Arabian Journal of Geosciences (2022) 15: 723

https://doi.org/10.1007/s12517-022-09900-y

ORIGINAL PAPER

Prediction of groundwater‑level using novel SVM‑ALO, SVM‑FOA,

and SVM‑FFA algorithms at Purba‑Medinipur, India
Sandeep Samantaray1 · Abinash Sahoo2 · Deba Prakash Satapathy1

Received: 17 August 2021 / Accepted: 15 March 2022 / Published online: 7 April 2022
© Saudi Society for Geosciences 2022

Abstract
Accurate and reliable prediction of groundwater level (GWL) fluctuation is vital and significant in water resources planning
and management. Because of intricacies in underground geological arrangement, efficiency of real-time GWL prediction is
inadequate. In the present study, a hybrid support vector machine (SVM) incorporated with ant lion optimizer (SVM-ALO)
was employed to estimate monthly GWL using data collected from two observational wells in Purba-Medinipur, India,
considering ten input combinations. The accuracy of hybrid SVM-ALO model is assessed against hybrid SVM-FOA (fruit
fly optimization algorithm), SVM-FFA (firefly algorithm), and conventional SVM models using five statistical performance
indices, i.e. root mean-squared error (RMSE), Willmott Index (WI), Pearson’s correlation coefficient (PCC), coefficient of
determination (R2), and graphical analysis. Analysis of results indicated that SVM-ALO model showed superior performance
than SVM-FOA and SVM-FFA models for all scenarios at both stations. The SVM-ALO-M10 model had the lowest value
of RMSE = 7.6638/8.838 and the highest value of PCC = 0.9815/0.98001, and WI = 0.98215/0.98067 for testing period at
Sherkhanchawk/Basantia high school, respectively. The developed robust SVM-ALO model is more efficient and appropri-
ate compared to SVM-FOA, SVM-FFA, and standalone SVM models for estimating monthly ground water levels in the
study region.

Keywords Ant lion optimizer · Fruit fly optimization · Firefly algorithm · Ground water · India

Introduction a major cause of worry for engineers, particularly in emerg-

ing countries such as India (Konikow and Kendy 2005).
One of the major sources of supply for drinking, industrial, Absence of optimum management of accessible resources
and agricultural water demands is groundwater. In many and change in climatic conditions lead to an increasing crisis
semiarid and arid regions around the world, it is considered of water (Ehteram et al. 2019a, b). Estimation accuracy of
the single dependable source of water. Due to ever increasing GWLs has an important impact on effectiveness of a sup-
requirement of water for different necessities and its overex- porting decisive system and misuse from accessible water
ploitation, management of groundwater resources has been resources. Prediction of GWL fluctuation is highly nonlinear
and very complicated as it is governed by several complex
factors like evapotranspiration, precipitation, soil character-
Responsible Editor: Broder J. Merkel
istics, different lag value of GWL, and topography of the
* Sandeep Samantaray catchment.
sandeep1139_rs@civil.nits.ac.in Groundwater simulation models can be classified into two
Abinash Sahoo major classifications: physical models and machine learning
abinash_rs@civil.nits.ac.in models. In case of physical models, a suitable combination
Deba Prakash Satapathy of aquifer parameters is utilised for determining spatial vari-
dpsatapathy@cet.edu.in ations in underground space. Even by costly on-site inves-
tigations that additionally increase the cost of computation,
1
Department of Civil Engineering, College of Engineering information regarding underground geological structures is
and Technology, Bhubaneswar, Odisha, India
challenging to obtain. Hence, physical domain needs to be
2
Department of Civil Engineering, National Institute divided for obtaining numerical solutions (Taormina et al.
of Technology, Silchar, Assam, India

13
Vol.:(0123456789)
723 Page 2 of 22 Arab J Geosci (2022) 15: 723

2012). A suitable alternate to these models is machine learn- people’s livelihood in an area. Non-stationary and non-lin-
ing models that can provide precise prediction outcomes earity of hydrological applications cannot be captured by
with less cost and effective calibration time where avail- conventional models which are linear. For improving accu-
ability of data is scarce and physical process is not the main rateness of predicted results, a common practice adopted is
study motivation (Mohanty et al. 2015). Among several integrating conventional techniques with a meta-heuristic
data-driven techniques, artificial neural networks (ANNs) optimization algorithm for forming a robust hybrid model
and SVM are considered conventional nonlinear predic- where distinctive features of all models are integrated for
tors, surpassing barriers of physical models (Behzad et al. capturing various patterns in data. Several nature-inspired
2010; Samantaray et al. 2020b, Das et al. 2019). Yet, we are algorithms like particle swarm optimisation (PSO); FFA;
familiar with the fact that ANN model sustains problems of FOA; Multi-Verse Optimiser (MVO); Grey Wolf Optimiser
over fitting and drawbacks of local minima. To overcome (GWO); ALO, and Whale Optimisation Algorithm (WOA)
these shortcomings, SVM model can be utilised (Hosseini have successfully been employed in fields of hydrology and
and Mahjouri 2016). Relative studies on ANN and SVM water resources for optimization, modelling, prediction,
have been conducted in perspective of groundwater level and forecasting (Samantaray and Sahoo 2020; Ferreira and
predictions. Cunha 2020; Malik et al. 2020; Mattar 2018; Samantaray
et al. 2020c; Granata 2019; Sanikhani et al. 2018; Karbasi
Previous studies 2018). Assessments based on theoretical and empirical out-
comes recommend that hybrid techniques can be efficient
Performances of several data driven techniques were evalu- and effective in increasing forecasting capability (Malekza-
ated for GWL forecasting at different locations for differ- deh et al. 2019). In the current study, SVM method coupled
ent prediction periods (Jalalkamali et al. 2011; Moosavi with ALO optimisation algorithm for increasing efficacy of
et al. 2013; Shirmohammadi et al. 2013; Samantaray et al. the model.
2020a). Yoon et al. (2016) utilised a weighted error func- Dash et al. (2010) proposed an integrated ANN-GA
tion method for improving performances of ANN and SVM (genetic algorithm) model to accurately predict GWL in
to predict GWL on long-term basis in South Korea with lower Mahanadi river of Orissa, India, and compared its
respect to rainfall. Outcomes revealed that weighted error performance with other ANN algorithms. Results revealed
function improved accuracy and stability of prediction mod- that hybrid ANN-GA algorithm produced better perfor-
els, with SVM model performing superiorly compared to mance and accuracy in medium and high GWL prediction
ANN model. Nie et al. (2017) applied radial basis function than conventional ANN methods. Suryanarayana et al.
network (RBFN) and SVM for simulating GWL variations in (2014) proposed wavelet based SVM (W-SVM) and ANN
Da’an region of China. Based on assessment of results using (W-ANN) models for predicting monthly GWL fluctua-
statistical indices, it was found that SVM model had superior tions in Visakhapatnam city situated in Andhra Pradesh,
capability to simulate and predict GWL at a specified loca- India. Results indicated that W-SVM model provided bet-
tion. Sattari et al. (2018) utilised M5 decision tree and SVM ter accuracy in GWL prediction than conventional ANN,
for GWL prediction in Ardebil plain, Iran. Results specified SVM, and auto-regressive integrated moving average
that performances of both models are better for GWL predic- (ARIMA) models. Al-Shammari et al. (2016) proposed
tion at proposed location. Tang et al. (2019) proposed a new SVM-FFA model for predicting dew-point temperature
two-stage data-driven structure for modelling time-series ­( T dew) of Isfahan Province in Iran on daily basis and
GWL in north county of UK with spatiotemporal study and compared its prediction capability against ANN, genetic
least square SVM (LS-SVM). Results of LS-SVM are com- programming (GP), and SVM models. Ghorbani et al.
pared with ANN, k-nearest neighbour, random forest, and (2017) applied hybrid SVM-FFA model for predicting
conventional SVM methods. They concluded that LS-SVM permanent wilting point and soil field capacity utilising
outperformed other data driven models in GWL prediction. certain easily accessible soil properties. Zare and Koch
Wang et al. (2018) established a customary decision-tree- (2018) comparatively evaluated the potential of ANFIS
based model for GWL prediction appropriate for conditions and Wavelet-ANFIS for simulating and predicting GWL
where low-dimension input data are obtainable. Findings fluctuations in Miandarband plain, Iran. Results showed
from their study indicated that proposed model produced that applied models can be utilised with suitable accu-
higher accuracy in forecasting GWL variations. rateness for GWL prediction, with W-ANFIS performing
While deciding management of water, results from vari- slightly better. Huang et al. (2017) applied SVM-PSO
ous kinds of techniques are considered by hydrologists which model based on chaotic theory for predicting GWLs at
help them in achieving their goal. Depending on a particular two locations in China and assessed its performance by
approach can be too dangerous, predominantly in hydro- comparing linear SVM-PSO and chaotic BPNN mod-
logical field of study since outcomes of the decisions reflect els. Results revealed that chaotic SVM-PSO model had

13
Arab J Geosci (2022) 15: 723 Page 3 of 22 723

higher prediction accurateness compared to other models. the best among all in GWL prediction. Seidu et al. (2021)
Ebrahimi and Rajaee (2017) explored potential of wavelet applied WT-SaDE-ELM (wavelet transform-self adap-
based ANN (W-ANN), SVM (W-SVM) and multi linear tive diferential evolutionary-extreme learning machine),
regression (W-MLR) models in simulating GWL consid- variational mode decomposition-self adaptive diferential-
ering groundwater data from two wells located in Qom ELM, and empirical wavelet transform-self adaptive difer-
floodplain, Iran. They observed that implementation of ential ELM to estimate and predict GWLs in sub-Saharan
wavelet analysis improved performance of simple neural Africa. Rezaei et al. (2021) integrated MODFLOW with
network models in GWL simulation. Moazenzadeh et al. shark smell optimization (SSO), PSO, and FFA to esti-
(2018) aimed at predicting evaporation on a daily scale at mate monthly GWL in Dezful-Andimeshk plain, Iran.
two meteorological stations in northern Iran using hybrid Results revealed that MODFLOW-SSO estimated GWL
SVM-FFA and simple SVM models. Mehr et al. (2019) better than MODFLOW-PSO, MODFLOW-FFA, and
proposed a hybrid SVM-FFA technique for 1-month ahead MODFLOW. Cui et al. (2021) proposed ANFIS-IAGWO
rainfall forecasts at two rain gauging stations situated in (Improved Alpha-Guided Grey Wolf optimisation) algo-
north-western region of Iran. Obtained results from the rithm for predicting GWL reliably in a severely irrigated
above studies indicated that proposed SVM-FFA model area of Northwest Bangladesh. Effectiveness of projected
provided most encouraging predictions with more accu- model was compared with ANN, ANFIS, and ANFIS-PSO
racy and reliability outperforming other applied models. models and found that ANFIS-IAGWO showed better per-
Wang et al. (2019) developed a water quality mechanism formance in GWL prediction.
model based on FOA which focuses on studying eutrophi- This study aims to predict monthly GWL at coastal region
cation risk assessment techniques of reservoirs and lakes. of Purba Medinipur situated in West Bengal, India, using a
Moayedi et al. (2019) applied ANN-SHO (spotted hyena novel SVM-ALO model. The efficiency of the hybrid SVM-
optimizer) and ANN-ALO hybrid models for simulating ALO model is evaluated by comparing it with hybrid SVM-
shear strength of soil. Findings based on comparison of FOA, SVM-FFA, and conventional SVM models. Strength
results revealed that ANN-ALO model performed more of SVM lies in its capability to optimise SVM parameters
effectively than ANN-SHO model in the prediction of like bias values and weight connections. In the present work,
soil shear strength. Samadianfard et al. (2019) proposed an innovative ALO algorithm was utilised for training SVM
SVM, SVM-FOA, and M5 model tree for forecasting river model. It is worth stating that in the present work, perfor-
flow in Urmia basin, Iran. Performance of applied models mance of applied hybrid SVM models was studied for the
was evaluated based on different statistical indicators, and first time for monthly GWL prediction at the selected study
results showed that SVM-FOA model had best forecast- area. The working procedure of the present study is illus-
ing performance of river flow at proposed study location. trated in the form of a flowchart, as shown in Fig. 1.
Chen et al. (2020) developed a model for predicting dam
deformation by applying a hybrid neural network model
integrating LS-SVM with ALO to assess operational state Study area
of concrete dams. Obtained results confirmed that pro-
posed hybrid model outperformed other applied models The district of Purba Medinipur (Fig. 2) is a part of lower
in dam deformation prediction. Tikhamarine et al. (2020) Eastern coastal plains and Indo-Gangetic Plain. It is sur-
employed hybrid SVM-WOA, SVM-ALO, and SVM- rounded by South 24 Parganas district and River Hooghly
MVO (Multi-Verse Optimizer) for estimating monthly to east; Bay of Bengal in south; Paschim Medinipur dis-
­ETo at Tlemcen and Algiers meteorological stations situ- trict at north and west border and Odisha state to southwest.
ated in northern Algeria. Based on assessment of proposed Elevation of this district is inside 10 m above mean sea level
model performances, they found that SVM-WOA model (MSL). Coastal region of East Midinipur district is consid-
performed superiorly in all scenarios than SVM-ALO and ered for this study which is situated near Bay of Bengal and
SVM-MVO models. Cao et al. (2020) intended at predict- at foothill of Rajmahal and Singhbhum. Precise location of
ing GWL of Tangjiao landslide, China by applying sin- wells (Sherkhanchawk and Basantia high school) consid-
gle-factor SVM-GA, multi-factor back propagation neural ered for the present study in East Midinipur is illustrated in
network (BPNN), and SVM-GA. Findings suggested that Fig. 2, which falls between geographical location 21°38′ N
multi-factor SVM-GA gave the best performance than sin- and 22°30′ N latitudes to 87°27′ E and 88°11′ E longitudes.
gle-factor SVM-GA and multi-factor BPNN. Banadkooki Coastal region of East Midinipur was mostly built by strati-
et al. (2020) applied genetic programming (GP), hybrid fied arranged layers formed by steady deposition of boulders,
RBF-WOA, and hybrid multilayer perception (MLP- silts, cobbles, and gravels transported through Ganga, Bhagi-
WOA) models for predicting GWL at Yazd province in rathi, and other tenacious rivers. Basic statistical indicators
Iran. Results revealed that MLP-WOA model performed of applied data are specified in Table 1. Figure 3 shows the

13
723 Page 4 of 22 Arab J Geosci (2022) 15: 723

( )
di −(𝜔𝜑) xi + bi ≤ ε + ξk
subject to { 𝜔𝜑 xi + bi − di ≤ ε + ξ∗k
ξk , ξ∗k ≥ 0, i = 1 … l

where 𝜉k and 𝜉k∗- slack variables penalising training errors

using loss function above tolerance of error ε, and Cis the
positive trade off constraint maintaining balance between
the considered ε and ω in an optimisation problem. Basic
architecture of SVM model is demonstrated in Fig. 4.

FFA

Yang (2009) introduced a nature-inspired algorithm called

FFA which simulates social behaviour of fireflies. Each
firefly is treated as an agent for developing a solution to
an optimisation problem. In 6, each firefly illuminates with
respect to its quality. Regardless of its gender is enticed to
the brightest firefly, which makes investigation within search
space more effective compared to traditional search tech-
niques under assumption of swarm intelligence (Yaseen
et al. 2017). A firefly’s attractiveness is directly propor-
tional to its light intensity, and hence, fireflies come together
nearer to the brightest firefly (which is the optimum solu-
tion) (Samantaray and Ghose 2020b). Formulation of FFA
is done by developing an objective function and discrepancy
in intensity of light. Amid any two fireflies xi and xj, carte-
sian distance rij is expressed in mathematical terms given by
(Yang et al. 2012):
Fig. 1  Study flowchart of present research
�∑
n
rij =⇑ xi − xj ⇑= (x − xj,k )
k=1 i,k (3)
graphical representation of ground water level at proposed
gauge station (Table 2).
(4)
2
I = Io e−r

Methodology 𝛽(r) = 𝛽0 e−𝛾r

2
(5)

SVM where 𝛾 is the light absorption factor, I(r) is the intensity of

light at distance r; β(r) is the attractiveness β at distance r,
In regression, procedure of an SVM estimator (f) can be Io is the firefly’s primary light intensity, and βo is the attrac-
articulated using the following equation: tiveness at distance r = 0. Subsequent movement of firefly i
is given by:
f(x) = 𝜔𝜑(x) + b (1)
xi+1 = xi + Δxi (6)
where ω is the weight vector; b is the bias. i

φ signifies a nonlinear transfer function, mapping input

2( )
vectors into a high-dimension feature space. Here, in theoret- Δxi = ω0 e−𝛾r xj − xi + 𝛼𝜀i (7)
ical terms, a modest LR can deal with an intricate nonlinear
regression of input space (Samantaray and Ghose 2020a). where α is the randomisation coefficient having a value amid
Subsequent convex optimisation problem having an insen- 0 to 1;εi is the arbitrary vector normally found using Gauss-
sitivity loss function (ε) is introduced for obtaining solution: ian distribution.
� � 1 1 ∑n
Minimise RSVMs w, 𝜉 (∗) = ⇑ 𝜔 ⇑2 + C (𝜉 , 𝜉 ∗ )
2 n k=1 k k
(2)

13
Arab J Geosci (2022) 15: 723 Page 5 of 22 723

Fig. 2  Location of observation wells in the study area

13
723 Page 6 of 22 Arab J Geosci (2022) 15: 723

Table 1  Basic statistics of training, testing, and total available data- FOA
sets
Statistical Training set (480) Testing set (120) Total data set Fruitfly is an insect existing broadly in tropical climatic
parameters (600) regions around the world. It is superior to other species in
Sherkhanchawk terms of osphresis and vision. In the course of food hunt-
Min 0.02 1.0090576 0.02 ing, a fruitfly primarily smells a specific odour utilising its
Max 7.913582903 6.986976797 7.913582903 osphresis organs, sends and obtains information from its
Mean 3.581144709 3.320436032 3.529002974 neighbours, and relates fitness with present best position.
Kurt − 0.767513811 − 0.956915936 − 0.764489914 Fitness values are identified by flies using their taste and
Skew 0.372053368 0.417876741 0.393185553 then flying towards location having improved fitness. They
SD 1.789036933 1.586420352 1.752236204 utilise their delicate vision for finding food and fly further in
Basantia high school that direction. Pan (2012) introduced an evolutionary com-
Min 0.056 1.02826832 0.056 putational technique called FOA, based on characteristics of
Max 7.889394405 7.839543172 7.889394405 food finding by fruitfly swarm. Process involved in an FOA
Mean 3.560599631 3.493485156 3.547176736 can be described in the following steps.
Kurt − 0.850979953 − 0.80434208 − 0.847847889
Skew 0.368161153 0.54065221 0.400512799 Step 1: Essential factors for initialisation comprise size of
SD 1.841958962 1.815904858 1.835466534 population (Sizepop), number of iteration (MaxFEs), and
location of preliminary population ­(Xaxis,Yaxis).
Yaxis = rand(),
Yaxis = rand(). (8)
Table 2  Initial parameter settings for the algorithms
Algorithm Parameter Value Step 2: Entities perform global search in arbitrary direc-
FFA Population (𝛽) 50, 100, 150,
tion and arbitrary distance.
200, 250,
Xi = Xaxis + RandValue,
300 (9)
Yi = Yaxis + RandValue.
Attractiveness (𝛽0) 0.9
Light absorption coefficient (𝛾 ) 0.7 Step 3: Compute concentration food’s taste in air.
Randomisation parameter (𝛼) 0.3
√
Number of iterations 500 Dist i = X2i + Y2i (10)
FOA Location range (LR) [0, 1]
maxgen 300
sizepop 30 1
Si =
(Dist i ) (11)
Flight range (FR) [− 10, 10]
ALO Number of antlions 50
Step 4: Decision value of every individual’s position is
Maximum number iterations 100
substituted to fitness function for figuring taste value.

Fig. 3  Monthly groundwater

level at two stations

13
Arab J Geosci (2022) 15: 723 Page 7 of 22 723

Fig. 4  Architecture of conven-

tional SVM

where X(t) is the arbitrary walks of ants, t is the(stage

) of
Smelli = Function (Si ) (12)
arbitrary walk, T is the maximum_Iteration, and r ti is the
Step 5: In the population obtain individual having opti- stochastic function expressed as:
mal value of taste. {
1 if rand > 0.5
[bestSmell, best Index] = min (Smell) (13)
r(t) =
0 if rand ≤ 0.5 (16)

Step 6: All entities utilise their visual organs for flying to where rand is the arbitrary number between 0 and 1. Hunting
best individual location. procedure of ALO can be mathematically explained by the
Smellbest = bestSmell, following equation:
Xaxis = X(bestIndex), (14) ( )
Xi (t) − αi ∗ (di (t) − ci (t))
Yaxis = Y(bestIndex). Xi (t) = + ci (t) (17)
(bi − αi )
Step 7: For iterative search repeat steps 2 to 5. When
concentration of smell reaches present accuracy value or where ­bi, αi is the maximum and minimum of random walk
iterative number reaches maximum iteration number, the in correspondence to ith variable in respective order. ­Di(t)
circulation halts. and ­Ci(t) are the maximum and minimum of ith variables at
ith iteration which can be described as:
ALO ci (t) = Antlionj (t) + c(t) (18)

Mirjalili (2015) first introduced a new optimisation algo-

rithm called the ALO, which mimics hunting behaviour of
di (t) = Antlionj (t) + d(t) (19)
ant-lions (insects). For hunting prey, working procedure of
ALO is divided into five steps. The steps involved are arbi- c(t) = c(t)∕I (20)
trary walk of ants, building traps, entrapment of ants in traps,
hunting prey, and rebuilding traps. ALO has been effectively d(t) = d(t)∕I (21)
used and assessed in various optimization and engineering
problems providing robust performance in comparison to where Antlionj is the location of specified jth antlion at ith
other optimisation algorithms (Dinkar and Deep 2019; Kose iteration. I denotes ratio defined as:
2018; Tharwat and Hassanien 2018; Mouassa et al. 2017). t
Mathematically, various steps involved in ALO mecha- I = 10w (22)
T
nism are described using subsequent equations as follows,
starting with random walk: where w is the constant, which varies on basis of present
[ ( ( ) ) ( ( ) ) ] iteration. After being trapped in antlion’s trap, hunting of
X(t) = 0, cumsum 2r t1 − 1 , cumsum 2r t1 − 1 , … , cumsum(2r(T) − 1)
(15)

13
723 Page 8 of 22 Arab J Geosci (2022) 15: 723

ant by Antlion and reconstructing pit for catching new prey data, utilised for training, and second set, corresponding to
can be defined using the following formula: a period from 2010 to 2019 (20% of the data) was used for
{ testing the network, permitting inside cross-validation on
Ant i (t) if f(Ant
( i (t)) >) f(Antlionj (t)) a part of data set. GWL data (1970–2019) was collected
Antlionj (t) =
Antlionj (t) if f Ant i (t) ≤ f(Antlionj (t)) from central ground water board (CGWB), Eastern Region,
(23) Kolkata. In the present investigation, to forecast GWL utilis-
where ­Anti(t) is the location of specified ith ant at ith ing proposed methods (SVM, SVM-FFA, SVM-FOA, and
iteration. SVM-ALO), various time delays of historic GWL data are
Figures 5, 6, and 7 show the pseudo-code of FFA, FOA, collected from observation wells and taken into considera-
and ALO, respectively. Figure 8 represents a general execution tion as input combination. Number of variables is from 1 to
procedure of SVM-based models optimised by FFA, FOA, and 6 because of statistical lagging effect of monthly and annu-
ALO optimisation techniques. In addition, it shows optimisa- ally varying patterns in historic GWL dataset. Considered
tion processes employed by each optimisation technique, i.e. models based on input combinations are given below:
FFA, FOA, and ALO, and their roles in three hybrid models of
SVM-FFA, SVM-FOA, and SVM-ALO for GWL prediction. M1 (Model 1) ∶ h(t + 1) = f(h(t − 1))
M2 (Model 2) ∶ h(t + 1) = f(h(t − 1), h(t − 2))
Model construction and performance measures M3 (Model 3) ∶ h(t + 1) = f(h(t − 1), h(t − 2), h(t − 3))
M4 (Model 4) ∶ h(t + 1) = f(h(t − 1), h(t − 2), h(t − 3), h(t − 4))
Time series were distributed into two groups: first set cor- M5 (Model 5) ∶ h(t + 1) = f(h(t − 1), h(t − 2), h(t − 3), h(t − 4), h(t − 5))
responds to a period of 1970–2009, which is 80% of total M6 (Model 6) ∶ h(t + 1) = f(h(t − 1), h(t − 12))

Fig. 5  Pseudo-code of FFA

13
Arab J Geosci (2022) 15: 723 Page 9 of 22 723

Fig. 6  Pseudo-code of FOA

13
723 Page 10 of 22 Arab J Geosci (2022) 15: 723

Fig. 7  Pseudo-code of ALO

M7 (Model 7) ∶ h(t + 1) = f(h(t − 1), h(t − 2), h(t − 12)) xi − xmax

M8 (Model 8) ∶ h(t + 1) = f(h(t − 1), h(t − 2), h(t − 3), h(t − 12))
xn =
xmax − xmin (24)
M9 (Model 9) ∶ h(t + 1) = f(h(t − 1), h(t − 2), h(t − 3), h(t − 4), h(t − 12))
M10 (Model 10) ∶ h(t + 1) = f(h(t − 1), h(t − 2), h(t − 3), h(t − 4), h(t − 5), h(t − 12))
where ­Xi is the measured data, ­Xn is the normalised data,
and Xmin and Xmax are the minimum and maximum data val-
Where, h is the average ground water level (monthly) ues, respectively.
(t − 1): Lag of 1 month Subsequently, estimation accuracy of conventional and
(t − 2): Lag of 2 months hybrid SVM models was assessed using five quantita-
(t − 3): Lag of 3 months tive statistical indicators like RMSE (Al-Abduljabbar
(t − 4): Lag of 4 months et al. 2020; Abro et al. 2021; Akpovi et al. 2022), WI
(t − 5): Lag of 5 months (Samantaray and Sahoo 2021; Mohanta et al. 2022), PCC
(t − 12): Lag of 6 months (Gokhale et al. 2021; Singh et al. 2018), R 2 (Agnihotri
For achieving faster and better learning, collected data et al. 2022; Onyutha 2022; Aawar and Khare 2020), and
were normalised between 0 and 1 before giving input into graphical interpretations (scatter plots, time series plot,
proposed models. It also enables us to have a common range box plot, histogram plot). The RMSE, WI, PCC, and R2
and improves modelling performances. Following equation are expressed as follows:
is utilised for normalisation process: �
1 ∑N
RMSE = (f − fo )2 (0 < RMSE < ∞) (25)
N i=1 i

13
Arab J Geosci (2022) 15: 723 Page 11 of 22 723

Prediction of GWL

Data interpretation

Training data set Testing data set

(80%) (20%)

Initialization of
parameter

SVM-FFA Model
SVM-FOA Model
SVM-ALO Model

Fitness evaluation Satisfy terminal condition

Performance evaluation Optimal SVM

of model predictive model

Fig. 8  Flowchart of hybrid models based on SVM optimised by FFA, FOA, and ALO techniques

∑N
k=1
(fo − fi )2 Results
WI = 1 − [ 2
](0 < WI ≤ 1) (26)
∑N � � � �
i=1
(�fi − fo � + �fi − fi �)
� � � � For assessing the impact of various input lag time on per-
formance of prediction models, a primary investigation of
∑N GWL prediction on 2 wells is reported in this section. Based
i=1 (fo − fo )(f i − fi ) on the results of preliminary investigation, a more in depth
PCC = � (−1 < PCC < 1)
∑N 2 ∑N 2 regional analysis was carried out where two observational
(fo − fo ) (fi − fi )
i=1 i=1 wells are situated. For this purpose, hybrid neural network
(27)
models were applied as a nonlinear modelling system for
∑N
i=1
(fi − fo )2 producing a model which can predict GWL based on input
R2 = 1 − − 2
(0 < R2 < 1) (28) combinations. Efficiency of models was assessed by estimat-
∑N
(f − fi )
i=1 i ing RMSE, WI, PCC, NSE, and R2 values. Five performance
measures (Eqs. 25, 26, 27, and 28) were implemented for
where fi is the predicted GWL values assessing performance of proposed SVM models. GWL pre-
fo is the observed GWL values diction results of hybrid SVM models at Sherkhanchawk and
fi is the mean of predicted GWL values Basantia high school wells are presented in Tables 3 and 4.
fo is the mean of observed monthly GWL values It was observed that there was a significant improvement
N is the number of data points in the performance of conventional SVM model when its
Moreover, developed hybrid SVM models with the lowest parameters were optimised utilising ALO in comparison to
RMSE value and the highest value of PCC, NSE, R2, and WI FOA and FFA.
are better performing for monthly GWL estimation. In the testing period, RMSE value was found to be
42.7583–39.5638 for SVM model, while it was ranged
from 38.6627 to 28.7893 for SVM-FFA, from 29.9672 to

13
723 Page 12 of 22 Arab J Geosci (2022) 15: 723

17.3726 for the SVM-FOA models, and 13.2872–7.6638 for Likewise, the results obtained in testing period for Basantia
SVM-ALO models, respectively, at Sherkhanchawk station. high school are provided in Table 4. For ten models, RMSE
Similarly, best values of WI, PCC, and R2 were found to value ranges from 43.5728 to 39.7724 for the SVM model-
be 0.98215, 0.9815, and 0.97846 respectively, with model ling approach, while for SVM-FFA, SVM-FOA, and SVM-
SVM-ALO-M10. On the contrary, best values of WI, PCC, ALO models, it ranged from 39.0275 to 32.3358, 31.6894
and R2 were found with SVM-FFA-M10 (0.9558, 0.95211, to 18.9063, and from 17.0025 to 8.838 in respective order.
and 0.94347) and SVM-FOA-M10 (0.97031, 0.96891, and Best values of WI, PCC, and R2 were found to be 0.98067,
0.96223). 0.98001, and 0.97628, respectively, considering model SVM-
ALO-M10. On the contrary, best values of RMSE, WI, PCC,

Table 3  Performance of models Algorithm Model RMSE WI PCC R2 RMSE WI PCC R2

at Sherkhanchawk station
SVM M1 39.2186 0.93507 0.93384 0.9283 42.7583 0.91702 0.91573 0.90649
M2 38.8236 0.93895 0.93682 0.93006 41.9965 0.91887 0.91658 0.90885
M3 37.9901 0.94103 0.93801 0.93318 41.8836 0.91924 0.91684 0.90906
M4 37.0126 0.9437 0.94082 0.93471 41.2905 0.92156 0.91896 0.9118
M5 36.2854 0.94701 0.9436 0.93685 41.0863 0.92281 0.91989 0.91253
M6 35.893 0.94806 0.94482 0.93764 40.7082 0.92546 0.92189 0.91427
M7 34.5052 0.95036 0.94784 0.9399 40.6185 0.92589 0.92305 0.91552
M8 30.2285 0.95362 0.95088 0.94247 40.3904 0.92787 0.92564 0.9171
M9 27.3493 0.95662 0.95426 0.94602 39.8905 0.93084 0.92892 0.91874
M10 25.9095 0.95839 0.95667 0.94886 39.5638 0.93382 0.93197 0.92074
SVM-FFA M1 24.8867 0.96005 0.95802 0.9504 38.6627 0.93897 0.93748 0.9301
M2 23.4523 0.96266 0.96049 0.95229 37.6627 0.94229 0.93995 0.93396
M3 22.8903 0.96384 0.96101 0.95304 37.4834 0.9427 0.94003 0.9341
M4 21.001 0.9657 0.9638 0.95542 36.6729 0.94505 0.94224 0.93599
M5 20.9212 0.96667 0.96405 0.95686 36.0287 0.94775 0.94378 0.93703
M6 19.3256 0.96927 0.96705 0.95993 35.1007 0.94978 0.9459 0.93864
M7 17.5738 0.97008 0.96847 0.9621 34.2753 0.95104 0.9479 0.93995
M8 15.0025 0.9744 0.97286 0.9685 32.8795 0.95173 0.94872 0.94071
M9 14.9827 0.97449 0.9729 0.96967 30.7859 0.95317 0.95082 0.9422
M10 12.002 0.97728 0.9758 0.97118 28.7893 0.9558 0.95211 0.94347
SVM-FOA M1 13.9063 0.97497 0.97303 0.96993 29.9672 0.95385 0.95164 0.94265
M2 12.5331 0.97681 0.97539 0.97039 29.2186 0.95494 0.95207 0.94304
M3 10.9273 0.97854 0.97685 0.9726 27.6718 0.95617 0.95379 0.94582
M4 9.2664 0.98001 0.97892 0.97432 26.3214 0.95806 0.95623 0.9477
M5 8.1648 0.9818 0.98017 0.97764 25.0912 0.95896 0.95745 0.94994
M6 8.005 0.9819 0.98053 0.97805 23.9091 0.9618 0.95964 0.95129
M7 7.3742 0.98326 0.9818 0.97911 20.0031 0.96784 0.96553 0.9583
M8 7.0327 0.98379 0.98308 0.97948 19.5289 0.96858 0.96599 0.95991
M9 7.0023 0.98379 0.98427 0.98003 18.2803 0.96952 0.96782 0.96038
M10 6.7826 0.98451 0.98546 0.9817 17.3726 0.97031 0.96891 0.96223
SVM-ALO M1 6.016 0.98583 0.98662 0.98274 13.2872 0.97538 0.97385 0.96998
M2 5.2996 0.9878 0.98704 0.98431 13.037 0.97582 0.97402 0.97001
M3 5.001 0.98942 0.98826 0.98497 12.1907 0.97717 0.97548 0.97064
M4 4.8328 0.9895 0.98843 0.98534 11.396 0.97774 0.97602 0.97189
M5 3.4834 0.99047 0.98976 0.98698 11.0037 0.97815 0.97602 0.97225
M6 3.008 0.99102 0.9901 0.98742 10.7822 0.97884 0.97703 0.97297
M7 2.648 0.99275 0.99098 0.9883 9.9951 0.97907 0.9774 0.9736
M8 1.743 0.99413 0.99224 0.98906 9.2546 0.98014 0.97906 0.97443
M9 0.1746 0.99516 0.99352 0.99189 8.9912 0.98054 0.97993 0.97572
M10 0.0891 0.99627 0.99543 0.99325 7.6638 0.98215 0.9815 0.97846

13
Arab J Geosci (2022) 15: 723 Page 13 of 22 723

Table 4  Performance of models Algorithm Model RMSE WI PCC R2 RMSE WI PCC R2

at Basantia high school station
SVM M1 39.4275 0.9349 0.93304 0.92683 43.5728 0.91683 0.91345 0.9034
M2 38.9994 0.93774 0.9359 0.92997 42.8806 0.91702 0.91479 0.90582
M3 38.4905 0.93952 0.9376 0.93102 42.5734 0.91883 0.91574 0.90702
M4 37.8552 0.94185 0.93868 0.93348 41.6507 0.92001 0.91735 0.91006
M5 36.9921 0.94401 0.94117 0.93512 41.402 0.92107 0.91773 0.9117
M6 36.3986 0.94556 0.94284 0.93669 40.9973 0.92325 0.91993 0.91285
M7 35.208 0.94978 0.94557 0.93807 40.8079 0.92405 0.92008 0.91423
M8 31.0385 0.95273 0.95037 0.9416 40.5588 0.92773 0.92475 0.91689
M9 28.3785 0.95583 0.95219 0.94438 40.1867 0.92984 0.92696 0.9175
M10 26.675 0.9579 0.95589 0.94681 39.7724 0.93289 0.93058 0.92014
SVM-FFA M1 25.5289 0.95895 0.95712 0.94991 39.0275 0.93683 0.93509 0.92893
M2 24.1906 0.9612 0.95925 0.9509 38.2892 0.93996 0.93785 0.9314
M3 23.7746 0.96206 0.9604 0.95195 38.1263 0.9406 0.93796 0.93261
M4 22.5226 0.96405 0.96211 0.9536 37.2182 0.94368 0.94006 0.93447
M5 21.5745 0.96415 0.96213 0.95508 36.8095 0.94468 0.94196 0.9354
M6 20.2732 0.96703 0.96476 0.95824 36.462 0.94521 0.94278 0.93618
M7 17.9736 0.97007 0.96846 0.96047 35.6579 0.94867 0.94486 0.93799
M8 16.9027 0.97108 0.97004 0.96359 34.9526 0.95002 0.94629 0.93982
M9 16.3298 0.97214 0.97058 0.96601 33.4832 0.95107 0.94807 0.94005
M10 15.8469 0.97315 0.97147 0.96774 32.3358 0.9519 0.94893 0.94076
SVM-FOA M1 15.5375 0.97326 0.97229 0.9682 31.6894 0.95219 0.94987 0.94102
M2 13.7723 0.9752 0.97347 0.96996 29.487 0.95473 0.95168 0.94285
M3 11.7382 0.97773 0.97598 0.97147 28.0014 0.95593 0.95362 0.94447
M4 9.3867 0.97993 0.9785 0.97398 26.9971 0.95682 0.95488 0.94626
M5 9.0896 0.98018 0.97993 0.9755 25.8856 0.95882 0.9568 0.9497
M6 8.4639 0.98106 0.98004 0.97673 24.4702 0.9605 0.95809 0.95084
M7 7.929 0.98213 0.98066 0.97812 21.1893 0.9652 0.96346 0.95531
M8 7.5287 0.98276 0.98164 0.97874 20.6283 0.96697 0.96406 0.95749
M9 7.1043 0.98365 0.98279 0.97936 19.8822 0.96822 0.9659 0.95983
M10 6.991 0.98404 0.98438 0.98009 18.9063 0.9693 0.9671 0.9602
SVM-ALO M1 6.1794 0.98572 0.9858 0.98214 17.0025 0.97074 0.96932 0.9635
M2 5.8902 0.98691 0.98673 0.98335 16.7386 0.97156 0.97007 0.96483
M3 5.564 0.98768 0.98703 0.98419 16.537 0.97208 0.97007 0.96502
M4 5.2785 0.98885 0.98792 0.98476 16.1182 0.97281 0.9711 0.9669
M5 4.3744 0.98996 0.98853 0.98593 15.2583 0.9735 0.97279 0.96836
M6 4.073 0.9904 0.98953 0.98692 14.4561 0.97494 0.97302 0.9699
M7 2.9961 0.99106 0.99092 0.98805 12.8903 0.9765 0.97416 0.97004
M8 2.2867 0.99368 0.99187 0.98897 11.9083 0.97763 0.97586 0.97125
M9 1.1994 0.99491 0.9924 0.99074 9.6536 0.97991 0.97843 0.97381
M10 0.1005 0.99614 0.995 0.9921 8.838 0.98067 0.98001 0.97628

and R2 obtained with SVM-FFA-M10 and SVM-FOA-M10 are be understood from Tables 3 and 4 that SVM-ALO performs
32.3358, 0.9519, 0.94893, and 0.94076 and 18.9063, 0.9693, best for all models (different input combinations) with mini-
0.9671, and 0.9602 respectively. Best performing models dur- mum RMSE value and SVM showing the worst performance
ing testing phase at Sherkhanchawk and Basantia high school in GWL prediction. Based on Tables 3 and 4, it is evident
stations are highlighted in bold in Tables 3 and 4. that SVM-ALO contributes to higher prediction accuracy
Robustness of SVM-ALO model was assessed in regional than SVM-FOA and SVM-FFA algorithms. The overall
investigation, attaining suitable values of statistical indica- computational outcomes demonstrate that SVM-ALO is
tors in training and in testing phase that proved to be the more efficient compared to other hybrid and conventional
most challenging circumstances for GWL prediction. It can models tested in monthly GWL prediction. It is evident that

13
723 Page 14 of 22 Arab J Geosci (2022) 15: 723

as the time delay increases prediction accuracy of SVM- models (Fig. 9). Closeness of R2 value to 1 is a suggestion
ALO models (M1–M10) increases. Hence, SVM-ALO has that both observed and predicted values are fitted well.
proven to be a reliable tool to investigate fluctuations in From Fig. 9, it is observed that SVM-ALO model fits the
wells positioned on various hydrogeological arrangements. data well with precise GWL predictions. High value of
In particular, sub-regional investigation emphasised differ- R2 and less value of RMSE obtained by SVM-ALO model
ences among wells located in arid regions. indicate that data are well fitted. Outcomes found from
Figure 9 illustrates the coefficient of determination assessment of data indicated that SVM-ALO model is the
(R2) plot for the best performing model of four proposed best fitting for GWL prediction. This recommends that
techniques at Sherkhanchawk and Basantia high school. proposed SVM-ALO model can be used in other regions
Firstly, R2 values for four applied algorithms in ten dif- of similar geology.
ferent scenarios/models were found. It is observed that Figure 10 demonstrates plot between observed and simu-
R 2 for SVM-ALO-M10 model is superior to other nine lated GWLs for the best results computed with GWL data

(a)

Fig. 9  Regression plots of predicted over observed GWL data using SVM, SVM-FFA, SVM-FOA, and SVM-ALO at a Sherkhanchawk and b
Basantia high school stations for testing phases

13
Arab J Geosci (2022) 15: 723 Page 15 of 22 723

(b)

Fig. 9  (continued)

by SVM, SVM-FFA, SVM-FOA, and SVM-ALO in two Comparison between observed and estimated GWL for
observational wells. However, utilised datasets were differ- two reference wells are presented using the box plot repre-
ent for training and testing. Outcomes reveal that simulated sentation. The best performance was obtained for “SVM-
peak GWL is 8.562671 m, 8.774913 m, 8.949114 m, and ALO,” and SVM-FOA gives slightly lower predictions but
9.099785 m for SVM, SVM-FFA, SVM-FOA, and SVM- performs pre-eminently for two reference wells. Figure 11
ALO, contrary to observed peak of 9.30068 m for Sherkhan- shows the boxplot comparison between results obtained by
chawk station. The estimated peak GWL is 7.982215 m, SVM, SVM-FFA, SVM-FOA, and SVM-ALO algorithms
8.161093 m, 8.329735 m, and 8.468535 m for SVM, SVM- at both well stations. Representations indicated that SVM-
FFA, SVM-FOA, and SVM-ALO against observed peak of ALO was the most accurate model illustrating the effective-
8.675 m for Basantia high school division. Time series plots ness of applied ALO algorithm. The reason behind this is
of observed versus simulated GWL for proposed models at due to improvement on updating mechanism and combina-
specified locations are presented in Fig. 10. tion of tournament selection of ALO. It is also seen from

13
723 Page 16 of 22 Arab J Geosci (2022) 15: 723

Fig. 10  Prediction perfor-

mances of SVM, SVM-FFA,
SVM-FOA, and SVM-ALO at a
Sherkhanchawk and b Basantia
high school stations

(a)

(b)

Fig. 10 that this fluctuation is negligible. Several studies discussed earlier, we consider ten models for each applied
reveal that the impact of arbitrary collection of initial popu- technique based on different input combinations. It is quite
lation and investigation space is negligible in meta-heuristic evident from Fig. 13 that for all models (M1–M10), SVM-
algorithms. ALO shows the best results compared to other hybrid and
Figure 12 illustrates that the histogram plot of GWL esti- simple techniques, with model (SVM-ALO-M10) showing
mated with hybrid SVM models. Based on the assessment of the highest accuracy in groundwater level prediction at both
the above results, it is indicated that specified hybrid SVM- wells. There are several reasons for the selection and bet-
ALO model provides better and consistent predictions of ter performance of ALO algorithm over FOA and FFA: a
average GWL of proposed well stations. Also, it shows reli- small number of tuning parameters, appropriate for various
able results owing to low information content and stand- kinds of optimisation problems, simple to hybridise with
ard deviation values in comparison to other models. Even data-driven techniques, simple implementation, fast conver-
though average GWL is a primary and suitable standard for gence speed, accelerated process to find optimal solutions,
assessment of variations all over the plain, spatial variations lesser chance of getting stuck in local optima, efficient global
across piezometer network must be known across the plain. process with extensive search space making it more feasible,
For that purpose, hybrid models have been employed for and efficient in getting global optima and reasonable execu-
simulating GWL fluctuations in all wells. tion time.
Comparison of various model with respect to proposed
algorithms at two well stations are given in Fig. 13. As

13
Arab J Geosci (2022) 15: 723 Page 17 of 22 723

Fig. 11  Comparison of box plot at station a Sherkhanchawk and b Basantia high school

Discussions of groundwater, its shortage, and inadequate figures for

budget occurred. Such circumstances call for application
Assortment of fitting estimators is vital for enhancing pre- of hybrid SVM approaches for making precise predictions
diction performances (Quilty et al. 2016; Maier and Dandy regarding GWL variation and for improved management
2000; Galelli et al. 2014). Therefore, a robust prediction of water reserves in the region besides providing suitable
model can offer an effective feature selection method utilis- tools to formulate operational strategies. However, it must
ing optimisation techniques for selecting appropriate predic- be assured that accurate GWL forecasting is vital for sustain-
tion variables. In addition, an effective tuning procedure on able and better controlling of groundwater storage around
basis of bio-inspired optimisation algorithms to obtain opti- the world. If integration of FFA, FOA, and ALO algorithms
mal model parameters in training period is one more benefit with SVM is appropriately done, it can assist water resource
of applying hybrid ML models in comparison to conven- planners and decision makers to attain precise results. This
tional ones (Sharafati et al. 2020; Salih et al. 2020; Banad- has turned out to be enormously significant due to inad-
kooki et al. 2020). Taking another perspective concerning equate global water reserves and enormous water demand
different ML techniques to simulate GWL, an assumption in agricultural, industrial, and domestic needs. Present work
can be made that one specific form is not factually recom- has applied time series data to examine fluctuations in GWL
mended of ML models for a specific problem. Also, it is evi- and arrangements of such variations. This suggests a per-
dent that a robust integrated model will perform much better ception of natural hydro-geological and hydrological cycles.
than a stand-alone ML model (Rajaee et al. 2019; Zare and Any changes occurring in GWL and any variations in natural
Koch 2018; Seidu et al. 2021; Rezaei et al. 2021). Results cycles caused due to external aspects (like extreme usage of
of the proposed models demonstrate correspondence with groundwater reserves, lack of natural water restoration, and
different studies, which suggested that hybrid ML techniques others) may be recognised by utilising these methods.
performed better than standalone techniques in improving Hence, it may be specified that using an integration
accurateness of forecasts (Yadav and Eliza 2017; Gupta and of neural networks and optimisation algorithms effec-
Barua 2018; Malik et al. 2020; Cui et al. 2021; Samantaray tively predicted GWL on different lag time combinations.
and Sahoo 2021). Furthermore, performed forecasting can assist water
Recently, with rapid consumption of existing groundwater resource planners and managers to appropriately strate-
reserves because of imbalance between supply and demand, gize on enhancing areas of water resources to a higher
India is facing a huge issue of groundwater declination. extent. Indeed, prediction models are significant tools for
Excess usage of groundwater in Purba Medinipur area led showing decrease in GWL at the time of droughts. Correct
to low GWL. During last decade, GWL even dropped in sev- predictions by decision-makers allow them to control and
eral regions; therefore, problems such as poor management manage basins and watersheds. It should be deliberated

13
723 Page 18 of 22 Arab J Geosci (2022) 15: 723

Fig. 12  Histogram plot of algo-

rithms at station a Sherkhan-
chawk and b Basantia high
school

(a)

(b)

that several climatic factors can have an impact on GWL. rivers can be investigated for other case studies where
Integration of data driven techniques with pre-processing these would impact GWLs.
data can ascertain value and time of different GWLs. Man-
agement of groundwater with precise prediction results
assists decision makers in allocating water equitably to Conclusions
stakeholders. In addition, groundwater and its index can
be utilised to predict drought periods. While investigat- For sustainable water resources management, prediction
ing these outcomes, it must be noted that, because models of groundwater level fluctuation plays a crucial role.
are developed considering water level from wells with no On the other hand, estimating precise GWL forecasts
observation of either water bodies or streams, the models is mostly challenging because of nonlinear relationship
seem to estimate drier scenarios than anticipated. Future between descriptive variables and GWL and their multi-
work may focus more on implementing an integrated scale behaviour, which varies with time. For developing
monitoring design for evaluating surface and groundwater an accurate ML model, one of the prerequisites is the
resources using field and remotely sensed data. Moreover, selection of optimal input variables as well as optimising
the influence of variables such as seasonal precipitation, model parameters. To address these problems, present
pumping from agricultural wells, and runoff from seasonal research investigated the capability of meta-heuristic opti-
mization algorithms (such as ALO, FOA, and FFA) that

13
Arab J Geosci (2022) 15: 723 Page 19 of 22 723

Fig. 13  Comparison of models

at a Sherkhanchawk and b
Basantia high school during
training and testing phase

(a)

(b)

can internally perform selection of optimal input vari- statistical indices and graphical interpretation. Compari-
ables. Proposed hybrid models capable of capturing non- son of results indicated that SVM-ALO model performed
linear relationships between model inputs were provided superiorly than SVM-FOA, SVM-FFA, and simple SVM
with multiple information to enhance the forecasting abil- models in all models at both stations. Consequently, the
ity of the proposed models. most efficient model was SVM-ALO-M10, with the low-
Present study aimed to examine accurateness of a meta- est value of RMSE = 7.6638/8.838 and the highest value
heuristic algorithm, ALO, for enhancing performance of of NSE = 0.98006/0.97896, PCC = 0.9815/0.98001, and
SVM in monthly GWL estimation at coastal region of WI = 0.98215/0.98067 for testing period at Sherkhan-
Purba Medinipur district located in West Bengal state of chawk/Basantia high school, respectively. The hydrolo-
India. A comparison was made between proposed algo- gists, agronomists, and agriculturists can construct a
rithm against FOA and FFA to assess its prediction per- truthful smooth decision support system with obtained
formance. Developed hybrid SVM-ALO, SVM-FOA, and outcomes of hybrid SVM-ALO model for precise predic-
SVM-FFA models were trained and tested by exploiting tion of GWL in the selected study area.
ten different input combinations. Results of SVM-ALO
model were appraised and compared with SVM-FOA,
SVM-FFA, and conventional SVM models on basis of

13
723 Page 20 of 22 Arab J Geosci (2022) 15: 723

Declarations Dinkar SK, Deep K (2019) Accelerated opposition-based antlion opti-

mizer with application to order reduction of linear time-invariant
systems. Arab J Sci Eng 44:2213–2241. https://​doi.​org/​10.​1007/​
Ethics approval All authors are equally contributed for development
s13369-​018-​3370-4
of model and preparation of manuscripts.
Ebrahimi H, Rajaee T (2017) Simulation of groundwater level varia-
tions using wavelet combined with neural network, linear regres-
Conflict of interest The authors declare no competing interests. sion and support vector machine. Glob Planet Change 148:181–
191. https://​doi.​org/​10.​1016/j.​glopl​acha.​2016.​11.​014
Ehteram M, Afan HA, Dianatikhah M et al (2019) Assessing the pre-
dictability of an improved ANFIS model for monthly streamflow
References using lagged climate indices as predictors. Water 11:1130. https://​
doi.​org/​10.​3390/​w1106​1130
Aawar T, Khare D (2020) Assessment of climate change impacts on Ehteram M, Singh VP, Ferdowsi A et al (2019) An improved model
streamflow through hydrological model using SWAT model: a based on the support vector machine and cuckoo algorithm for
case study of Afghanistan. Model Earth Syst Environ 6:1427– simulating reference evapotranspiration. PLoS ONE 14:e0217499.
1437. https://​doi.​org/​10.​1007/​s40808-​020-​00759-0 https://​doi.​org/​10.​1371/​journ​al.​pone.​02174​99
Abro MI, Zhu D, Elahi E, Majidano AA, Solangi BK (2021) Hydro- Ferreira LB, da Cunha FF (2020) New approach to estimate daily refer-
logical simulation using multi-sources precipitation estimates in ence evapotranspiration based on hourly temperature and relative
the Huaihe River Basin. Arab J Geosci 14:1–12. https://​doi.​org/​ humidity using machine learning and deep learning. Agric Water
10.​1007/​s12517-​021-​08254-1 Manag 234:106113. https://​doi.o​ rg/1​ 0.1​ 016/j.​agwat.2​ 020.​106113
Agnihotri A Sahoo A Diwakar MK (2022) Flood prediction using Galelli S, Humphrey GB, Maier HR et al (2014) An evaluation frame-
hybrid ANFIS-ACO model: a case study. In Inventive Computa- work for input variable selection algorithms for environmental
tion and Information Technologies (pp. 169–180). Springer, Sin- data-driven models. Environ Model Softw 62:33–51. https://​doi.​
gapore. https://​doi.​org/​10.​1007/​978-​981-​16-​6723-7_​13 org/​10.​1016/j.​envso​ft.​2014.​08.​015
Akpovi BA, Zhu D, Abro MI, Lawin AE, Houngnibo M, Bessou J Ghorbani MA, Shamshirband S, Haghi DZ et al (2017) Application
(2022) Hydrological appraisal using multi-source rainfall data in of firefly algorithm-based support vector machines for prediction
PDM model over the Qinhuai River basin in China. Arab J Geosci of field capacity and permanent wilting point. Soil Tillage Res
15:1–14. https://​doi.​org/​10.​1007/​s12517-​022-​09545-x 172:32–38. https://​doi.​org/​10.​1016/j.​still.​2017.​04.​009
Al-Abduljabbar A, Gamal H, Elkatatny S (2020) Application of artifi- Gokhale AA, Dobhal DP, Nainwal HC (2021) Source characterization
cial neural network to predict the rate of penetration for S-shape of suspended sediments transported from debris-covered Chora-
well profile. Arab J Geosci 13:1–11. https://​doi.​org/​10.​1007/​ bari Glacier in Central Himalaya, India. Arab J Geosci 14:1–16.
s12517-​020-​05821-w https://​doi.​org/​10.​1007/​s12517-​021-​08474-5
Al-Shammari ET, Mohammadi K, Keivani A et al (2016) Prediction Granata F (2019) Evapotranspiration evaluation models based on
of daily dewpoint temperature using a model combining the sup- machine learning algorithms—a comparative study. Agric Water
port vector machine with firefly algorithm. J Irrig Drain Eng Manag 217:303–315. https://d​ oi.o​ rg/1​ 0.1​ 016/j.a​ gwat.2​ 019.0​ 3.0​ 15
142:4016013. https://​doi.​org/​10.​1061/​(ASCE)​IR.​1943-​4774.​ Gupta H, Barua MK (2018) A novel hybrid multi-criteria method for
00010​15 supplier selection among SMEs on the basis of innovation ability.
Banadkooki FB, Ehteram M, Ahmed AN et al (2020) Enhancement of Int J Logist Res Appl 21:201–223. https://​doi.​org/​10.​1080/​13675​
groundwater-level prediction using an integrated machine learning 567.​2017.​13824​57
model optimized by whale algorithm. Nat Resour Res 29:3233– Hosseini SM, Mahjouri N (2016) Integrating support vector regression
3252. https://​doi.​org/​10.​1007/​s11053-​020-​09634-2 and a geomorphologic artificial neural network for daily rainfall-
Behzad M, Asghari K, Coppola EA Jr (2010) Comparative study of runoff modeling. Appl Soft Comput 38:329–345. https://​doi.​org/​
SVMs and ANNs in aquifer water level prediction. J Comput Civ 10.​1016/j.​asoc.​2015.​09.​049
Eng 24:408–413. https://​doi.​org/​10.​1061/​(ASCE)​CP.​1943-​5487.​ Huang F, Huang J, Jiang S-H, Zhou C (2017) Prediction of ground-
00000​43 water levels using evidence of chaos and support vector machine.
Cao Y, Yin K, Zhou C, Ahmed B (2020) Establishment of landslide J Hydroinformatics 19:586–606. https://​doi.​org/​10.​2166/​hydro.​
groundwater level prediction model based on GA-SVM and influ- 2017.​102
encing factor analysis. Sensors 20:845. https://​doi.​org/​10.​3390/​ Jalalkamali A, Sedghi H, Manshouri M (2011) Monthly groundwater
s2003​0845 level prediction using ANN and neuro-fuzzy models: a case study
Chen Y Gu C Shao C et al (2020) An approach using adaptive weighted on Kerman plain. Iran J Hydroinformatics 13:867–876. https://d​ oi.​
least squares support vector machines coupled with modified ant org/​10.​2166/​hydro.​2010.​034
lion optimizer for dam deformation prediction. Math Probl Eng Karbasi M (2018) Forecasting of multi-step ahead reference evapo-
2020 https://​doi.​org/​10.​1155/​2020/​94340​65 transpiration using wavelet-Gaussian process regression model.
Cui F Al-Sudani ZA Hassan GS Afan HA Ahammed SJ Yaseen ZM Water Resour Manag 32:1035–1052. https://​doi.​org/​10.​1007/​
(2021) Boosted artificial intelligence model using improved alpha- s11269-​017-​1853-9
guided grey wolf optimizer for groundwater level prediction: com- Kisi O, Cimen M (2011) A wavelet-support vector machine conjunc-
parative study and insight for federated learning technology. J tion model for monthly streamflow forecasting. J Hydrol 399:132–
Hydrol :127384. https://​doi.​org/​10.​1016/j.​jhydr​ol.​2021.​127384 140. https://​doi.​org/​10.​1016/j.​jhydr​ol.​2010.​12.​041
Dash NB, Panda SN, Remesan R, Sahoo N (2010) Hybrid neural Konikow LF, Kendy E (2005) Groundwater depletion: a global
modeling for groundwater level prediction. Neural Comput Appl problem. Hydrogeol J 13:317–320. https://​d oi.​o rg/​1 0.​1 007/​
19:1251–1263. https://​doi.​org/​10.​1007/​s00521-​010-​0360-1 s10040-​004-​0411-8
Das UK, Samantaray S, Ghose DK, Roy P (2019) Estimation of aquifer Kose U (2018) An ant-lion optimizer-trained artificial neural network
potential using BPNN, RBFN, RNN, and ANFIS. Smart Intel- system for chaotic electroencephalogram (EEG) prediction. Appl
ligent Computing and Applications, Springer, Singapore, pp Sci 8:1613. https://​doi.​org/​10.​3390/​app80​91613
569–576 Maier HR, Dandy GC (2000) Neural networks for the prediction and
forecasting of water resources variables: a review of modelling

13
Arab J Geosci (2022) 15: 723 Page 21 of 22 723

issues and applications. Environ Model Softw 15:101–124. groundwater level. Arab J Geosci 14:1–15. https://​doi.​org/​10.​
https://​doi.​org/​10.​1016/​S1364-​8152(99)​00007-9 1007/​s12517-​021-​07349-z
Malekzadeh M, Kardar S, Saeb K et al (2019) A novel approach Salih SQ, Sharafati A, Ebtehaj I et al (2020) Integrative stochastic
for prediction of monthly ground water level using a hybrid model standardization with genetic algorithm for rainfall pattern
wavelet and non-tuned self-adaptive machine learning model. forecasting in tropical and semi-arid environments. Hydrol Sci J
Water Resour Manag 33:1609–1628. https://​doi.​org/​10.​1007/​ 65:1145–1157. https://​doi.​org/​10.​1080/​02626​667.​2020.​17348​13
s11269-​019-​2193-8 Samadianfard S, Jarhan S, Salwana E et al (2019) Support vector
Malik A, Tikhamarine Y, Souag-Gamane D et al (2020) Support vec- regression integrated with fruit fly optimization algorithm for
tor regression optimized by meta-heuristic algorithms for daily river flow forecasting in Lake Urmia Basin. Water 11:1934.
streamflow prediction. Stoch Environ Res Risk Assess 34:1755– https://​doi.​org/​10.​3390/​w1109​1934
1773. https://​doi.​org/​10.​1007/​s00477-​020-​01874-1 Samantaray S, Ghose DK (2020a) Modelling runoff in an arid water-
Mattar MA (2018) Using gene expression programming in monthly shed through integrated support vector machine. H2Open Journal,
reference evapotranspiration modeling: a case study in Egypt. IWA Publishing 3(1):256–275
Agric Water Manag 198:28–38. https://​doi.​org/​10.​1016/j.​agwat.​ Samantaray S, Ghose DK (2020b) Assessment of suspended sediment
2017.​12.​017 load with neural networks in arid watershed. Journal of The Insti-
Mehr AD, Nourani V, Khosrowshahi VK, Ghorbani MA (2019) A tution of Engineers (India): Series A 101:371–380
hybrid support vector regression–firefly model for monthly rain- Samanataray S, Sahoo A (2021) A comparative study on prediction
fall forecasting. Int J Environ Sci Technol 16:335–346. https://d​ oi.​ of monthly streamflow using hybrid ANFIS-PSO approaches.
org/​10.​1007/​s13762-​018-​1674-2 KSCE J Civ Eng 25:4032–4043. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 007/​
Mirjalili S (2015) The ant lion optimizer. Adv Eng Softw 83:80–98. s12205-​021-​2223-y
https://​doi.​org/​10.​1016/j.​adven​gsoft.​2015.​01.​010 Samantaray S, Sahoo A, Ghose DK (2020) Assessment of sediment
Moayedi H, Tien Bui D, Anastasios D, Kalantar B (2019) Spotted load concentration using SVM, SVM-FFA and PSR-SVM-
hyena optimizer and ant lion optimization in predicting the shear FFA in arid watershed, India: a case study. KSCE J Civ Eng
strength of soil. Appl Sci 9:4738. https://​doi.​org/​10.​3390/​app92​ 24:1944–1957
24738 Samantaray S, Sahoo A (2020) Assessment of sediment concentration
Moazenzadeh R, Mohammadi B, Shamshirband S, Chau K (2018) Cou- through RBNN and SVM-FFA in Arid Watershed, India. In Smart
pling a firefly algorithm with support vector regression to pre- intelligent computing and applications (pp. 701–709). Springer,
dict evaporation in northern Iran. Eng Appl Comput Fluid Mech Singapore. https://​doi.​org/​10.​1007/​978-​981-​13-​9282-5_​67
12:584–597. https://​doi.​org/​10.​1080/​19942​060.​2018.​14824​76 Samantaray S Sahoo A (2021) Prediction of suspended sediment con-
Mohanta NR, Panda SK, Singh UK, Sahoo A, Samantaray S (2022) centration using hybrid SVM-WOA approaches. Geocarto Int
MLP-WOA is a successful algorithm for estimating sediment load :1-27. https://​doi.​org/​10.​1080/​10106​049.​2021.​192063
in Kalahandi Gauge Station, India. In Proceedings of International Samantaray S Sahoo A Ghose DK (2019) June. Assessment of ground-
Conference on Data Science and Applications (pp. 319–329). water potential using neural network: a case study. In International
Springer, Singapore. https://​doi.​org/​10.​1007/​978-​981-​16-​5120- Conference on Intelligent Computing and Communication (pp.
5_​25 655–664). Springer, Singapore. https://​doi.​org/​10.​1007/​978-​981-​
Mohanty S, Jha MK, Raul SK et al (2015) Using artificial neural net- 15-​1084-7_​63
work approach for simultaneous forecasting of weekly groundwa- Samantaray S Sahoo A Ghose DK (2020a) Infiltration loss affects
ter levels at multiple sites. Water Resour Manag 29:5521–5532. toward groundwater fluctuation through CANFIS in arid water-
https://​doi.​org/​10.​1007/​s11269-​015-​1132-6 shed: a case study. In Smart Intelligent Computing and Applica-
Moosavi V, Vafakhah M, Shirmohammadi B, Behnia N (2013) A wave- tions (pp. 781–789). Springer, Singapore. https://d​ oi.o​ rg/1​ 0.1​ 007/​
let-ANFIS hybrid model for groundwater level forecasting for dif- 978-​981-​13-​9282-5_​76
ferent prediction periods. Water Resour Manag 27:1301–1321. Samantaray S, Tripathy O, Sahoo A, Ghose DK (2020c) Rainfall fore-
https://​doi.​org/​10.​1007/​s11269-​012-​0239-2 casting through ANN and SVM in Bolangir Watershed, India.
Mouassa S, Bouktir T, Salhi A (2017) Ant lion optimizer for solving In Smart intelligent computing and applications (pp. 767–774).
optimal reactive power dispatch problem in power systems. Eng Springer, Singapore. https://​doi.​org/​10.​1007/​978-​981-​13-​9282-
Sci Technol an Int J 20:885–895. https://​doi.​org/​10.​1016/j.​jestch.​ 5_​74
2017.​03.​006 Sanikhani H, Deo RC, Samui P et al (2018) Survey of different data-
Nie S, Bian J, Wan H et al (2017) Simulation and uncertainty analysis intelligent modeling strategies for forecasting air temperature
for groundwater levels using radial basis function neural network using geographic information as model predictors. Comput Elec-
and support vector machine models. J Water Supply Res Technol tron Agric 152:242–260. https://​doi.​org/​10.​1016/j.​compag.​2018.​
66:15–24. https://​doi.​org/​10.​2166/​aqua.​2016.​069 07.​008
Onyutha C (2022) A hydrological model skill score and revised Sattari MT, Mirabbasi R, Sushab RS, Abraham J (2018) Prediction of
R-squared. Hydrol Res 53:51–64. https://​doi.​org/​10.​2166/​nh.​ groundwater level in Ardebil plain using support vector regression
2021.​071 and M5 tree model. Groundwater 56:636–646. https://​doi.​org/​10.​
Pan W-T (2012) A new fruit fly optimization algorithm: taking the 1111/​gwat.​12620
financial distress model as an example. Knowledge-Based Syst Seidu J Ewusi A Kuma JSY Ziggah YY Voigt HJ (2021) A hybrid
26:69–74. https://​doi.​org/​10.​1016/j.​knosys.​2011.​07.​001 groundwater level prediction model using signal decomposition
Quilty J, Adamowski J, Khalil B, Rathinasamy M (2016) Bootstrap and optimised extreme learning machine. Model Earth Syst Envi-
rank-ordered conditional mutual information (broCMI): a nonlin- ron :1-18. https://​doi.​org/​10.​1007/​s40808-​021-​01319-w
ear input variable selection method for water resources modeling. Sharafati A, Tafarojnoruz A, Shourian M, Yaseen ZM (2020) Simula-
Water Resour Res 52:2299–2326. https://​doi.​org/​10.​1002/​2015W​ tion of the depth scouring downstream sluice gate: the validation
R0169​59 of newly developed data-intelligent models. J Hydro Environ Res
Rezaei M, Mousavi SF, Moridi A, EshaghiGordji M, Karami H 29:20–30. https://​doi.​org/​10.​1016/j.​jher.​2019.​11.​002
(2021) A new hybrid framework based on integration of optimi- Shirmohammadi B, Vafakhah M, Moosavi V, Moghaddamnia A (2013)
zation algorithms and numerical method for estimating monthly Application of several data-driven techniques for predicting

13
723 Page 22 of 22 Arab J Geosci (2022) 15: 723

groundwater level. Water Resour Manag 27:419–432. https://​doi.​ combination of random features. Appl Water Sci 8:1–12. https://​
org/​10.​1007/​s11269-​012-​0194-y doi.​org/​10.​1007/​s13201-​018-​0742-6
Singh A, Malik A, Kumar A, Kisi O (2018) Rainfall-runoff modeling in Wang X, Zhou Y, Zhao Z et al (2019) A novel water quality mechanism
hilly watershed using heuristic approaches with gamma test. Arab modeling and eutrophication risk assessment method of lakes and
J Geosci 11:1–12. https://​doi.​org/​10.​1007/​s12517-​018-​3614-3 reservoirs. Nonlinear Dyn 96:1037–1053. https://d​ oi.o​ rg/1​ 0.1​ 007/​
Sridharam S, Sahoo A, Samantaray S, Ghose DK (2021) Estimation s11071-​019-​04837-6
of water table depth using wavelet-ANFIS: a case study. In Com- Yadav B, Eliza K (2017) A hybrid wavelet-support vector machine
munication Software and Networks (pp. 747–754). Springer, Sin- model for prediction of lake water level fluctuations using hydro-
gapore. https://​doi.​org/​10.​1007/​978-​981-​15-​5397-4_​76 meteorological data. Measurement 103:294–301. https://​doi.​org/​
Suryanarayana C, Sudheer C, Mahammood V, Panigrahi BK (2014) An 10.​1016/j.​measu​rement.​2017.​03.​003
integrated wavelet-support vector machine for groundwater level Yang X-S (2009) Firefly algorithms for multimodal optimization. In:
prediction in Visakhapatnam, India. Neurocomputing 145:324– International symposium on stochastic algorithms. Springer, pp
335. https://​doi.​org/​10.​1016/j.​neucom.​2014.​05.​026 169–178. https://​doi.​org/​10.​1007/​978-3-​642-​04944-6_​14
Tang Y, Zang C, Wei Y, Jiang M (2019) Data-driven modeling of Yang X-S, Hosseini SSS, Gandomi AH (2012) Firefly algorithm for
groundwater level with least-square support vector machine and solving non-convex economic dispatch problems with valve load-
spatial–temporal analysis. Geotech Geol Eng 37:1661–1670. ing effect. Appl Soft Comput 12:1180–1186. https://​doi.​org/​10.​
https://​doi.​org/​10.​1007/​s10706-​018-​0713-6 1016/j.​asoc.​2011.​09.​017
Taormina R, Chau K-W, Sethi R (2012) Artificial neural network simu- Yaseen ZM, Ebtehaj I, Bonakdari H et al (2017) Novel approach for
lation of hourly groundwater levels in a coastal aquifer system of streamflow forecasting using a hybrid ANFIS-FFA model. J
the Venice lagoon. Eng Appl Artif Intell 25:1670–1676. https://​ Hydrol 554:263–276. https://​doi.​org/​10.​1016/j.​jhydr​ol.​2017.​09.​
doi.​org/​10.​1016/j.​engap​pai.​2012.​02.​009 007
Tharwat A, Hassanien AE (2018) Chaotic antlion algorithm for param- Yoon H, Hyun Y, Ha K et al (2016) A method to improve the stability
eter optimization of support vector machine. Appl Intell 48:670– and accuracy of ANN-and SVM-based time series models for
686. https://​doi.​org/​10.​1007/​s10489-​017-​0994-0 long-term groundwater level predictions. Comput Geosci 90:144–
Tikhamarine Y, Malik A, Pandey K et al (2020) Monthly evapotranspi- 155. https://​doi.​org/​10.​1016/j.​cageo.​2016.​03.​002
ration estimation using optimal climatic parameters: efficacy of Zare M, Koch M (2018) Groundwater level fluctuations simulation and
hybrid support vector regression integrated with whale optimiza- prediction by ANFIS-and hybrid Wavelet-ANFIS/Fuzzy C-Means
tion algorithm. Environ Monit Assess 192:1–19. https://​doi.​org/​ (FCM) clustering models: application to the Miandarband plain. J
10.​1007/​s10661-​020-​08659-7 Hydro-Environment Res 18:63–76. https://​doi.​org/​10.​1016/j.​jher.​
Wang X, Liu T, Zheng X et al (2018) Short-term prediction of ground- 2017.​11.​004
water level using improved random forest regression with a

13