15% 3. RBFNN performed very well in recognizing the most
associated input variables on prediction WQI, for the
both scenarios.
4. It can be seen from the summary of the results that
BOD, 21.536,
COD, 18.48285, there was a significant improvement in the modeling
21%
18%
performance when the RBFNN model was used in the
DO BOD COD SS NH3-N pH both scenarios, compared to using the BPNN model.
This is due to the fact that Gaussian radial basis
(b) Water Quality Index function, which is capable to mimic the pattern
pH, 11.7325, phenomena chaotic disturbances, complex nonlinear
12%
DO, 34.96584, dynamics, and randomness of the water quality
35%
parameters that influenced the WQI.
NH3-N,
12.7463, 13%

References
SS, 17.6594,
17% COD,
1. Kotti ME, Vlessidis AG, Thanasoulias NC, Evmiridis NP (2005)
22.8625485, Assessment of river water quality in Northwestern Greece. Water
23% Resour Manag 19:77–94. doi:10.1007/s11269-005-0294-z
DO COD SS NH3-N pH 2. Niemi GJ, DeVore P, Detenbeck N et al (1990) Overview of case
studies on recovery of aquatic systems from disturbance. Environ
Fig. 9 Sensitivity analysis percentage of each water quality variables Manage 14:571–587. doi:10.1007/BF02394710
on WQI prediction. a RBFNN modeling for the first scenario, 3. Najah A, Elshafie A, Karim OA, Jaffar O (2009) Prediction of
b RBFNN modeling for the second scenario Johor River water quality parameters using artificial neural net-
works. Eur J Sci Res 28:422–435
4. Fahmi M, Nasir M, Samsudin MS et al (2011) River water quality
modeling using combined principle component analysis (PCA)
and multiple linear regressions (MLR): a case study at Klang
administration frameworks, graphical presentations, and River, Malaysia Department of Environmental Sciences, Faculty
learning obtaining modules will upgrade the applicability of Environmental Studies, Department of Environment. World
of modeling systems in real practice. Appl Sci J 14:73–82
5. Zali MA, Retnam A, Juahir H et al (2011) Sensitivity analysis for
water quality index (WQI) prediction for Kinta River, Malaysia.
World Appl Sci J 14:60–65
5 Conclusion 6. Behboudian S, Tabesh M, Falahnezhad M, Ghavanini FA (2014)
A long-term prediction of domestic water demand using pre-
This manuscript described the implementation of ANN to processing in artificial neural network. J Water Supply Res
Technol 63:31. doi:10.2166/aqua.2013.085
predict water quality index in tropical environment based 7. Caselli M, Trizio L, Gennaro G, Ielpo P (2009) A simple feed-
on monthly measurement of water quality variables. The forward neural network for the PM10 forecasting: comparison
monthly data (i.e., DO, BOD, COD, NH3-N, SS, and pH) with a radial basis function network and a multivariate linear
for 10 years’ time period (2001–2010) were selected for regression model. Water Air Soil Pollut 201:365–377. doi:10.
1007/s11270-008-9950-2
this analysis. Two different models have been examined, 8. Albergaria JT, Martins FG, Alvim-Ferraz MCM, Delerue-Matos
namely feed forward back propagation neural network C (2014) Multiple linear regression and artificial neural networks
(BPNN) and radial basis function neural network to predict time and efficiency of soil vapor extraction. Water Air
(RBFNN). Results obtained from this research are posi- Soil Pollut. doi:10.1007/s11270-014-2058-y
9. Modarres R (2008) Multi-criteria validation of artificial neural
tively encouraging with high performances accuracy. network rainfall-runoff modeling. Hydrol Earth Syst Sci Discuss
Nevertheless, the following conclusions can be drawn. 5:3449–3477. doi:10.5194/hessd-5-3449-2008

123
 Neural Comput & Applic

10. Shrestha RR, Theobald S, Nestmann F (2005) Simulation of flood networks and canonical correlation analysis. J Hydrol
flow in a river system using artificial neural networks. Hydrol 405:277–287. doi:10.1016/j.jhydrol.2011.05.024
Earth Syst Sci 9:313–321. doi:10.5194/hess-9-313-2005 27. Najah A, Karim OA, Jaafar O, El-shafie AH (2011) An appli-
11. Mwale FD, Adeloye AJ, Rustum R (2014) Application of self- cation of different artificial intelligences techniques for water
organising maps and multi-layer perceptron-artificial neural net- quality prediction. Int J Phys Sci 6:5298–5308. doi:10.5897/
works for streamflow and water level forecasting in data-poor IJPS11.1180
catchments: the case of the Lower Shire floodplain, Malawi. 28. Amiri BJ, Nakane K (2009) Comparative prediction of stream
Hydrol Res 45:838. doi:10.2166/nh.2014.168 water total nitrogen from land cover using artificial neural net-
12. Cheng J, Li QS (2008) Reliability analysis of structures using work and multiple linear regression approaches. Pol J Environ
artificial neural network based genetic algorithms. Comput Stud 18(2):151–160
Methods Appl Mech Eng 197:3742–3750. doi:10.1016/j.cma. 29. Chebud Y, Naja GM, Rivero RG, Melesse AM (2012) Water
2008.02.026 quality monitoring using remote sensing and an artificial neural
13. Ahmad A, El-Shafie A, Mohd Razali SF, Mohamad ZS (2014) network. Water Air Soil Pollut 223:4875–4887. doi:10.1007/
Reservoir optimization in water resources: a review. Water s11270-012-1243-0
Resour Manag 28:3391–3405. doi:10.1007/s11269-014-0700-5 30. Lee JHW, Huang Y, Dickman M, Jayawardena AW (2003)
14. Hossain MS, El-shafie A (2013) Intelligent systems in optimizing Neural network modelling of coastal algal blooms. Ecol Model
reservoir operation policy: a review. Water Resour Manag 159:179–201. doi:10.1016/s0304-3800(02)00281-8
27:3387–3407. doi:10.1007/s11269-013-0353-9 31. Palani S, Liong SY, Tkalich P (2008) An ANN application for
15. Chau KW (2006) A review on integration of artificial intelligence water quality forecasting. Mar Pollut Bull 56:1586–1597. doi:10.
into water quality modelling. Mar Pollut Bull 52:726–733. 1016/j.marpolbul.2008.05.021
doi:10.1016/j.marpolbul.2006.04.003 32. Palani S, Liong S, Tkalich P, Palanichamy J (2009) Development
16. Rahim NA, Ahmad Z (2013) Features selection in water quality of a neural network model for dissolved oxygen in seawater.
prediction in neural network using canonical correspondence Indian J Geo-Mar Sci 38:151–159
analysis (CCA). The 6th international conference on process 33. Muttil N, Chau K-W (2006) Neural network and genetic pro-
systems engineering (PSE ASIA), pp 25–27 gramming for modelling coastal algal blooms. Int J Environ
17. Hipni A, El-shafie A, Najah A et al (2013) Daily forecasting of Pollut 28:223–238. doi:10.1504/IJEP.2006.011208
dam water levels: comparing a support vector machine (SVM) 34. Lek S, Guégan JF (1999) Artificial neural networks as a tool in
model with adaptive neuro fuzzy inference system (ANFIS). ecological modelling, an introduction. Ecol Model 120:65–73.
Water Resour Manag 27:3803–3823. doi:10.1007/s11269-013- doi:10.1016/S0304-3800(99)00092-7
0382-4 35. Zou R, Lung W-S, Guo H (2002) Neural network embedded
18. Kazemi Yazdi S, Scholz M (2010) Assessing storm water Monte Carlo approach for water quality modeling under input
detention systems treating road runoff with an artificial neural information uncertainty. J Comput Civ Eng 16:135–142. doi:10.
network predicting fecal indicator organisms. Water Air Soil 1061/(ASCE)0887-3801(2002)16:2(135)
Pollut 206:35–47. doi:10.1007/s11270-009-0084-y 36. Kralisch S, Fink M, Flügel W-A, Beckstein C (2003) A neural
19. Ye J, Zhang P, Hoffmann E et al (2014) Comparison of response network approach for the optimisation of watershed management.
surface methodology and artificial neural network in optimization Environ Model Softw 18:815–823. doi:10.1016/S1364-
and prediction of acid activation of bauxsol for phosphorus 8152(03)00081-1
adsorption. Water Air Soil Pollut. doi:10.1007/s11270-014-2225- 37. Maier HR, Morgan N, Chow CWK (2004) Use of artificial neural
1 networks for predicting optimal alum doses and treated water
20. Lesven L, Lourino-Cabana B, Billon G et al (2009) Water-quality quality parameters. Environ Model Softw 19:485–494. doi:10.
diagnosis and metal distribution in a strongly polluted zone of 1016/S1364-8152(03)00163-4
Deûle River (Northern France). Water Air Soil Pollut 198:31–44. 38. Diamantopoulou MJ, Papamichail DM, Antonopoulos VZ (2005)
doi:10.1007/s11270-008-9823-8 The use of a neural network technique for the prediction of water
21. Afan HA, El-Shafie A, Yaseen ZM et al (2014) ANN based quality parameters. Oper Res 5:115–125. doi:10.1007/BF02944165
sediment prediction model utilizing different input scenarios. 39. Elhatip H, Kömür MA (2008) Evaluation of water quality
Water Resour Manag 29:1231–1245. doi:10.1007/s11269-014- parameters for the Mamasin Dam in Aksaray City in the central
0870-1 Anatolian part of Turkey by means of artificial neural networks.
22. Yaseen ZM, El-Shafie A, Afan HA et al (2015) RBFNN versus Environ Geol 53:1157–1164. doi:10.1007/s00254-007-0705-y
FFNN for daily river flow forecasting at Johor River. Neural 40. Singh KP, Basant A, Malik A, Jain G (2009) Artificial neural
Comput Appl, Malaysia. doi:10.1007/s00521-015-1952-6 network modeling of the river water quality—a case study. Ecol
23. El-Shafie A, Abdin AE, Noureldin A, Taha MR (2009) Enhanc- Model 220:888–895. doi:10.1016/j.ecolmodel.2009.01.004
ing inflow forecasting model at Aswan high dam utilizing radial 41. Dong Y, Scholz M, Harrington R (2012) statistical modeling of
basis neural network and upstream monitoring stations mea- contaminants removal in mature integrated constructed wetland
surements. Water Resour Manag 23:2289–2315. doi:10.1007/ sediments. J Environ Eng 138:1009–1017. doi:10.1061/
s11269-008-9382-1 (ASCE)EE.1943-7870.0000572
24. Sulaiman M, El-Shafie A, Karim O, Basri H (2011) Improved 42. Wang L, Li X, Cui W (2012) Fuzzy neural networks enhanced
water level forecasting performance by using optimal steepness evaluation of wetland surface water quality. Int J Comput Appl
coefficients in an artificial neural network. Water Resour Manag Technol 44:235. doi:10.1504/IJCAT.2012.049087
25:2525–2541 43. Niroobakhsh M (2012) Prediction of water quality parameter in
25. Chen Q, Mynett AE (2003) Integration of data mining techniques Jajrood River basin: application of multi layer perceptron (MLP)
and heuristic knowledge in fuzzy logic modelling of eutrophi- perceptron and radial basis function networks of artificial neural
cation in Taihu Lake. Ecol Model 162:55–67. doi:10.1016/ networks (ANNs). Afr J Agric Res 7:4131–4139. doi:10.5897/
S0304-3800(02)00389-7 AJAR11.1645
26. Khalil B, Ouarda TBMJ, St-Hilaire A (2011) Estimation of water 44. Department of Environment (2005) Malaysia environmental
quality characteristics at ungauged sites using artificial neural quality report. Petaling Jaya, Malaysia, 2007

123
Neural Comput & Applic

45. Maier HR, Dandy GC (2000) Neural networks for the prediction Aerosp Sci Technol 5:125–134. doi:10.1016/S1270-
and forecasting of water resources variables: a review of mod- 9638(00)01087-7
elling issues and applications. Environ Model Softw 15:101–124. 50. Lowe D, Broomhead D (1988) Multivariable functional interpo-
doi:10.1016/S1364-8152(99)00007-9 lation and adaptive networks. Complex Syst 2:321–355
46. Brown M, Harris CJ (1995) A perspective and critique of adap- 51. Bishop CM (1995) Neural networks for pattern recognition. J Am
tive neurofuzzy systems used for modelling and control appli- Stat Assoc. doi:10.2307/2965437
cations. Int J Neural Syst 6:197–220 52. Ripley BD (1996) Pattern recognition and neural networks.
47. Lin G-F, Chen L-H (2004) A non-linear rainfall-runoff model Cambridge University Press, Cambridge
using radial basis function network. J Hydrol 289:1–8. doi:10. 53. Haykin S (1994) Neural networks: a comprehensive foundation.
1016/j.jhydrol.2003.10.015 Macmillan, Englewood Clifs, NJ
48. Ma L, Xin K, Liu S (2008) Using radial basis function neural 54. El-shafie A, Mukhlisin M, Najah AA, Taha MR (2011) Perfor-
networks to calibrate water quality model. World Acad Sci Eng mance of artificial neural network and regression techniques for
Technol Int J Environ Chem Ecol Geol Geophys Eng 2(2):9–17 rainfall-runoff prediction. Int J Phys Sci 6:1997–2003. doi:10.
49. Beckert A, Wendland H (2001) Multivariate interpolation for 5897/IJPS11.314
fluid–structure-interaction problems using radial basis functions.

123