Agricultural Water Management 245 (2021) 106625

Contents lists available at ScienceDirect

Agricultural Water Management

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

Research paper

Groundwater quality forecasting using machine learning algorithms for

irrigation purposes
Ali El Bilali a, *, 1, Abdeslam Taleb a, Youssef Brouziyne b
a
Hassan II University of Casablanca, Faculty of Sciences and Techniques of Mohammedia, Morocco
b
Mohammed VI Polytechnic University (UM6P), International Water Research Institute, Ben Guerir, Morocco

A R T I C L E I N F O A B S T R A C T

Handling Editor - Dr. B.E. Clothier Using conventional methods to evaluate the irrigation water quality is usually expensive and laborious for the
farmers, particularly in developing countries. However, the applications of artificial intelligence models can
Keywords: overcome this issue through forecasting and evaluating the irrigation water quality indexes of aquifer systems
Machine learning using physical parameters as features. This paper aims forecasting the Total Dissolved Solid (TDS), Potential
Berrechid aquifer
Salinity (PS), Sodium Adsorption Ratio (SAR), Exchangeable Sodium Percentage (ESP), Magnesium Adsorption
Irrigation water quality
Ratio (MAR), and the Residual Sodium Carbonate (RSC) parameters through Electrical Conductivity (EC),
Sensitivity
Uncertainty Temperature (T), and pH as inputs. To achieve this purpose, we developed and evaluated Adaptive Boosting
Prediction performance (Adaboost), Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Regression (SVR) models
using 520 samples of data related to fourteen Groundwater quality parameters in Berrechid aquifer, Morocco.
The results revealed that the overall prediction performances of Adaboost and RF models are higher than those of
SVR and ANN. However, the generalization ability and sensitivity to the inputs analyses show that the ANN and
SVR models are more generalizable and less sensitive to input variables than Adaboost and RF. Globally, the
developed models are valuable in forecasting the irrigation water quality parameters and could help the farmers
and decision-makers in managing the irrigation water strategies. The developed approaches in this study have
been revealed promising in low-cost and real-time forecast of groundwater quality through the use of physical
parameters as input variables.

1. Introduction demonstrated good results in evaluating the groundwater quality,

mapping pollution risk, and assessing the health risk. These methods
Groundwater resources are considered among the main natural re­ include the index-based approaches, statistical, and the Geographic In­
sources which support the socio-economic development of countries. formation System-based methods that are frequently used in assessing
However, agriculture is the main sector that consumes groundwater in and mapping the groundwater quality (Adimalla et al., 2019; Adimalla
the world (Siebert et al., 2010). These natural resources are confronting and Taloor, 2020; Das et al., 2020; El Mountassir et al., 2020; Gao et al.,
different issues that threaten their sustainability such as climate change 2020; Heiß et al., 2020; Jahin et al., 2020; Jamaa et al., 2020; Taloor
effects, anthropogenic activities, and natural processes (Alabjah et al., et al., 2020; Tyagi and Sharma, 2014; Wu et al., 2020).
2018; Baghvand et al., 2010; Burri et al., 2019; El Asri et al., 2019; To serve agricultural purposes, the irrigation water quality (IWQ) is
Houéménou et al., 2020; Mountadar et al., 2018). Generally, these commonly evaluated based on many indexes and parameters adopted by
problems deteriorate the chemical composition of water and make it FAO guidelines 29 (Ayers and Westcot, 1994). Indeed, to reduce
unsuitable for drinking and/or irrigation purposes. However, moni­ subjectivity in evaluating the suitability of the water for irrigation
toring and evaluating innovative approaches may be valuable to manage purposes, several studies have investigated the index-based and statis­
water quality and reduce pollution and health risks (Chowdury et al., tical approaches. Recently, Singh et al. (2018) have used Saaty’s Ana­
2019; Krupková et al., 2019; Sakai et al., 2018; Zhu et al., 2019). For this lytic Hierarchy Process (SAHP) to develop the Irrigation Water Quality
purpose, many methods and approaches have been applied and have Index (IWQI). Jahin et al. (2020) have developed an IWQI for surface

* Corresponding author.
E-mail address: [email protected] (A. El Bilali).
1
ORCID ID: 0000-0002-5744-3086

https://doi.org/10.1016/j.agwat.2020.106625
Received 30 July 2020; Received in revised form 29 October 2020; Accepted 31 October 2020
Available online 7 November 2020
0378-3774/© 2020 Elsevier B.V. All rights reserved.
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

water in Egypt using the multivariate analysis. Ewaid et al. (2019) have systems in the country is deteriorating and experiencing an extensive
developed the irrigation water quality guide software to determine the salinization due to the seawater intrusion, climate change effects, and
water suitability for irrigation purposes based on the FAO guidelines anthropogenic activities (Alabjah et al., 2018; Fadili et al., 2015, 2016;
(Ayers and Westcot, 1994) and (Meireles et al., 2010) IWQI. Although Fakir et al., 2002; Gamar et al., 2018; Lamrani Alaoui et al., 2008; Molle
the results of all these studies have stated that using conventional ap­ and Tanouti, 2017; Mountadar et al., 2018; Moyé et al., 2017; Seif-
proaches is rapid and cheap while evaluating and controlling processes, Ennasr et al., 2016). This is a major problem in the agricultural sector
they also required a large amount of datasets to be involved. Therefore, that threats the national food and water security. Consequently, Irriga­
using conventional methods in evaluating irrigation water quality could tion Water Quality Indexes (IWQI) must be predicted to evaluate and
be laborious and expensive for farmers particularly in developing manage water suitability for irrigation uses and to ensure agricultural
countries. Their high cost can be one of the plausible reasons for the poor sustainability. The main aims of this study are: (1) to evaluate the
water quality assessment (Amado et al., 2019). Nowadays, sustainable capability of ML models in predicting the irrigation water quality pa­
management plans addressed to groundwater need innovative and smart rameters such as the potential salinity (PS), the Sodium Adsorption Ratio
methods with a reasonable cost in evaluating and forecasting the (SAR), the Exchangeable Sodium Percentage of (ESP), and the Magne­
groundwater quality. For this purpose, the prediction-based approaches sium Adsorption Ratio (MAR) in Berrechid aquifer in Morocco using the
may be valuable in overcoming this challenge in groundwater planning physical parameters as input variables: (2) compare the performances of
and management. four models such as Random Forest, Support Vector Regression (SVR),
During the last two decades, the application of Artificial Intelligence ANN, and adaptive boosting (Adaboost) and selecting the best suitable
(AI) techniques has increased in many fields. For instance, in hydro­ models, and (3) analyze the uncertainty and the sensitivity of the
logical forecasting fields, Machine Learning (ML) models have proved to models.
be highly accurate in predicting the river flow (Yaseen et al., 2019). For
the groundwater modeling, AI techniques have started to be satisfactory 2. Irrigation water quality parameters
robust using various ML model in forecasting the groundwater level
(Rajaee et al., 2019). Concerning the water quality prediction, several Evaluation of water suitability for irrigation purposes relies on
studies have been carried out using AI techniques as reviewed by several parameters and indexes that are suggested by many organiza­
Tiyasha and Yaseen (2020). Lu and Ma (2020) applied extreme gradient tions and agencies. In this study, we focus on the TDS, SAR, ESP, PS,
boosting and Random Forest (RF) models to predict six water quality RSC, and the MAR parameters that are calculated as follow:
indicators in the River Tualatin. Castrillo and García (2020) used linear The TDS is the sum of ion concentrations in the water as given by the
and RF models to estimate the high frequent nutrient concentrations in 1(1) (Sorensen, 1977).
the River Thames. Meyers et al. (2017) used Artificial Neural Network ∑( )
(ANN), Support Vector Machine (SVM), and RF models to predict the TDS = cations + anions (1)
water turbidity of a trunk network in United Kingdom. Chou et al.
(2018) determined the Carlson’s (1977) Trophic State Index (CTSI) of The potential salinity (PS) (in meq L− 1) is defined as the chloride
twenty reservoirs in Taiwan using ML models. El Bilali et al. (2020) used concentration plus half of the sulfate concentration and is given by 2(2)
ANN models to predict the chemical groundwater quality for drinking (Doneen, 1964).
purposes. Fijani et al. (2019) designed and implemented AI methods to SO2−4
support the reservoir water quality monitoring in real time. Di et al. PS = Cl− + (2)
2
(2019) applied classification ML models to evaluate water quality in the
Yangtze River in China. Wagh et al. (2016) used the ANN model to The Sodium Adsorption Ratio (SAR) in (meq0.5 L− 0.5) is calculated by
determine the groundwater suitability for irrigation purposes and 3(3) (Richards, 1954). The high SAR value reduces the hydraulic con­
showed perfect performances using 13 physic-chemical parameters as ductivity of the soil texture and, therefore, decreases the irrigation ef­
features. Importantly, all these studies demonstrated that the ML models ficiency. Water is considered unsuitable for irrigation purposes if the
are highly accurate in predicting and evaluating water quality. How­ SAR value is more than 10 (meq0.5 L− 0.5) according to the FAO
ever, the ML model efficiency does not only depend on prediction ac­ guideline.
curacy but also on the nature and number of the used predictors. In this Na+
regard, increasing input variables evaluated in the laboratory decreases SAR = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (3)
Mg2+ +Ca2+
the efficiency of ML models and their field application. The literature 2

review shows that most predictors of the applied models must be ideally The exchangeable sodium percentage (ESP in %) parameter assesses
evaluated in the laboratory. Besides, the generalization ability and the the effect of the sodium on soil texture. It is calculated by the 4(4)
sensitivity of the ML models to input variables are not analyzed enough (Kopittke et al., 2006).
even though they are a crucial issue in the practical implication.
Interestingly, using physical parameters such as electrical conduc­ ESP =
Na+
∗ 100 (4)
tivity, pH, and Temperature that can be measured by sensor technolo­ Ca2+ + Mg2+ + Na+ + K +
gies as predictors could significantly improve machine learning The residual sodium carbonate (RSC meq L− 1) is calculated as
efficiency (Ayadi et al., 2020; Chowdury et al., 2019). Therefore, it follow:
encourages decision-makers to implement AI technologies for water ( ) ( )
quality planning and management. It is vital to investigate the machine RSC = CO2−3 + HCO−3 − Ca2+ + Mg2+ (5)
learning models for forecasting the IWQ parameters using only the
The magnesium absorption ratio is calculated by the 6(6):
physical parameters as input variables without relying on reducing the
performance of the models using archived data. Consequently, in this Mg2+
study, we attempt to contribute to overcoming the conventional method MAR = ∗ 100 (6)
Mg2+ + Ca2+
limitations using ML models to forecast groundwater quality for irriga­
The ionic concentrations are given in meq L− 1, except for TDS
tion purposes under extensive salinization.
parameter that is given in mg L− 1.
In Morocco, continuous decline in groundwater levels (0.5–2 m per
year) due to overexploitation and climate change impacts has become a
serious challenge in planning and management water resources
(Hssaisoune et al., 2020). However, the water quality of the most aquifer

2
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

3. Materials and methods aquifer Berrechid and consist of sandstones, sands, sandy limestone, and
minor conglomerate with a total thickness that ranges from 5 to 40 m
3.1. Study area (Droubi et al., 2008; Lyazidi et al., 2003). 6) Quaternary: These deposits
are the dominant facies and essentially consist of silt and conglomerate
Berrechid aquifer is located in the north-west of Morocco between series followed by red silty clays, and then pebbles and gravel with a
the Settat and the Casablanca cities and covers a surface of about thickness of 0–50 m. Regarding the water quality, this aquifer is char­
15,000 Km2 (Fig. 1). This area is a semi-arid region with an annual acterized by three hydro-geochemical facies (Na-Cl), (Na-Mg-Ca-Cl),
rainfall that ranges from 280 to 320 mm and more 90% of the precipi­ and (Ca-Mg-HCO3-Cl) (El Ghali et al., 2020).
tation is occurs during the rainy seasons (between October to April) This area is characterized by fertile soil and productive groundwater
while the temperature ranges from 6.5 ◦ C (January) to 38 ◦ C (August). that lead to the groundwater depletion and water quality deterioration,
Geologically, this site is part of the marine area of the western Moroccan as today it is unsuitable for drinking purposes (El Bouqdaoui et al., 2010;
Meseta. The study area has been investigated for over six decades (El El Ghali et al., 2020; Elfarrak et al., 2014). Meanwhile, the water is
Houssine et al., 2014; El Mansouri, 1992, 1993; Hazan and Moullard, mainly used by the farmers for vegetable (carrot) and livestock pro­
1962; Lyazidi et al., 2003; Ruhard, 1975). Indeed, various geological ductions (Ouassissou et al., 2019). However, the water quality is
formations were deposited in the form of a large broth, under semi-arid monitored by the River Basin Agency of Bouregreg and Chaouia through
climatic contexts combined with the old subsidence and sedimentations fourteen monitoring stations.
that range from the primary to the recent deposits. However, the litho-
stratigraphic sequence of these formations is as follows: 1) primary: the
3.2. Machine learning models
bedrock comprises of shale intercalated with the quartzite and sand­
stone layers wherein the outcrops, with 150 m thick, are Silurian-
The used models are the Adaptive Boosting, Support Vector
Devonian and Green Acadian along the southeast and the northwest,
Regression (SVR), Random Forest (RF) and Artificial Neural Network
respectively (El Mansouri et al., 1992). 2) Triassic: these sediments are
(ANN) approaches. However, extensive details on machine learning al­
featured by Saliferous red clays and Siliciclastic intercalated by evapo­
gorithms used and data mining can be found in Aggarwal (2015), Bishop
rate and basalt. They are spread in the east part of the aquifer (Bensalah
(2006), Bonaccorso (2018), Freund and Schapire (1997, 1996), Hastie
et al., 2011). 3) infra-Cenomanian comprises detrital red clays, with a
et al. (2009), and Kubat (2017).
total thickness of about 40 m. These depots are rich in Gypsum, with
some layers of conglomerate followed by limestone and white to yellow
3.2.1. Adaptive boosting (adaboost)
marl layers. 4) Cenomanian: these consist of dolomitic limestone and
Adaboost is a boosting algorithm introduced by Freund and Schapire
yellow marls with the intercalation of the greenish marl for a thickness
(1997). It can be used for classification and regression tasks. For a
of about 120 m (Ruhard, 1975). 5) Pliocene: form the system of the
dataset S={(xi,yi), i=1,2,3..,N} where each xi is in some instance X and

Fig. 1. Study area and monitoring station used localizations.

3
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

each yi is in some label Y, and for a series of round (M). Wiener, 2002). These parameters play an important role in the stability
The algorithm initializes the distribution (D) (or weight) as follow: of the model and, therefore, in prediction performances (Oshiro et al.,
2012). In this study we determine the model parameters by trial error
1
D1i = for i = {1, ….N} (7) method.
N
Then for j=1 to M, Adaboost algorithm builds weak models hj from 3.2.4. Artificial neural network
the training dataset using D that minimizes εj and satisfies εj <0.5 ANN is commonly used as ML model in groundwater modeling
conditions. (Rajaee et al., 2019). ANN model is composed of three layers namely:
εj is a weighted error of the jth model and is given by 8(8). input layer, hidden layers and output layer. These layers are inter­
∑ connected through the neurons that are characterized by weight and
Dij (8)
εj =
i:hj (xi)∕
=yi
bias. The variables weighted summed with the bias of the layer are
transformed from the layer jth to the j + 1th layer through an activation
The weight “confidence” αj of the jth model is calculated by the 9(9). function (f) and so on until the target layer (Dawson and Wilby, 1998).
( ) The training process is iteratively repeated by changing the weights and
1 1 − εj
αj = ln (9) the biases of the layers until good preliminary performance (coefficient
2 εj
of correlation). To simplify this approach, we take the models with three
The distributions for next iteration were updated as follow: MLP layers and the outputs (Yk ) are given by the flowing equation
(Schalkoff, 1997):
Dj+1
i = e− yi hj (xi )αj
Dji (10) ( ( ))
∑
m ∑
n
Yk = fk Wjk ∗ fj Xi Wij + W0 (17)
Dj+1
Dj+1
i = i
(11) i=1 i=1
∑
N
Dj+1
i With n is the feature numbers, m is the neuron numbers in the hidden
i=1
layer and p is the neuron numbers of the target layers, W0 is the bias, Wjk
The prediction for new dataset is conducted by combining weighted and Wij are the weights between the jth neuron and the kth target neuron,
majority vote of the models hj. and between the ith neuron and jth neuron respectively. Whereas fk and fj
[∑M ]
are the transfer functions of the neurons k and j of the output and hidden
H(x′ ) = sign αj hj (x′ ) (12)
j layers respectively.

3.2.2. Support vector machine (SVM) 3.3. Data processing and model performance evaluation methods
Support vector Machine (SVM) is a discriminative method applied to
both the classification and regression (SVR) tasks based on the con­ 3.3.1. Data collection and analysis
struction of a hyper-plane to minimize errors (Vapnik, 1995). For Five hundred and twenty samples of Groundwater quality from
modeling system (S) with observation dataset (Ds), Ds = (xi , yi )ni=1 , xi Berrechid aquifer have been considered in this study. All samples were
represents the inputs and yi the outputs with a linear function as shown provided by River Basin Agency of Bouregreg and Chaouia and were
in 7(7). collected from the fourteen monitoring stations illustrated in Fig. 1.
Indeed, the depth of the wells ranges from 40 m to 140 m. The sampling
f (x) = 〈ω ∗ ∅(x) + b〉 (13)
protocol was carried out during the pumping for irrigation purposes two
The optimal function is the minimization of the function (14) (sub­ times during spring and the summer for the period from 2009 to 2019.
ject to 15(15)). Hence, the loss functions such as ϵ-insensitive, quadratic These samples include the measurements of the parameters: the Elec­
and Hubber methods can be used. trical conductivity (EC), pH, Temperature, Cl-, SO2- 2- -
4 , CO3 , HCO3, NO3,
-
- 2+ 2+
∑n NO2, NH4 ,Na ,K ,Ca , and Mg
+ + +
concentrations. The physical pa­
1 ⃒⃒⃒⃒ 2 ⃒⃒⃒⃒
min(ω, b, ξ− , ξ+ ) = ∗ ω +C ∗ (ξi− + ξi+ ) (14) rameters were measured in situ during the sampling. Meanwhile, the
2 i=1 other parameters were analyzed in the laboratory using the flame
⎧ ⎫ photometer for Na+ and K+, UV spectrophotometer for the nitrate and
⎪
⎪ y − ωT ∗ ∅(x) − b ≤ ε + ξi− ⎪⎪ sulfate concentrations, and volumetric titration method for the other ion
⎨ i T +⎬
− yi + ω ∗ ∅(x) + b ≤ ε + ξi
S.t (15) concentrations. Fig. 2 summarizes the distribution of the parameter
⎪ ξi− , ξi+ ≥ 0 ⎪
⎪
⎩ ⎪
⎭ values by sample, all ion concentrations are in mg L− 1, Temperature is in
i = 1, 2…..n ◦
C, and EC is in (µs cm− 1).
With ∅(x) is a Kernel function (k) such as polynomial, radial basis It can be seen that there are some outlier values on almost all pa­
and linear functions, ω and b represent weigh and basis vectors, and C is rameters. Consequently, it is necessary to analyze these data before their
a pre-specified value to penalize the training error while ξi− and ξi+ are input. Moreover, in ML models development the exploration and
the lower and upper constraints on the output. cleaning of the data are important steps to have accurate and reliable
This study adopted the Radial Basis Function (RBF) given by the 16 models. In this study, data pre-processing was conducted following five
(16) as Kernel Function (Ghosh, 2010). steps:
( )
k xi , xj = exp(− γ|xi − x j |2 )
(16) 1) Data reliability checking: This operation was carried out through the
calculation of the absolute value of the Charge Balance Error (CBE)
3.2.3. Random forest as given by 17(17) of the groundwater samples and those which have
Random Forest (RF) is an ensemble of ML model based on decision more than 5% in absolute value were rejected (Freeze and Cherry,
trees (Breiman, 2001). Regression and classification are conducted by 1979). Moreover, the outlier value detections of physical parameters
aggregating a technique that operates by constructing an ensemble of were conducted using the boxplots.
decision trees in training by swapping and changing the covariates to ⃒∑ ∑ ⃒
⃒ cations − anions ⃒
improve the prediction performance. The target is achieved by weighted CBE = ⃒⃒∑ ∑ ∗ 100⃒⃒ (18)
average of tree outputs. Executing this model requires a number of cations + anions
trained trees and an amount of the variable used in each tree (Liaw and

4
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

Fig. 2. plots of the raw data related to fourteen parameters provided by River Basin agency of Bouregreg and Chaouia, X line represents the identification of samples
while Y line represents the parameters.

2) Calculation of the IWQ parameters and data analysis: Statistical the lower value of r indicates a worse fit model.
analysis of data was carried out; Descriptive statistic characteristic of ⎛ ∑n ( )( ) ⎞
IWQ parameters are summarized and plotted as boxplot to detect the i=1 Xo i − Xo X p i − Xp
r = ⎝[ ⎠ (19)
outlier values. The correlation matrix was plotted and analyzed using ∑n ( )2 ∑ (
n
)2 ]0.5
Pearson method. i=1 Xo i − Xo i=1 Xp i − X p

3) Data normalization from 0 to 1 to reduce the dimensionality influ­

The RMSE is the square root of the variance of the residual errors.
ence of the variables on the model performances and to improve the
The lower value of the RMSE compared to the output ranges indicates
model generalization.
the better fit of the model.
4) Model training and validation processes
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
5) Uncertainty and sensitivity analysis of the ML model ∑( )2
Xp i − Xo i
RMSE = (20)
n
3.3.2. Metric evaluation models
The model performances are evaluated using three statistical crite­ The RBIAS was used to state if the model tended to underestimate
rions: Coefficient of correlation (r), Root Mean Square Error (RMSE), (RBIAS>0) or overestimate (RBIAS<0) the target magnitude and was
and the Relative Bias (RBIAS). calculated by 20(20). If the RBIAS equals 0, the model is perfect while
The coefficient of correlation (r) was calculated by 18(18). The co­ higher absolute value of RBIAS indicates that the model is biased.
efficient of correlation 1 show the best correlation between the observed
and predicted values but it does not indicate the best model. Meanwhile,

5
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

∑n ) (
Xo i − Xp i than 5% in absolute value.
RBIAS = i=1∑n (21) Fig. 4 presents the boxplots of irrigation water quality parameters
i=1 Xo i
and studied variables. This figure demonstrates that most parameters
Xo, Xp, and Xo are observed and the model predicted, mean of have some outlier values. TDS, pH, and Temperature parameters have 5
observed values respectively. outlier values while the RSC, PS, Temperature, and SAR parameters
In developing of ML models, the great challenge is to guarantee the have 16, 11, 8, and 2 outlier values respectively. Wang et al. (2020)
generalization ability (GA) of the models. Yoon et al. (2011) defined the demonstrated that removing outlier values does not improve the pre­
GA in predicting the groundwater level by the 21(21). GA values are diction accuracy of ML models. However, some outliers of physical pa­
unity indicate the ML model is perfect. If the GA is less than unity, the rameters cannot be acceptable; i.e. pH=0.75 and Temperature=41.5 ◦ C,
models are under trained while higher than unity, the models are over and therefore we removed their corresponded samples (six samples).
trained. Overall, 476 samples were retained as suitable data in terms of reli­
RMSE during the validation phase ability for this study. To reduce dimensionality impacts on model per­
GA = (22) formances the data were normalized in range from 0 to 1.
RMSE during the training phase
Table 1 presents basic statistical characteristics (mean, minimum,
However, the total uncertainty of the fitted models was evaluated maximum, ith quartile Qi, and the standard deviation (SD)) of irrigation
through comparing the observed and simulated values and calculating water quality indexes calculated based on the 476 of the retained sam­
of the standard error and the confidence bound as follow: ples. This Table shows that more than 50% of samples have EC and TDS
( ) values considered high according to Ayers and Westcot (1994) salinity
ei = Xo i − Xp i (23)
restrictions. Meanwhile for the other parameters, except for the extreme
√∑
̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ values, the restrictions are moderate to none for irrigation purposes.
n 2
i=1 (ei − e)
SD = (24)
(n − 1) 4.1.2. Exploratory data analysis
For further exploration of the variable, a matrix correlation analysis
SD
CB = z ∗ √̅̅̅ (25) and the importance evaluation of input variables have been carried out.
n Fig. 5 presents the matrix correlation (Fig. 5a) and variable ranks
With ei is the error. z is z-score of confidence level (for 95% is about (Fig. 5b). The results show that the electrical conductivity (EC) has
1.96), and e is the mean prediction error, IE is the interval prediction strong correlations with TDS (r=0.96), PS (r=0.95), SAR (r=0.78), and
error, and n=number of the observations used. RSC (r=− 0.79) parameters; and it has a moderate correlations with the
Furthermore, sensitivity analysis of the models was evaluated to ESP (r=0.50) and MAR (r=0.41) parameters. The pH has weak corre­
determine the input variables that influence significantly the model lations with all parameters. Meanwhile, the temperature has the lowest
predictions of the water irrigation parameters. This was carried out correlations with all parameters. Such results reflect that the electrical
using One-Factor-At-Time (OFAT) method. Here, the RMSE values of the conductivity is more correlated input variable with the predicted pa­
model performances were evaluated with one input variable omitted a rameters than the pH, and Temperature. However, strong correlations
time by generating randomly the uncertainties using the Monte Carlo do not imply causality, as the complex combinations of the features can
method, while the other variables were held constants. Then, the impact have influences on the target variable. Meanwhile, the lower correla­
of each input on the RMSE value was evaluated by calculating the ab­ tions between the T, pH, and EC demonstrate that these parameters are
solute value of the difference in RMSE (|ΔRMSE|). Accordingly, the separable and not redundant input variables and, therefore, are useful to
model sensitivity to an input increases with increasing |ΔRMSE|. improve the prediction accuracy of the ML models (Kuhn and Johnson,
2019). Besides, the importance of the input variables was evaluated
4. Results and discussion using Relief algorithm (Kira and Rendell, 1992). Fig. 5b presents the
ranks of the EC, Temperature, and pH variables for predicting the IWQ
4.1. Data pre-processing parameters. Globally, it was observed that there is no significant dif­
ference between the variables’ importance. Consequently, we adopted
4.1.1. Data cleaning the three parameters as features of the ML models.
Checking data reliability by 17(17) shows that 38 samples have CBE
more than 5% in absolute value and they are considered unsuitable in 4.2. Model training results
this study. Fig. 3 illustrates the scatterplot of sum cations vs sum anions.
This figure demonstrates a good reliability of samples since CBE is less Before the training phase, the selected data are organized in CSV.
files and were divided into two datasets: 300 samples were used for the
training-testing process and 176 samples were used for validations
phase. Then, the ML models were built in Jupyter using anaconda
150
platform tools (www.anaconda.com/products/individual). To avoid the
over-fitting of the models, we used the k-fold cross-validation method
sum of cation (meq/l)

(k = 5) during the training process (Arlot and Celisse, 2010). The

100 optimal architectures, functions, and hyper-parameters of each model
were determined through a trial and error analysis by their changing
during the training phase. Table 2 presents optimal parameters and
y = 1.1039x - 1.3674
50 functions selected.
R² = 0.9898
Table 3 represents the RMSE, r, and the RBIAS values of the models
n=482
during the training process. These results show that the SVR model has
significant values of RBIAS and RMSE compared to other models in
0 predicting the TDS and RSC parameters. ANN, RF, and Adaboost models
0 50 100 150
sum of anions (meq/l) demonstrated high accuracy in predicting the TDS parameter during the
training process with r 0.95, RMSE 457 mg L− 1 (for data ranging from
Fig. 3. Scatter plot between the sum of anions and cations of the consid­ 2000 to 9400 mg L− 1), and RBIAS 1% in average. For the PS parameter,
ered samples. all models had excellent performances with average coefficients of

6
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

Fig. 4. Boxplots of the used variables for 482 samples.

Table 1
Descriptive statistical characteristics of the used variables.
Te (◦ C) pH EC (dS/m) TDS (g/l) PS (meq/L) SAR RSC ESP (%) MAR (%)

Mean 21.8 7.3 3.0 2.0 24.1 5.7 -12.0 45.9 43.0
Min 10.0 6.5 0.31 0.4 0.5 0.1 -83.1 0.5 3.1
Q1 20.1 7.1 1.3 1.0 7.1 2.2 -15.5 34.5 30.8
Q2 22.2 7.3 2.4 1.8 17.2 4.7 -8.8 47.6 44.6
Q3 23.6 7.6 4.1 2.7 34.7 8.5 -4.7 57.5 55.7
Max 30 9.9 13.0 9.4 150.0 19.9 8.8 78.4 84.8
SD 3.7 0.6 2.1 1.4 22.6 3.9 11.4 15.3 16.9

Fig. 5. Pearson correlation matrix analysis (a) and variable ranks (b).

correlations 0.93, RBIAS less than 3% in absolute value, and average (SAR) and Exchangeable Sodium Percentage (ESP) parameters. Co­
RMSEs of about 7.9 meq L− 1 (for PS values range from 0.5 to efficients of correlations ranged from 0.75 to 0.82, RMSE less than
150 meq L− 1). Regarding the Sodium hazard prediction, all models have 3 meq0.5 L− 0.5, and RBIASs ranged from 0.33% to 5.63% for SAR
satisfactory performance in predicting the Sodium Adsorption Ratio parameter. Meanwhile, for the ESP parameter, the coefficients of

7
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

Table 2 Table 4
Hyper-parameter and functions selected during the training phase. Model performances during the validation phase.
Model Parameters and functions Parameter Statistical SVR ANN RF AdaBoost
indices
ANN 3 layers
12 neurons in hidden layer r 0.94 0.96 0.96 0.99
Algorithm: Levenberg- Marquardt TDS RBIAS 15.37% -1.11% 0.23% 0.01%
Function activation: Sigmoid RMSE(mg l− 1) 1270.68 400.63 343.92 182.0
Identity in output layer r 0.95 0.95 0.95 0.99
Epoch number: 1000 PS RBIAS 2.09% -0.87% 0.42% -0.04%
Learning rate: 0.01 RMSE(meq l− 1) 6.73 6.77 5.76 3.7
Momentum coefficient: 0.85 r 0.83 0.83 0.83 0.98
SVR C=200 SAR RBIAS 0.92% -5.85% -0.44% -0.12%
Kernel function: RBF (γ=1.2) RMSE 2.18 2.14 1.35 0.9
ε-function loss, ε = 0.002 r 0.65 0.65 0.65 0.98
Random Forest Number of trees: 20 ESP RBIAS -3.41% -2.89% 0.40% 0.09%
Adaboost Loss function : Exponential RMSE(%) 12.18 11.68 7.49 4.2
Learning rate: 0.5 r 0.51 0.52 0.52 0.98
Estimator number: 50 MAR RBIAS 0.52% -0.17% -0.43% 0.05%
RMSE(%) 14.74 14.29 8.93 2.9
r 0.77 0.82 0.82 0.99
RSC RBIAS 10.09% 4.09% -0.46% 0.09%
Table 3 RMSE(meq L− 1) 7.17 6.38 4.49 2.8
Model performances during the Training process.
Parameter Statistical SVR ANN RF AdaBoost RBIAS less than 1% in absolute value. Additionally, the Adaboost and RF
indices
have performances more excellent than those evaluated during the
r 0.93 0.96 0.93 0.95 training phase for predicting all parameters. Meanwhile, ANN and SVR
TDS RBIAS 11.81% 1% 1.80% 0.17%
models present prediction accuracy closer to that during the training
RMSE (mg L− 1) 1357.69 390.33 543.60 440.78
r 0.95 0.95 0.92 0.92 phase. Compared to the performances presented in Table 3, Table 4
PS RBIAS 2.40% -0.10% 1.92% -0.36% shows that SVR model still has unacceptable performances for simu­
RMSE (meq L− 1) 7.16 6.91 8.84 8.75 lating TDS and MAR parameters with high values of RMSE
r 0.81 0.82 0.79 0.75 (1270.68 mg l− 1) and RBIAS (15.37%) for TDS parameter; low coeffi­
SAR RBIAS 5.63% -0.33% 1.31% 2.11%
RMSE 2.38 2.23 2.42 2.72
cient of correlation (r = 0.51) and high RMSE (14.74%) for MAR
r 0.62 0.65 0.62 0.69 parameter. Similarly, the ANN and RF models failed to predict the MAR
ESP RBIAS -0.35% 0.47% 2.99% 2.98% parameter during the validation process due to the high RMSE and low
RMSE 12.53 11.78 12.35 13.40 coefficient of correlation.
r 0.49 0.51 0.47 0.64
Fig. 6 shows the scatterplots of the observed and simulated values by
MAR RBIAS -0.83% -0.34% -1.47% 0.22%
RMSE 15.18 14.64 15.36 16.15 the models during the validation process. The accuracy of the models is
r 0.63 0.77 0.75 0.74 satisfactory when the values are distributed over or evenly on both slides
RSC RBIAS 9.00% 0.80% -3.03% -3.67% the line X˭Y, showing that the errors obey the Gaussian distribution. This
RMSE 9.34 7.33 7.66 7.94 figure reflects that the predicted values by the ensemble models (RF and
Adaboost) are distributed over the X˭Y line better than those of the ANN
correlations ranged from 0.59 to 0.65 and RBIAS was less than 3%. In and SVR models. Furthermore, the Adaboost showed better perfor­
predicting the MAR parameter, all models gave relatively good perfor­ mances where the predicted values are very closer to the observed
mances with RBIAS less than 2% in absolute value. However, the co­ values; and for the Random Forest model, except for the MAR param­
efficients of correlations are weak except for the Adaboost model which eter, the predicted values are quite evenly distributed over the X˭Y line.
has r, 0.64. As for the RSC parameter, all other models have significant Concerning the SVR and ANN models, they showed fairly good value
performances (average of r is about 0.75 and |RBIAS| less than 4%). distributions over the X˭Y line for three (PS, SAR, and RSC) and four
Comparing these models, the Adaboost shows excellent to good (TDS, PS, SAR, and RSC) parameters respectively. Importantly, the SVR
performances for predicting all parameters followed by Random Forest and ANN models failed to reproduce the ESP parameter, as the predicted
and Artificial Neural Network models which failed to predict the MAR vales are far away from the Y˭X line and they have RMSE more than 10%
parameters. Interestingly, the SVR model also shows significantly worse even when they showed relatively good performances during the
performances for predicting the TDS and MAR parameters during the training phase. Such results demonstrate the superiority of the Adaboost
training phase. Importantly, during the training process, it was observed and Random forest models compared to the ANN and SVR models in
that there is no significant superiority observed between the ensemble predicting the studied irrigation water quality parameters.
models (Adaboost and Random Forest) and ANN. However, the valida­ By comparing the performances presented in Table 4, it can be
tion process, generalization ability evaluation, the sensitivity, and un­ concluded that (1) the Adaboost model has better performances in
certainty analysis are important issues in the application of ML in predicting all parameters followed by (2) Random Forest in predicting
groundwater resource planning and management. the TDS, PS, SAR, ESP, RSC parameters, by (3) the ANN in predicting the
TDS, PS, SAR, and the RSC parameters; then by (4) the SVR model in
predicting the PS, SAR, and RSC parameter. Except for the SVR in pre­
4.3. Validation and generalization ability results dicting the RSC parameter, all other models have RBIAS values less than
5% in absolute value indicating that the fitted models are unbiased.
After the training process, all models were used for simulating 176 These results showed that the bagging and boosting learning models (RF
samples to evaluate their validation and generalization. Table 4 presents and Adaboost) are superior to traditional ANN and SVR models in terms
the RMSE, r, and RBIAS values of the models during the validations of accuracy. Similarly, Zhao et al. (2013) found that the Adaboost is
process. Compared to ANN and SVR models, RF and Adaboost models more accurate than the SVM and ANN models in predicting the short-
showed remarkably high performances for predicting the TDS, PS, SAR, term ionospheric. Saghafi and Arabloo (2017) demonstrated that the
ESP and RSC parameters with r values ranging from 0.65 to 0.99, and Adaboost model provided high prediction accuracy for modelling the

8
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

9000
Y=X 120 Y=X
SVM SVM
Simulated TDS (mg/l) ANN ANN

Simulated PS (meq/l)
RF RF
6000
Adaboost 80 Adaboost

3000 40

0 0
0 3000 6000 9000 0 40 80 120
observed TDS (mg/L) observed PS (meq/l)

Y=X Y=X
80 SVM
21 SVM
ANN
ANN

Simulated ESP %
RF
Simulated SAR (meq/l)0.5

RF
60 Adaboost
Adaboost
14

40

7
20

0 0
0 7 14 21 0 20 40 60 80
observed SAR (meq/l)0.5 observed ESP %

Y=X
80 SVM -80 -60 -40 -20 0
ANN
Simulated RSC (meq/L)

RF -15
Simulated MAR %

60 Adaboost

-35
40 Y=X
SVM
ANN
20 -55
RF
Adaboost

0 -75
0 20 40 60 80 observed RSC (meq/L)
observed MAR %

Fig. 6. Scatter plots of the observed and simulated values for predicting the TDS, PS, SAR, ESP, MAR and RSC parameters during the validation phase.

CO2 solubility compared to the ANN and SVM models. Walker and Jiang et al., 2019; Chen et al., 2020b). However, forecasting IWQ parameters
(2019) applied Adaboost model for prediction of demand-driven in real-time requires accurate, generalizable, and stable models. Some
acquisition and found that it is higher accurate than the regression- models have a high sensitivity to the input variables that could be un­
based model. Regarding to the RF model, Castrillo and García (2020) suitable for forecasting purposes even they good overall performances.
showed a high prediction accuracy of the RF model compared to the Besides, the noisy data related to input variables caused by sensor
multiple linear regression approach. Besides, there is a further agree­ technologies are very likely, particularly during real-time. For example,
ment with previously published studies in classification tasks (Ahmed if a small fluctuation (±1.5 ◦ C) in Temperature significantly influences

9
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

prediction accuracy, the model is considered unsuitable for forecasting Table 6

the water quality in real times. Therefore, the stability of the ML models Uncertainty analysis of the models.
in the forecasting of the IWQ parameters in real-time is critical, partic­ Parameter Error SVR ANN RF AdaBoost
ularly when the maker-decision and the farmers have strategies to − 1
TDS (mg L ) e 320.458 -23.121 4.835 11.500
develop this approach in managing the saline water for irrigation pur­
CB (95%) 182.214 59.091 50.805 26.890
poses. In this study, we extend the exploration analysis of the applied ML PS (meq L− 1) e 0.526 -0.220 0.106 -0.09
models by discussing the generalization ability and the sensitivity to CB (95%) 0.993 1.000 0.851 0.540
different input variables. SAR (meq0.5 L− 0.5
) e 0.052 -0.332 -0.025 -0.070
The Generalization Ability (GA) indices of the models were calcu­ CB (95%) 0.323 0.313 0.199 0.140
lated by the 21(21) and presented in Table 5. This table shows, except ESP (%) e -1.549 -1.313 0.180 0.400
for the ANN model in predicting the TDS, all other models have GA CB (95%) 1.791 1.714 1.107 0.620
values less than 1 indicating that they do not concentrate on training the MAR(%) e 0.228 -0.075 -0.186 0.210

given data rather than a more general system; and the models were CB (96%) 2.184 2.111 1.319 0.420
RSC(meq L− 1) e -1.285 -0.521 0.059 -0.110
undertrained (Chen et al., 2020a; Yoon et al., 2011). Although the ANN
CB (95%) 1.045 0.940 0.664 0.410
and SVR models have prediction performances not good than those of
Adaboost and RF models, their overall generalization abilities are more
perfect than those of the RF and Adaboost models in this study area. parameters, the largest values of |ΔRMSE| were obtained for the Ada­
boost and RF models. Such results indicate that these models are more
4.4. Uncertainty analysis results sensitive to the features than the ANN and SVR models. However,
overall sensitivity analysis results demonstrate that the models are fairly
The problem of uncertainty of the conceptual-based models in water stable in forecasting the IWQ parameters.
quality modeling is inevitable and has been discussed in many studies
(Kardos and Obropta, 2011; Moreno-Rodenas et al., 2019; Radwan et al.,
4.6. Field practical implication in agricultural water management and
2004; Shojaei et al., 2015). In this study, we analyzed the total uncer­
future works
tainty of the fitted ML models. Table 6 presents the mean prediction
error (e) and the 95% Confidence Band in predicting the TDS, PS, SAR,
The water scarcity issue is leading several countries to increase the
ESP, MAR, and RSC parameters for 176 samples used during the vali­
use of the brackish water in their agricultural sector to ensure food se­
dation process. Based on the uncertainty analysis results, the Adaboost
curity, particularly in the arid and semi-arid regions (Bortolini et al.,
model has a smaller 95% confidence band values followed by RF, ANN,
2018; Gosling and Arnell, 2016). However, to reduce the impacts of the
and SVR models respectively.
salinity and the water sodicity on the soil and the yield productivity,
managing saline water for irrigation purposes based on two main
4.5. Model sensitivity analyses methods, namely: blending fresh and saline water and using Gypsum
(Filho et al., 2020; Kan and Rapaport-Rom, 2012; Yaduvanshi and
The prediction accuracy and the GA analysis show that the ML Swarup, 2005; Zaman et al., 2018). These processes require assessing
models are suitable for forecasting the IWQ parameters using the Tem­ the IWQ parameters of the saline water (coastal aquifer, low water
perature, pH, and EC as input variables. However, the sensitivity anal­ quality) that increases the cost related to the laboratory analysis when
ysis of the applied models to the inputs variables is required, as a high using the conventional method, particularly for the small farmers.
sensitivity of a model to noisy data can dramatically distort the fore­ In our study, we trained and validated the ML models for forecasting
casting results and, therefore, the field application of this model is the IWQ parameters using only physical parameters as predictors. In
questionable. Therefore, in this study, the sensitivity analysis of the accordance with the results of this study, ML models appeared being
fitted models was carried out by adding randomly Gaussian noises to the accurate and robust tools in forecasting the TDS, PS, SAR, ESP, MAR,
input variables and then re-simulation of the 176 samples (±2 ◦ C for the and RSC parameters. Besides, we analyzed the uncertainty and the
temperature, ±0.3 pH unity, and ±100 µs cm− 1 for the EC). The sensi­ sensitivity of these models to input variables. These analyses show
tivity analysis results of the models for forecasting irrigation water highlight that the overall generalization abilities of ANN and SVR
quality parameters are presented in Table 7. It was observed that the models are higher than those of the RF and Adaboost models. Also, the
model sensitivities to each input variable were uneven. For forecasting two last models are more sensitive to the input variables than the ANN
the TDS parameter, the models are sensitive to the EC followed by and SVR ones. However, selecting the ML models for forecasting the
Temperature, and pH with average values of |ΔRMSE| about 59.90, TDS, PS, SAR, ESP, MAR, and RSC parameters should take into account
56.16, and 38.3 mg/l respectively. These values are acceptable the sensitivity of these models. Indeed, applying sensitive models re­
compared to the range of the TDS parameter. The models are more quires a high precision of sensor technologies to measure the input
sensitive to the EC than to pH and Temperature for forecasting the PS variables. Overall, the performances and sensitivity of the models are
parameter. The pH is the most important input variable in predicting the acceptable for forecasting the studied IWQ parameters and, therefore,
ESP and RSC parameters. However, it can be seen that the model sen­ will help the farmers to manage the saline water by measurement only
sitivities to the features vary from one input to another, from one the physical parameters as input variables at low cost. Specifically,
parameter to another, and from one model to another. For all predicting the IWQ parameters of the saline water in real-time allows the
optimization of the blending of the fresh and saline water process and
Table 5 the Gypsum quantity.
Generalization ability (GA) of the models. Considering that the Electrical Conductivity, Temperature, and pH
Parameter SVR ANN RF AdaBoost parameters are measured in quasi real-time by River Basin Agency of
Bouregreg and Chaouia through its automatic hydrological information
TDS 0.94 1.03 0.63 0.40
PS 0.94 0.98 0.65 0.40 system. Therefore, using the applied ML models is efficient in fore­
SAR 0.91 0.96 0.56 0.30 casting the groundwater quality parameters. Also, nowadays, smart
ESP 0.97 0.97 0.61 0.30 technologies based on the Internet Of Thing (IOT) are innovative
MAR 0.97 0.98 0.58 0.20 methods for monitoring and analyzing data (Chowdury et al., 2019; Liu
RSC 0.77 0.87 0.59 0.40
et al., 2019). Therefore, implementing of the developed ML models and

10
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

Table 7
Sensitivity analysis results based on change in RMSE values (|ΔRMSE|) for all models with respect to input variables (EC, Te, and pH).
Variable Parameter TDS (mg l− 1) PS (meq l− 1) SAR (meq0.5 l− 0.5
) ESP (%) MAR (%) RSC (meq l− 1)

Min-Max 368–8165 1.3–127 0.5–20 12–80 6–84 (− 70)− 0.3

Model |ΔRMSE|

EC SVR 0.01 0.064 0.008 0.058 0.014 0.038

ANN 4.922 0.109 0.002 0.069 0.012 0.075
RF 60.527 1.501 0.149 0.738 0.874 0.564
Adaboost 174.3 6.5 0.6 2.7 6.0 2.2
Average |ΔRMSE| 59.90 2.0 0.2 0.8 1.7 0.7
pH SVR 0.03 0.000 0.012 0.014 0.017 0.005
ANN 0.008 0.002 0.001 0.002 0.001 0.054
RF 26.900 0.013 0.138 0.541 1.385 0.224
Adaboost 126.467 0.933 0.839 3.976 5.587 3.026
Average |ΔRMSE| 38.3 0.2 0.2 1.1 1.7 0.8
Te SVR 0.012 0.002 0.014 0.008 0.027 0.076
ANN 4.922 0.005 0.005 0.008 0.119 0.078
RF 45.436 0.030 0.235 0.799 0.587 0.134
Adaboost 174.283 0.760 0.820 3.193 5.076 1.439
Average |ΔRMSE| 56.16 0.20 0.27 1.00 1.38 0.35

mapping the prediction results in a monitoring platform will help the Declaration of Competing Interest
farmers to manage the saline water and to improve yield productivity in
this area. Alternatively, integration of the ML models in web or mobile The authors declare that they have no known competing financial
applications that could be available for the farmers improves the field interests or personal relationships that could have appeared to influence
applicability of these models in water quality management in the the work reported in this paper.
agriculture.
However, the management of the saline water can integrate two Acknowledgement
components the soil types and the crop’s tolerance. In this study, the
applied ML models are powerful tools to forecast the IWQ parameter This work is supported by River Basin Agency of Bouregreg and
rather than to evaluate the water suitability for irrigation purposes ac­ Chaouia (ABHBC) by providing further data required. Authors thank the
cording to the tolerance restrictions and the soil types. Importantly, teams of the ABHBC for their great assistance.
using the classification ML models and embedding the soil types and the
salinity tolerance of crops could be fruitful tools to manage the saline References
water with the aim to ensure sustainability agriculture. Therefore, the
integration of these aspects can be considered an emerging topic for Adimalla, N., Qian, H., Li, P., 2019. Entropy Water Quality Index and Probabilistic
Health Risk Assessment from Geochemistry of Groundwaters in Hard Rock Terrain of
future research. Nanganur County, South India. Chemie der Erde. https://doi.org/10.1016/j.ch
emer.2019.125544.
5. Conclusion Adimalla, N., Taloor, A.K., 2020. Hydrogeochemical investigation of groundwater
quality in the hard rock terrain of South India using geographic information system
(GIS) and groundwater quality index (GWQI) techniques. Groundw. Sustain. Dev.
The main aims of this study were to evaluate whether the ML models 10, 100288 https://doi.org/10.1016/j.gsd.2019.100288.
are valuable tools to forecast the IWQ parameters using physical pa­ Aggarwal, C.C., 2015. Data Mining. Springer International Publishing,. https://doi.org/
10.1007/978-3-319-14142-8.
rameters as features and to analyze their sensitivity to the input vari­
Ahmed, U., Mumtaz, R., Anwar, H., Shah, A.A., Irfan, R., 2019. Efficient water quality
ables. Therefore the Adaptive boosting, Random Forest, ANN, and SVR prediction using supervised machine learning. Water 11, 1–14. https://doi.org/
models were developed and evaluated in predicting the TDS, PS, SAR, 10.3390/w11112210.
ESP, MAR, and RSC parameters using the physical parameters as fea­ Alabjah, B., Amraoui, F., Chibout, M., Slimani, M., 2018. Assessment of saltwater
contamination extent in the coastal aquifers of Chaouia (Morocco) using the electric
tures. Through this study the major findings are as follow: recognition. J. Hydrol. 566, 363–376. https://doi.org/10.1016/j.
jhydrol.2018.09.003.
1. Adaboost model is suitable for forecasting all parameters followed by Amado, C.M., Minahk, C.J., Cilli, E., Oliveira, G., Dupuy, F.G., 2019. Economic and
efficiency based optimisation of water quality monitoring network for land use
RF model in forecasting 5 parameters (except MAR parameter). The impact assessment. BBA - Biomembr. 1862, 183135 https://doi.org/10.1016/j.
ANN and SVR models have good performances for forecasting 4 bbamem.2019.183135.
(TDS, PS, SAR, and RSC parameters) and 3 (PS, SAR, and RSC pa­ Arlot, S., Celisse, A., 2010. A survey of cross-validation procedures for model selection.
Stat. Surv. 4, 40–79.
rameters) of 6 parameters respectively; El Asri, H., Larabi, A., Faouzi, M., 2019. Climate change projections in the Ghis-Nekkor
2. The ensemble models Adaboost and Random Forest had prediction region of Morocco and potential impact on groundwater recharge. Theor. Appl.
performances better than the traditional ANN and SVR models dur­ Climatol. 138, 713–727. https://doi.org/10.1007/s00704-019-02834-8.
Ayadi, A., Ghorbel, O., BenSalah, M.S., Abid, M., 2020. A framework of monitoring water
ing the validation performance; pipeline techniques based on sensors technologies. J. King Saud. Univ. - Comput. Inf.
3. The SVR and ANN models presented higher generalization ability Sci. https://doi.org/10.1016/j.jksuci.2019.12.003.
than those of the Adaboost and RF models; Ayers, R.S., Westcot, D.W., 1994. Food, Agriculture Organization of the United Nations
(FAO), Water Quality for Agriculture, Irrigation and Drainage, Rome, Paper No. 29.
4. Overall, the sensitivity analysis shows that the developed models had
Rev1, M-56.
minor sensitivities to the input variables used compared to the range Baghvand, A., Nasrabadi, T., Bidhendi, G.N., Vosoogh, A., Karbassi, A., Mehrdadi, N.,
of each forecast parameter. 2010. Groundwater quality degradation of an aquifer in Iran central desert.
Desalination 260, 264–275. https://doi.org/10.1016/j.desal.2010.02.038.
Bensalah, M.K., Youbi, N., Mahmoudi, A., Bertrand, H., Mata, J., El Hachimi, H.,
ML models relying on physical parameters as features are efficient Madeira, J., Martins, L., Marzoli, A., Bellon, H., 2011. The central atlantic magmatic
tools and should be recommended for forecasting the IWQ parameters to province (CAMP) volcanic sequences of Berrechid and Doukkala basins (Western
manage the saline water. They will improve groundwater quality Meseta, Morocco): volcanology and geochemistry. Comun. Geol. 98, 98.
El Bilali, A., Abdeslam, T., Mazigh, N., Moukhliss, M., 2020. Prediction of chemical water
monitoring for irrigation purposes in real-time at a low cost. quality used for drinking purposes based on artificial neural networks. Moroc. J.
Chem. 3, 665–672.

11
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

Bishop, C.M., 2006. Pattern Recognition and Machine Learning. springer,. Gamar, A., Zair, T., El Kabriti, M., El Hilali, F., 2018. Study of the impact of the wild
Bonaccorso, G., 2018. Machine Learning Algorithms: Popular Algorithms for Data dump leachates of the region of El Hajeb (Morocco) on the physicochemical quality
Science and Machine Learning. Packt Publishing Ltd. of the adjacent water table. Karbala Int. J. Modern Sci. 4, 382–392. https://doi.org/
Bortolini, L., Maucieri, C., Borin, M., 2018. A tool for the evaluation of irrigation water 10.1016/j.kijoms.2018.10.002.
quality in the arid and semi-arid regions. Agronomy 8, 1–15. https://doi.org/ Gao, Y., Qian, H., Ren, W., Wang, H., Liu, F., Yang, F., 2020. Hydrogeochemical
10.3390/agronomy8020023. characterization and quality assessment of groundwater based on integrated-weight
El Bouqdaoui, K., Aachib, M., Blaghen, M., Kholtei, S., 2010. Modélisation de la pollution water quality index in a concentrated urban area. J. Clean. Prod. 260 https://doi.
par les nitrates de la nappe de Berrechid, au Maroc. Afr. Sci. Rev. Int. Des. Sci. org/10.1016/j.jclepro.2020.121006.
Technol. 5, 99–113. https://doi.org/10.4314/afsci.v5i1.61711. El Ghali, T., Marah, H., Qurtobi, M., Raibi, F., Bellarbi, M., Amenzou, N., El Mansouri, B.,
Breiman, L., 2001. Random forests. Mach. Learn. 45, 5–32. https://doi.org/10.1023/A: 2020. Geochemical and isotopic characterization of groundwater and identification
1010933404324. of hydrogeochemical processes in the Berrechid aquifer of central Morocco.
Burri, N.M., Weatherl, R., Moeck, C., Schirmer, M., 2019. A review of threats to Carbonates Evaporites 35, 1–21. https://doi.org/10.1007/s13146-020-00571-y.
groundwater quality in the anthropocene. Sci. Total Environ. 684, 136–154. https:// Ghosh, S., 2010. SVM-PGSL coupled approach for statistical downscaling to predict
doi.org/10.1016/j.scitotenv.2019.05.236. rainfall from GCM output. J. Geophys. Res. Atmos. 115, 1–18. https://doi.org/
Carlson, R.E., 1977. A trophic state index for lakes 1. Limnol. Oceanogr. 22, 361–369. 10.1029/2009JD013548.
Castrillo, M., García, Á.L., 2020. Estimation of high frequency nutrient concentrations Gosling, S.N., Arnell, N.W., 2016. A global assessment of the impact of climate change on
from water quality surrogates using machine learning methods. Water Res. 172, water scarcity. Clim. Change 134, 371–385. https://doi.org/10.1007/s10584-013-
115490 https://doi.org/10.1016/j.watres.2020.115490. 0853-x.
Chen, K., Chen, H., Zhou, C., Huang, Y., Qi, X., Shen, R., Liu, F., Zuo, M., Zou, X., Hastie, T., Tibshirani, R., Friedman, J., 2009. The Elements of Statistical Learning: Data
Wang, J., Zhang, Y., Chen, D., Chen, X., Deng, Y., Ren, H., 2020b. Comparative Mining, Inference, and Prediction. Springer Science & Business Media,.
analysis of surface water quality prediction performance and identification of key Hazan, R.M., Moullard, L., 1962. Notice Hydrogéologique de la Plaine de Berrechid. Off.
water parameters using different machine learning models based on big data. Water National Des Irrigation Services des Ressources en Eau,, Rabat.
Res. 171, 115454 https://doi.org/10.1016/j.watres.2019.115454. Heiß, L., Bouchaou, L., Tadoumant, S., Reichert, B., 2020. Index-based groundwater
Chen, C., He, W., Zhou, H., Xue, Y., Zhu, M., 2020a. A comparative study among machine vulnerability and water quality assessment in the arid region of Tata city (Morocco).
learning and numerical models for simulating groundwater dynamics in the Heihe Groundw. Sustain. Dev. 10, 100344 https://doi.org/10.1016/j.gsd.2020.100344.
River Basin, northwestern China. Sci. Rep. 10, 1–13. https://doi.org/10.1038/ Houéménou, H., Tweed, S., Dobigny, G., Mama, D., Alassane, A., Silmer, R., Babic, M.,
s41598-020-60698-9. Ruy, S., Chaigneau, A., Gauthier, P., Socohou, A., Dossou, H.J., Badou, S.,
Chou, J.S., Ho, C.C., Hoang, H.S., 2018. Determining quality of water in reservoir using Leblanc, M., 2020. Degradation of groundwater quality in expanding cities in West
machine learning. Ecol. Inform. 44, 57–75. https://doi.org/10.1016/j. Africa. A case study of the unregulated shallow aquifer in Cotonou. J. Hydrol. 582,
ecoinf.2018.01.005. 124438 https://doi.org/10.1016/j.jhydrol.2019.124438.
Chowdury, M.S.U., Emran, T., Bin, Ghosh, S., Pathak, A., Alam, M.M., Absar, N., El Houssine, E.L., Mridekh, A., EL Mansouri, B., Tammal, M., EL Bouhaddioui, M., 2014.
Andersson, K., Hossain, M.S., 2019. IoT based real-time river water quality Apport des données géophysiques et géologiques à la mise en évidence de nouveaux
monitoring system. Procedia Comput. Sci. 155, 161–168. https://doi.org/10.1016/j. éléments structuraux associés à la flexure de Settat (Maroc central) Contribution of
procs.2019.08.025. geophysical and geological data for the identification of new structural elements
Das, A., Das, S.S., Chowdhury, N.R., Joardar, M., Ghosh, B., Roychowdhury, T., 2020. related to the Settat flexure (central Morocco). Bull. l′ Institut Sci. Rabat 36,
Quality and health risk evaluation for groundwater in Nadia district, West Bengal: an 109–121.
approach on its suitability for drinking and domestic purpose. Groundw. Sustain. Hssaisoune, M., Bouchaou, L., Sifeddine, A., Bouimetarhan, I., Chehbouni, A., 2020.
Dev. 10, 100351 https://doi.org/10.1016/j.gsd.2020.100351. Moroccan groundwater resources and evolution with global climate changes.
Dawson, C.W., Wilby, R., 1998. Une approche de la modélisation pluie-deblt par ies Geoscience 10. https://doi.org/10.3390/geosciences10020081.
réseaux neuronaux artificiels. Hydrol. Sci. J. 43, 47–66. https://doi.org/10.1080/ Jahin, H.S., Abuzaid, A.S., Abdellatif, A.D., 2020. Using multivariate analysis to develop
02626669809492102. irrigation water quality index for surface water in Kafr El-Sheikh Governorate,
Di, Z., Chang, M., Guo, P., 2019. Water quality evaluation of the Yangtze River in China Egypt. Environ. Technol. Innov. 17 https://doi.org/10.1016/j.eti.2019.100532.
using machine learning techniques and data monitoring on different time scales. Jamaa, H., El Achheb, A., Namr, K.I., 2020. Spatial variation of groundwater quality and
Water 11. https://doi.org/10.3390/w11020339. assessment of water table fluctuations in Plio-Quaternary aquifer formations in
Doneen, L.D., 1964. Water Quality for Agriculture. Department of Irrigation University Doukkala Plain, Morocco. Groundw. Sustain. Dev. 11, 100398 https://doi.org/
California,, Davis, p. 48. 10.1016/j.gsd.2020.100398.
Droubi, A., Al-Sibai, M., Abdallah, A., Zahra, S., Obeissi, M., Wolfer, J., Huber, M., Kan, I., Rapaport-Rom, M., 2012. Regional blending of fresh and saline irrigation water:
Hennings, V., Schelkes, K., 2008. A decision support system (DSS) for water Is it efficient? Water Resour. Res. 48, 1–14. https://doi.org/10.1029/
resources management,–design and results from a pilot study in Syria. In: Climatic 2011WR011285.
Changes and Water Resources in the Middle East and North Africa. Springer, Kardos, J.S., Obropta, C.C., 2011. Water quality model uncertainty analysis of a point-
pp. 199–225. point source phosphorus trading program. J. Am. Water Resour. Assoc. 47,
Elfarrak, H., Hakdaoui, M., Fikri, A., 2014. Development of vulnerability through the 1317–1337. https://doi.org/10.1111/j.1752-1688.2011.00591.x.
DRASTIC method and geographic information system (GIS) (case groundwater of Kira, K., Rendell, L.A., 1992. A Practical Approach to Feature Selection, Machine
Berrchid), Morocco. J. Geogr. Inf. Syst. 06, 45–58. https://doi.org/10.4236/ Learning Proceedings 1992. Morgan Kaufmann Publishers, Inc,. https://doi.org/
jgis.2014.61006. 10.1016/b978-1-55860-247-2.50037-1.
Ewaid, S.H., Kadhum, S.A., Abed, S.A., Salih, R.M., 2019. Development and evaluation of Kopittke, P.M., So, H.B., Menzies, N.W., 2006. Effect of ionic strength and clay
irrigation water quality guide using IWQG V.1 software: a case study of Al-Gharraf mineralogy on Na–Ca exchange and the SAR–ESP relationship. Eur. J. Soil Sci. 57,
Canal, Southern Iraq. Environ. Technol. Innov. 13, 224–232. https://doi.org/ 626–633.
10.1016/j.eti.2018.12.001. Krupková, L., Havránková, K., Krejza, J., Sedlák, P., Marek, M.V., 2019. Impact of water
Fadili, A., Mehdi, K., Riss, J., Najib, S., Makan, A., Boutayab, K., 2015. Evaluation of scarcity on spruce and beech forests. J. Res. 30, 899–909. https://doi.org/10.1007/
groundwater mineralization processes and seawater intrusion extension in the s11676-018-0642-5.
coastal aquifer of Oualidia, Morocco: hydrochemical and geophysical approach. Kubat, M., 2017. An Introduction to Machine Learning. Springer International
Arab. J. Geosci. 8, 8567–8582. https://doi.org/10.1007/s12517-015-1808-5. Publishing,. https://doi.org/10.1007/978-3-319-63913-0.
Fadili, A., Najib, S., Mehdi, K., Riss, J., Makan, A., Boutayeb, K., Guessir, H., 2016. Kuhn, M., Johnson, K., 2019. Feature Engineering and Selection: A Practical Approach
Hydrochemical features and mineralization processes in coastal groundwater of for Predictive Models. CRC Press,. https://doi.org/10.1201/9781315108230.
Oualidia, Morocco. J. Afr. Earth Sci. 116, 233–247. https://doi.org/10.1016/j. Lamrani Alaoui, H., Oufdou, K., Mezrioui, N., 2008. Environmental pollutions impacts on
jafrearsci.2016.01.014. the bacteriological and physicochemical quality of suburban and rural groundwater
Fakir, Y., Zerouali, A., Aboufirassi, M., 2002. Exploitation et salinitk des aquiferes de la supplies in Marrakesh area (Morocco). Environ. Monit. Assess. 145, 195–207.
Chaouia c & i & e , littoral atlantique , Maroc ( Potential exploitation and salinity of https://doi.org/10.1007/s10661-007-0029-0.
aquifers , Chaouia coast, Atlantic shoreline, Morocco). J. Afr. Earth Sci. 32, 791–801. Liaw, A., Wiener, M., 2002. Classification and regression by randomForest. R News 2,
https://doi.org/10.1016/S0899-5362(02)00055-6. 18–22.
Fijani, E., Barzegar, R., Deo, R., Tziritis, E., Konstantinos, S., 2019. Science of the total Liu, P., Wang, J., Sangaiah, A., Xie, Y., Yin, X., 2019. Analysis and prediction of water
environment design and implementation of a hybrid model based on two-layer quality using LSTM deep neural networks in IoT environment. Sustainability 11,
decomposition method coupled with extreme learning machines to support real-time 2058. https://doi.org/10.3390/su11072058.
environmental monitoring of water quality parameters. Sci. Total Environ. 648, Lu, H., Ma, X., 2020. Hybrid decision tree-based machine learning models for short-term
839–853. https://doi.org/10.1016/j.scitotenv.2018.08.221. water quality prediction. Chemosphere 249, 126169. https://doi.org/10.1016/j.
Filho, F.G., da Silva Dias, N., Suddarth, S.R.P., Ferreira, J.F.S., Anderson, R.G., dos Santos chemosphere.2020.126169.
Fernandes, C., de Lira, R.B., Neto, M.F., Cosme, C.R., 2020. Reclaiming tropical Lyazidi, A., El Wartiti, M., Fadli, D, 2003. Évolution géodynamique du bassin triasique de
saline-sodic soils with gypsum and cow manure. Water 12. https://doi.org/10.3390/ Berrechid-ElGara-BenSlimane: Dynamique sédimentaire et géométrie des dépôts
w12010057. (Méseta nord occidentale, Maroc) Geodynamic evolution of the Berrechid-ElGara-
Freeze, R.A., Cherry, J.A., 1979. Groundwater. Prentice-Hall Inc,, Englewood Cliffs, NJ. BenSlimane triassic basin (north-western Meseta, Morocco): Depositional dynamics
Freund, Y., Schapire, R.E., 1997. A decision-theoretic generalization of on-line learning and geometry of sedimentary bodies.
and an application to boosting. J. Comput. Syst. Sci. 55, 119–139. https://doi.org/ El Mansouri, B., 1993. Structure et Modélisation Quantitative de l′ aquifère de Berrechid
10.1006/jcss.1997.1504. (Maroc): Validation par l′ approche géostatistique.
Freund, Y., Schapire, R.E., 1996. Experiments with a New Boosting Algorithm, in: Icml.
Citeseer, pp. 148–156.

12
A. El Bilali et al. Agricultural Water Management 245 (2021) 106625

El Mansouri, B., Dzikowski, M., Delay, F., Carlier, E., Crampon, N., 1992. Calage d′ un Siebert, S., Burke, J., Faures, J.-M., Frenken, K., Hoogeveen, J., Döll, P., Portmann, F.T.,
modèle mathématique en régime permanent appliqué à la nappe de Berrechid 2010. Groundwater use for irrigation – a global inventory. Hydrol. Earth Syst. Sci.
(Maroc). Ann. Soc. Géol. du Nord 189–193. 14, 1863–1880.
Meireles, A.C.M., Andrade, E.M., de, Chaves, L.C.G., Frischkorn, H., Crisostomo, L.A., Singh, S., Ghosh, N.C., Gurjar, S., Krishan, G., Kumar, S., Berwal, P., 2018. Index-based
2010. A new proposal of the classification of irrigation water. Rev. Ciênc. Agron. 41, assessment of suitability of water quality for irrigation purpose under Indian
349–357. https://doi.org/10.1590/s1806-66902010000300005. conditions. Environ. Monit. Assess. 190, 190. https://doi.org/10.1007/s10661-017-
Meyers, G., Kapelan, Z., Keedwell, E., 2017. Short-term forecasting of turbidity in trunk 6407-3.
main networks. Water Res. 124, 67–76. https://doi.org/10.1016/j. Sorensen, D.L., 1977. Suspended and Dissolved Solids Effects on Freshwater Biota: A
watres.2017.07.035. Review. US Environmental Protection Agency, Office of Research and Development
Molle, F., Tanouti, O., 2017. Squaring the circle: agricultural intensification vs. water ….
conservation in Morocco. Agric. Water Manag. 192, 170–179. https://doi.org/ Taloor, A.K., Pir, R.A., Adimalla, N., Ali, S., Manhas, D.S., Roy, S., Singh, A.K., 2020.
10.1016/j.agwat.2017.07.009. Spring water quality and discharge assessment in the Basantar watershed of Jammu
Moreno-Rodenas, A.M., Tscheikner-Gratl, F., Langeveld, J.G., Clemens, F.H.L.R., 2019. Himalaya using geographic information system (GIS) and water quality Index(WQI).
Uncertainty analysis in a large-scale water quality integrated catchment modelling Groundw. Sustain. Dev. 10 https://doi.org/10.1016/j.gsd.2020.100364.
study. Water Res. 158, 46–60. https://doi.org/10.1016/j.watres.2019.04.016. Tiyasha, T.M., Yaseen, Z.M., 2020. A survey on river water quality modelling using
Mountadar, S., Younsi, A., Hayani, A., Siniti, M., Tahiri, S., 2018. Groundwater artificial intelligence models: 2000–2020. J. Hydrol. 585, 124670 https://doi.org/
salinization process in the coastal aquifer Sidi Abed-Ouled Ghanem (Province of El 10.1016/j.jhydrol.2020.124670.
Jadida, Morocco). J. Afr. Earth Sci. 147, 169–177. https://doi.org/10.1016/j. Tyagi, S., Sharma, B., 2014. Water quality assessment in terms of water quality index
jafrearsci.2018.06.025. water. Am. J. Water Resour. 1 (3), 34–38. https://doi.org/10.12691/ajwr-1-3-3.
El Mountassir, O., Bahir, M., Ouazar, D., Ouhamdouch, S., Chehbouni, A., Ouarani, M., Vapnik, V.N., 1995. The Nature Of Statistical Learning Theory.
2020. The use of GIS and water quality index to assess groundwater quality of krimat Wagh, V.M., Panaskar, D.B., Muley, A.A., Mukate, S.V., Lolage, Y.P., Aamalawar, M.L.,
aquifer (Essaouira; Morocco). SN Appl. Sci. 2 https://doi.org/10.1007/s42452-020- 2016. Prediction of groundwater suitability for irrigation using artificial neural
2653-z. network model: a case study of Nanded tehsil, Maharashtra, India. Model. Earth Syst.
Moyé, J., Picard-Lesteven, T., Zouhri, L., El Amari, K., Hibti, M., Benkaddour, A., 2017. Environ. 2, 1–10. https://doi.org/10.1007/s40808-016-0250-3.
Groundwater assessment and environmental impact in the abandoned mine of Walker, K.W., Jiang, Z., 2019. Application of adaptive boosting ( AdaBoost ) in demand-
Kettara (Morocco). Environ. Pollut. 231, 899–907. https://doi.org/10.1016/j. driven acquisition ( DDA ) prediction : a machine-learning approach. J. Acad.
envpol.2017.07.044. Librariansh. 45, 203–212. https://doi.org/10.1016/j.acalib.2019.02.013.
Oshiro, T.M., Perez, P.S., Baranauskas, J.A., 2012. How many Trees in a Random Forest? Wang, L., Long, F., Liao, W., Liu, H., 2020. Prediction of anaerobic digestion performance
Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes and identification of critical operational parameters using machine learning
Bioinformatics) 7376 LNAI, 154–168. https://doi.org/10.1007/978–3-642–3153 algorithms. Bioresour. Technol. 298, 122495 https://doi.org/10.1016/j.
7-4_13. biortech.2019.122495.
Ouassissou, R., Kuper, M., Dugué, P., Amrani, M. El, Hammani, A., Ameur, F., 2019. Wu, J., Zhang, Y., Zhou, H., 2020. Groundwater chemistry and groundwater quality
Rivalries and cooperative arrangements for access to groundwater in the Berrechid index incorporating health risk weighting in dingbian County, Ordos basin of
plain in Morocco. Cah. Agric. 28 https://doi.org/10.1051/cagri/2019006. northwest China. Chem. der Erde, 125607. https://doi.org/10.1016/j.
Radwan, M., Willems, P., Berlamont, J., 2004. Sensitivity and uncertainty analysis for chemer.2020.125607.
river quality modelling. J. Hydroinform. 6, 83–99. https://doi.org/10.2166/ Yaduvanshi, N.P.S., Swarup, A., 2005. Effect of continuous use of sodic irrigation water
hydro.2004.0008. with and without gypsum, farmyard manure, pressmud and fertilizer on soil
Rajaee, T., Ebrahimi, H., Nourani, V., 2019. A review of the artificial intelligence properties and yields of rice and wheat in a long term experiment. Nutr. Cycl.
methods in groundwater level modeling. J. Hydrol. 572, 336–351. https://doi.org/ Agroecosyst. 73, 111–118. https://doi.org/10.1007/s10705-005-3361-1.
10.1016/j.jhydrol.2018.12.037. Yaseen, Z.M., Sulaiman, S.O., Deo, R.C., Chau, K.W., 2019. An enhanced extreme
Richards, L.A., 1954. Diagnosis and Improvement of, Saline and Alkali Soils. Handbook. learning machine model for river flow forecasting: State-of-the-art, practical
https://doi.org/ttps://www.ars.usda.gov/ARSUserFiles/20360500/hb60_pdf/ applications in water resource engineering area and future research direction.
hb60complete.pdf. J. Hydrol. 569, 387–408. https://doi.org/10.1016/j.jhydrol.2018.11.069.
Ruhard, J.P., 1975. Chaouia et plaine de Berrechid. Ressources en eau du Maroc 2. Yoon, H., Jun, S.C., Hyun, Y., Bae, G.O., Lee, K.K., 2011. A comparative study of artificial
Saghafi, H., Arabloo, M., 2017. Modeling of CO 2 solubility in MEA , DEA , TEA , and neural networks and support vector machines for predicting groundwater levels in a
MDEA aqueous solutions using adaboost-decision tree and artificial neural network. coastal aquifer. J. Hydrol. 396, 128–138. https://doi.org/10.1016/j.
Int. J. Greenh. Gas Control 58, 256–265. https://doi.org/10.1016/j. jhydrol.2010.11.002.
ijggc.2016.12.014. Zaman, M., Shahid, S.A., Heng, L., 2018. Guideline for Salinity Assessment, Mitigation
Sakai, N., Mohamad, Z.F., Nasaruddin, A., Abd Kadir, S.N., Mohd Salleh, M.S.A., and Adaptation Using Nuclear and Related Techniques, Guideline for Salinity
Sulaiman, A.H., 2018. Eco-heart index as a tool for community-based water quality Assessment, Mitigation and Adaptation Using Nuclear and Related Techniques. https
monitoring and assessment. Ecol. Indic. 91, 38–46. https://doi.org/10.1016/j. ://doi.org/10.1007/978–3-319–96190-3.
ecolind.2018.03.079. Zhao, X., Ning, B., Liu, L., Song, G., 2013. Correspondence to : Advances in Space
Schalkoff, R.J., 1997. Artificial Neural Networks. McGraw-Hill Higher Education,. Research https://doi.org/10.1016/j.asr.2013.12.001.
Seif-Ennasr, M., Zaaboul, R., Hirich, A., Caroletti, G.N., Bouchaou, L., El Morjani, Z.E.A., Zhu, J. jun, Yu, L. zhong, Xu, T. le, Wei, X., Yang, K., 2019. Comparison of water quality
Beraaouz, E.H., McDonnell, R.A., Choukr-Allah, R., 2016. Climate change and in two catchments with different forest types in the headwater region of the Hun
adaptive water management measures in Chtouka Aït Baha region (Morocco). Sci. River, Northeast China. J. Res. 30, 565–576. https://doi.org/10.1007/s11676-018-
Total Environ. 573, 862–875. https://doi.org/10.1016/j.scitotenv.2016.08.170. 0688-4.
Shojaei, M., Nazif, S., Kerachian, R., 2015. Joint uncertainty analysis in river water
quality simulation: a case study of the Karoon River in Iran. Environ. Earth Sci. 73,
3819–3831. https://doi.org/10.1007/s12665-014-3667-x.

13