Science of the Total Environment 785 (2021) 147319

Contents lists available at ScienceDirect

Science of the Total Environment

journal homepage: www.elsevier.com/locate/scitotenv

Efficacy of machine learning techniques in predicting groundwater

fluctuations in agro-ecological zones of India
Janaki B. Mohapatra a, Piyush Jha b, Madan K. Jha a,⁎, Sabinaya Biswal a
a
AgFE Department, Indian Institute of Technology Kharagpur, Kharagpur 721 302, India
b
University of Waterloo, 200 University Avenue West, Waterloo, Ontario N2L 3G1, Canada

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• Scientific framework is developed to

predict groundwater levels at large
scales.
• Efficacy of the ANFIS, DNN and SVM
models is assessed in predicting ground-
water levels.
• The models predict groundwater levels
in each Agro-Ecological Zone (AEZ) of
India.
• The DNN model's prediction ability is su-
perior for most AEZs followed by the
ANFIS model.
• The study calls for improved ground-
water-monitoring and data acquisition
across India.

a r t i c l e i n f o a b s t r a c t

Article history: In the 21st century, groundwater depletion is posing a serious threat to humanity throughout the world, partic-
Received 22 January 2021 ularly in developing nations. India being the largest consumer of groundwater in the world, dwindling ground-
Received in revised form 17 April 2021 water storage has emerged as a serious concern in recent years. Consequently, the judicious and efficient
Accepted 19 April 2021
management of vital groundwater resources is one of the grand challenges in India. Groundwater modeling is
Available online 24 April 2021
a promising tool to develop sustainable management strategies for the efficient utilization of this treasured re-
Editor: Christian Herrera source. This study demonstrates a pragmatic framework for predicting seasonal groundwater levels at a large
scale using real-world data. Three relatively powerful Machine Learning (ML) techniques viz., ANFIS (Adaptive
Neuro-Fuzzy Inference System), Deep Neural Network (DNN) and Support Vector Machine (SVM) were
Keywords: employed for predicting seasonal groundwater levels at the country scale using in situ groundwater-level and
Groundwater-level prediction pertinent meteorological data of 1996–2016. ANFIS, DNN and SVM models were developed for 18
Machine learning/artificial intelligence Agro-Ecological Zones (AEZs) of India and their efficacy was evaluated using suitable statistical and graphical in-
techniques dicators. The findings of this study revealed that the DNN model is the most proficient in predicting seasonal
Data-driven modeling
groundwater levels in most AEZs, followed by the ANFIS model. However, the prediction ability of the three
ANFIS
models is ‘moderate’ to ‘very poor’ in 3 AEZs [‘Western Plain and Kutch Peninsula’ in Western India, and ‘Deccan
DNN
SVM Plateau (Arid)’ and ‘Eastern Ghats and Deccan Plateau’ in Southern India]. It is recommended that groundwater-
Agro-Ecological Zone monitoring network and data acquisition systems be strengthened in India in order to ensure efficient use of
modeling techniques for the sustainable management of groundwater resources.
© 2021 Elsevier B.V. All rights reserved.

⁎ Corresponding author.
E-mail address: [email protected] (M.K. Jha).

https://doi.org/10.1016/j.scitotenv.2021.147319
0048-9697/© 2021 Elsevier B.V. All rights reserved.
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

1. Introduction et al., 2012; Shirmohammadi et al., 2013; Shiri et al., 2013; Nourani
et al., 2014; Jha and Sahoo, 2015; Nourani et al., 2015; Nourani and
The imprudent use and continued mismanagement of freshwater as Mousavi, 2016; Sun et al., 2016; Barzegar et al., 2017; Sahoo et al.,
well as rapid urbanization and industrialization due to ever-increasing 2017; Alizamir et al., 2018; Yu et al., 2018; Nadiri et al., 2019; Rajaee
anthropogenic activities have caused detrimental effects on the water et al., 2019; Roshni et al., 2019; Roshni et al., 2020). The data-driven
cycle, which have led to water scarcity, increasing pollution of freshwa- models are based on relationships between observed inputs and out-
ter and degraded ecosystems across the globe (e.g., UN Water, 2007; puts and do not require system parameters and any information about
Vörösmarty et al., 2010). Additionally, the impacts of global warming the underlying processes occurring in a system (ASCE, 2000b). This is
and climate change on the hydrosphere and biosphere are already the reason that the data-driven models fall in the category of “Empirical
being experienced and they threaten the sustainability of water re- Models” and offer a “Black-Box” approach for prediction or forecasting.
sources and ecosystems (e.g., Döll and Zhang, 2010; Kløve et al., 2014; Thus, the data-driven models are useful modeling tools even for natural
Liu et al., 2020). Groundwater is a fundamental component of the systems when the primary goal of research is not to investigate or con-
water cycle that permeates life on the earth. It plays a pivotal role in pro- ceptualize processes involved in hydrologic and hydrogeologic systems.
viding a reliable source of water supply for human consumption (sup- In the digital age of 21st century, many soft-commuting techniques
plying nearly 50% of the global drinking water in the world) even have emerged in the recent past under a new umbrella of ‘Machine
during drought and emergency situations as well as it sustains irrigated Learning (ML)’ or ‘Artificial Intelligence (AI)’ due to significant advance-
agriculture, groundwater-dependent ecosystems, and socio-economic ments in computing technologies (both hardware and software). These
development of a country (IAH, 2008). days, ML/AI techniques are attracting the attention of scientific commu-
The irrigated land in India is about 30% of the global irrigated land nities of various fields as well as policymakers and managers of different
(AQUASTAT, 2010) and groundwater is the backbone of India's agricul- sectors at an accelerated pace for finding out practical solutions to
ture (accounting for >60% of irrigation water demand) and drinking real-world problems faced by humanity. They are expected to offer
water security in urban and rural areas (World Bank, 2010). It has signif- more avenues in the future for the data-driven modeling of water re-
icantly contributed in increasing agricultural production and productiv- sources systems, which in turn can substantially improve the planning
ity, and has played a crucial role in achieving food security in India and management of water and related resources (Savic, 2019; Keck
(Sharma, 2009). However, India ranks first in the world by abstracting and Lee, 2021).
251 km3 of groundwater per year, which is greater than a quarter of The review of past literature revealed that the application of poten-
the total global groundwater withdrawals (WWAP, 2012; Margat and tial ML/AI techniques for simulating/predicting seasonal groundwater
Van der Gun, 2013). The abstraction of groundwater in India has in- fluctuations at a large scale using in situ (real-world) data is very limited
creased many folds during the post-green revolution period, which in the field of Subsurface Hydrology. In Subsurface Hydrology, the
has resulted in overexploitation of groundwater in states like Punjab, widely used ML/AI models, i.e., different types of ANN models have
Haryana, Rajasthan, Tamil Nadu, Gujarat, Odisha, West Bengal, etc. been mostly applied for predicting groundwater levels at limited num-
(CGWB, 2006) and this trend is gradually expanding in other parts of ber of sites in an area using in situ data (e.g., Coulibaly et al., 2001;
India. Thus, India's water security and food security are under a serious Coppola et al., 2005; Nayak et al., 2006; Nourani et al., 2008; Mohanty
threat, and the lives and livelihoods of millions are at risk, which high- et al., 2010; Adamowski and Chan, 2011; Taormina et al., 2012;
lights the urgent need for managing dwindling groundwater resources Shirmohammadi et al., 2013; Shiri et al., 2013; Sahoo and Jha, 2013;
efficiently and judiciously. Providing sustainable water supplies for do- Jha and Sahoo, 2015; Nourani et al., 2015; Nourani and Mousavi,
mestic and agricultural sectors under changing climate is one of the big- 2016; Sun et al., 2016; Barzegar et al., 2017; Sahoo et al., 2017;
gest challenges in the 21st century. To address this challenge, the Alizamir et al., 2018; Kouziokas et al., 2018; Yu et al., 2018; Nadiri
efficient planning and management of water resources using modern et al., 2019; Rajaee et al., 2019; Roshni et al., 2019; Roshni et al.,
scientific tools and techniques are of utmost importance. 2020). On the other hand, Mukherjee and Ramachandran (2018) evalu-
Groundwater modeling has emerged as a promising scientific tool to ated Support Vector Regression (SVR), Artificial Neural Network (ANN)
help decision-makers develop sustainable management strategies for and Linear Regression techniques by using GRACE-derived terrestrial
the efficient utilization and protection of this vital resource. The rela- water change and meteorological data as inputs for predicting ground-
tionships between hydro-meteorological variables (i.e., rainfall, water levels in five observation wells located in different geographical
temperature, river stage, pumping, evapotranspiration, etc.) and regions of India. It is also apparent from the review of literature that
groundwater levels are very complex and highly non-linear. In addition, fewer researchers have evaluated the performance of SVM models in
groundwater processes exhibit a high degree of spatial and temporal simulating/predicting groundwater levels at selected sites using in situ
variability within a basin as well as among basins, which renders input and output data (Yoon et al., 2011; Shiri et al., 2013;
groundwater modeling a challenging task for the researchers and Suryanarayana et al., 2014; Mirzavand et al., 2015; Gong et al., 2016;
water managers (Anderson and Woessner, 1992). Although process- Yoon et al., 2016; Huang et al., 2017; Guzman et al., 2019; Tang et al.,
based (or physically based) models offer an efficient tool for modeling 2019). In these studies, the performance of SVM models has been
complex natural systems, the extent of data and information required found better than different ANN models and conventional time series
for developing process-based groundwater-flow/transport models is models. As to the ANFIS application in groundwater-level prediction,
enormous; many of these data and information are usually expensive only seventeen studies are reported to date (Kholghi and Hosseini,
and/or difficult to obtain, particularly in developing countries. Under 2009; Shiri and Kisi, 2011; Jalalkamali et al., 2011; Sreekanth et al.,
such circumstances, data-driven/empirical models, which require rela- 2011; Shirmohammadi et al., 2013; Moosavi et al., 2013; Shiri et al.,
tively less data and efforts, serve as attractive alternatives for modeling 2013; Maiti and Tiwari, 2014; Moosavi et al., 2014; Emamgholizadeh
natural systems (e.g., Coulibaly et al., 2001; Coppola et al., 2005; et al., 2014; Mirzavand et al., 2015; Khaki et al., 2015; Nourani and
Daliakopoulos et al., 2005; Nayak et al., 2006; Solomatine and Ostfeld, Mousavi, 2016; Gong et al., 2016; Zhang et al., 2017; Das et al., 2017;
2008; Mohanty et al., 2010; Sahoo and Jha, 2013; Jha and Sahoo, 2015; Zare and Koch, 2018) wherein the capability of ANFIS models has
Roshni et al., 2019; Rajaee et al., 2019; Roshni et al., 2020). Conse- been found superior to ANN models and conventional time series
quently, machine learning (ML) techniques such as Artificial Neural models. Among these studies, only three studies (Shiri et al., 2013;
Network (ANN) and its different versions have been extensively used Mirzavand et al., 2015; Gong et al., 2016) have compared the
in hydrological sciences including groundwater hydrology (e.g., ASCE, performance of ANFIS and SVM models. In contrast, only two recent
2000a; Maier and Dandy, 2000; Coppola et al., 2005; Nayak et al., studies (Jeong and Park, 2019; Kumar et al., 2020) have used advanced
2006; Nourani et al., 2008; Adamowski and Chan, 2011; Taormina ML/AI models such as Deep Neural Network (DNN) for predicting

2
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

groundwater levels at one site and four sites, respectively. However, no of its kind at least in the developing world. The data-driven modeling
studies have explored the performances of ANFIS (Adaptive Neuro- framework developed in this study will be useful for predicting ground-
Fuzzy Inference System), Deep Neural Network (DNN) and Support water levels at larger scales under data-scarce conditions.
Vector Machine (SVM) models for predicting groundwater fluctuations
even at a site scale using in situ (field-measured) data. 2. Methodology
The poor availability of groundwater-level data at suitable spatial
and temporal scales, together with data discontinuity is one of the 2.1. Study area
greatest challenges in most developing nations (including India). As a
result, groundwater modeling (even data-driven modeling) at a large The study area for the present study is India (Fig. 1), which is a trop-
scale is a daunting task in several parts of the world, especially in devel- ical country located in South Asia. It lies between 6° 44′ N and 37° 30′ N
oping nations. Taking cognizance of the current constraints and review latitude and between 68° 07′ E and 97° 25′ E longitude. India is the
of literature, the major goal of this study is to demonstrate a pragmatic world's seventh largest country that encompasses a geographical area
scientific framework for predicting seasonal groundwater fluctuations of 3,287,263 km2. The average annual rainfall of the country is around
at larger scales (i.e., country scale) using real-world data. To achieve 1190 mm with appreciable spatial and temporal variability. The regions
this goal, three relatively powerful supervised ML/AI techniques viz., of the Western Ghats and the Northeast and Meghalaya Hills receive
ANFIS (Adaptive Neuro-Fuzzy Inference System), DNN (Deep Neural heavy annual rainfall greater than 2500 mm. On the contrary, the re-
Network), and SVM (Support Vector Machine) were considered for gions of Northern Kashmir and Western Rajasthan receive an annual
predicting seasonal groundwater levels using in situ groundwater- rainfall of less than 400 mm. More than 75% of the annual rainfall occurs
level and pertinent meteorological data of 1996–2016 period. In this in four months (June to September) only, and there is a significant var-
study, the groundwater-level data from 9941 monitoring wells were iation of rainfall within a year as well as among years (Purohit and Kaur,
used (based on the availability, continuity and reliability of the field 2016). The average temperature in the winter season ranges from 10 to
data) for predicting groundwater levels across India considering Agro- 15 °C in the northwest region with peaks around 20–25 °C, whereas the
Ecological Zones (AEZs) as spatial units. Agro-Ecological Zones vary sig- average temperature in the summer season is about 32–40 °C in most
nificantly in terms of climatic settings, topography, land use/land cover, parts of the country. The geological formations of the country are
soil, agricultural activities, hydrogeologic settings, hydrologic regimes, broadly divided into three classes (CGWB, 2017): (a) Unconsolidated
etc. In addition, the availability of groundwater-monitoring sites, conti- formations (Alluvial aquifers), (b) Semi-consolidated formations (Sand-
nuity of field data, duration of field data, and the quality of available field stone aquifers), and (c) Hard rock/consolidated formations (Crystalline
data vary considerably from one AEZ to another, which render the pre- aquifers, Sandstone, Limestone, Shale, Basalt and Volcanic aquifers).
processing of country-scale field data cumbersome and arduous. These Considering significant hydro-climatic, hydrogeologic, agricultural,
natural and anthropogenic factors pose challenges not only for the and ecological variability over the country, Agro-Ecological Zones
process-based modeling of groundwater flow but also for the data- (AEZs) were selected as spatial units in this study for simulating/
driven modeling of groundwater fluctuations. This study is an attempt predicting groundwater levels across the country. Based on physio-
to address these challenges by developing a pragmatic framework for graphic features, soil characteristics, bio-climatic types (rainfall, poten-
predicting seasonal groundwater levels at larger scales by using techni- tial evapotranspiration, and soil-water storage) and length of the
cally robust ML/AI techniques (ANFIS, DNN and SVM). As revealed from growing period, the National Bureau of Soil Survey and Land Use Plan-
the detailed review of past studies, the performance evaluation of ning divided entire India into 20 AEZs (Sehgal et al., 1992) which are
ANFIS, DNN and SVM models in predicting groundwater levels either depicted in Fig. 1.
at a site scale or country scale has not been reported to date. Further-
more, the efficient hyperparameter tuning technique (Bayesian optimi- 2.2. Data acquisition and pre-processing
zation), non-linear functions (Leaky ReLU and its variants), and
regularization technique (Dropout) have been used in this study for de- The ‘depth to groundwater level’ data were downloaded from the
veloping DNN models; it is a novel application in the field of Subsurface Water Resources Information System of India (WRIS) database. As of
Hydrology. To the best of authors' knowledge, the present study is first 2018, a network of about 29,553 observation wells (dug wells,

Fig. 1. (Left) Map of Study Area depicting Agro-Ecological Zones (modified from Shukla et al., 2017); (Right) locations of dug wells in the Agro-Ecological Zones of India.

3
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

tubewells, borewells, and dug-cum-bore wells) located in different the neurons (Haykin, 2004). Each of these connections has a certain
parts of the country were monitored by Central Ground Water Board weight that signifies the importance provided to a given input. The
(CGWB). The duration of groundwater-level data availability is from weight of each input is then added together along with a ‘Bias’ term.
1996 to 2016. The ‘depth to groundwater level’ (m bgl) is usually mea- Suitable ‘Activation Function’ is used to introduce non-linearity in the
sured four times in a year during 1–10 January (Post-Monsoon Rabi sea- output (Fig. 2); it helps in the generalization of an ANN model because
son), 20–30 March/April/May (Pre-Monsoon season), 20–31 August most of the real-world problems are very complex and non-linear. Dur-
(Monsoon season) and 1–10 November (Post-Monsoon Kharif season). ing the training of an ANN model, the weights and biases of every neu-
The month for taking groundwater-level observation in the Pre- ron are adjusted iteratively (using a suitable training algorithm) in such
Monsoon season is based on the onset of Monsoon rainfall in a particu- a way that the output(s) of the ANN model has a reasonably low error
lar state, i.e., the month of March for the northeastern states, April for with respect to the target value(s). Thereafter, the testing of an ANN
Kerala, Odisha and West Bengal states, and May for the remaining model is carried out to verify its accuracy for a given task.
states.
Dug wells, which usually tap swallow unconfined aquifer systems, 2.3.2. Adaptive Neuro-Fuzzy Inference System (ANFIS)
constitute about 75% of the observation wells monitored by CGWB. Artificial Neural Network (ANN) and Fuzzy Logic are two comple-
Hence, the data of dug wells were extracted for this study, out of mentary technologies. Unlike ANNs, Fuzzy Logic does not have a
which only those dug wells were considered that have at least 9 years learning algorithm. Jang (1993) introduced a novel architecture
of continuous data in at least one of the seasons. This resulted in a and learning product for Fuzzy Inference System (FIS) that employs
total of 9941 dug wells, which were divided into 20 sets based on a neural network learning algorithm for constructing a set of fuzzy
their locations in different Agro-Ecological Zones (AEZs). Owing to the “If-then” rules with appropriate membership functions from the
non-availability of dug wells in two Agro-Ecological Zones namely, stipulated input-output pairs. It is known as Adaptive Neuro-Fuzzy
‘Western Himalayas (Arid)’ and ‘Andaman & Nicobar and Lakshadweep Inference System (ANFIS), which is a class of adaptive networks
Islands’ (Fig. 1), the groundwater-level data are available for only 18 equivalent to the Sugeno or Tsukamoto fuzzy inference systems.
zones, which were used in this study; the locations of the dug wells Thus, Neuro-fuzzy hybridization results in a hybrid intelligent sys-
are shown in Fig. 1. It is noteworthy that the unconfined aquifers are tem that synergizes these two techniques by integrating the
most susceptible to climate and they sustain surface water bodies human-like reasoning style of fuzzy systems with the learning and
(e.g., lakes, rivers, wetlands, and springs). Furthermore, the meteorolog- connecting structure of ANNs (Nguyen et al., 2003). The ANFIS archi-
ical data of the 1996–2016 period were obtained from India Meteoro- tecture used in this study is the first-order Sugeno fuzzy model
logical Department (IMD), New Delhi. Like the groundwater-level consisting of five layers (Fig. 3): (a) Layer 1: Fuzzy Membership
data, the gridded rainfall (0.25° × 0.25°) and temperature (1° × 1°) Layer, (b) Layer 2: Fuzzification Layer, (c) Layer 3: Normalization
data were also divided into 20 sets based on their location in different Layer, (d) Layer 4: Defuzzification Layer, and (e) Layer 5: Output
AEZs. The pre-processing of the groundwater-level data was carried Layer (Jang, 1993; Nayak et al., 2004). In Fig. 3, each circle indicates
out using MATLAB R2016a 9.0 (www.mathworks.com/help/matlab/) a fixed node, whereas each square indicates an adaptive node
and MS-Excel software. (i.e., change in the parameters during training). Grid Partition
method is preferred for the models with a few inputs and hence, it
2.3. Machine Learning techniques used: overview and application to was implemented in this study.
study area ANFIS applies a hybrid-learning algorithm consisting of a combina-
tion of “Gradient Descent” and “Least Squares” methods, which outper-
In the present study, three relatively powerful supervised Machine forms the original backpropagation algorithm. To obtain the best output
Learning techniques namely, Adaptive Neuro-Fuzzy Inference System from a Neuro-Fuzzy model, it is mandatory to select the best ‘member-
(ANFIS), Deep Neural Network (DNN), and Support Vector Machine ship function’. A ‘membership function’ is used to generate a member-
(SVM) were employed to predict seasonal groundwater fluctuations in ship grade. The membership functions most commonly used are:
individual Agro-Ecological Zones (AEZs). The ANFIS, DNN and SVM Triangular, Trapezoidal, Gaussian, Two-sided Gaussian, Generalized
models were developed for each AEZ with the help of MATLAB Bell, and Sigmoidal Z- and S-functions (Nguyen et al., 2003). Further-
R2016a 9.0 software (www.mathworks.com/help/matlab/), Python 3 more, there is no standard rule to define the number of membership
(www.python.org) and TensorFlow 2 (www.tensorflow.org) library. functions of a Neuro-Fuzzy model, and the use of a large number of
Brief descriptions of these techniques are provided in the following membership functions is generally avoided to save time and computa-
sub-sections for the benefit of the readers/researchers, together with tional effort (Keskin et al., 2004). According to Shiri and Kisi (2011),
their application in this study. two, three or four membership functions are adequate for predicting
groundwater-depth fluctuations. In this study, the type of membership
2.3.1. Basic concept of Artificial Neural Network (ANN) modeling function and its optimal number for each input were finalized by using
Artifical Neural Network (ANN) is one of the oldest Machine Learn- the trial-and-error method.
ing (ML) algorithms, which is built on a simplified imitation of comput- The merits of ANFIS models are (Nayak et al., 2004): (i) they have
ing performed by the neurons of a human brain. It is a simplified better ability to handle large amounts of noisy data from nonlinear
computational model of the complex biological neuron system and dynamic systems; (ii) they use fuzzy partitions to discount outlier
(Fig. 2), which resembles the brain in two aspects: (a) knowledge is ac- effect; (iii) they have a strong computational complexity restriction;
quired by the network from its environment through a learning process, (iv) they have faster convergence than the typical feedforward ANN
and (b) interneuron connection strengths (known as ‘synaptic models and better generalization capability even with a smaller training
weights’) are used to store the acquired knowledge (Haykin, 2004). dataset; and (v) they have less chance of being stuck in local minima
The ANN models ascertain the relationship between a set of inputs during training as they use a hybrid-learning algorithm.
and desired outputs based on the “pattern recognition” process without
providing any information about the actual processes involved. Thus, 2.3.3. Deep Neural Network (DNN)
they provide a non-linear “Black-Box” modeling approach for simula- Deep Neural Network (DNN) (LeCun et al., 2015) is an advanced
tion/prediction or forecasting. Machine Learning algorithm. It is an extension of Artificial Neural
The ANN structure is typically comprised of an Input Layer, a Hidden Network (ANN) and includes multiple hidden layers with a combination
Layer, and an Output Layer (Fig. 2). Each of the circles in this figure rep- of different activation functions after each layer to capture the most com-
resents a biological neuron, and the arrows indicate connections among plex features in the input data. A typical architecture of DNN is illustrated

4
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

Fig. 2. (Top) Schematic of a biological neuron (Beale and Jackson, 1990); (Bottom) architecture of ‘Artificial Neural Network’.

Fig. 3. Architecture of ANFIS.

(Redrawn from Jang, 1993).

5
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

Fig. 4. Typical architecture of ‘Deep Neural Network’.

(Modified from Bahi and Batouche, 2018).

in Fig. 4, which looks somewhat similar to the basic structure of ANN The salient features of DNN models are as follows (LeCun et al.,
(Fig. 2). Although each of the neurons of DNN has the same functionality 2015):
as that of ANN, a DNN has multiple hidden layers between the input layer
(i) The Deep Neural Network (DNN) model consists of a multi-layer
and the output layer as opposed to usually one hidden layer in case
stack of simple modules which by the process of learning, in-
of ANN.
creases the selectivity and invariance of the representation, and
In this study, RMSE (Root Mean Square Error) was chosen as an error
thereby captures complex input-output mappings effectively
function for the predicted and target values. Due to the increase in com-
without being affected by large irrelevant variations (noises) in
plexity, it becomes difficult to reach global minima for the error function
inputs.
using basic mathematical operations. This is why the DNN uses fast
(ii) Multiple non-linear hidden layers help the DNN model to learn
matrix-based gradient calculation methods (e.g., backpropagation or
better representations using multiple abstraction levels.
gradient descent) to compute optimal weights and biases in the
(iii) Each of the multiple levels of abstraction (representation) is able
network in order to obtain the minimum error. In this study, a multi-
to transform the raw input into a higher level (slightly more ab-
layer fully connected network was used, wherein each neuron is con-
stract). With the composition of enough such transformations,
nected to every neuron in the next layer. The output of the network
very complex functions can be learned efficiently.
can be mathematically expressed as (Goodfellow et al., 2016):
(iv) With the help of backpropagation algorithm, the deep network is

 able to adjust the internal parameters of each layer based on the
i i
ai ¼ f wi  ai−1 þ b ð1Þ representation learned in the previous layer.

2.3.4. Support Vector Machine (SVM)

where, ai = output vector of the ith layer, fi = activation function of the
Support Vector Machine (SVM) is one of the artificial intelligent (AI)
ith layer, wi = weight matrix of the ith layer, and bi = bias vector of the
techniques based on the theory of statistical learning. SVMs are a set of
ith layer.
related supervised learning methods used for solving classification and
In this study, different numbers of hidden layer nodes per layer were
regression problems. SVMs were initially developed to solve classifica-
used for each Agro-Ecological Zone. Unlike the past research, in this
tion problems, but later on they were extended to the domain of regres-
study, the Glorot/Xavier kernel initialization (Glorot and Bengio, 2010)
sion problems as well (Vapnik et al., 1997). Support Vector Regressor
was used to ensure efficient gradient descent, the Dropout method
(SVR) is a regression algorithm based on Support Vectors (Drucker
(Gal and Ghahramani, 2016) was used to randomly drop units (neu-
et al., 1996). The primary objective of this algorithm is to find a hyper-
rons) along with their connections after each hidden layer to avoid
plane that has the maximum margin between the nearest training
the problem of overfitting during training, and the variants of Leaky
data points to increase the generalization in such a way that the hyper-
ReLU (Rectified Linear Unit) activation function (Maas et al., 2013)
plane deviates from the original values by less than epsilon (a tolerance
were used to introduce non-linearity. Further, the hyperparameters of
value).
the DNN models for different Agro-Ecological Zones viz., choice of num-
Mathematically, given the following discriminant function,
ber of hidden layers, number of output nodes in each hidden layer, type
i.e., equation of the hyperplane (Drucker et al., 1996):
of activation function in each hidden layer, and dropout probability for
each node were decided based on the validation dataset by tuning the
parameters using an efficient technique known as ‘Bayesian optimiza-
tion’ as opposed to the commonly used brute-force technique. f ðxÞ ¼ wT x þ b ð2Þ

6
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

We have to minimize: Table 1

Months of the meteorological variables considered as possible inputs for simulating sea-
1 Xn sonal groundwater levels.
kwk2 þ C ξi ð3Þ
2 i¼1
Rainfall and mean temperature (input) Groundwater level (output)

November, December Post-Monsoon Rabi (1–10 January)

  T 
yi w xi þ b ≥1−ξi February/March, March/April, April/May Pre-Monsoon (20–30 March/April/May)
such that ð4Þ July, August Monsoon (20–30 August)
ξi ≥0
September, October Post-Monsoon Kharif (1–10 November)

In this study, Radial Basis Function (RBF) kernel was used to trans-
form the data into a higher-dimensional space to allow a linear separa-
tion. This is how non-linearity is handled by SVM. As a result, Eq. (2) month, the rainfall and temperature of November and December (last two
changes as follows: months) along with the previous season's [i.e., Post-Monsoon Kharif season
(November)] groundwater level were considered as possible model inputs.
f ðxÞ ¼ wT φðxÞ þ b ð5Þ The months of the meteorological variables which were considered as pos-
sible inputs for simulating groundwater levels of four seasons in a year are
where φ is the radial basis function. shown in Table 1. Various combinations of these possible input data were
The hyperparameters such as ‘soft margin parameter (C)’, ‘gamma pa- tried as inputs to the Machine Learning (ML) models developed in this
rameter’ (kernel coefficient), and ‘epsilon parameter’ (specifies the study, and the combination of the input data that yielded the best results
epsilon-tube within which no penalty is associated in the training loss func- was chosen as the final inputs for a particular Agro-Ecological Zone. Further,
tion with points predicted within a distance epsilon from the actual value) due to different units of inputs and output time series, the input data of the
were tuned according to the validation dataset for each of the Agro- models were normalized between zero and one prior to be used for predic-
Ecological Zones. Intuitively, the ‘gamma parameter’ defines how far the in- tion to eliminate the effects of dimension. In general, the networks/models
fluence of a single training example reaches, with low values indicating ‘far’ trained using normalized data achieve better performance and faster con-
and high values indicating ‘close’. Usually, the ‘gamma parameter’ is se- vergence (Shanker et al., 1996; Luk et al., 2000). Finally, the results of the
lected as the inverse of the standard deviation of the RBF kernel, which is developed ML models were de-normalized for retrieving the groundwater
used as a similarity measure between two points. Furthermore, the C pa- levels predicted by the individual models.
rameter trades off correct classification of training examples against the
maximization of the decision function's margin. For larger values of C, a 2.5. Training, testing and performance evaluation of the developed models
smaller margin is accepted if the decision function is better in correctly clas-
sifying all the training points. On the other hand, a lower value of C encour- The total available data of 21 years (1996 to 2016) were divided into
ages a larger margin leading to a more straightforward decision function at two sets for training and testing of the developed ANFIS, DNN and SVM
the cost of training accuracy. Thus, the C parameter behaves as a regulariza- models. A time series of 15 years (1996–2010) were used for training,
tion parameter in the SVM model (Pedregosa et al., 2011). while two years data (2011−2012) were used as a validation dataset for
The main merits of SVM models are (Auria and Moro, 2008): (i) the checking the models during the training phase to prevent overfitting.
SVM models gain flexibility in the choice of the form of the threshold Further, 4 years (2013–2016) data were used for testing the developed
separating by introducing the kernel because the kernel implicitly con- models.
tains a non-linear transformation; (ii) the transformation occurs implic- Moreover, the performance evaluation of all models developed for
itly on a robust theoretical basis and human judgment beforehand is not 18 Agro-Ecological Zone was carried out during training and testing pe-
needed; (iii) unlike the ANN models, the SVM models are inherently riods to critically examine their efficacy in predicting seasonal ground-
convex optimization problems and hence they offer a unique solution water levels in individual Agro-Ecological Zones. Three goodness-of-fit
to a given problem; and (iv) they are less prone to overfitting and criteria (also known as ‘statistical indicators’) were used in this study
noise, thereby ensuring better results. for evaluating the developed models, which are (Table 2): (a) Root

2.4. Selection of model inputs

Table 2
Summary of the goodness-of-fit criteria used for model evaluation.
The selection of significant input variables (i.e., selection of nodes in
the input layer) is one of the most critical steps in the development of Goodness-of-fit criteria Mathematical expression
data-driven models, especially for simulating/predicting hydrologic vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1. Root Mean Square Error (RMSE) u n
u1 X
RMSE ¼ t
2
and hydrogeologic variables. The datasets used in this study include ðhoi −hpi Þ
n i¼1
monthly mean rainfall, monthly mean temperature, and mean seasonal
groundwater level for predicting groundwater fluctuations in 18 Agro- RMSE indicates the overall discrepancy between the observed values and the predicted
Ecological Zones of India. Rainfall was used as the surrogate for ground- values. The lower the value of RMSE, the more accurate the prediction is.
Pn
water recharge, and temperature was used as the surrogate for ground- 2. Percent Bias (PBIAS) ðhoi −hpi Þ
PBIAS ¼ i¼1 Pn  100
water discharge; they significantly influence groundwater and hence i¼1 hoi

they are often used as inputs of the data-driven models for predicting A model with a positive bias will consistently under-estimate and a model with a
groundwater levels. For each Agro-Ecological Zone (AEZ), the mean negative bias will consistently over-estimate. The model is said to be unbiased, if
‘depth to water table’ was calculated for a particular season of a year ‘bias’ = 0 (indicative of an overall agreement between the observed values and
the predicted values).
to obtain mean seasonal groundwater level over each AEZ. Similarly, Pn 2
3. Nash-Sutcliffe Efficiency (NSE) ðhoi −hpi Þ
the mean of gridded rainfall and temperature for all the grids falling in NSE ¼ 1− Pi¼1 2
n
a particular AEZ was computed for each month of a particular year. i¼1 ðhoi −ho Þ

Thus, a set of representative time series of groundwater level, rainfall, The range of NSE lies between 1 (perfect fit) and -∞ (worst fit). Usually, the closer the NSE
and temperature for each AEZ was obtained. is to 1, the more efficient the model is. A zero value of NSE indicates that the observed
mean is as good a predictor as the model, while negative values indicate that the observed
Owing to discontinuity in the seasonal groundwater-level time series,
mean is a better predictor than the model.
the meteorological variables of all the months of a year were not considered Where, hoi = observed groundwater level at the ith time, hpi = predicted groundwater
as model inputs in this study. For example, for simulating groundwater level at the ith time, h = mean of the observed groundwater levels, h = mean of the pre-
o p
levels of Post-Monsoon Rabi season measured in the initial days of January dicted groundwater levels, and n = total number of observations.

7
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

Mean Square Error (RMSE), (b) Percent Bias (PBIAS), and (c) Nash- performance during testing is rated moderate (0.5 ≥ NSE < 0.8) in 6
Sutcliffe Efficiency (NSE) or Modeling Efficiency. Besides these statisti- AEZs [‘Western Himalayas (Sub-Humid)’, ‘Assam and Bengal Plains’,
cal indicators, two graphical indicators such as ‘simultaneous plots of ‘Eastern Himalayas’, ‘Deccan Plateau (Semi-Arid)’, ‘Deccan Plateau and
observed and predicted groundwater levels’ and ‘scatter plots of ob- Eastern Ghats’, and ‘Eastern Coastal Plains’]. On the contrary, the perfor-
served and predicted groundwater levels with the 1:1 line’ were also mance of all the three models during testing is rated very poor (NSE ≤
prepared for each Agro-Ecological Zone during training and testing pe- 0.4) in one AEZ (‘Eastern Ghats and Deccan Plateau’). As to the remain-
riods to rigorously evaluate the performances of the developed models. ing 2 AEZs, the prediction accuracy of only SVM model is poor in these
two AEZs [NSE = 0.417 for ‘Western Plain and Kutch Peninsula’, and
3. Results and discussion NSE = 0.421 for ‘Deccan Plateau (Arid)’], while the prediction accuracy
of both ANFIS and DNN models is moderate in the ‘Western Plain and
3.1. Training and testing results of the ANFIS, DNN and SVM models Kutch Peninsula’ zone but the prediction accuracy of only DNN model
is moderate in the ‘Deccan Plateau (Arid)’ zone and the ANFIS model's
The values of three statistical indicators, i.e., Root Mean Square prediction accuracy in this zone can be rated somewhat moderate
Error (RMSE), Percent Bias (PBIAS) and Nash-Sutcliffe Efficiency (NSE = 0.483).
(NSE) during training and testing of the ANFIS, DNN, and SVM Comparing the individual performance of the models, it is evident
models for all the 18 Agro-Ecological Zones (AEZs) are summarized from Table 3 that the performance of the DNN model during both train-
in Table 3, together with influential model inputs. It is apparent for ing and testing periods is superior to that of the ANFIS and SVM models
this table that for the ANFIS models during training period, RMSE in 13 out of 18 Agro-Ecological Zones (nearly 72.22% of the AEZs), while
values range from 0.138 to 0.649 m, PBIAS from -1.34 to 1.52%, and the ANFIS model's performance during training period is superior to the
NSE from 0.736 to 0.991. For the DNN models during training, the DNN and SVM models in 5 AEZs (about 27.78% of the AEZs) but its per-
RMSE, PBIAS, and NSE values vary from 0.116 to 0.565 m, -1.72 to formance during testing period is superior to the DNN and SVM models
0.73%, and 0.800 to 0.995, respectively. For the SVM models during in 4 AEZs (about 22.22% of the AEZs) only. However, the performance of
training, the values of RMSE, PBIAS, and NSE range from 0.238 to the SVM model during testing period is superior to the DNN and ANFIS
1.041 m, -13.15 to 2.17% and 0.565 to 0.949, respectively. Thus, all models only in 2 AEZs [‘Eastern Himalayas’ and ‘Deccan Plateau (Semi-
the three ML models were trained successfully in almost all the Arid)’], though the performance of the DNN model could be practically
Agro-Ecological Zones. considered at par with that of the SVM model in these two zones. Thus,
On the other hand, during the testing period, RMSE values of the the DNN and ANFIS models are able to predict seasonal groundwater
ANFIS models range from a minimum of 0.256 m for the ‘Northern levels reasonably well in almost all the Agro-Ecological Zones of India.
Plain and Central Highlands’ zone to a maximum of 0.891 m for the The best performance of the DNN model in a majority of AEZs is attrib-
‘Western Himalayas (Sub-Humid)’ zone, whereas RMSE values of the uted to the fact that Deep Neural Networks have greater number of hid-
DNN models vary from a minimum of 0.208 m for the ‘Northern Plain den layers compared to the traditional artificial neural networks, which
and Central Highlands’ zone to a maximum of 0.713 m for the ‘Deccan can easily capture complex relationships between inputs and outputs
Plateau (Semi-Arid)’ zone (Table 3). In the case of the SVM models, (i.e., strong non-linear relationships between meteorological variables
RMSE values vary from a minimum of 0.279 m for the ‘Western Ghats and groundwater levels). In addition, efficient hyperparameter tuning
and Coastal Plains’ zone to a maximum of 0.721 m for the ‘Eastern technique (Bayesian optimization), non-linear functions (Leaky ReLU
Ghats and Deccan Plateau’ zone. Similarly, the PBIAS values of the and its variants), and regularization technique (Dropout) have been
ANFIS models during testing period range from a minimum of -6.33% used in this study for developing DNN models. These merits of the
for the ‘Eastern Plateau (Chhota Nagpur) and Eastern Ghats’ zone to a DNN models help enhance their learning capability and generalization
maximum of -10.97% for the ‘Eastern Himalayas’ zone, while these ability to a greater extent. Further details of the DNN model
values of the DNN models vary from a minimum of -1.85% for the (i.e., theoretical background, its efficient application in this study, and
‘Assam and Bengal Plains’ zone to a maximum of 4.48% for the ‘Eastern salient features) are given in Section 2.3.3 (Methodology). On the
Coastal Plains’ zone. Further, the PBIAS values of the SVM models during other hand, the ANFIS models integrate Fuzzy Logic principles
testing vary from a minimum of -6.23% for the ‘Assam and Bengal Plains’ (human-like reasoning) with the artificial neural networks, which im-
zone to a maximum of 4.07% for the ‘Central Highlands’ zone. Interest- prove their generalization capability compared to the traditional artifi-
ingly, Table 3 reveals that during testing period, all the three models cial neural network models. The merits of the ANFIS model and other
(ANFIS, DNN and SVM) have minimum NSE values (0.407, 0.378 and details are given in Section 2.3.2 (Methodology). Consequently, the
0.350, respectively) in the ‘Eastern Ghats and Deccan Plateau’ zone, ANFIS model is the second most accurate/efficient model in this study,
and maximum NSE values (0.978, 0.981 and 0.931, respectively) in and the SVM model is the least efficient among the three ML models
the ‘Central (Malwa) Highlands and Kathiawar Peninsula’ zone. used in this study, except for few exceptions.
Overall, it can be inferred from the training and testing results that Moreover, a close perusal of Table 3 reveals that in some Agro-
the performance of the developed models during training period is usu- Ecological Zones, two models can be practically rated as ‘almost similar
ally better than that during testing period for most of the Agro- prediction capability’ based on the statistical indicators as highlighted in
Ecological Zones, with a few exceptions (Table 3). This trend is very this table. Also, all the three models can be rated as ‘almost similar pre-
common for the simulation models. diction capability’ in 3 AEZs [‘North-Eastern Hills (Purvachal)’, ‘Eastern
Plateau (Chhota Nagpur) and Eastern Ghats’, and ‘Western Ghats and
3.2. Comparing prediction capability of the ANFIS, DNN and SVM models Coastal Plains’].

3.2.1. Based on statistical indicators 3.2.2. Based on graphical indicators

The modeling results (Table 3) indicate that the performance of all Apart from the goodness-of-fit criteria (statistical indicators), two
the three models in simulating seasonal groundwater levels during test- graphical indicators viz., simultaneous plots and scatter plots (with
ing period is rated the best (NSE ≥ 0.8) for 9 Agro-Ecological Zones 1:1 line) of predicted groundwater levels versus observed groundwater
(AEZs) such as ‘Central Highlands’, ‘Central (Malwa) Highlands and Ka- levels for all the 18 Agro-Ecological Zones (AEZs) were used to further
thiawar Peninsula’, ‘Eastern Plateau (Chhota Nagpur) and Eastern evaluate the efficacy of the developed models. The simultaneous plots
Ghats’, ‘Eastern Plateau (Chhattisgarh)’, ‘Northern Plain’, ‘Eastern and scatter plots (with 1:1 line) of the observed groundwater levels
Plain’, ‘North-Eastern Hills (Purvachal)’, ‘Western Ghats and Coastal and the groundwater levels predicted by the ANFIS, DNN, and SVM
Plains’, and ‘Northern Plain and Central Highlands’, whereas their models during training and testing periods for four selected AEZs are

8
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

Table 3
Goodness-of-fit statistics for the ANFIS, DNN and SVM models of 18 Agro-Ecological Zones along with influential model inputs.

Sl. Agro-Ecological Zone (AEZ) Model inputs Model RMSE (m) PBIAS (%) NSE
no. Training Testing Training Testing Training Testing
1 Western Himalayas (Sub-Humid) R-1, T-1, G-1 ANFIS 0.313 0.891 -0.118 -4.636 0.971 0.772
DNN 0.424 0.394 -1.108 1.815 0.947 0.955
SVM 0.470 0.679 0.763 -1.047 0.934 0.867
2 Western Plain and Kutch Peninsula R-1, R-2, T-1, T-2, G-1 ANFIS 0.339 0.547 -0.490 -1.922 0.975 0.587
DNN 0.426 0.568 0.267 1.705 0.960 0.554
SVM 1.041 0.650 -0.926 -1.558 0.762 0.417
3 Northern Plain and Central R-1, T-, G-1 ANFIS 0.277 0.256 0.270 0.240 0.946 0.936
Highlands DNN 0.212 0.208 -0.420 -0.951 0.968 0.958
SVM 0.384 0.288 0.306 0.758 0.897 0.919
4 Northern Plain R-1, R-2, T-1, G-1 ANFIS 0.138 0.308 0.234 -0.413 0.981 0.862
DNN 0.233 0.246 -0.361 0.908 0.944 0.911
SVM 0.351 0.377 -0.036 2.479 0.874 0.793
5 Eastern Plain R-1, R-, T-1 ANFIS 0.281 0.340 1.267 5.592 0.907 0.842
DNN 0.244 0.222 -1.011 0.310 0.930 0.933
SVM 0.335 0.323 -0.841 2.434 0.868 0.858
6 Assam and Bengal Plains R-1, T-1, G-1 ANFIS 0.144 0.494 0.430 -1.374 0.974 0.755
DNN 0.116 0.426 0.098 -1.846 0.983 0.819
SVM 0.566 0.628 -13.147 -6.232 0.603 0.606
7 Eastern Himalayas R-1, T-1, T-2 ANFIS 0.360 0.637 0.372 10.967 0.916 0.718
DNN 0.264 0.600 -0.444 0.790 0.955 0.750
SVM 0.698 0.584 -9.782 3.808 0.684 0.763
8 North-Eastern Hills (Purvachal) R-1, R-2, T-1 ANFIS 0.361 0.321 1.520 -3.478 0.793 0.813
DNN 0.347 0.308 -1.717 -1.564 0.808 0.827
SVM 0.523 0.330 0.194 -1.662 0.565 0.802
9 Central (Malwa) Highlands and R-1, R-2, T-1, G-1 ANFIS 0.260 0.300 0.448 0.545 0.984 0.978
Kathiawar Peninsula DNN 0.225 0.280 0.034 -0.520 0.988 0.981
SVM 0.519 0.532 -0.089 -3.927 0.935 0.931
10 Central Highlands R-1, R-2, T-1, T-2, G-1 ANFIS 0.178 0.460 -0.207 -3.211 0.991 0.939
DNN 0.127 0.299 -0.029 0.617 0.995 0.974
SVM 0.581 0.534 2.173 4.068 0.901 0.918
11 Eastern Plateau (Chhattisgarh) R-1, T-1 ANFIS 0.301 0.449 -0.257 -1.045 0.969 0.920
DNN 0.269 0.378 -0.275 0.547 0.975 0.943
SVM 0.386 0.523 1.534 1.266 0.949 0.891
12 Eastern Plateau (Chhota Nagpur) and R-1, T-1 ANFIS 0.286 0.373 -0.158 -6.331 0.960 0.916
Eastern Ghats DNN 0.235 0.405 -0.300 0.095 0.973 0.901
SVM 0.505 0.384 1.527 -2.292 0.875 0.911
13 Western Ghats and Coastal Plains R-1, T-1 ANFIS 0.195 0.264 -0.120 2.589 0.964 0.933
DNN 0.147 0.244 0.120 0.021 0.980 0.942
SVM 0.238 0.279 0.331 3.290 0.947 0.924
14 Deccan Plateau (Semi-Arid) R-1, T-1, G-1 ANFIS 0.297 0.666 -0.019 -2.425 0.953 0.736
DNN 0.225 0.609 0.236 -0.384 0.973 0.779
SVM 0.389 0.601 -0.172 -2.468 0.919 0.785
15 Deccan Plateau (Arid) R-1, T-1, T-2, G-1 ANFIS 0.340 0.502 1.124 2.380 0.904 0.483
DNN 0.277 0.444 -0.013 -1.346 0.936 0.595
SVM 0.320 0.531 0.111 -0.084 0.915 0.421
16 Deccan Plateau and Eastern Ghats R-1, T-1, G-1 ANFIS 0.296 0.752 -0.521 -4.388 0.951 0.594
DNN 0.339 0.498 -1.438 2.407 0.936 0.822
SVM 0.442 0.547 -0.001 -1.647 0.892 0.785
17 Eastern Coastal Plains R-1, T-1 ANFIS 0.306 0.406 -1.342 -4.794 0.867 0.700
DNN 0.324 0.425 -0.463 4.478 0.852 0.671
SVM 0.402 0.431 -0.549 0.492 0.770 0.662
18 Eastern Ghats (TN Uplands) and R-1, T-1 ANFIS 0.649 0.689 -0.705 2.112 0.736 0.407
Deccan Plateau DNN 0.565 0.705 0.727 0.437 0.800 0.378
SVM 0.632 0.721 -0.489 2.817 0.750 0.350
Note: 1. R: rainfall; T: temperature; and G: seasonal groundwater level.
2. R- 1 and T- 1 denote monthly rainfall and temperature of December and October months in case of Post-Monsoon Rabi and Post-
Monsoon Kharif seasons, respectively, while they denote monthly rainfall and temperature of March/April/May and August
months in case of Pre-Monsoon and Monsoon seasons, respectively. Similarly, R- 2 and T- 2 denote monthly rainfall and
temperature of November and September months in case of Post-Monsoon Rabi and Post-Monsoon Kharif seasons,
respectively, whereas they denote monthly rainfall and temperature of February/March/April and July months in case of Pre-
Monsoon and Monsoon seasons, respectively.
3. The models highlighted in bold (black color) with dark grey background rows denote the best model during testing period
for a particular AEZ.
4. The models highlighted in bold (pink color) with light grey background rows indicate almost similar performance during
testing period for a particular AEZ.

9
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

illustrated in Figs. 5 to 8 as an example. Since it is not possible to present due to highly dynamic rainfall and pumping patterns in these arid/semi-
the graphs of all the 18 AEZs due to space constraint, the AEZs having arid regions. In the absence of adequate information about pumping pat-
the least and most efficient prediction of groundwater levels among terns, it appears that the temperature (considered as one of the model in-
others from northern, central, and southern parts of India are presented. puts) is not able to represent groundwater discharge reasonably in these
These graphical indicators vividly support the findings obtained by the Agro-Ecological Zones.
interpretation of the modeling results based on the statistical indicators
discussed in the previous sub-section. The graphical indicators also 4. Conclusions
clearly reveal that the groundwater levels simulated/predicted by the
DNN model during training and testing periods are more accurate Groundwater is a hidden and vital natural resource that not only
(closer to the observed groundwater levels) in most of the AEZs than supports irrigated agriculture, food security, and ecological security
the groundwater levels simulated/predicted by the ANFIS and SVM but also serves as the most reliable source for drinking water across
models, and that the performances of the models are normally better the globe. Unfortunately, groundwater depletion is posing increasingly
during training than that during testing with some exceptions. The per- severe water and environment problems in different parts of the
formances of all the three models are reasonably good for all the AEZs world, including India. Groundwater modeling has emerged as a useful
during training period, except for the SVM model's relatively inferior tool to help policymakers in formulating management strategies for the
performance in the ‘Assam and Bengal Plains’, ‘Eastern Himalayas’, and efficient/sustainable utilization of this treasured natural resource. How-
‘North-Eastern Hills (Purvachal)’, whereas the performances of the ever, groundwater modeling at a large scale is very challenging in sev-
three models during testing period vary considerably from one AEZ to eral parts of the world, particularly in developing countries. Unlike
another AEZ. past studies, the present study aims to demonstrate a pragmatic sci-
Furthermore, the graphical indicators evidently support the findings entific framework for predicting seasonal groundwater levels at a
that the DNN model is the most efficient in predicting groundwater levels large scale (i.e., country scale) using field-measured data. Consider-
in most AEZs, followed by the ANFIS model. The worst performance of the ing the research need and the data constraints in developing nations,
models during testing period in the ‘Western Plain and Kutch Peninsula’ three relatively powerful Machine Learning (ML)/Artificial
zone located in Western India as well as in two AEZs of Southern India Intelligence (AI) techniques viz., ANFIS (Adaptive Neuro-Fuzzy In-
namely ‘Deccan Plateau (Arid)’, and ‘Eastern Ghats (TN Uplands) and Dec- ference System), DNN (Deep Neural Network), and SVM (Support
can Plateau’ is also evident from the graphical indicators. These AEZs expe- Vector Machine) were considered for assessing their capability in
rience a high degree of temporal fluctuations in the seasonal groundwater predicting seasonal groundwater levels in different Agro-Ecological
levels over the years as compared to other AEZs. This variability could be Zones (AEZs) of India. The ANFIS, DNN, and SVM models were

Fig. 5. (Top) Comparison of observed and predicted groundwater levels by ANFIS, DNN and SVM models for ‘Western Himalayas (Sub-Humid)’ AEZ; (Bottom) scatter plots of observed and
predicted groundwater levels by ANFIS, DNN and SVM models for ‘Western Himalayas (Sub-Humid)’ AEZ.

10
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

Fig. 6. (Top) Comparison of observed and predicted groundwater levels by ANFIS, DNN and SVM models for ‘Eastern Himalayas’ AEZ; (Bottom) scatter plots of observed and predicted
groundwater levels by ANFIS, DNN and SVM models for ‘Eastern Himalayas’ AEZ.

Fig. 7. (Top) Comparison of observed and predicted groundwater levels by ANFIS, DNN and SVM models for ‘Central (Malwa) Highlands and Kathiawar Peninsula’ AEZ; (Bottom) scatter
plots of observed and predicted groundwater levels by ANFIS, DNN and SVM models for ‘Central (Malwa) Highlands and Kathiawar Peninsula’ AEZ.

11
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

Fig. 8. (Top) Comparison of observed and predicted groundwater levels by ANFIS, DNN and SVM models for ‘Eastern Ghats and Deccan Plateau’ AEZ; (Bottom) scatter plots of observed and
predicted groundwater levels by ANFIS, DNN and SVM models for ‘Eastern Ghats and Deccan Plateau’ AEZ.

developed for each AEZ using groundwater-level data and pertinent predicting seasonal groundwater levels in the majority of AEZs,
meteorological data of 1996–2016, and the performance of the followed by the ANFIS model. The SVM model is the least efficient
models was rigorously evaluated. among the three machine learning models used in this study, except
Based on the results of this study, the following conclusions and rec- for few exceptions. Hence, the DNN or ANFIS models are recom-
ommendations are made: mended for the reliable prediction of seasonal groundwater levels in
different Agro-Ecological Zones of India.
• The modeling results revealed that the performances of all the three • This study explored the possibility of using ML/AI-based models in
models (ANFIS, DNN and SVM) are reasonably superior for all the predicting seasonal groundwater levels at larger scales, when ade-
Agro-Ecological Zones (AEZs) during training period, with an excep- quate data are not available for the process-based modeling of
tion of the SVM model's unusually inferior performance in 2 AEZs hydrogeologic systems. It also emphasizes the need for strengthening
[‘Assam and Bengal Plains’, and ‘North-Eastern Hills (Purvachal)’] groundwater-monitoring network and data acquisition systems
compared to the ANFIS and DNN models. On the other hand, the per- across India for the efficient application of emerging ML/AI techniques
formances of the models during testing period vary significantly from and/or state-of-the-art process-based modeling so as to formulate
one AEZ to another AEZ. strategies for the sustainable management of depleting groundwater
• The performance of the DNN model during both training and testing resources.
periods is superior to the ANFIS and SVM models in predicting sea- • The modeling framework demonstrated in this study can serve as a
sonal groundwater levels in 13 out of 18 AEZs (about 72.22% of the useful tool for the simulation/prediction of groundwater levels at
AEZs). However, the performance of the ANFIS model in predicting larger scales under data-scarce conditions. It can easily be replicated
seasonal groundwater levels during testing period is superior to that in other agro-climatic and hydrogeologic settings of the developing
of the DNN and SVM models only in 4 AEZs (about 22.22% of the or developed world.
AEZs), while the performances of the SVM and DNN models are al-
most similar in 2 AEZs [i.e., Eastern Himalayas', and ‘Deccan Plateau
(Semi-Arid)’]. CRediT authorship contribution statement
• The performances of all the three models during testing period vary
from ‘moderate’ to ‘very poor’ in 3 AEZs [‘Western Plain and Kutch Janaki B. Mohapatra: Methodology, Writing – original draft, Writ-
Peninsula’ located in Western India, and ‘Deccan Plateau (Arid)’ and ing – review & editing, Visualization, Software, Formal analysis. Piyush
‘Eastern Ghats and Deccan Plateau’ located in Southern India]. Jha: Methodology, Writing – original draft, Visualization, Software,
• Overall, it is concluded that the DNN model is the most efficient in Formal analysis. Madan K. Jha: Conceptualization, Supervision,

12
Methodology, Resources, Investigation, Writing – review & editing. vector machines. Environ. Model. Assess. 24, 223–234. https://doi.org/10.1007/
s10666-018-9639-x182.
Sabinaya Biswal: Investigation, Data curation, Writing – original draft, Haykin, S., 2004. Neural Networks: A Comprehensive Foundation. Prentice Hall, New
Writing – review & editing. Jersey.
Huang, F., Huang, J., Jiang, S., Shou, C., 2017. Prediction of groundwater levels using evi-
Declaration of competing interest dence of chaos and support vector machine. J. Hydroinf. 19 (4), 586–606.
IAH, 2008. Groundwater for Life and Livelihood: A Treasure of Planet Earth. International
Association of Hydrogeologists (IAH), Kenilworth, U.K.
The authors declare that they have no known competing financial Jalalkamali, A., Sedghi, H., Manshouri, M., 2011. Monthly groundwater level prediction
interests or personal relationships that could have appeared to influ- using ANN and neuro-fuzzy models: a case study on Kerman plain. J. Hydroinf. 13
(4), 867–876.
ence the work reported in this paper.
Jang, J.S.R., 1993. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans. Syst.
Manage. Cybern. 23 (3), 665–685.
Acknowledgements Jeong, J., Park, E., 2019. Comparative applications of data-driven models representing
water table fluctuations. J. Hydrol. 572, 261–273. https://doi.org/10.1016/j.
jhydrol.2019.02.051.
The authors are very thankful to the three anonymous reviewers for Jha, M.K., Sahoo, S., 2015. Efficacy of neural network and genetic algorithm techniques in
their constructive comments and suggestions, which improved the simulating spatio-temporal fluctuations of groundwater. Hydrol. Process. 29,
previous version of this manuscript. 671–691.
Keck, J., Lee, J., 2021. Embracing analytics in the water industry. J. Water Resour. Plan.
Manag. 147 (5), 02521002. https://doi.org/10.1061/(ASCE)WR.1943-5452.0001375.
References Keskin, M.E., Terzi, O., Taylan, D., 2004. Fuzzy logic model approaches to daily pan evap-
oration estimation in Western Turkey. Hydrol. Sci. J. 49 (6), 1001–1010.
Adamowski, J., Chan, F.H., 2011. A wavelet neural network conjunction model for ground- Khaki, M., Yusoff, I., Islami, N., 2015. Simulation of groundwater level through artificial in-
water level forecasting. J. Hydrol. 407, 28–40. telligence system. Environ. Earth Sci. 73, 8357–8367.
Alizamir, M., Kisi, O., Zounemat-Kermani, M., 2018. Modelling long-term groundwater Kholghi, M., Hosseini, S.M., 2009. Comparison of groundwater level estimation using
fluctuations by extreme learning machine using hydro-climatic data. Hydrol. Sci. J. Neuro-fuzzy and Ordinary Kriging. Environ. Model. Assess. 14, 729–737.
63 (1), 63–73. Kløve, B., Ala-Aho, P., Bertrand, G., Gurdak, J.J., Kupfersberger, H., Kvaerner, J., Muotka, T.,
Anderson, M.P., Woessner, W.W., 1992. Applied Groundwater Modeling. Academic Press, Mykrä, H., Preda, E., Rossi, P., Uvo, C.B., Velasco, E., Pulido-Velazquez, M., 2014. Cli-
San Diego. mate change impacts on groundwater and dependent ecosystems. J. Hydrol. 518,
AQUASTAT, 2010. Water resources development and management service. Food and Ag- 250–266.
riculture Organization of the United Nations, Rome, Italy. http://www.fao.org/nr/ Kouziokas, G.N., Chatzigeorgiou, A., Perakis, K., 2018. Multilayer feed forward models in
water/aquastat/main/index.stm. groundwater level forecasting using meteorological data in public management.
ASCE, 2000a. Artificial neural networks in hydrology, part II: hydrologic application. Water Resour. Manag. 32, 5041–5052.
J. Hydrol. Eng. ASCE 5 (2), 124–137. Kumar, D., Roshni, T., Singh, A., Jha, M.K., Samui, P., 2020. Predicting groundwater depth
ASCE, 2000b. Artificial neural networks in hydrology, part I: preliminary concepts. fluctuations using deep learning, extreme learning machine and Gaussian process:
J. Hydrol. Eng. ASCE 5 (2), 115–123. a comparative study. Earth Sci. Inf. 13, 1237–1250. https://doi.org/10.1007/s12145-
Auria, L., Moro, R.A., 2008. Support Vector Machines (SVM) as a technique for solvency 020-00508-y.
analysis. DIW Discussion Papers, No. 811. Deutsches Institut für Wirtschaftsforschung LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436–444.
(DIW), Berlin. Liu, W., Bailey, R.T., Andersen, H.E., Jeppesen, E., Nielsen, A., Peng, K., Molina-Navarro, E.,
Bahi, M., Batouche, M., 2018. Deep Learning for Ligand-based virtual screening in drug Park, S., Thodsen, H., Trolle, D., 2020. Quantifying the effects of climate change on hy-
discovery. Third International Conference on Pattern Analysis and Intelligent Systems drological regime and stream biota in a groundwater-dominated catchment: a
(PAIS), IEEE, pp. 1–5. modeling approach combining SWAT-MODFLOW with flow-biota empirical models.
Barzegar, R., Fijani, E., Asghari Moghaddam, A., Tziritis, E., 2017. Forecasting of groundwa- Sci. Total Environ. 745, 140933. https://doi.org/10.1016/j.scitotenv.2020.140933.
ter level fluctuations using ensemble hybrid multi-wavelet neural network-based Luk, K.C., Ball, J.E., Sharma, A., 2000. A study of optimal model lag and spatial inputs to ar-
models. Sci. Total Environ. 599–600, 20–31. tificial neural network for rainfall forecasting. J. Hydrol. 227, 56–65.
Beale, R., Jackson, T., 1990. In: Hilger, Adam (Ed.), Neural Computing: An Introduction. IOP Maas, A.L., Hannun, A.Y., Ng, A.Y., 2013. Rectifier nonlinearities improve neural network
Publishing Ltd., Techno House, Bristol. acoustic models. Proc. International Conference on Machine Learning. Vol. 30, p. 3
CGWB, 2006. Dynamic Groundwater Resources of India. (as on March 2004). Central No. 1.
Ground Water Board (CGWB), Ministry of Water Resources, New Delhi, India. Maier, H.R., Dandy, G.C., 2000. Neural networks for the prediction and forecasting of
CGWB, 2017. Dynamic Groundwater Resources of India. Central Ground Water Board water resources variables: a review of modeling issues and applications. Environ.
(CGWB), Ministry of Water Resources, River Development and Ganga Rejuvenation, Model. Softw. 15, 101–124.
Government of India, Faridabad. Maiti, S., Tiwari, R.K., 2014. A comparative study of artificial neural networks, Bayesian
Coppola, E.A., Rana, A.J., Poulton, M.M., Szidarovszky, F., Uhl, V.W., 2005. A neural network neural networks and Adaptive Neuro-Fuzzy Inference System in groundwater level
model for predicting aquifer water level elevations. Groundwater 43 (2), 231–241. prediction. Environ. Earth Sci. 71, 3147–3160.
Coulibaly, P., Anctil, F., Aravena, R., Bobée, B., 2001. Artificial neural network modeling of Margat, J., Van der Gun, J., 2013. Groundwater Around the World: A Geographic Synopsis.
water table depth fluctuations. Water Resour. Res. 37, 885–896. CRC Press, New York.
Daliakopoulos, I.N., Coulibaly, P., Tsanis, I.K., 2005. Groundwater level forecasting using Mirzavand, M., Khoshnevisan, B., Shamshirband, S., Kisi, O., Ahmad, R., Akib, S., 2015. Eval-
artificial neural networks. J. Hydrol. 309, 229–240. uating groundwater level fluctuation by support vector regression and neuro-fuzzy
Das, U.K., Roy, P., Ghose, D.K., 2017. Modeling water table depth using adaptive neuro- methods: a comparative study. Nat. Hazards https://doi.org/10.1007/s11069-015-
fuzzy inference system. ISH J. Hydraul. Eng., 1–7 https://doi.org/10.1080/ 1602-4.
09715010.2017.1420497. Mohanty, S., Jha, M.K., Kumar, A., Sudheer, K.P., 2010. Artificial neural network modeling
Döll, P., Zhang, J., 2010. Impact of climate change on freshwater ecosystems: a global-scale for groundwater level forecasting in a river island of eastern India. Water Resour.
analysis of ecologically relevant river flow alterations. Hydrol. Earth Syst. Sci. 14, Manag. 24 (9), 1845–1865.
783–799. Moosavi, V., Vafakhah, M., Shirmohammadi, B., Behnia, N., 2013. A wavelet-ANFIS hybrid
Drucker, H., Burges, C.J.C., Kaufman, L.C., Smola, A.J., Vapnik, V.N., 1996. Support vector re- model for groundwater level forecasting for different prediction periods. Water
gression machines. Proceedings of the 9th International Conference on Neural Infor- Resour. Manag. 27, 1301–1321.
mation Processing Systems (NIPS'96), December 1996, pp. 155–161. Moosavi, V., Vafakhah, M., Shirmohammadi, B., Ranjbar, M., 2014. Optimization of
Emamgholizadeh, S., Moslemi, K., Karami, G., 2014. Prediction the groundwater level of wavelet-ANFIS and wavelet-ANN hybrid models by Taguchi method for groundwater
Bastam Plain (Iran) by artificial neural network (ANN) and adaptive neuro-fuzzy in- level forecasting. Arab. J. Sci. Eng. 39, 1785–1796.
ference system (ANFIS). Water Resour. Manag. 28, 5433–5446. Mukherjee, A., Ramachandran, P., 2018. Prediction of GWL with the help of GRACE TWS
Gal, Y., Ghahramani, Z., 2016. Dropout as a Bayesian approximation: representing model for unevenly spaced time series data in India: analysis of comparative performances
uncertainty in Deep Learning. Proceedings of the International Conference on Ma- of SVR, ANN and LRM. J. Hydrol. 558, 647–658.
chine Learning, pp. 1050–1059. Nadiri, A.A., Naderi, K., Khatibi, R., Gharekhani, M., 2019. Modelling groundwater level
Glorot, X., Bengio, Y., 2010. Understanding the difficulty of training deep feedforward variations by learning from multiple models using fuzzy logic. Hydrol. Sci. J. 64 (2),
neural networks. Proceedings of the Thirteenth International Conference on Artificial 210–226.
Intelligence and Statistics, pp. 249–256. Nayak, P.C., Sudheer, K.P., Rangan, D.M., Ramasastri, K.S., 2004. A neuro-fuzzy computing
Gong, Y., Zhang, Y., Lan, S., Wang, H., 2016. A comparative study of artificial neural net- technique for modeling hydrological time series. J. Hydrol. 291 (1–2), 52–66.
works, support vector machines and adaptive neuro-fuzzy inference system for fore- Nayak, P.C., Satyaji Rao, Y.R., Sudheer, K.P., 2006. Groundwater level forecasting in a shal-
casting groundwater levels near Lake Okeechobee, Florida. Water Resour. Manag. 30, low aquifer using artificial neural network approach. Water Resour. Manag. 20,
375–391. 77–90.
Goodfellow, I., Bengio, Y., Courville, A., 2016. Deep Learning. The MIT Press, Cambridge, Nguyen, H.T., Prasad, N.R., Walker, C.L., Walker, E.A., 2003. A First Course in Fuzzy and
Massachusetts. Neural Control. CRC Press (Chapman & Hall/CRC), New York.
Guzman, S.M., Paz, J.O., Tagert, M.L.M., Mercer, A.E., 2019. Evaluation of seasonally classi- Nourani, V., Mousavi, S., 2016. Spatiotemporal groundwater level modeling using hybrid
fied inputs for the prediction of daily groundwater levels: NARX networks vs support artificial intelligence-meshless method. J. Hydrol. 536, 10–25.
J.B. Mohapatra, P. Jha, M.K. Jha et al. Science of the Total Environment 785 (2021) 147319

Nourani, V., Mogaddam, A.A., Nadiri, A.O., 2008. An ANN-based model for spatiotemporal Sreekanth, P.D., Sreedevi, P.D., Ahmed, S., Geethanjali, N., 2011. Comparison of FFNN and
groundwater level forecasting. Hydrol. Process. 22, 5054–5066. ANFIS models for estimating groundwater level. Environ. Earth Sci. 62, 1301–1310.
Nourani, V., Hosseini Baghanam, A., Adamowski, J., Kisi, O., 2014. Applications of hybrid Sun, Y., Wendi, D., Kim, D.E., Liong, S.Y., 2016. Application of artificial neural networks in
wavelet-artificial intelligence models in hydrology: a review. J. Hydrol. 514, 358–377. groundwater table forecasting: a case study in a Singapore swamp forest. Hydrol.
Nourani, V., Alami, M.T., Vousoughi, F.D., 2015. Wavelet-entropy data pre-processing ap- Earth Syst. Sci. 20, 1405–1412.
proach for ANN-based groundwater level modeling. J. Hydrol. 524, 255–269. Suryanarayana, C., Sudheer, C., Mahammood, V., Panigrahi, B.K., 2014. An integrated
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., et wavelet-support vector machine for groundwater level prediction in Visakhapatnam,
al., 2011. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830. India. Neurocomputing 145, 324–335.
Purohit, M.K., Kaur, S., 2016. Rainfall Statistics of India. Hydromet Division, India Meteoro- Tang, Y., Zang, C., Wei, Y., Jiang, M., 2019. Data-driven modeling of groundwater level with
logical Department (IMD), Ministry of Earth Sciences, Government of India, New least-square support vector machine and spatial-temporal analysis. Geotech. Geol.
Delhi. Eng. 37, 1661–1670. https://doi.org/10.1007/s10706-018-0713-6.
Rajaee, T., Ebrahimi, H., Nourani, V., 2019. A review of the artificial intelligence methods in Taormina, R., Chau, K., Sethi, R., 2012. Artificial neural network simulation of hourly
groundwater level modeling. J. Hydrol. 572, 336–351. groundwater levels in a coastal aquifer system of the Venice lagoon. Eng. Appl.
Roshni, T., Jha, M.K., Deo, R.C., Vandana, A., 2019. Development and evaluation of hybrid Artif. Intell. 25, 1670–1676.
artificial neural network architectures for modeling spatio-temporal groundwater UN Water, 2007. Coping with water scarcity: challenge of the twenty-first century. Report
fluctuations in a complex aquifer system. Water Resour. Manag. 33 (7), 2381–2397. for World Water Day 2007. http://www.unwater.org/wwd07/downloads/docu-
Roshni, T., Jha, M.K., Drisya, J., 2020. Neural network modeling for groundwater-level fore- ments/escarcity.pdf.
casting in coastal aquifers. Neural Comput. & Applic. 32, 12737–12754. https://doi. Vapnik, V., Golowich, S., Smola, A., 1997. Support vector method for function approxima-
org/10.1007/s00521-020-04722-z. tion, regression estimation, and signal processing. In: Mozer, M., Jordan, M., Petsche,
Sahoo, S., Jha, M.K., 2013. Groundwater-level prediction using multiple linear regression T. (Eds.), Advances in Neural Information Processing Systems. vol. 9. The MIT Press,
and artificial neural network techniques: a comparative assessment. Hydrogeol. J. Cambridge, MA, pp. 281–287.
21 (8), 1865–1887. Vörösmarty, C.J., McIntyre, P.B., Gessner, M.O., Dudgeon, D., Prusevich, A., Green, P.,
Sahoo, S., Russo, T.A., Elliott, J., Foster, I., 2017. Machine learning algorithms for modeling Glidden, S., Bunn, S.E., Sullivan, C.A., Reidy, L.C., Davies, P.M., 2010. Global threats to
groundwater level changes in agricultural regions of the U.S. Water Resour. Res. 53, human water security and river biodiversity. Nature 467 (7315), 555–561.
3878–3895. World Bank, 2010. http://www.worldbank.org/en/news/feature/2012/03/06/india-
Savic, D., 2019. Artificial Intelligence: How Can water Planning and Management Benefit groundwater-critical-diminishing (Accessed on October 24, 2018).
from It? IAHR White Papers. International Association for Hydro-Environment Engi- WWAP, 2012. The United Nations World Water Development Report 4: Managing Water
neering and Research (IAHR), Madrid, Spain, pp. 1–3. under Uncertainty and Risk. World Water Assessment Program (WWAP), Paris,
Sehgal, J., Mondal, D.K., Mondal, C., Vadivelu, S., 1992. Agro-ecological zones of India. France.
Technical Bulletin, NBSS Publication. 24 Nagpur, Maharashtra, India. Yoon, H., Jun, S.C., Hyun, Y., Bae, G.O., Lee, K.K., 2011. A comparative study of artificial neu-
Shanker, M., Hu, M.Y., Hung, M.S., 1996. Effect of data standardization on neural network ral networks and support vector machines for predicting groundwater levels in a
training. Int. J. Sci. Manage. 24, 385–397. coastal aquifer. J. Hydrol. 396 (1), 128–138.
Sharma, K.D., 2009. Groundwater management for food security. Curr. Sci. 96 (11), Yoon, H., Hyun, Y., Ha, K., Lee, K.K., Kim, G.B., 2016. A method to improve the stability and
1444–1447. accuracy of ANN- and SVM-based time series models for long-term groundwater
Shiri, J., Kisi, O., 2011. Comparison of genetic programming with neuro-fuzzy systems for level predictions. Comput. Geosci. 90, 144–155. https://doi.org/10.1016/j.
predicting short-term water table depth fluctuations. Comput. Geosci. 37, cageo.2016.03.002.
1692–1701. Yu, H., Wen, X., Feng, Q., Deo, R.C., Si, J., Wu, M., 2018. Comparative study of hybrid-
Shiri, J., Kisi, O., Yoon, H., Lee, K.-K., Nazemi, A.H., 2013. Predicting groundwater level fluc- wavelet artificial intelligence models for monthly groundwater depth forecasting in
tuations with meteorological effect implications: a comparative study among soft extreme arid regions, Northwest China. Water Resour. Manag. 32 (1), 301–323.
computing techniques. Comput. Geosci. 56, 32–44. Zare, M., Koch, M., 2018. Groundwater level fluctuations simulation and prediction by
Shirmohammadi, B., Vafakhah, M., Moosavi, V., Moghaddamnia, A., 2013. Application of ANFIS and hybrid Wavelet-ANFIS/Fuzzy C-Means (FCM) clustering models: applica-
several data-driven techniques for predicting groundwater level. Water Resour. tion to the Miandarband plain. J. Hydro Environ. Res. 18, 63–76. https://doi.org/
Manag. 27 (2), 419–432. 10.1016/j.jher.2017.11.004.
Shukla, R., Chakraborty, A., Joshi, P.S., 2017. Vulnerability of agro-ecological zones in India Zhang, N., Xiao, C., Liu, B., Liang, X., 2017. Groundwater depth predictions by GSM, RBF
under the earth system climate model scenarios. Mitig. Adapt. Strateg. Glob. Chang. and ANFIS models: a comparative assessment. Arab. J. Geosci. 10, 189. https://doi.
22, 399–425. org/10.1007/s12517-017-2954-8.
Solomatine, D.P., Ostfeld, A., 2008. Data-driven modelling: some past experiences and
new approaches. J. Hydroinf. 10, 3–22.

14