1 s2.0 S092523122200282X Main

Neurocomputing 489 (2022) 271–308
Contents lists available at ScienceDirect
Neurocomputing
journal homepage: www.elsevier.com/locate/neucom
Groundwater level prediction using machine learning models:

A comprehensive review
Hai Tao a,b,c, Mohammed Majeed Hameed d, Haydar Abdulameer Marhoon e,f, Mohammad Zounemat-
Kermani g, Salim Heddam h, Sungwon Kim i, Sadeq Oleiwi Sulaiman j, Mou Leong Tan k, Zulfaqar Sa’adi l,
Ali Danandeh Mehr m, Mohammed Falah Allawi j, S.I. Abba n,o, Jasni Mohamad Zain c,p, Mayadah W. Falah q,
Mehdi Jamei r, Neeraj Dhanraj Bokde s, Maryam Bayatvarkeshi t, Mustafa Al-Mukhtar u,
Suraj Kumar Bhagat v, Tiyasha Tiyasha v, Khaled Mohamed Khedher w,x, Nadhir Al-Ansari y,
Shamsuddin Shahid z, Zaher Mundher Yaseen aa,ab,⇑
a
School of Electronics and Information Engineering, Ankang University, China
b
School of Computer Sciences, Baoji University of Arts and Sciences, Shaanxi, China
c
Institute for Big Data Analytics and Artificial Intelligence (IBDAAI), Universiti Teknologi MARA, Shah Alam, Selangor, Malaysia
d
Department of Civil Engineering, Al-Maarif University College, Ramadi, Iraq
e
Information and Communication Technology Research Group, Scientific Research Center, Al-Ayen University, Thi-Qar, Iraq
f
College of Computer Sciences and Information Technology, University of Kerbala, Karbala, Iraq
g
Department of Water Engineering, Shahid Bahonar University of Kerman, Kerman, Iran
h
Faculty of Science, Agronomy Department, Hydraulics Division University, 20 Août 1955, Route EL HADAIK, 26 Skikda, BP, Algeria
i
Department of Railroad Construction and Safety Engineering, Dongyang University, Yeongju 36040, South Korea
j
Dams and Water Resources Engineering Department, College of Engineering, University of Anbar, Ramadi, Iraq
k
GeoInformatic Unit, Geography Section, School of Humanities, Universiti Sains Malaysia, 11800 USM, Penang, Malaysia
l
Centre for Environmental Sustainability and Water Security (IPASA), School of Civil Engineering, Faculty of Engineering, Universiti Teknologi Malaysia, 81310 UTM Sekudai, Johor,
Malaysia
m
Civil Engineering Department, Antalya Bilim University, Antalya, Turkey
n
Interdisciplinary Research Center for Membrane and Water Security, King Fahd University of Petroleum and Minerals, Dhahran 31261, Saudi Arabia
o
Faculty of Engineering, Department of Civil Engineering, Baze University, Abuja, Nigeria
p
Faculty of Computer and Mathematical Sciences, University Technology MARA, Malaysia
q
Building and Construction Engineering Technology Department, AL-Mustaqbal University College, Hillah 51001, Iraq
r
Faculty of Engineering, Shohadaye Hoveizeh Campus of Technology, Shahid Chamran University of Ahvaz, Dashte Azadegan, Iran
s
Department of Civil and Architectural Engineering, Aarhus University, Denmark
t
Department of Geography and Environmental Management, the faculty of Environment, the university of Waterloo, Canada
u
Civil Engineering Department. University of Technology, Baghdad, Iraq
v
Faculty of Civil Engineering, Ton Duc Thang University, Ho Chi Minh City, Viet Nam
w
Department of Civil Engineering, College of Engineering, King Khalid University, Abha 61421, Saudi Arabia
x
Department of Civil Engineering, High Institute of Technological Studies, Mrezgua University Campus, Nabeul 8000, Tunisia
y
Civil, Environmental and Natural Resources Engineering, Lulea University of Technology, 97187 Lulea, Sweden
z
School of Civil Engineering, Faculty of Engineering, Universiti Teknologi Malaysia (UTM), 81310 Johor Bahru, Malaysia
aa
Adjunct Research Fellow, USQ’s Advanced Data Analytics Research Group, School of Mathematics Physics and Computing, University of Southern Queensland, QLD 4350, Australia
ab
New era and Development in Civil Engineering Research Group, Scientific Research Center, Al-Ayen University, Thi-Qar 64001, Iraq
a r t i c l e i n f o a b s t r a c t
Article history: Developing accurate soft computing methods for groundwater level (GWL) forecasting is essential for
Received 12 July 2021 enhancing the planning and management of water resources. Over the past two decades, significant pro-
Revised 7 January 2022 gress has been made in GWL prediction using machine learning (ML) models. Several review articles have
Accepted 10 March 2022
been published, reporting the advances in this field up to 2018. However, the existing review articles do
Available online 14 March 2022
not cover several aspects of GWL simulations using ML, which are significant for scientists and practition-
ers working in hydrology and water resource management. The current review article aims to provide a
Keywords:
clear understanding of the state-of-the-art ML models implemented for GWL modeling and the mile-
State-of-the-art
Machine learning
stones achieved in this domain. The review includes all of the types of ML models employed for GWL
⇑ Corresponding author at: Adjunct Research Fellow, USQ’s Advanced Data Analytics Research Group, School of Mathematics Physics and Computing, University of Southern
Queensland, QLD 4350, Australia.
E-mail address: [email protected] (Z.M. Yaseen).
https://doi.org/10.1016/j.neucom.2022.03.014
0925-2312/Ó 2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
H. Tao, Mohammed Majeed Hameed, Haydar Abdulameer Marhoon et al. Neurocomputing 489 (2022) 271–308
Groundwater level modeling from 2008 to 2020 (138 articles) and summarizes the details of the reviewed papers, including
Input parameters the types of models, data span, time scale, input and output parameters, performance criteria used, and
Prediction performance the best models identified. Furthermore, recommendations for possible future research directions to
Catchment sustainability
improve the accuracy of GWL prediction models and enhance the related knowledge are outlined.
Ó 2022 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://
creativecommons.org/licenses/by/4.0/).
1. Introduction studied. Furthermore, Fig. 1 highlights the four major countries

which have done extensive GWL modeling-related studies,
1.1. Research background whereas the black color zone reveals the areas where the applica-
tion of AI has not yet gained in popularity. Around 70% of areas
Groundwater resources, as one of the most valuable and impor- have not yet used GWL, as many do not need GW-related studies,
tant sources of water in the world, play a direct and crucial role in due to a sufficient amount of surface water or less habitants, such
various aspects of human lives, such as agriculture, industrial as in polar areas, Russia, and so on. Moreover, some underdevel-
development, and potable water supply [1,2]. In addition, the indi- oped countries, such as Africa, and some parts of Asia and North
rect effects of groundwater resources on the environment and America, may not have explored AI techniques yet. As per Fig. 2,
communities are undeniable. The groundwater level (GWL) is a there has been a significant increase in studies in this field in the
direct and simple measure of groundwater availability and accessi- last few years; however, more studies should be done, based on
bility. Having a proper understanding of the past, current, and different geographical locations, to test the efficiency of the pro-
future situations of GWL can provide policy-makers and practition- posed models. The usability and reliability of AI models in dealing
ers in water sectors with better insight and perception to develop with complex and high-dimensional engineering problems have
strategies for the planning and management of water resources, to been proven in the last few decades [13–15]. AI consists of multi-
ensure sustainable socioeconomic development [2]. However, dimensional systems combining various mathematical and statisti-
GWL consists of an integrated response to several climatic, topo- cal components and arithmetic and heuristic algorithms. AI has
graphic, and hydrogeological factors and their interactions, which been extensively employed in different fields of science, engineer-
makes the simulation of GWL a challenging task [3,4]. ing design, energy, robotics, and economics [16–18]. It has also
Numerous studies using different simulation approaches have been intensively used for solving various civil and environmental
been conducted for the quantitative and qualitative prediction of engineering problems [19,15]. Some examples include soft com-
GWL. These methods cover a wide range of physically based con- puting techniques [20], Machine Learning (ML) methods [21–23],
ceptual models, experimental models [5–7], and numerical mod- probabilistic analysis [24], and Fuzzy-based systems [25]. In recent
els. Modeling groundwater using numerical models consists of years, more attention has been paid to the successful use of AI in
several approaches, such as finite difference [8], finite volume different hydrological fields, including water resources [26], sur-
[9], finite element [10], and element-free [11] methods. Even face and groundwater hydrology [27], sediment contamination
though these classical models are robust and reliable, the precision [28], and hydraulics [29].
and accuracy of numerical models are confined by several factors,
such as their high dependency on large volumes of data related to
aquifer properties, the geology of the porous media, and basement 1.2. Research significance
topography [12]. Moreover, properly demarcating domain bound-
aries, defining an efficient grid size for solving the associated dif- Proper measurement, nowcasting, and forecasting of GWL in
ferential equations, and calibrating/validating the executed aquifers are highly important for the sensible management of
model have made numerical modeling a complex and sophisti- groundwater resources [30]. Monitoring GWL can provide hydrol-
cated task. ogists and hydrogeologists valuable information to understand the
In last two decades, artificial intelligence (AI) models have been short- and long-term variations in groundwater availability. The
widely used to overcome the drawbacks of conventional numerical ability of AI models to simulate and predict GWL without requiring
models for GWL simulation. Fig. 1 presents the goal map, depicting deep and comprehensive knowledge of the underlying topograph-
the two major pieces of information, one being the most studied ical and hydro-geophysical parameters makes them appealing
geographical locations and other those which have not yet been methods compared to physically based and numerical methods
Fig. 1. Map representation of sampling location of GWL data all over the globe with Fig. 2. Arithmetical conceptualization of growth observed in GWL research using AI
specified area with no related research on GWL modeling using AI models. based model during 2008–2020.
272
[31]. A large volume of studies have already investigated and It is a major sort of AI applications, which is capable of handling
reported the applicability of AI in modeling GWL over the last complex issues which are difficult, according to statistical and
two decades [32]. Most of the early works included simple and human standards [19]. Furthermore, ANNs have efficient abilities
standard AI methods, such as perceptron Artificial Neural Net- to approximate functions that are commonly unknown or to pre-
works (ANNs) [33]. However, in the last decade, the application dict future values based on potentially noisy time-series data
of a variety of ML models for GWL simulation has been witnessed; [50–52]. The structure of an ANN is comprised of several simple
examples include different types of ANNs [34], fuzzy-based models elements working in parallel. The determination of the function
[35], Support Vector Machines (SVMs) [36], tree-based models of an ANN mainly depends on the connections between elements,
[37], Genetic Programming (GP) [38], and Gene Expression Pro- as in natural processing [53]. In general, an ANN is comprised of
gramming (GEP) models [39]. three layers, including the input layer (which is used to input the
Most recently, along with the application of novel AI models, variables) and the output layer (which is used to compute the
including Deep Learning (DL) [40], Extreme Learning Machine desired target) [54]. The hidden layer is an important component
(ELM) [41], and Long Short-Term Memory (LSTM) [42], novel of an ANN, due to its location between the input layer and output
strategies, such as integrated and hybrid AI models [43], ensemble layer, where the neurons receive a set of weighted inputs and,
learning [44], and AI-GIS (Artificial Intelligence-Geographic Infor- hence, generate an output by applying a certain activation function
mation System)-based models [45], have been implemented for [15]. The information transfers from one layer to another through
modeling GWL. Rajaee et al. [46], for instance, studied 67 journal neurons (processing elements). An activation function is always
papers and provided a bibliographic review of the applications of used, regardless of using an ANN with a single or several hidden
AI in GWL simulation and forecasting. Considering the outcomes layers. Feed-forward neural networks (FFNNs), which are often
of different classic AI methods, such as ANNs, Adaptive Neuro- called multilayer perceptrons (MLPs) are one of the most famous
fuzzy Inference System (ANFIS), SVM, GP, and hybrid AI methods, and powerful types of ANN and have been widely used for solving
the study concluded that AI methods can be successfully used to hydrological issues [55–58]. An FFNN has three layers, as shown in
model and forecast GWL in aquifers located in regions with differ- Fig. 3. In a classical FFNN, the initial weight and bias values are
ent geology and climate. Some studies have attempted to combine assigned randomly and, then, the algorithm starts to correct the
the advantages of AI and numerical methods to develop hybrid values, in order to minimize the loss function. Gradient descent
models. For example, Nourani and Mousavi [47] introduced a back-propagation and Levenberg–Marquardt back-propagation
hybrid AI-meshless model for modeling GWL. They used AI meth- algorithms have increasingly been employed in training FFNNs,
ods, such as ANN and ANFIS, for temporal modeling of GWL, while in order to optimize the magnitudes of the weights and biases
the meshless method was used for solving the governing differen- [59–62]. In a traditional FFNN, three parameters need to be consid-
tial equations to estimate the GWL in places with no observations ered, in order to accomplish more accurate predictions: (1) The
[48]. Chen et al. [49] carried out a comparative study using a finite number of hidden nodes and transfer functions; (2) the initial
difference numerical model versus three ML models, including weight and bias values; and (3) choosing a sufficient number of
ANNs and SVM, for simulating GWL. Comparing the general perfor- epochs [63].
mance of the two distinct approaches revealed that the ML models A large volume of literature is available on the application of
acted better than the numerical model. Nevertheless, they also ANN models to forecast GWL in different regions. Nair and Sindhu
mentioned the superiority of the finite difference method, due to [65] conducted a study for estimating the GWL in the Mamom river
its generalization ability in including the physical mechanism of basin in Trivandrum region, India. They used an ANN model based
the aquifer. on hydrological parameters to estimate the GWL in three wells
during monsoon and non-monsoon seasons. The predictive models
were constructed using only four meteorological factors as predic-
1.3. Research objectives
tors: Rainfall (Raf), potential Evapotranspiration (EVP), Tempera-
As complete descriptions and detailed analyses on the applica-
tion of ML models for GWL prediction are provided in the following
sections, giving more information on this matter herein-that is, as
the literature review-would be repetitive and unnecessary. It is
quite understandable that many hydrologists and hydrogeologists
have recognized the potential capability of ML models, in particu-
lar, for their use in GWL simulation. Even though there have been a
few comprehensive review studies published on the subject of
GWL modeling using ML models, such as the recent one conducted
by Rajaee et al. [46], this review article tries to fill in the literature
gap regarding the emergence and application of novel AI models in
groundwater simulation. Furthermore, the focus of the present
article is the recent developments, progresses, restrictions, and
shortcomings of advanced AI methods in dealing with GWL. Thus,
this article is aimed at researchers, groundwater engineers, envi-
ronmentalists, and hydrogeologists who find the prospects of AI
in the groundwater domain attractive.
2. Artificial intelligence models for GWL simulation
2.1. Application of artificial neural network models
An ANN is a computer system designed to mimic the manner in

which information is processed and analyzed by the human brain. Fig. 3. The structure of the classical ANN model [64].
273
ture (T), and Humidity (H%). Moreover, the Levenberg–Marquardt work (RBFNN). Quantitative assessments revealed that both mod-
(LM) algorithm was applied to train several ANN model sturctures, els can provide accurate GWL estimates; however, the RBFNN
in order to optimally choose the weight and bais values. Another model was more competent in forecasting GWL, compared to the
study in India used an ANN approach to estimate the monthly GM(1, 1) model. The comparable model, GM(1, 1), demonstrated
GWL at four sites located in south-east Punjab for the period of an inability to generate highly accurate estimations, especially in
2006 to 2013 [66], where Raf and preceding GWL were used as short-term forecasting, compared to the adopted model (RBFNN).
inputs. The results showed that the use of an ANN using the LM The study concluded that, despite the efficiency of the RBFNN
back-propagation algorithm provides more accurate predictions, model, it might still need to be improved to overcome the over-
compared to other algorithms. fitting issue.
In the African continent, Nouiri and Malek [67] developed an The process of understanding and predicting the fluctuations of
ANN model to estimate monthly GWL in four aquifers of the Neb- GWL is usually very complex, as several parameters play signifi-
hana watershed, located in Tunisia. Only three input parameters cant roles in determining the storage capacity of water in a certain
(Raf, antecedent GWL, and EVP) were used to develop the forecast- aquifer. Shamsuddin et al. [74] carried out an attempt to forecast
ing models. The ANN model gave higher prediction accuracy and daily GWL using the MLP technique in a tropical region, Jenderam
was also able to capture the dynamic fluctuations in piezometric Hilir in Selangor state of Malaysia. To achieve better estimation
levels. The study revealed that the monthly GWL depended mainly performance, different ANN structures were chosen, based on dif-
on monthly precipitation, EVP, and antecedent values of GWL. ferent input parameters including meteorological and hydrological
Iqbal et al. [68] developed an ANN model for forecasting daily factors. It is important to mention that the LM algorithm was uti-
GWL from three meteorological variables in a study area located lized to train the predictive models. The outcomes of the study
between the Ravi and Sutlej Rivers in Pakistan. The input parame- exhibited the robustness of Multiple linear regression (MLR) in
ters were Raf, T max ; T min , solar radiation, relative H%, wind speed forecasting the GWL, based on several statistical criteria. Addition-
(WS), area elevation, and polygon area. In order to select the most ally, the study presented a good relationship between the hydro-
accurate GWL prediction model, the authors applied different logical parameters and GWL.
types of ANN architectures with different transfer functions, hid- The extreme learning machine, an advanced version of an ANN,
den layers, and different percentages of data in the training, valida- was invented in 2006 and has gained good popularity, in recent
tion, and testing phases. LM back-propagation was used as a years, in solving water resource issues and groundwater estima-
learning algorithm. The study revealed that the proposed model tion tasks. The main structure of the classical ELM model is about
was able to estimate GWL more accurately, in terms of different the same as a single-layer FFANN model. However, the input
statistical criteria. Furthermore, the results revealed that the tan- weights and biases in the ELM algorithm are always assigned ran-
gent sigmoid transfer function was most efficient and the data domly and the output weights are calculated using the singular
division with 80% for training, 10% for validation, and 10% for test- value decomposition (SVD) method. ELM models have many
ing was more effective and optimistic, compared to other data advanced aspects, which make them superior to traditional ANN
divisions. models in solving complex engineering problems. ELM models
Guzman et al. [69] employed two predictive models based on are easier to train and have the characteristics of faster conver-
two approaches-support vector regression (SVR) and Recurrent gence, better generalization, and a lower chance of becoming stuck
Neural Network (RNN)-for the estimation of GWL for a well located in a local minimum, compared to other types of ANN models.
in Sunflower county, Mississippi, USA. Daily GWL was the only Moreover, the ELM algorithm can be trained quickly with minimal
variable used for developing both models. The authors concluded data [75]; hence, it has shown promising successes in several sec-
that both models produced preferable predictions, but the RNN tors of engineering, especially hydrological areas.
gave higher accuracy. Alizamir et al. [41] published a paper on modeling GWL using
Hong [70] presented a study to estimate hourly GWL using a hydrological and climatic data. The study developed an ELM model
feed-forward back-propagation neural network (FFBPNN) to and three other ML models (i.e., RBFNN, MLR, and autoregressive-
achieve two main aims: (1) prediction of next hour GWL and (2) moving-average; ARMA) to predict monthly GWL in the Shamil-
forecasting the fluctuations and changes in GWL between the cur- Ashekara Plain, Iran. The study found that the ELM model obtained
rent and one-lag-ahead GWL. The outcomes of the study illustrated much higher estimation accuracy than the other models. More-
that the proposed model managed to simulate the fluctuations in over, the proposed model showed better performance in predicting
GWL between a lag much more accurately than one-lag-ahead pre- multi-month GWL than the other employed models.
diction of GWL. A study has been carried out [76] for the prediction of GWL in
Kouziokas et al. [71] used an FFANN model to estimate the daily Canada using two different approaches: ELM and SVR models.
GWL from meteorological variables in the district of Montgomery Meteorological and hydrological data were used as inputs to
country in Pennsylvania, USA. Different ANN model structures develop both forecasting models. The obtained results showed
were investigated, with different transfer functions and learning the superiority of ELM over SVR in forecasting monthly GWL. A list
algorithms, such as Resilient Back-propagation, Scaled Conjugate of articles on the prediction of GWL using the ANN models are tab-
Gradient, LM, and BFGS Quasi-Newton. The simulated GWL values ulated in Table 1.
were found to be higher than those of comparable models found in
the literature. 2.2. Fuzzy logic and neuro-fuzzy models applications
An FFANN was employed by [72] to forecast hourly GWL at
eight wells in South Korea. The study considered the surface water Neuro-Fuzzy models belong to the category of hybrid models,
level and groundwater abstraction parameters as input. The esti- which combine two paradigms-ANNs and fuzzy logic (FL)-to utilize
mated GWL values were very accurate, in accordance with the the relative advantages of each algorithm. FL is mainly used to
actual magnitudes of GWL and, therefore, the proposed model transform linguistic labels into mathematical expressions using
was considered to be efficient in capturing the non-linear relation- if-then rule formulations [77]. The combination of ANN and FL
ship between the targets and predictors. has helped to form the famous ANFIS model [78]. The ANFIS model
In Longyan city, Fujian Province of southeast China, a study has uses an ensemble of if-then rules and membership functions (MFs)
been conducted [73] to forecast the monthly GWL using two to map a set of input variables (xi ) to an output variable (y). It con-
robust approaches: GM(1, 1) and radial basis function neural net- sists of five layers; namely, a fuzzification layer, a multiplication
274
Table 1
The established research on the GWL prediction using the applications of ANN models.
Research Applied AI Models Case study Data Span Input parameters Output parameters Performance metrics
[65] MLP India 2002–2016 Raf, P, T, H GWL R2
[66] MLP India 2006–2013 Raf, GWL Monthly GWL R2 , RMSE
[67] ANN Tunisia 2000–2018 Raf, GW, P Monthly GWL RE, RMSE R2 , NASH
[68] MLP Pakistan 2003–2014 Raf, SR,T max , T min , polygon area,
area elevation, RH Daily GWL MAE,MSE,R
[69] RNN, SVR USA 1987–1994 GWL Daily GWL MSE
[70] FFBPNN Taiwan 2008 GWL Hourly GWL, GWF R2 , RMSE
[71] FFBPNN USA 2014 RH, T, Raf Daily GWL RMSE
[72] FFBPNN South Korea 2014 SWL, GWA-WCC, GWA- GWHP Hourly GWL ME, RMSE, R, NASH
[73] RBFN, GM(1,1) China 2003–2011 GWL Monthly GWL R2 ,RMSE, MAE
[74] MLP Malyasia 2015–2016 GWL, Raf, RS, WL, SFR, T Daily GWL R, R2 , MSE RMSE
[41] ELM, RBFN, MLP, ARMA Iran Not mentioned GWL, Raf, eV Monthly GWL MSE, MAE, NASH, R2
[76] ELM Canada 2006–2014 GWL, Raf, T, EVP, Monthly GWL R2 , RMSE, NASH
layer (i.e., firing strength layer), a normalization layer, a defuzzifi- Moravej et al. [85] applied ANFIS and GP models to predict
cation layer and, finally, the output layer (i.e., summation) [79]. monthly GWL from evaporation (EP) and precipitation (P). In addi-
Like any other machine learning model, ANFIS possesses a set of tion, they exploited the abilities of metaheuristics optimization
parameters to be optimized during the learning process [19]. The algorithms, interior search algorithm (ISA), and genetic algorithm
structure of an ANFIS model is very close to that of an ANN, except (GA) to improve the performance of the least-squares support vec-
that it contains two kinds of parameters-linear and non-linear- tor machine (LSSVM). A relative performance analysis of ANFIS, GP,
making its training very difficult, especially for large numbers of GA-LSSVM, and ISA-LSSVM models showed that the highest accu-
input variables. The linear parameters (i.e., consequent parame- racy was achieved using the ISA-LSSVM algorithm. The study also
ters) are those of the fuzzy rules, while the non-linear parameters reported that the inclusion of P and EP did not contribute to the
are those of the MFs. During the training process, the two kinds of improvement of model performance. Bak and Bae [86] used the
parameters are optimized simultaneously. An example of the ANFIS model to predict GWL using P and T mean and reported accept-
ANFIS model architecture is presented in Fig. 4. Fuzzy logic and able results, with RMSE and MAPE of 0.1381 and 37.869%, respec-
Neuro-Fuzzy are among the most common models used in the area tively. The feasibility of fuzzy logic (FL) for GWL prediction has
of hydrological sciences [80]. However, the majority of the been demonstrated in a recent study [35]. They proposed FL mod-
reported studies have used the famous ANFIS model; for example, els, including Sugeno (SFL), Mamdani (MFL), and Larsen (LFL). In
for precipitation forecasting [81], soil moisture simulation [82], addition, multiple models were proposed, according to three prin-
modeling of reference EVP [83], sediment load modeling [14], cipal forms: Simple averaging, weighted averaging, and committee
and for modeling total dissolved solids [84]. machine techniques. The results showed the superiority of the sim-
The number of studies focused on modeling GWL using neuro- ple committee fuzzy logic (SCFL) model, with an R2 value ranging
fuzzy models has constantly increased over the last two decades from 0.690 to 0.940 and RMSE ranging from 0.252 to 0.103. see
(as shown in Table 2); however, the ANFIS model remains the Table 3.
most-used model, with or without the inclusion of meteorological An alternative model was developed, by Jahanara and Kho-
variables as predictors. On the other hand, our literature review dashenas [87], by combining three paradigms: Neuro-fuzzy (NF),
revealed that only a few studies have investigated the application group method of data handling (GMDH) and metaheuristics opti-
of ANFIS in predicting GWL using only antecedent GWL data. mization algorithms; that is, particle swarm optimization (PSO)
Fig. 4. The structure of ANFIS model.
275
Table 2
List of reviewed papers related to modeling groundwater level using ANFIS models.
Research Applied AI models Case study Data Time Input Output Performance Commentary
Location span scale parameters parameter metrics (Best model)
Moravej et al. ANFIS, GA-LSSVR, ISA-LSSVR Iran 2002– Monthly P, EP GWL R2 , NSE, RMSE ISA-LSSVR*
[85] 2008
Nourani and ANFIS, FFNN, WANFIS WFFNN Iran 2001– Monthly P, Q GWL R2 , RMSE ANFIS-RBF*
Mousavi [47] 2011
Zhang et al. [88] ANFIS, RBFNN, GSM China 2000– Monthly GWL GWL R, NSE, MARE, ANFIS*
2009 RMSE
Wen et al. [105] ANFIS, WANFIS, FFNN China 2007– Weekly GWL GWL R, MARE, WANFIS*
2009 RMSE
Khaki et al. [91] ANFIS, FFNN, CFN Malaysia 2007– Monthly GWL, P, EP, H GWL R, MSE ANFIS*
2013 %, T max , T min
Bak and Bae [86] ANFIS Korea 2015– - P, T mean GWL MAPE, RMSE Unknown
2017
Nadiri et al. [35] SFL, MFL, LFL, CFL-SA, CFL-WA, SCFL Iran 2007– Monthly GWL, Q, P, GWL R2 , RMSE SCFL
2016 T mean
Shiri et al. [100] ANFIS, FFNN, GEP, SVM, ARMA Korea 2001– Daily ET, P GWL R, NSE, RMSE, GEP*
2008 CO
Sridharam et al. ANFIS, WANFIS India 1990– Daily ET, P, T mean , IL, GWL R2 , MAE, WANFIS*
[106] 2017 GWL RMSE
Kisi and Shiri [99] ANFIS, WANFIS USA 2001– Daily GWL GWL R2 , RMSE WANFIS*
2008
Sreekanth et al. ANFIS, FFNN India 2000– Monthly P, EP, H%, T max , GWL R2 , RMSE, EV FFNN*
[102] 2006 T min
Emamgholizadeh ANFIS, FFNN Iran 2002– Monthly GWL GWL R2 , MAE, ANFIS*
et al. [95] 2011 RMSE
Fallah-Mehdipour ANFIS, GP Iran Not Monthly GWL, P, EP GWL R2 , NSE, RMSE GP*
et al.[104] reported
Gong et al.[89] ANFIS, FFNN, SVM USA 1998– Monthly GWL, P, LL, GWL R, NSE, RMSE, ANFIS*
2009 T mean NMSE, AIC
Shirmohammadi ANFIS, ARMA, ARIMA, SARIMA, Fuzzy- Iran 1992– Monthly GWL, P GWL R2 , RMSE, AIC ANFIS*
et al.[103] ARX, Fuzzy-ARMAX, ARX, ARMAX, 2007
Jahanara and NF-GMDH-PSO, RBFNN-NF-GMDH-PSO USA 2001– Daily GWL GWL R, RMSE, NF-GMDH -
Khodashenas 2008 PSO
[87]
Jeihouni et al. ANFIS, WANFIS, W-LSSVM, LSSVM, Iran 2002– Monthly P, T mean GWL R2 , RMSE W-NARX*
[107] NARX, W-NARX 2016
Djurovic et al.[92] ANFIS, FFNN Serbia 1990– Daily P, EP, T max , GWL R, NSE, RMSE ANFIS*
2010 T min
Kholghi and ANFIS, OKR Iran Unknown Unknown LNG, LAT GWL NSE, MAE, ANFIS*
Hosseini [101] MSE
Samantaray et al. CANFIS, RNN India 1988– Monthly P, H%, T mean , IL GWL R2 , MSE, CANFIS
[106] 2017 RMSE
Maiti and Tiwari ANFIS, FFNN, BNN India 1972– Monthly P, T max , T min , GWL R, IA, RE, ANFIS*
[94] 2001 T mean RMSE
Zare and Koch ANFIS-FCM, WANFIS-FCM Iran 1991– Monthly P, GWL GWL 2
R , RMSE WANFIS -FCM
[98] 2013
Moosavi et al.[97] ANFIS, WANFIS, FFNN, WFFNN Iran 1992– Monthly GWL, P, EP, Q GWL R2 , NSE, RMSE WANFIS*
2007
Moosavi et al. ANFIS, WANFIS, FFNN, WFFNN Iran 1992– Monthly GWL, P, T max , GWL RMSE, IA WANFIS*
[96] 2007 T min , T mean
Raghavendra and ANFIS, GPR India 2000– Monthly GWL GWL R, NSE, RMSE GPR*
Deka [90] 2013
and gravitational search algorithm (GSA). Consequently, two a monthly time step, and reported that the ANFIS model was more
hybrid models were obtained; namely, NF-GMDH-PSO and NF- accurate.
GMDH-GSA. The accuracies of the two hybrid models were evalu- Raghavendra and Deka [90] proposed the multi-step-ahead
ated against RBFNN and it was found that the NF-GMDH-PSO per- forecasting of monthly GWL in the river basin near Sullia Taluk,
formed significantly better than the NF-GMDH-GSA and RBFNN, India, using ANFIS and Gaussian process regression (GPR)
showing higher R2 and RMSE values (0.969 and 0.618, respec- approaches. Four input variables, including the GWL measured in
tively). Zhang et al. [88] analysed the differences in GWL prediction the previous four months, were used to forecast GWL up to six
by three AI models: ANFIS, RBFNN, and the grey self-memory months in advance. The results showed that GPR had significantly
model (GSM). They found that all models could be successfully higher accuracy in prediction than the ANFIS model. In addition, it
applied to model the GWL. They also reported that the ANFIS was demonstrated that the performances of the two models (i.e.,
model was generally more accurate than the GSM and RBFNN ANFIS and GPR) decreased from one to three months ahead. Simi-
models, as it obtained the highest performance metrics (i.e., R2 , larly, Khaki et al. [91] quantified monthly GWL measured at Langat
NSE, MARE, and RMSE). Gong et al. [89] modeled GWL using previ- Basin, in the southeastern part of Selangor state, Malaysia. A large
ous GWL, P, lake level (LL), and T mean as predictors. The authors number of regressors were used; namely, previous GWL, P, EP, H %,
compared the results obtained using ANFIS, FFNN, and SVM with T max ; T min . Three models were used: ANFIS, FFNN, and the cascade
276
Table 3
List of reviewed papers related to modeling groundwater level using kernel function embedded models.
Research Applied Models Case Study Data Time Input Parameters Output Performance Used Kernal Function
Location Span Scale Parameter metrics
Fang et al. SOM-SVM, MOGA- Taiwan 1998– Monthly GWL, P, TAmean , WS, GWL RMSE, bias, MAE, Polynomial
[124] SVM 2007 SD, RH NS
Gong et al. ANN, SVM, ANFIS Florida, USA 1998– Monthly GWL, P, TAmean , GWL R, NMSE, RMSE, –
[89] 2009 TAmax , TAmin , LL NS, AIC
Guzman et al. NARX-ANN, RBF- Mississippi, 1985– Daily GWL, P, ET GWL MSE Polynomial Radial Basis
[125] SVR USA 1994 Function Sigmoid
Nie et al. RBF-NN Jilin, China 2003– Monthly P, TAmean , E GWL CC, RMSE, MAE, NS Radial Basis Function
[126] 2014
RBF-SVM
Sahoo et al. RBF-SVR, RF, GB Indo- 2002– Monthly P, SM GWL MAE, RMSE, bias, Radial Basis Function
[127] Gangetic, 2011 CV(RMSE)
India
Sattari et al. SVR, M5Tree Ardebil, Iran 1997– Monthly GWL, PV, WD GWL CC, RMSE Polynomial
[129] 2013
Tang et al. LS-SVM, SVM, United 2016 Hourly GWL* GWL MAE, MAPE, MSE, Radial Basis Function
[130] ANN, RF, kNN Kingdom RMSE
Yoon et al. ANN, SVM South Korea 2003– Daily GWL, P GWL CC, MAPE, ME, Radial Basis Function
[131] 2008 RMSE
forward network (CFN). Their performances were evaluated using (DWT) and introduced the decomposed signal as an input variable
R and MSE. The results showed that the ANFIS model had a signif- to the ANFIS and FFNN models. For example, Kisi and Shiri [99]
icantly higher accuracy (R2 = 0.94, MSE = 0.005) than that of the have demonstrated that the DWT helps in improving the perfor-
FFNN and CFN. Moreover, Djurovic et al. [92] applied ANFIS and mance of the ANFIS model, allowing R2 values close to 0.99 during
FFNN models to predict daily GWL in the riparian lands of the the testing phase.
Danube basin in Serbia. Four input variables were selected, includ- Despite the increased use of machine learning, several authors
ing, P, EP, T max , and T min , for GWL prediction. The obtained results have tried to apply several kinds of models with the overall objec-
revealed that both models could be applied successfully with a tive of acquiring high-level prediction accuracy and exploiting the
high level of accuracy. Samantaray et al. [93] demonstrated that advantages of the available information from input variables. For
co-adaptive neuro-fuzzy inference systems (CANFISs) are more instance, Shiri et al. [100] used P and ET for predicting daily GWL
accurate than RNNs in predicting monthly GWL using P, H %, using ANFIS, FFNN, SVM, and GEP models. Although the first three
T mean , and infiltration loss (IL) measured in the region of Odisha, models have been documented in the literature, the GEP was con-
India, with R2 , RMSE, and MSE equal to 0.953, 0.0393, and sidered as one of the novelties of the investigation. They reported
0.00378, respectively. Maiti and Tiwari [94] and Emamgholizadeh that the GEP model was more accurate in terms of prediction.
et al. [95] have conducted extensive studies and compared the per- Some investigations must be discussed here due to their uncon-
formance of ANFIS and FFNN models in different regions of the ventional and novel approaches. Kholghi and Hosseini [101] pub-
worlds in simulating GWL. They reported ANFIS as more suitable lished an article related to modeling GWL using an
for GWL prediction, compared to the FFNN model, with either only unconventional approach: Ordinary Kriging (OK). In fact, [101] pio-
antecedent GWL or meteorological variables with antecedent GWL neered the use of OK for modeling GWL using only the Latitude and
as inputs. Longitude as predictors. The performance of OK was compared
Data transformations, such as wavelet transforms, provide with ANFIS in predicting GWL in the Qazvin plain, located in the
methods for data decomposition allowing us to obtain highly west of Tehran, Iran. The results showed that the ANFIS model
improved data signals and, thus, leading to a significant improve- had higher accuracy than the OK model. Another group of
ment of model accuracy. Such a data pre-processing approach researchers-Sreekanth et al. [102], Shirmohammadi et al. [103],
has been employed, by Moosavi et al. [96,97], to simulate monthly and Fallah-Mehdipour et al. [104], have explored the significance
GWL using ANFIS, WANFIS, FFNN, and WFFNN in Iran from P, of GWL prediction using GP, the seasonal autoregressive integrated
T max ; T min ; T mean , EP, and the GWL measured at the previous lag. moving average (SARIMA) model, and hybrid fuzzy-SARIMA. For
They found that wavelet data transformation is an effective instance, Fallah-Mehdipour et al. [104] highlighted the superiority
method for capturing the non-linearity in the time-series by of GP, compared to ANFIS, in predicting monthly GWL in Iran.
removing noise from data. However, they highlighted that the The aim of the study was to develop a new modeling strategy
wavelet transform may become more suitable and further con- using GP based on two readily available meteorological variables:
tribute to the performance of ANFIS, compared to FFNN. In order P and EP. Interestingly, another study conducted by Shirmoham-
to ensure a robust prediction strategy of GWL, Zare and Koch madi et al. [103] in the same year proposed different models for
[98] have recently proposed a prior wavelet transform with several GWL prediction from precipitation and antecedent GWL data. They
decomposition levels to provide inputs to the ANFIS with fuzzy c- tested a suite of models (i.e., ANFIS, ARMA, ARIMA, SARIMA, Fuzzy-
mean clustering model (ANFIS-FCM). Although this strategy of ARX, Fuzzy-ARMAX, ARX, and ARMAX) and reported the superior-
decomposition, accompanied with a good choice of the number ity of the ANFIS model.
of clusters, worked very well, the authors highlighted the impor-
tance of adequate selection of the mother wavelets, which plays 2.3. Kernel models applications
a fundamental role in obtaining excellent accuracy. They showed
that the best performance (with R2 and RMSE values of 0.983 Due to their simplicity and generality, kernel functions play
and 0.18, respectively) was obtained during the testing phase using progressively outstanding roles in machine learning and its appli-
the sym4 mother wavelet. In light of this, many researchers have cation. This was inaugurated with the launching of support vector
simplified the daily GWL using the discrete wavelet transform machines and extended with the improvement of other kernel
277
function-embedded models [108,109]. As a kernel function can be dictive results using P, T, previous GWL, and lake level accom-
described on any input space, diverse kernel functions have been plished better performance than those using P, T, and previous
successfully applied to describe the non-linear relationships defin- groundwater level only. This demonstrates that the interaction
ing complicated data architectures. The support vector machine is between groundwater and surface water is required for predicting
a well-established Kernel model, which has been implemented GWL fluctuation. Besides, the ANFIS and SVM models accom-
widely in several fields, such as regression, bioinformatics, pattern plished better predictions than the ANN model.
recognition, and environmental engineering [110]. It is a novel Guzman et al. [125] developed the non-linear autoregressive
kind of categorizer, stimulated by two approaches: First, convert- with exogenous inputs-based ANN (NARX-ANN) and RBF-based
ing data into a high-dimensional area can reconstruct complicated support vector regression (RBF-SVR) models, in order to evaluate
problems into simple ones. Second, it is inspired by the theory of GWL in irrigation wells in Mississippi, USA. Three kernel functions
training and uses only relevant inputs [111–113]. The resolution (i.e., polynomial, radial basis function, and sigmoid) were
of a traditional ANN may fall into a regional optimized solution, employed to determine the optimal SVR model with the lowest
while an overall optimized solution is guaranteed by the SVM training error. Among them, the RBF kernel function furnished
model [114]. Kernel function embedded models have been utilized the best accuracy. They classified the total historical time-series
for the solution of hydrological processes, such as Raf-runoff into withdrawn (i.e., summer) and recharge (i.e., winter) seasons.
[115,116], EP [117,118], EVP [119,120], water stage [121,122], The RBF-SVR model accomplished better prediction than the
and so on. A detailed list of reviewed papers related to modeling NARX-ANN model for individual (summer or winter) seasons.
groundwater level using kernel function embedded models is given Therefore, the prediction of GWL by individual season was more
in Table 3. The non-linear SVM Vapnik’s -insensitivity loss func- accurate than that using the total time-series. Furthermore, the
tion is presented in Fig. 5. results indicated that the winter season presented as a linear prob-
Numerous researchers have recently developed SVM models to lem, which decreased the computational requirements of the RBF-
estimate, predict, and forecast GWL fluctuations. Fang et al. [124] SVR model.
employed a two-stage approach for the development of self- Nie et al. [126] employed the RBF-NN and RBF-SVM models to
organizing maps (SOM) and multi-objective genetic algorithm predict the fluctuation of monthly GWL in Jilin, China. The RBF ker-
(MOGA)-based SVM (i.e., SOM-SVM and MOGA-SVM) models to nel function was implemented to set the SVM model structure. The
forecast spatial–temporal GWL in the Choushui River Alluvial uncertainties generated from the measured errors of input and
Fan, Taiwan. A polynomial kernel function was implemented in output variables were computed based on 95 % confidence inter-
the construction of the SVM model. In the first stage, the SOM- vals. Their research reported that the RBF-SVM model achieved
SVM model applied a clustering method to segregate the geograph- more accurate and less uncertain results to predict the fluctuation
ical and hydrological components into spatial groundwater zones. of monthly GWL, compared to the RBF-NN model.
In the second stage, the MOGA-SVM model decided the optimal The Gravity Recovery and Climate Experiment (GRACE), an arti-
input combinations. The results showed that the MOGA-SVM ficial satellite, can follow the variation of groundwater level. Grid-
model was appropriate for forecasting the fluctuation of GWL with ded rainfall (GR) and soil moisture (SM) data can also support the
short and long lead-times, while the MOGA-SVM model was more development of hydrological models. Unique research using
efficient and accurate. GRACE, GR, and SM data to predict the variation of GWL in the
The accurate prediction of GWL fluctutations is a very compli- Indo-Gangetic basin, India, has been accomplished by Sahoo et al.
cated and non-linear phenomenon in the natural environment, as [127]. The RBF-SVR, random forest (RF), and gradient boosting
it depends on many diverse components such as P, EP, T, and so (GB) models computed the ground assignment of satellite data.
on. Gong et al. [89] developed three machine learning models The RBF kernel function was used to construct the SVR model
(i.e., ANN, SVM, and ANFIS) to predict the fluctuation of monthly architecture. The addressed research explained that nine pixels of
GWL in Florida, USA. The kernel function was used to form the the GRACE satellite data were applied to identify the relationship
SVM model architecture. They reported that the models were between well-measured and satellite-acquired data. The nine pix-
effective and accurate for predicting the fluctuation of GWL with els were categorized based on the presence or absence of hydrolog-
three month lead-times (i.e., one, two, and three months). The pre- ical features. The pixels identifying perennial streams provided
Fig. 5. Non-linear SVR vapnik’s -insensitivity loss function [123].
278
reasonable prediction, while the pixels displaying wells near the level simulation [132]. One of the major advantages of deep learn-
streams supplied poor prediction for the variation of groundwater ing is the capability of analyzing complex and high-dimensional
level. Additionally, the RBF-SVR model supported better prediction data in a relatively short period with minimal manpower, com-
than the RF and GB models. Another investigation was conducted pared to conventional data collection [133]. Deep learning models
on GWL using the feasibility of GB model within the Karnataka are comprised of an input layer, hidden layers, and an output layer,
state, India and the research approved the proposed model signif- where a neural network is used to map features into the output
icant on the prediction accuracy [128]. layer [134]. CNN and LSTM are the most widely applied deep learn-
Sattari et al. [129] resolved the SVR and M5 Model Tree ing algorithms in hydrology studies [134]. As this review focuses
(M5Tree) models, in order to predict the variation of monthly solely on the application of deep learning to groundwater quantity,
GWL in the Ardebil plain, Iran. The polynomial kernel function studies on groundwater contamination were excluded. In general,
was built in the networks of the SVR model. The variables of the the application of deep learning in groundwater, as reported in
input combination involved the previous groundwater level, P vol- the literature, can be divided into three major groups: (1) Compar-
ume, and well discharge, while GWL was considered as the output ison of the performance of different deep learning algorithms; (2)
variable. Both models (i.e., SVR and M5Tree) predicted the varia- filling missing data values; and (3) improvement of the simulation
tion of monthly GWL accurately. However, the addressed research framework.
explained that the prediction processes applying the M5Tree A total of 10 articles on deep learning and groundwater level
model were simpler and easier than those when employing the were identified, as listed in Table 4. Identification of the best deep
SVR model. learning algorithm is one of the most popular research topics in
A novel two-phase scheme to predict the variation of hourly this field. Kumar et al. [135] compared deep learning, ELM, and
GWL utilizing temporal-spatial analysis and the LSSVM model in GPR in predicting groundwater level in the Konan basin, Japan,
the United Kingdom has eben proposed by Tang et al. [130]. In with P, river stage, T, recharge, and groundwater level in the input
the first phase, the temporal analysis utilizing the autocorrelation data layer. Similar studies have been conducted by Supreetha et al.
function (ACF) and partial autocorrelation function (PACF) is car- [136] in the Udipi district in Karnataka state of India, who found
ried out to analyze the temporal behaviors of GWL. The LS-SVM that the Long Short-term Memory-Lion Algorithm (LSTM-LA) out-
model is also constructed to predict the variation of hourly GWL, performed the LSTM and Feedforward Neural Network (FNN) in
based on the results of the temporal analysis. The RBF kernel func- groundwater level prediction. The proposed predictive model is
tion was embedded in the architectures of the LS-SVM and SVM demonstrated in Fig. 6, as an example for the readers. Park and
models. The results of the LS-SVM model were compared with four Chung [137] and Shin et al. [138] both reported the reliable perfor-
machine learning models: SVM, ANN, RF, and k-nearest neighbors mance of the LSTM model in groundwater level simulations in
(kNN). In the second phase, the spatial analysis, employing the Southeast Korea. Interestingly, Shin et al. [138] found that the
cross-correlation technique, was carried out to compute the duration of the training period did not have a significant impact
cross-correlation of mean sea level between the attractive and on the simulation performance, where the NSE value only differed
neighbor-measured locations. The results revealed that the LS- by 0.02 between the 6-year and 19-year study periods. The model-
SVM model was superior to the other machine learning models ing scheme proposed by Shin et al. [138] is reported in Fig. 7.
for predicting the variation of hourly GWL. Missing GWL records is a common issue, particularly in devel-
Yoon et al. [131] investigated ANN and SVM models to predict oping and underdeveloped countries. Human factors, equipment
the variation of daily GWL in South Korea using the weighted error failure, and fluctuations of water level are among the factors that
function (WEF) approach. The ANN model was trained by the back- contribute to such missing records [139]. Therefore, reconstructing
propagation (BP) algorithm, while the SVM model was calibrated missing GWL values can help us to better understand aquifer sys-
by the sequential minimal optimization (SMO) algorithm. The pre- tems. Vu et al. [140] evaluated the capability of an LSTM for filling
diction scheme consisted of a direct and recursive strategy. The in the 50-year GWL data at 31 piezometers to predict future GWL
RBF kernel function was utilized in the SVM model design. The fluctuations in northwestern Normandy, Italy. They concluded that
WEF approach clearly enhanced the predictive accuracy of two the use of deep learning is viable to reconstruct GWL fluctuations,
models (i.e., ANN and SVM). A comparison of the models demon- with correlation coefficient and RMSE values of 0.64–0.99 and
strated that the SVM model with the recursive strategy was better 0.07–1.08 m, respectively. The integration of deep learning and
than the ANN model with the recursive strategy to predict the vari- hydrological models for fusing GRACE satellite data with NOAH-a
ation of hourly GWL. land surface model developed by NASA-has been conducted by
Our review of previous articles concerning kernel function Sun et al. (2019) [141] to improve groundwater storage prediction
embedded models (e.g., SVM, SVR, and LS-SVM) for predicting in India. The deep convolution neural network (CNN) model was
the fluctuation of GWL revealed the following: First, the RFB kernel applied to learn the Spatio-temporal mismatch in groundwater
function has frequently been employed to construct the model patterns between GRACE and NOAH. Their findings indicated that
architectures, as it has few adjusting parameters, compared with CNN improved the GRACE-NOAH match and successfully filled in
other kernel functions (e.g., polynomial and sigmoid) and can the data gaps between GRACE missions.
well-capture non-linear behavior to accomplish the accurate pre- Several studies have proposed strategies to improve deep
diction of GWL fluctuations. Second, the kernel function embedded learning-based GWL simulation frameworks; for instance, Bowes
models can predict the fluctuation of GWL effectively, based on et al. [42] explored the effects of algorithm selection, training data
hourly, daily, and monthly lead-times. Finally, considering the ker- type, and integration of forecast data on GWL predictive model
nel function embedded models, using few input data can provide accuracy in Norfolk, Virginia, USA. The proposed model consisted
more effective prediction than other models for predicting the fluc- of data pre-processing, a learning process, and result post-
tuation of GWL. processing, as reported in Fig. 8. The results showed that the LSTM
had better performance, as compared to the Recurrent Neural Net-
2.4. Deep learning models applications work (RNN), in this region. The architectures of the RNN and LSTM
models are displayed in Fig. 9. The authors also found that the
Deep learning, a subset of machine learning which specializes in model trained with a data set of only storm events outperformed
generating outputs from unstructured input data through unsuper- the models trained with continuous and forecast data using a thou-
vised learning approaches, has become popular in groundwater sand bootstrapped data set (see Fig. 10).
279
Table 4
The surveyed literature on the implementation of deep learning models for GWL modeling.
Applied AI models Case study Data Input parameters Output Performance Features
span parameter metrics
[135] DL, ELM, GPR Konan, Basin, Japan Not P, River Stage, T, Groundwater RMSE, R and DL is more reliable and robust than
reported Groundwater depth level NSE ELM and GPR
[136] LSTM, FFNN, Udupi district, Karnataka 2000– Raf, Groundwater level Groundwater RMSE, MAE LSTM-LA outperformed LSTM and
LSTM-LA state, India 2018 level FFNN
[137] LSTM Hankyung-myeon, Jeju 2001– Raf, Groundwater level Groundwater R2 Accuracy R2 - 0.98
Island, South Korea 2013 level
[138] LSTM Pyoseon watershed, Jeju 2001– Raf, Groundwater Groundwater NSE, RMSE Short and long training periods both
Island, South Korea 2019 withdrawal level gave similar model performance.
data,groundwater level
[42] LSTM, Recurrent Norfolk, Virginia USA 2010– Raf, groundwater level, Groundwater RMSE - LSTM outperformed RNN
Neural Network 2018 sea level level
(RNN)
- Model trained with only storm
events performed better than
continuous data
[142] Two-layer LSTM Hetao Irrigation District, 2000– Water diversion, EP, P, T, Groundwater R2 LSTM much better than FFNN
Northwestern China 2013 time level
[140] LSTM Upper Normandy, Italy 1970– Groundwater level Groundwater RMSE, R -Good performance in filling missing
2005 level values
-Unreliable input could lead to poor
prediction
[143] LSTM Gangjin-Seongjeon and 2005– Groundwater level Groundwater MSE -development of a cost function for
Pohang-Gibu, South 2016 level robust training.
Korea
[144] LSTM, gated Jindo Uisin and Pohang 2005– Groundwater level Groundwater RMSE, R Pre-processing data before applying
recurrent unit Gibuk, Jeju Island, South 2014 level DL to estimate groundwater level
(GRU) Korea
[141] CNN India 2002– Total Groundwater Total NSE CNN was trained to learn the TGWS
2016 Storage (TGWS) Groundwater spatial mismatch between GRACE
Storage and NOAH
Zhang et al. [142] developed a two-layer LSTM model to predict

groundwater level in the Hetao Irrigation District of Northwestern
China. Their findings showed that the proposed model could
achieve considerably higher R2 values (0.79–0.95), when compared
to the FNN (0.0–0.50). Another two improvement studies have
been conducted in South Korea, where Jeong et al. [143] developed
a cost function for robust data training purposes and Jeong and
Park [144] proposed a massive pre-processing data scheme before
the training step. Both studies showed that the removal of data
noises before implementing the deep learning simulations could
improve the model accuracy.
2.5. Hybrid ML models applications
Despite significant advancement in recent years, in terms of

handling non-stationary, dynamic, and non-linear time-series data
using ML models-particularly applied in hydro-environmental and
water resource management-there are still some weakness associ-
ated to such approaches. A variety of problems in hydrological sim-
ulation reserach related to a single AI-/machine learning-based
modeling has been addressed, regardless of their promising perfor-
mances demonstrated in various studies [145–154]. According to
[155–159], hybrid models have proved not only their merit and
superiority to the use of a single model, but can also address a vari-
ety of different problems associated with the use of single
techniques.
The concept of Hybrid methods implies the combination of one
or more AI-based models, computing machine learning models,
and/or classical regression models for improving the performance
accuracy or to obtain optimal outcomes. Hybrid methods could
be utilized in the prediction or optimization stages, based on their
Fig. 6. The proposed predictive model based on the hybridization of the LSTM-LA specific purposes. Hence, it can be justified that hybrid methods
for modeling GWL [136]. comprise several combined single techniques and/or optimization
280
Fig. 7. The proposed LSTM predictive model for GWL modeling by Shin et al. [138].
Fig. 8. The proposed modeling procedure consisting the data preprocessing, predictive model process and post-processing [42].
281
Fig. 9. The models’ architectures of the RNN and LSTM developed by Bowes et al. [42].
Fig. 10. The proposed Model training and evaluation with bootstrapped datasets by Bowes et al. [42].
282
algorithms, which have been proved to be more reliable and cap- the hybrid ANN-GA produced the best GWL simulation. Jalalkamali
able of outperforming single models, with regards to modeling and Jalalkamali [173] studied the potential of a hybrid ANN-GA
accuracy [160]. model for the prediction of GWL. The classical FFNN and RNN were
Among the different categories of hybrid methods reported by compared, in order to demonstrate the performance accuracy of
the various researchers in the field of water resource management the models when using several input variables. The obtained
and hydro-environmental engineering, ANNs have generally results indicated that ANN-GA can serve as a reliable model for
shown outstanding improvement in GWL prediction [161,162]. GWL simulation. Adamowski and Chan [174] coupled a hybrid
Nourani et al. [47] integrated a wavelet hybrid neural network WA-ANN for GWL prediction using different hydro-climatology
(WT-FFNN) for GWL simulation using SOM clustering techniques variables. Recorded monthly GWL data were obtained from the
at different piezometer positions in the Ardabil plain. The results Chateauguay watershed in Quebec, Canada during 2002–2009.
were compared with the traditional FFNN and ARIMA models, The GWL simulation results, based on evaluation and comparison,
where the output reported that the hybrid WT-FFNN increased indicated that the hybrid WA-ANN performed better than ANN and
the average performance by up to 15.3% over FFNN. Mathur ARIMA models. Yadav and Mathur [76] applied the new hybrid
[163] proposed single and hybrid SVM with POS (SVM-PSO), in Quantum behaved Particle Swarm Optimization function (SVM-
order to investigate the feasibility of GWL modeling at Andhra Pra- QPSO), in order to estimate the GWL in Rentachintala region,
desh, India. The comparison was made using ANFIS and ARIMA Andhra Pradesh, India. The results showed that SVM-QPSO per-
models, which indicated that SVM-PSO is far more reliable and formed better, with regards to performance evaluation, than the
has higher accuracy than other single models. Chang et al. [164] ANN model. Moosavi et al. [96] presented several soft-computing
developed a new hybrid soft-computing approach using SOM- models, including ANN and ANFIS, which were coupled with opti-
NARX techniques. The study used monthly regional data recorded mization algorithms to develop hybrid models (i.e., WA-ANN and
from 203 stations during 2000–2013 in Zhuoshui River basin, Tai- WA-ANFIS) to estimate monthly GWL. The forecasting skill of the
wan. The results, based on statistical indicators, demonstrated the models indicated that the hybrid ANFIS (WA-ANFIS) outperformed
suitability and reliability of the hybrid SOM-NARX method in mod- the other models, in term of performance criteria. Tapoglou et al.
eling GWL. The outcomes also depicted that the proposed tech- [175] introduced hybrid ANN-Kriging techniques for spatial–tem-
niques could provide an environmental solution toward water poral modeling of GWL at different places in Bavaria, Germany.
resource management. Huang et al. [165] determined the daily The results showed that hybrid model can fully achieve the
GWL using standalone and hybrid models (PSO-SVM and PSO- expected GWL prediction outcomes.
BPNN) in the Three Gorges Reservoir Area in China. Chaos theory More recently, Yaseen et al. [26] developed an evolutionary
was applied to obtain the best input combination for the non- hybrid algorithm based on the comparison of the bat algorithm
linear models. The results indicated that non-linear PSO-SVM had (BA) and PSO algorithm, called the hybrid bat-swarm algorithm
better prediction skill than the simple linear PSO-SVM and (HB-SA), for dam and reservoir optimization. The obtained out-
chaos-BPNN. Zare and Koch [98] employed AI and regression mod- comes showed the suitability and generalizability of the proposed
els for the simulation of GWL using different input combinations; HB-SA method. Malekzadeh et al. [176] presented a study using a
after that, a new hybrid wavelet technique, WA-ANFIS, was pro- hybridized wavelet of Self-Adaptive Extreme Learning Machine
posed. Observational data were obtained from a case study in the (SAELM) and Wavelet-SAELM (WA-SAELM) in Kabodarahang
Miandarband plain, Iran. The overall results confirmed the accu- region, Iran. The results of the proposed methods were compared
racy of the hybrid model. Rakhshandehroo et al. [166] estimated to standalone hybrid AI-based models (i.e., WA-ANN and WA-
GWL through a new hybrid model using wavelet neural networks SVM). The predictive results proved that hybrid WA-SAELM pro-
(WNNs) calibrated with an improved harmony search (IHS) algo- duced the best outcomes and, hence, served as the most reliabile
rithm. The efficiency of the model was compared with classical dif- approach. Supreetha et al. [177] investigated hybrid ANNs, includ-
ferential evolution (DE), PSO, RBFNN, MLP, and harmony search ing hybrid ABC and PSO algorithms, for the forecasting of GWL
(HS) models. The outcomes established for GWL modeling indi- using the observational GWL at Manipal from Udupi, Karnataka,
cated the dominance of the proposed hybrid model over other sin- India. The models were evaluated using RMSE, MAE, and R, and
gle models. Balavalikar et al. [167] used monthly GWL variation the results indicated the superiority of the hybrid ABD-PSO tech-
data from 2000–2013 in Brahmavar, Kundapur, and Hebri in Udupi niques. Roshni et al. [178] developed a traditional FFNN with a
district, India, in order to predict the GWL using the combinations hybrid WANN model for the prediction of complex GWL in an allu-
of ANN and hybrid POS-ANN. For this purpose, the models were vial aquifer. The results integrated the Gamma and M-tests (GT)
calibrated using different input combinations. The performance approach for the same purposes, while a different evaluation
results demonstrated the capability of PSO-ANN over ANN in mod- matrix was used to assess the model’s performance. The examined
eling GWL. POS and GA are popularly known hybrid learning tech- calibrated results justified the robustness of GTWANN for the esti-
niques, due to their effectiveness in determining the RBFNN mation of GWL. Kombo et al. [147] presented a long-term multi-
optimization parameters. The insensitive nature of GA in optimiz- step GWL estimation using a hybrid K-Nearest Neighbors-
ing the initial guess parameters is the main merit of employing GA Random Forest (KNN-RF) technique in eastern Rwanda using cli-
over other hybrid approaches [168,169]. The advantage of the mate variables (T, P, max. solar radiation, and GWL). The modeling
Whale algorithm (WA) has been attributed to its high convergence, results, based on NSE, RMSE, MAE, and R2 , confirmed that the
suitability with other optimization algorithms, and its capability to hybrid model provided a reliable approach. Moravej et al. [85]
handle a huge amount of decision variables [170,171]. Further lit- developed a novel hybrid model based on ISA and GA for the sim-
erature regarding the merits and supremacy of other hybrid opti- ulation of GWL, using Raf, EP, and GWL data obtained from an
mization algorithms are enclosed in this review. However, unconfirmed aquifer in the northwest of the Karaj plain, Iran. Com-
investigating the aforementioned studies coupling popular AI- parison of ANFIS and GP models in GWL prediction justified the
based hybrid models (e.g., ANN), it is worth mentioning that hybrid superiority of the novel hybrid ISA-LSSVR methods. Roshni el al.
techniques have achieved considerable attention over the last few [161] presented a comparison of novel hybrid Emotional ANN cou-
decades; for instance, Dash et al. [172] made an attempt to develop pled with a genetic algorithm (EANN-GA), generalized regression
a hybrid NN with GA (ANN-GA) using several learning algorithms neural network (GRNN), and FFNN for the estimation of GWL at
for GWL modeling in the Mahanadi river basin, India. The predic- three different sites in a coastal aquifer. The performance evalua-
tion outcomes using different performance criteria revealed that
283
tion of the models was carried out using several indicators. The duce offspring. Various studies have shown that the offspring
prediction results indicated that EANN-GA outperformed EANN, may solve a problem better than their parents [202–204].
GRNN, and FFNN for the simulation of spatial–temporal GWL. Mutation is an operator in which a parent’s node or branch is
Banadkooki et al. [179] explored a hybrid RBF with whale algo- replaced with a randomly created node/branch using the materials
rithm (WA) model (RBF-WA), GP, and MLP-WA for modeling existing in functional or terminal sets. As illustrated in Fig. 13, the
GWL under different scenarios based on temporal P data. The per- sub-tree 1 þ x2 in the parent gave its place to the branch tanðx1 Þ in
formance of the predictive models showed that MLP-WA emerged the offspring.
as the best hybrid model for the prediction of GWL in Yazd pro- In order for the best individual to survive in each iteration, it
vince, Iran. Natarajan and Sudheer (2020) [154] explored the capa- directly moves to the new population without any change. This
bility of different data-driven models (ANN, SVM, ELM, and GP) for is a natural selection process, which allows the modeler to keep
modeling GWL at six different locations in Vizianagaram, Andhra improvement continuing in each generation. All of the above-
Pradesh, India. The results confirmed the superiority of the ELM mentioned operations are carried out in any GP variant; however,
model over other single models; on the other hand, the hybrid the process may not necessarily be done using a tree shape. Like
SVM-QPSO was found to be the best predictive model on some other AI techniques, a set of training data is used to train the GP
occasions. Seifi et al. [180] explored the performance of different and the evolved solution must be generalized for unseen testing
optimization algorithms-GA, grasshopper optimization algorithm data sets. To minimize computational costs, a set of suitable func-
(GOA), cat swarm optimization (CSO), PSO, weed algorithm (WA), tions, input variables, evolutionary operation rates, and a maxi-
GA, and krill algorithm (KA)-integrated with AI-based models mum depth of the GP trees must also be considered in the
(ANN, ANFIS, SVM). Observed monthly data for 144 months modeling process (Mehr and Noyrani 2018 [196]; Tur 2020
recorded in the Ardebil plain (Iran) were used. The modeling [204]). To avoid over-fitting, a lower number of functions and short
results were evaluated based on a number of evaluation criteria, trees are recommended [201,205].
which proved the superiority of ANFIS-GOA over standalone mod- Comparison of GEP with ANFIS for short-term GWL prediction
els. The reported research in the literature on hybrid ML models for in Illinois State has been conducted by Shiri and Kisi [197], perhaps
GWL modeling are collected in Table 5. one of the earliest applications of EC-based regression models
using GEP. The authors used daily P and antecedent GWL data as
2.6. Decision tree and data mining and Evolutionary computing the predictors of GWL in two wells: Bondville and Perry. The
models applications results showed that GEP and ANFIS can be applied to predict
GWL. An explicit expression of GEP was highlighted as the advan-
Evolutionary computing is a subfield of AI, in which biological tage of GEP over ANFIS. In a similar study, Fallah-Mehdipour et al.
evolution (based on Darwin’s principle) is used to solve stochastic [104] compared the classic GP algorithm with ANFIS, in order to
problems. Despite the variety of existing EC techniques, such as GA, predict and simulate GWLs in three observation wells in the Karaj
evolutionary algorithm, particle swarm optimization, ant colony plain, Iran. The study focused on the monthly variation of GWLs
algorithm, GP, and so on, they use the same automated problem- and suggested the use of P and EP from a surface water body to fill
solving procedure, starting with an initial set of candidate solu- possible gaps in the GWL data sets. The results showed the effec-
tions and iteratively improving solutions through mutation, tive and promising role of GP in solving the problem, as it was
cross-over, and natural selection. Current reviews of the applica- superior to ANFIS by up to 11 percent in the testing period. Kasivis-
tion of AI methods in hydrology (see, e.g., [46,196]) have shown wanatha et al. [38] implemented the classic GP method to model
that GP is one of the most popular EC techniques for GWL simula- and forecast GWL variations and emphasized the quantification
tion. Various types of GP algorithms have been used for GWL pre- of uncertainties due to input selection. To this end, observations
diction, including classic GP (e.g., [104]), GEP [197], and multigene from three wells (K. Paramathy, Keeranur, and Kuthiraiyar) in India
GP (MGGP) [198]. All of these GP variants have been well- were used. The authors demonstrated that the quantification of
described in [199]. To avoid duplication, we briefly describe the uncertainty may help to improve the confidence of GP-based
main concept of classic GP herein, in order to secure the integrity GWL models. The authors discovered that the EP is not an effective
of the current review paper. input. By contrast, river stage data near the wells was introduced
Like other EC techniques, GP uses an automatic problem-solving as the input, which may improve the model accuracy. More
technique to attain the best solution among randomly generated recently, Sadat-Noori et al. (2020) [206] showed that GP can pro-
potential solutions called genes. In the classic GP method [200], vide more accurate GWL predictions for wells close to meteorolog-
each gene is represented by a tree structure having a root node, ical stations when P data are used as input. A list of the articles on
inner nodes, and terminal nodes (called leaves); see Fig. 11. the developed evolutionary computing predictive models for GWL
Fig. 11 illustrates a genome, including a root node (multiplica- modeling is given in Table 6.
tion), inner nodes of subtraction, and terminal nodes of X 1 ; X 2 , Data mining techniques are relatively new methods for discov-
and a random number (5.25). Each terminal node in a GP tree ering patterns and finding anomalies and connections relating to
can adopt an independent variable or any random floating point complex processes in large data sets, which can be exploited to
number. forecast future trends [207]. As hydrological criteria follow a com-
To solve any time-series or regression problem using GP, the plicated process, particularly in the long-term scale, finding novel
algorithm begins with the formation of an initial population of models with high accuracy seems to be essential and indispensable
genes. Then, the evolutionary operators of natural selection- [208]. Therefore, these techniques are uniquely able to leverage the
cross-over and mutation-are used to modify the existing genes large amounts of real-time, multivariate data now being collected
[201]. The modified genes, or offspring, that show the highest fit- for hydrological systems. Karthik and Vijayarekha [209] believe
ness survive to the next generation of the population (potential that data mining techniques can be employed for quicker classifi-
solutions). This is the evolutionary process, which must be iterated cation of water portability. A broad range of algorithms is used,
until an individual meets the desired accuracy. Fig. 12 demon- in these methods, for classification and prediction. The application
strates the cross-over operator between two parents creating two of each model has been evaluated in several investigations in the
offspring. Each offspring has the same materials as their parents, hydrology field. Further details of some models with the highest
but a different combination of them. Indeed, the parents exchange level of application are discussed in the following:
some of their branches (the red and blue parts in Fig. 12)) to pro-
284
Table 5
The list of the reported research over the literature on the hybrid ML mdoels for GWL modeling
No. Author (year) Location of study Hybrid AI models Input Frequency Data Span Performance metrics
Combination
1 Kholghi and Hosseini, Qazvin plain, Iran ANFIS-Kriging GWL Monthly Not mention R2 , MAE, MSE, RSS, CE
(2009) [101]
2 Dash et al. (2010) [172] Mahanadi river basin, India ANN-GA GWL Monthly 1993–2002 R, E, MAE, IOA, RMSE
3 Nourani et al. (2011) Shabestar plain, Iran ANN-GS LL, GWL, R Monthly 1995–2007 R2 , RMSE
[181]
4 Jalalkamali and Kerman plain (Kerman, Iran) GA-ANN, FFNN, RNN Piezometers, Monthly 1988–2009 R2 , RMSE, MAPE
Jalalkamali, (2011) [173] T, R
5 Adamowski and Chan, Quebec, Canada WA-ANN, ANN, P, E, GWL Monthly Nov2002- R2 , RMSE, E
(2011) [174] ARIMA Oct2009
6 Kisi and Shiri, (2012) [99] USA WA-ANFIS GWL Daily Jan 2001- R, RMSE, CO, NSE
Dec2008
7 Sudhee et al. (2012) Andhra Pradesh, India SVM-QPSO, ANN GWL Monthly Nov 1984- EFF, R, RMSE
Dec 2001
8 Moosavi et al. (2012) Mashhad plain, Iran WA-ANN, WA- GWL, P, E, Monthly 1992 to R2 , RMSE
[182] ANFIS average Q average 2007
9 Shiri et al. (2013) [100] South Korea GP (ANN, SVM, GWL, R, ET Daily Not mention AARE, MSE, MAE
ANFIS) average
10 Fallah-Mehdipour et al. Karaj plain, Iran GP, ANFIS GWL, P, E Monthly 84-month R2 , RMSE, E
(2013) [104]
11 Maheswaran and Khosa, Northern Saanich Peninsula, WA-ANN, ANN GWL Monthly May 1975- NSE, RMSE, MAE, MRE
(2013) [183] Canada average Apr 2002
12 Moosavi et al. (2013) [97] Mashhad plain, Iran WA-ANN, WA- GWL, P, E, Monthly 1992 to R2 , RMSE, E
ANFIS average Q average 2007
13 Emamgholizadeh et al. Bastam plain, Iran ANFIS, ANN R recharge, Monthly 2002–2011 C, MAE
(2014) [184] IRF, PR
14 Suryanarayana et al. Visakhapatnam, India ANN, SVR, WA-SVR GWL, P, T max , Monthly 12 Month R2 , RMSE, EC, NMSE,
(2014) [185] T mean MAPE
15 Tapoglou et al. (2014) Bavaria, Germany ANN-ANFIS- GS GWL, SWL, T, Daily 2008–2012 RMSE, RMSEE, MAE, Bias
[186] R
16 He et al. (2014) [187] Ganzhou region, China WA-ANN, ANN GWL Monthly 1994–2004 RMSE
17 Mathur, (2015) [163] Andhra Pradesh, India SVM-PSO, ANFIS, GWL, R, H,T Monthly 1985 to RMSE, EFF, CORR
ARIMA 2004
18 Jha and Sahoo, (2015) Konan basin, Kochi, India ANN-GA GWL, R, T, Monthly 1999 to R2 , RMSE, IOA, NSE, Bias,
[188] 2004 CV
19 Yang et al. (2015) [189] Fujian, China WA-ANN, ANN GWL Monthly 1985–2004 RMSE, R, EFF
average
20 Khalil et al. (2014) [190] Quebec, Canada WA-ANN, ANN P, T Daily 1991–2012 R2 , RMSE, E, MAPE, MAE
21 Nourani et al. (2015) Ardabil plain, Iran WA-ANN, ANN GWL, R, Monthly 1998–2012 R2 , RMSE
[191] runo?
22 Chang et al. (2016) [164] Zhuoshui River basin, Taiwan ANN (SOM-NARX) GWL, Q, R Monthly 1985–2004 RMSE, CORR, EFF
average
23 Han et al. (2016) [192] Northwest China ANN-SOM GWL, Q, Monthly 1998 to NSE, R, RMSE
climatic 2010
24 Hosseini et al. (2016) Shabestar plain, Iran ANN-Ant colony GWL, R, E, Q, Monthly 1996–2006 R, RMSE, RAE
[193] T
25 Nourani and Mousavi, Miandoab plain, Iran WA-ANFIS, WA- GWL, P, Q Monthly 2000–2009 R2 , RMSE
(2016) [47] ANN
26 Ebrahimi and Rajaee, Qom plain, Iran WA-ANN, WA-SVR, GWL Monthly 2002–2013 RMSE, E
(2017) [194] ANN, SVR
27 Huang et al. (2017) [165] Gorges Reservoir Area, China PSO-SVM, PSO- GWL Daily 2013–2014 R2 , RMSE, NSE
BPNN
28 Barzegar et al. (2017) Azarbaijan, Iran WA-ANN GWL Monthly 2001–2018 REVIEW
[195]
29 Zare and Koch, (2018) Miandarband plain, Iran WA-ANFIS, AI GWL, P Yearly 1991–2013 RMSE, R2
[98]
30 Balavalikar et al. (2018) Brahmavar, Kundapur and Hebri In POS-ANN, ANN GWL Monthly 2000–2015 R2 , RMSE, R, MAE, MAPE
[167] Udupi district, India
31 Rakhshandehroo et al. USA WA, DE, PSO, GWL Daily 2000–2005 RMSE, MAE,PCC,NSE
(2018) [166] RBFNN, MLP
32 Malekzadeh et al. (2019) Kabodarahang region, Iran WA-SAELM, WA- GWL Monthly 1990–2015 RMSE, R, MAE, MAPE,
[176] ANN, WA-SVM. RSMRE, BIAS, NSC
33 Supreetha et al. (2019) Karnataka, India PSO-ANN, ABC-ANN GWL, P Monthly 2000–2013 RMSE, MAE, MAPE
[177]
34 Roshni et al. (2019) [162] Shikoku Island of Japan GT-WANN GWL, P Monthly 1998–2004 BIAS, RMSE, R, NSE
35 Tang et al. (2019) [130] Northern United Kingdom SVM, ANN, random GWL Hourly 2016–2017 MAE, MAPE, MSE, RMSE
forest, k-NN
36 Kombo et al. (2020) [147] Rwanda KNN-RF T, P, GWL, Daily 2016–2018 R2 , RMSE, MSE, MAE
Max. RH
37 Moravej et al. (2020) [85] Karaj plain, Iran GA-ISA, GP, ISA- GWL Monthly 2002–2008 R2 , RMSE, NS
LSSVR
38 Roshni, (2020) [178] Konan groundwater basin, Japan EANN-GA, EANN, P, GWL Monthly 1998–2004 NSE, RMSE,Bias
GRNN, FFNN
(continued on next page)
285
Table 5 (continued)
No. Author (year) Location of study Hybrid AI models Input Frequency Data Span Performance metrics
Combination
39 Banadkooki et al. (2020) Yazd, Iran RBF-WOA, MLP- R, T, GWL Monthly 2000–2012 NSE, MAE, RSR
[179] WOA
40 Natarajan and Sudheer, Andhra Pradesh, India SVM-QPSO, ANN, GWL Monthly 1997–2013 RMSE, PCC, MAE, R2
(2020) [154] SVM, GP, ELM
41 Seifi et al. (2020) [180] Ardebil plain, Iran CSO, PSO, WDA, GA, GWL Monthly 2000–2012 RMSE, MAE, NSE, PBIAS
KA, AI
2.6.1. Gaussian process regression (GPR) about each other. This process directly clarifies priorities on the
A Gaussian Process is a collection of random variables, where function space, which is called a natural extension of the Gaussian
limited numbers of them are compatible with Gaussian distribu- function with mean vector and covariance matrix. It should be
tions. A Gaussian process is determined completely by an average noted that the Gaussian distribution applies to vectors, while the
function, ðmðxÞÞ, and a covariance function, ðkðx; x0ÞÞ. This process is Gaussian process applies to functions. As a result, Gaussian process
a natural generalization from the Gaussian distribution with mean models, due to prior knowledge concerning data and functional
vector and covariance matrix, as shown in Eq. (1): dependencies, do not require any validation technique for general-
izing. Moreover, Gaussian process regression models can predict
f GPðm; kÞ ð1Þ the distributions corresponding to inputs [210]. In Gaussian pro-
regression models of the Gaussian process are based on this hypoth- cesses, X and Y denote inputs and outputs ranges, and there are n
esis that adjustment observations must include information about (xi ; yi ) pairs, which independently and identically distributed. In
each other. This process clarifies priorities on function space, which the regression, if y 2 Re then a Gaussian process on x with mean
is called a natural extension of Gaussian function with mean vector function l : Y ! Re and Covariance function k : X X ! Re is
and covariance matrix. It should be noted that Gaussian distribution defined. The main assumption of GPR is based on Eq. (2):
is on vector; however, the Gaussian process goes on functions. As a y ¼ f ðxÞ þ ; Nð0; r2 Þ ð2Þ
result, Gaussian processes models owing to prior knowledge con-
cerning data and functional dependencies, do not require any vali-
dation technique for generalizing. Moreover, Gaussian process 2.6.2. GPR implementation in the hydrology
regression models can predict distribution corresponding to inputs Koo et al. [211] stated that the primary benefit of using GPR
[210]. In Gaussian processes, X and Y denote inputs and outputs models is that the model provides not only future predictions,
ranges, there are n pair of (xi ; yi ) independently and similarly. In but also the associated uncertainty. This distinguishes GPR models
regression, if y 2 Re then a Gaussian process on x with mean func- from other statistical models, yielding original high-fidelity results
tion l : Y ! Re and Covariance function k : X X ! Re would be and a probabilistic estimate of the approximate uncertainties
defined. The main assumption of GPR is based on this equation: [212], as well as its simple structure [213], flexibility [214], and
Regression models of the Gaussian process are based on the the ability to incorporate prior knowledge of the outputs in the
hypothesis that adjustment observations must include information meta-model construction process [211]. The application of GPR
has been reported in several investigations associated with hydrol-
ogy, such as the proficiency of GPR in forecasting Monthly stream-
flow [215], construction of data-driven hydrological models [216],
prediction of short-term soil moisture [217], Modeling Pan EP
[218], estimating chlorophyll concentrations in sub-surface waters
[219], Assessment of infiltration models [220], modeling of infiltra-
tion of sandy soil [221], forecasting of reference EVP [222,223],
prediction of water temperature of rivers [224], forecasting
short-term WS [225] and seepage through earth dams [226], mon-
itoring and fault detection of wastewater treatment processes
Fig. 11. An example of a gene (genome) expressing a function 5:25ðX 1 X 2 Þ. (Samuelsson et al., 2017 [227]), Predictive Control of Drinking
Fig. 12. An example of crossover operation acting on two parents and producing two offspring.
286
Fig. 13. Mutation operation acts on genetic programming (GP) chromosome.
Table 6
List of the researches on the developed evolutionary computing predictive models for GWL modeling.
Research Applied AI Case study Data span Input Output Performance Time
models parameters parameter metrics scale
Fallah-Mehdipour et al. (2013) GP, ANFIS Karaj aquifers, Iran 7-year (84-month) GWL, P, EP GWL RMSE, NSE, R2 monthly
[104]
Kasiviswanathan et al. (2016) GP Amarawathi basin, 30 years (1980– Raf GWL GWL, CC,NE, RMSE,MBE monthly
[38] India 2009)
Sadat-Noori et al. 2020 [206] GP Tabriz plain, Iran 8-year (96- GWL, P GWL, R2 and RMSE monthly
months)
Water Networks [228], prediction of sulfate content in lakes of struct locally accurate surrogates for Bayesian experimental design
China [229], water demand forecasting [230], seawater intrusion in groundwater contaminant source identification problems. In
prediction [231], modeling adsorption equilibrium of water on this survey, without sacrificing estimation accuracy, the new
zeolite Li-LSX [232], and oceanic chlorophyll prediction [233]. approach achieved about 200 times of speed-up compared to
MCMC. Bozorg-Haddad et al. [244] utilized a GA-SVR hybrid algo-
2.6.3. Applications in GWT rithm for the simulation and prediction of GWL. Rajabi and Ketab-
The application of linear regression in the prediction of GWL has chi [245] applied GP emulation as a valuable tool for solving the
been assessed by Maatta [234]. In this survey, the importance of computational challenges of uncertainty- based simulation–opti-
the combination of statistical models in predicting GWL was mization schemes in coastal groundwater management. Their
reported. Aburub and Hadi [235] applied several data mining tech- results indicated that GP emulation can provide an acceptable level
niques to predict GWL. Their findings indicated that the SVM algo- of accuracy with no bias and low statistical dispersion. Lal and
rithm outperformed other algorithms in terms of classification Datta [246] compared the capacity of GP and GPR models for
accuracy. groundwater salinity prediction. The GPR model outperformed
The high proficiency of GPR in forming reasonable predictions the GP model. A list of articles using Gaussian Process Regression
of groundwater quality data for the majority of linear trend cases, Models for GWL modeling are tabulated in Table 7.
with a few exceptions of severely non-Gaussian data, has been
reported by Koo et al. [211]. The efficiency of GPR in producing 2.7. Complementary AI models applications
maps of GWL variability and identifying GWL patterns for the
island of Crete has been investigated by Varouchakis and P Karat- The GWL is particularly non-stationary, and noisy GWL time-
zas [236]. Kolli and Seshadri [237] relayed the ability of data min- series may not be properly simulated by AI-based models
ing techniques for the assessment of groundwater quality. [39,248]. The hybridization of ML techniques, developed based
In some studies, the accuracies of other statistical models have on wavelet-decomposed data, has become a very active research
been found to be better than that of GPR. For instance, Colchester area to address this issue. It has shown better performance in sim-
et al. [238] compared three methods for representing accelerome- ulating the raw GWL data sets than their simple model counter-
try data (wavelets, splines, and Gaussian processes) with two sys- parts [249]. A representation of the local time-series data using
tems for estimating GWL (SVR and GPR). Their results showed that scaling and wavelet coefficients at various resolutions attained
the method using splines and a SVR model provided the lowest through the Mallats pyramidal algorithm has been shown to pro-
overall errors. Dolat Kordestani et al. [239] applied evidential belief vide the discrete wavelet decomposition [249] Generally,
function and boosted regression tree (EBF-BRT) algorithms for wavelet-based ML models are more precise, as the discrete wavelet
groundwater potential mapping. Their findings demonstrated that transform (DWT) can improve discrimination against the non-
the combination of the two techniques could increase the efficacy stationary and non-linear trends that occur at different time-
of these methods in groundwater potential mapping. Similarly, series scales of input variables [249].
Pourghasemi and Beheshtirad [240] combined EBF and GIS for Wavelets are a type of mathematical function (waveform) that,
groundwater potential mapping. Azimi et al. [241] exploited Gaus- with an average of about zero, may oscillate and decay within a
sian process classification (GPC) to address the association analysis short time [249]. In order to denoise non-linear and non-
of climate-related drought and a decline in groundwater level. Kim stationary time-series and extract information that is hidden in
et al. [242] utilized the capacity of the GPR model for long-term the signal, WTs provide efficient optical signal processing tech-
predictions of GWL. Comparing the GPR and ANFIS models in niques [195,248]. They were conceived as the continuous wavelet
multi-step lead time forecasting of GWL revealed that the GPR transform (CWT) [250]. The discrete wavelet transform (DWT) has
model provided reasonably accurate predictions compared to been proposed for practical applications, as hydrologists and mod-
those of ANFIS [90]. Zhang et al. [243] integrated Gaussian process elers generally only have access to discrete-time signals [39]. In the
(GP) and Markov Chain Monte Carlo (MCMC) to adaptively con- DWT, a signal is decomposed to an approximation at the first
287
H. Tao, Mohammed Majeed Hameed, Haydar Abdulameer Marhoon et al.
Table 7
The conducted researches using Gaussian Process Regression Model for GWL modeling.Abbreviations: Gaussian process (GP), Genetic Algorithm (GA), Support Vector Regression (SVR), Naïve Bayes (NB), K-Nearest Neighbor (kNN),
Classification Based on Association Rule (CBA), evidential belief function and boosted regression tree (EBF-BRT), receiver operating characteristics (ROC), Gaussian process classification (GPC), artificial neural network (ANN), Groundwater
Resources Index (GRI), Standardized Precipitation Index (SPI), Adaptive Neuro Fuzzy Inference System (ANFIS), Correlation Coefficient (CC) Root Mean Squared Error (RMSE), Nash–Sutcliffe Efficiency (NSE)
Research Applied AI models Case study Data span Input parameters Output parameter Performance
metrics
M. M. Rajabi and Gaussia GP emulation Kish Island the period of the KL, aL and RNet quantities of
Ketabchi [245] (Persian Gulf) simulated dataset
interest (QoI). Time saving ratio (TSR)
OmidBozorg Haddad GA-SVR hybrid algorithm Karaj plain 2002–2008 EP, prediction, groundwater and surface data groundwater level R2 and RMSE
et al. [244] aquifer (Iran)
KamakKolli and Data mining technique on Arc/ Tadepalle, – physico-chemical
Seshadri [237] View software Guntur district
parameters like TDS,
TH, Cl
and NO3 Water Quality Index –
288
[211] GPR Pyeongchang 2007–2012 groundwater level (GWL), pH, Total dissolved solid (TDS), Groundwater Quality confidence
Yuchyeon T, CL- SO4, NO3-N, and NH4 intervals
Farah Colchester et al. SVR, GPR Kenya April and November, S,V groundwater depth median error
[238] 2014
Aburub and Hadi SVM, NB, KNN, CBA. Jordan – elevation, faults, Raf, slope, T, wadis and outcrop groundwater areas Accuracy,
[247] Precision and F1
Emman Varouchakis GPR island of Crete, groundwater level spatio- temporal
et al. [236] Greece variability
Dolat Kordestani EBF-BRT algorithms Lordegan aquifer 2014 EBF values of the groundwater- conditioning factors (GCFs) groundwater potential mapping. ROC test
et al. [239] (Iran)
Azimi et al. [241] GPC, ANN 609 study plains 2017 to 2019 pair of a statistical average of the value of each SPI and GRI
in Iran for each
plain. groundwater drought
Kim et al. [211] GPR Han River Basin 2004 to 2015 monthly averaged groundwater level Groundwater Level Trend confidence
intervals
Raghav and Deka [90] GPR, ANFIS Sullia Taluk, 2000 to 2013 Monthly ground water level time series up to previous multistep lead time forecasting of CC, RMSE, NSE
India Four time steps groundwater levels
Neurocomputing 489 (2022) 271–308

decomposition level, which is then iteratively applied to subse- sample observations at each transformation step. Although an
quent decompositions [250]. accurate orthogonal decomposition of the time-series is not given
The DWT is considered to be a sampled version of the CWT. In by MODWT, it is more effective than the simple DWT, as it can be
association with a specific dyadic scale and time, any time-series performed for any sample size [195].
can be commonly re-expressed in terms of the DWT coefficient WT relies on a fully adaptable window function (named the
[249]. It divides a given function into different scale components, mother wavelet), which can be adjusted over time, depending on
where a frequency range can also be allocated to each scale compo- the compactness and shape of the signal [195]. In the literature,
nent, allowing the time-series to be viewed at multiple resolutions, many mother wavelets exist and their choice depends on the data
thus enabling hydrologists to analyze each component with a res- set(s) to be examined [19]. The type of time-series is the key crite-
olution suited to its scale [39,195,248,249]. The scale here refers to rion for choosing a mother wavelet [39]. The key elements of a
the time interval of that specified time series, while the number of mother wavelet comprise the support area, the association with
recurring oscillations over a unit of time is denoted as the fre- the length of the wavelet period, and the number of missing
quency [249]. A signal (time-series) is decomposed by the DWT moments, which regulate the wavelet’s ability to display informa-
into non-sinusoidal components that provide adequate information in a time-series [39].
tion for both synthesis and analysis of the raw signal (i.e., the Numerous researchers have employed the WT for GWL model-
time-series). It is possible to select or build the wavelet form to ing, as shown in the literature provided in Table 8. In predicting
fit the time-series signal outline [249]. To improve the accuracy multi-step lead time GWLs across two neighboring micro-
of AI-based models, the WT is usually suggested as a method to watersheds-namely, Pavanje and Gurpura-along the coastline of
pre-process the time-series [39]. Barzegar et al. [195], on the other Karnataka, Rezaie-balf et al. [249] employed the WT to develop
hand, used the maximal overlap discrete wavelet transform hybrid Wavelet-M5 Model Trees (W-MT) and Wavelet-
(MODWT) for time-series decomposition without a dyadic dura- Multivariate Adaptive Regression Splines (W-MARS) models. To
tion. The MODWT is comparable to the DWT, in that the input sig- decompose the input time-series, they utilized Haar, Daubechies,
nal at each step is added to low- and high-pass filters. Here, the Dmey, and Coiflets as mother wavelets. The W-MARS and W-MT
coefficients are not decimated by the MODWT, and the number models were found to provide accurate forecasting, as opposed
of wavelets and scaling coefficients are similar to the number of to the standard MARS and MT models. In improving the forecasting
Table 8
The literature review researches on the complementary ML models for GWL modeling
AI models Case study Data span Input Output Performance metrics

parameters parameter
[249] W-MARS, W-MT Karnataka, India August 1996 - July GWL, Raf, T GWL R2 , RMSE, NNSE
2006
[39] WGEP, WM5 Lorestan Province, 2002–2012 GWL, T, P GWL R2 , RMSE, rRMSE, BIAS, rBIAS, AIC
Iran.
[252] EEMD-GEP, EEMD- Delfan plain, Iran 2002 to 2012 GWL, P, T GWL R2 , RMSE, rRMSE, BIAS, rBIAS
M5, CEEMD-GEP,
CEEMD-M5
[195] WA-GMDH, WA- Maragheh- Bonab, Sep 1985- Mar 2016 GWL GWL R2 , RMSE, NSC
ELM Iran
[248] WP-SVR Mangalore, India 1996–2006 GWL, Raf, T, GWL NRMSE, Normalized Mean Bias, Absolute Relative Error,
Tidal Level NSC, Threshold Statistics, R2
[194] wavelet-ANN, Qom plain, Iran April 2002 - March GWL GWL E, RMSE
wavelet-MLR, 2013
wavelet-SVR
[107] LSSVM, ANFIS, Shabestar Plain, Climate Data (1951– P, T Future R2 , RMSE
NARX Iran 2016) GWL (April ground-
2002- March 2016) water level
[254] EMD, PSR, PSO, ELM, Heilongjiang 1998 to 2014 groundwater ground- posterior error ratio (C), small error frequency (p),
PSO-ELM, EMD-PSR- Province, China depth water depth relative mean square error (E1), fitting accuracy ratio
PSO-ELM prediction (E2), test forecast effect index (E3).
[176] SAELM, WA-SAELM Kabodarahang August 1990 to GWL GWL R, RMSE, NSC
region, Iran September 2015
[255] ANN, ANFIS, Mashhad, 1992 to 2007 total P, EP, GWL R2 , RMSE, NSC
Wavelet-ANN, Khorasan Razavi discharge,
Wavelet-ANFIS province, Iran GWL
[191] FFNN, WT-FFNN, Ardabil, north- GWL GWL RMSE, R2
ARIMAX western Iran
[256] XGBT, XGBL, WT- Kumamoto City, 1980–2017 GWL GWL MSE, MAE, RMSE, RSR, R2 , NSE, KGE
XGBT, WT-XGBL, Kyushu Island,
WT-RF Japan
[185] WA-SVR Visakhapatnam, May 2001- February groundwater GWL R2 , NSC, NMSE, RMSE, MAPE
India 2012 depth, P,
T max , T mean
[257] EEMD-ANN, EEMD- Lake Okeechobee, 1997 to 2012 GWL GWL R, NMSE, RMSE, NSC, AIC
SVM, EEMD-ANFIS Florida
[258] WA-ANN Zhangye basin, June 2003- December GWL, P, EP, GWL R, MAE, RMSE, NSC, RSR
China 2010 T mean
[259] wavelet- SAELM Kermanshah, Iran 2002–2015 GWL GWL R, RMSE, MAE, MAPE, NSC
(Sarab Qanbar),
[98] Wavelet-ANFIS Miandarband October 1991- June P, GWL R2 , RMSE
plain, Kermanshah 2013 piezometric
province, Iran head data
289
efficacy, the W-MARS models outperformed the W-MT and their els, and the same architecture coupled with the wavelet decompo-
respective simple corresponding models. Furthermore, compared sition, i.e., WANFIS, W-LSSVM, and W-NARX. The six models were
with the other models, the W-MARS models provided relatively developed using P and T mean . Tt was reported that both models
good six-month lead-time forecasts for GWL. Two hybrid models were able to predict GWL accurately; however, the W-NARX was
have been developed by Bahmani et al. [39]-wavelet gene expres- more accurate and exhibited high R2 (0.99) and low RMSE
sion programming (WGEP) and WMT-in order to simulate monthly (0.03) values. Wen et al. [105] analyzed the importance of the
GWL at three groundwater wells in Iran. To decompose the time wavelet transform in improving the accuracy of the ANFIS model
signals, Haar, Coif1, Sym3, Db4, and Db2 wavelets, which have used for predicting weakly GWL in the Laizhou bay, China, and
been widely used in hydrological studies, were adopted. The study found that both ANFIS and WANFIS can provide good prediction
revealed that the hybrid models-WGEP and WM5-showed an accuracy. They also reported the superiority of the WANFIS com-
improved performance over their simple models-GEP and M5- pared the FFNN and ANFIS having a R2 , RMSE and MARE of 0.983,
while the performance was comparable between the hybrid mod- 0.062, and 2.48, respectively. More recently, some interesting
els. It was also reported that the choice of an appropriate level of research has been conducted by combining ANFIS with wavelet
decomposition significantly affects the hybrid model’s accuracy. transform. Sridharam et al. [106] used several input variables
The use of WGEP to pre-process a time-series and simulate GWL, namely, IL, P, T mean , GWL, and EVP for modelling GWL at daily times
compared to the hybridization of GEP with Ensemble Empirical step. They compared he performance of two AI models: the ANFIS
Mode Decomposition (EEMD) [251] and Complementary Ensemble and a hybrid WANFIS combining wavelet transform and ANFIS. The
Empirical Mode Decomposition (CEEMD), was also supported in two models were evaluated and compared using several perfor-
another study by Bahmani and Ouarda [252]. The study by Barze-
mances metrics, i.e., RMSE, MAE and R2 , and it was found that
gar et al. [195] demonstrated the effectiveness of the hybrid
hybridizing the ANFIS using a wavelet decomposition method con-
wavelet-group data handling (WA-GMDH) and WA-EL models,
tributed significantly to the improvement of the models perfor-
with high-level wavelet filters, which delivered more reliable fore-
mances for which the RMSE and MAE were significantly
casts than those obtained using low-level wavelet filters. They also
observed that, when using the least-squares boosting (LSBoost) decreased and the R2 was increased from 0.924 to 0.962 during
algorithm, ensemble multi-wavelet models can improve the per- the testing phase.
formance of the single wavelet-based model and lessen the fore- Based on a recent review on the employment of WT to develop a
cast uncertainty. Sujay Raghavendra and Deka [248] hybrid model to simulate GWL, we found that the WT hybrid mod-
demonstrated that the Wavelet packet-SVR (WP-SVR) model per- els, in all cases, performed better than their corresponding simple
forms better than the classic SVR model for forecasting monthly counterparts. This is due to the capability of the WT to act as a pre-
GWL fluctuations. They also found that better results were pro- processing tool in discriminating the non-linear and non-
duced by the wavelet packet coefficients of the Daubechies 4 stationary trends in the time-series which usually persist in hydro-
wavelet with level 4 decomposition. Use of the hybrid wavelet- logical and climatological input variables. The WT hybrid models
neural network (WNN), wavelet-linear regression (WLR), and are also able to take in any sample size, with the ability to view
wavelet-SVR (WSVR) models for monthly GWL simulation was the scaled component of the time-series at multiple resolutions.
tested by Ebrahimi and Rajaee [194] for two wells in the Qom As a wide selection of mother wavelets exists, a further compara-
plain, Iran. They reported that, by analyzing information at two tive study is needed to improve the wavelet selection in the devel-
decomposition levels, the wavelet- transformed data enhanced opment of WT hybrid models. Previous findings have also stressed
the training of the TDNN, MLR, and SVR models. They also found the importance of selecting a suitable decomposition level, as it
that wavelet types, such as Meyer’s mother wavelet, showed a sim- may affect the model’s accuracy.
ilar behavior at different well locations. The results of Jeihouni
et al. [107] revealed that the hybrid technique of the wavelet 2.8. Statistical models applications
Non-linear Autoregressive Network with Exogenous inputs
(wavelet-NARX) gave the best results, in most cases, in comparison Statistical models have been widely used in various aspects of
with wavelet-ANFIS (WA-ANFIS) and other models. For daily water hydrological modeling [260,261]. In this type of modeling, the rela-
level prediction in reservoirs, Seo et al. [253] researched the effi- tionships between one or more variables are mathematically
ciency of the combination of WT with ANN and ANFIS. They discov- embodied, in order to mimic the behavior of the real system. These
ered that the efficiency of the hybrid models was higher than that relationships are mainly set using function minimization proce-
of their corresponding simple models. The details of other studies dures, which minimizes the sum of squared residuals between
that have demonstrated the capability of hybridization of WT with the observed and modeled target variables. As such, in statistical
other models are given in Table 8. modeling, regression analysis and time-series analysis are two
Several other researchers used this approach of transforming methods that employ such a minimization process. The bivariate
input signal using wavelet decomposition [47,107,105]. These analysis of time-series differs from that of regression analysis, in
examples highlight how wavelet transform has the potential to which the time is used as the independent or predictor variable.
be used for improving the ANFIS performances used for GWL pre- Meanwhile, in regression, the bivariate analysis is represented
diction. Nourani and Mousavi [47] compared all possible combina- between two or more statistically associated variables. Further,
tions of P; Q and GWL decomposed using wavelet transform independence among the individual measurements are assumed
coherence and combined for predicting monthly GWL data in the bivariate form of regression. In other words, the order of
acquired over the period ranging between 2001 and 2011 in the the predictor-predictand data pairs is not important in bivariate
Miandoab plain, northwest of Iran. The decomposed signal was regression; whereas, in time-series analysis, the time dependence
used as input for the ANFIS and FFNN models and the best accuracy is recognized and used to improve the understanding of the under-
was obtained using hybrid WANFIS with R2 and RMSE of 0.940 and lying physical processes and/or the prediction accuracy [262].
0.084, respectively, compared with the 0.93 and 0.095 obtained A time-series model is stochastically handled without consider-
using the hybrid WFFNN. Jeihouni et al. [107] tested the impor- ing the inherited physical nature of the time-series [263,264]; that
tance of the wavelet decomposition in improving the prediction is to say, it conceptualizes the physical process of any time-series
accuracy of AI models by comparingANFIS, LSSVM, and NARX mod- into a mathematical model. Thus, it requires adequate knowledge
of the mathematical approaches for identifying time-series pat-
290
terns. Autoregressive (AR), moving average (MA), ARMA, ARIMA, through the model identification phase, along with non-seasonal
and SARIMA are commonly used methods for time-series model- ones. The seasonal differencing is applied to a seasonal time-
ing. In the following, the individual description of each model is series to remove the seasonal component. Therefore, the SARIMA
presented: model is typically denoted as ARIMAðp; d; qÞðps; qs; dsÞ, where
ps; qs, and ds are the seasonal AR, MA, and differencing parameters,
2.8.1. Autoregressive process ARðpÞ respectively [103].
The serial dependence of data points in time series is repre- The Box–Jenkins method, developed by Box and Jenkins [264],
sented by AutoRegressive process ARðpÞ. ARðpÞ (Eq. 3) describes is used to identify the best fit of the time-series model on past
the linear combination of the highest autoregression order (p) coef- observation values for the five stochastic models (AR, MA, ARMA,
ficients of consecutive data points of the time series [265]: ARIMA, and SARIMA). The methodology of time-series modeling
can be summarized as follows [264].
xt ¼ x þ /1 xðt1Þ þ /2 xðt2Þ þ þ /p xðtpÞ þ t ð3Þ
where, xt is the variable value of x at time t; x is the sample variable Model identification: in this step, the ACF and PACF functions
means; /1 ; /2 ; /p are the autoregressive model parameters; t is the are employed to determine the order of AR and MA parameters.
white noise error; p refers to the order of the autoregression. The AR Parameter estimation: in this step, the computation algorithms
model is used when the time series is stationary. Therefore, it is are employed to find the model parameters coefficients that
worth to assess stationary where the AR model parameters should best fit the model. Generally, the minimization function of the
be within 1, hence the influence of antecedent values is hindered. sum of squares of the residuals is employed using either the
Otherwise, the accumulated error from the previous values shifts approximate maximum likelihood method [270], the approxi-
the time series into a non-stationary one. mate maximum likelihood method with backcasting, or the
exact maximum likelihood method [271].
2.8.2. Moving average process MAðqÞ Model-checking: when the model structure and estimation
Besides the serial dependence of data points considered in the parameters values are completed, it is of critical importance
autoregressive process, the time series might be influenced by to check whether the built model conforms to the stationary
the antecedent random error (white noise error) involved in prior univariate process. In other words, the reliable model should
data points. This could be accounted for through the moving aver- produce statistically independent residuals that contain only
age (MA)(q) process (4) which is made of the random error compo- white noise error and no systematic error. Besides, the model
nent and a linear combination of random shocks of the antecedent should provide accurate forecasts sufficiently. The portmanteau
values [266]. lack-of-fit test-statistic [272] is typically used for the diagnostic
purposes of the built model where the behavior of the esti-
xt ¼ x þ t h1 ðt1Þ h2 ðt2Þ hq ðtqÞ ð4Þ mated residual is checked to confirm that the realizations are
where, h1 ; h2 ; hq are the moving average parameters; t ; ðt1Þ ; ðtqÞ approximately from a white noise process. A comparison of
are the random error components at ðt 1Þ; ðt 2Þ; ðt qÞ, respec- the forecasts with the measured data points can be further used
tively; q is the order highest moving average process. The MA model to check the accuracy of generated forecasts.
parameters require to invert to overcome the duality of the moving Forecasting: In the last step, the model is employed to compute
average process and the autoregressive process [267]. The new data points, which beyond those included in the input time
inevitability condition of a moving average process is analogous series.
to the stationarity condition of an autoregressive process.
2.8.6. Literature review

2.8.3. Autoregressive moving average ARMAðp; qÞ
In the literature, many successful applications of statistical
The real stochastic process of a random variable xt is repre-
models, including logistic regression [273], k-NN [274], linear dis-
sented by the ARMA process. The ARMA model is a combination
criminant analysis [274], quadratic discriminate analysis [275],
of AR and MA of order p and q, respectively [268]. The general form
multivariate adaptive regression spline (MARS) [276], and regres-
of ARMA is as follows (5):
sion trees [275], in hydrology have been reported. However,
xt ¼ x þ /1 xðt1Þ þ /2 xðt2Þ þ þ /p xðtpÞ þ t h1 ðt1Þ time-series models such as autoregressive, Moving average,
h2 ðt2Þ hq ðtqÞ ð5Þ autoregressive moving average, autoregressive integrated moving
average, and seasonal autoregressive integrated moving averages
An ARMA model with ðp; 0Þ is an autoregressive process only. have been extensively applied to predict the present and to fore-
While an ARMA of ð0; qÞ is a purely moving average process. cast future values in GWL series.
Mirza and Ghazavi [277] applied the five time-series models of
2.8.4. Autoregressive integrated moving average ARIMAðp; q; dÞ AR, MA, ARMA, ARIMA, and SARIMA to predict the monthly GWL of
An ARIMA model is presented as combined process of AR and 36 wells located in Isfahan province, Iran. The monthly GWL from
MA. In contrast to ARMA, the differencing step is applied in the 1990 to 2004 were first clustered using the Vard algorithm of the
ARIMA model once or more to eliminate the non-stationarity in hierarchy method to classify the true groups of piezometric wells
the time-series points. Differencing, in statistics, is a transforma- into five groups, according to their similarities to each other. The
tion applied to a non-stationary time-series to make it stationary performance of the five models was investigated through 11 differ-
and to remove the non-constant trend. Therefore, the ARIMA ent structures, according to the lag time and differencing pro-
model has three specific parameters, p; q, and d, where d represents cesses. They concluded that time-series models are one of the
the number of differencing passes [269]. appropriate methods which could be of use to forecast the GWL.
The AR with 2-lag showed the best forecasting of GWL for
2.8.5. Seasonal Autoregressive integrated moving average 60 months ahead for the five clusters.
SARIMAðp; q; dÞðps; qs; dsÞ Choubin and Malekian [278] compared the results of ANN and
SARIMA is the generalized form of the ordinary ARIMA model. It ARIMA models for GWL forecasting 4 months ahead in the Shiraz
is used when there is a seasonal pattern in the time-series. The sea- basin, southwestern Iran. The monthly time-series of GWL over
sonal parameters in SARIMA are estimated once they are identified the period 1993 to 2010, in addition to that of total P, monthly
291
average streamflow, T, and EP, were employed to set up the mod- tant class within the discrete-time non-linear group models. NARX
els. Gamma and M tests were used to identify the optimal input networks have several advantages; for example, they are computa-
parameters and the length of the training data, respectively. They tionally powerful and are useful for modeling an extensive variety
reported the superior performance of ARIMA with p; q, and d values of dynamic systems [285]. NARX networks are recurrent dynamic
of 2, 1, and 2, respectively, in predicting GWL one month ahead. neural networks which have feedback connections. Based on previ-
Gibrilla et al. [279] applied the ARIMA model to measure GWLt ous studies, it was noted that the gradient descent learning algo-
every six hours in seven monitoring wells from 2005 to 2014 in rithm may be more efficient within NARX, compared to other
the Upper East Region of Ghana. The results revealed that the cur- networks [286]. NARX networks can be modeled using two differ-
rent demand in the region could be sustained under the current ent types of architectures: Parallel and series–parallel [287,288].
and predicted GWL trends. Sakizadeh et al. [280] investigated the The latter type of architecture is purely feed-forward and, thus, it
performance of SARIMA and Holt-Winters Exponential Smoothing can utilize static back-propagation during the training phase. The
(HWES) methods using GWL records from 28 representative series–parallel architecture is generally expected to provide more
piezometers between 1984 and 2012 in the Malayer Aquifer. They accurate results with the use of accurate inputs. Meanwhile, in
concluded that the SARIMA technique provides further improve- the parallel architecture, the past predicted output is utilized as
ments over HWES. Therefore, the optimized SARIMA model was part of the input combination to predict the output value at the
used to predict the time-series for the next 4 years (i.e., from next step [289]. As part of the current research, we report previous
2012 to 2016). Takafuji et al. [281] compared the performance of studies that have employed the NARX model for GWL prediction.
ARIMA and a geostatistical method using sequential Gaussian sim- The groundwater information is an important issue for
ulation (SGS) for predicting GWL at 49 wells in the Bauru Aquifer decision-makers in the agricultural area of Mississippi. Therefore,
System domain in Sao Paulo State, Brazil. They found that, for mon- [288] applied a new model, called NARX, for simulating daily
itoring the aquifer, the ARIMA models performed more favorably GWL in the Mississippi River Valley Alluvial (MRVA) aquifer in
than SGS, as they achieved the same accuracy level as SGS and a the southeastern United States. Two different algorithms-namely,
higher precision for all periods. Furthermore, they reported that LM and Bayesian Regularization (BR)-were employed to train the
time-series models can be optimized automatically by using the NARX network. Several modeling architectures were created, uti-
Akaike information criterion, which provides a precise and accu- lizing different hidden node combinations and delays (5, 25, 50,
rate trade-off to choose among the models. Goodarzi [282] evalu- 75, and 100). A comparison between the obtained results was car-
ated the prediction of monthly GWL data at 58 piezometric wells ried out to find the optimal network. Daily historical time series for
for the period of 1995 to 2010 using ANN, HARTT, and SARIMA P and GWL over eight years were considered for GWL forecasting,
models in the Najafabad plain, Iran. They concluded that, though up to three months ahead. The results showed that NARX-BR learn-
the three models were capable of predicting the GWL, the SARIMA ing was better than the NARX-LM network for daily GWL forecast-
models were more appropriate than the other evaluated methods, ing, according to several statistical indicators. The most accurate
as they showed lower error. Therefore, it can be summarized that forecasting results were attained by BR with two hidden nodes
time-series models are capable of estimating the GWL with rela- and 100 time delays.
tively good accuracy. A list of articles proposing applied statistical The performance of the NARX model has been investigated to
models for groundwater level modeling is given in Table 9. forecast GWL by [290]. The proposed model was applied to forecast
GWLs at several wells located in southwest Germany. Two differ-
2.9. Non-linear auto-regressive network with exogenous input (NARX) ent parameters, P and T, were considered as predictors. Several sta-
model for GWL modeling tistical indicators, such as RMSE, Nash, and R2 , were utilized to
evaluate the performance of proposed models. The results indi-
FFN is one of the most common types of ANN model. An FFN cated the outstanding efficiency of the NARX model for GWL fore-
consists of one input layer, several hidden layers, and one output casting under a small set of input parameters.
layer. A distinctive feature of an FFN model is that the connections In 2019, [287] employed an autoregressive neural network
between layers permit forward information flow only [283,284]. (NNARx) for GWL forecasting in an aquifer system. The accuracy
The NARX model is a special type of RNN model, and is an impor- of the proposed model was compared with the autoregressive with
Table 9
The applied statistical models for groundwater level modeling.
Research Applied AI models Case study Data span Input Output Performance metrics
parameters parameter
Mirzavand and AR, MA, ARMA, ARIMA, SARIMA Isfahan province, Iran 1990–2004 Antecedent Present R2 , AIC
Ghazavi values of value of
[277] GWL GWL
Choubin and ANN, ARIMA Shiraz basin, southwestern 1993–2010 Antecedent Present RMSE, MAE, R
Malekian Iran values of value of
[278] GWL GWL
Gibrilla et al. ARIMA Upper East Region of Ghana 2005–2014 Antecedent Present R2 , RMSE, MAPE, MAE, MaxPE,
[279] values of value of MaxAE, Ljung-Box Q statistics
GWL GWL
Sakizadeh SARIMA, HWES Malayer Aquifer 1984–2012 Antecedent Present ME, RMSE, MAE, MPE, MAPE,
et al. [280] values of value of MASE, ACF1
GWL GWL
Takafuji et al. ARIMA, a geostatistical method Bauru Aquifer System September 2014 Antecedent Present RMSE
[281] using sequential Gaussian domain in Sao Paulo State, until 30, 2015). values of value of
simulation Brazil GWL GWL
Goodarzi [282] the artificial neural networks, Najafabad plain, Iran 1995–2010 Antecedent Present RMSE, R2
HARTT model, SARIMA values of value of
GWL GWL
292
Table 10
The established research on the implementation of NARX model for GWL prediction.
Models Case Study Data Span Time Input Parameters Output Perform- ance Best
Location Scale Parameter Metrics Model
Guzman et al. (2017) [288] NARX-LM, USA 1987– Daily P, GWL GWL MSE, R2 , NSE NARX-
NARX-BR 1994 BR
Wunsch et al. (2018) [290] NARX, ANN Germany 1948– Weekly P, T GWL RMSE, R2 , NSE, NARX
2008 RMSEr
Zanotti et al. (2019) [287] NNARx, ARx Italy Not Daily P, T GWL MSE, RMSE, NSE, NNARx
reported KGE, AIC
Di Nunno and Granata (2020) NARX, NARX- Italy 2008– Daily Raf, Evapotran- GWL RMSE, R2 , MAE, RAE NARX-
[291] BR 2012 spiration BR
Al Jami et al. (2020) [292] NARX, ANN Bangladesh 1980– Monthly P, T, H% GWL MSE, NSE, R2 NARX
2013
exogenous input (ARx) model. Early stopping and Bayesian regu- Nalarajan and Mohandas (2015) [294] presented an M5-MT
larization methods, as well as a combination of both, were utilized model to predict monthly GWL in Jones Island, New Jersey, USA.
for training the forecasting models to avoid over-fitting. The results A historical data series of GWL values were collected from moni-
showed that, for short-term forecasting, the performances of toring wells between 1970 and 2010. The data from 1995–2009
NNARx and ARx models were comparable, with a slightly better was used to train the developed model, whereas the remaining
performance of ARx model. However, For the long-term, the were used to validate the model. Previous values of monthly
NNARx model which was trained by the Bayesian regularization GWL (up to 12 months) were selected as the appropriate input
method was superior to ARx and other NNARx models. The authors combinations. Observed and predicted results were compared
concluded that the linear model needs less time and does not using R2 and RMSE. The results showed that the developed M5-
require high computational power. They found that suitable and MT model has strong accuracy in predicting the GWL. However,
reliable models for short- and long-term GWL forecasting are lin- the authors indicated that the effects of some other parameters,
ear and neural network methods, respectively. such as P, the permeability of the soil, soil moisture, and soil tem-
Daily GWL prediction was carried out, using the NARX method perature, are important for GWL prediction.
[291], at 76 wells located in the Apulian territory. Several input Zhao et al. (2016) [295] presented a methodology to predict
parameters, including Raf, EVP, and input time delay, were consid- GWL, in which a CART model was developed with the lagged value
ered for modeling. A comprehensive analysis of the results was car- of Raf, reservoir level, and change of GWL as input variables. The
ried out to discern the optimal predictive model. The results required data were collected between 2005–2007 for the Three
supported the reliability of the NARX-BR model to predict GWL. Gorges Dam Reservoir area. An SVM model was also developed
The performance of NARX model, in terms of monthly GWL predic- for the same purpose, in order to compare the performance of
tion, was examined in [292]. Three different algorithms were used the proposed CART model. Absolute error (AE) and relative error
to train NARX and a comparison between their results was con- (RE) metrics were used to evaluate the model performance. Based
ducted by utilizing several statistical indicators. The results on the comparison of observed and predicted values of the two
revealed that combining Bayesian Regularization as a training models, it was concluded that the proposed CART model could pre-
algorithm with NARX could provide good prediction ability. Such dict GWL better than the SVM model.
a model can provide important information for GWL prediction. Kaya et al. (2018) [296] proposed an M5-MT model and a feed-
The established research on the implementation of NARX models forward back-propagation ANN model to predict GWL in the Rey-
for GWL prediction are tabulated in Table 10. hanlí region, Turkey. Historical GWL data were collected from a
well located in the study area between 2000 and 2015. Monthly
P and T data were taken from Antakya Meteorological Station,
2.10. Other ML models applied for GWL prediction which is governed by the DSI (General State of Hydraulic Work).
Evaluation of the proposed model’s performance was carried out
Naghibi et al. (2015) [293] developed three different machine using R, MSE, and MAE. The results showed that the proposed
learning models-BRT, classification and regression tree (CART), ANN method had a similar accuracy as the M5-tree model.
and RF models-for mapping groundwater spring potential in the Wang et al. (2018) [297] proposed a hybrid model to predict
Kohrang Watershed, Iran. To perform this study, a large set of fac- GWL, which combines the canonical correlation forest algorithm
tors, including hydrological, geological, and physiographical fac- with random features, referred to as the CCA-CRF model. The per-
tors, were selected as critical factors affecting spring occurrence. formance of the proposed model was compared with the random
A GIS-based spring location map was prepared, using the topo- forest regression (RFR) and LS-SVR models. Historical daily G and
graphic maps obtained from the National Cartographic Center of T measurements collected for the Daguhe River in Qingdao, China
Iran (NCCI), which included 864 spring locations in the study area. were used as predictors of GWL. Several input combinations were
The data was divided with a ratio of 70% and 30% for the training examined, in order to find the best time lags to predict 1-, 3-, 5-, 7-,
and model validation sets, respectively. The model performance and 10-day-ahead GWL. The results of the models were evaluated
was evaluated using the receiver operating characteristic (ROC) using the R, RMSE, and MAE metrics. It was observed that the LS-
curve. The results showed that the BRT model had higher accuracy SVR model had higher performance for 1-day ahead prediction,
than the CART and RF models, considering the model performance while the proposed CCA-CRF model was superior in predicting
evaluation results. Based on the discussion of the authors, the RF GWL in the longer-term. Furthermore, the authors indicated that
model was not superior to the others, although the RF has shown the proposed CCA-CRF model was faster than the others.
superior performance in the previous literature. It was also Sharafati et al. (2020) [298] predicted GWL over the Rafsanjan
reported that factors such as altitude, drainage density, and slope aquifer, Iran, using a gradient boosted regression (GBR) model.
degree were the most effective factors for predicting spring Satellite-based products, including hydro-geological and climatic
occurrence. factors, were derived as predictors and several input combinations
293
Table 11
The applied other ML models for GWL modeling
Research Applied AI Case study Data span Input parameters Output Performance Data mode
models parameter metrics
Naghibi et al. BRT, Koohrang Watershed, NONE A huge set of hydrological, Potential ROC NONE
(2015) [293] CART, RF Chaharmahal-e- Bakhtiari geological, and physiographical Ground water
Province, Iran factors. spring
Nalarajan and M5 - MT Jones Island, New Jersey, USA. 15 years GWL GWL R2 , RMSE monthly
Mohandas (1995–
(2015) [294] 2010)
Zhao et al. (2016) CART, Three Gorges Dam Reservoir, 2005– Raf, Reservoir Level, GWL Landslide AE, RE Both daily
[295] SVM China. 2007 change GWL and
monthly
Kaya et al. (2018) M5 - MT Reyhanli region, Turkey 15-years Raf, T, GWL GWL R, RMSE, monthly
[296] (2000– MAE
2015)
Wang et al. (2018) CCA-CRF, Daguhe River, Qingdao, China 1 year Raf, T, GWL GWL R, RMSE, daily
[297] LS-SVR, (2013 MAE
RFR 2014)
Sharafati et al. GBR Rafsanjan aquifer, Iran 9 years Satellite-based products of GWL R2 , NRMSE monthly
(2020) [298] (2007 hydro-geological, climatic
2016) factors.
Javadinejad et al. MLPNN, Micro-watershed, Gurpura 10-year Raf, T, GWL GWL R, RMSE, RAE monthly
(2020) [299] M5-MT River Basin, Ganjimatta Region, (1996–
India 2006)
were selected using the gamma test (GT) method. The developed variables to predict GWL. It was also found that the majority
GBR model was used to predict 1-, 3-, and 6-month-ahead GWL of the authors considered the monthly and daily time scale for
in the study area. Thus, both short- and long-term values were pre- GWL modeling.
dicted. The R2 and normalized RMSE (NRMSE) metrics were used ANN models can easily be extended from univariate to multi-
to evaluate model performance, by comparing observed and pre- variate cases, compared to other conceptual models. Moreover,
dicted GWL values. It was observed that the developed model the complexity of ANN models can be varied simply, through
can be accurately used for GWL studies. changing the learning algorithms, transfer functions, and model
Javadinejad et al. (2020) [299] developed two different machine structure. Similar to regression models, the input variables
learning techniques-MLP-NN and M5-MT models-to predict the might be assigned by using correlation analyses or empirical
monthly GWL fluctuations for a micro-watershed of the Gurpura proof. In accordance with the reviewed papers, the results
river basin, India. For this purpose, first, monthly GWL values dur- showed that ANN models can efficiently predict the GWL and
ing the period of 1996–2006 were collected from a well in the Gan- capture the non-linear behaviors of the GWL in different regions
jimatta region. Then, two different scenarios were applied, in order and case studies, compared to other models such as ARMA and
to determine appropriate input combinations of the MLP-NN GM(1, 1).
model. Based on these scenarios, the one lagged time and present Based on the reviewed papers collected from 2008 to 2020, MLP
values of monthly GWL, T, and Raf were found to be the appropri- has been much more popular than other modeling approaches,
ate predictors of the MLPNN model. Similarly, an investigation was such as RBFNN and ELM, in GWL modeling. The most interesting
performed to determine the best input combinations, using linear observation from these studies is that the LM algorithm is more
rules to develop the M5-MT model. The observed and predicted popular than other algorithms in training ANNs for forecasting
values were compared in terms of R2 , RMSE, NSE, and RE. The GWL. The LM algorithm is one of the most efficient learning
results showed that the developed M5-MT model was superior to algorithms, as it can interpolate between the Gauss–Newton
the MLP-NN model to predict GWL fluctuations. A list of articles algorithm and the gradient descent method. The robustness of
discussing other applied ML models for GWL modeling is given the LM algorithm, in many cases, comes from its capability to
in Table 11. reach an optimal solution, even if starts very far off the final
minimum. The LM algorithm has many other advantages, such
3. Literature review assessment as faster convergence and lower probability to become stuck
in local minima than other learning algorithms used to train
It is important for survey research to assess and evaluate the MLP.
established literature and provide a comprehensive debate that In modeling GWL using MLP with a single hidden layer,
can benefit the interested readers. In this section, several essential researchers most frequently used the sigmoid transfer function
points are abstracted and discussed, in light of the reported review in the output layer. It is important to mention that other
on machine learning-based models for GWL modeling. researchers have also used the hyperbolic tangent sigmoid
transfer function in the hidden layer; however, it is not as fre-
All tables (Tables 1–11) summarize the details of the reviewed quently used as the sigmoid transfer function. Moreover, during
papers, including author names, type of model, case study loca- the development phase of ANN models, scholars generally use
tion, data span, time scale, input and output parameters, perfor- trial-and-error methods to select the number of hidden nodes
mance criteria, and best model. It can be noted that the in the hidden layer(s). The assessment process of ANN models
researchers were often concerned with selecting the proper for assessing the qualification of the suggested models was car-
input combination for GWL prediction. It can be observed that ried out by using statistical measures. Moreover, very limited
most of the previous studies have used GWL, P, and T as input published papers have used parameters that had a significant
294
effect on the fluctuation of GWL, such as groundwater abstrac- management. In the future, collaborations among researchers
tion quantity. Researchers usually used error metrics such as and local stakeholders need to be strengthened, in order to
RMSE, MAE, R2 , R and NASH. develop a more practical framework, such that deep learning-
ELM was presented, in the last decade, as an alternative algo- extracted GWL information can be fully utilized by local gov-
rithm to train a single feedforward neural network or MLP. ernments; for instance, the development of tools with user-
Few studies have been conducted to forecast GWL using ELM. friendly graphical user interfaces may encourage the usage of
The main advantages of the ELM are its ability to train faster, deep learning in GWL simulation and forecasting by local stake-
its capability to obtain more accurate results even when a larger holders and users with little programming background.
number of hidden nodes is used, and its good generalizability. The Internet of Things (IoT) paradigm allows groundwater
The input data used is very important in developing GWL pre- information to be gathered and analyzed more smartly with
dictive models over a certain period. Some researchers have broader user communities. Advances in computational power,
used only the lags of GWL to establish their models for predict- internet speed, and coverage, together with rapid development
ing GWL. However, other researchers combined past GWL val- in software technologies, such as cloud storage systems and
ues and hydrological variables, such as Raf, EP, EVP, T, relative service-oriented architecture (SOA), have stimulated the devel-
H%, flow discharge of a river, and so on. Based on the present opment of smart water quantity and quality monitoring sys-
review, only a few researchers have used only meteorological tems [300,301]. Approximately 1.6 million monitoring wells
or hydrological parameters for GWL forecasting. are available worldwide to measure GWL; however, most of
The cited literature reported that GWL predicted using statisti- them are not recorded automatically [302,303]. Therefore, the
cal ARIMA models are more accurate, in comparison to other implementation of IoT is expected to reduce the difficulty in
applied AI models. Despite requiring a great deal of experience, obtaining GWL data. The integration of IoT and deep learning
due to their complexity, time-series models have been inher- can provide more accurate real-time GWL data collection, trans-
ently identified as very powerful techniques with great flexibil- fer, and analysis at low-cost. For instance, Su et al. [303] devel-
ity, as they are more likely to deal with non-stationary time- oped deep learning algorithms that combine both IoT and
series more effectively. Besides, these models can be used for groundwater model using a groundwater-related web GIS plat-
short-term GWL prediction without the need for other input form, in order to enhance groundwater data management.
data. This might be an advantageous feature in an area where The literature review presented and discussed above highlights
the availability of hydrological data is lacking. The monthly the great capabilities of the ANFIS model as a robust tool for
time step of collected GWL was used in most of the cited GWL prediction. ANFIS can provide high accuracy and precision
papers; this might be attributed to the high availability of for nearly all cases (i.e., monthly, daily, or weekly time steps).
monthly GWL records, in comparison to other time scales. However, an important conclusion that can be drawn, based
According to the reviewed papers, the input parameters that the above analysis, is that the selection of the variables used
have been used in GWL modeling are mostly antecedent values as inputs for the ANFIS model is an important factor controlling
(auto-correlated input variables) without any other exogenous the accuracy of the model. It has been demonstrated that, in
hydrological variables. However, in some other models (e.g., some cases, the inclusion of a large number of climate variables
ANNs), data such as T, river discharge (surface runoff), EVP, sur- is required for an accurate GWL estimation. As shown above,
face water (lake) level, pumping rates (extraction from wells), most research regarding the application of the ANFIS model
and H% have been used as input variables and the results were has been conducted without metaheuristic optimization algo-
compared with those of time-series models. However, the out- rithms, such as PSO, firefly optimization algorithm (FFA), or
comes of time-series models were superior, compared with Grey wolf optimizer (GWO), among others. These algorithms
multi-input variables models. remain relatively underexplored and they can certainly con-
A review of the previous studies utilizing NARX networks for tribute to the advancement of methods for GWL prediction,
GWL forecasting was also conducted, in order to enrich the helping to overcome some specific difficulties encountered dur-
review process. The main points highlighted from the related ing the training of the ANFIS model, especially rapid conver-
studies can be summarized in three points: (i) NARX networks gence and local minima.
suffer from several limitations, such as over-fitting problems The employment of the wavelet technique to develop comple-
and local minima; (iii) The Bayesian algorithm can be consid- mentary models can enhance the predictability efficiency and,
ered as a reliable training method to avoid over-fitting problem, in all cases, hybrid methods involving wavelets performed bet-
but its performance may be reduced when coupling it with ter than their regular counterparts. In general, wavelet-based
early stopping; and (iii) the literature has reported that linear machine learning models are more accurate, as the DWT pro-
models (ARX) could be a useful and easily applicable tool for vides better discrimination of the non-linear and non-
short-term GWL forecasting. stationary trends that exist at various scales in the time-series
With the advances in computational power and internet tech- of the input variables [249]. It has also been found to be useful
nology, deep learning applications in hydrology are expected for other aquifers and basins with different characteristics
to grow significantly in the long-term. Indeed, the application [194]. The generated wavelet models are simpler and more
of deep learning in groundwater prediction is still relatively interpretable, making it possible to understand which indepen-
low, as compared to other hydrological variables such as water dent variables have a greater effect in simulating the dependent
level and discharge [134]. It is no surprise that more deep learn- variable [39].
ing groundwater-related articles will become available in the An optimal number of lag times for the input values must be
near future. As stated in the United Nation’s Sustainable Devel- identified, as it plays an important role in GWL simulation
opment Goal 6 of Clean Water and Sanitation, more than 1.7 bil- [39]. Furthermore, selecting a suitable decomposition level
lion people live in river basins where the water usage exceeds affects the accuracy of hybrid models [39]. This is because a
water supply. In many parts of the world, groundwater is still high level of decomposition is not always helpful in increasing
unexplored and can provide an alternative solution to reduce the model’s accuracy. Further research needs to be conducted,
water stress issues, if used wisely. Reliable groundwater infor- in order to improve higher lead-time forecasting. The selection
mation and forecasting are essential for groundwater resource of significant hydrological variables as inputs for the models
must also be considered, which is very specific to the region
295
Fig. 14. Applied AI models for GWL modelling: Pie-chart demonstration of the various AI models with extended view of the three most widely preferred AI model types
including Hybrid-type models, Wavelet assimilated models and others.
of study and the climate in question, as it significantly affects ous and current literature. The effects of the various inherent
the performance of the models [39,194]. Therefore, it is recom- parameters in the various employed models also need further
mended to carry out exhaustive studies in the future to evaluate investigation.
the effects of other hydrological variables on groundwater sim- In our review of the numerous studies carried out in the field of
ulation and to understand the dynamics involved to find suit- GWL, it was observed that there has been a significant increase
able variables for groundwater modeling in different climates in the number of published works over the years. It is evident
and conditions. The employed model can be used for monitor- that increases in groundwater contamination, the dependence
ing seasonal GWL fluctuations by dividing the data, according of groundwater utilization, changes in water level, and advance-
to monsoon and non-monsoon periods. More comparative stud- ment in machine learning and AI tools have led to more promi-
ies to investigate the performance of the models at different nent research using AI-based models in this field, as shown in
stages of modeling, as well as comparisons with other time- Fig. 2. As per Fig.2 a slow growth was observed during 2008–
series pre-processing tools, need to be explored. In future 2011; however, the number of publications doubled during
works, the hybrid models should able to explain the physical 2011–2013, while steady progress was detected in the subse-
phenomena involved, which are currently lacking in the previ- quent years. Furthermore, an exponential growth can be
observed from 2017–2020 due to an increase in the popularity
of AI models. Even though 2020 was a difficult year for every-
one, 28 papers were still published in the field, thus demon-
strating that such computational tools are the need of the
hour, allowing for the progress of such important research.
Fig. 15. Time scale analysis: Pie-chart demonstration of the usage of the dataset
time scale for GWL modelling and bar chart presenting the application of the
specified time scale in total number of studies. Fig. 16. Depiction of the preferred input selection for GWL modeling.
296
Fig. 17. Statistical analysis of applied performance metrics (PMs): A. Combination of PMs utilized in total studies. B. Individual PMs examined. C. presenting the other PMs
examined except the prevalent PMs.
Over the period of the survey, AI models have advanced, leading ies), and 15-year data sets (8 studies) have also been majorly
to more complex and integrated type of models, which have applied. It should be considered that larger data sets can reveal
given researchers the means to develop more accurate and the time-series pattern trend changes, whereas smaller data
robust models. A similar conclusion was drawn after visualiza- sets cannot. A total of 10 research papers used larger data sets
tion of the various models applied to GWL modeling in the past (between 20–60 years), which are important in understanding
decade. Fig. 14 presents an in-depth description of the most long-term trend changes; however, such data may be ineffec-
applied model types in GWL simulation, in pie-chart form. It tive for near-future prediction, as recent data are more useful
can be seen, from the chart, that the most implemented were for prediction analyses.
hybrid-type (36%) models, mostly consisting of the base models The selection of input was mostly based on the influence of
such as ANN, SVM, fuzzy-based models, and tree-based models those variables on the selected output. Fig. 16 presents such
(presented in the blue box). Their popularity can be explained preferred input variables for GWL modeling. As per this figure
based on their higher accuracy, efficiency, and ability to deal (Fig. 16), GWL (34%), precipitation, and temperature were the
with non-linear data sets. The second extended pie-chart from most preferable input variables. Most studies utilized a combi-
the main chart denotes the wavelet-type models, which consist nation of these input variables, as mentioned in the figure;
of the integration of wavelet tools with the base models (sec- however, most researchers chose GWL in those combinations.
ondary pie-chart). The ability of wavelets to reduce data noise In addition, 1% also included sea level, in order to consider pos-
has played an important role in their widespread application sible seawater intrusion into a permeable aquifer. Such consid-
(8%). The third most applied model was distinguished as others, erations depend on the geographical location, which may affect
as it covered various other types of models which, apparently, (either directly or indirectly) the GWL.
are difficult to place in other groups (presented in the orange Performance metrics (PMs) help to understand the performance
box). This also indicates that there is a lot more to be explored of models, where each PM can reveal different information of
in this area. the model, leading to a wide range of PMs being applied, as
AI models has been proven to be effective tools when provided shown in Fig. 17. Fig. 17a shows the combination of PMs applied
with systematic and consistent data over a continuous time in the studies, which indicates that most preferred combination
period. Moreover, the time scale also plays a vital role in the was 3 and least preferred was 6. However, it is always beneficial
model performance; thus, a visual representation of the time to utilize various PMs to overcome the limitations of individual
scales used in the published papers is presented in Fig. 15. PMs. Fig. 17b shows the most applied PM; among all, the RMSE
The pie chart depicts that the most applied data collection (77) was the most-used error type PM, while R2 (46) was the
was monthly data (69%), followed by daily data (24%). Studies most-used accuracy assessment PM. Fig. 17c is an extension
that applied yearly data (1%) can be utilized for GWL capacity of Fig. 17b, showing the other PMs applied.
change studies but may be rendered insignificant when study- The literature review also emphasised on the integration cli-
ing seasonal changes. Similarly, depending on the type of study, mate change on watershed GWL and more research shall be
hourly data (4%) can also be used. However, hourly data collec- adopted for this kind of modeling interaction between climate
tion can be expensive and may not be a preferable choice for and geo-science water behaviour [304]. Climate change has
low-budget research projects. When considering the duration been remarkably observed especially in tropical region and
of the data utilized for the study, 1 year (8 studies) seemed to has substantial influence on ground temperature in which lead
be the smallest duration which allowed the model to work effi- for the GWL fluctuation [305].
ciently. Following that, 7- (8 studies), 9- (9 studies), 11- (8 stud-
297
4. Future research direction decision-makers to develop strategic plans and policies for
water management and to ensure water sustainability, espe-
Based on the presented literature review and the gaps identified cially in dry areas.
in the previous section, recommendations of possible future Our review showed that, among the genetic programming tech-
research directions to improve the accuracy of GWL prediction niques, Classic GP, MGGP, and GEP have been successfully used
models and to enhance the related knowledge are outlined in this for GWL prediction to date. Future studies could be designed to
section. assess the efficiency of other GP variants, such as Linear GP
[308] or multi-stag GP [309], for GWL prediction.
Exogenous parameters, such as sea level and groundwater Further studies should take into consideration the filling in (i.e.,
abstraction, have vital influences on the GWL in different imputation) of missing groundwater values using deep learning
regions around the world. Increased attention should be paid techniques [310–313]. Groundwater observations are usually
to the rapid changes in sea level rise when simulating GWL in sparse and the existence of missing values (e.g., due to instru-
coastal aquifers. Global warming has undeniably major effects ment failure or poor monitoring management systems) are
on sea level by melting large quantities of ice, thus leading to common. These missing values may degrade the data quality
a rise in sea and ocean water levels in some areas over the and increase the uncertainty in spatio-temporal groundwater
world. Consequently, the amount of seawater that will seep into analysis and simulation [314]. Hence, reliable groundwater data
the aquifer increases, which has a great impact on the sharp filling algorithms are needed, in order to reproduce the actual
fluctuations in changing GWL, especially in coastal areas. On conditions for groundwater forecasting. As the groundwater
the other hand, the rise in global temperatures leads to evapo- pattern is usually non-linear or non-stationary, imputing the
ration of large quantities of water, which may lead to a decrease missing values is a complex issue. For this task, deep learning
in the GWL in wells. The other important parameter is ground- could provide a powerful tool to extract the non-linear spatio-
water abstraction quantity, which should also be taken into temporal groundwater patterns without considering their
account in future studies. In arid and semi-arid areas, which explicit forms [315]. The LSTM is specifically designed for
depend largely on groundwater for irrigation, agriculture, and long-term period data prediction, being equipped with memory
other purposes, the groundwater that is extracted may not be cells that retain important information regarding historical
replaced until after long periods of groundwater recharge, due events [316]. Therefore, the improvement of LSTM models in
to natural factors such as less rainfall. Therefore, abstraction groundwater missing value imputation is foreseen to be a pop-
has a great impact on GWL fluctuations. ular topic. Furthermore, groundwater sensors can employ deep
The selection of GWL lags should be given more attention in the learning methods to detect and correct some unreasonable or
development of the AI modeling approaches. Among different outlier readings; in this way, only useful information will be
mathematical and statistical approaches, ACF and PACF have transmitted to the central system.
been considered efficient methods to select the best GWL lags Hybrid versions of ML models that incorporate nature inspired
as inputs. Based on the reviewed papers collected in this study, algorithms for tuning the hyperparameters standalone ML mod-
few studies used ACF and/or PACF to select the most proper pre- els are highly empathised to be explored in this research
ceding GWL variables. The ACF and PACF approaches provide domain. As the optimization of the internal models parameters
much more information on the main characteristics of the is influencing the learning process and the prediction capacity
time-series of groundwater fluctuations over a certain period. [317,318].
The fluctuations of the GWL provide a direct measure of the
effects of groundwater development and valuable knowledge 5. Conclusions
about the dynamics of an aquifer in GWL time-series data.
Therefore, there is a high possibility to accurately forecast The current survey was established to provide an informative
future GWL from its previous data. milestone on the implementation of machine learning models in
To obtain more precise GWL predictions, many studies have the simulation of GWL. The survey covered the period of 2008–
used antecedent data of GWL, as well as hydrological and mete- 2020, where all the gathered studies were obtained from indexed
orological information, such as Raf, H%, T, EVP, and so on. How- journals in the Web of Science. Based on the reported review, ten
ever, increasing the number of input variables could hinder the versions of ML models have been applied for GWL modeling over
process of developing reliable and accurate models. In future the globe. The survey identified several essential elements in the
research, the researchers may have to apply feature selection existing GWL simulation models, including the applied algorithms,
techniques to select the most significant input parameters and input parameters, target parameters, investigated regions, data
to get rid of redundant information [306,307]. Moreover, the span, and performance metrics. The surveyed studies were
selection of the most appropriate variables can enhance the assessed and evaluated scientifically, and numerous findings were
model learning process, save time, and reduce computational discussed in detail. In accordance with the current status of the
costs [54]. Thus, applying a feature selection approach prior to conducted literature, various possible future research directions
the learning phase of AI models may help to achieve more pre- were recognized for the interested readers and practitioners in this
cise and reliable GWL prediction models. domain.
A majority of researchers have developed predictive models at
monthly and daily time scales. More attention should be paid Abbreviations
to the prediction of yearly GWL, as it is considered very signif-
icant for water resource management and planning in the long- The list of abbreviations used in this paper are tabulated in
term. Moreover, long-term GWL forecasting could help Table 12.
298
Table 12
The list of abbreviations used in this paper.
ACF Autocorrelation Function AE Absolute Error

AI Artificial Intelligence ANFIS Adaptive Neuro-Fuzzy Inference System
ANN Artificial Neural Networks AR Autoregressive
ARIMA Autoregressive Integrated Moving Average ARMA Autoregressive-Moving-Average
ARMAX Auto Regressive Moving Average With External Model ARX Auto-Regressive With External Model
BA Bat Algorithm BNN Bayesian Neural Networks
BP Backpropagation BR Bayesian Regularization
BRT Boosted Regression Tree CANFIS Co-Adaptive Neuro-Fuzzy Inference Systems
CART Classification And Regression Tree CBA Classification Based On Association Rule
CC Correlation Coefficient CEEMD Complementary Ensemble Empirical Mode Decomposition
CFL-WA Committee Fuzzy Logic Weighted Averaging CFN Cascade Forward Network
CSO Cat Swarm Optimization CWT Continuous Wavelet Transform
DE Differential Evolution DL Deep Learning
DWT Discrete Wavelet Transform EANN Emotional ANN
EBF Evidential Belief Function EEMD Ensemble Empirical Mode Decomposition
ELM Extreme Learning Machine EP Evaporation
EVP Evapotranspiration EV Error Variation
FFA Firefly Optimization Algorithm FFBPNN Feed-Forward Back-Propagation Neural Network
FFNN Feed Forward Neural Network FL Fuzzy Logic
GA Genetic Algorithm GB Gradient Boosting
GBR Gradient Boosted Regression GEP Gene Expression Programming
GIS Geographic Information System GMDH Group Method Of Data Handling
GOA Grasshopper Optimization Algorithm GP Genetic Programming
GP Gaussian Process GPC Gaussian Process Classification
GPR Aussian Process Regression GR Gridded Rainfall
GRACE Gravity Recovery And Climate Experiment GRI Groundwater Resources Index
GRNN Generalized Regression Neural Network GSA Gravitational Search Algorithm
GSM Grey Self-Memory Model GT Gamma And M-Tests
GWA-GWHP Groundwater Abstraction Associated With Operation Of The GWHP System GWF Groundwater Fluctuation
GWA-WCC Groundwater Abstraction Associated With Operation Of The WCC System GWL Groundwater Level
GWO Grey Wolf Optimizer H% Humidity
HBSA Hybrid Bat-Swarm Algorithm HIS Improved Harmony Search
HS Harmony Search HWES Holt-Winters Exponential Smoothing
IA Index Of Agreement IL Infiltration Loss
IoT Internet Of Things ISA Interior Search Algorithm
KA Krill Algorithm kNN K Nearest Neighbor
LAT Latitude LFL Larsen Fuzzy Logic
LM Levenberg–Marquardt LNG Longitude
LSBoost Last Squares Boosting LSSVM East-Squares Support Vector Machine
LSTM Long Short-Term Memory LSTM-LA Long Short-Term Memory-Lion Algorithm
M5tree M5 Model Tree MA Moving Average
MAE Mean Absolute Error MAPE Mean Absolute Percentage Error
MARS Multivariate Adaptive Regression Spline MCMC Markov Chain Monte Carlo
ME Mean Error MF Membership Functions
MFL Mamdani Fuzzy Logic MGGP Multigene GP
ML Machine Learning MLP Multilayer Perceptron
MLR Multiple Linear Regression MODWT Maximal Overlap Discrete Wavelet Transform
MOGA Multi-Objective Genetic Algorithm MRVA Mississippi River Valley Alluvial
NB Naïve Bayes NCCI National Cartographic Center Of Iran
NF Neuro-Fuzzy NMSE Normalized Mean Square Error
NRMSE Normalized Root Mean Square Error NS Nash–Sutcliffe Efficiency Coefficient
NSE Nash–Sutcliffe Efficiency OK Ordinary Kriging
P Precipitation PACF Partial Autocorrelation Function
PSO Particle Swarm Optimization Q Discharge
QPSO Quantum Behaved Particle Swarm Optimization Function R Pearson’S Correlation Coefficient
Raf Rainfall RBFNN Radial Basis Function Neural Network
RE Reduction Of Error Statistics RE Relative Error
RF Random Forest RFR Random Forest Regression
RMSE Root Mean Square Error RNN Recurrent Neural Network
ROC Receiver Operating Characteristics RS River Stage
SARIMA Seasonal Autoregressive Integrated Moving Average SCFL Simple Committee Fuzzy Logic
SD Sunshine Duration SFL Ugeno Fuzzy Logic
SFR Stream Flow Rate SGS Sequential Gaussian Simulation
SM Soil Moisture SMO Sequential Minimal Optimization
SOA Service-Oriented Architecture SOM Self-Organizing Map
SPI Standardized Precipitation Index SVD Singular Value Decomposition
SVM Support Vector Machines SVR Support Vector Regression
SWL Surface Water Level T Temperature
T max Maximum Temperature T mean Mean Air Temperature
T min Minimum Temperature WA Whale Algorithm
WA Weed Algorithm WANFIS Wavelet-Adaptive Neuro-Fuzzy Inference System
WEF Weighted Error Function WGEP Wavelet Gene Expression Programming
WL Water Level WLR Wavelet-Linear Regression
WMT Wavelet-M5 Model Tree WNN Wavelet- Neural Network
WS Wind Speed WSVR Wavelet-SVR
WT Wavelet Transform
299
Declaration of Competing Interest mechanism and grey model algorithms for streamflow forecasting over
multiple time horizons, Water Resour. Manage 32 (5) (2018) 1883–1899.
[22] Omeje, Otitodilichukwu Excel, Hamza Sabo Maccido, Yusuf Abdullahi
The authors declare that they have no known competing finan- Badamasi, and Sani Isah Abba. "Performance of Hybrid Neuro-Fuzzy Model
cial interests or personal relationships that could have appeared for Solar Radiation Simulation at Abuja, Nigeria: A Correlation Based Input
Selection Technique." Knowledge-Based Engineering and Sciences 2, no. 3
to influence the work reported in this paper.
(2021): 54-66.
[23] M. Jamei, I. Ahmadianfar, X. Chu, Z.M. Yaseen, Prediction of surface water
Acknowledgement total dissolved solids using hybridized wavelet-multigene genetic
programming: New approach, J. Hydrol. 589 (2020) 125335.
[24] Y. Zhu, Y. Liu, W. Wang, V.P. Singh, X. Ma, Z. Yu, Three dimensional
The authors would like to thank Al-Mustaqbal University Col- characterization of meteorological and hydrological droughts and their
lege for providing technical support for this research. probabilistic links, J. Hydrol. 578 (2019) 124016.
[25] M. Zounemat-Kermani, M. Scholz, Computing air demand using the takagi–
sugeno model for dam outlets, Water 5 (3) (2013) 1441–1456.
References [26] Z.M. Yaseen, S.O. Sulaiman, R.C. Deo, K.-W. Chau, An enhanced extreme
learning machine model for river flow forecasting: State-of-the-art, practical
applications in water resource engineering area and future research direction,
[1] M. Qadir, B.R. Sharma, A. Bruggeman, R. Choukr-Allah, F. Karajeh, Non-
J. Hydrol. 569 (2019) 387–408.
conventional water resources and opportunities for water augmentation to
[27] R.M. Adnan, Z. Liang, S. Heddam, M. Zounemat-Kermani, O. Kisi, B. Li, Least
achieve food security in water scarce countries, Agricult. Water Manage. 87
square support vector machine and multivariate adaptive regression splines
(1) (2007) 2–22.
for streamflow prediction in mountainous basin using hydro-meteorological
[2] Y. Wada, L.P. Van Beek, C.M. Van Kempen, J.W. Reckman, S. Vasak, M.F.
data as inputs, J. Hydrol. 586 (2020) 124371.
Bierkens, Global depletion of groundwater resources, Geophys. Res. Lett. 37
[28] S.K. Bhagat, T.M. Tung, Z.M. Yaseen, Heavy metal contamination prediction
(20).
using ensemble model: Case study of bay sedimentation, australia, J. Hazard.
[3] H. Afzaal, A.A. Farooque, F. Abbas, B. Acharya, T. Esau, Groundwater
Mater. 403 (2020) 123492.
estimation from major physical hydrology components using artificial
[29] A. Mahdavi-Meymand, M. Zounemat-Kermani, A new integrated model of the
neural networks and deep learning, Water 12 (1) (2020) 5.
group method of data handling and the firefly algorithm (gmdh-fa):
[4] S. Sadeghi-Tabas, S. Samadi, B. Zahabiyoun, et al., Application of bayesian
application to aeration modelling on spillways, Artif. Intell. Rev. 53 (4)
algorithm in continuous streamflow modeling of a mountain watershed,
(2020) 2549–2569.
European Water 57 (2017) 101–108.
[30] I.N. Daliakopoulos, P. Coulibaly, I.K. Tsanis, Groundwater level forecasting
[5] P.K. Gupta, B. Yadav, B.K. Yadav, Assessment of lnapl in subsurface under
using artificial neural networks, J. Hydrol. 309 (1–4) (2005) 229–240.
fluctuating groundwater table using 2d sand tank experiments, J. Environ.
[31] J. Adamowski, H.F. Chan, A wavelet neural network conjunction model for
Eng. 145 (9) (2019) 04019048.
groundwater level forecasting, J. Hydrol. 407 (1–4) (2011) 28–40.
[6] A. Izady, K. Davary, A. Alizadeh, A.N. Ziaei, A. Alipoor, A. Joodavi, M.L.
[32] S. Ranjithan, J. Eheart, J. Garrett Jr, Neural network-based screening for
Brusseau, A framework toward developing a groundwater conceptual model,
groundwater reclamation under uncertainty, Water Resour. Res. 29 (3)
Arab. J. Geosci. 7 (9) (2014) 3611–3631.
(1993) 563–574.
[7] J. Xue, Z. Huo, F. Wang, S. Kang, G. Huang, Untangling the effects of shallow
[33] L.L. Rogers, F.U. Dowla, Optimization of groundwater remediation using
groundwater and deficit irrigation on irrigation water productivity in arid
artificial neural networks with parallel solute transport modeling, Water
region: New conceptual model, Sci. Total Environ. 619 (2018) 1170–1182.
Resour. Res. 30 (2) (1994) 457–481.
[8] P.J. Omar, S. Gaur, S. Dwivedi, P. Dikshit, Groundwater modelling using an
[34] S. Lallahem, J. Mania, A. Hani, Y. Najjar, On the use of neural networks to
analytic element method and finite difference method: An insight into lower
evaluate groundwater levels in fractured media, J. Hydrol. 307 (1–4) (2005)
ganga river basin, J. Earth Syst. Sci. 128 (7) (2019) 195.
92–111.
[9] P. Jamin, M. Cochand, S. Dagenais, J.-M. Lemieux, R. Fortier, J. Molson, S.
[35] A.A. Nadiri, K. Naderi, R. Khatibi, M. Gharekhani, Modelling groundwater level
Brouyère, Direct measurement of groundwater flux in aquifers within the
variations by learning from multiple models using fuzzy logic, Hydrol. Sci. J.
discontinuous permafrost zone: an application of the finite volume point
64 (2) (2019) 210–226.
dilution method near umiujaq (nunavik, canada), Hydrogeol. J. (2020) 1–17.
[36] T. Zhou, F. Wang, Z. Yang, Comparative analysis of ann and svm models
[10] C. Ukpaka, S.N.-A. Adaobi, C. Ukpaka, Development and evaluation of trans-
combined with wavelet preprocess for groundwater depth prediction, Water
amadi groundwater parameters: The integration of finite element techniques,
9 (10) (2017) 781.
Chem. Int. 3 (2017) 306.
[37] D.D. Moghaddam, O. Rahmati, A. Haghizadeh, Z. Kalantari, A modeling
[11] T. Pathania, A. Bottacin-Busolin, A. Rastogi, T. Eldho, Simulation of
comparison of groundwater potential mapping in a mountain bedrock
groundwater flow in an unconfined sloping aquifer using the element-free
aquifer: Quest, garp, and rf models, Water 12 (3) (2020) 679.
galerkin method, Water Resour. Manage. 33 (8) (2019) 2827–2845.
[38] K. Kasiviswanathan, S. Saravanan, M. Balamurugan, K. Saravanan, Genetic
[12] B. Barnett, L. Townley, V. Post, R. Evans, R. Hunt, L. Peeters, S. Richardson, A.
programming based monthly groundwater level forecast models with
Werner, A. Knapton, A. Boronkay, Australian groundwater modelling
uncertainty quantification, Model. Earth Syst. Environ. 2 (1) (2016) 27.
guidelines, National Water Commission, Canberra.
[39] R. Bahmani, A. Solgi, T.B. Ouarda, Groundwater level simulation using gene
[13] F. Jiang, Y. Jiang, H. Zhi, Y. Dong, H. Li, S. Ma, Y. Wang, Q. Dong, H. Shen, Y.
expression programming and m5 model tree combined with wavelet
Wang, Artificial intelligence in healthcare: past, present and future, Stroke
transform, Hydrol. Sci. J. (2020) 1–13.
Vascul. Neurol. 2 (4) (2017) 230–243.
[40] H. Afzaal, A.A. Farooque, F. Abbas, B. Acharya, T. Esau, Groundwater
[14] M. Zounemat-Kermani, O. Kisi, J. Piri, A. Mahdavi-Meymand, Closure to
estimation from major physical hydrology components using artificial
”assessment of artificial intelligence–based models and metaheuristic
neural networks and deep learning, Water 12 (1) (2020) 5.
algorithms in modeling evaporation” by mohammad zounemat-kermani,
[41] M. Alizamir, O. Kisi, M. Zounemat-Kermani, Modelling long-term
ozgur kisi, jamshid piri, and amin mahdavi-meymand, J. Hydrol. Eng. 25 (9)
groundwater fluctuations by extreme learning machine using hydro-
(2020) 07020015.
climatic data, Hydrol. Sci. J. 63 (1) (2018) 63–73.
[15] S.K. Bhagat, T.M. Tung, Z.M. Yaseen, Development of artificial intelligence for
[42] B.D. Bowes, J.M. Sadler, M.M. Morsy, M. Behl, J.L. Goodall, Forecasting
modeling wastewater heavy metal removal: State of the art, application
groundwater table in a flood prone coastal city with long short-term memory
assessment and possible future research, J. Cleaner Prod. 250 (2020) 119473.
and recurrent neural networks, Water 11 (5) (2019) 1098.
[16] T. Hai, A. Sharafati, A. Mohammed, S.Q. Salih, R.C. Deo, N. Al-Ansari, Z.M.
[43] B.T. Pham, A. Jaafari, I. Prakash, S.K. Singh, N.K. Quoc, D.T. Bui, Hybrid
Yaseen, Global solar radiation estimation and climatic variability analysis
computational intelligence models for groundwater potential mapping,
using extreme learning machine based predictive model, IEEE Access 8 (2020)
Catena 182 (2019) 104101.
12026–12042.
[44] P.T. Nguyen, D.H. Ha, A. Jaafari, H.D. Nguyen, T. Van Phong, N. Al-Ansari, I.
[17] S.Q. Salih, A.A. Alsewari, Z.M. Yaseen, Pressure vessel design simulation:
Prakash, H.V. Le, B.T. Pham, Groundwater potential mapping combining
Implementing of multi-swarm particle swarm optimization, in: Proceedings
artificial neural network and real adaboost ensemble technique: the daknong
of the 2019 8th International Conference on Software and Computer
province case-study, Vietnam, Int. J. Environ. Res. Public Health 17 (7) (2020)
Applications, 2019, pp. 120–124.
2473.
[18] U. Beyaztas, S.Q. Salih, K.-W. Chau, N. Al-Ansari, Z.M. Yaseen, Construction of
[45] G.-B. Kim, A study on the establishment of groundwater protection area
functional data analysis modeling strategy for global solar radiation
around a saline waterway by combining artificial neural network and gis-
prediction: application of cross-station paradigm, Eng. Appl. Comput. Fluid
based ahp, Environ. Earth Sci. 79 (5) (2020) 1–17.
Mech. 13 (1) (2019) 1165–1181.
[46] T. Rajaee, H. Ebrahimi, V. Nourani, A review of the artificial intelligence
[19] T.M. Tung, Z.M. Yaseen, et al., A survey on river water quality modelling using
methods in groundwater level modeling, J. Hydrol. 572 (2019) 336–351.
artificial intelligence models: 2000–2020, J. Hydrol. 585 (2020) 124670.
[47] V. Nourani, S. Mousavi, Spatiotemporal groundwater level modeling using
[20] M. Fadaee, A. Mahdavi-Meymand, M. Zounemat-Kermani, Suspended
hybrid artificial intelligence-meshless method, J. Hydrol. 536 (2016) 10–25.
sediment prediction using integrative soft computing models: On the
[48] C. Chen, W. He, H. Zhou, Y. Xue, M. Zhu, A comparative study among machine
analogy between the butterfly optimization and genetic algorithms,
learning and numerical models for simulating groundwater dynamics in the
Geocarto International (just-accepted) (2020) 1–14.
heihe river basin, northwestern china, Sci. Rep. 10 (1) (2020) 1–13.
[21] Z.M. Yaseen, M. Fu, C. Wang, W.H.M.W. Mohtar, R.C. Deo, A. El-Shafie,
Application of the hybrid artificial neural network coupled with rolling
300
[49] C. Chen, W. He, H. Zhou, Y. Xue, M. Zhu, A comparative study among machine [76] B. Yadav, S. Ch, S. Mathur, J. Adamowski, Assessing the suitability of extreme
learning and numerical models for simulating groundwater dynamics in the learning machines (elm) for groundwater level prediction, J. Water Land
heihe river basin, northwestern china, Sci. Rep. 10 (1) (2020) 1–13. Develop. 32 (1) (2017) 103–112.
[50] N. Bisoyi, H. Gupta, N.P. Padhy, G.J. Chakrapani, Prediction of daily sediment [77] A. Sharafati, A. Tafarojnoruz, D. Motta, Z.M. Yaseen, Application of nature-
discharge using a back propagation neural network training algorithm: A case inspired optimization algorithms to anfis model to predict wave-induced
study of the narmada river, india, Int. J. Sediment Res. 34 (2) (2019) 125–135. scour depth around pipelines, J. Hydroinform. 22 (6) (2020) 1425–1451.
[51] M.K. AlOmar, M.M. Hameed, M.A. AlSaadi, Multi hours ahead prediction of [78] J.-S. Jang, Anfis: adaptive-network-based fuzzy inference system, IEEE Trans.
surface ozone gas concentration: Robust artificial intelligence approach, Syst., Man, Cybern. 23 (3) (1993) 665–685.
Atmosp. Pollut. Res. 11 (9) (2020) 1572–1587. [79] L. Penghui, A.A. Ewees, B.H. Beyaztas, C. Qi, S.Q. Salih, N. Al-Ansari, S.K.
[52] M. Hameed, S.S. Sharqi, Z.M. Yaseen, H.A. Afan, A. Hussain, A. Elshafie, Bhagat, Z.M. Yaseen, V.P. Singh, Metaheuristic optimization algorithms
Application of artificial intelligence (ai) techniques in water quality index hybridized with artificial intelligence model for soil temperature
prediction: a case study in tropical region, malaysia, Neural Comput. Appl. 28 prediction: Novel model, IEEE Access 8 (2020) 51884–51904.
(1) (2017) 893–905. [80] T.M. Tung, Z.M. Yaseen, et al., A survey on river water quality modelling using
[53] Z.M. Yaseen, A. El-Shafie, H.A. Afan, M. Hameed, W.H.M.W. Mohtar, A. artificial intelligence models: 2000–2020, J. Hydrol. 585 (2020) 124670.
Hussain, Rbfnn versus ffnn for daily river flow forecasting at johor river, [81] A. Azad, H. Karami, S. Farzin, S.-F. Mousavi, O. Kisi, Modeling river water
malaysia, Neural Comput. Appl. 27 (6) (2016) 1533–1542. quality parameters using modified adaptive neuro fuzzy inference system,
[54] S.K. Bhagat, T. Tiyasha, S.M. Awadh, T.M. Tung, A.H. Jawad, Z.M. Yaseen, Water Sci. Eng. 12 (1) (2019) 45–54.
Prediction of sediment heavy metal at the australian bays using newly [82] S. Maroufpoor, E. Maroufpoor, O. Bozorg-Haddad, J. Shiri, Z.M. Yaseen, Soil
developed hybrid artificial intelligence models, Environ. Pollut. 268 (2020) moisture simulation using hybrid artificial intelligent model: Hybridization
115663. of adaptive neuro fuzzy inference system with grey wolf optimizer algorithm,
[55] R.C. Deo, M. S ß ahin, Application of the artificial neural network model for J. Hydrol. 575 (2019) 544–556.
prediction of monthly standardized precipitation and evapotranspiration [83] V. Nourani, G. Elkiran, J. Abdullahi, Multi-step ahead modeling of reference
index using hydrometeorological parameters and climate indices in eastern evapotranspiration using a multi-model approach, J. Hydrol. 581 (2020)
australia, Atmosph. Res. 161 (2015) 65–81. 124434.
[56] H.B. Abarghouei, M.R. Kousari, M.A.A. Zarch, Prediction of drought in dry [84] F.B. Banadkooki, M. Ehteram, F. Panahi, S.S. Sammen, F.B. Othman, E.-S.
lands through feedforward artificial neural network abilities, Arab. J. Geosci. 6 Ahmed, Estimation of total dissolved solids (tds) using new hybrid machine
(5) (2013) 1417–1433. learning models, J. Hydrol. 124989 (2020).
[57] H.A. Afan, A. El-Shafie, Z.M. Yaseen, M.M. Hameed, W.H.M.W. Mohtar, A. [85] M. Moravej, P. Amani, S.-M. Hosseini-Moghari, Groundwater level simulation
Hussain, Ann based sediment prediction model utilizing different input and forecasting using interior search algorithm-least square support vector
scenarios, Water Resour. Manage. 29 (4) (2015) 1231–1245. regression (isa-lssvr), Groundwater Sustain. Devel. 11 (2020) 100447.
[58] G.R. Rakhshandehroo, M. Vaghefi, M.A. Aghbolaghi, Forecasting groundwater [86] G.-M. Bak, Y.-C. Bae, Groundwater level prediction using anfis algorithm, J.
level in shiraz plain using artificial neural networks, Arab. J. Sci. Eng. 37 (7) Korea Inst. Electron. Commun. Sci. 14 (6) (2019) 1235–1240.
(2012) 1871–1883. [87] A.-A. Jahanara, S.R. Khodashenas, Prediction of ground water table using nf-
[59] X. Meng, M. Yin, L. Ning, D. Liu, X. Xue, A threshold artificial neural network gmdh based evolutionary algorithms, KSCE J. Civil Eng. 23 (12) (2019) 5235–
model for improving runoff prediction in a karst watershed, Environ. Earth 5243.
Sci. 74 (6) (2015) 5039–5048. [88] N. Zhang, C. Xiao, B. Liu, X. Liang, Groundwater depth predictions by gsm, rbf,
[60] M.M. Hameed, M.K. AlOmar, Prediction of compressive strength of high- and anfis models: a comparative assessment, Arab. J. Geosci. 10 (8) (2017)
performance concrete: Hybrid artificial intelligence technique, in: 189.
International Conference on Applied Computing to Support Industry: [89] Y. Gong, Y. Zhang, S. Lan, H. Wang, A comparative study of artificial neural
Innovation and Technology Springer, 2019, pp. 323–335. networks, support vector machines and adaptive neuro fuzzy inference
[61] S. Nacar, M.A. Hınıs, M. Kankal, Forecasting daily streamflow discharges using system for forecasting groundwater levels near lake okeechobee, florida,
various neural network models and training algorithms, KSCE J. Civil Eng. 22 Water Resour. Manage. 30 (1) (2016) 375–391.
(9) (2018) 3676–3685. [90] N.S. Raghavendra, P.C. Deka, Multistep ahead groundwater level time-series
[62] S. Ali, M. Shahbaz, Streamflow forecasting by modeling the rainfall– forecasting using gaussian process regression and anfis, in: Advanced
streamflow relationship using artificial neural networks, Model. Earth Syst. Computing and Systems for Security Springer, 2016, pp. 289–302.
Environ. 6 (3) (2020) 1645–1656. [91] M. Khaki, I. Yusoff, N. Islami, Simulation of groundwater level through
[63] Z.M. Yaseen, M. Fu, C. Wang, W.H.M.W. Mohtar, R.C. Deo, A. El-Shafie, artificial intelligence system, Environ. Earth Sci. 73 (12) (2015) 8357–8367.
Application of the hybrid artificial neural network coupled with rolling [92] N. Djurovic, M. Domazet, R. Stricevic, V. Pocuca, V. Spalevic, R. Pivic, E.
mechanism and grey model algorithms for streamflow forecasting over Gregoric, U. Domazet, Comparison of groundwater level models based on
multiple time horizons, Water Resour. Manage. 32 (5) (2018) 1883–1899. artificial neural networks and anfis, Sci. World J. (2015).
[64] A. Jalalkamali, H. Sedghi, M. Manshouri, Monthly groundwater level [93] S. Samantaray, A. Sahoo, D.K. Ghose, Infiltration loss affects toward
prediction using ann and neuro-fuzzy models: a case study on kerman groundwater fluctuation through canfis in arid watershed: A case study, in:
plain, iran, J. Hydroinform. 13 (4) (2011) 867–876. Smart Intelligent Computing and Applications, Springer, 2020, pp. 781–789.
[65] S.S. Nair, G. Sindhu, Groundwater level forecasting using artificial neural [94] S. Maiti, R. Tiwari, A comparative study of artificial neural networks, bayesian
network, Int. J. Sci. Res. Publ. 6 (1) (2016) 2250–3153. neural networks and adaptive neuro-fuzzy inference system in groundwater
[66] A. Lohani, G. Krishan, Groundwater level simulation using artificial neural level prediction, Environ. Earth Sci. 71 (7) (2014) 3147–3160.
network in southeast punjab, india, J. Geol. Geosci. 4 (3) (2015) 206. [95] S. Emamgholizadeh, H. Kashi, I. Marofpoor, E. Zalaghi, Prediction of water
[67] M. Derbela, I. Nouiri, Intelligent approach to predict future groundwater level quality parameters of karoon river (iran) by artificial intelligence-based
based on artificial neural networks (ann), Euro-Mediterranean J. Environ. models, Int. J. Environ. Sci. Technol. 11 (3) (2014) 645–656.
Integr. 5 (3) (2020) 1–11. [96] V. Moosavi, M. Vafakhah, B. Shirmohammadi, N. Behnia, A wavelet-anfis
[68] M. Iqbal, U.A. Naeem, A. Ahmad, U. Ghani, T. Farid, et al., Relating hybrid model for groundwater level forecasting for different prediction
groundwater levels with meteorological parameters using ann technique, periods, Water Resour. Manage. 27 (5) (2013) 1301–1321.
Measurement 166 (2020) 108163. [97] V. Moosavi, M. Vafakhah, B. Shirmohammadi, M. Ranjbar, Optimization of
[69] S.M. Guzman, J.O. Paz, M.L.M. Tagert, A. Mercer, Artificial neural networks and wavelet-anfis and wavelet-ann hybrid models by taguchi method for
support vector machines: Contrast study for groundwater level prediction., groundwater level forecasting, Arab. J. Sci. Eng. 39 (3) (2014) 1785–1796.
in: 2015 ASABE Annual International Meeting, Am. Soc. Agricul. Biol. Eng. [98] M. Zare, M. Koch, Groundwater level fluctuations simulation and prediction
2015 (2015) 1. by anfis-and hybrid wavelet-anfis/fuzzy c-means (fcm) clustering models:
[70] Y.-M. Hong, Feasibility of using artificial neural networks to forecast Application to the miandarband plain, J. Hydro-environment Res. 18 (2018)
groundwater levels in real time, Landslides 14 (5) (2017) 1815–1826. 63–76.
[71] G.N. Kouziokas, A. Chatzigeorgiou, K. Perakis, Multilayer feed forward models [99] O. Kisi, J. Shiri, Wavelet and neuro-fuzzy conjunction model for predicting
in groundwater level forecasting using meteorological data in public water table depth fluctuations, Hydrol. Res. 43 (3) (2012) 286–300.
management, Water Resour. Manage 32 (15) (2018) 5041–5052. [100] J. Shiri, O. Kisi, H. Yoon, K.-K. Lee, A.H. Nazemi, Predicting groundwater level
[72] S. Lee, K.-K. Lee, H. Yoon, Using artificial neural network models for fluctuations with meteorological effect implications–a comparative study
groundwater level forecasting and assessment of the relative impacts of among soft computing techniques, Comput. Geosci. 56 (2013) 32–44.
influencing factors, Hydrogeol. J. 27 (2) (2019) 567–579. [101] M. Kholghi, S. Hosseini, Comparison of groundwater level estimation using
[73] Z. Li, Q. Yang, L. Wang, J.D. Martín, Application of rbfn network and gm (1, 1) neuro-fuzzy and ordinary kriging, Environ. Model. Assessment 14 (6) (2009)
for groundwater level simulation, Appl. Water Sci. 7 (6) (2017) 3345–3353. 729.
[74] M.K.N. Shamsuddin, F.M. Kusin, W.N.A. Sulaiman, M.F. Ramli, M.F.T. [102] P. Sreekanth, P. Sreedevi, S. Ahmed, N. Geethanjali, Comparison of ffnn and
Baharuddin, M.S. Adnan, Forecasting of groundwater level using artificial anfis models for estimating groundwater level, Environ. Earth Sci. 62 (6)
neural network by incorporating river recharge and river bank infiltration, in: (2011) 1301–1310.
MATEC Web of Conferences, Vol. 103, EDP Sciences, 2017, p. 04007. [103] B. Shirmohammadi, M. Vafakhah, V. Moosavi, A. Moghaddamnia, Application
[75] S. Evans, G.P. Williams, N.L. Jones, D.P. Ames, E.J. Nelson, Exploiting earth of several data-driven techniques for predicting groundwater level, Water
observation data to impute groundwater level measurements with an Resour. Manage 27 (2) (2013) 419–432.
extreme learning machine, Remote Sens. 12 (12) (2020) 2044. [104] E. Fallah-Mehdipour, O.B. Haddad, M. Mariño, Prediction and simulation of
monthly groundwater levels by genetic programming, J. Hydro-Environ. Res.
7 (4) (2013) 253–260.
301
[105] X. Wen, Q. Feng, H. Yu, J. Wu, J. Si, Z. Chang, H. Xi, Wavelet and adaptive [132] A. Wunsch, T. Liesch, S. Broda, Groundwater level forecasting with artificial
neuro-fuzzy inference system conjunction model for groundwater level neural networks: a comparison of long short-term memory (lstm),
predicting in a coastal aquifer, Neural Comput. Appl. 26 (5) (2015) 1203– convolutional neural networks (cnns), and non-linear autoregressive
1215. networks with exogenous input (narx), Hydrol. Earth Syst. Sci. 25 (3)
[106] S. Sridharam, A. Sahoo, S. Samantaray, D.K. Ghose, Estimation of water table (2021) 1671–1687.
depth using wavelet-anfis: A case study, in: Communication Software and [133] S. Sengupta, S. Basak, P. Saikia, S. Paul, V. Tsalavoutis, F. Atiah, V. Ravi, A.
Networks, Springer, pp. 747–754. Peters, A review of deep learning with special emphasis on architectures,
[107] E. Jeihouni, M. Mohammadi, S. Eslamian, M.J. Zareian, Potential impacts of applications and recent trends, Knowl.-Based Syst. 105596 (2020).
climate change on groundwater level through hybrid soft-computing [134] M. Sit, B.Z. Demiray, Z. Xiang, G.J. Ewing, Y. Sermet, I. Demir, A comprehensive
methods: a case study–shabestar plain, iran, Environ. Monit. Assess. 191 review of deep learning applications in hydrology and water resources,
(10) (2019) 620. Water Sci. Technol.
[108] V. Vapnik, The nature of statistical learning theory, Springer Science & [135] D. Kumar, T. Roshni, A. Singh, M.K. Jha, P. Samui, Predicting groundwater
Business Media, 2013. depth fluctuations using deep learning, extreme learning machine and
[109] B. Scholkopf, A.J. Smola, Learning with kernels: support vector machines, gaussian process: a comparative study, Earth Sci. Inf. (2020) 1–14.
regularization, optimization, and beyond, Adaptive Computation and [136] B. Supreetha, N. Shenoy, P. Nayak, Lion algorithm-optimized long short-term
Machine Learning series (2018). memory network for groundwater level forecasting in udupi district, india,
[110] S.K. Bhagat, T. Tiyasha, T.M. Tung, R.R. Mostafa, Z.M. Yaseen, Manganese (mn) Appl. Computat. Intell. Soft Comput. (2020).
removal prediction using extreme gradient model, Ecotoxicol. Environ. Saf. [137] C. Park, I.-M. Chung, Evaluating the groundwater prediction using lstm
204 (2020) 111059. model, J. Korea Water Resour. Assoc. 53 (4) (2020) 273–283.
[111] J.C. Principe, N.R. Euliano, W.C. Lefebvre, Innovating adaptive and neural [138] M.-J. Shin, S.-H. Moon, K.G. Kang, D.-C. Moon, H.-J. Koh, Analysis of
systems instruction with interactive electronic books, Proc. IEEE 88 (1) groundwater level variations caused by the changes in groundwater
(2000) 81–95. withdrawals using long short-term memory network, Hydrology 7 (3)
[112] V. Vapnik, The nature of statistical learning theory, Springer Science & (2020) 64.
Business Media, 2013. [139] A. Dax, M. Zilberbrand, Imputing missing groundwater observations, Hydrol.
[113] S. Kim, Y. Seo, V.P. Singh, Assessment of pan evaporation modeling using Res. 49 (3) (2018) 831–845.
bootstrap resampling and soft computing methods, J. Comput. Civil Eng. 29 [140] M. Vu, A. Jardani, N. Massei, M. Fournier, Reconstruction of missing
(5) (2015) 04014063. groundwater level data by using long short-term memory (lstm) deep
[114] S. Haykin, Neural Networks and Learning Machines, 3/E, Pearson Education neural network, J. Hydrol. 125776 (2020).
India, 2010. [141] A.Y. Sun, B.R. Scanlon, Z. Zhang, D. Walling, S.N. Bhanja, A. Mukherjee, Z.
[115] Y. Tikhamarine, D. Souag-Gamane, A.N. Ahmed, S.S. Sammen, O. Kisi, Y.F. Zhong, Combining physically based modeling and deep learning for fusing
Huang, A. El-Shafie, Rainfall-runoff modelling using improved machine grace satellite data: Can we learn from mismatch?, Water Resour. Res. 55 (2)
learning methods: Harris hawks optimizer vs. particle swarm optimization, (2019) 1179–1195.
J. Hydrol. (2020) 125133. [142] J. Zhang, Y. Zhu, X. Zhang, M. Ye, J. Yang, Developing a long short-term
[116] M. Kwon, H.-H. Kwon, D. Han, A hybrid approach combining conceptual memory (lstm) based model for predicting water table depth in agricultural
hydrological models, support vector machines and remote sensing data for areas, J. Hydrol. 561 (2018) 918–929.
rainfall-runoff modeling, Remote Sensing 12 (11) (2020) 1801. [143] J. Jeong, E. Park, H. Chen, K.-Y. Kim, W.S. Han, H. Suk, Estimation of
[117] A. Malik, A. Kumar, S. Kim, M.H. Kashani, V. Karimi, A. Sharafati, M.A. groundwater level based on the robust training of recurrent neural networks
Ghorbani, N. Al-Ansari, S.Q. Salih, Z.M. Yaseen, et al., Modeling monthly pan using corrupted data, J. Hydrol. 582 (2020) 124512.
evaporation process over the indian central himalayas: application of [144] J. Jeong, E. Park, Comparative applications of data-driven models
multiple learning artificial intelligence model, Eng. Appl. Comput. Fluid representing water table fluctuations, J. Hydrol. 572 (2019) 261–273.
Mech. 14 (1) (2020) 323–338. [145] Z.M. Yaseen, I. Ebtehaj, S. Kim, H. Sanikhani, H. Asadi, M.I. Ghareb, H.
[118] J.-L. Chen, H. Yang, M.-Q. Lv, Z.-L. Xiao, S.J. Wu, Estimation of monthly pan Bonakdari, W.H.M. Wan Mohtar, N. Al-Ansari, S. Shahid, Novel hybrid data-
evaporation using support vector machine in three gorges reservoir area, intelligence model for forecasting monthly rainfall with uncertainty analysis,
china, Theoret. Appl. Climatol. 138 (1–2) (2019) 1095–1107. Water 11 (3) (2019) 502.
[119] M.Y. Chia, Y.F. Huang, C.H. Koo, Support vector machine enhanced empirical [146] Z.M. Yaseen, M.F. Allawi, H. Karami, M. Ehteram, S. Farzin, A.N. Ahmed, S.B.
reference evapotranspiration estimation with limited meteorological Koting, N.S. Mohd, W.Z.B. Jaafar, H.A. Afan, et al., A hybrid bat–swarm
parameters, Comput. Electron. Agricul.s 175 (2020) 105577. algorithm for optimizing dam and reservoir operation, Neural Comput. Appl.
[120] L.B. Ferreira, F.F. da Cunha, R.A. de Oliveira, E.I. Fernandes Filho, Estimation of 31 (12) (2019) 8807–8821.
reference evapotranspiration in brazil with limited meteorological data using [147] O.H. Kombo, S. Kumaran, Y.H. Sheikh, A. Bovim, K. Jayavel, Long-term
ann and svm–a new approach, J. Hydrol. 572 (2019) 556–570. groundwater level prediction model based on hybrid knn-rf technique,
[121] M. Behzad, K. Asghari, E.A. Coppola Jr, Comparative study of svms and anns in Hydrology 7 (3) (2020) 59.
aquifer water level prediction, J. Comput. Civil Eng. 24 (5) (2010) 408–413. [148] K.-H. Chang, G. Lin, Optimal design of hybrid renewable energy systems using
[122] S.-Y. Liong, C. Sivapragasam, Flood stage forecasting with support vector simulation optimization, Simul. Model. Pract. Theory 52 (2015) 40–51.
machines 1, JAWRA J. Am. Water Resour. Assoc. 38 (1) (2002) 173–186. [149] Q.B. Pham, S.I. Abba, A.G. Usman, N.T.T. Linh, V. Gupta, A. Malik, R. Costache,
[123] N.D. Bokde, Z.H. Ali, M.T. Al-Hadidi, A.A. Farooque, M. Jamei, A.A. Al Maliki, B. N.D. Vo, D.Q. Tri, Potential of hybrid data-intelligence algorithms for multi-
H. Beyaztas, H. Faris, Z.M. Yaseen, Total dissolved salt prediction using station modelling of rainfall, Water Resour. Manage. 33 (15) (2019) 5067–
neurocomputing models: Case study of gypsum soil within iraq region, IEEE 5087.
Access 9 (2021) 53617–53635. [150] V. Nourani, A.H. Baghanam, J. Adamowski, O. Kisi, Applications of hybrid
[124] H.-T. Fang, B.-C. Jhong, Y.-C. Tan, K.-Y. Ke, M.-H. Chuang, A two-stage wavelet–artificial intelligence models in hydrology: a review, J. Hydrol. 514
approach integrating som-and moga-svm-based algorithms to forecast (2014) 358–377.
spatial-temporal groundwater level with meteorological factors, Water [151] G.P. Zhang, Time series forecasting using a hybrid arima and neural network
Resour. Manage. 33 (2) (2019) 797–818. model, Neurocomputing 50 (2003) 159–175.
[125] S.M. Guzman, J.O. Paz, M.L.M. Tagert, A.E. Mercer, Evaluation of seasonally [152] M. Ghorbani, R.C. Deo, Z.M. Yaseen, M.H. Kashani, B. Mohammadi, Pan
classified inputs for the prediction of daily groundwater levels: Narx evaporation prediction using a hybrid multilayer perceptron-firefly
networks vs support vector machines, Environ. Model. Assess. 24 (2) algorithm (mlp-ffa) model: case study in north iran, Theor. Appl. Climatol.
(2019) 223–234. 133 (3–4) (2018) 1119–1131.
[126] S. Nie, J. Bian, H. Wan, X. Sun, B. Zhang, Simulation and uncertainty analysis [153] T. Kashiwao, K. Nakayama, S. Ando, K. Ikeda, M. Lee, A. Bahadori, A neural
for groundwater levels using radial basis function neural network and network-based local rainfall prediction system using meteorological data on
support vector machine models, J. Water Supply: Res. Technol.–AQUA 66 (1) the internet: A case study using data from the japan meteorological agency,
(2017) 15–24. Appl. Soft Comput. 56 (2017) 317–330.
[127] M. Sahoo, A. Kasot, A. Dhar, A. Kar, On predictability of groundwater level in [154] N. Natarajan, C. Sudheer, Groundwater level forecasting using soft computing
shallow wells using satellite observations, Water Resour. Manage. 32 (4) techniques, Neural Comput. Appl. 32 (12) (2020) 7691–7708.
(2018) 1225–1244. [155] A. Jain, A.M. Kumar, Hybrid neural network models for hydrologic time series
[128] S.R. Naganna, B.H. Beyaztas, N. Bokde, A.M. Armanuos, On the evaluation of forecasting, Appl. Soft Comput. 7 (2) (2007) 585–592.
the gradient tree boosting model for groundwater level forecasting, [156] V. Nourani, A.H. Baghanam, J. Adamowski, O. Kisi, Applications of hybrid
Knowledge-Based Eng. Sci. 1 (01) (2020) 48–57. wavelet–artificial intelligence models in hydrology: a review, J. Hydrol. 514
[129] M.T. Sattari, R. Mirabbasi, R.S. Sushab, J. Abraham, Prediction of groundwater (2014) 358–377.
level in ardebil plain using support vector regression and m5 tree model, [157] F. Fahimi, Z.M. Yaseen, A. El-shafie, Application of soft computing based
Groundwater 56 (4) (2018) 636–646. hybrid models in hydrological variables modeling: a comprehensive review,
[130] Y. Tang, C. Zang, Y. Wei, M. Jiang, Data-driven modeling of groundwater level Theor. Appl. Climatol. 128 (3–4) (2017) 875–903.
with least-square support vector machine and spatial–temporal analysis, [158] A. Sharafati, M. Haghbin, M.S. Aldlemy, M.H. Mussa, A.W. Al Zand, M. Ali, S.K.
Geotech. Geol. Eng. 37 (3) (2019) 1661–1670. Bhagat, N. Al-Ansari, Z.M. Yaseen, Development of advanced computer aid
[131] H. Yoon, Y. Hyun, K. Ha, K.-K. Lee, G.-B. Kim, A method to improve the model for shear strength of concrete slender beam prediction, Appl. Sci. 10
stability and accuracy of ann-and svm-based time series models for long- (11) (2020) 3811.
term groundwater level predictions, Comput. Geosci. 90 (2016) 144–155.
302
[159] S.Q. Salih, M. Habib, I. Aljarah, H. Faris, Z.M. Yaseen, An evolutionary network trained with particle swarm optimization, Hydrol. Sci. J. 59 (6)
optimized artificial intelligence model for modeling scouring depth of (2014) 1225–1239.
submerged weir, Eng. Appl. Artif. Intell. 96 (2020) 104012. [187] Z. He, Y. Zhang, Q. Guo, X. Zhao, Comparative study of artificial neural
[160] S. Ardabili, A. Mosavi, A.R. Várkonyi-Kóczy, Advances in machine learning networks and wavelet artificial neural networks for groundwater depth data
modeling reviewing hybrid and ensemble methods, in: International forecasting with various curve fractal dimensions, Water Resour. Manage. 28
Conference on Global Research and Education Springer, 2019, pp. 215–227. (15) (2014) 5297–5317.
[161] A.Y. Sun, D. Wang, X. Xu, Monthly streamflow forecasting using gaussian [188] M.K. Jha, S. Sahoo, Efficacy of neural network and genetic algorithm
process regression, J. Hydrol. 511 (2014) 72–81. techniques in simulating spatio-temporal fluctuations of groundwater,
[162] T. Roshni, M.K. Jha, R.C. Deo, A. Vandana, Development and evaluation of Hydrol. Processes 29 (5) (2015) 671–691.
hybrid artificial neural network architectures for modeling spatio-temporal [189] T.-H. Yang, W.-H. Shi, P.-T. Wang, H.-L. Liu, Q.-L. Yu, Y. Li, Numerical
groundwater fluctuations in a complex aquifer system, Water Resour. simulation on slope stability analysis considering anisotropic properties of
Manage 33 (7) (2019) 2381–2397. layered fractured rocks: a case study, Arab. J. Geosci. 8 (8) (2015) 5413–5421.
[163] S. Mathur, Groundwater level forecasting using svm-pso, International [190] B. Khalil, S. Broda, J. Adamowski, B. Ozga-Zielinski, A. Donohoe, Short-term
Journal of Hydrology, Sci. Technol. 2 (2) (2012) 202–218. forecasting of groundwater levels under conditions of mine-tailings recharge
[164] F.-J. Chang, L.-C. Chang, C.-W. Huang, I.-F. Kao, Prediction of monthly regional using wavelet ensemble neural network models, Hydrogeol. J. 23 (1) (2015)
groundwater levels through hybrid soft-computing techniques, J. Hydrol. 541 121–141.
(2016) 965–976. [191] V. Nourani, M.T. Alami, F.D. Vousoughi, Wavelet-entropy data pre-processing
[165] F. Huang, J. Huang, S.-H. Jiang, C. Zhou, Prediction of groundwater levels using approach for ann-based groundwater level modeling, J. Hydrol. 524 (2015)
evidence of chaos and support vector machine, J. Hydroinform. 19 (4) (2017) 255–269.
586–606. [192] J.-C. Han, Y. Huang, Z. Li, C. Zhao, G. Cheng, P. Huang, Groundwater level
[166] G. Rakhshandehroo, H. Akbari, M. Afshari Igder, E. Ostadzadeh, Long-term prediction using a som-aided stepwise cluster inference model, J. Environ.
groundwater-level forecasting in shallow and deep wells using wavelet Manage. 182 (2016) 308–321.
neural networks trained by an improved harmony search algorithm, J. [193] Z. Hosseini, S. Gharechelou, M. Nakhaei, S. Gharechelou, Optimal design of bp
Hydrol. Eng. 23 (2) (2018) 04017058. algorithm by aco r model for groundwater-level forecasting: A case study on
[167] S. Balavalikar, P. Nayak, N. Shenoy, K. Nayak, Particle swarm optimization shabestar plain, iran, Arab. J. Geosci. 9 (6) (2016) 436.
based artificial neural network model for forecasting groundwater level in [194] H. Ebrahimi, T. Rajaee, Simulation of groundwater level variations using
udupi district, in: AIP Conference Proceedings, Vol. 1952, AIP Publishing LLC, wavelet combined with neural network, linear regression and support vector
2018, p. 020021. machine, Global Planet. Change 148 (2017) 181–191.
[168] K. Muralitharan, R. Sakthivel, R. Vishnuvarthan, Neural network based [195] R. Barzegar, E. Fijani, A.A. Moghaddam, E. Tziritis, Forecasting of groundwater
optimization approach for energy demand prediction in smart grid, level fluctuations using ensemble hybrid multi-wavelet neural network-
Neurocomputing 273 (2018) 199–208. based models, Sci. Total Environ. 599 (2017) 20–31.
[169] J. Wu, J. Long, M. Liu, Evolving rbf neural networks for rainfall prediction [196] A.D. Mehr, An improved gene expression programming model for streamflow
using hybrid particle swarm optimization and genetic algorithm, forecasting in intermittent streams, J. Hydrol. 563 (2018) 669–678.
Neurocomputing 148 (2015) 136–142. [197] J. Shiri, Ö. KisßI, Comparison of genetic programming with neuro-fuzzy
[170] H.M. Hasanien, Performance improvement of photovoltaic power systems systems for predicting short-term water table depth fluctuations, Comput.
using an optimal control strategy based on whale optimization algorithm, Geosci. 37 (10) (2011) 1692–1701.
Electric Power Syst. Res. 157 (2018) 168–176. [198] M. Cobaner, B. Babayigit, A. Dogan, Estimation of groundwater levels with
[171] B.S. Yıldız, A.R. Yıldız, Comparison of grey wolf, whale, water cycle, ant lion surface observations via genetic programming, J.-Am. Water Works Assoc.
and sine-cosine algorithms for the optimization of a vehicle engine 108 (6) (2016) E335–E348.
connecting rod, Mater. Testing 60 (3) (2018) 311–315. [199] A.D. Mehr, V. Nourani, E. Kahya, B. Hrnjica, A.M. Sattar, Z.M. Yaseen, Genetic
[172] N.B. Dash, S.N. Panda, R. Remesan, N. Sahoo, Hybrid neural modeling for programming in water resources engineering: A state-of-the-art review, J.
groundwater level prediction, Neural Comput. Appl. 19 (8) (2010) 1251– Hydrol. 566 (2018) 643–667.
1263. [200] J.R. Koza, J.R. Koza, Genetic programming: on the programming of computers
[173] A. Jalalkamali, N. Jalalkamali, Groundwater modeling using hybrid of artificial by means of natural selection, vol. 1, MIT Press, 1992.
neural network with genetic algorithm, Afr. J. Agric. Res. 6 (26) (2011) 5775– [201] B. Hrnjica, A. Danandeh Mehr, Optimized Genetic Programming Applications:
5784. Emerging Research and Opportunities, Emerging Research and Opportunities,
[174] J. Adamowski, H.F. Chan, A wavelet neural network conjunction model for IGI global (2018).
groundwater level forecasting, J. Hydrol. 407 (1–4) (2011) 28–40. [202] A.D. Mehr, V. Nourani, Season algorithm-multigene genetic programming: a
[175] E. Tapoglou, G.P. Karatzas, I.C. Trichakis, E.A. Varouchakis, A spatio-temporal new approach for rainfall-runoff modelling, Water Resour. Manage 32 (8)
hybrid neural network-kriging model for groundwater level simulation, J. (2018) 2665–2679.
Hydrol. 519 (2014) 3193–3203. [203] B. Hrnjica, A.D. Mehr, Š. Behrem, N. Ağıralioğlu, Genetic programming for
[176] M. Malekzadeh, S. Kardar, K. Saeb, S. Shabanlou, L. Taghavi, A novel approach turbidity prediction: hourly and monthly scenarios, Pamukkale Üniversitesi
for prediction of monthly ground water level using a hybrid wavelet and non- Mühendislik Bilimleri Dergisi 25 (8) (2019) 992–997.
tuned self-adaptive machine learning model, Water Resour. Manage. 33 (4) [204] R. Tür, Maximum wave height hindcasting using ensemble linear-nonlinear
(2019) 1609–1628. models, Theoret. Appl. Climatol. 141 (2020) 1434–4483.
[177] B. Supreetha, K.P. Nayak, K.N. Shenoy, Hybrid artificial intelligence based abc- [205] A. Rahmani-Rezaeieh, M. Mohammadi, A.D. Mehr, Ensemble gene expression
pso system for ground water level forecasting in udupi region, J. Eng. Sci. programming: a new approach for evolution of parsimonious streamflow
Technol. 14 (2) (2019) 797–809. forecasting model, Theoret. Appl. Climatol. 139 (1–2) (2020) 549–564.
[178] T. Roshni, M.K. Jha, J. Drisya, Neural network modeling for groundwater-level [206] M. Sadat-Noori, W. Glamore, D. Khojasteh, Groundwater level prediction
forecasting in coastal aquifers, Neural Comput. Appl. (2020) 1–18. using genetic programming: the importance of precipitation data and
[179] F.B. Banadkooki, M. Ehteram, A.N. Ahmed, F.Y. Teo, C.M. Fai, H.A. Afan, M. weather station location on model accuracy, Environ. Earth Sci. 79 (2020) 37.
Sapitang, A. El-Shafie, Enhancement of groundwater-level prediction using an [207] T. Hastie, R. Tibshirani, J. Friedman, The elements of statistical learning: data
integrated machine learning model optimized by whale algorithm, Nat. mining, inference, and prediction, Springer Science & Business Media, 2009.
Resour. Res. (2020) 1–20. [208] X. Liang, Y. Liang, Applications of data mining in hydrology, in: Proceedings
[180] A. Seifi, M. Ehteram, V.P. Singh, A. Mosavi, Modeling and uncertainty analysis 2001 IEEE International Conference on Data Mining IEEE, 2001, pp. 617–620.
of groundwater level using six evolutionary optimization algorithms [209] D. Karthik, K. Vijayarekha, Multivariate data mining techniques for assessing
hybridized with anfis, svm, and ann, Sustainability 12 (10) (2020) 4023. water potability, Rasayan J. Chem. 7 (3) (2014) 256–259.
[181] V. Nourani, R.G. Ejlali, M.T. Alami, Spatiotemporal groundwater level [210] M. Pal, S. Deswal, Modelling pile capacity using gaussian process regression,
forecasting in coastal aquifers by hybrid artificial neural network- Comput. Geotech. 37 (7–8) (2010) 942–947.
geostatistics model: a case study, Environ. Eng. Sci. 28 (3) (2011) 217–228. [211] M.-H. Koo, E. Park, J. Jeong, H. Lee, H.G. Kim, M. Kwon, Y. Kim, S. Nam, J.Y. Ko,
[182] V. Moosavi, M. Vafakhah, B. Shirmohammadi, N. Behnia, A wavelet-anfis J.H. Choi, et al., Applications of gaussian process regression to groundwater
hybrid model for groundwater level forecasting for different prediction quality data, J. Soil Groundwater Environ. 21 (6) (2016) 67–79.
periods, Water Resour. Manage. 27 (5) (2013) 1301–1321. [212] C.E. Rasmussen, Gaussian processes in machine learning, in: Summer School
[183] R. Maheswaran, R. Khosa, Long term forecasting of groundwater levels with on Machine Learning, Springer, 2003, pp. 63–71.
evidence of non-stationary and nonlinear characteristics, Comput. Geosci. 52 [213] P. Moonen, J. Allegrini, Employing statistical model emulation as a surrogate
(2013) 422–436. for cfd, Environ. Modell. Software 72 (2015) 77–91.
[184] S. Emamgholizadeh, K. Moslemi, G. Karami, Prediction the groundwater level [214] N.E. Owen, P. Challenor, P.P. Menon, S. Bennani, Comparison of surrogate-
of bastam plain (iran) by artificial neural network (ann) and adaptive neuro- based uncertainty quantification methods for computationally expensive
fuzzy inference system (anfis), Water Resour. Manage. 28 (15) (2014) 5433– simulators, SIAM/ASA J. Uncertainty Quantif. 5 (1) (2017) 403–435.
5446. [215] A.Y. Sun, D. Wang, X. Xu, Monthly streamflow forecasting using gaussian
[185] C. Suryanarayana, C. Sudheer, V. Mahammood, B.K. Panigrahi, An integrated process regression, J. Hydrol. 511 (2014) 72–81.
wavelet-support vector machine for groundwater level prediction in [216] Y. Wang, C. Ocampo-Martínez, V. Puig, J. Quevedo, Gaussian-process-based
visakhapatnam, india, Neurocomputing 145 (2014) 324–335. demand forecasting for predictive control of drinking water networks, in:
[186] E. Tapoglou, I.C. Trichakis, Z. Dokou, I.K. Nikolos, G.P. Karatzas, Groundwater- International Conference on Critical Information Infrastructures Security
level forecasting under climate change scenarios using an artificial neural Springer, 2014, pp. 69–80.
303
[217] M. Liu, C. Huang, L. Wang, Y. Zhang, X. Luo, Short-term soil moisture water resources management, in: 2018 IEEE Conference on Technologies for
forecasting via gaussian process regression with sample selection, Water 12 Sustainability (SusTech) IEEE, 2018, pp. 1–7.
(11) (2020) 3085. [247] F. Aburub, W. Hadi, Predicting groundwater areas using data mining
[218] S. Shabani, S. Samadianfard, M.T. Sattari, A. Mosavi, S. Shamshirband, T. Kmet, techniques: Groundwater in jordan as case study, Int. J. Comput. Inform.
A.R. Várkonyi-Kóczy, Modeling pan evaporation using gaussian process Eng. 10 (9) (2016) 1621–1624.
regression k-nearest neighbors random forest and support vector [248] N.S. Raghavendra, P.C. Deka, Forecasting monthly groundwater level
machines; comparative analysis, Atmosphere 11 (1) (2020) 66. fluctuations in coastal aquifers using hybrid wavelet packet–support vector
[219] L. Pasolli, F. Melgani, E. Blanzieri, Gaussian process regression for estimating regression, Cogent Eng. 2 (1) (2015) 999414.
chlorophyll concentration in subsurface waters from remote sensing data, [249] M. Rezaie-balf, S.R. Naganna, A. Ghaemi, P.C. Deka, Wavelet coupled mars and
IEEE Geosci. Remote Sens. Lett. 7 (3) (2010) 464–468. m5 model tree approaches for groundwater level forecasting, J. Hydrol. 553
[220] P. Sihag, M. Kumar, B. Singh, Assessment of infiltration models developed (2017) 356–373.
using soft computing techniques, Geol., Ecol., Landscapes (2020) 1–11. [250] S.G. Mallat, A theory for multiresolution signal decomposition: the wavelet
[221] P. Sihag, N. Tiwari, S. Ranjan, Modelling of infiltration of sandy soil using representation, IEEE Trans. Pattern Anal. Mach. Intell. 11 (7) (1989) 674–693.
gaussian process regression, Model. Earth Syst. Environ. 3 (3) (2017) 1091– [251] N. Bokde, A. Feijóo, D. Villanueva, K. Kulat, A review on hybrid empirical
1100. mode decomposition models for wind speed and wind power prediction,
[222] M. Karbasi, Forecasting of multi-step ahead reference evapotranspiration Energies 12 (2) (2019) 254.
using wavelet-gaussian process regression model, Water Resour. Manage. 32 [252] R. Bahmani, T.B. Ouarda, Groundwater level modeling with hybrid artificial
(3) (2018) 1035–1052. intelligence techniques, J. Hydrol. 125659 (2020).
[223] D. Holman, M. Sridharan, P. Gowda, D. Porter, T. Marek, T. Howell, J. [253] Y. Seo, S. Kim, O. Kisi, V.P. Singh, Daily water level forecasting using wavelet
Moorhead, Gaussian process models for reference et estimation from decomposition and artificial intelligence techniques, J. Hydrol. 520 (2015)
alternative meteorological data sources, J. Hydrol. 517 (2014) 28–35. 224–243.
[224] R. Grbić, D. Kurtagić, D. Slišković, Stream water temperature prediction based [254] D. Liu, G. Li, Q. Fu, M. Li, C. Liu, M.A. Faiz, M.I. Khan, T. Li, S. Cui, Application of
on gaussian process regression, Expert Syst. Appl. 40 (18) (2013) 7407–7414. particle swarm optimization and extreme learning machine forecasting
[225] J. Hu, J. Wang, Short-term wind speed prediction using empirical wavelet models for regional groundwater depth using nonlinear prediction models as
transform and gaussian process regression, Energy 93 (2015) 1456–1466. preprocessor, J. Hydrol. Eng. 23 (12) (2018) 04018052.
[226] K. Roushangar, S. Garekhani, F. Alizadeh, Forecasting daily seepage discharge [255] V. Moosavi, M. Vafakhah, B. Shirmohammadi, N. Behnia, A wavelet-anfis
of an earth dam using wavelet–mutual information–gaussian process hybrid model for groundwater level forecasting for different prediction
regression approaches, Geotech. Geol. Eng. 34 (5) (2016) 1313–1326. periods, Water Resour. Manage. 27 (5) (2013) 1301–1321.
[227] O. Samuelsson, A. Björk, J. Zambrano, B. Carlsson, Gaussian process regression [256] A.S. Rahman, T. Hosono, J.M. Quilty, J. Das, A. Basak, Multiscale groundwater
for monitoring and fault detection of wastewater treatment processes, Water level forecasting: Coupling new machine learning approaches with wavelet
Sci. Technol. 75 (12) (2017) 2952–2963. transforms, Adv. Water Resour. 103595 (2020).
[228] Y. Wang, C. Ocampo-Martínez, V. Puig, J. Quevedo, Gaussian-process-based [257] Y. Gong, Z. Wang, G. Xu, Z. Zhang, A comparative study of groundwater level
demand forecasting for predictive control of drinking water networks, in: forecasting using data-driven models based on ensemble empirical mode
International Conference on Critical Information Infrastructures Security decomposition, Water 10 (6) (2018) 730.
Springer, 2014, pp. 69–80. [258] X. Wen, Q. Feng, R.C. Deo, M. Wu, J. Si, Wavelet analysis–artificial neural
[229] J. Zhao, H. Guo, M. Han, H. Tang, X. Li, Gaussian process regression for network conjunction models for multi-scale monthly groundwater level
prediction of sulfate content in lakes of china, J. Eng. Technol. Sci. 51 (2) predicting in an arid inland river basin, northwestern china, Hydrol. Res. 48
(2019) 198–215. (6) (2017) 1710–1729.
[230] P. Vijai, P.B. Sivakumar, Performance comparison of techniques for water [259] F. Yosefvand, S. Shabanlou, Forecasting of groundwater level using ensemble
demand forecasting, Procedia Comput. Sci. 143 (2018) 258–266. hybrid wavelet–self-adaptive extreme learning machine-based models, Nat.
[231] G. Kopsiaftis, E. Protopapadakis, A. Voulodimos, N. Doulamis, A. Mantoglou, Resour. Res. (2020) 1–18.
Gaussian process regression tuned by bayesian optimization for seawater [260] S.A. Naghibi, H.R. Pourghasemi, K. Abbaspour, A comparison between ten
intrusion prediction, Comput. Intell. Neurosci. (2019). advanced and soft computing models for groundwater qanat potential
[232] A. Desai, V. Schwamberger, T. Herzog, J. Jaanchen, F.P. Schmidt, Modeling of assessment in iran using r and gis, Theoret. Appl. Climatol. 131 (3–4)
adsorption equilibria through gaussian process regression of data in (2018) 967–984.
dubinin’s representation: Application to water/zeolite li-lsx, Ind. Eng. [261] M. Al-Mukhtar, Random forest, support vector machine, and neural networks
Chem. Res. 58 (37) (2019) 17549–17554. to modelling suspended sediment in tigris river-baghdad, Environ. Monitor.
[233] K. Blix, G. Camps-Valls, R. Jenssen, Gaussian process sensitivity analysis for Assess. 191 (11) (2019) 673.
oceanic chlorophyll estimation, IEEE J. Selected Top. Appl. Earth Observ. [262] R. McCuen, Modeling Hydrologic Change: Statistical Methods, CRC Press
Remote Sens. 10 (4) (2017) 1265–1277. (2016), URL: https://books.google.dk/books?id=iI_MBQAAQBAJ.
[234] S. Maatta, Predicting groundwater levels using linear regression and neural [263] S. Mahdouh, H. van Oorschot, S. de Lange, et al., Statistical analysis in water
networks, Ground Water 3 (2011) 3–5. resources engineering., Statistical analysis in water resources engineering.
[235] F. Aburub, W. Hadi, A new associative classification algorithm for predicting [264] G.E. Box, G.M. Jenkins, G.C. Reinsel, G.M. Ljung, Time series analysis:
groundwater locations, J. Inform. Knowl. Manage. 17 (04) (2018) 1850043. forecasting and control, John Wiley & Sons, 2015.
[236] E.A. Varouchakis, G.P. Karatzas, Gaussian process regression for [265] X. Yang, J. Zhou, W. Fang, Y. Wang, An ensemble flow forecast method based
spatiotemporal analysis of groundwater level variations, EGU General on autoregressive model and hydrological uncertainty processer, Water 12
Assembly Conference Abstracts (2020) 6665. (11) (2020) 3138.
[237] K. Kolli, R. Seshadri, Ground water quality assessment using data mining [266] A. Katimon, S. Shahid, M. Mohsenipour, Modeling water quality and
techniques, Int. J. Comput. Appl. 76 (15). hydrological variables using arima: a case study of johor river, malaysia,
[238] F.E. Colchester, H.G. Marais, P. Thomson, R. Hope, D.A. Clifton, Accidental Sustain. Water Resources Manage. 4 (4) (2018) 991–998.
infrastructure for groundwater monitoring in africa, Environ. Modell. [267] F. Cui, S.Q. Salih, B. Choubin, S.K. Bhagat, P. Samui, Z.M. Yaseen, Newly
Software 91 (2017) 241–250. explored machine learning model for river flow time series forecasting at
[239] M.D. Kordestani, S.A. Naghibi, H. Hashemi, K. Ahmadi, B. Kalantar, B. Pradhan, mary river, australia, Environ. Monit. Assess. 192 (12) (2020) 1–15.
Groundwater potential mapping using a novel data-mining ensemble model, [268] S. Mehdizadeh, Using ar, ma, and arma time series models to improve the
Hydrogeol. J. 27 (1) (2019) 211–224. performance of mars and knn approaches in monthly precipitation modeling
[240] H.R. Pourghasemi, M. Beheshtirad, Assessment of a data-driven evidential under limited climatic data, Water Resour. Manage 34 (1) (2020) 263–282.
belief function model and gis for groundwater potential mapping in the [269] P. Aghelpour, V. Varshavian, Evaluation of stochastic and artificial
koohrang watershed, iran, Geocarto Int. 30 (6) (2015) 662–685. intelligence models in modeling and predicting of river daily flow time
[241] S. Azimi, M.A. Moghaddam, S.A.H. Monfared, Large-scale association analysis series, Stoch. Env. Res. Risk Assess. 34 (1) (2020) 33–50.
of climate drought and decline in groundwater quantity using gaussian [270] A. McLeod, P.H. Sales, Algorithm as 191: An algorithm for approximate
process classification (case study: 609 study area of iran), J. Environ. Health likelihood calculation of arma and seasonal arma models, Appl. Stat. (1983)
Sci. Eng. 16 (2) (2018) 129–145. 211–223.
[242] H.G. Kim, E. Park, J. Jeong, W.S. Han, K.-Y. Kim, Groundwater level trend [271] G. Melard, Algorithm as 197: A fast algorithm for the exact likelihood of
analysis for long-term prediction basedon gaussian process regression, J. Soil autoregressive-moving average models, J. R. Stat. Soc. Ser. C (Appl. Stat.) 33
Groundwater Environ. 21 (4) (2016) 30–41. (1) (1984) 104–114.
[243] J. Zhang, W. Li, L. Zeng, L. Wu, An adaptive gaussian process-based method for [272] G.E. Box, D.A. Pierce, Distribution of residual autocorrelations in
efficient bayesian experimental design in groundwater contaminant source autoregressive-integrated moving average time series models, J. Am. Stat.
identification problems, Water Resour. Res. 52 (8) (2016) 5971–5984. Assoc. 65 (332) (1970) 1509–1526.
[244] O. Bozorg-Haddad, M. Delpasand, H.A. Loáiciga, Self-optimizer data-mining [273] H.-C. Chan, P.-A. Chen, J.-T. Lee, Rainfall-induced landslide susceptibility
method for aquifer level prediction, Water Supply 20 (2) (2020) 724–736. using a rainfall–runoff model and logistic regression, Water 10 (10) (2018)
[245] M.M. Rajabi, H. Ketabchi, Uncertainty-based simulation-optimization using 1354.
gaussian process emulation: application to coastal groundwater [274] J. Ren, B. Ren, Q. Zhang, X. Zheng, A novel hybrid extreme learning machine
management, J. Hydrol. 555 (2017) 518–534. approach improved by k nearest neighbor method and fireworks algorithm
[246] A. Lal, B. Datta, Genetic programming and gaussian process regression for flood forecasting in medium and small watershed of loess region, Water
models for groundwater salinity prediction: Machine learning for sustainable 11 (9) (2019) 1848.
304
[275] S. Janizadeh, M. Avand, A. Jaafari, T.V. Phong, M. Bayat, E. Ahmadisharaf, I. [301] R. Martínez, N. Vela, A. e. Aatik, E. Murray, P. Roche, J.M. Navarro, On the use
Prakash, B.T. Pham, S. Lee, Prediction success of machine learning methods of an iot integrated system for water quality monitoring and management in
for flash flood susceptibility mapping in the tafresh watershed, iran, wastewater treatment plants, Water 12 (4) (2020) 1096.
Sustainability 11 (19) (2019) 5426. [302] Y. Fan, H. Li, G. Miguez-Macho, Global patterns of groundwater table depth,
[276] O. Rahmati, F. Falah, K.S. Dayal, R.C. Deo, F. Mohammadi, T. Biggs, D.D. Science 339 (6122) (2013) 940–943.
Moghaddam, S.A. Naghibi, D.T. Bui, Machine learning approaches for spatial [303] Y.-S. Su, C.-F. Ni, W.-C. Li, I.-H. Lee, C.-P. Lin, Applying deep learning
modeling of agricultural droughts in the south-east region of queensland algorithms to enhance simulation of large-scale groundwater flow in iots,
australia, Sci. Total Environ. 699 (2020) 134230. Appl. Soft Comput. 106298 (2020).
[277] M. Mirzavand, R. Ghazavi, A stochastic modelling technique for groundwater [304] R. Dehghani, H.T. Poudeh, Z. Izadi, The effect of climate change on
level forecasting in an arid environment using time series methods, Water groundwater level and its prediction using modern meta-heuristic model,
Resour. Manage. 29 (4) (2015) 1315–1328. Groundwater Sustain. Devel. 16 (2022) 100702.
[278] B. Choubin, A. Malekian, Combined gamma and m-test-based ann and arima [305] B. Halder, M. Haghbin, A.A. Farooque, An assessment of urban expansion
models for groundwater fluctuation forecasting in semiarid regions, Environ. impacts on land transformation of rajpur-sonarpur municipality, Knowl.-
Earth Sci. 76 (15) (2017) 538. Based Eng. Sci. 2 (3) (2021) 34–53.
[279] A. Gibrilla, G. Anornu, D. Adomako, Trend analysis and arima modelling of [306] H.A. Afan, M.F. Allawi, A. El-Shafie, Z.M. Yaseen, A.N. Ahmed, M.A. Malek, S.B.
recent groundwater levels in the white volta river basin of ghana, Koting, S.Q. Salih, W.H.M.W. Mohtar, S.H. Lai, et al., Input attributes
Groundwater for, Sustain. Develop. 6 (2018) 150–163. optimization using the feasibility of genetic nature inspired algorithm:
[280] M. Sakizadeh, M.M. Mohamed, H. Klammler, Trend analysis and spatial Application of river flow forecasting, Sci. Rep. 10 (1) (2020) 1–15.
prediction of groundwater levels using time series forecasting and a novel [307] S.Q. Salih, A. Sharafati, I. Ebtehaj, H. Sanikhani, R. Siddique, R.C. Deo, H.
spatio-temporal method, Water Resour. Manage 33 (4) (2019) 1425–1437. Bonakdari, S. Shahid, Z.M. Yaseen, Integrative stochastic model
[281] E.H. de Moraes Takafuji, M.M. da Rocha, R.L. Manzione, Groundwater level standardization with genetic algorithm for rainfall pattern forecasting in
prediction/forecasting and assessment of uncertainty using sgs and arima tropical and semi-arid environments, Hydrol. Sci. J. 65 (7) (2020) 1145–1157.
models: A case study in the bauru aquifer system (brazil), Nat. Resour. Res. 28 [308] A. Uyumaz, A. Danandeh Mehr, E. Kahya, H. Erdem, Rectangular side weirs
(2) (2019) 487–503. discharge coefficient estimation in circular channels using linear genetic
[282] M. Goodarzi, Application and performance evaluation of time series, neural programming approach, J. Hydroinform. 16 (6) (2014) 1318–1330.
networks and hartt models in predicting groundwater level changes, [309] A.D. Mehr, M.J.S. Safari, Multiple genetic programming: a new approach to
najafabad plain, iran, Sustain. Water Resour. Manage. 6 (4) (2020) 1–10. improve genetic-based month ahead rainfall forecasts, Environ. Monit.
[283] M.F. Allawi, O. Jaafar, M. Ehteram, F.M. Hamzah, A. El-Shafie, Synchronizing Assess. 192 (1) (2020) 25.
artificial intelligence models for operating the dam and reservoir system, [310] M. Das, S.K. Ghosh, A deep-learning-based forecasting ensemble to predict
Water Resour. Manage. 32 (10) (2018) 3373–3389. missing data for remote sensing analysis, IEEE J. Selected Top. Appl. Earth
[284] Q.A. Aljanabi, Z. Chik, M.F. Allawi, A.H. El-Shafie, A.N. Ahmed, A. El-Shafie, Observ. Remote Sens. 10 (12) (2017) 5228–5236.
Support vector regression-based model for prediction of behavior stone [311] P.J. García-Laencina, J.-L. Sancho-Gómez, A.R. Figueiras-Vidal, Pattern
column parameters in soft clay under highway embankment, Neural Comput. classification with missing data: a review, Neural Comput. Appl. 19 (2)
Appl. 30 (8) (2018) 2459–2469. (2010) 263–282.
[285] Y. Raptodimos, I. Lazakis, Application of narx neural network for predicting [312] J.M. Jerez, I. Molina, P.J. García-Laencina, E. Alba, N. Ribelles, M. Martín, L.
marine engine performance parameters, Ships Offshore Struct. 15 (4) (2020) Franco, Missing data imputation using statistical and machine learning
443–452. methods in a real breast cancer problem, Artif. Intell. Med. 50 (2) (2010) 105–
[286] A.A. Alsumaiei, A nonlinear autoregressive modeling approach for forecasting 115.
groundwater level fluctuation in urban aquifers, Water 12 (3) (2020) 820. [313] Q. Yu, Y. Miche, E. Eirola, M. Van Heeswijk, E. SéVerin, A. Lendasse,
[287] C. Zanotti, M. Rotiroti, S. Sterlacchini, G. Cappellini, L. Fumagalli, G.A. Stefania, Regularized extreme learning machine for regression with missing data,
M.S. Nannucci, B. Leoni, T. Bonomi, Choosing between linear and nonlinear Neurocomputing 102 (2013) 45–51.
models and avoiding overfitting for short and long term groundwater level [314] H. Ren, E. Cromwell, B. Kravitz, X. Chen, Using deep learning to fill spatio-
forecasting in a linear system, J. Hydrol. 578 (2019) 124015. temporal data gaps in hydrological monitoring networks, Hydrol. Earth Syst.
[288] S.M. Guzman, J.O. Paz, M.L.M. Tagert, The use of narx neural networks to Sci. Discuss. (2019) 1–20.
forecast daily groundwater levels, Water Resour. Manage. 31 (5) (2017) [315] M. Längkvist, L. Karlsson, A. Loutfi, A review of unsupervised feature learning
1591–1603. and deep learning for time-series modeling, Pattern Recogn. Lett. 42 (2014)
[289] R. Ezzeldin, A. Hatata, Application of narx neural network model for discharge 11–24.
prediction through lateral orifices, Alexand. Eng. J. 57 (4) (2018) 2991–2998. [316] X. Ma, Z. Tao, Y. Wang, H. Yu, Y. Wang, Long short-term memory neural
[290] A. Wunsch, T. Liesch, S. Broda, Forecasting groundwater levels using network for traffic speed prediction using remote microwave sensor data,
nonlinear autoregressive networks with exogenous input (narx), J. Hydrol. Transp. Res. Part C: Emerging Technol. 54 (2015) 187–197.
567 (2018) 743–758. [317] F. Cui, Z.A. Al-Sudani, G.S. Hassan, H.A. Afan, S.J. Ahammed, Z.M. Yaseen,
[291] F. Di Nunno, F. Granata, Groundwater level prediction in apulia region Boosted artificial intelligence model using improved alpha-guided grey wolf
(southern italy) using narx neural network, Environ. Res. 190 (2020) 110062. optimizer for groundwater level prediction: Comparative study and insight
[292] A. Al Jami, M.U. Himel, K. Hasan, S.R. Basak, A.F. Mita, Narx neural network for federated learning technology, J. Hydrol. 127384 (2021).
approach for the monthly prediction of groundwater levels in sylhet sadar, [318] N. Shiri, J. Shiri, Z.M. Yaseen, S. Kim, I.-M. Chung, V. Nourani, M. Zounemat-
bangladesh, J. Groundwater Sci. Eng. 8 (2) (2020) 118–126. Kermani, Development of artificial intelligence models for well groundwater
[293] S.A. Naghibi, H.R. Pourghasemi, B. Dixon, Gis-based groundwater potential quality simulation: Different modeling scenarios, Plos one 16 (5) (2021)
mapping using boosted regression tree, classification and regression tree, and e0251510.
random forest machine learning models in iran, Environ. Monitor. Assess. 188
(1) (2016) 44.
[294] N.A. Nalarajan, C. Mohandas, Groundwater level prediction using m5 model
trees, J. Inst. Eng. (India): Series A 96 (1) (2015) 57–62. Tao Hai, graduated with B.Sc at the Department of
[295] Y. Zhao, Y. Li, L. Zhang, Q. Wang, Groundwater level prediction of landslide Computer and Information Science of Northwest
based on classification and regression tree, Geodesy Geodyn. 7 (5) (2016) University of Nationalities from 2000 to 2004. He
348–355. completed his M.S at School of Mathematics and
[296] Y. Kaya, F. Ünesß, M. Demirci, B. Tasßar, H. Varçin, Groundwater level prediction Statistics of Lanzhou University in 2009. From 2009 to
using artificial neural network and m5 tree models, Aerul si Apa, 2012, he obtained his PhD at Faculty of Computer Sys-
Componente ale Mediului (2018) 195–201. tem and Software Engineering in University Malaysia
[297] X. Wang, T. Liu, X. Zheng, H. Peng, J. Xin, B. Zhang, Short-term prediction of Pahang. He is an associate professor in Baoji University
groundwater level using improved random forest regression with a of Arts and Sciences. His current research interests
combination of random features, Appl. Water Sci. 8 (5) (2018) 125. include machine learning, Internet of things and opti-
[298] A. Sharafati, S. Asadollah, A. Neshat, A new artificial intelligence strategy for mization computation.
predicting the groundwater level over the rafsanjan aquifer in iran, Journal of
Hydrology 591, cited By 0. doi:10.1016/j.jhydrol.2020.125468. URL:https://
www.scopus.com/inward/record.uri?eid=2-s2.0-85090282789& doi=10.1016 %
2fj.jhydrol.2020.125468&partnerID=40&md5=f361fd8d18bc5bd44aa23db338e
60ca7.
[299] S. Javadinejad, R. Dara, F. Jafary, Modelling groundwater level fluctuation in
an indian coastal aquifer, Water SA 46 (4) (2020) 665–671, cited By 0.
doi:10.17159/wsa/2020.v46.i4.9081. URL:https://www.scopus.com/inward/
record.uri?eid=2-s2.0-85094923349&doi=10.17159%2fwsa%2f2020.v46.i4.9081
&partnerID=40&md5=d146c404ee3e532056e679e64eadea92.
[300] A.A. Maroli, V.S. Narwane, R.D. Raut, B.E. Narkhede, Framework for the
implementation of an internet of things (iot)-based water distribution and
management system, Clean Technol. Environ. Policy (2020) 1–13.
305
Mr. Mohammed Majeed Hameed has received his Sadeq Oleiwi Sulaiman holds a Ph.D. degree in Civil
master’s degree from the National University of Engineering from the University of Technology (Iraq).
Malaysia in 2014. Currently, he is assigned as a lecturer Currently, he is a Full Professor in the Department of
at Al-Maarif university college, department of civil Dams & Water Resources Engineering, University of
engineering. He is interested in water resources engi- Anbar, Iraq. He has also worked as a postdoctoral
neering, environment and water pollutants, water researcher at the Lulea University of Technology (Swe-
quality, and applying machine learning models to solve den). Dr. Sulaiman has published more than 30 refereed
different engineering problems. publications and collaborated as subject editor/reviewer
for different international scholarly journals. His current
research focuses on rainfall-runoff, streamflow, water
stage, stochastic hydrology, optimization, hydrosystems
engineering, sediment transport, and using machine
learning and deep learning techniques to solve different problems in water
resources engineering.
Dr. Haydar A. Marhoon holds a Bachelor in Control and
System Engineering from University of Technology,
Baghdad in 2003. His Master’s degree and Ph.D in Dr. Mou Leong Tan is a senior lecturer in the field of
Information Technology are both from Universiti Utara physical geography. He has published more than pub-
Malaysia in 2012 and 2017 respectively. He works at the lished more than 70 scientific articles in predominantly
Computer Science, University of Kerbala and currently Q1 and Q2 journals that have been cited over 1600
in Al-Ayen University, Iraq. Research interests are times based on citation data reported by Google scholar,
Computer Networks, Communication Engineering and with an H-index of 19. The impact of his research also
computer graphics processing. extends across multiple disciplines, as confirmed by
citing studies published in the fields of hydrologic,
remote sensing, climatic, environmental engineering
and others.
Dr. Mohammad Zounemat-Kermani is a university

lecturer and researcher specializing in hydrosciences.
His research areas are focused on developing soft com- Dr Zulfaqar Sa’adi is postdoctoral researcher in the field
puting techniques and machine learning models in of hydro-climatology and currently working at Centre
hydraulics and hydrology applications. for Environmental Sustainability and Water Security
(IPASA), Universiti Teknologi Malaysia. The scope of his
research is in the field of hydro-climatological mod-
elling and climate change future projection for hydro-
logic resilience. He has published in various reputed
international journal. He has served as a reviewer for
more than 30 international journals.
Salim HEDDAM is a lecture and researcher in the field

of water resources management and civil engineering.
His current research interests include Modelling using
Artificial Intelligence Technique, Hydrology, Water Ali Danandeh Mehr holds Ph.D. degree in Civil Engi-
Resources Analysis Planning and management, Reser- neering from Istanbul Technical University (Turkey).
voir Water Quality Modeling, Application of statistical Currently, he is an associate professor at Civil Engi-
methods in Water and Environmental Resources, Envi- neering Department of Antalya Bilim University, Turkey.
ronmental and Water Quality, Hydrologic forecasting He has also worked as a postdoctoral researcher at
and modelling; suspended sediment modelling; fore- University of Tabriz (Iran) and University of Oulu (Fin-
casting hydrological variables such as precipitation, land). Dr. Danandeh Mehr has published more than 50
streamflows, evaporation, evapotranspiration, ground- refereed publications and collaborated as subject editor/
water, lake level; and also Hydro-Informatics. He is a reviewer for different international scholarly journals.
Reviewer for more than 50 journals indexed in Science Citation Index (SCI) in the His current research focuses on the stochastic hydrol-
fields of hydrology, irrigation, water resources and hydro-informatics. He has ogy, hydroinformatics, and developing evolutionary
authored more than 75 research articles, and 6 chapters (Web of Science h-index = algorithms to solve different problems in water
20, Scopus h-index 22), and a total number of citation over 1600 (Google Scholar h- resources systems.
index = 24).
Dr. Mohammed Falah Allawi is a Lecturer and

Professor Dr. Sungwon KIM is currently a Full Profes- Researcher in the water resources engineering field. His
sor in Department of Railroad Construction and Safety research is focused on developing and optimizing
Engineering, Dongyang University, South Korea. He has machine learning methods for analyzing and processing
conducted the international collaborative studies with patterns of hydrological variables. He is reviewer of
renowned international researchers on the impact of many journals some of them are Journal of Hydrology,
geophysical modeling including rainfall-runoff, Water Resources Management Journal and Water Jour-
streamflow, water stage, evaporation, evapotranspira- nal.
tion, and meteorological variables using machine
learning, deep learning, and data-driven techniques. He
has developed efficient methods and models that can be
used to characterize these natural processes. He has
published over 250 research papers including leading
international journals and total number of citation over 2100 (Google scholar H-
Index=26). In addition, he has served as reviewer more than 135 international
journals including hydrology, water resources, geophysics, artificial intelligence,
and machine learning fields and so on.
306
Dr S. I. Abba is a lecture and researcher in Civil Engi- Dr. Maryam Bayatvarkeshi is an Assistant Professor in
neering with a specialization in water resource and the Department of Soil Science at the University of
environmental engineering. He received a B.Sc. degree Malayer. She has majored in Irrigation and Drainage;
from Bayero University Kano (BUK), the MTech. Degree however, is passionate about hydrology in particular
from Sharda University, India, and from Near East modeling. Most of her researches focuses on the appli-
University (NEU), Cyprus. He is currently a researcher cation of intelligent models for tackling complicated
with King Fahad University of Petroleum and Minerals problems such as evapotranspiration, weather variables,
and a lecturer at Baze University, Nigeria. His research soil physical and chemical properties in the hydrology
interests include artificial intelligence, machine learn- and water resource management domain. She has
ing, hydrology, hydro-climatology, hydro- published over 12 research manuscripts in high-ranked
environmental modelling, computational engineering, journals so that they have been cited in over 300 pub-
soft computing, and optimization algorithm. In addition, lications.
he has excellent expertise in remote sensing, GIS, and advanced data analysis and
classification.
Dr. Mustafa Al-Mukhtar is working as associate pro-

Jasni Mohamad Zain received the bachelor’s degree in fessor of Water Resources Engineering at the Civil
computer science from the University of Liverpool, Engineering Department of the University of
Liverpool, U.K., in 1989 and the Ph.D. degree from Bru- Technology-Baghdad. His research interests focus
nel University, West London, U.K., in 2005., She started mainly on using the artificial intelligence methods in
her career as a Tutor in 1997 with the University of water resources management. He is interested in
Technology Malaysia, Johor Bahru, Malaysia. She is machine learning, statistical and stochastic modelling
currently a Professor with the Faculty of Computer and and he manipulated with variety of statistical software
Mathematical Sciences, Universiti Teknologi MARA, since 2011. He published several papers in peer
Shah Alam, Malaysia, and was the Dean of the Faculty of reviewed international journals in this field. With
Computer Systems and Software Engineering, Univer- respect to his academic experience profile, He started
sity Malaysia Pahang, Pahang, Malaysia, for eight years. his academic job in 2006 as assistant lecturer at the
She has graduated 15 Ph.D. students and 6 master’s University of Technology, Baghdad. In 2015, he completed his PhD degree from TU
students by research under her supervision. She has authored or co-authored over Freiberg in Germany. He taught several subjects related to the water resources and
100 refereed papers. She has a patent pending for digital watermarking. Her current hydrological engineering such as water management, fluid mechanics, hydrology,
research interests include digital watermarking and image processing, as well as dams engineering, and statistics. In addition, He was involved in some consultancy
data and network security. Dr. Zain has been actively presenting papers and key- projects such as designing wastewater treatment plant, and Environmental Impact
note addresses at national and international conferences. Assessment of several projects. He has supervised on many under and post graduate
students. In 2017, he honorably was selected for the Fulbright Visiting Scholar
Program.
Dr. Mayadah Waheed Falah is a lecturer and researcher
in the field of civil engineering. She has worked exten-
sively with diverse civil engineering applications and TIYASHA successfully secured National eligibility test
published with multiple disciplines. She has more in conducted by government of India. She awarded honors
data analytics and material engineering. in master’s degree in Environmental Engineering from
the Rajasthan Technical University, India in 2014. She
joined the Department of Civil Engineering, Institute of
Technology, Ambo University, Ethiopia, as a Lecturer,
where, she also led a funded project which won the
Energy Globe Award, issued by Austrian energy pioneer
Wolfgang Neumann. At present, She is pursuing a Ph.D.
degree in Faculty of Civil Engineering from Ton Duc
Thang University, Ho Chi Minh City, Vietnam.
Mehdi Jamei received the M.Sc. and Ph.D. degrees in

civil engineering from Shahid Chamran University of SURAJ KUMAR BHAGAT received Gold Medal (First
Ahvaz, Iran, in 2005 and 2015, respectively. He worked Rank in the University) during his master’s degree in
12 years as a signor engineering and a project manager Environmental Engineering from the Rajasthan Techni-
in the famous consulting companies in Iran. He is cur- cal University, India in 2014. He also earned a research
rently an Assistant Professor with the Department of fellowship at the Indian Institute of Technology (IIT
Civil engineering, Shahd Chamran University of Ahvaz, Delhi), India in 2013. In 2014, He joined the Department
Iran. His research contributions are in the domain of of Civil Engineering, Institute of Technology, Ambo
data mining topics, focused majorly on applied soft University, Ethiopia, as a Lecturer, where, he also led a
computing, numerical method in porous media, and funded project which won the Energy Globe Award at
prediction applications in hydrology, scouring, water national level, issued by Austrian energy pioneer Wolf-
quality, nanofluids, and energy. gang Neumann. At present, he is pursuing a Ph.D. degree
in Faculty of Civil Engineering from Ton Duc Thang
University, Ho Chi Minh City, Vietnam.
Dr. Neeraj Dhanraj Bokde received the M.E. degree in
embedded systems from the EEE Department, BITS
Pilani, Pilani Campus, India, and the Ph.D. degree in data Dr. Khaled Mohamed Khedher is currently Assistant
science from the Visvesvaraya National Institute of Professor in Civil Engineering Department, in College of
Technology, Nagpur, India. He is currently working as a Engineering at King Khalid University since 2018.
Postdoctoral Researcher with the Department of Before, he worked as Assistant Professor in Civil Engi-
neering for many years at the Universities (France,
Mechanical and Production Engineering - Renewable
Tunisia, and Saudi Arabia). He has participated in many
Energy and Thermodynamics, Aarhus University, Aar-
hus, Denmark. His major research contributions are in international and national conferences and symposiums
the domain of data science topics, focused majorly on linked to Civil Engineering and new technologies (GIS,
time series analysis, software package development, Remote Sensing, and Geosciences). He has also worked
and prediction applications in renewable energy. He is on many articles and papers in ISI journals with high
serving in editorial positions in Data in Brief, Frontiers in Energy Research, Energies, impact factor. Until now, more than 60 papers online
and Information journals. published in the International publishers (Elsevier,
Taylor & Francis, MDPI, Wiley, Springer, and more. . .).
307
Nadhir Al-Ansari obtained his BSc and MSC degrees using an alternative approach for better projection of climate change, modeling
from the University of Baghdad in 1968 and 1972 economical impacts of global warming induced changes in water resources, mod-
respectively. He obtained his PhD degree from Dundee eling droughts during crop-growing seasons, projection of drought severity-area-
University in Water Resources Engineering in 1976. frequency curves for adaptation to climate change impacts on droughts and water
Now he is a Professor at the department of Civil, Envi- stress, multi-criteria decision making for selection of adaptation and mitigation
ronmental and Natural Resources Engineering at Lulea strategies, determination of unidirectional trends to distinguish global climate
Technical University Sweden. Previously worked at change from natural variability. Dr. Shahid has successfully completed about 10
Baghdad University 1976-1995 then at Al-Bayt Univer- national and international research projects as a principal investigator. He pub-
sity in Jordan (1995- 2007). Research interests are lished about 100 research articles in internationally reputed indexed journals and 5
mainly in Geology, Water Resources and Environment. academic books.
Served several academic administrative post (Dean,
Head of Department). Publications include more than
424 articles in international/national journals, chapters in books and 13 books. He Dr. Zaher Mundher Yaseen is a lecturer and researcher
executed more than 60 major research projects in Iraq, Jordan and UK. Awarded in the field of civil engineering. The scope of his research
several scientific and educational awards, among them is the British Council on its is quite abroad, covering water resources engineering,
70th Anniversary awarded him top 5 scientists in Cultural Relations. One patent on environmental engineering, knowledge-based system
Physical methods for the separation of iron oxides. Supervised more than 66 development, and the implementation of data analytic
postgraduate students at Iraq, Jordan, UK Australia universities. Member of several and artificial intelligence models. He has published over
scientific societies e.g., International Association of Hydrological Sciences, Char- 260 research articles within international journals and
tered Institution of Water and Environment Management, Network of Iraqi Scien- total number of citations over 7500 (Google Scholar H-
tists Abroad, Founder and president of the Iraqi Scientific Society for Water Index = 49). He has collaborated with over 45 interna-
Resources . . . etc. Member of the editorial board of ten international journals. tional countries and more than 430 researchers. He has
served as a reviewer for more than 135 international
journals.
DR. SHAMSUDDIN SHAHID is an Associate Professor in
the Department of Hydraulic & Hydrology, Faculty of
Civil Engineering, Universiti Teknologi Malaysia (UTM).
He is also the head of the Integrated Water Resources
Management (IWRM) research group of UTM. His major
research interests include Water Resources Manage-
ment, Climate Change Impacts and Adaptation to Water
Resources, Statistical Hydrology, Hydrological Disasters,
and Groundwater Hydrology. He uses statistical and
mathematical tools for innovative solutions of hydro-
logical problems for adaptation to global environmental
changes. His major research achievements in recent
years include forecasting water demand in a holistic way, downscaling climate
308

1 s2.0 S092523122200282X Main

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S092523122200282X Main

Uploaded by

Copyright:

Available Formats

Neurocomputing 489 (2022) 271–308

Contents lists available at ScienceDirect

Groundwater level prediction using machine learning models:

1. Introduction studied. Furthermore, Fig. 1 highlights the four major countries

2. Artificial intelligence models for GWL simulation

2.1. Application of artificial neural network models

An ANN is a computer system designed to mimic the manner in

Fig. 4. The structure of ANFIS model.

Fig. 5. Non-linear SVR vapnik’s -insensitivity loss function [123].

Zhang et al. [142] developed a two-layer LSTM model to predict

2.5. Hybrid ML models applications

Despite significant advancement in recent years, in terms of

(continued on next page)

Fig. 13. Mutation operation acts on genetic programming (GP) chromosome.

Neurocomputing 489 (2022) 271–308

AI models Case study Data span Input Output Performance metrics

2.8.6. Literature review

ACF Autocorrelation Function AE Absolute Error

Dr. Mohammad Zounemat-Kermani is a university

Salim HEDDAM is a lecture and researcher in the field

Dr. Mohammed Falah Allawi is a Lecturer and

Dr. Mustafa Al-Mukhtar is working as associate pro-

Mehdi Jamei received the M.Sc. and Ph.D. degrees in

You might also like