remote sensing

Article
Machine Learning Algorithms for Modeling and Mapping of
Groundwater Pollution Risk: A Study to Reach Water Security
and Sustainable Development (Sdg) Goals in a Mediterranean
Aquifer System
Safae Ijlil 1 , Ali Essahlaoui 1 , Meriame Mohajane 1,2, * , Narjisse Essahlaoui 1 , El Mostafa Mili 1
and Anton Van Rompaey 3

1 Laboratory of Geoengineering and Environment, Research Group “Water Sciences and Environment
Engineering”, Department of Geology, Faculty of Sciences, Moulay Ismail University,
Meknes B.P.11201, Morocco; [email protected] (S.I.); [email protected] (A.E.);
[email protected] (N.E.); [email protected] (E.M.M.)
2 Research Group “Soil and Environment Microbiology”, Department of Biology, Faculty of Sciences,
Moulay Ismail University, Meknes B.P.11201, Morocco
3 Geography and Tourism Research Group, Department Earth and Environmental Science, KU Leuven,
Celestijnenlaan 200E, 3001 Heverlee, Belgium; [email protected]
* Correspondence: [email protected]

Abstract: Groundwater pollution poses a severe threat and issue to the environment and humanity
overall. That is why mitigative strategies are urgently needed. Today, studies mapping ground-
Citation: Ijlil, S.; Essahlaoui, A.; water risk pollution assessment are being developed. In this study, five new hybrid/ensemble
Mohajane, M.; Essahlaoui, N.; Mili, machine learning (ML) models are developed, named DRASTIC-Random Forest (RF), DRASTIC-
E.M.; Van Rompaey, A. Machine
Support Vector Machine (SVM), DRASTIC-Multilayer Perceptron (MLP), DRASTIC-RF-SVM, and
Learning Algorithms for Modeling
DRASTIC-RF-MLP, for groundwater pollution assessment in the Saiss basin, in Morocco. The per-
and Mapping of Groundwater
formances of these models are evaluated using the Receiver Operating Characteristic curve (ROC
Pollution Risk: A Study to Reach
curve), precision, and accuracy. Based on the results of the ROC curve method, it is indicated that
Water Security and Sustainable
Development (Sdg) Goals in a
the use of hybrid/ensemble machine learning (ML) models improves the performance of the indi-
Mediterranean Aquifer System. vidual machine learning (ML) algorithms. In effect, the AUC value of the original DRASTIC is 0.51.
Remote Sens. 2022, 14, 2379. Furthermore, both hybrid/ensemble models, DRASTIC-RF-MLP (AUC = 0.953) and DRASTIC-RF-
https://doi.org/10.3390/rs14102379 SVM, (AUC = 0.901) achieve the best accuracy among the other models, followed by DRASTIC-RF
(AUC = 0.852), DRASTIC-SVM (AUC = 0.802), and DRASTIC-MLP (AUC = 0.763). The results de-
Academic Editors: Biswajeet Pradhan
lineate areas vulnerable to pollution, which require urgent actions and strategies to improve the
and Seyed Amir Naghibi
environmental and social qualities for the local population.
Received: 20 April 2022
Accepted: 10 May 2022 Keywords: groundwater pollution risk; vulnerability; DRASTIC; machine learning algorithms;
Published: 15 May 2022
bivariate statistic; Saiss basin
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations. 1. Introduction
Groundwater resources represent a precious source for the life of both humans and
animals [1,2]. According to UNICEF [3], it is estimated that about 2.2 billion people globally
still have no access to drinking water supply [4]. Groundwater pollution is a worldwide
Copyright: © 2022 by the authors.
challenge, specifically in arid and semi-arid regions [5,6].
Licensee MDPI, Basel, Switzerland.
Groundwater pollution is associated with the uncontrolled and irrational use of
This article is an open access article
agrochemicals for agricultural activities [7,8]. Therefore, the applicability of pesticides
distributed under the terms and
conditions of the Creative Commons
leads to dangerous effects on both human life and environmental ecosystems [9]. Thus,
Attribution (CC BY) license (https://
the use of pesticides for agricultural practices presents a double-edged sword and farmers
creativecommons.org/licenses/by/
should be sensitize about following an integrated approach of crop insurance and farmers’
4.0/).

Remote Sens. 2022, 14, 2379. https://doi.org/10.3390/rs14102379 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2022, 14, 2379 2 of 19

selection of pesticide input [10]. This may be complicated by drought related to climate
change impacts, which affect crop yields and decrease groundwater resources [10].
The uncontrolled use of nitrogen fertilizers catalyzes the nitrate leaching, leading to
groundwater pollution [11]. The nitrates (NO3 −N) are the most popular indicator for
assessing groundwater pollution due to their severe effect on the aquifer; furthermore,
they can lead to many effects including methaemoglobinaemia. Ranking the risk of nitrate
concentration pollution on a groundwater aquifer is done by the World Health Organization
(WHO), i.e., an aquifer receiving a concentration of nitrate exceeding 50 mg/L, is considered
polluted [12]. This situation forces researchers to develop a strategy to select areas with
the highest risk to vulnerability. Thus, the establishment of spatial vulnerability tools
constitutes a crucial opportunity to urgently and easily delineate areas that are highly
vulnerable to groundwater pollution and, in doing so, to highlight a need for adequate
strategies and decision-making actions to sustainably manage the groundwater resources.
A major portion of modeling studies is complex due to the processing of different data
and their size. In this sense, with the many opportunities of the geographic information
system (GIS) tools and spatial operations, GIS tools have shown its applicability as a useful
space for preprocessing different kinds of spatiotemporal datasets [13,14]. These opportu-
nities greatly assist researchers in achieving their purposes in environmental studies.
Numerous studies have been conducted to assess groundwater vulnerability to pol-
lution based on different methods [15]. Among these, the DRASTIC approach developed
by Aller et al. [16], based on seven parameters, including depth to water (D), recharge
(R), aquifer media (A), soil media (S), topography (T), impact of the vadose zone (I), and
hydraulic conductivity (C), has yielded promising outcomes in many research cases [17,18].
However, the original version of the DRASTIC method has its limitations [19]. As pointed
out by Nadiri et al. [20], each aquifer, unconfined or confined, presents its own character-
istics (i.e., the aquifer type in terms of geological composition and confined/unconfined
nature and the water table depth). Thus, we argue that the applicability of the DRAS-
TIC method in an area presenting both unconfined or confined aquifers becomes difficult
regarding the assignments of vulnerability risk and it cannot be generalized.
The fact that each user adopts and adjusts the methodology based on the conditions
of the background of the area studied may be another drawback of the applicability of the
DRASTIC method. There is no conventional guideline to follow to decide whether the
vulnerability index is “right” or “wrong” and the level of the exactitude of the vulnerability
index remains unknown. Because of these uncertainties, many researchers have modified
the original DRASTIC method to increase its accuracy using several techniques. These in-
clude frequency ratio [21], Analytical Hierarchy Process (AHP) [22], Fuzzy-AHP model [1],
Single-Parameter Sensitivity Analysis [23], Mamdani Fuzzy Logic (MFL) model [24] and
Supervised Committee with Fuzzy Logic model [24], projection pursuit dynamic clustering,
and also anthropogenic influence, such as land cover/land use impacts [25].
More recently, hybrid ensemble machine learning (ML) models (i.e., a combination of
single methods and statistical techniques) have been developed with promising results in
different environmental hazard fields [26,27], such as flood [28,29], landslide [30,31], forest
fire susceptibility [32], and groundwater level prediction [33].
To date, there has been no prior study on the use of hybrid machine learning (ML)
algorithms (i.e., combination of machine algorithm models and statistical tests) as an
opportunity for groundwater pollution studies. In the Saiss basin, where this research was
conducted, the local population is largely dependant on agriculture to serve as a source of
their livelihood. However, the unsustainable use of pesticides is increasingly affecting the
quality and the quantity of water resources in this aquifer.
In this ecosystem, previous efforts have been made by other authors. For instance,
Sadkaoui et al. [34] used the DRASTIC, GOD, and PRK methods to assess groundwater
vulnerability. They reported the sensitivity of water resources due to the pollution coming
from the ground surface. Moreover, a recent article developed by Lahjouj et al. [35] used
random forest (RF) to map groundwater vulnerability with encouraging outcomes. With
 In this ecosystem, previous efforts have been made by other authors. For instance,
Sadkaoui et al. [34] used the DRASTIC, GOD, and PRK methods to assess groundwater
Remote Sens. 2022, 14, 2379 vulnerability. They reported the sensitivity of water resources due to the pollution coming 3 of 19
from the ground surface. Moreover, a recent article developed by Lahjouj et al. [35] used
random forest (RF) to map groundwater vulnerability with encouraging outcomes. With
respect to
respect to the
the above
abovecited
citedarticles, thethe
articles, work presented
work herehere
presented takes a step
takes a further in thein
step further field
the
of hybridization of machine learning (ML) models and statistical tests. In this
field of hybridization of machine learning (ML) models and statistical tests. In this case, case, we
proposed
we proposed five new
five newensemble
ensemblemodels,
models,DRASTIC-RF
DRASTIC-RF(random
(randomforest),
forest), DRASTIC-SVM
DRASTIC-SVM
(support vector machine), DRASTIC-MLP (multilayer perceptron), DRASTIC-RF-SVM,
(support vector machine), DRASTIC-MLP (multilayer perceptron), DRASTIC-RF-SVM,
and DRASTIC-RF-MLP
and DRASTIC-RF-MLP for for groundwater
groundwater vulnerability
vulnerability pollution
pollutionriskriskassessment
assessmentininthethe
Saiss basin.
Saiss basin.

2.2.Materials
Materials and
and Methods
Methods
2.1.Study
2.1. Study Area
Area
This research
This research was
was conducted
conducted in the Saiss
Saiss basin,
basin, located
locatedin inthe
theFez-Meknes
Fez-Meknesregion,
region,
Morocco (Figure
Morocco (Figure 1),
1), between
between latitudes
latitudes 33 ◦ 380 N and 34°4′
33°38′ 34◦ 40 N N and
and longitudes 5◦ 490W
longitudes5°49′ Wand
and
◦ 0
4°53′ W. The basin occupies
4 53 W. The basin occupies an area of about 2100 km 2
2 .. The elevation ranges from 212toto
The elevation ranges from 212
1047m.
1047 m. According
According to to Koppen
Koppen Climate Classification,
Classification, the the study
studyarea
areaisischaracterized
characterizedby by
Mediterranean climate
Mediterranean climate and it has an average annual
annual temperature
temperatureof of17.2 ◦ C,recorded
17.2°C, recordedinin
Meknes station.
Meknes station. The average annual
annual precipitation
precipitation isis about
about 589.3
589.3 mm.
mm. According
Accordingtotothethe
censusreport
census reportofof 2014,
2014, the
the population
population is about 2.3 million inhabitants.
inhabitants.

Figure1.1.Location
Figure Location and
and geological
geological map
map of
of the
the Saiss
Saiss basin,
basin, (a)
(a) location
locationof
ofthe
thestudy
studyarea,
area,Fes-Meknes
Fes-Meknes
region, Morocco. (b) Geological map of the studied area (modified from [36]).
region, Morocco. (b) Geological map of the studied area (modified from [36]).

From the geological point of view, the Saiss comprises several formations extending
from the Palaeozoic to the Quaternary, the majority of which are Pliocene Lake limestones
and fawn sands [37]. The hydrographic network of the basin consists of four main wadis,
which are Oued El Kell, Oued Mikkes, Oued R’Dom, and Oued Fes. According to Es-
sahlaoui et al. [37], the Saiss aquifer system presents a complex structure of two aquifers:
Remote Sens. 2022, 14, 2379 4 of 19

the Plioquaternary phreatic aquifer composed of Pliovillafranchian sandstone, conglomer-

ated sand and lacustrine limestone, and the liassic aquifer composed of dolomitic limestone.
In this aquifer, as agriculture is still going to be one of the main sources of the local popula-
tion, to meet the livelihood’s resource requirements, more comprehensive studies should be
carried out to promote the sustainable use of natural resources for the proper management
and planning of them.

2.2. Data Used and Methodology

The methodology followed in this work is composed of the following stages. The
first step of this study was the use of DRASTIC method to map and assess groundwater
vulnerability to pollution. Next, the bivariate statistic frequency ratio was applied to
modify the result of DRASTIC method. Then, five new hybrid/ensemble machine learning
models, namely DRASTIC-RF, DRASTIC-SVM, DRASTIC-MLP, DRASTIC-RF-SVM, and
DRASTIC-RF-MLP, were proposed and developed in Python and the Jupyter Notebook
Environment. Finally, groundwater vulnerability pollution risk maps were carried out.

2.3. Assessment of Groundwater Vulnerability to Pollution Using DRASTIC Method

2.3.1. DRASTIC Method and Parameters Description
DRASTIC method is a widely applied approach to assess groundwater vulnerabil-
ity [21]. It is based on seven factors (Figure 2), including depth to groundwater (D), net
recharge (R), aquifer media (A), soil media (S), topography (T), impact of the vadose zone (I),
and hydraulic conductivity (C). Following the classification established by Aller et al. [16],
different rates (r) and weights (w) are assigned to each class of each DRASTIC parameter.
The DRASTIC method was calculated using Equation (1):

DRASTIC = Dr Dw + Rr R + Ar Aw + Sr S + Tr Tw + Ir Iw + Cr C (1)

where r is the parameter rate and w is the parameter weight.

• Depth to groundwater (D)
According to Khosravi [21], the (D) parameter presents the distance measured from
ground surface to water table. It is considered a limiting factor for groundwater vul-
nerability, because it conditions the transfer process of pollutant and its possibility of
degradation [17]. The greater the depth of water level, the lower the risk of groundwater
vulnerability to pollution [16]. In our case, the depth to water map (Figure 2a) was derived
from the Digital Elevation Model (DEM) with a pixel size of 30 m and the piezometric map.
Based on the DRASTIC framework range value, the map generated was classified into
7 classes, including 0–1.5 m, 1.5–4.5 m, 4.5–9 m, 9–15 m, 15–33 m, 23–31 m, and >31 m.
• Recharge (R)
The recharge is a hydrological process and corresponds to the amount of water that in-
filtrates through the surface of the ground and contributes to the recharge of the aquifer [16].
The increase in net recharge leads to an increase in the risk of contamination of groundwater.
This parameter is related to the topography and the nature of the geological formations.
Net Recharge of the Plio-Quaternary aquifer of the Saiss basin is mainly contributed by
precipitation and infiltration of irrigation water, as well as by the drainage of the Liasic
aquifer of the Middle Atlassic Causse in the southern part of the Saiss basin [34]. The
equation for calculating the net recharge was given by Scanlon et al. [38] as:

∆h
R = Sy (2)
∆t
where Sy represents the specific yield, ∆h represents the differences of the water-table
height for the highest and lowest tables, and ∆t represents the interval time for those tables.
The net recharge map obtained was divided into 3 classes, including 0–50, 50–100, and
100–180 mm (Figure 2b).
Remote
RemoteSens. 2022,14,
Sens.2022, 14,2379
x FOR PEER REVIEW 5 of2119
5 of

Figure 2.
Figure 2. Maps
Mapsof groundwater vulnerability
of groundwater conditioning
vulnerability factors:factors:
conditioning (a) depth(a)
to groundwater, (b) net
depth to groundwa-
recharge, (c) aquifer media, (d) Soil, (e) topography, (f) impact of vadose zone, (g) hydraulic
ter, (b) net recharge, (c) aquifer media, (d) Soil, (e) topography, (f) impact of vadose zone,
conductivity.
(g) hydraulic conductivity.

•• Depth tomedia
Aquifer groundwater
(A) (D)
According to Khosravi [21], the (D) parameter presents the distance measured from
The aquifer environment or the saturated zone (see in Figure 2c) influences the vulner-
ground surface to water table. It is considered a limiting factor for groundwater
ability to pollution because its properties make it possible to control the concentration of
vulnerability, because it conditions the transfer process of pollutant and its possibility of
pollutants by diluting them. The Saiss aquifer consists of lacustrine formations, conglomer-
degradation [17]. The greater the depth of water level, the lower the risk of groundwater
ates, sandstones, and sands. These formations are extracted from the Hydrogeological map
vulnerability to pollution [16]. In our case, the depth to water map (Figure 2a) was derived
of the Meknes-Fes basin obtained from Morocco’s National Irrigation Office, Directorate
from the Digital Elevation Model (DEM) with a pixel size of 30 m and the piezometric
General of Studies (1/100,000).
map. Based on the DRASTIC framework range value, the map generated was classified
•into Soil (S) including 0–1.5 m, 1.5–4.5 m, 4.5–9 m, 9–15 m, 15–33 m, 23–31 m, and >31
7 classes,
m. Soil texture affects the amount of infiltration from ground surface. In this study, this
• Recharge
parameter was(R)
constructed using the pedological map of central Morocco obtained from
the national institute
The recharge is aofhydrological
agronomic research, physical
process and environment
corresponds to thedepartment
amount of water(1/500,000).
that
The soils of
infiltrates this study
through area were
the surface divided
of the groundinto
and3contributes
classes (seetoFigure 2d): clay,
the recharge of theclay loam,
aquifer
and
[16].sand. Areas with
The increase in sand are characterized
net recharge by increase
leads to an high permeability,
in the riskwhereas clay areas of
of contamination are
characterized by lower permeability.
groundwater. This parameter is related to the topography and the nature of the geological
•formations.
TopographyNet Recharge
(T) of the Plio-Quaternary aquifer of the Saiss basin is mainly
contributed by precipitation and infiltration of irrigation water, as well as by the drainage
Topography
of the parameter
Liasic aquifer plays an
of the Middle important
Atlassic rolein
Causse inthe
the southern
infiltration at the
part ground
of the Saisssurface.
basin
A lower slope results in more infiltration and therefore a higher potential for
[34]. The equation for calculating the net recharge was given by Scanlon et al. [38] as: contamination.
The slope of the Saiss basin (see in Figure 2e) was extracted from Digital Elevation Model
with pixel size 30 m and it ranges from 0% to 112%.
Remote Sens. 2022, 14, 2379 6 of 19

• Impact of Vadose zone (I)

The vadose zone (see in Figure 2f) represents the zone located between the surface
of the earth and the upper part of the aquifer; using its property of permeability, it can
determine the potential of groundwater contamination. In this research, this layer is
constituted by Quaternary formation, such as: impermeable clays, alluviums, travertines,
and basalts.
Geological data are obtained from the Hydrogeological map of the Meknes-Fes
basin obtained from Morocco’s National Irrigation Office, Directorate General of Stud-
ies (1/100,000).
• Hydraulic conductivity (I)
Hydraulic conductivity or permeability presents the capacity of an aquifer to transmit
water in a porous medium. The higher C is, the faster the pollutant will be transported. In
our case, the hydraulic conductivity is given by Equation (3) [39]:

t
K= , (3)
b

where K represents the hydraulic conductivity (m/s), t represents the transmissivity (m2 /s),
and b is the aquifer thickness (m). The hydraulic conductivity values in the Sais basin
ranged from 0 to 30 m/day (Figure 2g). These values were classified into 4 classes: 0.04–4,
4–12, 12–29, and 29–41 (m/s).

2.3.2. Frequency Ratio

Due to its ease of applicability and understanding [40], the bivariate statistic test
frequency ratio is widely used in several hazard monitoring studies [41]. In this study,
it was used to understand the spatial correlation between each parameter of DRASTIC
method and the nitrate points.
Sf
∆Sf
FR = Nf
, (4)
εN
where Sf represent the area of each class for each DRASTIC parameter; ∆Sf represent the
total area of each parameter; Nf represent concentration of nitrate occurrence in the class of
each parameter; εN represent the number of total nitrates in the study area. The correlation
between the independent variable, i.e., DRASTIC parameter and the target, i.e., nitrate
sample is highest if the FR is >1, whereas the correlation is low if the FR is <1.
The FR was calculated for the rate of each class of each DRASTIC parameter and
reclassified using natural breaks method in spatial analyst tools in ArcGIS 10.5 (ESRI;
Redlands, CA, USA). Thus, we proposed an improved result for groundwater vulnerability
based on statistical approach (Table 1). The frequency ratio of each DRASTIC parameter
was normalized between 0 and 1 using Equation (5):

x − minx
x0 = , (5)
max(x) − min(x)

where x is the current value of the variable and x’ is the normalized value.
In this study, these seven FR-DRASTIC parameters served as the explicative vari-
ables for groundwater vulnerability modeling. Thus, the dataset was randomly split into
70 points which served as training sample (70%) and 30 points which served as validation
sample (30%). The seven DRASTIC parameters were applied as the model inputs and the
nitrate values used as the target of the model after normalization process were randomly
split into 70% training data and 30% validation data.

2.4. Preparation of Nitrate Locations’ Data and Validation

In this study, a total of 100 well samples (Figure 3) were selected to collect nitrate
concentration. They were taken from the Sebou Hydraulic Basin Agency (SHBA). For the
Remote Sens. 2022, 14, 2379 7 of 19

analysis, the inverse distance weighted (IDW) interpolation was performed for the samples.
The result of nitrate mapping is presented in Figure 3. Following the recommendations
of the WHO, i.e., an aquifer receiving a concentration of nitrate exceeding 50 mg/L is
considered polluted. To validate the machine learning models, nitrate concentration data,
which served as target in this study, were divided into 2 groups: locations with nitrate
concentrations higher than 50 mg/L were classified as polluted areas, whereas those
with nitrate concentrations less than 50 mg/L were classified as unpolluted areas. We
randomly separated groundwater polluted areas and groundwater unpolluted areas; out
of 100 samples of nitrate concentrations, 70% were used as training dataset and 30% were
used as validation dataset.

Table 1. Frequency ratio of DRASTIC parameters.

DRASTIC Factors Class Weight No of Nitrate Point No. of Pixels in Class FR

>31 11 684,246 0.20
23–31 3 509,395 0.07
15–23 15 1,051,753 0.18
D: Depth to water table (m) 9–15 5 11 468,340 0.29
4.5–9 3 139,093 0.27
1.5–4.5 0 13,853 0
0–1.5 0 994 0
0–50 4 303,233 0.09
R: Net Recharge (mm) 50–100 4 38 2,371,682 0.88
100–180 1 191,799 0.02
Limestone 34 2,497,029 0.79
A: Aquifer media Conglomerate 3 1 67,596 0. 02
Sand and gravel 8 302,011 0.19
Clay 15 915,874 0.35
S: Soil Clay loam 2 26 1,712,526 0.60
Sand 2 239,594 0.05
>18 1 220,003 0.08
12–18 2 267,960 0.14
T: Slope (◦ ) 6–12 1 16 940,573 0.31
2–6 22 1,148,493 0.35
0–2 2 290,069 0.13
Alluvium 7 361,430 0.27
Vindobonion clays 1 282,347 0.05
Limestone 24 1,693,703 0.20
I: Impact of vadose zone 5
Sandstone and
10 484,631 0.28
Conglomerates
Basalt 1 67,932 0.20
0.04–4 21 1,182,660 0.49

C: Hydraulic conductivity 4–12 14 1,255,673 0.33

3
(m/day) 12–29 8 361,236 0.19
29–41 0 1459 0
 nitrate concentration data, which served as target in this study, were divided into 2
groups: locations with nitrate concentrations higher than 50 mg/L were classified as
polluted areas, whereas those with nitrate concentrations less than 50 mg/L were classified
as unpolluted areas. We randomly separated groundwater polluted areas and
Remote Sens. 2022, 14, 2379 groundwater unpolluted areas; out of 100 samples of nitrate concentrations, 70% were 8 of 19
used as training dataset and 30% were used as validation dataset.

Spatialdistribution
Figure3.3.Spatial
Figure distributionofofnitrate
nitrateconcentrations
concentrationsin
inthe
theSaiss
Saissaquifer.
aquifer.

2.5. Algorithm Background and Implementations

2.5. Algorithm Background and Implementations
2.5.1. Support Vector Machine (SVM)
2.5.1. Support Vector Machine (SVM)
SVM was developed for the first time by Vapnik with colleagues [42,43]. It is used
Remote Sens. 2022, 14,
SVM x FORwas
PEERdeveloped
REVIEW for theThe
firstgoal
timeofby Vapnik with colleagues 9 of 21
for classification and problems. SVM algorithm is to find a[42,43]. It is used
hyperplane in an
for classification space
N-dimensional and problems. The
(Figure 4). goalisof
SVM SVM algorithm
widely used with issuccessful
to find a hyperplane in an
results in ground-
N-dimensional space (Figure 4). SVM is widely used with successful results
water potential mapping [44,45], landslide mapping [46], and flood susceptibility assess- in
groundwater potential mapping [44,45], landslide mapping [46], and flood susceptibility
ment [47,48]. In assessment
this study,[47,48].
the kernel function
In this was
study, the used
kernel due towas
function its robustness
used due to in
its previous
robustness in
studies [49,50]. previous studies [49,50].

Figure 4. Representation of the SVM Margin.

Figure 4. Representation of the SVM Margin.
2.5.2. Random Forest (RF)
Developed for the first time by Breiman [51], RF is a supervised machine learning
algorithm, based on ensemble technique [52], used for both regression and classification
tasks [47]. According to Rahmati et al. [6], running the RF algorithm requires two
optimized parameters: the number of variables/factors to be used in each tree-building
process (mtry) and the number of trees to be built in the forest to run it (ntree).
Remote Sens. 2022, 14, 2379 9 of 19

2.5.2. Random Forest (RF)

Developed for the first time by Breiman [51], RF is a supervised machine learning
algorithm, based on ensemble technique [52], used for both regression and classification
tasks [47]. According to Rahmati et al. [6], running the RF algorithm requires two optimized
parameters: the number of variables/factors to be used in each tree-building process (mtry)
and the number of trees to be built in the forest to run it (ntree).
In this research, the number of trees was considered as 123 with one random division
variable, which led to the highest precision.
RF algorithm was used to evaluate many natural environmental hazard studies, be-
cause of its reliability and efficiency [29,53]. RF algorithm was used to evaluate many
natural environmental hazards studies, such as gully erosion modelling [54], forest fire
mapping [53], and flood susceptibility prediction [55,56].

2.5.3. Multilayer Perceptron-Neural Network (MLP-NN)

Multilayer Perceptron-Neural network MLP-NN is a class of Neural Networks algo-
rithms (ANN) with a structure similar to biological neurons. The structure of MLP-NN
models consists of three layers: an input layer, an output layer, and a hidden layer [57,58].
The input layers represent the factors conditioning the risk of pollution, the output layers
are considered as classified results that infer the classes of pollution or non-pollution by
nitrates, whereas the hidden layer is located between these two, in which the function
applies weights. Meanwhile, the hidden layers perform non-linear transformations by
applying weights to the inputs and directing them through an activation function as output.
The training of this algorithm takes place in three steps: feedforward, backpropagation,
and weight adjustment [59,60].
In the current study, in order to achieve the highest performance and to avoid the
overfitting problem, two hidden layers were used with 37 neurons in each layer, the Linear
Unit Rectification (Relu) was used as activation function, and the optimizer was set to
Adam, which was developed by Kingma and Ba [61].

2.6. Validation of Groundwater Vulnerability Models

Evaluating the performance of machine learning models is a mandatory step in the
modeling research [52,62]. Common statistical approaches were used by several researchers
in related studies [15,55].
In this article, using the training and validation datasets, the receiver operating curve
(ROC) was applied to evaluate the performance of the groundwater pollution risk models
developed in this study. The ROC is a graphical presentation in which the specificity
is plotted on the X-axis and the sensitivity is plotted on the Y-axis at different cut-off
thresholds classifications [15,63]. To quantitatively validate the models, the area under
the curve (AUC) is often used. It is defined as “the probability of a classifier to correctly
anticipate the occurrence or non-occurrence of predefined events”.
In addition, other statistical measures were used, including Sensitivity, Specificity,
Accuracy (ACC), and Precision:

TP + TN
Accuarcy= , (6)
TP + TN + FP + FN
TP
Sensitivity= , (7)
TP + FN
TN
Specificity= , (8)
TN + FP
TP
Precision= , (9)
TP + FP
where FP (false positive) is the number of groundwater points incorrectly predicted and
considered as polluted and FN (false negative) is the number of groundwater points
 Remote Sens. 2022, 14, 2379 10 of 19

considered and predicted as non-polluted; TP (true positive) is the number of nitrate

pollution samples classified correctly; and TN (true negative) is the number of non-pollution
samples classified correctly.
According to Pham et al. [64], to measure the quality of the model, a threshold classi-
fication can be applied to classify the area under the ROC curve (AUC) into five classes,
including: values ranging from 0.5 to 0.6, indicating that the prediction of the model is
poor, 0.6 to 0.7, indicating a fair quality of prediction, 0.7 to 0.8, indicating a good quality of
prediction, 0.8 to 0.9, indicating a very good quality of prediction, and 0.9 to 1, indicating
an excellent quality of prediction.

3. Results
3.1. DRASTIC Vulnerability
After multiplying each of the seven DRASTIC index maps by their standard rat-
ings and weights, the overlays of these seven parameters of the DRASTIC index using
Equation (1) were observed to range from 53 to 143.
Remote Sens. 2022, 14, x FOR PEER REVIEWFollowing the classification provided by Aller et al. [16], the generated map was 11 of 21
divided into three classes of groundwater vulnerability, including very low (14%), low
(83%), and medium (3%) (Figure 5).

Figure 5. Vulnerability map using original DRASTIC method.

Figure 5. Vulnerability map using original DRASTIC method.
It can be seen from Figure 5 that the least vulnerable areas are located in the eastern
part Itofcan
the be
study
seenarea.
fromWhereas
Figure 5the south
that and the
the least west parts
vulnerable of the
areas arestudy areainwere
located the eastern
part of the study area. Whereas the south and the west parts of the study the
considered to be without risk, probably due to the high depth (>15) and clay soil, area were
northeast and the center parts of the basin were considered to be the most vulnerable
considered to be without risk, probably due to the high depth (>15) and clay soil, the
areas. Due to the low slope (0–12), low depth (0–4.5 m), and high hydraulic conductivity
northeast and the center parts of the basin were considered to be the most vulnerable
(12–41 m/day), the vadose zone in this area is represented by sandstone and conglomerate.
areas. Due to the low slope (0–12), low depth (0–4.5 m), and high hydraulic conductivity
(12–41 m/day),Ratio
3.2. Frequency the vadose zone in this area is represented by sandstone and conglomerate.
Using the frequency ratio bivariate statistic test, the correlations between groundwater
3.2. Frequency
pollution Ratio
risk and each DRASTIC parameter were calculated. As can be seen from Table 1,
the net recharge
Using the class, ranging from
frequency ratio 50bivariate
to 100 mm,statistic
has the highest FR (0.88).
test, the The classes
correlations between
0–1.5 and 1.5–4.5 of depth to water table and the class 29–41 of hydraulic conductivity
groundwater pollution risk and each DRASTIC parameter were calculated. As can be seen
present the FR value of zero, i.e., no probability of groundwater pollution with nitrate.
from Table 1, the net recharge class, ranging from 50 to 100 mm, has the highest FR (0.88).
The classes 0–1.5 and 1.5–4.5 of depth to water table and the class 29–41 of hydraulic
conductivity present the FR value of zero, i.e., no probability of groundwater pollution
with nitrate.

3.3. Groundwater Pollution Risk Maps

 Remote Sens. 2022, 14, 2379 11 of 19

3.3. Groundwater Pollution Risk Maps

The spatial distribution of groundwater pollution risk maps (GPRMs) for the study
area were produced by the DRASTIC method and the applied machine learning models.
mote Sens. 2022, 14, x FOR PEER REVIEWUsing equal interval reclassification within the ArcGIS environment, the generated maps 12 of
were categorized into five different classes, including very high, high, moderate, low, and
very low (Figures 6 and 7).

100
7 6
Groundwater pollution risk classes (%)

Remote Sens. 2022, 14, x FOR PEER REVIEW 14 12 of 21

90 12
30 30 12
80 Very High
17 15
70
100 15
6
High
7
Groundwater pollution risk classes (%)

14
90 11 Moderate
15 30 13 12 12
60 30 13 15
80 Very High
9 17 15
Low
50 9 15 High
70 15
1211 Moderate
Very Low
10 15 13
40 60 13 15
9 Low
50 9
30 15
Very Low
40 10 12 52 53
20 36 38 41
30
52 53
10 41
20 36 38
0 10
DRASTIC-RF-MLP DRASTIC-RF-SVM DRASTIC-RF DRASTIC-SVM DRASTIC-MLP
0
DRASTIC-RF-MLP Machine DRASTIC-RF
learning DRASTIC-SVM
DRASTIC-RF-SVM Model DRASTIC-MLP

Machine learning Model

Figure 6. Percentages of groundwater pollution risk classes for the applied models in this study
Figure 6. Percentages of groundwater pollution risk classes for the applied models in this study.
Figure 6. Percentages of groundwater pollution risk classes for the applied models in this study.

Figure 7. Cont.
Remote
RemoteSens.
Sens.2022,
2022,14,
14,x2379
FOR PEER REVIEW 13 of1221
of 19

Figure 7. Groundwater pollution risk maps using the ML models: (a) DRASTIC-MLP, (b) DRASTIC-
Figure 7. Groundwater pollution risk maps using the ML models: (a) DRASTIC-MLP, (b) DRASTIC-
SVM, (c) DRASTIC-RF, (d) DRASTIC-RF-SVM and (e) DRASTIC-RF-MLP.
SVM, (c) DRASTIC-RF, (d) DRASTIC-RF-SVM and (e) DRASTIC-RF-MLP.
It can clearly be seen from Figure 7 that the west and the center parts of the basin are
It can clearly be seen from Figure 7 that the west and the center parts of the basin are
at risk of being more contaminated, whereas the eastern part is considered an unpolluted
at risk of being more contaminated, whereas the eastern part is considered an unpolluted
area. For the DRASTIC-MLP model, the pollution risk map (GPRM) elaborated showed
area. For the DRASTIC-MLP model, the pollution risk map (GPRM) elaborated showed
that 53% of the basin was characterized as having very low vulnerability risk. Whereas
that 53% of the basin was characterized as having very low vulnerability risk. Whereas low,
low, medium, high, and very high contributed 15%, 15%, 12%, and 6%, respectively
medium, high, and very high contributed 15%, 15%, 12%, and 6%, respectively (Figure 7a).
(Figure 7a). For the DRASTIC-SVM model, the GPRM elaborated showed that 52% of the
For the DRASTIC-SVM model, the GPRM elaborated showed that 52% of the basin was
basin was characterized as having very low vulnerability risk. Whereas low, medium,
characterized as having very low vulnerability risk. Whereas low, medium, high, and
high, and very high contributed 13%, 15%, 12%, and 7%, respectively (Figure 7b). As can
very high contributed 13%, 15%, 12%, and 7%, respectively (Figure 7b). As can be seen
be seen from Figure 7c, the generated groundwater pollution risk map (GPRM) produced
from Figure 7c, the generated groundwater pollution risk map (GPRM) produced by
by RF model showed that 41% of the study area was classified as very low vulnerability,
RF model showed that 41% of the study area was classified as very low vulnerability,
followed by the class of low vulnerability (15%). Whereas medium, high, and very high
followed by the class of low vulnerability (15%). Whereas medium, high, and very high
contributed 13%, 17%, and 14%, respectively. The GPRM generated by the DRASTIC-FR-
contributed 13%, 17%, and 14%, respectively. The GPRM generated by the DRASTIC-FR-
SVM was divided into five classes: very low vulnerability (38%), low vulnerability (12%),
SVM was divided into five classes: very low vulnerability (38%), low vulnerability (12%),
moderate vulnerability (9%), high vulnerability (11%), and very high vulnerability (30%)
moderate vulnerability (9%), high vulnerability (11%), and very high vulnerability (30%)
(Figure 7d). The GPRM generated by the DRASTIC-FR-MLP was divided into five classes:
(Figure 7d). The GPRM generated by the DRASTIC-FR-MLP was divided into five classes:
very low vulnerability (36%), low vulnerability (10%), moderate vulnerability (9%), high
very low vulnerability (36%), low vulnerability (10%), moderate vulnerability (9%), high
vulnerability (15%), and very high vulnerability (30%) (Figure 7e).
vulnerability (15%), and very high vulnerability (30%) (Figure 7e).
3.4.
3.4.Validation
Validation
Using
Usingthethetraining
trainingand
andthe
thevalidation
validationdatasets,
datasets,thethe
performance
performanceof of
thethe
developed
developed
models
modelsininthis
thisstudy
studywas
wasevaluated.
evaluated.The
TheDRASTIC-RF-MLP
DRASTIC-RF-MLP ensemble
ensembleoutperformed
outperformed all all
other models developed in this study. Thus, based on the training dataset (Table
other models developed in this study. Thus, based on the training dataset (Table 2), the 2), the
Remote Sens. 2022, 14, 2379 13 of 19

following values were obtained: Accuracy = 0.957, Precision = 0.943, Sensitivity = 0.971,
and Specificity = 0.944. Based on the validation dataset, the following statistical values were
obtained: Accuracy = 0.7952, Precision = 0.969, Sensitivity = 0.939, and Specificity = 0.966.

Table 2. Model statistical measures assigned to training and validation datasets.

Models Sample Accuracy Precision Sensitivity Specificity

Training 0.743 0.718 0.800 0.686
DRASTIC-SVM
Validation 0.733 0.706 0.800 0.667
Training 0.786 0.750 0.857 0.733
DRASTIC-MLP
Validation 0.767 0.750 0.800 0.681
Training 0.886 0.865 0.914 0.875
DRASTIC-RF
Validation 0.871 0.857 0.875 0.867
Training 0.914 0.892 0.943 0.886
DRASTIC-RF-SVM
Validation 0.900 0.933 0.875 0.929
Training 0.957 0.943 0.971 0.944
DRASTIC-RF-MLP
Validation 0.952 0.969 0.939 0.966

The DRASTIC-RF-SVM ensemble performed second best, with the training dataset
revealing the following values shown in Table 2: Accuracy = 0.914, Precision = 0.892,
Sensitivity = 0.943, and Specificity = 0.886; whereas the implication of the validation dataset
yields the following statistical values: Accuracy = 0.900, Precision = 0.933, Sensitivity = 0.875,
and Specificity = 0.929.
Based on the training dataset, the performances obtained by the DRASTIC-RF model re-
vealed: Accuracy = 0.886, Precision = 0.865, Sensitivity = 0.914, and Specificity = 0.875 (Table 2).
Thus, for the validation dataset: Accuracy= 0.871, Precision= 0.857, Sensitivity = 0.875, and
Specificity = 0.867.
The performances obtained by the DRASTIC-SVM ensemble were as follows for
the training dataset (Table 2): Accuracy (0.743), Precision (0.718), Sensitivity (0.840), and
Specificity (0.686). Thus, the validation dataset achieved the following statistical values:
Accuracy (0.733), Precision (0.706), Sensitivity (0.800), and Specificity (0.667). The DRASTIC-
MLP model has the lowest performance in comparison to the other developed models.
Thus, for the training dataset, as shown in Table 2, the following statistical values were
obtained: Accuracy (0.786), Precision (0.750), Sensitivity (0.857), and Specificity (0.733). For
the validation dataset, our results show: Accuracy (0.767), Precision (0. 750), Sensitivity
(0.800), and Specificity (0.681).
The AUC value of the original DRASTIC method was equal to 0.53. However, the
applied machine learning models, including DRASTIC-MLP, DRASTIC-SVM, DRASTIC-RF,
DRASTIC-RF-SVM, and DRASTIC-RF-MLP performed better with AUC values of 0.747,
0.804, 0.857, 0.892, and 0.940, respectively, in term of Success Rate (Figure 8a). In addition,
based on the prediction rate (Figure 8b), the AUCs values were as follows: 0.763, 0.803, 0.852,
0.901, and 0.953 for DRASTIC-MLP, DRASTIC-SVM, DRASTIC-RF, DRASTIC-RF-SVM, and
DRASTIC-RF-MLP, respectively. It should be noted that the DRASTIC-RF-MLP ensemble
model had the highest performance. Our results indicate that both in terms of success and
prediction rate, the use of machine learning algorithms improves the performance of the
DRASTIC method.
Remote Sens. 2022, 14, x FOR PEER REVIEW 15 of 21
Remote Sens. 2022, 14, 2379 14 of 19

Figure
Figure 8. (a)
8. (a) ROC ROC curvesofofsuccess
curves success rate;
rate; (b)
(b) ROC
ROCcurves
curvesofof
prediction rate.
prediction rate.
3.5. Variable Importance
3.5. Variable Importance
In all modeling research purposes, variables’ importance should be applied to find
Insuitable
the all modeling research
predictive purposes,
variables variables’research
for the modelling importance should
[65,66]. In thisbe applied
study, to find
the RF
the model
suitable
waspredictive variables
used to evaluate the for the modelling
importance research parameters.
of the DRASTIC [65,66]. In this
Ourstudy, the RF
findings
showed
model wasthat
used both
to the depth ofthe
evaluate groundwater
importance (1.8)
of and
the the hydraulicparameters.
DRASTIC conductivityOur(1.6) findings
had
the highest
showed importance
that both in groundwater
the depth of groundwatervulnerability
(1.8) andassessment, followed
the hydraulic by net recharge
conductivity (1.6) had
(1.49), aquifer media (1.38), topography (1.27), soil (1.21), and
the highest importance in groundwater vulnerability assessment, followed the impact of the vadoseby net
zone (0.97).
recharge (1.49), aquifer media (1.38), topography (1.27), soil (1.21), and the impact of the
vadose zone (0.97).
4. Discussion
Groundwater pollution is one of the challenging issues in the world, especially in arid
4. Discussion
and semi-arid ecosystems [20]. To overcome this issue, researchers have deployed great
efforts to preventpollution
Groundwater and manage water
is one ofcontamination.
the challenging Theissues
DRASTICin themethod
world, is aespecially
widely in
applied approach to assess groundwater pollution.
arid and semi-arid ecosystems [20]. To overcome this issue, researchers have deployed
Our results
great efforts showedand
to prevent thatmanage
the northwater
and central parts tendedThe
contamination. to have a moderate
DRASTIC methodrisk is a
of pollution, whereas the least vulnerable areas are located in the eastern parts of the
widely applied approach to assess groundwater pollution.
study area. The south and west parts of the study area are considered to be without
OurThis
risk. results
couldshowed
be due to that
thethe north andbackground
hydrological central parts tended
of the studytoarea
have a moderate
(i.e., low slope, risk
of pollution,
lowest depth, highest hydraulic conductivity, and the lithological nature) [67]. of
whereas the least vulnerable areas are located in the eastern parts the study
Similar
area.
to The south and
our results, west research
previous parts ofreported
the study area
that are considered
a higher groundwater to vulnerability
be without risk. was This
could be dueinto
observed the hydrological
downslope areas [68],background
meaning theofhigher
the study area (i.e.,
the hydraulic low slope,the
conductivity, lowest
higher
depth, the risk
highest of groundwater
hydraulic [67]. According
conductivity, and theto lithological
Baghapour etnature)
al. [68], [67].
the slope has an
Similar to our
important
results, impact.
previous In addition,
research reported [21]that
indicated that
a higher the lowest depth
groundwater is more affected
vulnerability by
was observed
groundwaterareas
in downslope vulnerability.
[68], meaning the higher the hydraulic conductivity, the higher the
Due to its uncertainty, many researchers have modified the original DRASTIC method
risk of groundwater [67]. According to Baghapour et al. [68], the slope has an important
using statistical techniques and machine learning models to increase its accuracy. In this
impact. In addition, [21] indicated that the lowest depth is more affected by groundwater
study, five new individual models and their ensemble machine learning models were
vulnerability.
developed, namely DRASTIC-RF, DRASTIC-SVM, DRASTIC-MLP, DRASTIC-RF-SVM,
Due
and to its uncertainty,
DRASTIC-RF-MLP, many researchers
for groundwater have
pollution riskmodified
assessmentthe original
in the Saiss DRASTIC
basin,
method using statistical techniques and machine learning models to increase its accuracy.
in Morocco.
Through
In this study, its new
five ability to handle models
individual large datasets, its low
and their aptitudemachine
ensemble to overfitting,
learningand models
its
ability to learn non-linear relationships between the nitrate
were developed, namely DRASTIC-RF, DRASTIC-SVM, DRASTIC-MLP, DRASTIC-RF- polluted samples and the
SVM, and DRASTIC-RF-MLP, for groundwater pollution risk assessment in the Saiss
basin, in Morocco.
Through its ability to handle large datasets, its low aptitude to overfitting, and its
ability to learn non-linear relationships between the nitrate polluted samples and the
Remote Sens. 2022, 14, 2379 15 of 19

input layers [15], random forest has shown greater accuracy in previous studies [26,53]. In
addition, Ref. [44] confirms that RF is one of the successful machine learning models for
groundwater mapping.
Hybrid/ensemble approaches have been proposed recently to improve the performance
of both individual machine learning models and bivariate statistical techniques [24,42,51]. In
the current research, RF was used to improve the accuracy of the other applied models. In
term of the accuracy, our results show that the DRASTIC-RF-MLP hybrid model provided the
best accuracy in comparison to the other models. MLP has been used in several studies with
successful results because of its ability to deal with complex non-linear problems and ease in
processing large numbers of input data [24].
In terms of the DRASTIC-RF-SVM model, the SVM algorithm has significant ad-
vantages in solving linear and non-linear problems and works well in high dimensional
spaces [48]. It has already revealed great potential for groundwater mapping [45,52].
It should be highlighted that selecting the best model is not an easy task because all the
aforementioned models present their own advantages and drawbacks. In addition, we can
add that the outcomes of this research are limited and related to the study area background
and the data used. Thus, it can be concluded that all the above discussed models can be
used for any hazard environment monitoring studies, including groundwater pollution
risk mapping, for other environment backgrounds with promising results.
Our findings showed that both depth of groundwater and hydraulic conductivity
had the highest importance in groundwater vulnerability assessment, followed by net
recharge, aquifer media, topography, soil, and impact of the vadose zone. Similar findings
were also found by [52,69]. From our results, depth of groundwater is the most important
variable; this is because the groundwater could easily be contaminated by surface runoff
and contaminants. These findings agree with Pham et al. [52], who also confirm that
depth of groundwater is the most significant factor for groundwater potential mapping.
Hydraulic conductivity is considered an important factor for groundwater management
strategies because it controls the contaminants’ migration rate from the source to the aquifer.
With the rapid population growth in the Fes-Meknes region, urbanization and exten-
sive use of soils for agriculture activities have become serious challenges for environmental
agency, leading to the groundwater quality deterioration in the Saiss basin. In the same
geographical area, in a recent work El Hafyani et al. [70] reported that the increase in
water consumption is linked to Meknes city’s urban growth and the agricultural activities’
system adopted. In addition, the impact of climate change on natural resources has been
investigated in this study area [71]. Laraichi et al. [72] demonstrated a weakening in the
transmission of information and communication about groundwater, which might lead to
several issues including overexploitation and pollution. For instance, Benaabidate et al. [73]
reported in their study a decrease in piezometric level, which is mainly due to several
factors, including the decrease in precipitation, the reduced natural aquifer recharge, and
the increased pumping, mainly for irrigation. Likewise, a recent work by Berni et al. [9]
highlighted that a negative effect of pesticides in the Saiss basin was explained by farmers’
safety behavior. Additionally, expert interpretations confirmed that a clear link between
groundwater vulnerability risk and geological background of the Saiss basin was shown,
highlighting that Meknes city is more susceptible to groundwater pollution, which is
probably due to the existence of permeable Paleocene sands in this area.
Our results indicate that the most vulnerable areas are located in the northeast and
the center of the basin, because of low depth, low slope, high recharge, and high hydraulic
conductivity, whereas the high depth, low recharge, and low conductivity of the western
areas of the Saiss cause this part to be considered as without risk. These findings are in
line with previous works [9,35]. Thus, the delineation of highly affected areas to pollution
is urgently needed to avoid the deterioration of groundwater resources. In this aquifer,
as agriculture is still going to be one of the main sources of the local population, to meet
the livelihood’s resource requirements, more comprehensive studies should be carried
out to promote the sustainable use of natural resources for the proper management and
Remote Sens. 2022, 14, 2379 16 of 19

planning of them. Therefore, the proposed methodology could be a newer, effective tool
and an emergent path to decision making for assessing groundwater pollution based on
the performance accuracy of hybrid machine learning algorithms.
Marking vulnerable areas to pollution based on the opportunities of the developed
models will be very helpful in encouraging the ongoing efforts to develop a geoportal
platform in the framework of the VLIR-UOS project through a collaborative effort with
scientists and different environmental agencies (https://www.vliruos.be/en/projects/
project/ (accessed on 14 December 2021).Furthermore, this work contributes to the aim of
the Sustainable Development Goals (SDGs) framework on its sixth goal, which is dedicated
to water sustainability and different indicators related to water quality.

5. Conclusions
In this study, five new hybrid models were developed based on the modified DRASTIC
method by the frequency ration bivariate statistic test and the random forest machine
learning algorithm, namely DRASTIC-RF, DRASTIC-SVM, DRASTIC-MLP, DRASTIC-RF-
SVM, and DRASTIC-RF-MLP, for groundwater pollution assessment in the Saiss basin.
Furthermore, three conclusions can be highlighted.
- The results obtained indicate that the most vulnerable areas are located in the west and
the center parts of the basin, because of the low depth, low slope, and high hydraulic
conductivity, whereas the high depth, low recharge, and low conductivity of the
western areas of the Saiss basin mean that this area is considered to be without risk;
- As expected, the locations subject to high vulnerability risk are associated with a high
concentration of nitrate;
- The spatial distribution of groundwater pollution risk maps (GPRMs) for the study
area show that the west and the center parts of the basin are the most vulnerable areas;
- The results highlight that the hybrid/ensemble machine learning (ML) model outper-
forms the individual based model.
It should be noted that the overall goal beyond this research is to implement a ma-
chine learning algorithm to understand groundwater pollution. Furthermore, the output
will help the authorities and government agencies in designing appropriate decision-
making strategies.
Finally, in a vision to protect our environment, the methodology developed here could
be applied in other case studies with similar background.

Author Contributions: Conceptualization, S.I., A.E. and M.M.; methodology S.I., A.E. and M.M.;
software, S.I. and M.M.; validation, S.I., A.E., N.E. and M.M.; formal analysis, S.I. and M.M.; resources,
A.E. and A.V.R.; data curation, S.I., A.E. and M.M.; writing—original draft preparation, S.I. and
M.M.; writing—review and editing, S.I., A.E., A.V.R., E.M.M., M.M. and N.E.; visualization, S.I.,
M.M. and N.E.; supervision, A.E., E.M.M. and A.V.R.; project administration, A.E. and A.V.R.;
funding acquisition, A.E. and A.V.R. All authors have read and agreed to the published version of
the manuscript.
Funding: This work was funded by the VLIR-UOS project of the CUI program partnership between
KU Leuven Belgium and the University of Moulay Ismail Morocco.
Data Availability Statement: Not applicable.
Acknowledgments: This work was supported by the VLIR-UOS project of the CUI program partner-
ship between KU Leuven Belgium and the University of Moulay Ismail Morocco, which the authors
would like to thank. They would also like to acknowledge the Sebou Hydraulic Basin Agency (SHBA)
for providing the necessary data for this research.
Conflicts of Interest: The authors declare no conflict of interest.
Remote Sens. 2022, 14, 2379 17 of 19

References
1. Jesiya, N.P.; Gopinath, G. A Fuzzy Based MCDM–GIS Framework to Evaluate Groundwater Potential Index for Sustainable
Groundwater Management—A Case Study in an Urban-Periurban Ensemble, Southern India. Groundw. Sustain. Dev. 2020,
11, 100466. [CrossRef]
2. Naghibi, S.A.; Pourghasemi, H.R.; Dixon, B. GIS-Based Groundwater Potential Mapping Using Boosted Regression Tree,
Classification and Regression Tree, and Random Forest Machine Learning Models in Iran. Environ. Monit. Assess. 2016, 188, 44.
[CrossRef] [PubMed]
3. Organisation Mondiale de la Santé; UNICEF. Progress on Drinking Water, Sanitation and Hygiene: 2017 Update and SDG Baselines;
World Health Organization: Geneva, Switzerland, 2017; ISBN 978-92-4-151289-3.
4. Omarova, A.; Tussupova, K.; Hjorth, P.; Kalishev, M.; Dosmagambetova, R. Water Supply Challenges in Rural Areas: A Case
Study from Central Kazakhstan. Int. J. Environ. Res. Public. Health 2019, 16, 688. [CrossRef] [PubMed]
5. Kammoun, S.; Trabelsi, R.; Re, V.; Zouari, K.; Henchiri, J. Groundwater Quality Assessment in Semi-Arid Regions Using Integrated
Approaches: The Case of Grombalia Aquifer (NE Tunisia). Environ. Monit. Assess. 2018, 190, 87. [CrossRef] [PubMed]
6. Rahmati, O.; Melesse, A.M. Application of Dempster–Shafer Theory, Spatial Analysis and Remote Sensing for Groundwater
Potentiality and Nitrate Pollution Analysis in the Semi-Arid Region of Khuzestan, Iran. Sci. Total Environ. 2016, 568, 1110–1123.
[CrossRef] [PubMed]
7. Chen, R.; Teng, Y.; Chen, H.; Hu, B.; Yue, W. Groundwater Pollution and Risk Assessment Based on Source Apportionment in a
Typical Cold Agricultural Region in Northeastern China. Sci. Total Environ. 2019, 696, 133972. [CrossRef] [PubMed]
8. Serio, F.; Miglietta, P.P.; Lamastra, L.; Ficocelli, S.; Intini, F.; De Leo, F.; De Donno, A. Groundwater Nitrate Contamination and
Agricultural Land Use: A Grey Water Footprint Perspective in Southern Apulia Region (Italy). Sci. Total Environ. 2018, 645,
1425–1431. [CrossRef]
9. Berni, I.; Menouni, A.; El Ghazi, I.; Godderis, L.; Duca, R.-C.; Jaafari, S.E. Health and Ecological Risk Assessment Based on
Pesticide Monitoring in Saïss Plain (Morocco) Groundwater. Environ. Pollut. 2021, 276, 116638. [CrossRef]
10. Möhring, N.; Dalhaus, T.; Enjolras, G.; Finger, R. Crop Insurance and Pesticide Use in European Agriculture. Agric. Syst. 2020,
184, 102902. [CrossRef]
11. Sanchezperez, J.; Antiguedad, I.; Arrate, I.; Garcialinares, C.; Morell, I. The Influence of Nitrate Leaching through Unsaturated
Soil on Groundwater Pollution in an Agricultural Area of the Basque Country: A Case Study. Sci. Total Environ. 2003, 317, 173–187.
[CrossRef]
12. Biddau, R.; Cidu, R.; Da Pelo, S.; Carletti, A.; Ghiglieri, G.; Pittalis, D. Source and Fate of Nitrate in Contaminated Groundwater
Systems: Assessing Spatial and Temporal Variations by Hydrogeochemistry and Multiple Stable Isotope Tools. Sci. Total Environ.
2019, 647, 1121–1136. [CrossRef] [PubMed]
13. Meng, L.; Zhang, Q.; Liu, P.; He, H.; Xu, W. Influence of Agricultural Irrigation Activity on the Potential Risk of Groundwater
Pollution: A Study with Drastic Method in a Semi-Arid Agricultural Region of China. Sustainability 2020, 12, 1954. [CrossRef]
14. Oliveira, A.; Lopes, A.; Niza, S. Local Climate Zones in Five Southern European Cities: An Improved GIS-Based Classification
Method Based on Copernicus Data. Urban Clim. 2020, 33, 100631. [CrossRef]
15. Rodriguez-Galiano, V.; Mendes, M.P.; Garcia-Soldado, M.J.; Chica-Olmo, M.; Ribeiro, L. Predictive Modeling of Groundwater
Nitrate Pollution Using Random Forest and Multisource Variables Related to Intrinsic and Specific Vulnerability: A Case Study in
an Agricultural Setting (Southern Spain). Sci. Total Environ. 2014, 476–477, 189–206. [CrossRef] [PubMed]
16. Aller, L.; Lehr, J.H.; Petty, R.; Bennett, T. DRASTIC: A Standardized System for Evaluating Ground Water Pollution Potential Using
Hydrogeologic Settings; Robert, S., Ed.; Kerr Environmental Research Laboratory, Office of Research and Development, U.S.
Environmental Protection Agency: Ada, OK, USA, 1987.
17. Arya, S.; Subramani, T.; Vennila, G.; Roy, P.D. Groundwater Vulnerability to Pollution in the Semi-Arid Vattamalaikarai River
Basin of South India Thorough DRASTIC Index Evaluation. Geochemistry 2020, 80, 125635. [CrossRef]
18. Sinan, M.; Razack, M. An Extension to the DRASTIC Model to Assess Groundwater Vulnerability to Pollution: Application to the
Haouz Aquifer of Marrakech (Morocco). Environ. Geol. 2009, 57, 349–363. [CrossRef]
19. Arshad, A.; Zhang, Z.; Zhang, W.; Dilawar, A. Mapping Favorable Groundwater Potential Recharge Zones Using a GIS-Based
Analytical Hierarchical Process and Probability Frequency Ratio Model: A Case Study from an Agro-Urban Region of Pakistan.
Geosci. Front. 2020, 11, 1805–1819. [CrossRef]
20. Nadiri, A.A.; Sedghi, Z.; Khatibi, R.; Gharekhani, M. Mapping Vulnerability of Multiple Aquifers Using Multiple Models and
Fuzzy Logic to Objectively Derive Model Structures. Sci. Total Environ. 2017, 593–594, 75–90. [CrossRef]
21. Khosravi, K.; Sartaj, M.; Tsai, F.T.-C.; Singh, V.P.; Kazakis, N.; Melesse, A.M.; Prakash, I.; Tien Bui, D.; Pham, B.T. A Comparison
Study of DRASTIC Methods with Various Objective Methods for Groundwater Vulnerability Assessment. Sci. Total Environ. 2018,
642, 1032–1049. [CrossRef]
22. Sarkar, T.; Mishra, M. Soil Erosion Susceptibility Mapping with the Application of Logistic Regression and Artificial Neural
Network. J. Geovisualization Spat. Anal. 2018, 2, 8. [CrossRef]
23. Neshat, A.; Pradhan, B. An Integrated DRASTIC Model Using Frequency Ratio and Two New Hybrid Methods for Groundwater
Vulnerability Assessment. Nat. Hazards 2015, 76, 543–563. [CrossRef]
Remote Sens. 2022, 14, 2379 18 of 19

24. Fijani, E.; Nadiri, A.A.; Asghari Moghaddam, A.; Tsai, F.T.-C.; Dixon, B. Optimization of DRASTIC Method by Supervised
Committee Machine Artificial Intelligence to Assess Groundwater Vulnerability for Maragheh–Bonab Plain Aquifer, Iran. J. Hydrol.
2013, 503, 89–100. [CrossRef]
25. Asfaw, D.; Mengistu, D. Modeling Megech Watershed Aquifer Vulnerability to Pollution Using Modified DRASTIC Model for
Sustainable Groundwater Management, Northwestern Ethiopia. Groundw. Sustain. Dev. 2020, 11, 100375. [CrossRef]
26. Hosseini, F.S.; Choubin, B.; Mosavi, A.; Nabipour, N.; Shamshirband, S.; Darabi, H.; Haghighi, A.T. Flash-Flood Hazard
Assessment Using Ensembles and Bayesian-Based Machine Learning Models: Application of the Simulated Annealing Feature
Selection Method. Sci. Total Environ. 2020, 711, 135161. [CrossRef] [PubMed]
27. Wang, Y.; Feng, L.; Li, S.; Ren, F.; Du, Q. A Hybrid Model Considering Spatial Heterogeneity for Landslide Susceptibility Mapping
in Zhejiang Province, China. CATENA 2020, 188, 104425. [CrossRef]
28. Chapi, K.; Singh, V.P.; Shirzadi, A.; Shahabi, H.; Bui, D.T.; Pham, B.T.; Khosravi, K. A Novel Hybrid Artificial Intelligence
Approach for Flood Susceptibility Assessment. Environ. Model. Softw. 2017, 95, 229–245. [CrossRef]
29. Costache, R. Flash-Flood Potential Assessment in the Upper and Middle Sector of Prahova River Catchment (Romania). A
Comparative Approach between Four Hybrid Models. Sci. Total Environ. 2019, 659, 1115–1134. [CrossRef]
30. Pham, B.T.; Prakash, I.; Singh, S.K.; Shirzadi, A.; Shahabi, H.; Tran, T.-T.-T.; Bui, D.T. Landslide Susceptibility Modeling Using
Reduced Error Pruning Trees and Different Ensemble Techniques: Hybrid Machine Learning Approaches. CATENA 2019, 175,
203–218. [CrossRef]
31. Pham, B.T.; Tien Bui, D.; Prakash, I.; Dholakia, M.B. Hybrid Integration of Multilayer Perceptron Neural Networks and Machine
Learning Ensembles for Landslide Susceptibility Assessment at Himalayan Area (India) Using GIS. CATENA 2017, 149, 52–63.
[CrossRef]
32. Tien Bui, D.; Bui, Q.-T.; Nguyen, Q.-P.; Pradhan, B.; Nampak, H.; Trinh, P.T. A Hybrid Artificial Intelligence Approach Using
GIS-Based Neural-Fuzzy Inference System and Particle Swarm Optimization for Forest Fire Susceptibility Modeling at a Tropical
Area. Agric. For. Meteorol. 2017, 233, 32–44. [CrossRef]
33. Yadav, B.; Gupta, P.K.; Patidar, N.; Himanshu, S.K. Ensemble Modelling Framework for Groundwater Level Prediction in Urban
Areas of India. Sci. Total Environ. 2020, 712, 135539. [CrossRef] [PubMed]
34. Sadkaoui, N.; Boukrim, S.; Bourak, A.; Lakhili, F.; Mesrar, L.; Chaouni, A.-A.; Lahrach, A.; Jabrane, R.; Akdim, B. Groundwater
pollution of SAÏS basin (Morocco), vulnerability mapping by drastic, god and PRK methods, involving geographic information
system (GIS). Present Environ. Sustain. Dev. 2013, 7, 298–308.
35. Lahjouj, A.; El Hmaidi, A.; Bouhafa, K.; Boufala, M. Mapping Specific Groundwater Vulnerability to Nitrate Using Random
Forest: Case of Sais Basin, Morocco. Model. Earth Syst. Environ. 2020, 6, 1451–1466. [CrossRef]
36. Margat, J. Hydrogeological Map of the Meknes-Fes Basin; Edition of the Office of Irrigation: Rabat, Morocco, 1960.
37. Essahlaoui, A.; Sahbi, H.; Bahi, L.; El-Yamine, N. Reconnaissance de la structure géologique du bassin de saïss occidental, Maroc,
par sondages électriques. J. Afr. Earth Sci. 2001, 32, 777–789. [CrossRef]
38. Scanlon, B.R.; Healy, R.W.; Cook, P.G. Choosing Appropriate Techniques for Quantifying Groundwater Recharge. Hydrogeol. J.
2002, 10, 18–39. [CrossRef]
39. Khosravi, K.; Sartaj, M.; Karimi, M.; Levison, J.; Lotfi, A. A GIS-Based Groundwater Pollution Potential Using DRASTIC, Modified
DRASTIC, and Bivariate Statistical Models. Environ. Sci. Pollut. Res. 2021, 28, 50525–50541. [CrossRef]
40. Tehrany, M.S.; Jones, S.; Shabani, F.; Martínez-Álvarez, F.; Tien Bui, D. A Novel Ensemble Modeling Approach for the Spatial
Prediction of Tropical Forest Fire Susceptibility Using LogitBoost Machine Learning Classifier and Multi-Source Geospatial Data.
Theor. Appl. Climatol. 2019, 137, 637–653. [CrossRef]
41. Pradhan, B.; Lee, S. Landslide Risk Analysis Using Artificial Neural Network Model Focussing on Different Training Sites. Int. J.
Phys. Sci. 2009, 4, 1–15.
42. Guyon, I.; Weston, J.; Barnhill, S. Gene selection for cancer classification using support vector machines. Mach. Learn. 2002, 46,
389–422. [CrossRef]
43. Boser, B.E.; Guyon, I.M.; Vapnik, V.N. A Training Algorithm for Optimal Margin Classifiers | Proceedings of the Fifth Annual
Workshop on Computational Learning Theory. Available online: https://dl.acm.org/doi/abs/10.1145/130385.130401 (accessed
on 6 May 2022).
44. Naghibi, S.A.; Ahmadi, K.; Daneshi, A. Application of Support Vector Machine, Random Forest, and Genetic Algorithm Optimized
Random Forest Models in Groundwater Potential Mapping. Water Resour. Manag. 2017, 31, 2761–2775. [CrossRef]
45. Yousefi, S.; Sadhasivam, N.; Pourghasemi, H.R.; Ghaffari Nazarlou, H.; Golkar, F.; Tavangar, S.; Santosh, M. Groundwater Spring
Potential Assessment Using New Ensemble Data Mining Techniques. Measurement 2020, 157, 107652. [CrossRef]
46. Han, H.; Shi, B.; Zhang, L. Prediction of Landslide Sharp Increase Displacement by SVM with Considering Hysteresis of
Groundwater Change. Eng. Geol. 2021, 280, 105876. [CrossRef]
47. Costache, R.; Hong, H.; Pham, Q.B. Comparative Assessment of the Flash-Flood Potential within Small Mountain Catchments
Using Bivariate Statistics and Their Novel Hybrid Integration with Machine Learning Models. Sci. Total Environ. 2020, 711, 134514.
[CrossRef] [PubMed]
48. Tehrany, M.S.; Pradhan, B.; Jebur, M.N. Flood Susceptibility Analysis and Its Verification Using a Novel Ensemble Support Vector
Machine and Frequency Ratio Method. Stoch. Environ. Res. Risk Assess. 2015, 29, 1149–1165. [CrossRef]
Remote Sens. 2022, 14, 2379 19 of 19

49. Kavzoglu, T.; Colkesen, I. A Kernel Functions Analysis for Support Vector Machines for Land Cover Classification. Int. J. Appl.
Earth Obs. Geoinf. 2009, 11, 352–359. [CrossRef]
50. Pourghasemi, H.R.; Jirandeh, A.G.; Pradhan, B.; Xu, C.; Gokceoglu, C. Landslide Susceptibility Mapping Using Support Vector
Machine and GIS at the Golestan Province, Iran. J. Earth Syst. Sci. 2013, 122, 349–369. [CrossRef]
51. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. Available online: https://link.springer.com/article/10.1023/A:
1010933404324 (accessed on 17 April 2021). [CrossRef]
52. Rahmati, O.; Choubin, B.; Fathabadi, A.; Coulon, F.; Soltani, E.; Shahabi, H.; Mollaefar, E.; Tiefenbacher, J.; Cipullo, S.;
Ahmad, B.B.; et al. Predicting Uncertainty of Machine Learning Models for Modelling Nitrate Pollution of Groundwater Using
Quantile Regression and UNEEC Methods. Sci. Total Environ. 2019, 688, 855–866. [CrossRef]
53. Mohajane, M.; Costache, R.; Karimi, F.; Bao Pham, Q.; Essahlaoui, A.; Nguyen, H.; Laneve, G.; Oudija, F. Application of Remote
Sensing and Machine Learning Algorithms for Forest Fire Mapping in a Mediterranean Area. Ecol. Indic. 2021, 129, 107869.
[CrossRef]
54. Jiang, T.; Gradus, J.L.; Lash, T.L.; Fox, M.P. Addressing Measurement Error in Random Forests Using Quantitative Bias Analysis.
Am. J. Epidemiol. 2021, 190, 1830–1840. [CrossRef]
55. Chen, W.; Li, Y.; Xue, W.; Shahabi, H.; Li, S.; Hong, H.; Wang, X.; Bian, H.; Zhang, S.; Pradhan, B.; et al. Modeling Flood
Susceptibility Using Data-Driven Approaches of Naïve Bayes Tree, Alternating Decision Tree, and Random Forest Methods.
Sci. Total Environ. 2020, 701, 134979. [CrossRef] [PubMed]
56. Schoppa, L.; Disse, M.; Bachmair, S. Evaluating the Performance of Random Forest for Large-Scale Flood Discharge Simulation.
J. Hydrol. 2020, 590, 125531. [CrossRef]
57. Kavzoglu, T.; Mather, P.M. The Use of Backpropagating Artificial Neural Networks in Land Cover Classification. Int. J. Remote
Sens. 2003, 24, 4907–4938. [CrossRef]
58. Rosenblatt, F. The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain. Psychol. Rev. 1958, 65,
386–408. [CrossRef] [PubMed]
59. Basheer, I.A.; Hajmeer, M. Artificial Neural Networks: Fundamentals, Computing, Design, and Application. J. Microbiol. Methods
2000, 43, 3–31. [CrossRef]
60. Fausett, L. Fundamentals Of Neural Networks: Architectures, Algorithms, and Applications; Prenctice-Hall: Hoboken, NJ, USA, 1994.
61. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. arXiv 2017, arXiv:1412.6980.
62. Yen, H.P.H.; Pham, B.T.; Phong, T.V.; Ha, D.H.; Costache, R.; Le, H.V.; Nguyen, H.D.; Amiri, M.; Tao, N.V.; Prakash, I. Locally
Weighted Learning Based Hybrid Intelligence Models for Groundwater Potential Mapping and Modeling: A Case Study at Gia
Lai Province, Vietnam. Geosci. Front. 2021, 12, 101154. [CrossRef]
63. Costache, R.; Bui, D.T. Spatial prediction of flood potential using new ensembles of bivariate statistics and artificial intelligence: A
case study at the Putna river catchment of Romania. Sci. Total Environ. 2019, 691, 1098–1118. [CrossRef]
64. Pham, B.T.; Jaafari, A.; Prakash, I.; Singh, S.K.; Quoc, N.K.; Bui, D.T. Hybrid Computational Intelligence Models for Groundwater
Potential Mapping. CATENA 2019, 182, 104101. [CrossRef]
65. Costache, R.; Popa, M.C.; Tien Bui, D.; Diaconu, D.C.; Ciubotaru, N.; Minea, G.; Pham, Q.B. Spatial Predicting of Flood Potential
Areas Using Novel Hybridizations of Fuzzy Decision-Making, Bivariate Statistics, and Machine Learning. J. Hydrol. 2020,
585, 124808. [CrossRef]
66. Hong, H.; Pourghasemi, H.R.; Pourtaghi, Z.S. Landslide Susceptibility Assessment in Lianhua County (China): A Comparison
between a Random Forest Data Mining Technique and Bivariate and Multivariate Statistical Models. Geomorphology 2016, 259,
105–118. [CrossRef]
67. Bera, A.; Mukhopadhyay, B.P.; Chowdhury, P.; Ghosh, A.; Biswas, S. Groundwater Vulnerability Assessment Using GIS-Based
DRASTIC Model in Nangasai River Basin, India with Special Emphasis on Agricultural Contamination. Ecotoxicol. Environ. Saf.
2021, 214, 112085. [CrossRef] [PubMed]
68. Baghapour, M.A.; Fadaei Nobandegani, A.; Talebbeydokhti, N.; Bagherzadeh, S.; Nadiri, A.A.; Gharekhani, M.; Chitsazan, N.
Optimization of DRASTIC Method by Artificial Neural Network, Nitrate Vulnerability Index, and Composite DRASTIC Models
to Assess Groundwater Vulnerability for Unconfined Aquifer of Shiraz Plain, Iran. J. Environ. Health Sci. Eng. 2016, 14, 13.
[CrossRef]
69. Knoll, L.; Breuer, L.; Bach, M. Large Scale Prediction of Groundwater Nitrate Concentrations from Spatial Data Using Machine
Learning. Sci. Total Environ. 2019, 668, 1317–1327. [CrossRef] [PubMed]
70. El Hafyani, M.; Essahlaoui, A.; Van Rompaey, A.; Mohajane, M.; El Hmaidi, A.; El Ouali, A.; Moudden, F.; Serrhini, N.-E.
Assessing Regional Scale Water Balances through Remote Sensing Techniques: A Case Study of Boufakrane River Watershed,
Meknes Region, Morocco. Water 2020, 12, 320. [CrossRef]
71. Brouziyne, Y.; Abouabdillah, A.; Bouabid, R.; Benaabidate, L. SWAT Streamflow Modeling for Hydrological Components’
Understanding within an Agro—Sylvo—Pastoral Watershed in Morocco. J. Mater. Environ. Sci. 2018, 9, 128–138. [CrossRef]
72. Laraichi, S.; Hammani, A. How Can Information and Communication Effects on Small Farmers’ Engagement in Groundwater
Management: Case of SAISS Aquifers, Morocco. Groundw. Sustain. Dev. 2018, 7, 109–120. [CrossRef]
73. Benaabidate, L.; Cholli, M. Groundwater stress and vulnerability to pollution of SAISS basin shallow aquifer, Morocco.
In Proceedings of the Fifteenth International Water Technology Conference, Alexandria, Egypt, 28–30 May 2011.