Monitoring and prediction of high fluoride concentrations in groundwater in Pakistan

Abstract

Concentrations of naturally occurring fluoride in groundwater exceeding the WHO guideline of 1.5 mg/L have been detected in many parts of Pakistan. This may lead to dental or skeletal fluorosis and thereby poses a potential threat to public health. Utilizing a total of 5483 fluoride concentrations, comprising 2160 of new measurements as well as those from other sources, we have applied machine learning techniques to predict the probability of fluoride in groundwater in Pakistan exceeding 1.5 mg/L at a 250 m spatial resolution. Climate, soil, lithology, topography, and land cover parameters were identified as effective predictors of high fluoride concentrations in groundwater. Excellent model performance was observed in a random forest model that achieved an Area Under the Curve (AUC) of 0.92 on test data that were not used in modeling. The highest probabilities of high fluoride concentrations in groundwater are predicted in the Thar Desert, Sargodha Division, and scattered along the Sulaiman Mountains. Applying the model predictions to the population density and accounting for groundwater usage in both rural and urban areas, we estimate that about 13 million people may be at risk of fluorosis due to consuming groundwater with fluoride concentrations >1.5 mg/L in Pakistan, which corresponds to ~6% of the total population. Both the fluoride prediction map and the health risk map can be used as important decision-making tools for authorities and water resource managers in the identification and mitigation of groundwater fluoride contamination.

Full text

Monitoring and prediction of high fluoride concentrations in groundwater
in Pakistan
Yuya Ling
a
,Joel Podgorski
a,
⁎,Muhammad Sadiq
b
, Hifza Rasheed
c
,
Syed Ali Musstjab Akber Shah Eqani
b
, Michael Berg
a
a
Eawag, Swiss Federal Institute of Aquatic Science and Technology, Department Water Resources and Drinking Water, 8600 Dübendorf, Switzerland
b
Public Health and Environment Division, Department of Biosciences, COMSATS University, Islamabad, Pakistan
c
National Water Quality Laboratory, Pakistan Council of Research in Water Resources (PCRWR), Islamabad, Pakistan
HIGHLIGHTS
•Groundwater fluoride risk maps (>1.5
mg/L) created with >5000 data and
machine learning for all of Pakistan.
•Arid climate and soil composition are
statistically important predictors of
geogenic fluoride contamination.
•The high-resolution maps reveal the
vulnerable areas and the number of peo-
ple exposed.
•An estimated13 million people (6% of the
population) are at risk of fluorosis.
•Most affected areas are in the Thar Desert,
the Thal Desert, and scattered along the
Sulaiman Mountain Range.
GRAPHICAL ABSTRACT
ABSTRACTARTICLE INFO
Editor: Daniel Alessi Concentrations of naturally occurring fluoride in groundwater exceeding the WHO guideline of 1.5 mg/L have been
detected in many parts of Pakistan. This may lead to dental or skeletal fluorosis and thereby poses a potential threat
to public health. Utilizing a total of 5483 fluoride concentrations, comprising 2160 new measurements as well as
those from other sources, we have applied machine learning techniques to predict the probability of fluoride in
groundwater in Pakistan exceeding 1.5 mg/L at a 250 m spatial resolution. Climate, soil, lithology, topography, and
land cover parameters wereidentified as effectivepredictors of high fluorideconcentrations ingroundwater. Excellent
model performance was observed in a random forest model that achieved an Area Under the Curve (AUC) of 0.92 on
test data that were not used in modeling. The highest probabilities of high fluoride concentrations in groundwater are
predictedin the Thar Desert, Sargodha Division, and scattered alongthe Sulaiman Mountains. Applying the modelpre-
dictions to the population density and accounting for groundwater usage in both rural and urban areas, we estimate
that about 13 million people may be at risk of fluorosis due to consuming groundwater with fluoride concentrations
>1.5 mg/L in Pakistan, which corresponds to ~6% of the total population. Both the fluoride prediction map and the
health risk map can be used as important decision-making tools for authorities and water resource managers in the
identification and mitigation of groundwater fluoride contamination.
Keywords:
Aquifers
Geogenic groundwater pollution
Drinking water quality
Human health threat
Fluorosis
Random forest modeling
Science of the Total Environment 839 (2022) 156058
⁎Corresponding author.
E-mail address: joel.podgorski@eawag.ch (J. Podgorski).
http://dx.doi.org/10.1016/j.scitotenv.2022.156058
Received 9 March 2022; Received in revised form 14 May 2022; Accepted 15 May 2022
Available online 20 May 2022
0048-9697/© 2022 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Contents lists available at ScienceDirect
Science of the Total Environment
journal homepage: www.elsevier.com/locate/scitotenv

1. Introduction
Fluorine is greatly abundant in Earth's crust (Amini et al., 2008;
Edmunds and Smedley, 2013). It is usually present in the ionic form in
the natural environment as a consequence of having the highest electroneg-
ativity among the chemical elements. Fluorine is prevalent in a wide variety
of minerals but mainly in fluorite, fluorapatite, topaz, and to a moderate
degree in biotite and muscovite (Farooqi, 2015;Kumar et al., 2020). Fluo-
rites are commonly found as cements in carbonate rocks (García and
Borgnino, 2015). Aside from natural sources, anthropogenic activities can
introduce fluoride to the environment, for example, through fluoride phos-
phate fertilizer effluents (Ali et al., 2019) or fossil fuel combustion (García
and Borgnino, 2015). These fluorine-bearing minerals gradually become
enriched in groundwater primarily through dissolution and desorption
from metal oxides (Kumar et al., 2020), the processes of which are
promotedby high alkalinity, low calcium concentration, and the sodium bi-
carbonate water type in groundwater (Banerjee, 2015;Farooqi, 2015;
Kumar et al., 2016;Kumar et al., 2020;Rafique et al., 2009;Singh and
Mukherjee, 2015). Elevated fluoride concentrations are also tied to arid cli-
matic conditions (Handa, 1975;Rasool et al., 2018), which are associated
with processes such as a higher cation exchange rate, faster dissolution
from fluoride-containing minerals and prolonged groundwater residence
times and therefore water-rock interactions (Ali et al., 2019;Podgorski
et al., 2018).
While lower concentrations of fluoride in drinking water (0.5–1.0
mg/L) are known to prevent tooth decay (World Health Organization,
1994), excessive exposure to fluoride (above 1.5 mg/L) can lead to dental
and/or skeletal fluorosis (Wang et al., 2004). It has been estimated that
more than 200 million people in the world are at risk of fluorosis (Ayoob
and Gupta, 2006). Given that drinking water is one of the major sources
of fluoride for humans, the World Health Organization (WHO) maintains
a health-based guideline of 1.5 mg/L for fluoride in drinking water
(World Health Organization, 2011), which is also the permissible limit in
Pakistan (Khwaja and Aslam, 2018). However, some countries such as
India (Podgorski et al., 2018;Shah and Bandekar, 1998) and China (Bo
et al., 2003) have adopted the fluoride concentration limit of 1.0 mg/L to
account for arid climates and other fluoride intake pathways, e.g. food
(Ozsvath, 2009).
Groundwater is used extensively to serve the growing population of
Pakistan, supplying around 39% of drinking water (World Health
Organization and UNICEF, 2019) and 73% of irrigation water (Qureshi,
2020). Having a predominantly semi-arid to arid climate, the conditions
in Pakistan are favorable for fluoride accumulation in groundwater. Fur-
thermore, fluoride-bearing rocks such as granites are present in many
parts of the country (Naseem et al., 2010), and many cities contain high
levels of Na
+
and K
+
in groundwater (Raza et al., 2017), which promote
calcite precipitation by cation exchange and reinforce fluoride release
from minerals (Farooqi, 2015).
Elevated fluoride concentrations in groundwater (>1.5 mg/L) have
been identified in many places in Pakistan, though mainly confined to the
populous and flat-lying Punjab and Sindh Provinces (Ali et al., 2019;
Brahman et al., 2013;Farooqi, 2015;Farooqi et al., 2007;Iwasaki, 2007;
Khattak et al., 2022;Rafique, 2008;Rafique et al., 2008;Rafique et al.,
2009). For instance, 27.2% of the samples in a study along the riverine
systems in the Punjab and Sindh exceeded 1.5 mg/L (Ali et al., 2019). The
aquifers of these two provinces are recharged primarily from rainwater
along with infiltration from the five major rivers in Punjab: the Sutlej,
Ravi, Jhelum, Chenab, and Indus Rivers, all of which come together as the
Indus that flows further south through Sindh. Very high fluoride concentra-
tions up to 30 mg/L have been measured in the Thar Desert (Iwasaki, 2007;
Rafique et al., 2009), which is situated at the southeastern corner of the
country where sand dunes and kaolin/granite abound. Widespread concerns
have been raised about dental and skeletal fluorosis that has been detected,
particularly among children, in the cities of Mianwali, Quetta, Lahore, Kara-
chi, Peshawar, and the Thar Desert (Ahmad et al., 2020;Khan et al., 2004;
Khan et al., 2015;Rafique et al., 2015;Sami et al., 2016).
To protect public health, it is essential to determine if wells and springs
contain safe or hazardous levels of fluoride. As such, maps of affected areas
can provide a key first step in determining the locations of safe and hazard-
ous wells. The most comprehensive nationwide analysis previously
conducted of the distribution of fluoride in groundwater in Pakistan
consisted of some 1000 measurements that were summarized at the sub-
tehsil scale (Khan et al., 2002). This corresponds to an average size of
about 2100 km
2
per administrative unit, which provides only very limited
spatial resolution of fluoride contamination. Furthermore, no attempt had
been made to make predictions beyond the data collected.
Thanks to the availability of high-resolution data sets of environmental
predictor variables, machine learning approaches have been used to create
accurate geospatial prediction models of various groundwater as well as
soil parameters (Chen et al., 2017;Erickson et al., 2021;Hengl et al.,
2017;Podgorski and Berg, 2020;Podgorski et al., 2020;Reichstein et al.,
2019;Winkel et al., 2008;Wu et al., 2021). In contrast to conventional
geostatistical techniques based upon interpolation among observations,
machine learning models have the potential to be applied with high
accuracy to larger scales (regional to country scale) where sufficient data
are available.
Machine-learning methods, such as random forest, have proved effec-
tive in modeling a binary target variable (e.g. fluoride concentrations
above a threshold) that is effectively unchanging in time and predicting
its occurrence on the basis of relevant predictors (e.g., climate, geology)
(Podgorski et al., 2017;Podgorski et al., 2018;Winkel et al., 2008). Since
machine learning models learn from data, relationships can be inferred
between predictors and the target variable from modeling results and
thus learn more about how hydro-geochemical conditions regulate fluoride
concentrations in groundwater.
In this paper, we analyzed fluoride in over 2100 groundwater samples
from all over Pakistan, including in the as yet poorly examined Thal Desert,
and combine these with previously published data to produce with
machine learning a first-ever high-resolution prediction map of high fluo-
ride concentrations in Pakistan. This allows us to generate a health-risk
map, estimating the locations and number of people at risk of excessive
fluoride exposure. Furthermore, the statistically important predictor vari-
ables are also discussed for their insight into the environmental conditions
associated with high fluoride concentrations. The main purpose of these
hazard and risk maps is to identify areas with greater and lesser chances
of containing high fluoride concentrations in groundwater. They can
thereby provide valuable guidance for authorities and water resource
managersin testing for and ultimately mitigating hazardously high concen-
trations of fluoride for the protection of health.
2. Materials and methods
2.1. Study area
Pakistan is characterized by a variable geomorphology, with the flat-
lying Indus plain that comprises the provinces of Punjab and Sindh in the
east; the Hindu Kush, Karakoram, and Himalaya ranges in the north; and
the vast BaluchistanPlateau in the west.The climatic conditions areprimar-
ily arid to semi-arid, with temperate conditions in the northwest and arctic
conditions in the northern mountain ranges. Desert areas cover around
one-third of the country, with mountain areas, grasslands and agricultural
regions covering the otherparts (Greenman et al., 1967). Pakistan's geology
is dominated by young (Quaternary age) alluvial and deltaic deposits
(Sanaullah et al., 2019) that outcrop across much of the Indus plain (Pun-
jab, Sindh) and Baluchistan basin, while older formations (granites, meta-
morphic rocks) are mainly restricted to the Khyber Pakhtunkhwa region
(WAPDA/EUAD, 1989). Most of Pakistan's population is concentrated in
Punjab and Sindh due to fertile soil and an abundant water supply. Over
100 million inhabitants of Punjab and Sindh rely on groundwater
replenished by the Indus River and its tributaries (Jhelum, Chenab, Ravi,
and Sutlej) for drinking water and agricultural uses (Bhowmik et al.,
2015). There are more than one million private tube wells in the country,
Y. Ling et al. Science of the Total Environment 839 (2022) 156058
2

of which 3.8% are in Khyber-Pakhtunkhwa, 6.4% in Sindh, 4.8% in Balu-
chistan and more than 80% inPunjab, with a total groundwater abstraction
of about 60 billion m
3
per year (Qureshi, 2020). Unregulated groundwater
abstraction, unsustainable pumping, agricultural irrigation and increasing
water demands in urban areas result in fluctuating groundwater levels
and generally impact the water quality of the aquifers of the Indus plain
(Rasheed et al., 2022;Ullah et al., 2022).
2.2. Groundwater samples
We sampled 2160 groundwater wells between 2013 and 2019 through-
out Pakistan (Fig. S2) and measured fluoride with a portable photometer
field test kit (FTK; HI96739, Hanna Instruments) and verified values by
ion chromatography (IC). The majority of measurements was taken in
well-populated Punjab and Sindh Provinces, especially in the Thal Desert
where a comprehensive study of fluoride contamination has been lacking.
To test the accuracy of the FTK, water samples with known fluoride con-
centrations ranging from 0 to 8 mg/L were preparedand subsequentlymea-
sured by IC and the FTK. This comparison showed that the FTK is generally
able to recover the fluoride concentrations with high accuracy (Table S3),
in particular in the range of 0 to 2 mg/L. In addition, 54 of the field samples
were also verified by IC. These mostly plot along the 1:1 line (Fig.S3), with
deviations generally being due to higher FTK measurements. With respect
to the threshold of 1.5 mg/L, 83% of these samples were classified the
same by both measurement methods.
The original 2160 fluoride measurements were combined with data
from other sources to form a dataset of 5543 geolocated concentrations of
fluoride in groundwater throughout Pakistan (see Fig. 1, Fig. S5, and
Table S1), which were later used in modeling. Of these, 2814 samples
were collected between October 2013 and March 2014 by the Rural
Water Quality Monitoring Program (RWQMP) of the Pakistan Council of
Research in Water Resources (PCRWR, 2015). Approximately 573 of the
2865 villages in Punjab and Sindh Provinces were covered in this program,
with 4 to 5 samples collected in each village. In addition, we were able to
include a few hundred data points from the Indus plain in Punjab and
Sindh (Ali et al., 2019) as well as the Thar Desert (Rafique, 2008;Rafique
et al., 2008;Rafique et al., 2009), where 79% of the samples exceeded
1.5 mg/L.
Overall, 911 of the 5543 groundwater samples (16%) are greater than
1.5 mg/L, and 1451 samples (26%) are above 1 mg/L. The clusters of
samples reported at the same location in a village were averaged after
removing outliers, the points of which are distinctly different from their
neighboring data (e.g. in the RWQMP project). This reduced the number
of the total data points from 5543 to 5483, with 16% of points having a
high fluoride level after outlier removal and averaging. The fluoride con-
centrations were then converted into binary levels “high”(F >1.5 mg/L)
and “low”(F ≤1.5 mg/L).
2.3. Predictor data sets
A variety of environmental data sets dealing with climate, lithology,
land cover and soil (n= 48, see Fig. S11, and Table S4) were considered
for modeling based on known or possible links with fluoride contamination
(García and Borgnino, 2015;Handa, 1975;Podgorski et al., 2017;
Podgorski et al., 2018). All of the variables represent data at the surface,
or up to 2 m depth in the case of soil parameters (Hengl et al., 2017).
Each level of the categorical predictors of soil groups, geologic age, and
lithology was converted into a binary data set to enable testing the signifi-
cance of each category in relation to fluoride in subsequent variable selec-
tion procedures. Pearson correlations were determined between each
continuous predictor variable and fluoride concentration, whereas the
fraction of fluoride concentrations exceeding 1.5 mg/L was calculated for
categorical variables.
Predictor values were extracted at the locations of the 5483 fluoride
measurements and added to the fluoride data set. As most predictors
were available in 250 m resolution, rasters with a coarser (1000 m) or
finer (100 m) resolution were resampled by nearest neighbor and bilinear
methods respectively, to 250 m for subsequent prediction procedures,
allowing the resolution of the prediction maps to be 250 m.
2.4. Training and testing data sets
Training and test data sets were split at the ratio of 80% to 20%. Tightly
clustered data points selected by random sampling may result in data rows
that do not contain much variance in the predictor variables. Therefore, a
method was developed that ensuresgreater predictor variability as follows:
i). Selection of continuous predictors with the highest correlations with
fluoride (Table 1, Table S4): coarse fragments fraction, nitrogen
fraction, organic carbon density, potential evapotranspiration (PET),
aridity index, compound topographic index, and slope.
ii). The quantile interval (0–25%, 25–50%, 50–75%, 75–100%) of each
predictor in step i) was determined for each of the 5483 fluoride
measurements.
iii). The data points were then divided into groups that share the same
predictor quantile interval combinations (1714 unique groups).
iv). Training dataset (80% of data) created by random sampling, ensuring
that at leastone data point is selected from each group. The remaining
data points are assigned to the test dataset (20% of data).
Fig. 1. Groundwater fluoride concentrations (n= 5543) from original and existing
measurements. a) High (F >1.5) and low (F ≤1.5) fluoride levels are plotted with
topography.About 16% of the wells exceed the WHO drinking-water guideline of
1.5 mg/L. b) Population density map of Pakistan with labels of provinces and rivers.
Y. Ling et al. Science of the Total Environment 839 (2022) 156058
3

Applying this procedure, the proportion of “high”to “low”fluoride clas-
ses in the training and test sets remains approximately 16% (Table S5).
2.5. Random forest modeling
Random forest is a machine-learning algorithm that constructs an
ensemble of decision trees, which recursively partition predictor variables
to predict a dependent variable (Breiman, 2001). Individual decision
trees consider a random subset of candidate predictors at each split,
which reduces correlation between decision trees and helps avoid
overfitting (James et al., 2013). Increasing the number of trees can further
reduce overfitting (Breiman, 2001). Randomness is also introduced by
growing trees with bootstrapped samples (sampling with replacement) of
the training set, which results in approximately one-third of sample data
being left out of each tree. The unselected, out-of-bag samples can also be
used to estimate the generalization error of the random forest model
(Breiman, 2001).
The R programming language (R Core Team, 2013) was used with the
“randomForest”package (Liaw and Wiener, 2002) to create random forest
classification models of groundwater fluoride. The output of the random
forest is therefore the probability of the occurrence of high fluoride concen-
trations. As such, the WHO drinking water guideline of 1.5 mg/L was
generally used to define the cut-off between high and low fluoride concen-
trations. Bootstrapped sampling was made with a balancebetween the two
categories of high and low fluoride.
The performance of a model can be evaluated through the accuracy of
the predictions for a given probability cut-off value. The Area Under the
ROC (receiver operator characteristic) Curve (AUC) overcomes the subjec-
tivity of choosing a threshold by summing up over true positive and false
positive rates at all cut-off values (Huang and Ling, 2005). The AUC score
is bounded between 0 and 1 (perfect classifiers); an uninformative classifier
that uses random guessing may yield 0.5 (Tharwat, 2020).
Variance importance plots of a random forest model help evaluate the
influence of individual variables. They are composed of two importance in-
dices, namely, Mean Decrease Accuracy (MDA) and Mean Decease Impurity
(MDI). MDA is calculated by averaging the change in out-of-bag error
estimates after permuting an individual variable's values in out-of-bag ob-
servations (Biau and Scornet, 2016). In general, permuting the values of
an important variable will lead to deterioration of model performance,
thus a decrease in accuracy. Likewise, MDI is defined as the average
decrease in node Gini impurity from splitting an individual variable over
all grown decision trees (Biau and Scornet, 2016). Gini impurity indicates
the homogeneity of a node, which is lower after splitting an important var-
iable that divides observations roughly into the same class.
To simplify the model but retain its predictive power, variable selection
by recursive feature elimination (RFE) was conducted with the caret
package in R (Kuhn, 2009), which backwardly reduces the number of
variables by removing the least important one at each step. Random forests
were created starting with all 48 variables (Table S4) and working down to
just one variable. The predictor subset ultimately chosen was that with
which the corresponding model was simplest and obtained high test AUC
score.
For all models, 1000 trees were grown, and the default number of vari-
able candidates at each split (i.e. square root of the number of features) was
used. Sampling for each tree was made with replacement and an even
balance between high and low fluoride classes. The optimal number of sam-
ples from each class was determined by trying 70%, 80%, 90%, and 100%
of the minority class (high fluoride) and selecting the sample size with the
highest AUC score in 10-fold cross validation. The final random forest
model was built with the RFE-selected variables and optimal sample size.
A prediction map of the occurrence of fluoride contamination >1.5 mg/L
was then created by applying the final random forest model to the predictor
datasets using the raster package (Hijmans, 2021).
We then estimated the population in Pakistan at risk of exposure to high
fluoride concentrations from groundwater used as drinking water by multi-
plying the population densit y in 2020 (Gao, 2017;Jones and O’Neill, 2016)
by the probability of high fluoride concentrations and accounting for the
average rate of domestic groundwater usage (39.1%) (World Health
Organization and UNICEF, 2019). Only areas above the probability cut-
off value at which the accuracy rates for the two classes (sensitivity and
specificity) are equal were taken into consideration.
3. Results and discussion
3.1. Fluoride concentrations in groundwater
Summary statistics, box plots and spatial distributions of the new
groundwater fluoride measurements (n= 2160) are shown in Table S2,
Fig. S1-and Fig. S3, respectively. Whereas the data points from previously
published studies (Ali et al., 2019;Brahman et al., 2013;Khattak et al.,
Table 1
Correlations and significance of environmental parameters used in the model with fluoride concentrations.
Type Variable Resolution Correlation (p)
Climate (continuous) Aridity index (Zomer et al., 2007;Zomer et al., 2008) 1000 m −0.1276 (3.09E-21)
Actual evapotranspiration (AET) (mm/year) (Trabucco and Zomer, 2010) 1000 m 0.0276 (4.17E-02)
Potential evapotranspiration (PET) (mm/year) (Zomer et al., 2007;Zomer et al., 2008) 1000 m 0.1550 (1.07E-30)
Precipitation (mm/year) (Fick and Hijmans, 2017) 1000 m 0.0412 (2.31E-03)
Temperature (°C)(Fick and Hijmans, 2017) 1000 m 0.0902 (2.53E-11)
Soil (continuous) Silt fraction (g/kg) (Hengl et al., 2017) 250 m −0.0319 (1.84E-2)
Nitrogen fraction (cg/kg) (Hengl et al., 2017) 250 m −0.1363 (4.83E-24)
Coarse fragments fraction (cm
3
/dm
3
)(Hengl et al., 2017) 250 m 0.2515 (1.75E-79)
Organic carbon density (g/dm
3
)(Hengl et al., 2017) 250 m 0.1111 (1.91E-16)
Soil (categorical) Arenosols (Hengl et al., 2017) 250 m 0.6677
Calcisols (Hengl et al., 2017) 250 m 0.3010
Cambisols (Hengl et al., 2017) 250 m 0.0993
Gypsisols (Hengl et al., 2017) 250 m 0
Solonchaks (Hengl et al., 2017) 250 m 0.0376
Solonetz (Hengl et al., 2017) 250 m 0.0500
Lithology (categorical) Carbonate sedimentary rocks (Hengl, 2018) 250 m 0.2468
Metamorphic rocks (Hengl, 2018) 250 m 0.0566
Mixed sedimentary rocks (Hengl, 2018) 250 m 0.2025
Evaporite (Hengl, 2018) 250 m 0.1644
Siliciclastic sedimentary rocks (Hengl, 2018) 250 m 0.0222
Topography (continuous) Elevation (m) ( Verdin, 2017) 100 m −0.0275 (4.22E-2)
Land cover (categorical) Shrubland (Buchhorn et al., 2020) Polygon 0.5664
Cropland (Buchhorn et al., 2020) Polygon 0.0885
Herbaceous vegetation (Buchhorn et al., 2020) Polygon 0.3839
Y. Ling et al. Science of the Total Environment 839 (2022) 156058
4

2022;Naseem et al., 2010;Rafique, 2008;Rafique et al., 2008;Rafique
et al., 2009)areconfined almost exclusively to the Punjab and Sindh prov-
inces, our data span the entire country (Fig. S5).
High fluoride concentrations were detected in all regions, including
maximum values of 27.5 mg/L in Punjab and 33.3 mg/L in Sindh. How-
ever, the average concentrations by province are all less than 1.5 mg/L
(Table S2). As seen in the inset of Fig. S3, high-fluoride areas were discov-
ered in the Sargodha Division in northwest Punjab, which had an average
measured fluoride concentration of 1.8 mg/L. Particularly affected are the
upper-Thal Desert districts of Bhakkar, Khushab and Mianwali.
Correlations between fluoride and other geochemical indicators are
shown in Fig. S2. The alkaline environment and the presence of bicarbon-
ates create a favorable condition for high fluoride waters, while calcium
ions suppress fluoride in groundwater.
3.2. Prediction modeling
3.2.1. Random forest model
Of the 48 variables considered, eight were selected by the RFE process:
actual evapotranspiration (AET), aridity index, coarse fragments fraction,
elevation, nitrogen fraction, PET, precipitation, and temperature. The RFE
algorithm selects variables based on their importance, which is generally
lower in categorical variables than in continuous ones due to only a fixed
(generally small) number of possible values in the former. The following
binary predictors were therefore manually added: arenosols, calcisols,
cambisols, carbonate sedimentary rocks, cropland, evaporite, gypsisols,
herbaceous vegetation, mixed sedimentary rocks, shrubland, siliciclastic
(non‑carbonate) rocks, solonchaks, solonetz, and herbaceous vegetation,
which are considered to be important from a geochemical standpoint, for
example, by acting as a sink or source of fluoride (Ali et al., 2019;García
and Borgnino, 2015;Podgorski et al., 2018). The optimal sample size deter-
mined by tuning was 562, which is 80% of the number of the minority class
(high fluoride) in the training dataset.
The final random forest model attained an AUC score of 0.92 as deter-
mined with the test dataset (Fig. 2b). The cut-off value of 0.47 was found
at the point at which sensitivity equals specificity, that is, where the accu-
racy rates for the two classes are evenly balanced. Using this cut-off, the
overall accuracy with the test set is 0.83, which is comparable to the
out-of-bag accuracy of 0.82 (accuracy calculated with out-of-bag samples
during training), confirming that the distribution of data in the training
and testing datasets is generally similar.
Unsurprisingly, the measured importance of the continuous variables is
higher than that of the binary ones (Fig. 2a). The climate predictors
temperature, precipitation, PET, aridity index, and AET as well as nitrogen
content, coarse fragments fraction, and elevation received the highest
importance. Among the binary predictors, calcisols and cropland are the
most important.
3.2.2. Prediction map
The prediction map derived from the randomforest model (Fig. 3)indi-
cates that approximately 30% of Pakistan is at-risk of fluoride concentra-
tions in groundwater exceeding 1.5 mg/L. Two particularly high-hazard
areas include the Thar Desert (Sindh Province) and the Bhakkar, Mianwali,
and Khushab districts in the upper Thal Desert (Sargodha Division, Punjab
Province). Both locations have an arid climate and are situated within the
flat-lying Indus Plain. Higher probabilities are also found in the Sulaiman
Mountains in eastern Balochistan. While most areas of the prediction map
clearly show higher or lower hazard, the determination ismore ambiguous
in southwestern Balochistan, where the probabilities are around 0.50. This
may result from a relative lack of measurements to adequately sample the
distribution of fluoride in this region, whereas most of the data used in
the model stem from the Indus Plain.
The fluoride prediction map was compared with a similar fluoride map
of India (Podgorski et al., 2018) (Fig. S8). High probabilities of fluoride
exceeding 1.5 mg/L in northern Punjab align very well with those across
the border in Indian Punjab. Similarly in the Thar Desert, which forms a
natural boundary between Pakistan and India, the predictions on either
side of the border are compatible. Whereas these sections are well
constrained by fluoride data, border areas in between match less well,
which may be due in part to a lack of measurements in southern
Pakistani Punjab.
Since much of Pakistan has a warm, dry climate and people conse-
quently drink more water, a prediction model was also created for fluoride
concentrations exceeding 1.0 mg/L (Fig. S7). However, with only 543 of
the 5483 measurements falling in the range of 1.0–1.5 mg/L, the
high-hazard regions of the prediction maps for 1.5 mg/L (Fig. 3)and
1.0 mg/L (Fig. S7) are largely similar. For example, higher probabilities
for 1.0 mg/L were found for central Pakistan near the borders of Punjab,
Khyber Pakhtunkhwa, and Balochistan. The selection of variables by RFE
was also similar for both models (1.0 mg/L and 1.5 mg/L).
3.2.3. Predictor importance
Some of the most important predictors are the climate parameters
(Fig. 2a). The correlations in Table S4 also confirm that drier climate condi-
tions favor fluoride release. High importance was also found for the eleva-
tion variable, which is strongly negatively correlated with temperature.
Fig. 2. Random forest modeling results of groundwater fluoride concentrations in Pakistan exceeding the WHO guideline of 1.5 mg/L. a) Variable importance plots of Mean
Decrease Accuracy and Mean Decrease Gini Impurity for each variable in the random forest model. Due to having only two degrees of possible values, the binary variables
(bin.) show lower importance scores relative to theother continuous variables. b) ROC curve (AUC score: 0.92) of the model response using the test dataset plotted with the
true positive rate (sensitivity) against the false positive rate (1 –specificity). The red dot indicates the cut-off value of 0.47.(For interpretationof the references to colorin this
figure legend, the reader is referred to the web version of this article.)
Y. Ling et al. Science of the Total Environment 839 (2022) 156058
5

The two continuous variables of soil parameters, fraction of coarse soil
fragments and nitrogen fraction, also have high importance measures.
The fraction of coarse fragments is high in the Sulaiman Mountains of
eastern Balochistan and western Punjab and Sindh and are composed of
mixed or carbonate sedimentary rock that may contain fluoride-bearing
minerals (Fig. 4). On the other hand, nitrogen fraction is associated closely
with the presence of forested mountains in the north, which has lower tem-
peratures, and where the sparse measurementsgenerally show low fluoride
concentrations. The “calcisols”binary soil predictor, which is associated
with substantial accumulations of lime, is connected to the presence of
high fluoride concentrations, as the precipitation of calcite removes
calcium from dissolution and results in higher fluoride concentrations
(Banerjee, 2015).
Carbonate sedimentary rocks can represent an important source of fluo-
rite (García and Borgnino, 2015), and silicate minerals may contain small
amounts of fluorine. Shrubland and herbaceous vegetation are the principal
vegetation cover type in the highly contaminated Thar Desert, and thus
assigned high importance in the model, though it is unclear if this relation-
ship exists elsewhere.
In the Thal Desert (Punjab), calcisols likely control Ca
2+
levels and
enhance the dissolution of F-bearing minerals. Moreover, alkaline pH
conditions can promote F
−
leaching by ion exchange processes. This is
particularly the case along the Jhelum River where water samples are dom-
inated by the Na-Cl type of water (Ali et al., 2019). A rapid change in aqui-
fer recharge in recentyears has resulted in increasedfluoride levels (Younas
et al., 2019). High alkaline waters in the Lahore and Kasur areas, which
experience excessive pumping, are also associated with high fluoride
leaching (Farooqi et al., 2007).
While calcisols are also the dominant soil type in the Sulaiman Moun-
tains (Fig. 2d), fluoride levels are not as high as in the Thal Desert, possibly
due to a stable water table.In the south in the Thar Desert (Sindh), dolomite
dissolution and arid climatic conditions promote evaporation process and
the dissolution of evaporites, contributing to the formation of saline
groundwater.
3.3. Health risk map
A health risk map was made using the optimal cut-off value of 0.47
(Fig. 5). It identifies numerous densely populated regions where people
rely on groundwater associated with high fluoride. In total, over 13 mil-
lion people are estimated to be potentially affected by fluoride contam-
ination in groundwater, which is 6.0% of the total population in
Pakistan. However, this number would be even larger if the population
in regions with probabilities under the cut-off of 0.47 would be taken
into consideration. With the increasing population in Pakistan (popula-
tion growth rate around 2% in the year 2020) (World Bank, 2020), the
problem may become more severe in the future if the reliance on
groundwater remains high. Large at-risk populations are found in north-
ern Punjab, Islamabad, and Khyber Pakhtunkhwa. Furthermore, the
fluoride risk map (Fig. 5) indicates that residents in the cities of Lahore,
Sargodha, Depalpur, Peshawar, Bannu, Karachi, Quetta, and others are
at high risk, which is confirmed by the high prevalence of fluorosis in
some of these cities (Ahmad et al., 2020;Mohsin et al., 2014;Rahman
et al., 2018;Sami et al., 2016). Conversely, the total number of people
at risk in the Thar Desert (SE Sindh), which has high probabilities of
fluoride contamination, is far smaller owing to its sparse population,
yet residents there are still under high risk.
The presented probability and health risk maps (Figs. 3 and 5)raise
awareness about fluoride contamination and its adverse health impacts
in Pakistan. Furthermore, they can help authorities in prioritizing areas
Fig. 3. Prediction map at a resolution of approximately 250 m of groundwater fluoride inPakistan exceedingthe WHO guideline of1.5 mg/L. (AUC scoreis 0.92). a) Full map
with locations of insets. b) Fluoride hotspot of the Sargodha Division in the upper Thal Desert (Bhakkar, Mianwali, Khushab districts). c) Hotspots in the Kasur and densely
populated Lahore districts. d) Hotspot of theThar Desert. The prediction map is also available for viewing at high-resolution on the GIS-based Groundwater Assessment Plat-
form (GAP), www.gapmaps.org.
Y. Ling et al. Science of the Total Environment 839 (2022) 156058
6

Fig. 4. Maps of selected predictor variables of a) coarse fragments fraction, b) AET, c) lithology, and d) soil groups.
Fig. 5. Population at risk of exposure to fluoride concentrations in groundwater exceeding 1.5 mg/L.
Y. Ling et al. Science of the Total Environment 839 (2022) 156058
7

for implementing mitigation measures. These could include monitoring
programs for drinking water wells or fluoride removal, e.g. adsorption
treatment (Bhatnagar et al., 2011) or membrane separation processes
(Waghmare and Arfin, 2015), and improving health management sys-
tems. Compared to the previous nationwide representation of fluoride
at the sub-tehsil-scale (Khan et al., 2002), the novel maps presented
here have a 3–4 order of magnitude higher spatial resolution (250 m),
are based on much larger new datasets, and predict the probability of
high groundwater fluoride for areas where data are lacking. Also in re-
lation to a recent study by Khattak et al. that contains clusters of many
groundwater fluoride measurements across much of Punjab (Khattak
et al., 2021), the new maps identify hotspots, e.g. in the Sargodha Divi-
sion, that that study did not uncover.
4. Conclusions
This study, which presents a large new dataset of fluoride in ground-
water across Pakistan, combined with geospatial modeling and risk
mapping using various environmental predictors, highlights several re-
gions where exposure to high fluoride levels pose a significant public
health risk. Hot spots include the Thal Desert in Punjab (Sargodha Divi-
sion), the Thar Desert in Sindh, and the Sulaiman Mountains in the west-
ern part of the country. Analysis of the importance of the predictor
variables and their correlation with fluoride show that high fluoride
concentrations in groundwater benefit from arid climatic conditions
with high temperatures and evapotranspiration, the presence of
fluoride-bearing minerals (e.g. carbonate sedimentary rock), and the
presence of calcisols.
Knowing the countrywide groundwater fluoride risk and affected popu-
lations shall be helpful for authorities and water resource managers in iden-
tifying fluoride-contaminated wells and mitigating the risk for residents. All
groundwater wells in areas with a high probability (e.g., above the cut-off
value of 0.47) should be tested, for instance, in the Thar Desert and the
SargodhaDivision (especially the Bhakkar, Mianwali,and Khushab districts
in the upper Thal Desert). Particular attention should also be paid to risk
areas with a high population density such as Lahore, Sargodha, Depalpur,
Peshawar, and Bannu. Mitigation measures include monitoring, provision
of alternative sources of drinking water, fluoride removal treatment, and
awareness-raising campaigns. These maps are not a replacement for actual
groundwater testing but indicate hazard and risk for drinking water use.
Future work could consider additional groundwater contaminants, e.g.
uranium, nitrate, pesticides or salinity in order to obtain a more compre-
hensive understanding of the safety of groundwater. Model accuracy
could be further improved by incorporating additional data and other
predictor variables, such as hydrological parameters, if available.
CRediT authorship contribution statement
Yuya Ling:Data curation, Formal analysis, Investigation, Methodology,
Software, Validation, Visualization, Writing –original draft, Writing –
review &editing. Joel Podgorski: Conceptualization, Data curation,
Formal analysis, Funding acquisition, Investigation, Methodology, Project
administration, Software, Validation, Visualization, Writing –review &
editing. Muhammad Sadiq: Data curation, Formal analysis, Validation.
Hifza Rasheed: Data curation, Formal analysis, Validation. Syed Ali
Musstjab Akber Shah Eqani: Conceptualization, Funding acquisition,
Project administration, Supervision, Data curation, Validation, Writing –
review &editing. Michael Berg: Conceptualization, Funding acquisition,
Investigation, Methodology, Project administration, Supervision, Visualiza-
tion, Writing –review &editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial inter-
ests or personal relationships that could have appeared to influence the
work reported in this paper.
Acknowledgments
We are grateful to Peter Molnar for valuable input and feedback on this
work. This project benefitted from financial support of the Swiss Agency
for Development and Cooperation (project no. 7F-09963.01.01). M.S. and
S.A.M.A.S.E. acknowledge the Higher Education Commission of Pakistan
for project funding (project no. 20-14825/NRPU/R&D/HEC/2021 2021)
and the scholarship award (project no. 1-8/HEC/HRD/2021/10887).
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://doi.
org/10.1016/j.scitotenv.2022.156058.
References
Ahmad, M., Jamal, A., Tang, X.W., Al-Sughaiyer, M.A., Al-Ahmadi, H.M., Ahmad, F., 2020.
Assessingpotable water quality and identifying areas of waterborne diarrheal and fluoro-
sis health risks using spatial interpolation in Peshawar, Pakistan. Water 12 (8).
Ali, W., Aslam, M.W., Junaid, M., Ali, K., Guo, Y.K., Rasool, A., Zhang, H., 2019. Elucidating
various geochemical mechanisms drive fluoride contamination in unconfined aquifers
along the major rivers in Sindh and Punjab, Pakistan. Environ. Pollut. 249, 535–549.
Amini, M., Mueller, K., Abbaspour, K.C., Rosenberg, T., Afyuni, M., Moller, K.N., et al., 2008.
Statisti cal modeling of g lobal geo genic fluoride contamination in groundwaters. Environ.
Sci. Technol. 42 (10), 3662–3668.
Ayoob, S., Gupta, A.K., 2006. Fluoride in drinking water: a review on the status and stress
effects. Crit. Rev. Environ. Sci. Technol. 36 (6), 433–487.
Banerjee, A., 2015. Groundwater fluoride contamination: a reappraisal. Geosci. Front. 6 (2),
277–284.
Bhatnagar, A., Kumar, E., Sillanpää, M., 2011. Fluoride removal from water by adsorption—a
review. Chem. Eng. J. 171 (3), 811–840.
Bhowmik, A.K., Alamdar, A., Katsoyiannis, I., Shen, H., Ali, N., Ali, S.M., Eqani, S.A.M.A.S.,
2015. Mapping human health risks from exposure to trace metal contamination of drink-
ing water sources in Pakistan. Sci. Total Environ. 538, 306–316.
Biau, G., Scornet, E., 2016. A random forest guided tour. TEST 25 (2), 197–227.
Bo, Z., Mei, H., Yongsheng, Z., Xueyu, L., Xu elin, Z., Jun, D., 2003. Distribution and risk
assessment of fluoride in drinking water in th e west plain regio n of Jilin province,
China. Environ. Geochem. Health 25 (4), 421–431.
Brahman, K.D., Kazi, T.G., Afridi, H.I., Naseem, S., Arain, S.S., Ullah, N., 2013. Evaluation of
high levels of fluoride, arsenic species and other physicochemical parameters in under-
ground water of two sub districts of tharparkar, Pakistan: a multivariate study. Water
Res. 47 (3), 1005–1020.
Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32.
Buchhorn, M., Smets, B., Bertels, L., De Roo, B., Lesiv, M., Tsendbazar, N.-E., et al., 2020. Co-
pernicusGlobal Land Service:Land Cover 100m: Collection 3 Epoch 2015, Globe. Version
V3. 0.1) [Data set].
Chen, W., Xie, X.S., Wang, J.L., Pradhan, B., Hong,H.Y., Bui, D.T., et al.,2017. A comparative
study of logistic model tree, random forest, and classification and regression tree models
for spatial prediction of landslide susceptibility. Catena 151, 147–160.
Edmunds, W.M., Smedley, P.L., 2013. Fluoride in natural waters. Essentials of Medical Geol-
ogy. Springer, pp. 311–336.
Erickson, M.L., Elliott, S.M., B rown, C.J., Stac kelberg, P.E., Ransom, K.M., Red dy, J.E.,
Cravotta III, C.A., 2021. Machine-learning predictions of high arsenic and high manga-
nese at drinking water depths of the glacial aquifer system, northern continental United
States. Environ. Sci. Technol. 55 (9), 5791–5805.
Farooqi, A., 2015. Arsenic and Fluoride Contamination. A Pakistan Perspective.
Farooqi, A., Masuda, H., Firdous, N., 2007. Toxic fluoride and arsenic contaminated ground-
water in the Laho re and Kasur dist ricts, Punjab, P akistan and poss ible contamina nt
sources. Environ. Pollut. 145 (3), 839–849.
Fick, S.E., Hijmans, R.J., 2017. WorldClim 2: new 1-km spatial resolution climate surfaces for
global land areas. Int. J. Climatol. 37 (12), 4302–4315.
Gao, J., 2017. Downscaling Global Spatial Population Projections From 1/8-Degree to 1-km
Grid Cells. National Center for Atmospheric Research, Boulder, CO, USA.
García, M.G., Borgnino, L., 2015. Fluoride in the Context of the Environment.
Greenman,D.W., Swarzenski,W.V., Bennett, G.D., 1967. Ground-water Hydrology of the Pun-
jab, West Pakistan, With Emphasis on Problems Caused by Canal Irrigation. Government
Printing Office.
Handa, B., 1975. Geochemistry and genesis of fluoride-containing ground waters in India.
Groundwater 13 (3), 275–281.
Hengl, T., 2018. Global Landform and Lithology Class at 250 m Based on the USGS Global
Ecosystem Map. Zenodo.
Hengl, T., deJesus, J.M., Heuvelink, G.B.M., Gonzalez, M.R., Kilibarda, M., Blagotic, A., et al.,
2017. SoilGrids250m: global gridded soil information based on machine learning. PLoS
One 12 (2).
Hijmans, R.J., 2021. Geographic Data Analysis and Modeling [R Package Raster Version 3.4-
10].
Huang, J., Ling, C.X., 2005. Using AUC and accuracy in evaluating learning algorithms. IEEE
Trans. Knowl. Data Eng. 17 (3), 299–310.
Iwasaki, A., 2007. Mucosal dendritic cells. Annu. Rev. Immunol. 25, 381–418.
James, G., Witten, D., Hastie, T., Tibshirani, R., 2013. An Introduction to Statistical Learning.
112. Springer.
Y. Ling et al. Science of the Total Environment 839 (2022) 156058
8

Jones, B., O’Neill, B.C., 2016. Spatially explicit globalpopulation scenariosconsistent with the
shared socioeconomic pathways. Environ. Res. Lett. 11 (8), 084003.
Khan, A., Whelton, H., O'Mullane, D., 2002. A map of natural fluoride in drinking water in
Pakistan. Int. Dent. J. 52 (4), 291–297.
Khan, A., Whelton, H., O'Mullane, D., 2004. Determining the optimal concentration of fluo-
ride in drinking water in Pakistan. Community Dent. Oral Epidemiol. 32 (3), 166–172.
Khan, S., Moheet, I.A., Farooq, I., Farooqi, F.A., ArRejaie, A.S., Al Abbad,M.H.A., Khabeer, A.,
2015. Prevalence of dental fluorosis in school going children of Dammam, Saudi Arabia.
J. Dent. Allied Sci. 4 (2), 69.
Khattak, J.A., Farooqi, A., Hussain, I., Kumar, A., Singh, C.K., Mailloux, B.J., et al., 2021.
Groundwater fluoride acrossthe Punjab plains of Pakistan and India: Distribution and un-
derlying mechanisms. Sci. Total Environ. 151353.
Khattak, J.A., Farooqi, A., Hussain, I., Kumar, A., Singh, C.K., Mailloux, B.J., et al., 2022.
Groundwater fluoride across the Punjab plains of Pakistan and India: distribution andun-
derlying mechanisms. Sci. Total Environ. 806, 151353.
Khwaja, M.A., Aslam, A.,2018. Comparative Assessment of Pakistan NationalDrinking Water
Quality Standards With Selected Asian Countries and World Health Organization.
Kuhn, M., 2009. The caret package. J. Stat. Softw. 28 (5).
Kumar, M., Das, N., Goswami, R., Sarma, K.P., Bhattacharya, P., Ramanathan, A.L., 2016.
Coupling fractio nation and batch deso rption to understand arsenic and fluoride co-
contamination in the aquifer system. Chemosphere 164, 657–667.
Kumar, M., Goswami, R., Patel, A.K., Srivastava, M., Das, N., 2020.S cenario, perspecti ves and
mechanism of arsenic and fluoride co-occurrence in t he groundwater : a review.
Chemosphere 249.
Liaw, A., Wiener, M., 2002. Classification and regression by randomForest. R News 2 (3),
18–22.
Mohsin, A.,Hakeem, S., Arain, A.H.,Ali, T., Mirza, D., 2014.Frequency and severity of dental
fluorosis among school children in Gadap Town, Karachi. Pakistan Oral &Dental Journal
34 (4).
Naseem, S.,Rafique, T., Bashir,E., Bhanger, M.I., Laghari, A., Usmani, T.H.,2010. Lithological
influenceson occurrence of high-fluoride groundwater in Nagar Parkar area, Thar Desert,
Pakistan. Chemosphere 78 (11), 1313–1321.
Ozsvath, D.L., 2009. Fluoride and Environmental Health: A Review. 8, pp. 59–79 1.
PCRWR, 2015. Pakistan Council of Research in Water Resources PCRWR, Pakistan.
Podgorski, J., Berg, M., 2020. Global threat of arsenic in groundwater. Science 368 (6493),
845.
Podgorski,J., Eqani, S.A.M.A.S., Khanam, T., Ullah, R., Shen, H.Q., Berg, M., 2017. Extensive
arsenic contaminationin high-pH unconfined aquifers in the IndusValley. Sci. Adv. 3 (8).
Podgorski,J., Labhasetwar, P., Saha, D., Berg, M., 2018. Prediction modeling andmapping of
groundwater fluoride contamination throughout India. Environ. Sci. Technol. 52 (17),
9889–9898.
Podgorski,J., Wu, R., Chakravorty, B., Polya,D.A., 2020. Groundwater arsenic distribution in
India by machine learning geospatial modeling. Int. J. Environ. Res. Public Health 17
(19), 711 9.
Qureshi, A.S., 2020. Groundwater go vernance in Pakistan: from col ossal development to
neglected management. Water 12 (11).
R Core Team, 2013. R: A Language and Environment for Statistical Computing.
Rafique, T., 2008. Occurrence, Distribution and Origin of Fluoride-rich Groundwater in the
Thar Desert. Sindh University Jamshoro, Pakistan.
Rafique, T., Naseem, S., Bhanger, M.I., Usmani, T.H., 2008. Fluoride ion contamination in the
groundwater of Mi thi sub-distri ct, the Thar Deser t, Pakistan. Envi ron. Geol. 56 (2),
317–326.
Rafique, T., Naseem, S., Usmani, T.H., Bashir, E., Khan, F.A., Bhanger, M.I., 2009. Geochem-
ical factors controlling the occurrence of high fluoride groundwater in the Nagar Parkar
area, Sindh, Pakistan. J. Hazard. Mater. 171 (1–3), 424–430.
Rafique, T., Ahmed, I., Soomro, F., Khan, M.H., Shirin, K., 2015. Fluoride levels in urin e,
blood plasma and serum of people living in an endemicfluorosis area in the Thar Desert,
Pakistan. J. Chem. Soc. Pak. 37 (6), 1212–1219.
Rahman, Z.U., Khan, B., Ahmada, I., Mian, I.A., Saeed, A., Afaq, A., et al., 2018. Areviewof
groundwater fluoride contamination in Pakistan and an assessment of the risk of fluoro-
sis. Fluoride 51 (2), 171–181.
Rasheed, H., Iqbal, N., Ashraf, M., ul Hasan, F., 2022. Groundwater quality and availability
assessment: a case study of District Jhelum in the Upper Indus, Pakistan. Environ. Adv.
7, 100148.
Rasool, A., Farooqi, A., Xiao, T., Ali, W., Noor, S., Abiola, O., et al., 2018. A review of global
outlook on fluoride contamination in groundwater with prominence on the Pakistan cur-
rent situation. Environ. Geochem. Health 40 (4), 1265–1281.
Raza, M., Hussain, F., Lee, J.Y., Shakoor, M.B., Kwo n, K.D., 2017. Groundwater status in
Pakistan: a review of contamination, health risks, and potential needs. Crit. Rev. Environ.
Sci. Technol. 47 (18), 1713–1762.
Reichstein, M., Camps-Valls, G., Stevens, B., Jung, M., Denzler, J., Carvalhais, N., Prabhat,
2019. Deep learning and process understanding for data-driven earth system science. Na-
ture 566 (7743), 195–204.
Sami, E., Vichayanrat, T., Satitvipawee, P., 2016. Caries with dental fluorosis and oral health
behaviour among 12-year school children in moderate-fluoride drinking water commu-
nity in Quetta, Pakistan. J. Coll. Phys. Surg. Pakistan 26 (9), 744–747.
Sanaullah, M., Me hmood, Q., Ahmad , S.R., Rehman, H .U., 2019. Arseni c contamination
trends of abandoned river banks: a case study at the left bank of river Ravi, Lahore. Int.
J. Econ. Environ. Geol. 21–24.
Shah, S., Bandekar, K., 1998. Drinking water compared to WHO guidelines (1993). Indian
Waterworks Assoc. 30, 179–184.
Singh, C.K., Mukherjee, S., 2015. Aqueous geochemistry of fluoride enriched groundwater in
arid part of Western India. Environ. Sci. Pollut. Res. 22 (4), 2668–2678.
Tharwat, A., 2020. Classification assessment methods. Appl. Comput. Informatics 17 (1).
Trabucco, A., Zomer, R.J., 2010. Global soil water balance geospatial database. CGIAR Con-
sortium for spatial information. CGIAR Consort ium for Spatial Information. https://
cgiarcsi.community.
Ullah, Zahid, et al., 2022. Arsenic contamination, water toxicity, source apportionment, and
potential heal th risk in groundw ater of Jhelum Basin, Punjab, Pakist an. Biol. Trace
Elem. Res. 1–11.
Verdin, K., 2017. Hydrologic Derivatives for Modeling and Applications (HDMA) Database:
US Geological Survey Data Release.
Waghmare, S.S., Arfin, T., 2015. Fluoride removal from water by various techniques. Int.
J. Innov. Sci. Eng. Technol. 2 (9), 560–571.
Wang, B.B., Zheng, B.S., Zhai, C., Yu, G.Q., Liu, X.J., 2004. Relationship between fluorine in
drinking water and dental health of residents in some large cities in Ch ina. Environ.
Int. 30 (8), 1067–1073.
WAPDA/EUAD, 1989. Booklet on Hydrogeological Map of Pakistan, 1:2,000,000 Scale.Water
&Power Development Authority, Lahore and Environment &Urban Affairs Division,
Govt. of Pakistan, Islamabad.
Winkel, L., Berg, M., Amini, M., Hug, S.J., Johnson, C.A., 2008. Predicting groundwater arse-
nic contaminat ion in Southeas t Asia from surf ace parameter s. Nat. Geosci. 1 (8),
536–542.
World Bank, 2020. P opulation grow th (annual), Pakistan. https://dat a.worldbank. org/
indicator/SP.POP.GROW?locations=PK.
World Health Organization, 1994. Expert committee on oral health status and fluoride use.
Fluorides and Oral Health. WHO Technical Report Series. 846.
World Health Organization, 2011. Guidelines for drinking-water quality. WHO Chronicle 38
(4), 104–108.
World Health OrganizationUNICEF, 2019. Jonit Monitoring Programme for Water Supply,
Sanitation and Hy giene: Estima tes on the Use of Water, Sanitation and Hygiene in
Pakistan.
Wu, R., Podgorski, J., Berg, M., Polya, D.A., 2021. Geostatistical model of the spatial distribu-
tion of arsenicin groundwaters in Gujarat State, India. Environ. Geochem. Health 43 (7),
2649–2664.
Younas, A., Mushtaq, N., Khattak, J.A., Javed, T., Rehman, H.U., Fa rooqi, A., 2019.
High levels of fluoride contamination in groundwater of the semi-arid alluvial aquifers,
Pakistan: evaluating the recharge sources and geochemical identification via stable
isotopes and othe r major elemental da ta. Environ. Sci. Po llut. Res. 26 (35),
35728–35741.
Zomer, R.J., Bossio, D.A., Trabucco, A., Yuanjie, L., Gupta, D.C., Singh, V.P., 2007. Trees and
Water: Smallholder Agroforestry on Irrigated Lands in Northern India. 122. IWMI.
Zomer, R.J., Trabucco, A., Bossio, D.A., Verchot, L.V., 2008. Climate changemitigation: a spa-
tial analysis of global land suitability for clean development mechanismafforestation and
reforestation. Agric. Ecosyst. Environ. 126 (1–2), 67–80.
Y. Ling et al. Science of the Total Environment 839 (2022) 156058
9