See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260724302

Groundwater-level prediction using multiple linear regression and artificial

neural network techniques: A comparative assessment

Article in Hydrogeology Journal · October 2013

DOI: 10.1007/s10040-013-1029-5

CITATIONS READS

170 2,286

2 authors, including:

Madan Kumar Jha

Indian Institute of Technology Kharagpur
337 PUBLICATIONS 10,024 CITATIONS

SEE PROFILE

All content following this page was uploaded by Sasmita Sahoo on 28 April 2015.

The user has requested enhancement of the downloaded file.

Groundwater-level prediction using multiple linear regression
and artificial neural network techniques: a comparative assessment

Sasmita Sahoo & Madan K. Jha

Abstract The potential of multiple linear regression (MLR) Introduction

and artificial neural network (ANN) techniques in predicting
transient water levels over a groundwater basin were The accurate prediction of groundwater levels is essential for
compared. MLR and ANN modeling was carried out at 17 sustainable utilization and management of vital groundwater
sites in Japan, considering all significant inputs: rainfall, resources. Since groundwater is hidden and groundwater
ambient temperature, river stage, 11 seasonal dummy processes exhibit a high degree of temporal and spatial
variables, and influential lags of rainfall, ambient tempera- variability, modeling groundwater fluctuations is a very
ture, river stage and groundwater level. Seventeen site- difficult task. There exist conceptual and process-based
specific ANN models were developed, using multi-layer modeling techniques for simulating groundwater flow in a
feed-forward neural networks trained with Levenberg- variety of hydrogeological settings. However, the data
Marquardt backpropagation algorithms. The performance requirements for the process-based models used to simulate
of the models was evaluated using statistical and graphical groundwater fluctuations are enormous, and generally
indicators. Comparison of the goodness-of-fit statistics of the difficult or expensive to obtain (Coulibaly et al. 2001;
MLR models with those of the ANN models indicated that Nikolos et al. 2008). Despite large investments in time and
there is better agreement between the ANN-predicted resources, the prediction accuracy possible with distributed
groundwater levels and the observed groundwater levels at numerical flow models has not improved satisfactorily for
all the sites, compared to the MLR. This finding was many types of water-management problems (Coppola et al.
supported by the graphical indicators and the residual 2005). Therefore, a dynamically predictive model that can
analysis. Thus, it is concluded that the ANN technique is tackle the persistent trend and time-variant behavior of
superior to the MLR technique in predicting spatio-temporal hydrological variables is desirable for the efficient planning
distribution of groundwater levels in a basin. However, and management of water resources. Under such circum-
considering the practical advantages of the MLR technique, stances, empirical models such as regression models and
it is recommended as an alternative and cost-effective artificial neural network (ANN) models serve as attractive
groundwater modeling tool. alternatives, because they can provide useful results using
relatively fewer data, and are less laborious and therefore
Keywords Groundwater-level prediction . Multiple linear cost-effective. Despite the inability of multiple linear regres-
regression . Artificial neural network . Statistical sion (MLR) models to cope with the non-linearity existing
modeling . Japan between model inputs and outputs, they have been used in
many hydrological studies, possibly due to the fact that the
results are quite easy to use and the interpretation of the
relationship between the parameters is easier (Heuvelmans et
Received: 14 September 2012 / Accepted: 22 July 2013 al. 2006; Adeloye 2009). On the other hand, the ANN
Published online: 4 October 2013
technique has been found to be very much suited to the
* Springer-Verlag Berlin Heidelberg 2013 modeling of non-linear and dynamic systems such as water-
resources systems (ASCE 2000b; Maier and Dandy 2000).
S. Sahoo ()) : M. K. Jha
The main advantage of the ANN technique over traditional
AgFE Department, Indian Institute of Technology Kharagpur, methods is that it does not require the complex nature of
Kharagpur – 721 302, West Bengal, India underlying processes to be explicitly described in mathemat-
e-mail: [email protected] ical form. After proper training, ANN models can yield
Tel.: +91-9749935147 satisfactory results for many prediction problems in the field
Fax: +91-3222-282244
of hydrology (ASCE 2000a, b).
M. K. Jha In the last decade or so, tools such as ANN and statistical
e–mail: [email protected] techniques such as MLR have attracted the attention of some

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1866
hydrologists and hydrogeologists for the purposes of predic- The inputs to the ANN model consisted of daily tempe-
tion/forecasting, due to their parsimony in data requirement, rature, rainfall, pumping rate and hydraulic head. A
simplicity and cost-effectiveness. Although several studies are differential-evolution algorithm was used to optimally
reported in the literature that use MLR as a modeling define the time lag in the rainfall measurements, as well
technique in the field of surface-water hydrology (e.g., as the ANN architecture and training parameters. Sethi et al.
McCuen et al. 1979; Sinnakaudan et al. 2006), the use of (2010) developed multi-layer feed-forward networks, trained
the MLR technique in groundwater modeling is very limited. with back-propagation algorithms using gradient-descent with
Hodgson (1978) used MLR for the simulation of groundwa- momentum, to predict groundwater depths in 64 dug wells
ter-level responses in the Vryburg aquifer of South Africa by located in the Munijhara micro-watershed in Orissa, eastern
considering rainfall and pumping as input parameters. India. Data on monthly rainfall, potential evapotranspiration,
A comprehensive review of the application of ANN to and groundwater depth (from nearby wells) were used as inputs
hydrology can be found in the ASCE Task Committee reports to the model, with the output being groundwater depths,
(ASCE 2000a, b) and in Maier and Dandy (2000). In the 1 month ahead. Similarly, several researchers have used ANN
recent past, several researchers have successfully used ANN for predicting groundwater levels in confined aquifers (Hani et
for the prediction of groundwater levels in unconfined al. 2006; Uddameri 2007), in leaky aquifers (Coppola et al.
aquifers (e.g., Coulibaly et al. 2001; Lallahem et al. 2005; 2003, 2005; Mohanty et al. 2009), and in multi-layer aquifer
Daliakopoulos et al. 2005; Nayak et al. 2006; Affandi et al. systems (Coppola et al. 2003; Feng et al. 2008; Nourani et al.
2007; Krishna et al. 2008; Trichakis et al. 2009; Sethi et al. 2008).
2010). Coulibaly et al. (2001) developed four types of ANN As far as the combined application of statistical and ANN
models, namely: input-delay neural network (IDNN), recur- techniques to the simulation of groundwater levels is
rent neural network (RNN), generalized radial-basis function concerned, only two studies are reported to date (Uddameri
(RBF) network, and probabilistic neural network (PNN) to 2007; Izady et al. 2012). Uddameri (2007) applied MLR and
simulate groundwater-level fluctuations in four observation ANN techniques to forecast monthly and quarterly time-series
wells in the Gondo aquifer, Burkina Faso, Africa. Monthly of piezometric levels in one deep well in South Texas, by using
water-level depth, precipitation, temperature and river-water bias, time and dummy variables as inputs to the models. In
level were used as inputs to the networks, and it was found this study, no real-world data were used for developing the
that the generalized RBF network is not suitable for deep ANN and MLR models, and the study concluded that the
groundwater-level modeling, whereas IDNN and PNN are ANN model is more suitable than the MLR model for
effective for predictions up to 2 months ahead. Lallahem et al. forecasting piezometric levels at the selected well. In a recent
(2005) developed a multi-layer feed-forward network, trained study, Izady et al. (2012) investigated the performance of
with a standard back-propagation algorithm, for estimating fixed-effects and random-effects panel-data regression
groundwater levels in 13 piezometers installed in the models for predicting groundwater levels in six representa-
unconfined chalky aquifer of northern France, using monthly tive wells in the Neishaboor Plain, Iran, using monthly
rainfall, temperature and potential evapotranspiration as precipitation and evapotranspiration as inputs to the models.
inputs. Daliakopoulos et al. (2005) and Affandi et al. (2007) The panel-data model results were compared with those of
compared different types of back-propagation algorithms for the ANN models and the former performed better than the
predicting water-level fluctuations, and found that the latter.
performance of the ANN model improved from the gradi- The aforementioned review of literature reveals that a
ent-descent algorithm to the Levenberg-Marquardt algorithm, multi-layer feed-forward neural network, trained with the
but decreased in the case of the RBF algorithm. Nayak et al. Levenberg-Marquardt algorithm, is more efficient than
(2006) developed multi-layer feed-forward networks, trained other types of ANN (e.g., traditional gradient-descent
with standard back-propagation algorithms, for forecasting ANN and RBF ANN). The number and types of input, and
groundwater-level fluctuations at two sites in a shallow the number of sites used for predicting groundwater levels,
unconfined aquifer of the Central Godavari Delta System, vary considerably from one study to another. Further-
South India, using groundwater levels of different lags, more, the review reveals that in all likelihood, no studies
rainfall and canal releases as inputs to the model. Krishna et have been conducted to date comparing the predictive
al. (2008) used feed-forward neural networks and RBF capabilities of MLR and ANN techniques in simulating
networks, with Levenberg-Marquardt and Bayesian Regu- groundwater levels using real-world data. Therefore, the
larization training algorithms, to predict water levels focus of this study was on evaluating the efficacy of two
1 month ahead in six wells installed in an unconfined data-driven approaches such as MLR and ANN, for
aquifer in Andhra Pradesh, South India. Monthly rainfall, predicting the spatio-temporal distribution of water-levels
temperature and evapotranspiration were used as inputs, in a groundwater basin, using relevant real-world data. In
and the feed-forward neural network with Levenberg- the present study, standard protocols of MLR and ANN
Marquardt algorithm was reported to be suitable for modeling have been strictly followed and all the
predicting water levels in the six wells. Trichakis et al. significant real-world data have been considered as inputs
(2009) developed an ANN model using a multi-layer to the models. Thus, this study demonstrates a scientif-
perception network, trained with an error back-propagation ically sound methodology for the evaluation of two data-
algorithm, to predict water levels in two wells installed in driven approaches (modeling tools) in simulating ground-
an unconfined karstic aquifer in Mavrosouvala, Greece. water levels using real-world data.

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1867
Methodology (significant) input variables. Generally, all of the potential
input variables are not equally informative, because of the
Overview of study area fact that some variables may be correlated, noisy, or have
The study area selected for the present study is known as no significant relationship with the output to be modeled
the Konan groundwater basin, which is located in Kochi (Maier and Dandy 1998). In order to study the influence
Prefecture of Shikoku Island, Japan (Fig. 1). The Konan of rainfall, temperature, and river stage on groundwater
groundwater basin is bounded by the Monobe River levels at the 17 selected sites, cross-correlation analyses
(perennial) in the west and the Koso River (intermittent) in were performed. The values of the correlation coefficients
the east. Mountains demarcate the northern boundary and are summarized in Table 1, which shows that rainfall,
the southern boundary is demarcated by the Pacific Ocean. temperature and river stage have good correlation with the
There are two further intermittent rivers called the Karasu groundwater levels of all 17 sites. Similarly, the effect of
River and the Yamakita River. Land use mainly comprises the previous month’s groundwater levels on the current
agricultural land (paddy fields, greenhouses and fisheries), month’s groundwater levels was analyzed by autocorrela-
industry and built-up areas. The study area encompasses a tion and partial-autocorrelation techniques (Fig. 2) with
total geographical area of 2,200 ha, of which about 1,502.5 ha the help of STATISTICA 6.0 software (Statistica 2001).
is paddy fields, 488.0 ha is upland, and 186.5 ha is under Apart from the three hydro-meteorological variables, 11
greenhouse cultivation. seasonal dummy variables (D1, D2, D3, …D11) were also
Cold dry winters, and warm humid summers charac- considered, in order to capture the seasonal fluctuation of
terize the regional climate. The average daily maximum groundwater levels in the monthly datasets. Each of these
temperature is 37 °C during summer, and the average dummy variables is equivalent to a new explanatory
daily minimum temperature is −4 °C during winter. The variable, and 11 of them (i.e., 12–1) were used to denote
mean annual rainfall and evapotranspiration in the study monthly periods, as per the rule in Makridakis et al.
area are about 2,600 mm and 800 mm, respectively. (2008). These 11 monthly variables are defined as: D1 =1,
Unconfined aquifers consisting of alluvial sand and gravel if the month is January and the value of D1 is zero for the
and/or diluvial silty sand and gravel are predominant over remaining months; D2 =1, if the month is February and zero
the Konan groundwater basin (Jha et al. 1999). The otherwise and so on, with the value of D11 =1 for November
unconfined aquifer system encompasses the area between and zero for the remaining months. In the regression
the Koso River in the east and the Monobe River in the equations, the regression coefficients associated with these
west (Fig. 1). The hydraulic conductivity of these aquifers variables reflect the average difference in the groundwater-
varies from 65 to 804 m/day (Jha et al. 1999). level between those months and the omitted month (i.e.,
December), which is considered as a base period. Therefore,
the coefficient associated with D1 is a measure of the effect
Study sites and data sources of January–February change in groundwater-level at a site
In the present study, 17 sites (A-2, B-3, C-2, C-7, D-6, E-2, compared to December. Similarly, D2 refers to the effect of
E-4, E-5, F-1, F-6, G-2, GH-4.5, H-2, H-3, H-4, H-5 and I-2) February–March change in groundwater-level at a site, D3
over the Konan groundwater basin of Kochi Prefecture, refers to the effect of March–April change in groundwater-
Japan (Fig. 1) were selected, based on the availability and level, and so on, with D11 representing the effect of
continuity of groundwater-level time-series data. The select- November–December change in groundwater-level. Thus,
ed sites (encircled observation wells in Fig. 1) are located in a total of 22 model inputs were selected for each of the 17
the unconfined aquifer of the basin, and are more-or-less sites, which are: monthly rainfall, monthly mean ambient
representative of the unconfined aquifer underlying the temperature, monthly mean river stage, 11 seasonal dummy
Konan basin. 1999). Further details of the Konan ground- variables, and significant lags (1-month and 2-month lags) of
water basin can be found in Jha et al. (1999). rainfall, ambient temperature, river stage and groundwater-
Daily groundwater-level data for a 6-year period (1999 level.
to 2004) for the 17 sites were collected from Kochi
Prefectural Office, Kochi City, Japan. The daily rainfall,
and maximum and minimum ambient temperature data for MLR modeling
the 1999–2004 period were obtained from the Gomen
meteorological station, situated in Nankoku-shi, Kochi MLR model
Prefecture, about 7 km west of the Monobe River. Besides The MLR technique models the relationship between two or
these data, daily river-stage data for the Monobe River at more explanatory (independent) variables and a response
the Fukabuchi stream-gauging station (Fig. 1) for the (dependent) variable by fitting a linear equation to the
1999–2004 period were also collected from Kochi Work observed data. The generic form of an MLR model is as
Office, Ministry of Construction, Japan. follows (Makridakis et al. 2008):
Y i ¼ β 0 þ β1 X 1;i þ β 2 X 2;i þ ⋯⋯⋯⋯ þ βk X k;i þ εi ð1Þ
Selection of model inputs
One of the most important steps in the development of where Yi represents the ith observation of the dependent
MLR and ANN models is the selection of influential variable Y; X1, i, X2, i, ……., Xk, i represent the ith

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1868

Kochi Prefecture

Source: http://wikitravel.org/en/File:Shikoku.png

Study Area

Fig. 1 Location of the study area and observed sites in the Konan groundwater basin

observations of each of the independent variables X1, Hence, the pragmatic form of the statistical regression
X2,…, Xk respectively; β0, β1, β2........, βk are fixed but model obtained after applying the least-square technique
unknown parameters; and εi is a random variable that is is given as (Makridakis et al. 2008):
normally distributed.
The task of MLR modeling is to estimate the unknown
parameters (β0, β1, β2........, βk) of an MLR model (Eq. 1). Y i ¼ b0 þ b1 X l;i þ b2 X 2;i þ ⋯⋯⋯⋯ þ bk X k;i þ ei ð2Þ

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1869
Table 1 Cross-correlation between water levels in selected observation where i=1, 2,…., n; b0, b1, b2.......bk are the estimates or
wells and three independent variables (rainfall, temperature and river stage) unstandardized regression coefficients of β0, β1, β2,...., βk
Observation well Values of correlation coefficient (r) respectively; and ei is the estimated error (residual) for the
Rainfall Temperature River stage
ith observation.
Therefore, estimate of Y is given by:
A-2 0.818 0.881 0.904
B-3
C-2
0.801
0.651a
0.687a
0.881
0.744
0.694a
Yb ¼ b0 þ b1 X l;i þ b2 X 2;i þ ⋯⋯⋯⋯ þ bk X k;i ð3Þ
C-7 −0.330b −0.108b −0.333b
D-6 0.721 0.875 0.797 with the difference between the observed Y and the estimated
E-2 0.679a 0.888 0.757 Yb being known as the ‘residual’ (or ‘residual error’).
E-4 0.676a 0.917 0.739
E-5 0.682a 0.937 0.765
F-1 0.821 0.866 0.918 Development of MLR models
F-6 0.700 0.905 0.766
G-2 0.713 0.921 0.801 Site-specific MLR models were developed for all the 17
GH-4.5 0.593a 0.771 0.687a sites, using influential hydrological variables, 11 seasonal
H-2 0.666a 0.874 0.778 dummy variables and significant multi-period lags of
H-3 0.664a 0.902 0.770 hydrological and hydrogeological variables as indepen-
H-4 0.635a 0.854 0.719
H-5 0.558a 0.688a 0.692a dent variables, and the groundwater levels of these sites as
I-2 0.664a 0.867 0.793 dependent variables. In all, 70 % of the data, covering the
period 1999–2002 was used for training and 30 % of the
Strongly correlated (r>0.7) data (for the period 2003–2004) was used for testing. The
a
Moderately correlated (0.5≤r≤0.7) step-wise-regression method was employed using
b
Poorly correlated (r<0.5) STATISTICA 6.0 software to identify significant inputs
to the site-specific MLR models, in order to predict
groundwater levels at the 17 sites.

Evaluation of the developed MLR models

The effectiveness of the developed MLR models was
measured by a set of standard statistical indicators, namely:
standardized regression coefficient (βj), standard error (SE),
t-test, F-test, coefficient of multiple determination (R2),
multiple correlation coefficient (R), adjusted R2, p-level and
standard error of estimate (SEE). The mathematical expres-
sions for these indicators are as follows:
σj
β j¼ b j ð4Þ
σY
where bj is the unstandardized regression coefficient and
σj and σY are the standard deviations associated with the
jth independent variable and the dependent variable Y,
respectively.
σ
SE ¼ pffiffiffiffiffiffiffiffiffi ð5Þ
n –1
where σ=sample standard deviation, and n=total number
of observations. In the t-test, the value of t is given as:

bj
t¼  ð6Þ
SE b jÞ

where bj =unstandardized regression coefficients of un-

known parameters. In the F-test, the value of F is given as:


. 
SSR p MSR
F¼ . ¼ ð7Þ
Fig. 2 Correlograms for water levels at site A-2 for various time lags, SSE ðn – p – 1Þ MSE
showing: a autocorrelation function; b partial-autocorrelation function

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1870
where SSR=sum of squares regression, SSE=sum of 2001). Finally, the standard error of estimate (SEE) is the
squared error, SST=total sum of squares, MSR=mean measure of the amount of error in the prediction of a
square regression, MSE=mean squared error, p=number of dependent variable (Y) for each independent variable (X) in
independent variables, and n=total number of observations. the regression equation.

X 2
Yb i −Y SSR Design of ANN models
R2 ¼ X  2 ¼ ð8Þ
SST A neural network is made up of a number of nodes, which
Y i −Y are the processing elements of the network and are usually
called ‘neurons’. Each neuron is connected to other neurons,
receives an input signal, processes it and transforms it into an
where Yi =observed value of a dependent variable, Yb i = output signal (Haykin 1994). In order to execute the function
estimated or predicted value of Y, and Y = mean value of a of biological neurons artificially, a proper design of artificial
dependent variable. neural network (ANN) is necessary. Hence, designing ANN
  (or ANN models) follows a systematic procedure involving
n–1   five basics steps: (1) selection of influential inputs; (2)
Adjusted R ¼ 1−
2
1−R2
n–p–1 selection of suitable ANN architecture; (3) building the
0 . 1 neural network; (4) training and testing of the developed
SSE ðn – p – 1Þ ANN models; and (5) performance evaluation of the ANN
¼ 1−@ . A ð9Þ models. The first step is described in section “Study sites and
SST ðn – 1Þ data sources” and a brief description of the remaining steps is
given in the following.

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Selection of ANN architecture

uX
u n  2 After selecting significant inputs for the ANN, the next step is
u Y – bi
Y
u i the selection of suitable network architecture. A multi-layer
t i¼1
SEE ¼ ð10Þ feed-forward network or multi-layer perceptrons (MLP) with
n–p–1 single hidden layer was selected for this study, because this
architecture has been widely used in simulation/prediction of
groundwater levels for its ease of implementation and also for
where all the symbols have the same meaning as defined being considered as a universal approximator (Hornik et al.
earlier. 1989). In a multi-layer feed-forward neural network, the
The standardized regression coefficient (βj) measures the nodes are generally arranged in layers, comprising an input
impact of a unit change in the standardized value of an layer, an output layer and one or more hidden layers. A
independent variable on the standardized value of a schematic diagram of a three-layer feed-forward neural
dependent variable. Standard error (SE) is a measure of the network having an input layer (with 22 input neurons), one
stability of regression coefficients. A t-test is performed for hidden layer and one output layer, as used in this study, is
each regression coefficient, to examine its significance in the illustrated in Fig. 3. It has input nodes (x1, x2, x3,……….., xi,..,
presence of all other independent variables. The overall xn) including bias (x0), hidden nodes (z1, z2, z3,………..,zj,.., zl)
statistical significance of the regression model is evaluated including bias (z0), and output node (yk), where x, z, and y
using an F-test in the format of ANOVA (analysis of represent the input, hidden, and output layers respectively,
variance; Makridakis et al. 2008). The coefficient of multiple and n, l, and m represent the number of the nodes in each layer
determination (R2), is a fraction that represents the propor- (Fig. 3). The weights associated with the connections between
tion of total variation of the dependent variable accounted for the input and the hidden nodes are denoted by wij and those
(explained) by the explanatory variables. The fit of a multiple between the hidden and the output nodes are denoted by wjk.
regression model can be assessed by R2. The multiple The addition of a bias term to an artificial neuron can be
correlation coefficient (R) is the correlation between the expressed by incorporating a bias element into the input and
observed value of a dependent variable (Y) and an estimated weight vectors to create extended vectors resulting in an
or predicted value of Y based on multiple independent increase of their dimensionality by one. It is useful as it helps
variables. The ‘adjusted R2’ is interpreted in the same the neural network in learning patterns by providing
manner as the R2 value, except that the adjusted R2 takes into additional weights to the connection between neurons and
consideration the number of degrees of freedom (Statistica the bias input is generally assigned a value of positive one
2001). Moreover, the p-level represents the probability of (Haykin 1994). The output of a node in the hidden layer (zj) is
error that is involved in accepting the observed results as determined by applying a nonlinear transformation (activa-
valid. In many fields of research, including the field of tion function) to the sum of the weighted inputs it received
hydrological sciences, a p-level of 0.05 (i.e., 5 % probability from neurons in the input layer. Then the weighted sum of
of error or 95 % confidence interval) is customarily treated as inputs at the output layer is transformed to the network output
a “border-line acceptable” error level (Haan 2002; Statistica using a linear activation function between hidden and output

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1871
Input Layer

x0 (Bias)

x1 (Rainfall)

x2 (Temperature)
Hidden Layer
x3 (River stage)

x4 (Dummy 1) z0
Output Layer
x5 (Dummy 2) .
.
.
.
. .
Output ( )
x14 (Dummy 11)
. Groundwater-level
.
x15 (Rainfall t-1) .
Output Node (m) = 1
x16
. Connection
(Rainfall t-2) zl
Weight

x17 (River stage t-1)

x18 (River stage t-2)

Hidden Nodes (l)
x19 (Temp t-1)

x20 (Temp t-2) Connection

Weight

x21 (GW t-1)

x22 (GW t-2)

Input Nodes (n) = 22

Fig. 3 The architecture for the multi-layer feed-forward network used in the study (parameters defined in the text)

layer. Thus, the prediction of the ANN model (yk) can be simulated values are as close as possible to the known output
expressed as follows (Haykin 1994): (observed) values. The most common method used for
training feed-forward neural networks is back-propagation
X
l
yk ¼ wjk z j þ b j ð11Þ training (Hagan and Menhaj 1994; Hagan et al. 1996), which
j¼1
is a two-step approach. In the first step, the input signal is
propagated forward to estimate the outputs. In the second step,
where wjk is the connection weight between jth node of
a backward step is performed to adjust the weight vectors
hidden layer and output node k; l is the number of neurons
between the layers to minimize the network error. Although
in the hidden layer; zj is the output of the jth hidden
various weight-optimization techniques are available for the
neuron resulting from all the input data and bj is the
training of ANNs, the Levenberg-Marquardt (LM) algorithm
connection weight for bias term.
was used in this study, owing to its ability to achieve
convergence quickly (Moré 1978; Press et al. 1992). The
Training algorithm LM algorithm trains neural networks at a rate 10 to 100 times
After selecting the ANN architecture, the next step is the faster than the usual gradient-descent back-propagation
training of the network. The purpose of training is to ensure algorithm, and is more effective at finding optimal results
that the network extracts the fundamental characteristics or than standard back-propagation (Hagan and Menhaj 1994).
pattern from the datasets used in ANN modeling. The training The LM algorithm combines the Gauss-Newton algorithm
process consists of determining the ANN weights and biases. and the method of gradient descent, and it uses an
The ANN is trained with a set of input and known output data approximation of the Hessian matrix to update weights as
to determine the connection weights, so that the ANN- given below (Bishop 1995):

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1872
Table 2 Categorization of sensitivity and associated ranks function and in Eq. (14) v is the weighted sum of inputs at the
Category of sensitivity Value of sensitivity index Rank output layer as considered in the present study.
Very high sensitivity >0.1 1
High sensitivity >0.05–0.1 2
Moderate sensitivity >0.01–0.05 3 Number of hidden-layer neurons
Low sensitivity >0.005–0.01 4 The function of the hidden-layer neurons is to detect
Very low sensitivity ≤0.005 5 relationships between network inputs and outputs. If there
is an insufficient number of hidden neurons, it may be
h difficult to obtain convergence within the training time.
Δw ¼ JT ðwÞJðwÞ þ l I−1 JT ðwÞeðwÞ ð12Þ On the other hand, if too many hidden neurons are used,
the network may lose its ability to generalize (ASCE
where, w=weight, J =Jacobian matrix, JT =transpose 2000a). Although a trial-and-error method is widely used
for determining the optimal number of hidden neurons,
matrix of J, JTJ=Hessian matrix, l=learning parameter,
some researchers have demonstrated the effectiveness of a
I=identity matrix and e=vector of network errors. The
genetic algorithm (GA) for optimizing hidden neurons
parameter l governs the step size and is automatically (Jones 1993; Castillo et al. 2000; Gerken et al. 2006). In
adjusted based on the direction of the error at each particular, GA has been found to be quite useful and
iteration, in order to secure convergence. In this study, an efficient when the exploration space of the network is
initial l value of 0.15 was used for optimizing weights extensive. Consequently, the heuristic optimization tech-
with the LM algorithm. nique GA was used in this study for optimizing the hidden
neurons for all the 17 site-specific ANN models.
Activation function In the GA optimization, initially, chromosomes (consisting
The activation (transfer) function determines the response of of ‘genes’ of binary numbers) were characterized (encoded) by
a node to the total input signal it receives. The most a set of unknown parameters (i.e., hidden neurons). The number
commonly used activation function, named logistic sig- of hidden neurons was set in a range between a lower bound of
moid-type function (Haykin 1994; Dekker et al. 2001) was 1 and an upper bound of 30. Thereafter, the GA parameters such
used in this study for the hidden layer. However, a linear- as selection method, crossover probability, mutation probability,
type activation function was used for the output layer, as population size, and number of generations, were decided based
suggested by Maier and Dandy (2000) and Rumelhart et al. on trial and error and standard literature (Goldberg 1989; Reed
(1995). The sigmoid function is a bounded, monotonic, non- et al. 2000 and Mayer et al. 2001). The GAwas initiated with 50
decreasing function that provides a graded, non-linear randomly generated chromosomes, with a maximum number of
response (Shamseldin 1997), whereas a linear-transfer generations of 200, with gene structures as mentioned in the
function calculates a neuron’s output by simply returning preceding. The best-fit chromosomes were chosen by roulette-
the value passed to it. The mathematical expressions for wheel selection, based on the ranking-algorithm method, and
these two functions are as follows: the crossover and mutation operators were applied with a
crossover probability of 0.9–1.0 and a mutation probability of
1 0.001–0.01, to produce a new population (of chromosomes) for
Logistic sigmoid function : φðvÞ ¼ ð13Þ
1 þ expð−vÞ the next generation. The new chromosomes thus reproduced
Linear function : φðvÞ ¼ v ð14Þ were evaluated based on their fitness values, and this procedure
of evaluation and reproduction of the chromosomes was
where v in Eq. (13) represents the weighted sum for a node in repeated until a pre-defined termination criterion (i.e., ‘fitness
the hidden layer and exp denotes the natural exponential threshold’) of 0.001 m2 was satisfied. The combination of

Table 3 MLR coefficients and their significance for the MLR model for site F-6 using a step-wise regression technique (parameters defined
in the text). SE standard error; Temp temperature
Parameter Beta SE (m) B SE (m) t (37) p-level
Intercept 1.625 0.238 6.831 0.000000
Temp 0.259 0.106 0.025 0.010 2.448 0.001921
D4 (Apr–May) 0.236 0.027 0.618 0.070 8.840 0.000000
GWt–1 0.497 0.080 0.495 0.080 6.175 0.000000
D7 (Jul–Aug) 0.173 0.027 0.453 0.071 6.356 0.000000
Rainfall 0.141 0.027 0.001 0.000 5.160 0.000009
Rainfallt –2 −0.074 0.028 0.0003 0.000 −2.621 0.000637
D10 (Oct–Nov) −0.036 0.024 −0.095 0.064 −1.483 0.006509
D2 (Feb–Mar) −0.065 0.025 −0.170 0.066 −2.593 0.001354
Tempt–2 −0.266 0.102 −0.026 0.010 −2.620 0.001267
Tempt–1 0.329 0.157 0.032 0.015 2.089 0.004366
Regression summary: multiple R=0.941, multiple R2 =0.885, adjusted R2 =0.869, F-statistic=216.414, p-level=0.0000 and SEE=0.107 m;
D2, D4 etc. refer to seasonal dummy variables

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1873
Table 4 MLR models for predicting groundwater levels at 17 sites using training datasets

Site MLR models based on training datasets

A-2 GWL ¼ 16:1572 þ 0:0003  R þ 0:0129  T þ 0:8803  0:2878  D4 þ 0:0917  D5 −0:1040

D9 −0:1020  D10 þ 0:0003  Rt−1 þ 0:0133  T t−1 −0:2404  St−1 −0:0647  St−2

B-3 GWL ¼ 14:2696 þ 0:0004  R þ 0:0375  D3 þ 0:0586  D4 þ 0:0663  D8 −0:2739  GWt−1

þ0:0003  Rt−1 −0:1529  St−1

C-2 GWL ¼ 4:9457 þ 0:0011  R þ 0:1033  T þ 0:2836  D2 þ 1:6220  D4 þ 0:6288  D3 −0:7022

D8 −0:3603  D9 −0:4169  D10 þ 0:5431  GWt−1 −0:0007  Rt−2 −0:4311  St−2

C-7 GWL ¼ 4:0716−0:0002  R þ 0:0519  D4 þ 0:0489  D6 þ 0:0666  D9 þ 0:4985  GWt−1

−0:0003  Rt−1 þ 0:1618  St−1 þ 0:0729  St−2

D-6 GWL ¼ 5:4199 þ 0:0003  R þ 0:0432  T þ 0:6289  S−0:212  D1 þ 0:0620  D4 −0:5236

D2 −0:3439  D3 −0:163135  D8 þ 0:000358  Rt−1 −0:0186  T t−2

E-2 GWL ¼ 2:9648 þ 0:0235  T þ 1:3327  S−0:2348  D1 þ 0:2587  D2 þ 0:4599  D3 þ 1:4086

D4 þ 0:3248  D5 þ 0:4926  D7 þ 0:4037  GWt−1 þ 0:0008  Rt−1 −1:4295  St−1 þ 0:0314  T t−1

E-4 GWL ¼ 2:6132 þ 0:0011  R þ 0:0748  T þ 1:0500  D4 −0:3106  D1 þ 0:4598  D7 −0:5201

D2 −0:1631  D8 þ 0:1911  D11 þ 0:3776  GWt−1 þ 0:0951  GWt−2 −0:0004  Rt−1 −0:0009
Rt−2 −0:0248  T t−2
E-5 GWL ¼ 2:0439 þ 0:0011  R þ 0:0349  T þ 1:0076  D4 −0:2406  D2 þ 0:9641  D7 þ 0:5285
GWt−1 þ 0:0005  Rt−2 þ 0:0421  T t−1 −0:0421  T t−2

F-1 GWL ¼ 2:3141 þ 0:0004  R−0:0079  T þ 1:0911  S−0:2895  D4 þ 0:0426  D5 þ 0:1080

D6 þ 0:1486  D7 þ 0:2649  GWt−1 −0:2179  GWt−2 −0:0003  Rt−t þ 0:0003  Rt−2 −0:5139
St−1 þ 0:0357  T t−1 þ 0:0121  T t−2
F-6 GWL ¼ 1:6247 þ 0:0006  R þ 0:0252  T þ 0:6183  D4 −0:1702  D2 þ 0:4527  D7 −0:0948
D10 þ 0:4591  GWt−1 −0:0003  Rt−2 þ 0:0322  T t−1 −0:0262T t−2

G-2 GWL ¼ 0:4252 þ 0:0008  R þ 0:0615  T þ 0:3063  S þ 0:1820  D1 −0:2577  D2 −0:8261

D4 þ 0:2460  D7 þ 0:2432  D11 þ 0:2498  GWt−1 þ 0:1022  GWt−2 þ 0:0005  Rt−1 −0:2998
St−1 −0:0006  Rt−2
GH-4.5
GWL ¼ −0:9839 þ 0:0010  R þ 0:2445  D1 þ 0:4051  D2 þ 0:7339  D3 þ 0:9149  D4
þ0:3509  D11 þ 0:6470  GWt−1 −0:2079  GWt−2 þ 0:0005  Rt−1 þ 0:0902  T t−1 −0:6591  St−1
H-2 GWL ¼ −0:0774 þ 0:0003  R þ 0:0257  T þ 0:0861  S þ 0:1150  D4 −0:1496  D5 −0:1080
D6 þ 0:0900  D11 þ 0:5011  GWt−1 þ 0:3718  GWT −2 þ 0:0004  Rt−2 þ 0:0097  T t−1 −0:0199
T t−2 þ 0:2208  St−1
H-3 GWL ¼ −0:1759 þ 0:0007  R þ 0:0261  T þ 0:2050  D3 þ 0:7763  D4 þ 0:0758  D6
þ0:2191  D7 þ 0:2100  D11 þ 0:3995  GWt−1 þ 0:0005  Rt−1 −0:0002  Rt−2 þ 0:0373  T t−1
−0:3886  St−1
H-4
GWL ¼ −0:0068 þ 0:0007  R þ 0:2399  D3 þ 0:6511  D4 þ 0:1570  D11 þ 0:1432  D7
þ0:4849  GWt−1 −0:0927  GWt−2 þ 0:0005  Rt−1 þ 0:0584  T t−1 −0:0102  T t−2 þ 0:4554  St−1
H-5
GWL ¼ −0:5338 þ 0:0003  R þ 0:2817  S þ 0:1116  D2 −0:1053  D7 þ 0:2362  D3
þ0:2652  D4 −0:0859  D5 þ 0:5748  GWt−1 þ 0:0356  T t−1 −0:4258  St−1
I-2 GWL ¼ −0:0626 þ 0:0175  T þ 0:2311  S þ 0:0004  Rt−2 þ 0:0748  D5 þ 0:4272  GWt−1
−0:0002  Rt−1 −0:0033  T t−1 þ 0:1847  St−2

GWL groundwater-level, R rainfall, T temperature, S river stage, D1, D2, D3,..,D11 seasonal dummy variables

parameter that produced the lowest fitness value during training criteria, such as fixing the number of epochs setting a
runs was selected as the optimal ANN network. target error-goal and fixing the minimum performance
gradient (derivatives of network error with respect to those
Optimization of neuron weights weights and bias). An initial weight range of −0.5 to 0.5 was
During the training process, the three factors that are selected in this study, as suggested by earlier researchers
associated with optimization of weights are: (1) initial (e.g., Sietsma and Dow 1991; Looney 1996). These weights
weight matrix; (2) error function; and (3) termination were then optimized using Eq. (11). The following objective

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1874
Table 5 Statistical indicators for the developed MLR models based on training datasets
Site Multiple R Multiple R2 Adjusted R2 F-statistic p-level SEE (m)
A-2 0.981 0.963 0.951 84.546 0.0000 0.090
B-3 0.884 0.781 0.742 20.337 0.0000 0.057
C-2 0.973 0.947 0.931 58.217 0.0000 0.351
C-7 0.738 0.544 0.436 5.040 0.0002 0.067
D-6 0.969 0.939 0.923 57.305 0.0000 0.163
E-2 0.944 0.891 0.848 72.630 0.0000 0.227
E-4 0.986 0.973 0.962 93.055 0.0000 0.227
E-5 0.975 0.950 0.936 210.716 0.0000 0.177
F-1 0.958 0.917 0.897 99.452 0.0000 0.089
F-6 0.941 0.885 0.869 216.414 0.0000 0.107
G-2 0.986 0.973 0.962 93.319 0.0000 0.174
GH-4.5 0.979 0.958 0.946 75.270 0.0000 0.221
H-2 0.975 0.951 0.932 50.661 0.0000 0.090
H-3 0.949 0.901 0.885 154.325 0.0000 0.117
H-4 0.974 0.948 0.932 146.570 0.0000 0.104
H-5 0.927 0.860 0.851 49.270 0.0000 0.120
I-2 0.933 0.870 0.864 45.805 0.0000 0.070

function (E) was used for the optimization of weights error (RMSE), coefficient of determination (r2) and Nash-
(Principe et al. 2000): Sutcliffe efficiency (NSE), in order to examine their
effectiveness in predicting transient groundwater levels at
1 X 2
n
individual sites. MAE indicates the average of the total
E¼ hoi −hpi ð15Þ model errors (both under-prediction and over-prediction)
2n i¼1
and is used to measure how close forecasts/predictions are to
where, hoi =observed groundwater-level at the ith time [L], the eventual outcomes (observed values). RMSE indicates
hpi =predicted groundwater-level at the ith time [L], and an overall (global) discrepancy between the observed values
n=total number of observations. and the calculated (predicted or simulated) values. r2
Monthly groundwater-level data for the 4-year period describes the proportion of the total variance in the observed
1999–2002 were used for training the 17 site-specific ANN data that can be explained by a model. NSE, on the other
models, with the 2-year period 2003–2004 being used for hand, is an index for assessing the predictive accuracy of a
testing. The entire ANN modeling exercise was carried out model, and it represents an improvement over r2 for model
using the Neural Builder Wizard of NeuroSolutions (version evaluation, because it is sensitive to the differences in
5.05) software (Principe et al. 2000). observed and model-simulated or predicted means and
variances (Legates and McCabe 1999). The mathematical
expressions for MAE, RMSE, r2 and NSE are as follows:
1X
n
Model evaluation criteria
Performance evaluation of the 17 site-specific MLR and MAE ¼ hoi −hpi ð16Þ
n i¼1
ANN models was carried out using four statistical indicators,
namely: mean absolute error (MAE), root mean squared

Fig. 4 Plot of residuals versus predicted groundwater levels for

MLR model of site GH-4.5 Fig. 5 Histogram of residuals for MLR model of site GH-4.5

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1875
Table 7 Neural network architecture and optimal GA parameters
for the 17 ANN models
Site ANN No. of Best Fitness
architecture generations chromosome (m2)
(n-l-m)
A-2 22-14-1 2 29 0.001
B-3 21-8-1 1 34 0.001
C-2 22-8-1 1 32 0.004
C-7 21-11-1 2 30 0.001
D-6 22-10-1 1 29 0.059
E-2 22-6-1 2 35 0.048
E-4 22-7-1 2 40 0.017
E-5 22-6-1 1 26 0.017
F-1 22-7-1 2 20 0.008
F-6 22-11-1 1 19 0.017
G-2 22-6-1 1 39 0.003
GH-4.5 22-10-1 2 34 0.0005
H-2 22-7-1 1 32 0.0002
H-3 22-9-1 2 21 0.021
Fig. 6 Normal probability plot of residuals for MLR model of site H-4 22-10-1 2 39 0.0057
GH-4.5 H-5 22-9-1 1 42 0.0079
I-2 22-12-1 1 32 0.0016
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi n number of input neurons, l number of hidden layer neurons, m
1X 2
n
number of output neurons
RMSE ¼ hoi −hpi ð17Þ
n i¼1 In addition to quantitative evaluation using statistical
indicators, the efficacy of the MLR and ANN techniques
2 32 in predicting transient groundwater levels at the 17 sites
n h
X  i
6 hpi −hp hoi −ho 7 was assessed by graphical indicators such as visual
6 7
6 7 checking of simulated and observed groundwater-level
r ¼6
2
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
6 u"
i¼1
! !# 7 ffi 7 ð18Þ hydrographs for individual sites, as well as scatter plots of
6u X n  2 X n  2 7
4t h −h h −h 5 simulated versus observed groundwater levels along with
pi p oi o
i¼1 i¼1 1:1 lines for individual sites. Prediction errors (as a
measure of model uncertainty) of all the MLR and ANN
X
n
 2 models were also analyzed.
hoi −hpi
i¼1
NSE ¼ 1− Xn  2 ð19Þ
hoi −ho Sensitivity analysis
i¼1 Sensitivity analyses were conducted for the ANN model
(i.e., MLP-LM) to investigate the relative importance of
where, hoi =observed groundwater-level at the ith time [L], each input variable for accurately predicting groundwa-
hpi =predicted groundwater-level at the ith time [L], ho = ter-levels. This analysis was carried out for all the 17
mean of the observed groundwater levels [L], hp = mean sites by imposing certain changes on individual inputs
of the predicted groundwater levels [L], and n=total and observing their effects on the model output. The
number of observations. change in the input was made by adding a random value

Table 6 Results of the multicollinearity test for site GH-4.5

Variable R2 Multicollinearity statistic t (36) p-Level
VIF Tolerance
Tempt–1 0.728 3.671 0.272 8.697 0.0000
D4 (Apr–May) 0.571 2.329 0.429 5.198 0.0000
GWt–1 0.712 3.478 0.288 5.601 0.0000
Rainfall 0.443 1.796 0.557 4.243 0.0001
D3 (Mar–Apr) 0.659 2.930 0.341 3.717 0.0007
D11 (Nov–Dec) 0.303 1.436 0.697 2.539 0.0056
GW(t–2) 0.752 4.032 0.248 3.346 0.0046
D2(Feb–Mar) 0.654 2.887 0.346 2.068 0.0025
River staget–1 0.763 4.214 0.237 4.841 0.0038
D1 (Jan–Feb) 0.611 2.568 0.389 3.323 0.0043
Rainfallt–1 0.733 3.749 0.267 3.178 0.0047
VIF variance inflation factor; D1, D2, D3 etc. refer to seasonal dummy variables

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1876

Table 8 Goodness-of-fit statistics for the ANN and MLR models of the 17 sites during training and testing periods
Site Dataset used ANN models MLR models
MAE (m) RMSE (m) r2 NSE MAE (m) RMSE (m) r2 NSE
A-2 Training 0.072 0.05 0.987 0.984 0.122 0.078 0.963 0.963
Testing 0.121 0.128 0.935 0.912 0.137 0.127 0.914 0.914
B-3 Training 0.023 0.033 0.925 0.907 0.041 0.052 0.780 0.780
Testing 0.038 0.04 0.825 0.822 0.056 0.050 0.736 0.736
C-2 Training 0.075 0.085 0.999 0.995 0.213 0.304 0.947 0.947
Testing 0.311 0.397 0.918 0.901 0.297 0.474 0.858 0.858
C-7 Training 0.021 0.036 0.832 0.828 0.057 0.059 0.544 0.544
Testing 0.042 0.038 0.781 0.776 0.061 0.080 0.007 0.007

Hydrogeology Journal (2013) 21: 1865–1887

D-6 Training 0.047 0.059 0.993 0.990 0.102 0.144 0.939 0.939
Testing 0.104 0.157 0.930 0.890 0.127 0.255 0.709 0.709
E-2 Training 0.221 0.304 0.910 0.905 0.246 0.394 0.891 0.891
Testing 0.241 0.220 0.935 0.995 0.278 0.263 0.930 0.930
E-4 Training 0.157 0.182 0.976 0.975 0.173 0.191 0.973 0.973
Testing 0.287 0.411 0.89 0.861 0.306 0.471 0.887 0.887
E-5 Training 0.245 0.186 0.973 0.973 0.265 0.257 0.950 0.950
Testing 0.323 0.264 0.944 0.931 0.317 0.269 0.928 0.928
F-1 Training 0.096 0.123 0.937 0.936 0.102 0.174 0.917 0.917
Testing 0.098 0.132 0.912 0.911 0.114 0.148 0.888 0.888
F-6 Training 0.212 0.186 0.937 0.934 0.228 0.246 0.885 0.885
Testing 0.221 0.187 0.922 0.918 0.243 0.071 0.989 0.989
G-2 Training 0.061 0.075 0.995 0.993 0.072 0.146 0.973 0.973
Testing 0.193 0.277 0.912 0.895 0.197 0.253 0.910 0.910
GH-4.5 Training 0.015 0.03 0.998 0.998 0.045 0.191 0.958 0.958
Testing 0.101 0.129 0.971 0.967 0.115 0.288 0.841 0.841
H-2 Training 0.017 0.020 0.997 0.997 0.036 0.075 0.951 0.951
Testing 0.057 0.100 0.937 0.914 0.062 0.152 0.802 0.802
H-3 Training 0.159 0.205 0.925 0.922 0.178 0.300 0.901 0.901
Testing 0.177 0.246 0.941 0.884 0.192 0.280 0.938 0.938
H-4 Training 0.055 0.107 0.970 0.969 0.076 0.121 0.948 0.948
Testing 0.112 0.161 0.919 0.917 0.127 0.217 0.849 0.849
H-5 Training 0.107 0.126 0.903 0.899 0.132 0.145 0.860 0.860
Testing 0.087 0.118 0.906 0.837 0.152 0.115 0.845 0.845
I-2 Training 0.045 0.056 0.926 0.924 0.076 0.073 0.870 0.870
Testing 0.055 0.068 0.905 0.893 0.089 0.077 0.863 0.863
Range of statistical Training Min: 0.015 (site GH-4.5) 0.02 (site H-2) 0.832 (site C-7) 0.828 (site C-7) 0.036 (site H-2) 0.052 (site B-3) 0.544 (site C-7) 0.544 (site C-7)
indicators Max: 0.245 (site E-5) 0.304 (site E-2) 0.999 (site C-2) 0.998 (site GH-4.5) 0.265 (site E-5) 0.394 (site E-2) 0.973 (site E-4) 0.973 (site E-4)
Testing Min: 0.038 (site B-3) 0.038 (site C-7) 0.781 (site C-7) 0.776 (site C-7) 0.056 (site B-3) 0.050 (site B-3) 0.007 (site C-7) 0.007 (site C-7)
Max: 0.323 (site E-5) 0.411 (site E-4) 0.971 (site GH-4.5) 0.995 (site E-2) 0.317 (site E-5) 0.474 (site C-2) 0.989 (site F-6) 0.989 (site F-6)

DOI 10.1007/s10040-013-1029-5
 1877

Fig. 7 Observed and predicted groundwater levels using ANN and MLR models at site B-3: a for the training period (1999–2002); b for
the testing period (2003–2004) (m MSL m above mean sea level)

of known variance to each sample, while keeping the where, Sk =sensitivity index for input k, yip ¼ ith
other inputs at their mean values and then calculating the output obtained with the fixed weights for the pth
change in the model output (i.e., simulated water level). pattern, n=number of network outputs, p=number of
The changes in the model results were assessed using a patterns, and σk2 = variance of the input k. The
sensitivity index, which is expressed as follows (Principe training dataset was used for computing sensitivity
et al. 2000): indices of each input. After calculating the sensitivity
Xp X n  2 index for all the inputs at each of the 17 sites, they
yip −yip were classified into five categories, depending upon
p¼1 i¼1 their sensitivity index values and ranked on a scale of
Sk ¼ ð20Þ
σk 2 1–5 (Table 2).

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1878

Fig. 8 Observed and predicted groundwater levels using ANN and MLR models at site E-4: a for the training period (1999–2002); b for
the testing period (2003–2004)

Results and discussion wise regression technique to predict groundwater levels

(GWL) at site F-6 is:
Developed MLR models
The results of MLR analysis for site F-6 using 22 GWLð F-6Þ ¼ 1:6247 þ 0:0006  R þ 0:0252  T
independent variables and a step-wise regression tech- þ 0:6183  D4  0:1702  D2 þ 0:4527
nique are summarized in Table 3 as an example. In  D7  0:0948  D10 þ 0:4951  GWt1 ð21Þ
Table 3, the ‘beta’ (standardized regression coefficients)
 0:003  Rt2 þ 0:0322  T t1 0:0262
values show the relative contribution of each independent
variable in prediction of groundwater levels. The ‘B’  T t2
refers to the raw or unstandardized regression coefficients
or weights, which are included in the regression It is obvious from Table 3 that D4 (t=8.84, p=0.0000),
equation. Thus, the MLR model obtained using a step- D7 (t=6.36, p=0.0000), GWt–1 (t=6.17, p=0.0000), rainfall

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1879
(t=5.16, p=0.0000) are highly significant explanatory variability in the fluctuation of groundwater-levels, and it is
variables, indicating higher t values and very low p-level also confirmed that the multiple R2 is highly significant for
(i.e., 0.0000) at the 95 % significance level, followed by this model. The results of other statistical indicators for site
rainfall t–2 (t = −2.62, p = 0.0006) ) , temp t–2 (t = −2.62, F-6, namely, multiple correlation coefficient (R), adjusted R2
p=0.001), D2(t=−2.59, p=0.001), temp (t=2.45, p=0.002), and standard error of estimate (SEE) are also summarized in
tempt–1 (t=2.08, p=0.004) and D10 (t=−1.48, p=0.006) in Table 3.
predicting groundwater levels at site F-6. The remaining Similarly, for the remaining sites, the MLR models
variables are excluded from the model during step-wise developed using training datasets (1999–2002) are
regression, because their regression coefficients were found presented in Table 4. It is apparent from Table 4 that
to be statistically insignificant in predicting groundwater- there are some big differences in the values of regression
levels. coefficient between the sites, which indicates the hetero-
Moreover, an F-test was used to check the overall sig- geneity of the aquifer and/or the variation in local
nificance of the developed MLR model for predicting conditions. This finding emphasizes the need to develop
groundwater levels at site F-6. From the F-statistic site-specific MLR models. Furthermore, the statistical
(F=216.414) and p-value (p=0.0000) of site F-6, it is indicators of step-wise regression for the 17 sites, using
concluded that it is indeed a significant model, i.e., the 22 inputs, are summarized in Table 5. It is apparent from
independent variables explain a significant amount of Table 5 that the MLR model for sites E-4 and G-2 have

Fig. 9 Observed and predicted groundwater levels using ANN and MLR models at site GH-4.5: a for the training period (1999–2002); b
for the testing period (2003–2004)

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1880
the highest multiple R (0.986) and adjusted R2 (0.962) environmental factors under consideration, and that the
values, and their significance is supported by the corre- groundwater-level fluctuation at this site is affected by
sponding F-statistic and p-level. Thus, the values of some anthropogenic factors such as the existence of a
multiple R, adjusted R2, F-statistic and p-level for the rubber dam very close to site C-7.
MLR models for almost all the sites are statistically
significant, which suggests that the developed MLR
models can simulate/predict groundwater-level fluctua- Results of diagnostic checks on the developed MLR
tions reasonably. However, the MLR model of site C-7 models
has the lowest multiple R (0.738) and adjusted R2 (0.436) After developing the MLR models for the 17 sites, the
among the 17 sites. This indicates that the groundwater- models were examined for basic assumptions to justify the
level at site C-7 is not significantly influenced by the use of MLR models for prediction. The four assumptions
of MLR models are (Makridakis et al. 2008): (1) linearity

Fig. 10 Scatter plots of observed and predicted groundwater levels Fig. 11 Scatter plots of observed and predicted groundwater levels
by ANN and MLR models with ± 1 % error band at site B-3: a for by ANN and MLR models with ± 10 % error band at site E-4: a for
the training period (1999–2002); b for the testing period (2003– the training period (1999–2002); b for the testing period (2003–
2004) (r correlation coefficient between observed and simulated 2004) (r correlation coefficient between observed and simulated
groundwater-level elevation) groundwater-level elevation)

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1881
the residuals are distributed in such a manner that they
resemble the normal distribution curve, thereby satisfying
the assumption of normality. Also, Fig. 6 reveals that a
majority of the residuals fall on the straight line; it can
therefore be inferred that the residuals obtained are normally
distributed. The autocorrelation of residuals was checked by
using the Durbin-Watson statistic (Makridakis et al. 2008).
For example, the value of Durbin-Watson statistic for the
residuals obtained at site GH-4.5 is 1.78, which is significant
at a 95 % confidence interval, thereby satisfying the
condition of no autocorrelation at lag 1 in the groundwater-
level time-series of this site. These diagnostic checks were
employed for all of the 17 sites, and the results indicated that
all 17 MLR models satisfy the basic assumptions of MLR
technique.
In addition to the basic assumptions, the degree of
multicollinearity was checked using the variance inflation
factor (VIF) and tolerance. As a rule of thumb, a variable
with a VIF value of greater than 5, or tolerance <0.2,
corresponds to high multicollinearity (Statistica 2001).
Table 6 shows the check for multicollinearity at site GH-
4.5 as an example, which reveals that the magnitude of
VIF is less than 5 and also the tolerance is greater than
0.2 for all the independent variables used in developing
the MLR models, and hence no multicollinearity exists
in the independent variables used. Similarly, the multi-
collinearity check was employed for all 17 sites, and it was
found that the multicollinearity is absent in the independent
variables used for developing the MLR models. Thus, the
developed MLR models are technically acceptable for
predicting groundwater levels at almost all the sites, except
for site C-7.

Optimal neural network parameters of the ANN models

The optimal number of hidden neurons, together with the
corresponding number of generations, population size and
fitness values for the 17 ANN models, is summarized
in Table 7. Clearly, the optimal number of hidden
Fig. 12 Scatter plots of observed and predicted groundwater levels neurons varies from 6 (sites E-2, E-5 and G-2) to 14
by ANN and MLR models with ± 2 % error band at site GH-4.5: a (site A-2) for the ANN architecture selected in this
for the training period (1999–2002); b for the testing period (2003–
2004) (r correlation coefficient between observed and simulated study. It is also evident from this table that the number of
groundwater-level elevation) generations required to obtain optimal neural network
parameters for the 17 ANN models varies from 1 (sites B-
3, C-2, D-6, E-5, F-6, G-2, H-2, H-5 and I-2) to 2 (sites A-2,
of the relationship between dependent and independent C-7, E-2, E-4, F-1, GH-4.5, H-3 and H-4), with the fitness
variables; (2) independence of the errors (no autocorrela- values ranging between 0.0016 m2 (site I-2) and 0.059 m2
tion); (3) homoscedasticity (constant variance) of the (site D-6).
errors with respect to the predicted values; and (4)
normality of the error distribution. The plots of the
standardized residuals (model errors) and predicted water Table 9 Percentage error bands for the ANN and MLR models of
tables at site GH-4.5 are shown in Fig. 4, in order to check 14 sites
the assumption of homoscedasticity. This figure reveals
that there is no increasing spread of residuals from left to Error band with respect to the 1:1 line Sites
right, and hence the assumption of constant variance of ± 2% A-2, C-7, D-6 and E-5
errors is satisfied. The assumption of normality was ± 5% C-2, E-2 and H-3
examined with the help of histogram of residuals ± 10 % F-1, F-6, G-2 and H-4
± 20 % H-2, H-5 and I-2
(Fig. 5) and normal probability plots (Fig. 6). In Fig. 5,

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1882

Fig. 13 Comparison of prediction errors of groundwater levels for ANN and MLR models at: a site B-3; b site E-4; c site GH-4.5; all for
the testing period (2003–2004) (Linear line represents best linear fit trendline for ANN and MLR models)

Comparative performance of the ANN and MLR models corresponding MLR models during both training and testing
periods by using MAE, RMSE, r2 and NSE goodness-of-fit
Quantitative evaluation statistics. The results of this comparison are presented in Table 8.
The performances of the 17 site-specific ANN models in Like the training results, the testing results also indicated
predicting groundwater levels were compared with those of the appreciably high values of r2 and NSE and considerably low

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 Table 10 Summary of sensitivity index values and level of sensitivity (rank) for different inputs at the 17 sites
Input Values of sensitivity index (Sk) and rank (R) at different sites
A-2 B-3 C-2 C-7 D-6 E-2 E-4 E-5 F-1 F-6 G-2 GH-4.5 H-2 H-3 H-4 H-5 I-2
Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R Sk R
0.052 2 0.035 3 0.304 1 0.045 3 0.049 3 0.053 2 0.333 1 0.034 3 0.303 1 0.155 1 0.161 1 0.166 1 0.121 1 0.048 3 0.045 3 0.132 1 0.028 3

Hydrogeology Journal (2013) 21: 1865–1887

GWt–1
GWt–2 0.061 2 – – 0.263 1 – – 0.062 2 0.070 2 0.070 2 0.051 2 0.066 2 0.073 2 0.033 3 0.010 4 0.022 3 0.008 4 0.007 4 0.023 3 0.052 2
Rainfall 0.073 2 0.036 3 0.124 1 0.004 5 0.025 3 0.024 3 0.174 1 0.037 3 0.179 1 0.073 2 0.148 1 0.075 2 0.020 3 0.022 3 0.040 3 0.047 3 0.007 4
Rainfalt–1 0.037 3 0.014 3 0.101 1 0.026 3 0.002 5 0.003 5 0.034 3 0.099 2 0.032 3 0.005 5 0.124 1 0.220 1 0.000 5 0.025 3 0.026 3 0.005 5 0.023 3
Rainfallt–2 0.012 3 0.004 5 0.069 2 0.005 4 0.004 5 0.004 5 0.076 2 0.022 3 0.077 2 0.019 3 0.046 3 0.142 1 0.044 3 0.007 4 0.040 3 0.008 4 0.069 2
Temperature 0.044 3 0.007 4 0.771 1 0.131 1 0.102 1 0.105 1 0.259 1 0.313 1 0.271 1 0.184 1 0.164 1 0.267 1 0.064 2 0.147 1 0.021 3 0.070 2 0.149 1
Tempt–1 0.262 1 0.038 3 0.324 1 0.029 3 0.076 2 0.078 2 0.341 1 0.042 3 0.341 1 0.172 1 0.173 1 0.820 1 0.021 3 0.079 2 0.087 2 0.080 2 0.041 3
Tempt–2 0.169 1 0.013 3 0.444 1 0.030 3 0.345 1 0.349 1 0.480 1 0.014 3 0.482 1 0.117 1 0.153 1 0.193 1 0.009 4 0.043 3 0.010 4 0.116 1 0.091 2
River stage 0.106 1 0.008 4 0.117 1 0.008 4 0.049 3 0.047 3 0.092 2 0.093 2 0.084 2 0.019 3 0.085 2 0.006 4 0.017 3 0.050 3 0.020 3 0.011 3 0.011 3
River staget–1 0.084 2 0.014 3 0.132 1 0.034 3 0.057 2 0.057 2 0.037 3 0.032 3 0.043 3 0.031 3 0.040 3 0.234 1 0.009 4 0.023 3 0.017 3 0.047 3 0.046 3
River staget–2 0.027 3 0.016 3 0.227 1 0.004 5 0.078 2 0.079 2 0.071 2 0.021 3 0.068 2 0.079 2 0.038 3 0.209 1 0.016 3 0.004 5 0.058 3 0.018 3 0.021 3
D1 (Jan–Feb) 0.145 1 0.031 3 0.090 2 0.020 3 0.070 2 0.070 2 0.230 1 0.045 3 0.228 1 0.011 3 0.024 3 0.178 1 0.009 4 0.036 3 0.023 3 0.015 3 0.004 5
D2 (Feb–Mar) 0.040 3 0.017 3 0.093 2 0.021 3 0.226 1 0.229 1 0.099 2 0.041 3 0.104 1 0.054 2 0.027 3 0.326 1 0.009 4 0.057 2 0.058 2 0.054 2 0.054 2
D3 (Mar–Apr) 0.011 3 0.022 3 0.080 2 0.013 3 0.124 1 0.125 1 0.074 2 0.049 3 0.074 2 0.044 3 0.054 2 0.303 1 0.007 4 0.038 3 0.037 3 0.047 3 0.036 3
D4 (Apr–May) 0.067 2 0.000 5 0.033 3 0.025 3 0.157 1 0.159 1 0.104 1 0.103 1 0.111 1 0.053 2 0.141 1 0.094 2 0.013 3 0.051 2 0.056 2 0.025 3 0.067 2
D5 (May–Jun) 0.051 2 0.008 4 0.165 1 0.047 3 0.114 1 0.108 1 0.052 2 0.094 2 0.037 3 0.061 2 0.008 4 0.184 1 0.019 3 0.004 5 0.083 2 0.010 4 0.022 3
D6 (Jun–Jul) 0.002 5 0.003 5 0.161 1 0.054 2 0.173 1 0.1741 1 0.086 2 0.029 3 0.091 2 0.042 3 0.057 2 0.224 1 0.001 5 0.009 4 0.040 3 0.004 5 0.010 4
D7 (Jul–Aug) 0.042 3 0.008 4 0.251 1 0.070 2 0.056 2 0.055 2 0.045 3 0.035 3 0.040 3 0.094 2 0.082 2 0.331 1 0.015 3 0.028 3 0.059 2 0.002 5 0.009 4
D8 (Aug–Sep) 0.034 3 0.016 3 0.227 1 0.073 2 0.070 2 0.072 2 0.161 1 0.089 2 0.168 1 0.070 2 0.055 2 0.259 1 0.001 5 0.031 3 0.078 2 0.033 3 0.006 4
D9 (Sep–Oct) 0.009 4 0.013 3 0.012 3 0.037 3 0.079 2 0.079 2 0.095 2 0.024 3 0.091 2 0.097 2 0.029 3 0.069 2 0.012 3 0.028 3 0.047 3 0.064 2 0.075 2
D10 (Oct–Nov) 0.003 5 0.004 5 0.065 2 0.015 3 0.031 3 0.030 3 0.014 3 0.080 2 0.010 4 0.027 3 0.049 3 0.133 1 0.027 3 0.009 4 0.051 2 0.095 2 0.031 3
D11 (Nov–Dec) 0.003 5 0.008 4 0.038 3 0.007 4 0.052 2 0.052 2 0.110 1 0.027 3 0.118 1 0.009 4 0.046 3 0.007 4 0.007 4 0.007 4 0.056 2 0.064 2 0.029 3

Rank 1: very high sensitivity (Sk >0.1); Rank 2: high sensitivity (Sk =>0.05−0.1); Rank 3: moderate sensitivity (Sk =>0.01–0.05); Rank 4: low sensitivity (Sk =>0.005–0.01); Rank 5: very
low sensitivity (Sk ≤0.005). The rank 5 values are italicized
D1–D11 indicate seasonal dummy variables

DOI 10.1007/s10040-013-1029-5
1883
1884

Fig. 14 Sensitivity of the MLP-LM model to the input variables for: a site E-4; b site G-2. In these graphs, D1–D11 indicate seasonal
dummy variables

values of MAE and RMSE for the ANN models compared to NSE=0.998) and site H-2 (MAE=0.017 m, RMSE=0.02 m,
those for the MLR models. The ANN (MLP-LM) models r2 =0.997, NSE=0.997), whereas the prediction accuracy at
have the highest overall prediction accuracy at site GH-4.5 site C-7 is the lowest (MAE=0.021 m, RMSE=0.036 m,
(MAE = 0.015 m, RMSE = 0.03 m, r 2 = 0.998 and r2 =0.832 and NSE=0.828), but it is still acceptable,

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1885
2
because both r and NSE are >0.8. Here, it is interesting ANN model show an increasing trend with time, while the
to note that the MLR model fails to predict reliable residuals of the MLR model have a decreasing trend.
groundwater levels at site C-7, while the ANN model is Moreover, the residuals of ANN models for site GH-4.5
capable of providing reasonable groundwater-level predic- exhibit a slightly decreasing trend with zero residual line,
tions at this site. whereas the residuals of MLR models have a distinct
increasing trend with time. Similarly, the residual analyses
Qualitative evaluation for the remaining 14 sites were performed, and it was
As far as the results of graphical indicators are concerned, a found that the residuals of the ANN models do not have a
comparison of observed and simulated/predicted groundwa- trend with time for almost all the sites, thereby indicating
ter levels by ANN and MLR models is shown in Figs. 7a,b, greater prediction accuracy for the ANN models, except
8, 9a,b at sites B-3, E-4 and GH-4.5 for training and testing for sites A-2, C-2, D-6, F-1, F-6, G-2 and H-5. On the
periods, as examples. It is apparent from these figures that contrary, the residuals of the MLR models show a distinct
the simulated groundwater levels obtained by ANN models trend at a majority of the sites, except for sites C-7, E-5,
match reasonably well with the observed groundwater- and F-6, thereby suggesting a lower prediction accuracy
levels, as compared to those obtained by MLR models. for the MLR models at most sites.
Besides the simultaneous plots, the match between observed Considering the quantitative and qualitative performance
groundwater levels and the groundwater levels predicted by indicators, it was found that the ANN technique performs
both the models was also investigated by scatter plots with much better than the MLR technique. The inferior perfor-
1:1 lines and error bands. For instance, the scatter plots with mance of the MLR models could be attributed to the fact that
1:1 lines and error bands at sites B-3, E-4, and GH-4.5 are the MLR modeling technique is based on the simple least-
illustrated in Figs. 10a,b, 11, 12a,b. In these figures, the square method, whereas the ANN modeling technique is
parallel lines indicate percentage upper and lower error based on artificial intelligence, which mimics the functioning
bands with respect to the 1:1 line. Obviously, the simulated of the human brain by using sophisticated neural networks
groundwater levels yielded by the ANN and MLR models and non-linear activation functions for establishing input–
fall within ±1 % error at site B-3, ±2 % error at site GH-4.5 output relationships.
and ±10 % error at site E-4. The error bands for the
remaining 14 sites are summarized in Table 9. It is apparent
from Table 9 that the groundwater levels simulated by both Results of sensitivity analysis
ANN and MLR techniques at sites A-2, C-7, D-6 and E-5 are Table 10 presents the results of the sensitivity analyses for the
within a much lower tolerance limit (i.e., ±2 % error band). ANN model, and the variations of sensitivity index values
The simulated groundwater levels by both the models at sites with the inputs at sites E-4 and G-2 are shown in Fig. 14a,b, as
C-2, E-2 and H-3 lie within an error band of ±5 %, whereas an example. It can be seen from Fig. 14a, that at site E-4,
they are within an error band of ±10 % at sites F-1, F-6, G-2 almost all the inputs have reasonably high values of
and H-4. However, both the models yield groundwater levels sensitivity index (the rank ranging between 1 and 3), and
at sites H-2, H-5 and I-2 with larger errors (i.e., within 20 %), hence they significantly affect the groundwater-level at this
which suggests somewhat greater uncertainty in the predict- site. However, at site G-2, input D5 has a low value of
ed groundwater levels at these three sites. Nevertheless, a sensitivity index (rank 4) as compared to those for the
close examination of Figs. 10a,b, 11, 12a,b reveals that the remaining inputs (rank=1–3; Fig. 14b). Therefore, input D5
spreading of the groundwater levels simulated by the MLR may be discarded at site G-2. Thus, it is evident from Table 10
models with respect to the 1:1 line is larger than the that at seven sites (C-2, E-4, E-5, F-1, G-2, GH-4.5, and H-4),
groundwater levels simulated by the ANN models for both all the inputs have a strong influence on groundwater-levels,
training and testing periods. Similar trends were found for showing greater sensitivity levels (rank <5). However, for the
the remaining sites. This suggests that the prediction remaining ten sites, only a small number of inputs have very
accuracy of the ANN technique is better than the MLR low sensitivity (rank=5), which are highlighted in Table 10.
technique. Therefore, the inputs having ‘very high’ to ‘moderate’
Furthermore, the residual analysis was performed by sensitivity should be considered with greater accuracy so as
plotting prediction errors against time, in order to detect to ensure reliable prediction of groundwater levels by the
any trend/spread present in the process, and also to assess ANN model.
the accuracy of the groundwater levels predicted by both
techniques. As an example, the residual plots for the ANN
and MLR models of sites B-3, E-4 and GH-4.5 during Conclusions
testing period are shown in Fig. 13a–c. It is obvious from
Fig. 13a that during testing period, the residuals of the This study was carried out to examine the potential of two
ANN model for site B-3 are evenly distributed with data-driven approaches, MLR and ANN, for simulating/
respect to the zero reference line, which implies that they predicting transient groundwater levels over a groundwa-
exhibit almost no trend with respect to time, i.e., the ter basin using relevant real-world data. It employed the
residuals are independent in nature. In contrast, the standard protocols of ANN and MLR modeling as well as
residuals of the MLR model for site B-3 follow an all the pertinent and influential input variables to achieve
increasing trend with time. At site E-4, the residuals of the this goal. Altogether, 17 site-specific MLR and ANN

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

1886
models were developed, using all relevant and influential ASCE (ASCE Task Committee on Application of Artificial Neural
input data—namely, rainfall, ambient temperature, river Networks in Hydrology) (2000b) Artificial neural networks in
hydrology, part II: hydrologic application. J Hydrol Eng ASCE
stage, 11 seasonal dummy variables, and influential lags of 5(2):124–137
rainfall, ambient temperature, river stage and water level. Bishop C (1995) Neural networks for pattern recognition.
The genetic algorithm (GA) technique was applied to Clarendon, Oxford, 504 pp
optimize the number of hidden neurons of the ANN models. Castillo PA, Merelo JJ, Prieto A, Rivas V, Romero G (2000) GProp:
global optimization of multilayer perceptrons using GAs.
The performance of the site-specific MLR and ANN models Neurocomputing 35:149–163
developed for the 17 sites was assessed both quantitatively Coppola EA, Szidarovszky F, Poulton MM, Charles E (2003)
and qualitatively by using appropriate statistical and graph- Artificial neural network approach for predicting transient water
ical indicators. levels in a multilayered groundwater system under variable
state, pumping, and climate conditions. J Hydrol Eng ASCE
Analysis of the results of MLR and ANN modeling 8(6):348–360
indicated that the developed ANN models provide more Coppola EA, Rana AJ, Poulton MM, Szidarovszky F, Uhl VW
accurate prediction of groundwater levels than the MLR (2005) A neural network model for predicting aquifer water
models at almost all the 17 sites, with considerably high level elevations. Ground Water 43(2):231–241
values of r2 and NSE and much lower values of MAE and Coulibaly P, Anctil F, Aravena R, Bobee B (2001) Artificial neural
network modeling of water table depth fluctuations. Water
RMSE. This is due to the fact that the MLR models have Resour Res 37(4):885–896
inherently limited prediction ability, particularly in the Daliakopoulos IN, Coulibaly P, Tsanis IK (2005) Groundwater level
presence of nonlinear relationships and/or noisy data. forecasting using artificial neural network. J Hydrol 309:229–240
Therefore, the ANN models developed in this study can be Dekker SC, Bouten W, Schaap MG (2001) Analyzing forest
transpiration model errors with artificial neural networks. J
regarded as a robust, powerful and reliable tool for simulating Hydrol 246:197–208
the dynamic and complex behavior of groundwater in a basin. Feng S, Kang S, Huo Z, Chen S, Mao X (2008) Neural networks to
Nevertheless, the performance of the MLR models is simulate regional ground water levels affected by human
reasonably good at most sites, except for site C-7, though activities. Ground Water 46:80–90
Gerken WC, Purvis LK, Butera RJ (2006) Genetic algorithm for
their residuals indicated a trend at a majority of the sites under optimization and specification of a neuron model.
study. The MLR technique has important practical advantages Neurocomputing 69:1039–1042
such as the fact that implementation is much easier and less Goldberg DE (1989) Genetic algorithms in search, optimization and
time-consuming and labor-intensive, and it requires minimal machine learning. Addison-Wesley, New York, pp 1–145
skill compared to the ANN technique. Consequently, the Haan CT (2002) Statistical methods in hydrology, 2nd edn. The
Iowa State University Press, Ames, IA, 496 pp
MLR technique can serve as an alternative and cost-effective Hagan MT, Menhaj MB (1994) Training feedforward networks with
tool for groundwater-level simulation/prediction in the the Marquardt algorithm. IEEE Trans Neural Network 5:989–993
situation where professional skill and time are the constraints, Hagan MT, Demuth HB, Beale MH (1996) Neural network design.
and the field data are favorable (i.e., of good quality and free PWS, Boston, MA
Hani A, Lallahem S, Mania J, Djabri L (2006) On the use of finite
from noise). The methodology presented in this study can be difference and neural-network models to evaluate the impact of
easily applied to other parts of the world as well, irrespective underground water overexploitation. Hydrol Processes
of the hydrogeological settings. Thus, the methodology and 20:4381–4390
findings demonstrated in this study are very useful to the Haykin S (1994) Neural Networks. Macmillan College Publishing,
New York
research community of both developing and developed Heuvelmans G, Muys B, Feyen J (2006) Regionalisation of the
nations involved in groundwater management or protection, parameters of a hydrological model: comparison of linear regression
as well as to practising hydrogeologists. models with artificial neural nets. J Hydrol 319:245–265
Hodgson FDI (1978) The use of multiple linear regression in
simulating ground-water level responses. Ground Water
Acknowledgements The authors are grateful to the staff of the 16(4):249–253
Kochi Prefecture Office and of other concerned local offices for Hornik K, Stinchcombe M, White H (1989) Multilayer feedforward
providing necessary data and kind assistance during field investi- networks are universal approximators. Neural Networks 2:359–
gations. The financial support to the second author from the Japan 366
Society for the Promotion of Science (JSPS) is also gratefully Izady A, Davary K, Alizadeh A, Ghahraman B, Sadeghi M,
acknowledged. The authors are also very much obliged to the three Moghaddamnia A (2012) Application of “panel-data” modeling
reviewers and the editor for their helpful comments. to predict groundwater levels in the Neishaboor Plain, Iran.
Hydrogeol J 20:435–447
Jha MK, Chikamori K, Kamii Y, Yamasaki Y (1999) Field
References investigations for sustainable groundwater utilization in the
Konan basin. Water Resour Manag 13:443–470
Adeloye AJ (2009) Multiple linear regression and artificial neural Jones AJ (1993) Genetic algorithms and their applications to the
network models for generalized reservoir storage–yield–reliabil- design of neural networks. Neural Comput Appl 1:32–45
ity function for reservoir planning. J Hydrol Eng ASCE Krishna B, Rao YRS, Vijaya T (2008) Modeling groundwater levels
14(7):731–738 in an urban coastal aquifer using artificial neural networks.
Affandi AK, Watanabe K, Tirtomihardjo H (2007) Application of an Hydrol Processes 22:1180–1188
artificial neural network to estimate groundwater level fluctua- Lallahem S, Mania J, Hani A, Najjar Y (2005) On the use of neural
tion. J Spat Hydrol 7(2):23–46 networks to evaluate groundwater levels in fractured media. J
ASCE (ASCE Task Committee on Application of Artificial Neural Hydrol 307:92–111
Networks in Hydrology) (2000a) Artificial neural networks in Legates DR, McCabe GJ (1999) Evaluating the use of goodness-of-
hydrology, part I: preliminary concepts. J Hydrol Eng ASCE fit measures in hydrologic and hydroclimatic model validation.
5(2):115–123 Water Resour Res 35(1):233–241

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

 1887
Looney CG (1996) Advances in feed-forward neural networks: Nourani V, Mogaddam AA, Nadiri AO (2008) An ANN-based
demystifying knowledge acquiring black boxes. IEEE Trans model for spatiotemporal groundwater level forecasting. Hydrol
Knowledge Data Eng 8(2):211–226 Processes 22:5054–5066
Maier HR, Dandy GC (1998) Understanding the behavior and Press WH, Teukolski SA, Vetterling WT, Flannery BP (1992)
optimizing the performance of back-propagation neural networks: Numerical recipes in Fortran, 2nd edn. Cambridge University
an empirical study. Environ Modeling Softw 13(2):179–191 Press, New York, pp 675–683
Maier HR, Dandy GC (2000) Neural networks for the prediction Principe JC, Euliano NR, Lefebvre WC (2000) Neural and adaptive
and forecasting of water resources variables: a review of systems: fundamentals through simulations. Wiley, New York, 656 pp
modeling issues and applications. Environ Modeling Softw Reed P, Minsker B, Goldberg DE (2000) Designing a competent
15:101–124 simple genetic algorithm for search and optimization. Water
Makridakis S, Wheelwright SC, Hyndman RJ (2008) Forecast- Resour Res 36(12):3757–3761
ing methods and applications, 3rd edn. Wiley, Singapore, Rumelhart DE, Durbin R, Golden R, Chauvin Y (1995)
656 pp Backpropagation: the basic theory. In: Chauvin Y, Rumelhart
Mayer DG, Belward JA, Burrage K (2001) Robust parameter DE (eds) Backpropagation: theory, architectures, and applica-
settings of evolutionary algorithms for the optimization of tions. Erlbaum, Hillsdale, NJ, pp 1–34
agricultural systems models. Agric Syst 69:199–213 Sethi RR, Kumar A, Sharma SP, Verma HC (2010) Prediction of
McCuen RH, Rawls WJ, Whaley BL (1979) Comparative evalua- water table depth in a hard rock basin by using artificial neural
tion of statistical methods for water supply forecasting. Water network. Int J Water Resour Environ Eng 2(4):95–102
Resour Bull 15(4):935–947 Shamseldin AY (1997) Application of a neural technique to rainfall-
Mohanty S, Jha MK, Kumar A, Sudheer KP (2009) Artificial runoff modeling. J Hydrol 199:272–294
neural network modeling for groundwater level forecasting in Sietsma J, Dow RJF (1991) Creating artificial neural networks that
a river island of eastern India. Water Resour Manag 24(9):1845– generalize. Neural Networks 4:67–69
1865 Sinnakaudan SK, Ghani AA, Ahmad MSS, Zakaria NA (2006)
Moré JJ (1978) The Levenberg-Marquardt algorithm: implementation Multiple linear regression model for total bed material load
and theory. In: Watson GA (ed) Lecture notes in mathematics 630. prediction. J Hydraul Eng ASCE 132(5):521–528
Springer, Berlin, pp 5–116 Statistica (2001) STATISTICA: system reference. StatSoft, Tulsa,
Nayak PC, Rao YRS, Sudheer KP (2006) Groundwater level OK, 1098 pp
forecasting in a shallow aquifer using artificial neural network Trichakis IC, Nikolos IK, Karatzas GP (2009) Optimal selection of
approach. Water Resour Manag 20:77–90 artificial neural network parameters for the prediction of a
Nikolos IK, Stergiadi M, Papadopoulou MP, Karatzas GP (2008) karstic aquifer’s response. Hydrol Processes 23:2956–2969
Artificial neural networks as an alternative approach to Uddameri V (2007) Using statistical and artificial neural network
groundwater numerical modeling and environmental design. models to forecast potentiometric levels at a deep well in South
Hydrol Processes 22(17):3337–3348 Texas. Environ Geol 51:885–895

Hydrogeology Journal (2013) 21: 1865–1887 DOI 10.1007/s10040-013-1029-5

View publication stats