Using hydro-informatics methods for Ground Water Simulation

Groundwater plays an important role in semi-arid region especially for water requirements in
agriculture, municipal and industrial uses [1], and on the other hand for exploration,
prediction and remediation of it [2]. Groundwater modelling helps to analyze many important
groundwater problems. many hydrologists, geologists, engineers and other researchers pay
special attention to numerical-based groundwater model. Groundwater modelling have been
performed by many computer programs. As a model is a simplified version of a real-world
system, the first step in building a model is to build a conceptual model. After that, by help of
boundary and primary conditions the conceptual models are translated into mathematical
models in terms of flow governing equations. A problem can be solved by translating it into
numerical models and writing computer programs (codes) [3].
Prediction of Groundwater Level (GWL) has an crucial role in groundwater resources
management [1], while it is very complex due to its nonlinear nature as depending on many
factors like: precipitation, evapotranspiration, soil characteristics and topography of the
watershed [2]. Due to involving a large number of variables like: clogging, design,
optimization, feasibility, water quality, geotechnical processes, groundwater management,
recovery efficiency, saltwater intrusion, and residence time modelling groundwater has many
challenges [4].
There are three main categories models based on their physical characteristics: black box
models, conceptual models, and physical- based models. The two last ones are the main in
predicting hydrological variable and understanding the system's physical processes [1].
In spite of developing in Artificial Intelligence application of hybrid simulation models by
optimizing technique is very sparse however demand for water specially groundwater
increase annually [2].
Recently development of Geographical Information System (GIS) and Remote Sensing (RS)
technology and the availability of spatially distributed data on climate, geomorphology and
geology make distributed watershed models useable that can be calibrated with available
stream flow data. The groundwater development and management can be revolutionized by
RS technology in the future due to providing unique and completely new hydrological and
hydrogeological data [5].
Every method has its own strengths and weaknesses, but by combining them they become
much stronger. So, by combining the advantages, limitations, and economy of each method,
the proper solution to the recharge estimation in different climatic environments can be
obtained [5].
Methods:
1. Time Series and Marcov Chain Methods
Markov is a tool which can describe the distribution of hydrostratigraphic units of
groundwater simulation and is an approach which model the spatial variability of geologic
formations, transition probability. This model (MTPG) introduced by Carle and Fogg1997
and while it is conceptually simple and easy to implement, but still its conditional simulation
has limitations [6].
Makrov chain receive much attention these years because of following reasons: (a) its
transition probability approaches are characterized by conceptual and mathematical (b) users
can explain complex cross-correlations and account for asymmetric juxtaposition tendencies
in geologic strata (c) allow users to incorporate subjective semi-quantitative geologic
interpretations based on observable attributes of the aquifer and finally (d) it can generate 3D
stochastic realizations of real-world aquifers and the conditional simulation of 3D lithologies
for groundwater flow and transport simulations [6].
Some factors like anthropologic effects makes it so hard to determine the hydrological
parameters, as a result it is preferred to apply time series model in these cases [7]. A time
series model is an empirical model for stochastic modeling and predicting temporal behavior
of hydrologic system. For medium- range forecasting and synthetic data generation the
stochastic time series models are popular tools. Several stochastic time series models for
these purposes are: the Markov, Box- Jenkins (BJ) Seasonal Autoregressive Integrated
Moving Average (SARIMA), deseasonalized Autoregressive Moving Average (ARMA),
Periodic Autoregressive (PAR), Transfer Function Noise (TFN) and Periodic Transfer
Function Noise (PTFN). The Markov, ARMA and SARIMA models are considered as
univariate models and the PAR, TFN and PTFN models are considered as multivariate
models, and the PAR and PTFN models are also periodic multivariate models [7].
The number of series which are being modelled, required accuracy, modelling costs, ease of
model usage and results interpretation, are some factors which cause the selection of special
modeling method. Time series modeling includes three common steps: identifying the
experimental model which needs at least 50 observations of the regarded series, estimating
the model parameters (fitting), and verification of the model. In the first step of modeling a
time series it must be drown as a time series diagram to make it easy for identification of
trends, variance non-stationarity, seasonality (periodicity) and other types of irregularities in
data. Box- Cox transformation and non-stationarity in mean or variance are considered as a
pretreatment step for normalizing data. These tests are divided in two main groups:
independence tests (for time), and normal distribution tests [7].

2. -Geostatistics Methods
Spatial mapping and interpolating of groundwater qualitative parameters can be done by
geostatistical methods as appropriate techniques. One of these techniques is simple kriging
which is a weighted linear combination that mean value of definitely characterized by the
second-order stability and used in the estimation process. This method needs variograms with
the threshold [8]. Kriging method is the best method among unbiased method for estimation
of the value of regionalized variables in unsampled location. This method is classified into 3
groups: simple kriging (SK), ordinary kriging (OK), and universal kriging (UK) [9].
Another method is the Co-Kriging estimator as a developed kriging process by secondary
variables. For improving the estimation convenience as the state of only one secondary
variable is used [8]. For improving the transmissivity data that are insufficient for the
database of the groundwater moel for flow simulation cokriging is a good approach [10].
Inverse distance weighting (IDW) method is another interpolating method which is related to
the weightings are solely a function of the distance between the point of interest and the
sampling points for i= 1, 2, ..., n [9].
The radial basis function (RBF) method is based on neural networks. which has a feed-
forward architecture and consists of three layers: one input layer, one hidden layer, and one
output layer, with a number of neurons in each. This network has self-organizing
characteristics that allows to adapt determination of the hidden neurons while the network is
under training. Each input neuron is totally connected to all hidden neurons, and hidden
neurons and output neurons are also interconnected to each other by a set of weights. For
feeding the information into the network input neurons are utilized and transmitted to the
hidden neurons; each hidden neuron then transforms the input signal using a transfer function
f. The output of hidden neurons has the form of a radial basis function [9].
The global polynomial interpolation (GPI) method is a trend surface analysis which has been
introduced into the earth sciences to analyze environments of sedimentation, contour-type
map and elevation. In trend surface analysis, a two-dimensional polynomial equation of the
first, second or a higher degree described the distribution of observational data [11].
The local polynomial interpolation (LPI) method introduce the concept of distance weight,
this method combined the advantage of GPI method (reflect of tendency variation) and the
advantage of IDW method (reflect local characters). Generally, LPI fits different
polynomials, to specify overlapping surface into which the study area has been subdivided.
For defined region the specified order of the polynomial is fitted. When the regions overlap,
the value of each prediction is the value of the fitted polynomial at the center of the region
[11].
For assessing the best method in each interpolation cross-validation and orthogonal-
validation methods are used. By these methods estimated values are tested with the existing
samples at the location. When the estimate is calculating it can be compared to the true
sample value using statistics. For this aim the error between the true value and estimated
value is calculated and by compering root mean square error (RMSE) and correlation
coefficient (R2) the best method is selected [9, 11].
In a study eight spatial interpolation methods were compared in simulating spatial
distribution of groundwater level based on GIS in China. The results indicated that among all
these eight method the ordinary kriging method is the optimal method for groundwater level
interpolation based on cross- validation and orthogonal-validation [11].

3. -GIS and Remote Sensing

water resource mapping could be done by Remote sensing. As these techniques can make
continuous and up-to-date measurements, users have been widely applied in different fields,
such as hydrology, agriculture and climate studies. One of the best known versatile tools,
especially in spatial analysis, modeling, visualization, data processing and management is
GIS. ArcGIS products software, are required to process the original remote sensing images
and videos[12]
Hydrologic models can be done by remote sensing and GIS and in this way it is possible to
provide a full spectrum of modeling what happened in the past and project what will happen
in the future and played increasingly important roles in the hydrologic community [12]
water resources, quantitatively measure the hydrologic flux, and monitor the working
conditions of hydraulic infrastructures, drought conditions and flooding inundation can be
done by remote sensing. better management of water resources, drought and flooding
disasters can be done by GIS, statistics and numerical models. Remote sensing has this ability
to prepare critical data for mapping water resources (snow and glaciers, water bodies, soil
moisture and groundwater), measuring hydrological fluxes (ET, precipitation and river
discharge), and monitoring drought and flooding duration; while GIS is the best tools for
water resource, drought and flood risk management and for hydrologic models’ setup, input
data processing, output analysis and visualization [12].
GIS and RS as an integrated approach in groundwater management have this ability to
support many distributed hydrological models and spatial variation in recharge [13].
ArcGIS also is able to link with other models like WetSpass model which computes and
generates flow hydrographs time series [13].

4. -Cluster Analysis
For understanding the natural processes, analyzing the monitoring data of natural objects is
essential. In the process of analysis, identifying the similar objects is required, observing
which you can better understand the laws by which changes in these objects occur.
Classification is one of the fundamental processes in science. where there is a need to classify
many objects according to several factors. For these multidimensional classifications, cluster
analysis methods are used. Clustering is considered as a procedure that, starting to work with
a special data type, converts them into cluster data. The most popular are hierarchical
agglomerative methods and iterative grouping methods. In the case of using the methods of
cluster analysis, giving unambiguous recommendations on the preference for using certain
methods is difficult. Cluster analysis in conditions of the need for a multivariate analysis of
the studied groundwater indicators and low data structure, is an effective method for
identifying similar features [14].

5. -Soft-Computing Methods
Artificial Neural Network (ANN) is used as a forecasting tool specially in water resource and
hydrology [1]. Artificial neural networks learning processes take place in biological systems.
They are composed of many artificial neurons that according to a specific network
architecture, are linked together. A neural network used for predicting future values of
possibly noisy multivariate time-series based on past histories. Transforming the inputs into
meaningful outputs is the aim of the neural network [15]. Typically, ANN uses input, hidden
and output three layers and 'nodes' are artificial neurons which each layer is composed to.
The model is assisted in simulations by the connection between neurons. Where each neuron
in a layer is connected to all the neurons of the next layer is called as feed forward
architecture. Each neuron multiplies every input through its interconnection weight, sums the
product, and then passes the sum by a transfer function to produce its result [16]. ANN's best
structure can be obtained by a trial-and-error approach. The correlation coefficient (R), root
mean square error (RMSE), standard error of the estimate (SEE), coefficient of determination
(R2) and mean absolute error (MAE) are parameters that are able to evaluate the
performance of artificial intelligence [17].
The autoregressive integrated moving average (ARIMA) method is used to identify complex
patterns in data and generate forecasts. This model has the ability to analyze and forecast
univariate time series data. The ARIMA model function is represented by (p, d, q), in which p
representing the number of autoregressive terms, d the number of non-seasonal differences,
and q the number of lagged forecast errors in the prediction equation. ARIMA models have
three steps: identification,
estimation and forecasting and defined as follows:
∆ 1 Z t=φ1 Z t−1 +…+ φ p Z t− p + at −at −1 θ1−…−at−q θq

where ∆ 1 Z tis a differenced series (i.e., zt -zt - 1), zt is the set of possible observations on the
time-sequenced random variable, at is the random shock term at time t, φ 1 . . . φ p are the
autoregressive parameters of order p and θ1 . . . θq are the moving average parameters of
order q [15].
SVM algorithm is a machine learning theory which does not have a pre-determined structure,
while the training samples are judged by their contributions. Only samples that are selected
contributed to the final model, which called “support vectors”. This model is known as a high
dimensional quadratic programming problem. So for avoiding “dimensional disaster”,
therefore high dimensional computing is converted into low dimensional computing by a
kernel function. Generic kernel functions include linear, radial basis function (RBF),
Gaussian, polynomial, and other kernel functions. Among these functions, the RBF kernel
has better performance in comparison with the linear kernel in the case of dealing with high
dimensional complex samples; and due to its simple function it is more applicable than
Gaussian and polynomial kernel functions [18].
the Analytical Hyporheic Flux model (AHF) allows estimating the seepage at river banks
through a riverbed using a simplified geometry and its numerical model was done in
MODFLOW. Therefore, the model requires extensive manipulation to create the packages,
by representing the river package of MODFLOW a test phase through the seepage package is
done. For developing the implementation FloPy tool is used due to the versatility of
manipulating the packages of MODFLOW through coding. A region where surface and
groundwater mix within the bed and banks of a river is the hyporheic zone as a base of AHF
model, which could be in centimeters to meters. Estimation of the flux through this zone is
complex and it simulated by numerical groundwater flow models [4]. The errors of AHF
model are significantly lower than the often-used model, which are based on vertical water
seepage through the streambed described by Darcy’s law. The AHF model is considered as a
suitable model due to the simple set of data needed to solve the problem and simple
implementation in any computing environment The AHF model geometry is its limitation: a
rectangular-shaped riverbed cross-section which is followed by the same shape of the
sediment layer under its bottom and alongside its bank. Consequently for small and deep
rivers, neglecting the flow in the banks leads to significant errors in estimation of total flow
[19].
As neural networks learning speed were slower than what is expected new learning algorithm
which is called Extreme Learning Machine (ELM) was developed for single-hidden layer
feedforward neural networks (SLFNs) [20]. Among soft computing techniques Extreme
learning machines (ELM) has rarely used and recently gained attention than other tools, due
to its less accuracy in comparison with empirical formula and few engineers familiarity with
that. For this two problems ELM model was developed in conjunction with db2 mother
wavelet transform to enhance the accuracy in some studies. ELM is a training algorithm for
the single layer feed-forward- neural network (SLFFNN). This algorithm is based on the
SLFFNN's concept which with H hidden nodes randomly determines values of input weights
and biases of the hidden layer according to continuous probability distribution with
probability of 1, so that to be able to train N separate samples [21]. In theory, this algorithm
has extremely fast learning speed. The experimental results based on a few artificial and real
benchmark function show that the new algorithm can produce good generalization
performance in most cases and can learn thousands of times faster than conventional popular
learning algorithms [20]. In the ELM structure the input weight and the bias values are
generated randomly. ELM calculates the output weight matrix between hidden layers and
output layers analytically among inverse operation of the hidden layer output matrix. ELM
has the ability of interpolating and universal approximation which made ELM as a tool for
time series prediction. The ELM models for groundwater level prediction were done in the
MATLAB program which is able to use for both a univariate and multivariate analysis. The
performance of the developed models can be evaluated by several used statistical tests. The
coefficient of determination (R2) shows the degree of correlation between two variables. It
indicates how well the model is developed a relationship between observed and predicted
variables. Root mean square error (RMSE) shows model performance by calculating the
difference between observed and predicted values. Nash–Sutcliffe efficiency criterion (NS) is
parameter for assessing the predictive ability of hydrological models. As a normalized
measure it compares the mean square error of the model to the variance of the target output
[22].
 6. -Stochastic Models
Among many approaches toward solving Groundwater contaminant source characterization
problem the stochastic- simulation statistic (S-S) method is very popular [23]. In groundwater
management studies there are different terms for optimization modeling, for example linear
programing techniques are used for groundwater management due to their simple formulation
and application. On the other hand, some researches have done by nonlinear programing
algorithms and dynamic programing techniques. The optimization models are deterministic
but some uncertainties cause stochastic in optimization modeling like inherent variability and
fundamental lack of knowledge about aquifers parameters [24].
Monte Carlo is a stochastic optimization model which the entire system is simulated for a
large number of times in each time using a different set of random values [24]. Mantoglou
and Kourakos (2007) developed this modeling to incorporate hydraulic conductivity
uncertainty in a multi-objective management model for optimal remediation of groundwater.
This target achieved by minimizing two objectives of the optimization model: contaminated
groundwater in the aquifer and remediation cost [25]. Although method with base of Monte
Carlo usually used in stochastic simulations, utilizing them in the real world, to large-scale
water resources management problems could be too computationally demanding [24].
7. -GSM Models
The Self Organizing Map (SOM) is a method for clustering the homogenous monitoring
piezometers in the plain by use of GWL and Universal Transverse Mercator (UTM) data. For
visualization of high- dimensional data it is an effective software tool[1].
Wavelet transforms is a tool for the analysis, de-noising and compression of signals and
images. Wavelets transform analysis, developed during the last two decades are mathematical
functions that use to analyze time series that contain non-stationarities. Their ability to
simultaneously obtain information on the time, location and frequency of a signal, their main
advantage [15]. Wavelet analysis is a tool to analyze localized variations of power by help of
time series. The wavelet transform has been used for numerous studies like: geophysics,
including tropical convection, atmospheric cold fronts, the dispersion of ocean waves, wave
growth and breaking, and coherent structures in turbulent flows, and etc. [26]. Wavelet as a
transformation tool is utilized for decomposition, compression and de-noising. This method is
a time independent spectral analysis for describing the time-scale, processes and their
relations. The Wavelet is able to transform considers time series as a linear combination of
multiple base functions. The Wavelet transform has this ability to achieve time, frequency
and situation data simultaneously [21]. Because this method has multi-resolution strength and
localization capability in time and frequency domains it is an important tool for spectral
analysis area [16].
There are some methods of estimating groundwater recharge around the world which divided
into five group by United States Geological Survey: 1) groundwater method (Groundwater
Modeling and Water Table Fluctuation method (WTF)), 2) streamflow methods (Seepage
Meters, Stream-flow Gain/ Loss Measurements), Recession-Curve Displacement Method and
Watershed Models), 3) Tracer methods (Chloride, Chlorofluorocarbons, Temperature and
Tritium), 4) Unsaturated zone methods (Darcian Method, Zero-Tension Lysimeters and Zero-
Flux Plane), and (5) water budget methods Deep Percolation Model and Hydrologic
Evaluation of Landfill Performance (HELP3) Model) [5].

The structure of ELM [27]

Models:
MODFLOW packages with independent modules in steady and transient state conditions
calibrate processes either manual or automated. This model is used by many researchers to
understand the groundwater dynamics to fix the optimal pumping limits from the aquifer [2].
Among its development many packages have been created to satisfy and model some
scenarios. The main challenge of modelling would be the high numerical instability, the high
costs of modelling, and the computational efficiency [4].
This model solves the 3-D groundwater flow equation numerically by finite-difference
method in porous media:

∂
∂x (
K xx
∂h
) (
+
∂x ∂y
∂
K yy
∂h
+
∂
∂ y ∂x) (
K zz
∂h
∂z )
−W = S s
∂h
∂t( )
where x, y, z are Cartesian coordinate axes, h = potentiometric head [L], Kxx, Kyy, Kzz =
hydraulic conductivities along x, y, and z axes [LT−1], W = volumetric flux/unit volume and
represents sources and/or sinks of water [T−1], SS = specific storage of the porous material
[L−1] and t = time. All capabilities of MODFLOW is available in McDonald and Harbaugh
1988 [28]. This model has some assumption which many aquifer layer systems are met them:
the flow is in saturated condition, (2) Darcy’s law applies, (3) the concentration of
groundwater is constant, and (4) the principal direction of hydraulic conductivity or
transmissivity does not change within the aquifer system [3].
The SWAT (Soil and Water Assessment Tool) watershed model with 3700 published studies
has been the most used watershed model in the world since 2000 [29]. To help water
resource manager to assess the impact of management on water supplies and nonpoint source
pollution in watersheds and large river basins this model was developed. This model is able
to predict the impact of management on water, sediment and agricultural chemical yields and
it also operates on a daily time step [30]. Within SWAT, the water balance equation is solved
for the shallow aquifer. This model also has an updated version called SWAT+ which include
MODFLOW for simulating interactions of groundwater flow and groundwater surface water.
SWAT+ gives recharge and stream stage to MODFLOW, and then spatial values of
groundwater head, groundwater flow rate, and exchange rates of groundwater and surface
water are provided by MODFLOW that are provided to SWAT+ stream channels for stream
routing. SWAT+(MODFLOW) can simulate all hydrologic pathways of watershed, for
mapping of water table elevation, water table depth, and stream-aquifer exchange rates for all
grid cells in the model domain. This model is an excellent tool for investigating water supply
and the impact of conservation practices and climate change in watersheds [29].
GMS as a three-dimensional groundwater simulation software is used for simulating ground-
water and solute transport problems in groundwater. This software is able to model the aq-
uifer in both steady and unsteady conditions in the 2D and 3D modes by imposing changes in
atmospheric parameters and conditions in the area for better management of groundwater in
the area. Compared to other similar software such as MODFLOW, PMWIN, the GMS
modules are better than its counterparts because of its better user efficiency and 3D
visualization [31].

Reviews
In Nourani et al. (2012) study's SOM, FFNN approaches were combined to develop a hybrid
black box model for groundwater level simulation. Their results showed that according to
rainfall uncertainty and Fuzzy concept ability, the conjuction of ANN and FIS (Fuzzy
Inference System) model as an ANFIS could be a suitable model for the model progress [1].
In a study wavelet transform (WA) and artificial neural networks (ANN) used for
groundwater level prediction and their performance were compared to ANN and
autoregressive integrated moving average (ARIMA) models for monthly groundwater level
prediction. It was revealed that the WA–ANN models are more accurate for monthly average
groundwater level forecasting in comparison to the ANN and ARIMA models [15].
In a study in order to assess the groundwater quality three optimization meta-heuristic
algorithms: imperialist competitive (ICA), election(EA), and grey wolf (GWO) were used. 10
years data were used and results indicate that GWO method with support vector regression
method (SVR) had higher accuracy than two other methods and this method also had higher
accuracy in simulation and estimation of groundwater quality [32].
support vector regression model(SVR) is used to estimate heavy metal in groundwater by two
input data HCO3, SO4, and Pb, Zn, and Cu as output parameter. The result showed that this
model is a reliable system for this propose [33].
In Africa estimation of groundwater recharge rate was done by method using groundwater
level data, the streamflow method, and the water balance methods. The results indicate that
Water-Table Fluctuations (WTF), Recession-Curve Displacement, and Chloride Methods can
be used by certainty to improve estimation of groundwater recharge. On the other hand,
recently developed water balance methods combined with GIS technology could be a
powerful tool for estimating groundwater recharge, when spatial temporal variability of
components in water balance is taken into account [5].
According to Diaz et al (2020) Neumann boundary condition is used for the implementation
of the AHF model through the recharge package. Then, for representing it as a Cauchy
boundary condition type for a river, an iterative algorithm that allows defining the AHF
fluxes through seepage in MODFLOW is used. So by help of AHF model implementing the
lateral contribution on the banks of a river, which is applied in a real case using a
groundwater modelling tool in three dimensions as MODFLOW was done. In their study the
groundwater model is performed in several steps of: fieldwork, information gathering, set up
of conceptual and numerical models, and calibration. The Biebrza River between Rogoz˙yn
and Rogoz˙ynek villages is a study area which the model is constructed for. At last their
methodology allows the fluxes that come from the bottom and the banks of the river for
artificial channel instead of rivers with complex natural geometry [4].

In 470 km2 Middle Bosque River Watershed (MBRW) in central Texas, USA, the SWAT+
(MODFLOW) model was used. The model was tested to measure: annual average hydrologic
fluxes trends between annual rainfall rates and hydrologic fluxes, stream discharge at the
watershed outlet, groundwater head and water table fluctuations, and groundwater-surface
water interactions. The results of simulation were fitted well with observed values [29].
Three model including MODFLOW, Extreme Learning Machin (ELM) and Wavelet Extreme
Learning Machin (WA-ELM) were utilized in a study to simulate groundwater level. In this
way that they defined ten different models for ELM and WA-ELM models separately and
examined the results of them to choose the best soft computing model. Also the best artificial
intelligence model was compared with MODFLOW and showed that WA-ELM model had
higher accuracy in simulating groundwater level doe to comparison of MAE and RMSE [21].
Hydrological and meteorological parameters like: rainfall, temperature, evapotranspiration
and groundwater level monthly data were used to forecast groundwater level in two wells by
ELM and SVM. The results showed that ELM model had better ability for simulating
monthly groundwater level [22].
In a study which was done in Maragheh-Bonab plain in Iran 367 months average
groundwater data were used to compare the performance of different hybrid wavelet-neural
network. About 85% of total data set were utilized for training and 15% for testing the
models. The results showed that according to R2, RMSE and NSC the ELM had better
performance than other model (GMDH) [27].

In a study WA-ANN and WA-SVR were developed to optimize the ANN and SVR models.
For investigating the feasibility of ANN, SVR, WA-ANN, and WA-SVR models in
simulating of 1, 2 and 3-month-ahead groundwater depth in 3 wells. The performances of
ANN, SVR, WA-ANN and WA-SVR models for 1-month-ahead forecasting were better than
2 and 3-months ahead forecasts, because as the lead time get longer the accuracy of simulated
groundwater depth become worse. Also it was observed that WA-SVR had better results than
WA-ANN model for 1, 2 and 3-month lead times and this model has the ability to forecast
groundwater under ecological water conveyance conditions [16].
The coupled wavelet Support Vector Machine (WSVM) model was tested in a study. partial
autocorrelation function of groundwater depth time series was introduced as input variables
lag times. For preprocessing and decomposing the original groundwater level time series to
four subseries by different frequencies a three level discrete wavelet transform (DWT) was
utilized. By the same historic data, the WSVM model was compared with ANN, SVM and
WANN and the results showed that according to the RAE, r coefficient, RMSE and NSE,
WSVM had more accurate results. The groundwater depth series was decomposed into four
subseries with better stationary for
model training by using three level DWT, so that the extraction of mainstream components
was improved the prediction performance. Also it was indicated that WSVM model had a
good prediction on monthly groundwater depth [18].
Bayat et al.,2020, used Groundwater Modelling System (GMS software) in 3D in Gavkhnoi
Basin, Isfahan Province, Iran to investigate decreases of ground water and potential
environmental hazards. In this study 86 months of steady and unsteady state was utilized for
calibrating and validating the model. Results showed that by use of a calibrated model,
reduction of almost 30% in well water harvesting leads to an increase in mean groundwater
level by 37 cm and an increase of 10% in well water harvesting; the mean groundwater level
decreased about 12 cm. Surface water results for this period showed that water levels in most
parts of the aquifer follow a uniform decreasing pattern, indicating if the wells are discharged
and recharged as in the previous years, most observational wells face a drastic drop in water
level. According to the scenarios, among the changes through artificial recharge, precipitation
and well water harvesting, precipitation has the most effective procedure in aquifer water
management [31].

In a study which had done by authors water- table according to different scenario was
simulated by Artificial Neural Network and ArcGIS. By applying artificial neural network,
water-table data and cropping pattern in plain, three land-use scenarios were designed: 1)
existing crop pater in study area, 2) increase fallowing every year, and 3) planting low-water
use crops. Water table was simulated for water year of 2008 with one-layer network,
Levenberg-Marquart algorithm, and three functions in MATLAB_R2012a. Water table map
was prepared by using simulated water table in ArcGIS 10.2, the model divided data to 3
groups one for training, another for variation of simulation, and other for testing so according
to covariance and RMSE the best model was selected. and zoning was performed according
to the costs of water pumping. Results showed that 61-86 percent of the plain in all three
scenarios had medium limitation. Also, using different management in field like fallowing
and planting low-water use crops caused 3 and 5 percent increase in acreage of "without
limitation" lands, respectively. Due to the results it seems that ANN is a strong tool for
simulating water table even in case of lack of data and in managing crop planning in different
region.

Uncategorized References
1. Nourani, V., et al., Classification of groundwater level data using SOM to develop ANN-
based forecasting model. 2012. 2(1): p. 2231-07.
2. Panda, P. and M.J.I.R. Narasimham, A Review on Modelling and Simulation of Ground Water
Resources in Urban Regions. 2020. 9: p. 235-244.
3. Chakraborty, S., et al., Investigation, simulation, identification and prediction of groundwater
levels in coastal areas of Purba Midnapur, India, using MODFLOW. 2020. 22(4): p. 3805-
3837.
4. Diaz, M., G. Sinicyn, and M.J.W. Grodzka-Łukaszewska, Modelling of Groundwater–
Surface Water Interaction Applying the Hyporheic Flux Model. 2020. 12(12): p. 3303.
5. Chung, I.-M., et al., Estimating groundwater recharge in the humid and semi-arid African
regions. 2016. 20(5): p. 731-744.
6. Langousis, A., et al., Markov based transition probability geostatistics in groundwater
applications: assumptions and limitations. Stochastic Environmental Research and Risk
Assessment, 2018. 32(7): p. 2129-2146.
7. Khorasani, M., et al., Simulation and analysis of temporal changes of groundwater depth
using time series modeling. Modeling Earth Systems and Environment, 2016. 2(2): p. 1-10.
8. Maroufpoor, S., A. Fakheri-Fard, and J. Shiri, Study of the spatial distribution of
groundwater quality using soft computing and geostatistical models. ISH Journal of
Hydraulic Engineering, 2019. 25(2): p. 232-238.
9. Sun, Y., et al., Comparison of interpolation methods for depth to groundwater and its
temporal and spatial variations in the Minqin oasis of northwest China. Environmental
Modelling & Software, 2009. 24(10): p. 1163-1170.
10. Drias, T., et al., Groundwater modelling of the Tebessa-Morsott alluvial aquifer
(northeastern Algeria): A geostatistical approach. Groundwater for Sustainable
Development, 2020. 11: p. 100444.
11. Yao, L., et al., Evaluation of spatial interpolation methods for groundwater level in an arid
inland oasis, northwest China. Environmental earth sciences, 2014. 71(4): p. 1911-1924.
12. Wang, X. and H. Xie, A review on applications of remote sensing and geographic
information systems (GIS) in water resources and flood risk management. Water, 2018. 10(5):
p. 608.
13. Rwanga, S. and J. Ndambuki, Approach to quantify groundwater recharge using gis based
water balance model: a review. Int J Adv Agric Environ Eng (IJAAEE), 2017. 4(1).
14. Kozhevnikova, T., I. Manzhula, and L. Kondratieva. Simulation in the tasks of environmental
monitoring of groundwater. in IOP Conference Series: Earth and Environmental Science.
2020. IOP Publishing.
15. Adamowski, J. and H.F.J.J.o.H. Chan, A wavelet neural network conjunction model for
groundwater level forecasting. 2011. 407(1-4): p. 28-40.
16. Yu, H., et al., Comparative study of hybrid-wavelet artificial intelligence models for monthly
groundwater depth forecasting in extreme arid regions, Northwest China. 2018. 32(1): p.
301-323.
17. Jeihouni, E., Eslamian, S., Mohammadi, M., & Zareian, M. J, Simulation of groundwater
level fluctuations in response to main climate parameters using a wavelet–ANN hybrid
technique for the Shabestar Plain, Iran. Environmental Earth Sciences, 2019. 78(10): p. 1-9.
18. Zhou, T., F. Wang, and Z.J.W. Yang, Comparative analysis of ANN and SVM models
combined with wavelet preprocess for groundwater depth prediction. 2017. 9(10): p. 781.
19. Nawalany, M., et al., Groundwater–Surface Water Interaction—Analytical Approach. 2020.
12(6): p. 1792.
20. Huang, G.-B., Q.-Y. Zhu, and C.-K.J.N. Siew, Extreme learning machine: theory and
applications. 2006. 70(1-3): p. 489-501.
21. Malekzadeh, M., S. Kardar, and S.J.G.f.S.D. Shabanlou, Simulation of groundwater level
using MODFLOW, extreme learning machine and Wavelet-Extreme Learning Machine
models. 2019. 9: p. 100279.
22. Yadav, B., et al., Assessing the suitability of extreme learning machines (ELM) for
groundwater level prediction. 2017(32): p. 103--112.
23. Wang, H., W. Lu, and J. Li, Groundwater contaminant source characterization with
simulation model parameter estimation utilizing a heuristic search strategy based on the
stochastic-simulation statistic method. Journal of Contaminant Hydrology, 2020. 234: p.
103681.
24. Joodavi, A., et al., Groundwater management under uncertainty using a stochastic multi-cell
model. Journal of hydrology, 2017. 551: p. 265-277.
25. Mantoglou, A. and G. Kourakos, Optimal groundwater remediation under uncertainty using
multi-objective optimization. Water resources management, 2007. 21(5): p. 835-847.
26. Torrence, C. and G.P.J.B.o.t.A.M.s. Compo, A practical guide to wavelet analysis. 1998.
79(1): p. 61-78.
27. Barzegar, R., et al., Forecasting of groundwater level fluctuations using ensemble hybrid
multi-wavelet neural network-based models. 2017. 599: p. 20-31.
28. Shukla, P. and R.M. Singh, Groundwater system modelling and sensitivity of groundwater
level prediction in Indo-Gangetic Alluvial Plains, in Groundwater. 2018, Springer. p. 55-66.
29. Bailey, R.T., et al., Enhancing SWAT+ simulation of groundwater flow and groundwater-
surface water interactions using MODFLOW routines. 2020. 126: p. 104660.
30. Arnold, J.G., et al., Large area hydrologic modeling and assessment part I: model
development 1. 1998. 34(1): p. 73-89.
31. Bayat, M., et al., Groundwater Level Prediction through GMS Software–Case Study of
Karvan Area, Iran. 2020. 39(3): p. 139–145-139–145.
32. Emami, S., et al., Modeling groundwater quality using three novel hybrid support vector
regression models. 2021. 6(1).
33. Ghadimi, F.J.J.o.T.V., Machine Learning Algorithm for Prediction of Heavy Metal
Contamination in the Groundwater in the Arak Urban Area. 5(2): p. 115-127.