Exploring soil pollution patterns in Ghana’s northeastern mining zone using
machine learning models

Daniel Kwayisi a,b, Raymond Webrah Kazapoe c, Seidu Alidu d, Samuel Dzidefo Sagoe e,
Aliyu Ohiani Umaru f, Ebenezer Ebo Yahans Amuah g,h,*, Prosper Kpiebaya i

a Department of Geology, University of Johannesburg, Auckland Park Kingsway Campus, South Africa
b Department of Earth Science, University of Ghana, Legon-Accra, Ghana
c Department of Geological Engineering, University for Development Studies, Nyankpala, Ghana
d Ghana Geological Survey Authority, P.O. Box M80, Accra, Ghana
e Department of Environment and Sustainability Sciences, University for Development Studies, Ghana
f Department of Geology, University of Maiduguri, Maiduguri, Borno State, Nigeria
g Department of Environmental Science, Kwame Nkrumah University of Science and Technology, Kumasi, Ghana
h Department of Civil Engineering, Takoradi Technical University, P. O. Box 256, Takoradi, Ghana
i Department of Soil Science, University for Development Studies, P. O. Box TL 1882, Nyanpkala, Ghana

A R T I C L E I N F O

Keywords:
Machine learning
Galamsey
Gold mining
Environmental degradation
Pollution indices

A B S T R A C T

This study assessed the pollution status and effectiveness of machine learning models in predicting pollution
indices in soils from a mining area in Northeastern Ghana. 552 soil samples were analysed with an Energy
Dispersive X-ray Fluorescence (ED-XRF) spectrometer for their elemental concentrations. Four pollution indices;
Nemerow Integrated Pollution Index (NIPI), degree of contamination (Cdeg), modified degree of contamination
(mCdeg) and Pollution Load Index (PLI). Additionally, the Multivariate Adaptive Regression Splines (MARS)
machine learning approach were used. The high CV%, skewness, and kurtosis values show a high degree of
variability and uneven distribution patterns which denotes dispersed hotspots that can be interpreted as an
influence of gold anomalies and illegal mining activities in the area. V (120.86 mg/L), Cr (242.42 mg/L), Co
(30.92 mg/L) Ba (337.62 mg/L), and Zn (35.42 mg/L) recorded values higher than the global and regional
contaminant thresholds. The NIPI shows that 46.74% and 26.81% of samples are slightly and moderately
polluted respectively. The Cdeg analysis supports these findings, with 36.96% and 41.49% of samples classified
as having “moderate” to “considerable” contamination, respectively. The PLI indicates progressive soil quality
deterioration (43.84%) of samples reflecting substantial environmental disturbance. The pollution indices show
the effect of illegal mining on Shaega, Buin and other areas in the eastern boundary of the study. The MARS
models developed for the study demonstrated high predictive capabilities with an R2 value of 0.9665 for model 1
(NIPI), and RMSE and MAE values of 0.8227 and 0.4287 respectively. For model 2 (Cdeg), R2 value of 0.9863,
RMSE and MAE of 1.0416 and 0.6181, respectively. Model 3 (mCdeg) produced an R2 value of 0.9844, RMSE and
MAE of 0.1225 and 0.0670. These findings suggest MARS models can be an integral tool for soil quality analysis
in cooperation with pollution indices. The study suggests that remedial and legislative measures be implemented
to address the issue of illegal mining in the area.

1. Introduction

Ghana’s vibrant gold mining industry is known to be associated with
significant environmental and ecological risk factors (Kazapoe et al.,
2023). These risks emanate from the industry’s small-scale and artisanal
mining (ASM) sector (Achina-Obeng and Aram, 2022). Some of these

activities are carried out with rudimentary equipment, are
labour-intensive, and often lack the regulation major mines are sub-
jected to (Bansah et al., 2016). This may result in the exposure of the
environment to mine waste or inputs which cause significant damage to
water bodies and the environment (Agboola et al., 2020; Amuah et al.,
2022a). Until recently, these activities have been particularly severe

* Corresponding author.
E-mail address: amuahyahans@gmail.com (E.E.Y. Amuah).

Contents lists available at ScienceDirect

Journal of Hazardous Materials Advances

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

https://doi.org/10.1016/j.hazadv.2024.100480
Received 17 August 2024; Received in revised form 17 September 2024; Accepted 21 September 2024

Journal of Hazardous Materials Advances 16 (2024) 100480 

Available online 22 September 2024 
2772-4166/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by- 
nc-nd/4.0/ ). 

mailto:amuahyahans@gmail.com
www.sciencedirect.com/science/journal/27724166
https://www.elsevier.com/locate/hazadv
https://doi.org/10.1016/j.hazadv.2024.100480
https://doi.org/10.1016/j.hazadv.2024.100480
https://doi.org/10.1016/j.hazadv.2024.100480
http://crossmark.crossref.org/dialog/?doi=10.1016/j.hazadv.2024.100480&domain=pdf
http://creativecommons.org/licenses/by-nc-nd/4.0/
http://creativecommons.org/licenses/by-nc-nd/4.0/


only in the Southern part of Ghana (Nunoo et al., 2022).
Gold mining in Northern Ghana is in its relative infancy, and

generally not as intensive as is the case in the Southern mining district.
However, the advent of mining communities in some Northern areas and
reported issues of adverse effects on the environment have brought it
under the spotlight. Studies in areas such as Nangodi, Sheaga and Lawra
areas have highlighted the risk the activities of such miners have on the
water bodies and soils of the local communities who rely on the water
and the crops which grow on possibly compromised soils for their daily
sustenance (Arhin et al., 2016; Moomen and Dewan, 2017; Isung, 2021;
Akoto et al., 2023). Equally as threatening are geogenic contaminants
such as fluoride and arsenic whose anomalous concentrations in soil and
groundwater have been reported in the study area (Arhin et al., 2016;
Arhin et al., 2017; Akoto et al., 2023).

This situation necessitates advanced environmental monitoring
techniques, such as spatiotemporal and multivariate analysis of ele-
ments, which have proven effective in capturing complex environmental
changes. Additionally, pollution and ecological indices have been suc-
cessfully used to assess environmental quality. Building on these
methods, this study incorporates machine learning models to enhance
monitoring accuracy and efficiency. Machine learning offers robust
analytical capabilities, identifying intricate patterns and relationships
within large datasets (Kamm et al., 2023; Strielkowski et al., 2023). This
integration allows for precise predictions and actionable insights,
contributing to more effective environmental management and
decision-making.

Combining machine learning with traditional techniques represents
a significant advancement, ensuring thorough assessments and a pro-
active approach to combating environmental challenges (Bibri et al.,
2024). To help in decoding the outcomes of the pollution indices, ML
models were used. Machine Learning (ML) is a contribution of various
algorithms that undertake predictions and tasks by learning from a given
“problem-specific” data and training on it. Machine learning analyzes
large amounts of data in a short time possibly providing precise pre-
diction. Javed et al. (2023) critically provides an understanding of how
machine learning can be utilized in improving speed and precision in the
medical sector. ML has made significant contributions to the advance-
ment of environmental science offering a wide range of innovative so-
lutions for the management and mitigation of environmental issues
plaguing the world. Through the analysis of vast amounts of data, ML
models are employed in the prediction of temperature changes, pre-
cipitation patterns, and disastrous weather with incredible accuracy
(Fister et al., 2023; Iglesias et al., 2024).

Multivariate Adaptive Regression Spline (MARS), a type of ML al-
gorithm, is a non-parametric variant of regression that extends the
ability of linear models to identify linear relationships to non-linear
relationships. The technique was developed by Friedman (1991) and
can automatically identify and model complex relationships without
requiring the usual assumptions that come with regression (Zhang and
Goh, 2016; Adiguzel and Cengiz, 2023). The technique operates through
the division of the dataset into separate sections with each section fitted
with its own piecewise linear regression. This is achieved with the help
of basis functions, the points where these pieces connect are known as
knots. Naser et al. (2022) was successful in employing the use of the
MARS algorithm in the prediction of compressive strength of
eco-friendly concrete and even found it superior to well-known ML
models like Random Forest (RF) and Support Vector Machine (SVM).
Gackowski et al. (2022) was also successful in predicting antitumor
activity in anthrapyrazole derivatives. These studies highlight MARS as
a suitable tool for us due to its high level of accuracy and flexibility
across several fields.

The objective of this study was to assess the environmental condition
of the soil in the Nangodi region concerning heavy metal pollution to
determine (i) the concentration of heavy metal pollutants; (ii) sources,
and interactions between these elements; (iii) spatial distribution; (iv)
ecological risk related to elements, and (v) whether MLmodels can serve

as reliable predictors of soil labelling such models as effective.

2. Methodology

2.1. Description of the study area

The Nangodi area (Fig. 1) is located within the Talensi-Nabdam
district of the Upper East Region (Amosah and Lukman, 2023). Ac-
cording to the Ghana Statistical Service (2010), the wider area has a
population of 94,650. It adjoins the Bolgatanga Municipality to the
North, the West and East Mamprusi Districts to the South, the
Kassena-Nanakana District to the West and the Bawku West District to
the East (Ghana Statistical Services, 2010). The district is located be-
tween the latitudes of 10. 15◦ and 10. It is located at latitude 60 ◦ north
of the equator and longitude 0. 31◦ and 10. 50. The area is predomi-
nantly an agricultural-based economy. A significant proportion of the
households (i.e. 85.9%) in the area are engaged in the cultivation of
crops, rearing of animals, and planting of trees among others. About
49.3% of those actively engaged in agricultural practices are males
while 50.7% are females (Ministry of Food and Agriculture, n.d.).
However, the relatively recent discovery of gold in the area has attracted
a lot of unemployed youth and small-scale mining ventures to the area,
whose activities are affecting the area (Tom-Dery et al., 2012; Kazapoe
et al., 2021).

The area is characterized by a gentle to medium slope of 1%-5%
while the upland slopes are about 10% (Ministry of Food and Agricul-
ture, n.d.). The district is drained mainly by the Red andWhite Volta and
their tributaries. The climate is warm and is characterized by two sea-
sons namely the Wet and the Dry. The rainfall is irregular and received
from May to October each year with an average annual rainfall of 950
mm. The dry season lasts for 7 months every year beginning in October
and ending in April (Sekyi-Annan, 2019). The temperature can rise to
45◦C in March and April and can drop to 12◦C in December (Ghana
Statistical Services, 2010). The district has three gazetted forest reserves
that cover a total area of 455.21 km; Nyokoko (established in 1954),
Tankwiddi East and The Red Volta were respectively established in the
year 1956 (Ministry of Food and Agriculture, n.d.). The forest reserves
unfortunately host enclaves of illegal mining pits which poses an envi-
ronmental challenge in the area (Okyere et al., 2021).

The area falls within the Nangodi greenstone belt, which is one of six
(6) Paleoproterozoic Birimian gold-bearing belts in Ghana (Dzikunoo
et al., 2021). These Paleoproterozoic formations are found in the Baoulé
Mossi domain of the West African Craton (Abouchami et al., 1990). The
stratigraphic package of the area is composed of volcanics including
andesites, spilites, greenstones, porphyries, meta-basalts, vesicular ba-
salts, rhyolites and tuffs along with metasediments (mostly phyllites,
greywackes and chlorite schist) (Kazapoe, 2014). The greenstones are
commonly massive and are often chloritised with the intrusives mostly
made up of granodiorites (Murray, 1960).

2.2. Soil sampling

For this study, a grid-based sampling method was adopted in order to
cover the entire area and obtain a good representation of the soil sam-
ples. This approach was selected to address natural factors including
topography and vegetation and anthropogenic factors including mining
and agriculture concerning soil parameters. The study area was divided
into equal-sized cells and soil samples were collected at regular intervals
thus 552 samples were obtained. The grid size and the sampling points
were computed according to the size of the study area and possible
variation in the soil type, which includes agricultural lands, mines, and
transitional area.

The samples were taken from the B-horizon by digging up to a depth
of 20 cm using soil augers and shovels. Materials such as stones, roots,
and other plant debris were sieved out in the field separately from the
soil particles and were neatly wrapped in polythene bags and labelled

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

2 



for laboratory analysis.

2.3. Analytical procedures

Composite samples were prepared using a mixture of three homo-
genised samples taken at each sampling point. The samples were sent to
the Ghana Geological Survey Authority (GGSA) for further treatment
and analysis. After the samples were brought to the laboratory, the soil
samples were left to dry at room temperature and in normal atmospheric
conditions for 72 h to remove moisture from the samples. These were
then dried in an oven and sieved through a 2mmmesh to ensure no large
particles and other debris were included in the analyses. To make the
samples as uniform as possible, 50 g of the sieved soil from each sample
was ground with a mechanical agate mortar and pestle, ensuring all the
particles were of small size. The powdered samples were then kept in
clean, airtight containers until the next step of the experiment. The
elemental analysis of the soil samples was done by using an Energy
Dispersive X-ray Fluorescence (ED-XRF) spectrometer. The XRF spec-
trometer was checked and standardized with standard samples to give
precise results for the readings taken. For each sample, approximately
10 grams of soil was placed in a die and compacted into a pellet at a
pressure of 10,000 kN for 1 min. The samples obtained were then placed
in the sample holder of the XRF spectrometer for analysis. All the sam-
ples were prepared and analysed in triplicate to ensure the results were
precise. The spectrometer was set to operate under the following con-
ditions: The instrument parameters used were an excitation voltage of
50 kV, a current of 1 mA, and a counting time of 300 s per sample. The
spectra of the emitted X-rays were recorded and processed to quantify
the concentrations of the materials present, including some of the heavy
metals. Some steps were taken in order to guarantee the accuracy and
credibility of the analysis. Blank measurements were done at appro-
priate time intervals in order to check for contamination and instru-
mental interferences. Control samples were used in each batch to check

the accuracy of the measurements due to sample repetition. To check the
reliability and precision of the instrument, CRMs with established
elemental concentrations were used alongside the soils. The XRF spec-
trometer was routinely run in the calibration mode with the CRMs to
check its stability and efficiency. All the procedures, settings, and
measurements were properly documented to create an organized record
of the quality control measures.

2.4. Quality Assurance and Quality Control Analysis (QA/QC)

Certified Reference Materials (CRMs) were applied as the control
samples in the course of the analysis to check the correctness of the
measurements. One (1) control sample was inserted for every twenty
(20) field samples based on international recommendations. The control
samples were used to check the validity of the elemental analysis done
on the samples of the soil. The measurement precision and the method
reproducibility were checked by including the duplicate samples in the
analysis. Consistent with normal protocol, about 5 per cent of the field
samples were duplicated. Hence, 28 duplicate samples were incorpo-
rated after every 20 samples when the samples were being analysed. As
both the original and duplicate samples were analyzed, the study was
able to assess the variability of the analytical procedures. The elements
show small average variations of 1.3–9.2% between duplicate sample
pairs, which was considered accepted following Arhin et al. (2019).

2.5. Machine Learning Model

2.5.1. Pollution indices
The NIPI, PLI, Cdeg and mCdeg were used in this study. Table 1

provides details of these.

2.5.2. Multivariate Adaptive Regression Splines (MARS)
The process through which the MARS model is built involves two

Fig. 1. Map of the study area showing the sampling locations.

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

3 



phases, forward and backwards. The forward phase involves the addi-
tion of basis functions considering the interactions that exist between
variables and knots (Ağyar et al., 2022). This usually results in the
creation of an overfitted model (Zhang and Goh, 2016; Ağyar et al.,
2022). The backward phase involves the pruning of basis functions that
contribute very little to the overall performance of the model. This re-
duces the possibility of overfitting thus enhancing model generalization
(Zhang and Goh, 2016).

The MARS model is expressed by the equation:

f(x) = β0 +
∑M

m=1
βm

∏Km

k=1
hkm

(
Xv(k,m)

)
(1)

Where f(x) represents the predicted value of the dependent variable y
based on the independent ones x, β0 the intercept, a constant repre-
senting the baseline value of the dependent variable when all predictors
are zero, βm the coefficient associated with the mth basis function to the

overall model,
∏Km

k=1
denotes the product of Kmhinge functions for the

mth basis function, hkm
(
Xv(k,m)

)
represents the hinge function, a piecewise

linear function of the variable Xv(k,m), and Xv(k,m)the predictor variable
indexed by v(k, m) (Şengül et al., 2020).

After the first phase, the basis function that did not significantly
contribute to model performance is removed using a method called
pruning. This is achieved using the generalized cross-validation error
(GCV). This is expressed by the Eq. (2):

GCV (λ) =

∑n
i=1

(
yi − yip

)2

(

1 −
M(λ)
n

)2 (2)

Where n represents the number of training cases, yi is the observed
values of the dependent variable, yip is the predicted values of the
dependent variable, and M(λ) is the penalty function for the model
complexity (Şengül et al., 2020).

2.5.3. Evaluation metrics
To determine the accuracy of the MARS model, several well-known

statistical evaluation metrics were employed. These metrics provide
important insights as to how well the model fits, the accuracy of the
predictions made by the models, and the errors generated by the model.

R squared/ Adjusted R squared (R2/Adj R2)
The coefficient of determination, also known as R2 measures the

proportion of the variances that exist within the dependent variable that
is explained by the independent variables. This metric is expressed by
the Eq. (3):

R2 = 1 −
SSres
SStot

(3)

Where SSres represents the residual sum of squares and SStot, the
total sum of squares. A coefficient of determination of 1 indicates the
model predicted perfectly and thus explains the variance within the
dependent variable, 0 indicates an inability to predict accurately thus
unable to explain the variance within the dependent variable (Kim,
2018). According to Hair et al. (2021), an R2 of 0.75 is described as
substantial, 0.50 as moderate, and 0.25 as weak.

Adj R2 is a modified version of the R2. It adjusts the R2 value based on
the number of predictors in the model. This helps in accounting for
overfitting. It is expressed by equation (Eq. (4)):

Adj R2 = 1 −
n − 1

n − m − 1
(
1 − R2

)
(4)

Where n represents the number of observations, and m, the number
of predictors (Trunfio et al., 2022).

2.5.4. Root Mean Square Error (RMSE)
RMSE determines the average errors generated between the pre-

dicted and observed values. It is denoted by the Eqn. 5:

RMSE =

̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n

i=1

(ŷi − yi)2

n

√

(5)

()Where yi represents the observed values and ŷi, the predicted
values, n the number of observations. A lower RMSE value represents a
more accurate model and vice versa (Chai and Draxler, 2014).

2.5.5. Mean Absolute Error (MAE)
MAE determines the average magnitude of absolute errors between

the observed and predicted values. It is represented by the Eq. (6):

MAE =
1
n
∑n

i=1
|yi − ŷi| (6)

Where yi represents the observed values and ŷi, the predicted values,
n the number of observations. Similar to RMSE, a lower MAE value
represents a more accurate model and vice versa (Chai and Draxler,
2014).

2.5.6. Implementation
The MARSmodel was implemented using R studio employing the use

of the earth package. Model selection was conducted to ensure all var-
iables were suitable for the model training. PLI variable failed to meet
the requirement for homoscedasticity after the Breusch-Pagan test was
conducted. Cu variable was also removed after a test to determine the
presence of multicollinearity was conducted. The train test split was
80:20. 80% of the data was used to train the models and 20% for testing.
Parameter optimization was conducted using grid search which runs
several iterations of the MARS models changing parameters at each
iteration to determine which combination yields the best set of results.
Three different MARS models were developed separately for three
different dependent variables. Model 1 had Nemerow Integrated Pollu-
tion Index (NIPI) as the dependent variable, model 2, degree of
contamination (Cdeg), and model 3, modified degree of contamination
(mCdeg). Vanadium (V), Chromium (Cr), Cobalt (Co), Zinc (Zn),
Strontium (Sr), Molybdenum (Mo), Barium (Ba), and Lead (Pb) were the
independent variables for all models. The main aim of this study is to
assess the environmental quality of the Nangodi region through the
identification of complex interactions that may exist within the heavy
metal concentrations in the area. MARS helps us achieve this goal
through the identification of a set of simple linear functions that
collectively results in the best predictive performance. The identification
of these complex relationships may have been particularly challenging
given the distribution of the dataset. By employing three separate MARS

Table 1
A table showing the details of Pollution indices used in this study.

Index Equation

NIPI Cf =

Ci

Cs

NIPI =

̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅

(Cf)2max + Cf2mean
2

√

Where Cf represents the concentration factor, Ci represents the measure
concentration, and Cs represents the evaluation standard (Wang et al., 2022).

PLI PLI =
̅̅̅̅̅̅̅̅̅̅̅
Cf1×n

√
Cf2..... × Cfn

Where Cf represents the concentration factor and n represents the number of heavy
metals considered (Tomlinson et al., 1980).

Cdeg Cdeg =
∑n

i=1
Ci
f

Cf represents the concentration factor and n represents the number of heavy metals
considered (Hakanson, 1980).

mCdeg mCdeg =
1
n

∑n
i=1

Ci
f

Cf represents the concentration factor and n represents the number of heavy metals
considered (Abrahim and Parker, 2008)

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

4 



models (NIPI, Cdeg, and mCdeg), we are able to identify distinct re-
lationships that may exist with regard to each dependent variable
providing us with a clearer and more interpretable set of results.

2.6. Statistical data analysis

The R software was used to perform geostatistical analysis with the
help of the “compositions” and “robcompositions” packages. Descriptive
measures, including minimum, maximum, mean, median, standard de-
viation (SD), coefficient of variation (CV%), kurtosis, and skewness,
were determined to understand the distribution of the elemental con-
centrations. For this purpose, the analysis concerned nine potential toxic
elements (PTEs): V, Cr, Co, Cu, Zn, Sr, Mo, Ba, and Pb, and their in-
terconnections with other geochemical factors, which were estimated
based on Pearson correlation coefficients (r) and their corresponding p-
values for both positive and negative relationships. PCA was used to
determine the sources of the elemental concentration in the soil samples.
PCA is not only useful in simplifying the analysis of large datasets
through the identification of the important components but also dem-
onstrates the relationships between the variables, reduces the number of
factors to consider and pinpoints the factors which have the most impact
on the quality of the soil (Kazapoe et al., 2021; Abu et al., 2021). Before
performing PCA, the variables were ranked using normal score trans-
formation to select the components with the highest signal.

3. Results

3.1. Spatial distribution of elements cross the study area

Table 2 presents the general distribution of elemental values across
the study area. Mo (144.66%), Cu (102.45%), Cr (94.74%), Sr (92.19%),
Co (87.13%), and Ba (78.50%) displayed very high CV%. Compara-
tively, Pb (43.71%) showed moderate variability. Positive skewness
exhibited by all the elements suggests that the concentrations of ele-
ments are significantly higher than their average values at most parts
across the area. Similarly, elements such as Co, Zn, Mo, and Ba record
high kurtosis values (> 3), signifying that their concentrations have
more extreme values in comparison to a normal distribution.

In terms of the general distribution of the PTEs, V ranges from 9.45-
528 mg/L with an average value of 120.86 mg/L (Fig. 2A). These values
exceed the average distribution of V set as the contaminant level when
compared to values from the European Union (2002), the United States
Environmental Protection Agency (USEPA) (2002) and other parts of
Europe. Cr averaged a concentration of 242.42 mg/L across the area
which exceeded the contaminant level reported by Kazapoe and Arhin
(2021) for the Birimian of Ghana (80.74 mg/L). This value is similarly
high as compared to the standards set by the EU (75 mg/L) and USEPA
(11 mg/L). Co distribution in the soils within the study area ranged from
3.3 to 311.8 mg/L with an average value of 30.92 mg/L. Also, 40% of the

samples fall above the contaminant threshold of 24 mg/L suggested by
Crommentuijn et al. (2000). Similarly, Cu recorded figures (3.25–249.9
mg/L) which far exceeded the average global concentration of 30 mg/L.
12% of the samples fell above the average of 100 mg/L set by
McLaughlin et al. (2000). The results of Zn varied between 9.8 and 228.2
mg/L (35.42 mg/L). The results show that 109 (19.74%) of the samples
were above the set global contaminant level Zn (30 mg/L). Sr and Mo
recorded average values of 98.17 mg/L and 0.95 mg/L respectively
which were below the average crustal values (Sr = 375 mg/L and Mo =

1.5 mg/L) reported by Taylor (1964). Ba showed values between 40.5
and 2735 mg/L. This indicates that all of the samples were above the
recommendation (9 mg/L) by Crommentuijn et al. (2000). Pb concen-
tration (Fig. 2B) ranged from 1.7 to 35.8 mg/L (Avg. 10.07 mg/L). The
results for Pb show that all the samples fell below the guideline value for
contaminants globally.

3.2. Relationships among the pollutants and trace elements

The eigenvalues and rotated sum of squares loadings from the PCA
are shown in Table 3. The first three components (PCs) explain 81.30 %
of the total variance. PC1 loadings contributed 42.12 % of the total
variance, PC2 contributed 23.39% while PC3 loadings contributed
15.79 % (Table 3). The factor analysis using R–Mode factor analysis

Table 2
Statistical summary of the results from the Nangodi area.

Elements V Cr Co Cu Zn Sr Mo Ba Pb

Min 9.45 11.6 3.3 3.25 9.8 6.9 0.4 40.5 1.7
Max 528 1294 311.8 249.9 228.2 567 14.2 2735 35.8
Mean 120.86 242.42 30.92 43.8 35.42 98.17 0.95 337.62 10.07
Standard Deviation (SD) 106.11 229.65 26.94 43.77 21.58 90.5 1.38 265.02 4.4
Skewness 1.6 1.65 3.8 1.86 2.51 2.1 5.94 2.64 1.26
Kurtosis 1.85 2.79 26.04 3.2 12.76 4.83 42.46 13.81 3.49
Coefficient of Variation (CV%) 87.79 94.74 87.13 102.45 60.93 92.19 144.66 78.50 43.71
World* ​ 100 ​ 30 50 ​ ​ ​ 42.5
EU* ​ 75 ​ ​ 1 ​ ​ ​ ​
USEPA* ​ 11 ​ 270 1100 ​ ​ ​ 200
Kazapoe and Arhin (2021) ​ 80.74 6.44 ​ ​ ​ ​ ​ 6.99
McLaughlin et al. (2000) 200 100 ​ 100 ​ ​ ​ ​ 150
Crommentuijn et al. (2000) 1.1 3.8 24 3.5 ​ ​ ​ 9 55

World (Vinogradov, 1959), EU (EU, 2002) and USEPA (USEPA, 2002).

Fig. 2. Box and whisker plots of the elemental distribution across the study
area for (A) V, Cr, Sr and Ba (B) Pb, Mo, Co, Cu and Zn.

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

5 



produced the rotated components (Table 4) depicting various compo-
nent elements of the PC. PC1 holds V, Co, Cu and Zn, PC2 holds Sr and Ba
while PC2 has Mo and Pb (Fig. 3; Table 4).

3.3. Heavy metal contamination in the Nangodi area

A summary of the results of the indices used to assess the level of
contamination or pollution in the soil samples of the study area is shown
in Table 5. The Nemerow Pollution Index reflects a mixed picture where
the majority of the samples (46.74%) fall under the Slight Polluted class
while 26.81% and 5.98% are classed in the Moderate and Heavy
Polluted categories respectively. The central part of the area around
Datoko is classed as clean under the NIPI while Shaega in the west and
other areas around the eastern margins of the study area are categorised
as moderate to heavily polluted respectively (Fig. 4A). These areas are
known centres of galamsey activities, which may explain the values
derived. The calculation for the Cdeg shows that 36.96% and 41.49%,
representing the majority of the samples, are classed as “moderate” to
“considerable” respectively, while 13.41% of the samples are classed as
“low” in terms of contamination. The results for the Cdeg further un-
derscore the effect of illegal mining on Shaega and the eastern part of the
study area which are classed as having considerable pollution (Fig. 4B).
The Pollution Load Index analysis shows that nearly half of the samples
(43.84%) indicate a progressive deterioration of the quality of the area
and the ecosystem. PLI shows that Shaega and Buin rank as showing a
progressive deterioration of the quality of the area and ecosystem
(Fig. 4C). Correspondingly, Table 6 shows that most of the samples
(39.67% and 4%) fall under the “moderate” and “high” classes respec-
tively depicting a moderate level of contamination among the soil
samples in the area in terms of the mCdeg Index. Mirroring the other
indices, the mCdeg shows that Shaega and Buin are moderately polluted
while the eastern part around Tula records some spots which are cat-
egorised as high (Fig. 4D). This is mainly due to the activities of the
illegal miners in the area A summary of the results of the indices used to
assess the level of contamination or pollution in the soil samples of the

study area is shown in Table 5. The Nemerow Pollution Index reflects a
mixed picture where the majority of the samples (46.74%) fall under the
Slight Polluted class while 26.81% and 5.98% are classed in the Mod-
erate and Heavy Polluted categories respectively. The central part of the
area around Datoko is classed as clean under the NIPI while Shaega in
the west and other areas around the eastern margins of the study area
are categorised as moderate to heavily polluted respectively (Fig. 4A).
The calculation for the Cdeg shows that 36.96% and 41.49%, repre-
senting the majority of the samples, are classed as “moderate” to
“considerable” respectively, while 13.41% of the samples are classed as
“low” in terms of contamination. The Pollution Load Index analysis
shows that nearly half of the samples (43.84%) indicate a progressive
deterioration of the quality of the area and the ecosystem. PLI shows that
Shaega and Buin rank as showing a progressive deterioration of the
quality of the area and ecosystem (Fig. 4C). Correspondingly, Table 6
shows that most of the samples (39.67% and 4%) fall under the “mod-
erate” and “high” classes respectively depicting a moderate level of
contamination among the soil samples in the area in terms of the mCdeg
Index.

Classification of the dataset was conducted using existing literature
as guidelines (Hakanson, 1980; Tomlinson et al., 1980; Abrahim and
Parker, 2008; Wang et el., 2022). Computation of the classification was
conducted using Microsoft Excel.

3.4. Machine Learning Results

Table 6 shows the performance of all three models after training and
testing, GCV, the degree of interactions, and the number of terms are
also listed. GCV is used for model validation, it assesses how well the

Table 3
The variance contribution of each principal component in the soil sample analysis, highlighting their significance in explaining the dataset’s variation.

Total Variance Explained

Component Initial Eigenvalues Extraction Sums of Squared Loadings Rotation Sums of Squared Loadings

Total % of Variance Cumulative % Total % of Variance Cumulative % Total % of Variance Cumulative %

1 3.46 43.27 43.27 3.46 43.27 43.27 3.37 42.12 42.12 ​
2 1.82 22.81 66.08 1.83 22.81 66.08 1.87 23.39 65.51 ​
3 1.22 15.22 81.30 1.22 15.22 81.30 1.26 15.79 81.30 ​
4 0.62 7.75 89.05 ​ ​ ​ ​ ​ ​ ​
5 0.47 5.82 94.87 ​ ​ ​ ​ ​ ​ ​
6 0.20 2.51 97.38 ​ ​ ​ ​ ​ ​ ​
7 0.14 1.70 99.07 ​ ​ ​ ​ ​ ​ ​
8 0.07 0.93 100.00 ​ ​ ​ ​ ​ ​ ​

Extraction Method: Principal Component Analysis.

Table 4
Rotated Component Matrix from the Principal Component Analysis (PCA) of Soil
Samples.

Component

1 2 3

V (mg/l) 0.94 0.01 0.08
Co (mg/l) 0.83 0.23 0.20
Cu (mg/l) 0.97 0.09 -0.01
Zn (mg/l) 0.83 0.31 -0.13
Sr (mg/l) -0.19 0.84 -0.23
Mo (mg/l) 0.04 0.04 0.89
Ba (mg/l) -0.23 0.88 -0.10
Pb (mg/l) -0.43 0.44 0.55

Extraction Method: Principal Component Analysis.
a. 3 components extracted.

Fig. 3. Plots of R-mode factor analysis of the PTEs.

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

6 



results from the model will generalize on unseen data (Bottegal and
Pillonetto, 2018). It achieves this by adjusting the number of terms in
the model penalizing complexity to prevent overfitting. The degree of
interactions specifies the maximum number of interactions allowed
between the predictor variables and the basis functions. The low GCV
values across all three models indicate a high level of generalizability by
the models (Oduro et al., 2015). This essentially means that the models
perform well on unseen data.

The study sought to examine the relationship between heavy metals
(Co, Cr, V, Mo, Sr, Ba, Pb, Zn) on water quality indices (NIPI, Cdeg,
mCdeg) using MARS algorithms. To this end, three MARS models were
developed one for each dependent variable. The results from Table 6
suggest that model 1 (NIPI) can explain a significant portion of the
variance specifically 96.65%. The RMSE and MAE associated with
model 1 are relatively low at 0.8227 and 0.4287 respectively, indicating
high prediction accuracy. Models 2 (Cdeg) and 3 (mCdeg) produced
similar outputs as well in terms of explanation of the variance. Model 2
explains 98.63% of the variance while model 3 explains 98.44% of the
variance. The RMSE and MAE for model 2 were slightly higher than that
of model 1, 1.0416 and 0.6181 indicating a slightly less accurate model.
Model 2 produced less RMSE and MAE as compared to the other models,
0.1225 and 0.0670 respectively making it the most accurate model.
Although model 3 produced the least errors, R2 value was only slightly
less than model 3′s. Model 1 had the lowest R2 value, it also had the least
number of interaction terms with only Cr, Co, and V being used by the
model, all other independent variables were pruned this was not the case
for models 2 and 3. The adjusted R2 values for all 3 models indicate
slight R2 change, this indicates the possibility of overfitting for all 3
models is highly unlikely. The high R2 values from all 3 models mirror
findings from existing research like Hlokoe et al. (2022) where MARS
models were used to examine the effects of biometric traits on the body
weight of Nguni cows and Zhang and Goh (2016) where MARS models
were used in the prediction of pile drivability thus supporting MARS
algorithms as a highly accurate predictive model across varying
variables.

Comparing our MARS model with other ML models used in the
prediction of pollution indices, our MARS models outperformed the ML
models used by Wang et al. (2022) which used ML models like Partial
Least Squares Regression, Support Vector Machines, and Gaussian Pro-
cess Regression in the prediction of pollution indices like NIPI and Po-
tential Ecological Risk Index (RI). This is consistent with Naser et al.
(2022) which found MARS models superior to other ML models like
Random Forest (RF) and Support Vector Machine (SVM) in the predic-
tion of compressive strength of eco-friendly concrete.

The results from Table 7 indicate that the model’s intercept which is
a representation of the baseline level of the dependent variable NIPI is
9.7502264. The coefficient associated with BF1 suggests a negative
contribution to NIPI when Cr values are less than 492.5. Therefore,
below the threshold of 492.5, as Cr decreases, NIPI is expected to
decrease. Above the same threshold, a positive coefficient is observed
specifically 0.0209554 for BF2 implying that above this threshold, an
increase in NIPI is expected as Cr increases. BF3 describes the combined
effects of V and Cr values above the thresholds of 263.7 and 492.5
respectively. A positive coefficient is observed although quite low sug-
gesting a slight increase in NIPI if both threshold conditions are met. BF4
and BF5 also indicate slight increases in NIPI upon meeting their
threshold conditions due to the low positive coefficient associated with
them.

The results from Table 8 indicate a value of 41.984806 for the
dependent variable Cdeg when all other variables are 0. BF1 and BF2
suggest a negative impact on Cdeg when V is below 67.2 and a positive
above the same threshold. Similar effects are observed for BF3 to BF10
with a negative impact being observed below their varying thresholds
and a positive one observed above their thresholds. The interaction term
associated with BF11 describes the combined effect of Cr values below
the threshold limit of 760 and that of Zn above 28. A positive coefficient
is observed suggesting a slight increase in Cdeg when both threshold
conditions are met.

From Table 9 the intercept value for mCdeg when all other terms are
0 is 4.4746867. BF1 and BF2 capture the effects of V at the threshold
limit of 67.2. Above the threshold a positive impact is observed, below
it, a negative one is observed. For BF3 and BF4 a similar effect is
observed for Cr but for the threshold of 760. A trend Is observed for the
rest of the interaction terms for BF5 to BF12. A negative impact is
observed below their various threshold limits and a positive one above.

Fig. 5 contains the subplots of each predictor variable relating to the
response variable NIPI. The black line represents the model’s prediction,
and the blue points are the observed data points. As part of its results, the
MARS model also provides data in terms of variable importance. In
terms of variable importance Co was the most influential followed by Cr,
V, Mo, Zn, and finally, Ba. Sr and Pb were unused by the model meaning
that it played no significant role in model output. This is consistent with
Table 7, the trend observed for Co and Cr is positive indicating a strong
relationship. The trend captured for Ba, Zn, and Mo is flat indicating a
weak relationship. It can be observed that the prediction for Cr was the
most accurate as well. Data points from the Cr subplot closely follow the
regression line, this effect is observed in decreasing intensity for Co, and
V indicating less prediction accuracy for the 2 variables. In terms of
feature importance, Cr was the most important variable followed by Co

Table 5
The table showing summary of the results of the indices.

Index Min Max Mean SD Skewness Kurtosis Class %

NIPI 0.55 5.69 1.70 0.75 0.86 0.80 NIPI≤0.7 – Clean
0.7<NIPI≤ 1- Warning limit
1<NIPI≤2 – Slight pollution
2<NIPI≤3 – Moderate
NIPI>3 - Heavy

8 (1.45%)
105 (19.02%)
258 (46.74%)
148 (26.81%)
33 (5.98%)

Cdeg 3.78 68.11 17.63 9.34 0.90 0.91 Cdeg < 8 - low degree
8<Cdeg<16 - moderate
16<Cdeg< 32 – considerable

74 (13.41%)
204 (36.96%)
229 (41.49%)
45 (8.15%)

PLI 0.27 2.85 0.97 0.34 1.05 2.11 <1
=1
>1

310 (56.16%)
0
242 (43.84%)

mCdeg 0.42 7.57 1.96 1.04 0.90 0.91 mCdeg<1.5(nil to very low) 1.5≤mCdeg<2(low)
2≤mCdeg<4(moderate)
4≤ mCdeg<8(high)
8≤mCdeg<16(very high)
16≤ mCdeg<32 (extremely
high)
mCdeg≥ 32 (ultra-high)

234 (42.39%)
75 (13.59%)
219 (39.67%)
24 (4.35%)
0
0
0

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

7 



and V respectively, this is expected due to the positive correlation be-
tween Cr, Co and NIPI.

Regarding feature importance for Cdeg, Co played the most impor-
tant role, followed by Cr, V, Zn, Mo, and Ba, Sr and Pb were unused by
the model. This differs from model 1 (NIPI) where Cr was the most
important variable. Another noticeable difference was the addition of
Zn, Mo, and Ba, these variables were unused by model 1, however,
although used in model 2, their input was really low given the level of
positive correlation between them and the dependent variable Cdeg.

The most important predictor variable was Co once again followed
by Cr, V, Mo, Zn, and Ba. This is expected given how close Cdeg and
mCdeg are, Sr and Pb were also unused by the model. Predictor variables
on the lower end of feature importance demonstrate weak relationships
between the dependent variable mCdeg thus Mo, Zn, and Ba did not play
a critical role in terms of model output.

Considering the Figs. 5–7, Co, Cr, and V are the most important
drivers in terms of model output. This finding differs from the findings
from Wang et al. (2022) which found Cu, Pb, and Zn to be the most
important contributors in predicting pollution indices in mining areas

within China.

4. Discussions

According to Amuah et al. (2022b) and Ahado et al. (2024), a high
coefficient of Variation (CV%), as recorded in the study area, suggests a
high degree of variability in elemental distribution and can be used to
infer the level of anthropogenic effect on elemental concentration or
distribution. The relatively moderate CV% shown by Pb suggests that
while there are fluctuations in Pb concentrations across the area, they
are not as pronounced as the other elements. The variability in the bulk
of the elements can be attributed to the gold anomalies recorded across
the area as seen in Table 2, occurring as localized hotspots both as pri-
mary sources and from contamination from illegal mining in the area.
This is corroborated by the skewness and kurtosis values of the elements.
The combination of high skewness and kurtosis values for most of the
elements depict uneven distribution patterns and localized contamina-
tion hotspots across the area, suggesting a combined effect of mineral-
ized anomalies and influences from illegal mining activities.

Fig. 4. Spatial distribution of (A) NIPI (B) Cdeg (C) PLI (D) mCdeg.

Table 6
Model performance for all three MARS models used in the study.

Model R2

(Adj)R2
RMSE MAE GCV Degree of int No. of terms

Training Testing Training Testing Training Testing

1(NIPI) 0.9812
(0.9808)

0.9665
(0.9638)

0.6339 0.8227 0.3723 0.4287 0.428 1 6

2 (Cdeg) 0.9921
(0.9920)

0.9863
(0.9852)

0.8419 1.0416 0.5702 0.6181 0.810 1 12

3 (mCdeg) 0.9918
(0.9916)

0.9844
(0.9832)

0.0958 0.1225 0.0607 0.0670 0.010 1 13

From the results provided in Table 7, the regression equation for the dependent variable NIPI is:
NIPI = 9.7502264 - 0.0191000 * max (0, 492.5 - Cr)+ 0.0209554 * max (0, Cr - 492.5)+ 0.0000481 * max (0, V - 263.7) * max (0, 492.5 - Cr) + 0.0001471 * max (0,
492.5 - Cr) * max (0, Co - 20.5) + 0.0001924 * max (0, Cr - 492.5) * max (0, Co - 16.5)

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

8 



The values recorded for V are comparable to findings by Reyes et al.
(2020) in Taltal, northern Chile which ranges between 58.00–663 mg/L.
Additionally, these values are similar to findings of Kazapoe et al. (2021)
in the Birimian of Southwestern Ghana which averaged 92.06 mg/L. The
study by Kazapoe et al. (2021) was conducted in an area with a similar
geological setting which is also populated with mining firms. These
relatively high values can be attributed to the underlying geology, which
has been described as composed of granitic intrusions high in alkaline
(Kesse, 1985; Kazapoe, 2014). The recorded values for Cr far exceeded

what was reported by Kazapoe et al. (2021) in the Southwestern part of
Ghana (17.00–331.00 mg/L). According to the World Health Organi-
zation (2000), the ingestion of crops which have high Cr concentrations
due to uptake from the soil may result in allergic eczematous and acute
irritative dermatoses, chrome ulcers, and other health complications.

The range of values identified for Pb is consistent with findings by
This was similar to the findings of Darko et al. (2019) in Gbani, Ghana
and Kazapoe et al. (2021) in Southwestern Ghana where Pb levels were
3.6–63.20 mg/L and 5.00 - 71.00 mg/L (Avg. 7.85 mg/L), respectively.
Machiwa (2010) also recorded similar mean values in the Lake Victoria
Basin of Tanzania.

In terms of the PCA; PC1, which accounts for the greatest variance
among the dataset and may therefore suggest the most dominant process
in terms of control on the distribution of heavy metals in the area, may
be linked to an external process that affects the distribution of some
metals in the soil, as indicated by the high CV% for elements like Cu
(102.45%) and V (87.79%). The source of Cu could be agrochemical
applications such as fertilizers and biocides (Lamichhane et al., 2018) or
metallurgical activities associated with gold mining (Kazapoe, 2023).
PC2 implies a geological or natural source, for instance, through the
process of weathering of rocks containing minerals. The presence of Ba
and Sr, which can be associated with fertilizers or pesticides, suggests a
mixed source with some contribution from low-level agricultural
pollution. The third factor (PC3) suggests human disturbances, espe-
cially from mining activities. The high CV% for Mo (144. 66%) and its
association with gold in the Birimian terrain may also imply that mining
waste can alter the soil chemistry. The relatively low standard deviation
of Pb (43. 71%) suggests a minor anthropogenic impact, following the
findings of Arhin et al. (2019) and Akoto et al. (2023), who link Pb to the
local gold-bearing rocks.

The results in Table 4 and Fig. 3 suggest lithological control of the
elements present in the samples, particularly those associated with mafic
rocks. The elemental associations in the main cluster of Fig. 3 addi-
tionally show the elemental assemblages typically associated with gold
mineralisation in Ghana (Kazapoe et al., 2022).

The pollution indices further indicated the pollution status of the
area. According to the Nemerow Pollution Index (NIPI), the area around

Table 7
Model 1 (NIPI) indicates the basis function, its related equations and
coefficients.

Basis function (BF) Equation Coefficients Std. Error Pr(>|t|)

​ (Intercept) 9.7502264 7.649e-02 0.000e+00
BF1 max (0, 492.5-

Cr)
-0.0191000 2.267e-04 1.579e-271

BF2 max (0, Cr-
492.5)

0.0209554 3.589e-04 5.878e-208

BF3 max (0, V-263.7)
* max (0, 492.5-
Cr)

0.0000481 3.019e-06 2.332e-45

BF4 max (0, 492.5-
Cr) * max (0, Co-
20.5)

0.0001471 5.137e-06 3.711e-102

BF5 max (0, Cr-
492.5) * max (0,
Co-16.5)

0.0001924 1.719e-05 9.839e-26

From Table 8, the regression equation for Cdeg is:
Cdeg= 41.984806 - 0.017853 * max(0, 67.2 - V)+ 0.029273 * max(0, V - 67.2) -
0.028772 * max(0, 760 - Cr)+ 0.029502 * max(0, Cr - 760) - 0.151267 * max(0,
92.1 - Co) + 0.083055 * max(0, Co - 92.1) - 0.596140 * max(0, 4.2 - Mo) +
0.661543 * max(0, Mo - 4.2) - 0.002323 * max(0, 590.5 - Ba) + 0.002591 * max
(0, Ba - 590.5) + 0.000108 * max(0, 760 - Cr) * max(0, Zn - 28).

Table 8
Results for model 2 (Cdeg) indicating the basis function, its related equations
and coefficients.

Basis Function
(BF)

Equation Coefficients Std.
Error

Pr(>|t|)

​ (Intercept) 41.984806 3.927e-
01

2.308e-
311

BF1 max (0, 67.2-V) -0.017853 3.502e-
03

5.144e-07

BF2 max (0, V-67.2) 0.029273 8.732e-
04

6.506e-
122

BF3 max (0, 760-Cr) -0.028772 2.221e-
04

0.000e+00

BF4 max (0, Cr-760) 0.029502 9.810e-
04

1.056e-
107

BF5 max (0, 92.1-Co) -0.151267 4.058e-
03

1.258e-
136

BF6 max (0, Co-92.1) 0.083055 3.773e-
03

2.029e-72

BF7 max (0, 4.2-Mo) -0.596140 6.622e-
02

7.274e-18

BF8 max (0, Mo-4.2) 0.661543 6.403e-
02

1.714e-22

BF9 max (0, 590.5-Ba) -0.002323 2.886e-
04

8.175e-15

BF10 max (0, Ba-590.5) 0.002591 3.288e-
04

2.753e-14

BF11 max (0, 760-Cr) * max
(0, Zn-28)

0.000108 6.493e-
06

1.945e-48

From Table 9, the regression equation for mCdeg is:
mCdeg = 4.4746867 - 0.0011756 * max(0, 67.2 - V) + 0.0033762 * max(0, V -
67.2) - 0.0031351 * max(0, 760 - Cr) + 0.0034487 * max(0, Cr - 760) -
0.0161624 * max(0, 92.1 - Co) + 0.0092752 * max(0, Co - 92.1) - 0.0069885 *
max(0, 38.3 - Zn) + 0.0054938 * max(0, Zn - 38.3) - 0.0785176 * max(0, 1.1 -
Mo) + 0.0680869 * max(0, Mo - 1.1) - 0.0002002 * max(0, 640.2 - Ba) +

0.0002763 * max(0, Ba - 640.2).

Table 9
Results for model 3 (mCdeg) indicating the basis function, its related equations
and coefficients.

Basis Function
(BF)

Equation Coefficients Std. Error Pr(>|t|)

​ (Intercept) 4.4746867 3.858e-
02

0.000e+00

BF1 max (0, 67.2-V) -0.0011756 4.363e-
04

7.332e-03

BF2 max (0, V-67.2) 0.0033762 9.560e-
05

6.735e-
129

BF3 max (0, 760-Cr) -0.0031351 2.506e-
05

0.000e+00

BF4 max (0, Cr-760) 0.0034487 1.094e-
04

1.129e-
113

BF5 max (0, 92.1-Co) -0.0161624 4.925e-
04

6.224e-
119

BF6 max (0, Co-92.1) 0.0092752 4.342e-
04

1.876e-69

BF7 max (0, 38.3-Zn) -0.0069885 9.805e-
04

4.378e-12

BF8 max (0, Zn-38.3) 0.0054938 4.048e-
04

3.791e-35

BF9 max (0, 1.1-Mo) -0.0785176 2.176e-
02

3.447e-04

BF10 max (0, Mo-1.1) 0.0680869 4.254e-
03

1.609e-45

BF11 max (0, 640.2-
Ba)

-0.0002002 3.575e-
05

3.801e-08

BF12 max (0, Ba-
640.2)

0.0002763 3.996e-
05

1.697e-11

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

9 



Datoko is relatively clean while Shaega in the western part and some
areas in the eastern part are moderately to highly polluted. These areas
are known centres of galamsey activities, which may explain the values
derived. The results for the Cdeg further underscore the effect of illegal
mining on Shaega and the eastern part of the study area which are
classed as having considerable pollution (Fig. 4B). Mirroring the other
indices, the mCdeg shows that Shaega and Buin are moderately polluted
while the eastern part around Tula records some spots which are cat-
egorised as high (Fig. 4D). This is mainly due to the activities of the
illegal miners in the area.

The MARS models identified several complex non-linear relation-
ships between the independent variables (V, Cr, Co, Zn, Mo, and Ba) and
the dependent variables (NIPI, Cdeg, and mCdeg). Cr, V, and Co

consistently emerged as the vital role players amongst the independent
variables with significant non-linear relationships. This differs from
findings in existing literature. Wang and Xu (2022) used ML models like
support vector machines (SVM), partial least squares (PSLR), and
gaussian process regression (GPR) in the prediction of pollution indices
like NIPI and ecological risk index (RI. Their findings found Cu, Pb, and
Zn to be the most significant contributors to their model output. The
common element between both studies is Zn, which was found to be
vital to model by Wang and Xu (2022) but not in our case.

Cr across all 3 models plays a dominant role with splits at multiple
thresholds across all 3 models suggesting a strong relationship between
the chemical element and all 3 dependent variables subject to their in-
dividual thresholds.

Fig. 5. Partial dependence plot for each predictor variable against NIPI.

Fig. 6. Partial dependence plot for each predictor variable against Cdeg (yhat = Cdeg).

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

10 



Results from Table 7 suggest V acts as a moderator to Cr enhancing
its effect on NIPI past specifics thresholds. In Tables 8 and 9 however, V
has a more active role directly impacting the dependent variables Cdeg
and mCdeg subject to their individual thresholds.

A similar effect is observed for Co where in model 1 is acts as an
amplifier to Cr’s ability to influence NIPI subject to the threshold but in
model 2 and 3 a more direct interaction between Co and the two
dependent variables Cdeg and mCdeg is observed at specific thresholds.

Mo, Ba, and Zn also play notable roles but only in models 2 and 3. No
notable relationships are observed between the 3 elements and model 1.
Mo has relationships with Cdeg and mCdeg but these interactions are
subtle at best.

Ba on the other hand although smaller in magnitude exhibits sig-
nificant threshold effect towards both Cdeg and mCdeg.

Zn has a direct although subtle interaction with mCdeg highlighting
a difference between the two dependent variables even though they
share similar features. No such direct interaction is observed in Cdeg but
rather a slight interaction between Cr and Zn for model 2.

The interactions identified by the MARS models highlight not only
interactions between the dependent variables and independent ones but
also between independent variables suggesting that changes in one el-
ement’s concentration may influence the effects of another towards the
dependent variables pointing to a highly dynamic system.

5. Conclusion

This paper examines the environmental quality of the Nangodi re-
gion in terms of the concentrations of heavy metals in the soil, the
sources and interactions between these elements, the spatial distribu-
tion, the ecological risk associated with these elements, and the effec-
tiveness of ML models as reliable predictors of soil.. Considering the
performance of the MARS algorithm in this study, the very high pre-
dictive ability observed between the variables considered both depen-
dent and independent highlights NIPI, Cdeg, and mCdeg as viable tools
for predicting soil quality. Based on the comprehensive analysis, Cr, Co,
and V are chosen as important predictors for the dependent variables;
they have high CV%, skewness, and kurtosis, which show a large range
of dispersion and non-normal distribution. This indicates scattered
hotspots, which may have been shaped by gold deposits and illicit

mining operations. This is further corroborated by the PCA which pro-
duced three principal components which cumulatively accounted for
81.30% of the total variance. PC1 (V, Co, Cu, and Zn), PC2 (Sr and Ba),
and PC3 (Mo and Pb) suggest lithological controls on the elemental
concentrations. The pollution indices reflected the state of soil
contamination at different levels. The Nemerow Pollution Index (NIPI)
indicated that the Shaega and certain parts of the eastern region of the
study area and Datoko had moderate to high levels of pollution, which is
consistent with areas with high galamsey activities. Cdeg analysis
further corroborated these findings. The Pollution Load Index (PLI) also
showed a decreasing trend in the quality of soil with 43.84% of samples
showing evidence of environmental disturbance, especially in the
Shaega, Buin and the eastern region close to Tula which are the most
influenced by mining.

Based on the above findings, the authors recommend the following
measures that if adopted, may assist in reducing the impacts of soil
contamination, enhancing public health and environmental stewardship
in the study area.

• There is a need to enhance the implementation of environmental
laws to address the challenges of galamsey in areas like Shaega, Buin,
and the eastern parts of Tula.

• There is a need to embark on immediate soil remediation measures in
the affected areas to restore soil fertility and prevent further pollu-
tion of the environment.

• Establish a consistent monitoring system to assess the soil’s condition
and detect potential contaminations. This will enable the imple-
mentation of early intervention measures and the reduction of
extreme pollution. To raise awareness among the local communities
regarding the detrimental effects of illegal mining on their health and
the environment. A routine monitoring system, as recommended in
this study, is highly feasible when implemented, as it will mitigate
the consequences of soil contamination on mining communities. A
well-organised monitoring schedule would enable the identification
of pollution surges in advance, thereby reducing the environmental
impact. It is advisable to conduct soil sampling on a bi-annual basis,
at the conclusion of the dry and rainy season. The implementation of
more stringent monitoring protocols is likely to be restricted by the
necessity of funding these exercises. These periods are significant

Fig. 7. Partial dependence plot for each predictor variable against mCdeg.

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

11 



because they can reveal varying pollution patterns, such as the
concentration of pollutants during periods of drought or precipita-
tion. In order to ensure that the results of the monitoring are acces-
sible and comprehensible to the residents whose health is being
impacted by the pollution, this routine monitoring must be supple-
mented with other aspects of data management and community
engagement.Educate and equip them with the right tools and tech-
niques to embrace sustainable mining. Encourage and sustain other
sources of income to lessen the reliance on galamsey. This can range
from agricultural projects, eco-tourism, and other income-generating
activities that are environmentally friendly.

• Further studies should be conducted to determine the effects of soil
pollution on crop yields and ecosystem services in the long run to
formulate effective measures for soil protection.

CRediT authorship contribution statement

Daniel Kwayisi: Writing – original draft, Validation, Software, Re-
sources, Methodology, Investigation, Formal analysis, Data curation,
Conceptualization. Raymond Webrah Kazapoe: Writing – original
draft, Validation, Supervision, Software, Resources, Methodology,
Investigation, Formal analysis, Data curation, Conceptualization. Seidu
Alidu:Writing – original draft, Validation, Methodology, Investigation,
Formal analysis, Data curation, Conceptualization. Samuel Dzidefo
Sagoe: Writing – original draft, Visualization, Validation, Software,
Methodology, Investigation, Data curation, Conceptualization. Aliyu
Ohiani Umaru: Writing – original draft, Visualization, Validation,
Methodology, Formal analysis, Data curation, Conceptualization. Ebe-
nezer Ebo Yahans Amuah: Writing – original draft, Visualization,
Software, Methodology, Formal analysis, Data curation, Conceptuali-
zation. Prosper Kpiebaya: Writing – original draft, Validation, Inves-
tigation, Data curation, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.

Data availability

Data will be made available on request.

Funding

This research did not receive any grant from any funding agency,
commercial or profit sectors.

Compliance with ethical standards

Conflict of interest
The authors declare that they have no competing interests.

Supplementary materials

Supplementary material associated with this article can be found, in
the online version, at doi:10.1016/j.hazadv.2024.100480.

References

Abouchami, W., Boher, M., Albarede, F., 1990. 2.1 mafic magmatism in West Africa: an
early stage of crustal accreation. In: Seventh international conference on
geochronology, cosmochronology and isotope geology.

Abrahim, G.M.S., Parker, R.J., 2008. Assessment of heavy metal enrichment factors and
the degree of contamination in marine sediments from Tamaki Estuary, Auckland,
New Zealand. Environ. Monit. Assess. 136 (1), 227–238.

Abu, M., Kalimenze, J., Mvile, B.N., Kazapoe, R.W., 2021. Sources and pollution
assessment of trace elements in soils of the central, Dodoma region, East Africa:
implication for public health monitoring. Environ. Technol. Innov. 23, 101705.

Achina-Obeng, R., Aram, S.A., 2022. Informal artisanal and small-scale gold mining
(ASGM) in Ghana: assessing environmental impacts, reasons for engagement, and
mitigation strategies. Resour. Policy. 78, 102907.

Adiguzel, M.B., Cengiz, M.A., 2023. Model selection in multivariate adaptive regressions
splines (MARS) using alternative information criteria. Heliyon. 9 (9).

Agboola, O., Babatunde, D.E., Fayomi, O.S.I., Sadiku, E.R., Popoola, P., Moropeng, L.,
Mamudu, O.A., 2020. A review on the impact of mining operation: Monitoring,
assessment and management. Results Eng. 8, 100181.

Ağyar, O., Tırınk, C., Önder, H., Şen, U., Piwczyński, D., Yavuz, E., 2022. Use of
multivariate adaptive regression splines algorithm to predict body weight from body
measurements of anatolian buffaloes in türkiye. Animals 12 (21), 2923.

Ahado, S.K., Agyeman, P.C., Borůvka, L., Kanianska, R., Nwaogu, C., 2024. Using
geostatistics and machine learning models to analyze the influence of soil nutrients
and terrain attributes on lead prediction in forest soils. Model. Earth. Syst. Environ.
10 (2), 2099–2112.

Akoto, O., Yakubu, S., Ofori, L.A., Bortey-Sam, N., Boadi, N.O., Horgah, J., Sackey, L.N.,
2023. Multivariate studies and heavy metal pollution in soil from gold mining area.
Heliyon. 9 (1).

Amosah, J., Lukman, T., 2023. From adaptation to resilience: the capability of women
smallholder farmers in The Nabdam district of the upper east region. Int. J. Manag.
Entrepr. Res. 5 (7), 483–502.

Amuah, E.E.Y., Fei-Baffoe, B., Sackey, L.N.A., Dankwa, P., Nang, D.B., Kazapoe, R.W.,
2022a. Remediation of mined soil using shea nut shell (Vitellaria paradoxa) as an
amendment material. J. Environ. Chem. Eng. 10 (6), 108598.

Amuah, E.E.Y., Fei-Baffoe, B., Sackey, L.N.A., Douti, N.B., Kazapoe, R.W., 2022b.
Understanding the distribution, source-pattern and geochemical controls of soils in
an artisanal mine site during a ban on illegal mining activities: Is a ban an absolute
solution? Soil Secur. 9, 100078.

Arhin, E., Boansi, A.O., Zango, M.S., 2016. Trace elements distributions at Datoko-Shega
artisanal mining site, northern Ghana. Environ. Geochem. Health 38, 203–218.

Arhin, E., Kazapoe, R., Zango, M.S., 2017. The Hidden Dangers of Unknowingly Ingesting
Harmful Trace Elements from Food Crops and Their Health Implications. a case
study at Talensi District in the Upper East Region, Ghana.

Arhin, E., Zhang, C., Kazapoe, R., 2019. Medical geological study of disease-causing
elements in Wassa area of Southwest Ghana. Environ. Geochem. Health 41 (6),
2859–2874.

Bansah, K.J., Yalley, A.B., Dumakor-Dupey, N., 2016. The hazardous nature of small
scale underground mining in Ghana. J. Sustain. Min. 15 (1), 8–25.

Bibri, S.E., Krogstie, J., Kaboli, A., Alahi, A., 2024. Smarter eco-cities and their leading-
edge artificial intelligence of things solutions for environmental sustainability: A
comprehensive systematic review. Environ. Sci. Ecotechnol. 19, 100330.

Bottegal, G., Pillonetto, G., 2018. The generalized cross validation filter. Automatica 90,
130–137.

Chai, T., Draxler, R.R., 2014. Root mean square error (RMSE) or mean absolute error
(MAE)?–Arguments against avoiding RMSE in the literature. Geosci. Model. Dev. 7
(3), 1247–1250.

Dzikunoo, E.A., Kazapoe, R.W., Agbetsoamedo, J.E., 2021. An integrated structural and
geophysical approach to defining the structures of part of the Nangodi greenstone
belt, northeastern Ghana. J. Afr. Earth Sci. 180, 104238.

European Union, 2002. European Union. Heavy Metals in Wastes. Eur. Commission
Environ. Retrieved from h.t.t.p://ec.e uropa.eu/environment/waste/studies/pdf/
heavy metals report.pdf.

Fister, D., Pérez-Aracil, J., Peláez-Rodríguez, C., Del Ser, J., Salcedo-Sanz, S., 2023.
Accurate long-term air temperature prediction with machine learning models and
data reduction techniques. Appl. Soft. Comput. 136, 110118.

Friedman, J.H., 1991. Multivariate adaptive regression splines. Ann. Stat. 19 (1), 1–67.
Gackowski, M., Szewczyk-Golec, K., Pluskota, R., Koba, M., Mądra-Gackowska, K.,

Woźniak, A., 2022. Application of multivariate adaptive regression splines
(MARSplines) for predicting antitumor activity of anthrapyrazole derivatives. Int. J.
Mol. Sci. 23 (9), 5132.

Ghana Statistical Services (GSS), 2010. Nabdam District. Provisional Results of the 2010
Population and Housing Census, pp. 33,826Retrieved from. https://unstats.un.org/
unsd/demographicsocial/census/documents/Ghana/Provisional_results.pdf.
Accessed July 7, 2024.

Hair Jr, J.F., Hult, G.T.M., Ringle, C.M., Sarstedt, M., Danks, N.P., Ray, S., ... & Ray, S.
(2021). Evaluation of the structural model. Partial least squares structural equation
modeling (PLS-SEM) using R: a workbook, 115–138.

Hakanson, L., 1980. An ecological risk index for aquatic pollution control. A
sedimentological approach. Water. Res. 14 (8), 975–1001.

Hlokoe, V.R., Mokoena, K., Tyasi, T.L., 2022. Using multivariate adaptive regression
splines and classification and regression tree data mining algorithms to predict body
weight of Nguni cows. J. Appl. Anim. Res. 50 (1), 534–539.

Iglesias, J., Cuesta, I., Salueña, C., Solé, J., Prevatt, D.O., Fabregat, A., 2024. Predictive
modeling of severe weather impact on individuals and populations using Machine
Learning. Int. J. Disaster Risk Reduc. 105, 104398.

Isung, C.B., 2021. The socio-economic implications of artisanal and small-scale mining
on mining communities in northern Ghana. Open Access Lib. J. 8 (03), 1.

Javed, H., Muqeet, H.A., Javed, T., Rehman, A.U., Sadiq, R., 2023. Ethical Frameworks
for Machine Learning in Sensitive Healthcare Applications. IEEE Access.

Kamm, S., Veekati, S.S., Müller, T., Jazdi, N., Weyrich, M., 2023. A survey on machine
learning based analysis of heterogeneous data in industrial automation. Comput. Ind.
149, 103930.

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

12 

https://doi.org/10.1016/j.hazadv.2024.100480
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0001
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0001
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0001
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0002
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0002
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0002
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0003
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0003
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0003
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0004
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0004
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0004
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0005
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0005
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0006
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0006
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0006
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0007
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0007
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0007
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0008
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0008
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0008
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0008
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0009
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0009
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0009
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0010
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0010
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0010
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0011
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0011
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0011
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0012
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0012
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0012
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0012
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0013
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0013
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0014
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0014
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0014
http://refhub.elsevier.com/S2772-4166(24)00081-0/optwoSyrNKmBK
http://refhub.elsevier.com/S2772-4166(24)00081-0/optwoSyrNKmBK
http://refhub.elsevier.com/S2772-4166(24)00081-0/optwoSyrNKmBK
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0015
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0015
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0016
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0016
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0016
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0017
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0017
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0018
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0018
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0018
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0019
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0019
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0019
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0020
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0020
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0020
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0021
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0021
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0021
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0022
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0023
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0023
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0023
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0023
https://unstats.un.org/unsd/demographicsocial/census/documents/Ghana/Provisional_results.pdf
https://unstats.un.org/unsd/demographicsocial/census/documents/Ghana/Provisional_results.pdf
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0026
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0026
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0027
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0027
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0027
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0028
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0028
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0028
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0029
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0029
http://refhub.elsevier.com/S2772-4166(24)00081-0/optsBKdMD95Ql
http://refhub.elsevier.com/S2772-4166(24)00081-0/optsBKdMD95Ql
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0030
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0030
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0030


Kazapoe, R.W., 2014. Geological Structural Interpretation of the Bongo and Nabdam
districts of the Bole Nangodi belt, North-Eastern Ghana, from Integrated Geophysical
and Field Mapping Datasets (MPhil Thesis). University of Ghana.

Kazapoe, R.W., 2023. Assessing the Lithium Potential of the Paleoproterozoic Rocks of
the West African Craton; the Case so Far. Geosystem Eng. 26 (6), 257–271.

Kazapoe, R.W., Amuah, E.E.Y., Dankwa, P., 2022. Sources and pollution assessment of
trace elements in soils of some selected mining areas of southwestern Ghana.
Environ. Technol. Innov. 26, 102329.

Kazapoe, R.W., Amuah, E.E.Y., Dankwa, P., Ibrahim, K., Mville, B.N., Abubakari, S.,
Bawa, N., 2021. Compositional and source patterns of potentially toxic elements
(PTEs) in soils in southwestern Ghana using robust compositional contamination
index (RCCI) and k-means cluster analysis. Environ. Challenges 5, 100248.

Kazapoe, R., Arhin, E., 2021. Determination of local background and baseline values of
elements within the soils of the Birimian Terrain of the Wassa Area of Southwest
Ghana. Geol. Ecol. Landsc. 5 (3), 199–208.

Kim, H.Y., 2018. Statistical notes for clinical researchers: simple linear regression
2–evaluation of regression line. Restor. Dent. Endod. 43 (3).

Lamichhane, J.R., Osdaghi, E., Behlau, F., Köhl, J., Jones, J.B., Aubertot, J.N., 2018.
Thirteen decades of antimicrobial copper compounds applied in agriculture. a
review. Agron. Sustain. Dev. 38 (3), 28.

Ministry of Food and Agriculture. (n.d.). Nabdam District. Retrieved from http://mofa.
gov.gh/site/sports/district-directorates/upper-east-region/269-talensi-nabdam.

Moomen, A.W., Dewan, A., 2017. Assessing the spatial relationships between mining and
land degradation: evidence from Ghana. Int. J. Min. Reclam. Environ. 31 (7),
505–518.

Murray, R.J., 1960. The geology of the Zuarungu 12◦ field sheet. Geol. Surv. Ghana, Bull.
25.

Naser, A.H., Badr, A.H., Henedy, S.N., Ostrowski, K.A., Imran, H., 2022. Application of
Multivariate Adaptive Regression Splines (MARS) approach in prediction of
compressive strength of eco-friendly concrete. Case Stud. Constr. Mater. 17, e01262.

Nunoo, S., Manu, J., Owusu-Akyaw, F.K., Nyame, F.K., 2022. Impact of artisanal small-
scale (gold and diamond) mining activities on the Offin, Oda and Pra rivers in
Southern Ghana, West Africa: a scientific response to public concern. Heliyon. 8
(12).

Oduro, S.D., Metia, S., Duc, H., Hong, G., Ha, Q.P., 2015. Multivariate adaptive
regression splines models for vehicular emission prediction. Visualiz. Eng. 3, 1–12.

Okyere, M., Ayitey, J.Z., Ajabuin, B.A., 2021. Large scale mining in Ghana: a review of
the implications on the host communities. Journal of Degraded and Mining Lands
Management 9 (1), 3193.

Sekyi-Annan, E., 2019. Performance Evaluation of Reservoir-Based Irrigation Schemes in
the Upper East Region of Ghana (Doctoral dissertation. Universitäts-und
Landesbibliothek Bonn.

Şengül, T., Çelik, Ş., Şengül, Ö., 2020. Use of Multivariate Adaptive Regression Splines
(MARS) for predicting parameters of breast meat in quails. JAPS: J. Animal Plant Sci.
30 (4).

Strielkowski, W., Vlasov, A., Selivanov, K., Muraviev, K., Shakhnov, V., 2023. Prospects
and challenges of the machine learning and data-driven methods for the predictive
analysis of power systems: a review. Energies. (Basel) 16 (10), 4025.

Tom-Dery, D., Dagben, Z.J., Cobbina, S.J., 2012. Effect of illegal small-scale mining
operations on vegetation cover of arid northern Ghana. Res. J. Environ. Earth Sci. 4
(6), 674–679.

Tomlinson, D.L., Wilson, J.G., Harris, C.R., Jeffrey, D.W., 1980. Problems in the
assessment of heavy-metal levels in estuaries and the formation of a pollution index.
Helgoländer Meeresuntersuchungen 33, 566–575.

Trunfio, T.A., Scala, A., Giglio, C., Rossi, G., Borrelli, A., Romano, M., Improta, G., 2022.
Multiple regression model to analyze the total LOS for patients undergoing
laparoscopic appendectomy. BMC. Med. Inform. Decis. Mak. 22 (1), 141.

United States Environmental Protection Agency, 2002. Supplemental Guidance for
Developing Soil Screening Levels For Superfund Sites. Office of Solid Waste and
Emergency Response, Washington, D.C. Retrieved from. http://www.epa.gov/s
uperfund/health/conmedia/soil/index.htm.

Vinogradov, A.P., 1959. The Geochemistry of Rare and Dispersed Chemical Elements in
Soils, 2nd ed. Consultants Bureau Enterprises, New York, p. 209. Revised and
enlarged.

Wang, Y., Zhao, Y., Xu, S., 2022. Application of VNIR and machine learning technologies
to predict heavy metals in soil and pollution indices in mining areas. J. Soils.
Sediments. 22 (10), 2777–2791.

Zhang, W., Goh, A.T., 2016. Multivariate adaptive regression splines and neural network
models for prediction of pile drivability. Geosci. Front. 7 (1), 45–52.

D. Kwayisi et al. Journal of Hazardous Materials Advances 16 (2024) 100480 

13 

http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0031
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0031
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0031
http://refhub.elsevier.com/S2772-4166(24)00081-0/optcZuAQecqK9
http://refhub.elsevier.com/S2772-4166(24)00081-0/optcZuAQecqK9
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0032
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0032
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0032
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0033
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0033
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0033
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0033
http://refhub.elsevier.com/S2772-4166(24)00081-0/optXyRPCUwHMn
http://refhub.elsevier.com/S2772-4166(24)00081-0/optXyRPCUwHMn
http://refhub.elsevier.com/S2772-4166(24)00081-0/optXyRPCUwHMn
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0034
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0034
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0035
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0035
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0035
http://mofa.gov.gh/site/sports/district-directorates/upper-east-region/269-talensi-nabdam
http://mofa.gov.gh/site/sports/district-directorates/upper-east-region/269-talensi-nabdam
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0037
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0037
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0037
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0038
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0038
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0039
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0039
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0039
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0040
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0040
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0040
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0040
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0041
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0041
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0042
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0042
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0042
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0043
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0043
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0043
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0044
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0044
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0044
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0045
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0045
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0045
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0046
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0046
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0046
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0047
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0047
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0047
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0048
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0048
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0048
http://www.epa.gov/superfund/health/conmedia/soil/index.htm
http://www.epa.gov/superfund/health/conmedia/soil/index.htm
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0050
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0050
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0050
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0051
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0051
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0051
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0052
http://refhub.elsevier.com/S2772-4166(24)00081-0/sbref0052

	Exploring soil pollution patterns in Ghana’s northeastern mining zone using machine learning models
	1 Introduction
	2 Methodology
	2.1 Description of the study area
	2.2 Soil sampling
	2.3 Analytical procedures
	2.4 Quality Assurance and Quality Control Analysis (QA/QC)
	2.5 Machine Learning Model
	2.5.1 Pollution indices
	2.5.2 Multivariate Adaptive Regression Splines (MARS)
	2.5.3 Evaluation metrics
	2.5.4 Root Mean Square Error (RMSE)
	2.5.5 Mean Absolute Error (MAE)
	2.5.6 Implementation

	2.6 Statistical data analysis

	3 Results
	3.1 Spatial distribution of elements cross the study area
	3.2 Relationships among the pollutants and trace elements
	3.3 Heavy metal contamination in the Nangodi area
	3.4 Machine Learning Results

	4 Discussions
	5 Conclusion
	CRediT authorship contribution statement
	Declaration of competing interest
	Data availability
	Funding
	Compliance with ethical standards
	Supplementary materials
	References