Journal of African Earth Sciences 206 (2023) 105024
Contents lists available at ScienceDirect 
Journal of African Earth Sciences 
journal homepage: www.elsevier.com/locate/jafrearsci 
Mineral prospectivity mapping over the Gomoa Area of Ghana’s southern 
Kibi-Winneba belt using support vector machine and naive bayes 
Eric Dominic Forson a,c, Prince Ofori Amponsah b,* 
a Department of Physics, School of Physical and Mathematical Sciences, University of Ghana, Ghana 
b Department of Earth Science, School of Physical and Mathematical Sciences, University of Ghana, Ghana 
c University of Tasmania, CODES - ARC Centre of Excellence in Ore Deposits and School of Earth Science, Hobart, Tasmania, 7001, Australia   
A R T I C L E  I N F O   A B S T R A C T   
Handling Editor: DR Damien Delvaux  Geospatial modeling of mineral prospective regions is essential, owing to its significant contribution towards the 
development and economic gains of many mineral-endowed countries including, Ghana. Thus, the primary 
Keywords: objective of this study is to delineate mineral potential zones in the Gomoa Area of Ghana’s southern Kibi- 
Machine learning Winneba belt in order to supplement mineral resources in Ghana’s existing mineral prospective zones. To ach-
Support vector machine ieve the aforementioned objective, researchers generated predictive models characterising gold mineralisation 
Naive bayes 
Geophysics prospects within the study area by employing machine learning techniques comprising support vector machines 
Remote sensing (SVM) and naive bayes (NB) classifiers on mineral-related conditioning factors. These mineral-related factors 
Mineral potential mapping (geoscientific thematic layers) were sourced from geophysical, remote sensing, and geological datasets. The 
resulting mineral prospective models (MPM) produced based on SVM and NB classifiers were exhibited in binary 
classes (prospective and non-prospective zones). Regions delineated as prospective zones within the study area 
were, respectively estimated to cover an area extent of 181.62 km2 and 296.02 km2 for the SVM-derived MPM 
and NB-derived MPM and analogously characterise 22.07% and 35.97% of the study area. The ability of these 
two models to predict was determined using the area under the receiver operating characteristic curve (AUC). 
The AUC scores obtained for the SVM-derived MPM and the NB-derived MPM were, respectively, 0.90 and 0.83. 
Outputs of the AUC scores generally indicate that the two models produced have good accuracy, although the 
SVM-derived MPM performed better than that of the NB-derived MPM. Thus, the machine learning-based 
mineral prospectivity models produced in this study are worthy outputs to guide the planning of detailed 
mineral exploration surveys within the study area.   
1. Introduction information systems (GIS) approaches (Carranza, 2008; Zeghouane 
et al., 2016; Zuo et al., 2016; Shabankareh and Hezarkhani, 2017; 
One of the initial stages of significant relevance in mineral pro- Amponsah and Forson, 2023). As a consequence, the goal of exploring 
specting is to localise various geologically important features associated for minerals by incorporating a variety of geoscientific and geospatial 
with the target mineral. Localising these important features is based on datasets is a multi-criteria decision-making (MCDM) task with the sole 
the availability and analysis of geological maps made up of lithological purpose of generating a predictive map comprising prospectively 
units, structures, alteration types, and locations, as well as indicator delineated zones of the said mineral (known as mineral prospectivity 
minerals. In recent years, the evolution of geological mapping methods, mapping (MPM)) (Forson et al., 2020). Additionally, the use of geo-
including geochemical analysis, coupled with the emergence of scientific information in MPM based on machine learning (ML) algo-
remotely sensed geoscientific datasets and advancements in data ana- rithms helps mineral exploration geoscientists overcome common 
lytic techniques have gained prominence, particularly in mineral challenges associated with traditional methods (where the influence or 
exploration programmes (McKay and Harris, 2016; Bachri et al., 2019; weight of various thematic layers used is determined in a subjective 
Shirmard et al., 2022). Visualisation, processing, and analysis of these manner) due to over-reliance on expert opinions (examples include 
datasets can be carried out with the support of computer and geographic PROMETHEE (Abedi et al., 2012b), the Analytical Hierarchy Process 
* Corresponding author. 
E-mail address: pamponsah@ug.edu.gh (P.O. Amponsah).  
https://doi.org/10.1016/j.jafrearsci.2023.105024 
Received 25 January 2023; Received in revised form 1 August 2023; Accepted 1 August 2023   
Available online 5 August 2023
1464-343X/© 2023 Elsevier Ltd. All rights reserved.
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
(AHP; Du et al., 2016), Fuzzy AHP (Forson et al., 2020; Amponsah et al., Hezarkhani, 2017), it was also selected as to produce a MPM in this 
2022b), and the best-worst method (Forson and Menyeh, 2023)). The study. Also, thematic layers sourced from geophysical, geological, and 
use of machine learning methods are devoid of this aforesaid subjec- remote sensing datasets were chosen for this study owing to their 
tivity, and thus, the influence of the various thematic layers used are capability in delineating alteration zones, indicator minerals (such as 
determined by assessing the spatial correlation of the thematic layers pyrite, arsenopyrite, magnetite, and chalcopyrite), as well as lineaments 
used with respect to the target labels (occurrences). Notwithstanding the associated with gold mineralisation within the study area (Klemd et al., 
difficulty associated with generating reliable models in geoscience 2002; Forson et al., 2021). It is expected that the outcome of this study 
owing to the inhomogeneity of the Earth (Zhan et al., 2023), these ML will contribute to efforts geared towards unravelling the mineral po-
algorithms also help reduce exploration costs, particularly over barren tential of the SKWB. 
or under-explored terrains, because of their capacity to classify the 
mineral prospects over an area into favourable and non-favourable 2. Study area and geological setting 
zones (Shirmard et al., 2022). Thus, machine learning techniques have 
also been found applicable by geoscientists in hydrocarbon exploration 2.1. Study area description 
(Zhang et al., 2021; Cheng and Fu, 2022), groundwater exploration (Liu 
et al., 2023; Al-Kindi and Janizadeh, 2022), and geohazard monitoring The 823 km2 study area (Fig. 1) is located in the Central Region of 
(Cao et al., 2020; Ma et al., 2022). Classification based on machine Ghana, 122 km southwest of Accra, the country’s capital, and spans 
learning algorithms generally come in two forms: supervised and un- between the Gomoa East, west, and Ekumfi districts along the Atlantic 
supervised classifications. ML-based supervised classification uses coast of the Gulf of Guinea. The study area is bounded by the coordinates 
training data containing locations of known mineral deposits and 730,000 mE, 610,000 mN and 754,000 mE, 586,000 mN within the 
non-mineral deposits and a set of thematic layers extracted from geo- World Geodetic System 1984 (WGS84) datum. 
scientific datasets (also referred to as predictors) to classify potential 
zones of mineral occurrences over an area of interest (Sun et al., 2019, 2.2. Geological setting 
2020). However, with unsupervised classification, the mineral potential 
orientation over an area is classified based on the feature statistics of According to the current 1:1,000,000 geological map published by 
each of the geoscientific thematic layers used (Zuo and Carranza, 2011). Agyei Duodu et al. (2009), the area under consideration is mainly un-
In the supervised classification of prospectively viable zones of mineral derlain by the southern portion of the Kibi-Winneba belt (Fig. 2), with 
occurrences over an area, machine learning offers several classifiers that small portions to the east and west underlain by the southern portions of 
can be employed. Application of supervised ML classifiers in carrying the Suhum and Cape-Coast basins, respectively. The Kibi-Winneba belt, 
out mineral prospectivity mapping over various geologic environments which outcrops in the study area, is part of several linear arcuate 
continues to rise in recent years owing to their ability to produce effi- greenstone belts which defines the architecture of the Birimian Paleo-
cient mineral prospectivity models that can be essential to guide future proterozoic terrane in Ghana (Leube et al., 1990; Jessell et al., 2012; 
mineral exploration activities (Abedi et al., 2012a; Zeghouane et al., Amponsah et al., 2015, 2016; Salvi et al., 2016; Feng et al., 2018, 2019; 
2016; Shabankareh and Hezarkhani, 2017; Sun et al., 2019, 2020; Parsa Asiedu et al., 2019; Sapah et al., 2021; Nunoo et al., 2022; Amponsah 
and Maghsoudi, 2021). It is worth acknowledging that each of these ML et al., 2023) formed during the Eburnean orogeny (Bonhomme, 1962; 
classifiers have their respective strengths and weaknesses and may be Baratoux et al., 2011). Rocks that crop out within the belt include am-
more or less suitable than the other in classifying the presence or phibolites, volcaniclastics, and volcanic flows with bimodal tholeiitic 
absence of a particular mineral based on the geological terrane as well as and calc-alkaline affinities (Anum et al., 2015; Forson et al., 2020). The 
the context of the exploration (Zuo and Carranza, 2011). The inference tholeiitic basalts predate the calc-alkaline series and represent an 
from this assertion by Zuo and Carranza (2011) suggests that for best immature volcanic arc setting (Ama Salah et al., 1996; Béziat et al., 
practices in MPM over a particular region of interest, the use of two or 2000; Lüdtke et al., 2000; Diatta et al., 2017). Volcanism from radio-
more ML classifiers is commendable. Thus, model comparison for min- metric dating in the Birimian terrane shows that volcanism peaked 
eral prospectivity mapping has been found applicable in other geolog- around 2190 - 2160 Ma (Davis et al., 1994). Syn-tectonically intruding 
ical terranes (Rodriguez-Galiano et al., 2015; Saberioon et al., 2018; the mafic volcanites above are the 2113 ± 1 Ma biotite granites with 
Cardoso-Fernandes et al., 2019). Archean signatures (Davis et al., 1994; Agyei Duodu et al., 2009) and the 
Of all the gold belts in Ghana, it is the southern Kibi-Winneba belt 2116 ± 2 Ma biotite (hornblende muscovite) granitoid. Unconformably 
(SKWB) where the study area falls, that is less explored in terms of its overlying the volcanites is the molassic deposit Tarkwaian sediments, 
prospectivity in spite of its geological resemblance with the Ashanti belt. 
In an attempt to unravel the mineral prospects over the belt, knowledge- 
driven (Fuzzy analytical hierarchy process (Forson et al., 2020) and 
Best-worst method (Forson and Menyeh, 2023)) methods have been 
used to delineate the mineral prospects of the area. Taking into cog-
nisance the superiority of machine learning models over these afore-
mentioned techniques employed over the area (Shirmard et al., 2022), 
the objective of this present study is to delineate the mineral prospects 
over the Gomoa Area of the SKWB using thirteen (13) thematic layers 
extracted from geophysical, geological, and remotely sensed datasets 
based on two machine-learning models comprising support vector ma-
chine (SVM) and naive bayes (NB) classifiers. The choice of the NB 
classifier as a suitable ML algorithm for generating the MPM model in 
this study is guided by its ability to perform well with multi-source 
thematic layers as well as fewer training datasets (which is the case in 
this study) and has been employed in the western part of India (Porwal 
et al., 2006) and other geological terranes (Bérubé et al., 2018; Bédard 
et al., 2022). SVM is fundamentally a two-class classifier, coupled with 
its application in several geological terranes (Zuo and Carranza, 2011; Fig. 1. Map of the Central Region of Ghana showing various administrative 
Abedi et al., 2012a; Rodriguez-Galiano et al., 2015; Shabankareh and districts (Study area is marked in red). 
2
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
lineations and symmetrical pressure shadows were produced. The 
transposition of bedding was not associated with the strain type pro-
duced during this deformation. The D2 deformational event is charac-
terized by maximum strain, which is manifested by F2 folds with 
horizontal or slightly plunging hinges, associated with a general 
east-northeast to west-southwest striking S2 cleavages and north-east to 
south-west sinistral ductile faults with reverse components. D3 is 
defined by folds generally associated with brittle shears, indicating that 
the crust was already exhumed to higher structural levels (Amponsah 
et al., 2016). 
2.3. Gold mineralisation style 
Gold mineralisation in the Kibi-Winneba belt is structurally 
controlled and occurs mainly in second-order structures, usually 10–15 
km of the main crustal-scale transcurrent shear zone. On the deposit 
scale, gold in the belt is hosted in intensely sheared and tightly folded 
volcanoclastic rocks (which are lithofacies controlled) or in chemical 
sediments usually intruded by syn-tectonic and late Eburnean age felsic, 
mafic, and intermediate dykes or as quartz-albite-carbonate-sulphide 
lodes or tensional stockworks usually developed at the margins of the 
dykes (Xtra-Gold Resource Corporation, 2021). Arsenopyrite in the 
metasediments and metavolcanics, as well as pyrites within the plutonic 
rocks, are the predominant sulphide minerals that control gold miner-
alisation in the area. These sulphides are mostly associated with regional 
greenschist facies metamorphism. The wall-rock alteration associated 
with the mineralisation zone in the metasediments are carbonitization, 
sulphidation and graphite formation through the reduction of CO2, CO 
and CH4 (Leube et al., 1990), dolomitization, and sericitization, whilst 
in the plutonic rock, k-feldspar alteration and mylonitization are key. 
Other sulphides associated with the gold mineralisation zones are pyr-
rhotite, chalcopyrite, marcasite, rutile, xenotime, bornite, and galena. 
Gold does occur as free gold within the vein selveges with the wall rocks 
or is interlocked as refractory gold within the lattices of the arsenopyrite 
and pyrites (Leube et al., 1990; Xtra-Gold Resource Corporation, 2021). 
Furthermore, hydrothermal gold mineralisation in the study area is 
mostly quartz-veins hosted in close association with disseminated 
auriferous sulphide and is therefore mostly structurally controlled. 
Fig. 2. Modified geological map of the study area (Agyei Duodu et al., 2009).  These quartz veins occur in steeply dipping shear zone contacts or 
boundaries between the meta-sediments and metavolcanics, or mostly 
composed of detrital sediments dominated by sandstones and con- as stockwork fracture-controlled quartz veins within the late plutons, 
glomerates. The granitoid complex Suhum Basin (Agyei Duodu et al., which intruded the metasediments (bsv and bmss) and metavolcanics 
2009) occurs to the east of the Kibi-Winneba belt as well as in the study (bv and bma). These shear zones act as fluid plumbing systems and 
area and is composed of four types of granitoids. Type 1 is composed of therefore control the path of gold mineralisation (Leube et al., 1990). 
biotite gneisses with subordinate biotite schist dated around 2165 ± 9 
Ma (Opare-Addo et al., 1993), while the Type 2 granitoid in the basin 3. Materials and methods 
was emplaced around 2134 ± 4 Ma (Agyei Duodu et al., 2009). Types 1 
and 2 granitoids have been intruded by both the 2106 1 Ma K-feldspar 3.1. Geoscientific thematic layers ±
rich Type 3 granitoid and the 2088 ± 1 Ma Type 4 two mica granites 
(Hirdes et al., 1992). Juxtaposing the southern Kibi-Winneba belt to the In carrying out predictive modeling for gold prospects over a 
west is the southern Cape Coast basin. The southern Cape Coast basin is designated area, the use of geoscientific-based thematic layers are 
underlain by 2187 1 Ma biotite gneisses, biotite schist and the 2116 essential because they are decisive conditional factors. In choosing ± ±
1 Ma to 2102 ± 1 Ma two mica (hornblende) granitoid intrusives. All the various geoscientific thematic layers that are to be included in a mineral 
rocks in the study area have been intruded by later Mesozoic mafic predictive modeling exercise over an area of interest, an understanding 
dolerite dykes. The tectonic evolution of the Birimian terrane has been of the mineralisation style of the said area is required. Thus, thirteen 
described by many workers as polycyclic events (Ledru et al., 1988; geoscientific-based thematic layers were derived from geophysical, 
Hirdes et al., 1992; Milési et al., 1992; Feybesse et al., 2006). This paper geological, and remote sensing datasets that had been obtained from 
will adopt the structural interpretation of Feybesse et al. (2006). Fey- reputable and publicly accessible institutions. The geophysical dataset, 
besse et al. (2006) identified three phases of deformation within the which comprises airborne magnetics and airborne radiometrics together 
Birimian terrane in Ghana. The first deformation event, termed D1 with the geological data, was obtained from the Ghana Geological 
deformation, is defined by S1 foliation, which is parallel to the axial Survey Authority. The gravity dataset used was obtained from the GFZ 
plane of F1 microfolds, and by an L1 stretching lineation. Foliations German Research Centre for Geoscience (www.gfz-potsdam.de). The 
observed during this stage differ in accordance with the intensity of main remote sensing data used (obtained from the United States 
metamorphism and the coaxial strain associated with the deformation. Geological Survey Earth Resources Observation and Science Centre) in 
Where the evidence of simple shear strain was pronounced stretching this study was based on data acquired by the Landsat 8 Operational Land 
Imager (OLI). From the geological data, the various lithologies within 
3
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
the study area were derived in ArcGIS 10.4 environment. By using the raster. Afterwards, each of these thirteen thematic layers were imported 
Geosoft Oasis Montaj software, each of the three radiometric elements into a GIS environment and were subsequently resampled to a cell res-
(potassium (K), thorium (eTh) and uranium (eU)) were gridded. After- olution of 75 m, resulting in a grid of 146,316 pixels in an R program-
wards, two radiometric channel ratios (K-eTh and eU-eTh ratios) were ming environment. To efficiently integrate these thematic layers (shown 
further generated by using the individual radioelements as inputs. In all, in Figs. 3(a-f), 4(a-f), and 5) towards the production of a mineral pro-
five layers were derived from the radiometric data. For the airborne spectivity model based on the machine learning classifiers to be used, 
magnetic data, various preprocessing and processing procedures were the original values of each of these thematic layers were normalised to 
carried out in the Geosoft Oasis Montaj, including regional-residual confine them within a range of 0–1 as shown in equation (1). 
separation to generate the residual magnetic intensity (RMI) grid. The TM − TM
study area lies in a low magnetic latitude region and thus the RMI was TM minnorm = (1) TM − TM
reduced to the magnetic pole to generate the reduction-to-pole (RTP) max min
magnetic intensity layer. By applying the analytic signal (AS) on the RMI TM is the geoscientific thematic layer that is to be normalised; TMmin 
and the first vertical derivative (FVD) filters on the RTP grid, the AS and TMmax depict, respectively, the minimum and maximum values of 
(which determines the distribution of the magnetic intensity gradient) the chosen geoscientific thematic layer (TM). TMnorm is the normalised 
and FVD (determines the rate at which the magnetic field changes in the layer, with values ranging from 0 to 1. 
vertical direction and subsequently enhances the resolution of closely It is noteworthy that the layers derived from the magnetic data are 
spaced magnetic anomalies) layers were respectively generated. The useful in mapping anomalously high magnetic regions (which could be 
lineament density (LD) layer (which comprises the occurrence of line- due to indicator minerals such as arsenopyrite, pyrite, and magnetite) 
aments such as faults, fractures, foliations, dykes, etc. with respect to a (Forson et al., 2020). The gravity layer has the capacity to delineate bulk 
given area) was also derived by applying the Centre for Exploration mineral deposits associated with sulphide minerals such as arsenopyrite, 
Targeting (CET) grid analysis technique on the RMI grid. In the case of chalcopyrite, and pyrite (Forson et al., 2021). Radiometrically derived 
the Bouguer anomaly data, regional-residual separation was carried out layers used in this study are essential in mapping out hydrothermal 
to obtain the residual gravity layer. In the case of the Landsat 8 OLI alteration zones with significant relevance towards mineralisation oc-
imagery data, an atmospheric correction was first carried out on the currences within the study area. The remote sensing-derived layers (OH 
remote sensing data to remove various effects that may have been and Fe concentration) also characterise sericite, chalcopyrite, and 
induced by the atmosphere on the reflectance values of the remote argillite associated minerals with association to gold mineralisation 
sensing images (Forson et al., 2021). Radiometric correction was, within the study area (Klemd et al., 2002; Forson et al., 2021). 
however, not applied since the acquisition of the data was at L1C 
(radiance to the sensor) without cloud cover. Afterwards, two layers of 3.2. Training and testing data 
band ratios B4/B2 and B6/B7 which respectively elucidate iron and 
hydroxyl alteration zones within the study area, were generated in QGIS Mineral deposit occurrence is a dichotomous variable that is 
software. After carrying out the various processing procedures that are expressed in terms of target labels with a value of 0 for non-deposit 
vital towards the generation of the thematic layers, the geological layer occurrence and 1 for deposit occurrence when carrying out the 
which was originally in a vector format was also transformed into a training and testing procedures (Carranza and Laborte, 2016). In this 
Fig. 3. Normalised image of (a) eTh concentration layer (b) eU/eTh ratio layer (c) eU concentration layer (d) K/eTh ratio layer (e) K concentration layer (f) Fe 
concentration layer. 
4
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
Fig. 4. Normalised image of (a) Analytic signal layer (b) First vertical derivative layer (c) Lineament density layer (d) Residual gravity layer (e) RTP magnetic 
intensity layer (f) Hydroxyl (OH) concentration layer. 
step in the aforementioned procedure that governs the selection of 
non-Au deposit sample locations dwells on the premise that the number 
of non-Au deposit locations should be the same as the number of Au 
deposit locations. This first procedure balances the negative and positive 
samples within the training and testing datasets to be employed in the 
generation of the mineral predictive model. In the second procedure, the 
locations of the non-Au deposits that would be selected should be situ-
ated such that they are distal from any known Au deposit location; this is 
because locations that are nearer to the Au deposit locations are more 
likely to possess synonymous mineralisation conditions. In this study, 
point pattern analysis was carried out to ascertain the sufficient distance 
beyond which the locations of non-Au deposit occurrence should be 
chosen owing to its usefulness in visualising and interpreting the spatial 
distribution patterns within point data (Zhang and Liu, 2019). Thus, for 
any two Au deposit occurrences, the maximum distance between them 
was statistically computed to be 1609 m using the average nearest 
neighbour analysis technique. This indicates that, for each location of 
the Au deposit, there is a 100% possibility of finding another Au 
occurrence within a 1609 m circumference around it. Non-Au deposit 
occurrences should be located beyond the 1609 m distance. It is note-
worthy that only a few locations can be chosen within the aforemen-
tioned range, and thus, a buffer distance of 1312 m, within which there 
is an 82% possibility of finding a neighbouring gold deposit close to any 
given Au deposit was selected. The third step of the aforementioned 
procedure concerns the fact that mineral occurrences within an area of 
interest are distributed in a spatially clustered manner, because they are 
Fig. 5. Normalised geological layer.  products whose occurrence is scarce and within non-random ore-form-
ing processes. Contrary to this, the locations of non-Au deposits selected 
study, 60 known locations of Au occurrence (Dove, 1991; Newmont should be spatially distributed in a random fashion as they result from 
Ghana Limited, 2006; Geodita Resources Ltd, 2007; 2012, 2013) within common geological processes. From the three steps outlined, 60 non-Au 
the Gomoa Area of the Southern Kibi-Winneba Belt were employed to deposit locations were selected randomly to attain a dichotomously 
represent the Au deposit occurrence. For the non-Au deposit occur- balanced target labels. The target labels were then split into two, rep-
rences, the location of the samples was chosen per the procedure out- resenting respectively the training data (70% of the target labels) and 
lined by Carranza et al. (2008) and Zuo and Carranza (2011). The first 
5
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
testing data (30% of the target labels). The training data was employed Table 1 
in generating the predictive models whereas the testing data was used to Parameter ranges for training the machine learning models.  
assess the performance of the predictive models generated. Machine Parameter Parameter Description Chosen 
learning Range 
3.3. Classifiers model 
Support Cost (C) Is the penalty factor due to 0.01–100 
In this study, two machine learning-based supervised classification vector misclassification error 
algorithmss have been employed, and they encompass the naive bayes machine Kernel Transforms dataset that are poly, rbf, 
and support vector machine. linearly insparable to linearly sigmoid separable data 
Gamma Outlines the degree of influence 1e-4 - 1.0 
3.3.1. Naive bayes classifier associated with each training 
The naive bayes classifier fundamentally dwells on the likelihood of example 
occurrence of a particular event; in this present study, it can analogously Naive bayes priors Prior probabilities of the classes None 
var_smoothing Portion of the largest variance of 1e-09 - 1  
be associated with the probability of occurrence of a mineral deposit all features that is added to 
over the study area, with prior knowledge learned from the training data variances for calculation stability 
(Bédard et al., 2022). In delineating mineral prospective zones based on 
the naive bayes classifier, class probabilities for each target (occurrence 
or non-occurrence) are calculated, and the observation is classified Table 2 
based on the class with the highest probability. Although this classifier Optimum parameters for training the machine learning models using the Grid 
primarily functions based on the assumption of independence among the Search CV.  
geospatial layers (predictors), it still exhibits good predictive ability Machine Learning Model Parameter Chosen Optimal Parameter Value 
even in instances where the predictors violate the independence rule 
(Bédard et al., 2022). Implementation of the naive nayes classifier for Support vector machine Cost (C) 10 Kernel rbf 
mineral prospectivity mapping in our present study was carried out in Gamma 0.1 
the Python programming language. Naive bayes priors None 
var_smoothing 0.00231012970008316  
3.3.2. Support vector machine classifier 
The support vector machine (SVM) is a supervised machine learning- training each of the two machine learning models incorporated in this 
based classifier essential for maximising the capacity of various math- study. The classification performance of each of the machine learning 
ematical functions with a particular set of datasets (Noble, 2006; Li and models was assessed by implementing a 10-fold cross-validation pro-
Sun, 2020). This classifier, which was originally proposed by Vapnik cedure which employs their respective optimal parameters attained 
(1999), has in recent years been employed successfully in mineral pro- based on the grid search cv approach. 
specting in many areas based on remote sensing, geophysical, and other To assess the importance of each thematic layer towards the MPM 
mineral-related geoscientific datasets (Abedi et al., 2012a; Rodri- models to be generated based on the SVM and NB, the permutation 
guez-Galiano et al., 2015; Chen and Wu, 2017; Shabankareh and feature importance was carried out using the testing data. For a given 
Hezarkhani, 2017). Supervised classification based on the SVM classifier number of layers, the permutation importance tool, which is hosted in 
was carried out in Python based on kernel functions comprising linear, the scikit-learn library, computes the feature importance of each layer as 
polynomial, radial basis function (rbf) and sigmoid. The fundamental expressed in equation (2) (Pedregosa et al., 2011). 
principle which underlies the SVM classifier can be found in many recent 
works of literature (Rodriguez-Galiano et al., 2015; De Boissieu et al., 1 ∑K
2018; Cardoso-Fernandes et al., 2020). In this study, the rbf kernel ij = s − sk,j (2)  K k=1
function with a penalty parameter and gamma values of 10 and 0.1, 
respectively, were identified as the most efficient parameters for the Where ij is the computed permutation importance for each layer, s and 
classification. This penalty parameter was essential, particularly in sk,j represent the scores of the model, K is a the number of repetitions 
dealing with non-separable classes during the classification. observed and k is a repetitive value within a number of repetitions. 
In predictive modeling, the efficacy of the natural resource 
3.4. Training and evaluation of the models (groundwater or mineral) to be predicted ought to be evaluated to build 
confidence in the outputs generated (Amponsah et al., 2022a). For this 
In this study, the generation of input data, which comprises the reason, the mineral prospectivity models generated based on the support 
thematic layers and the Au target labels were succeeded by the pro- vector machine and naive bayes classifiers were assessed by using the 
duction of SVM and NB-based machine learning models through a receiver operating characteristics (ROC) curve to determine the spatial 
training process. The training process primarily deals with determining correlation between the mineral-based predictive models produced and 
the essential parameters for the machine learning models to be gener- Au mineralisation occurrence within the study area by using the testing 
ated. During data-driven modeling, specifying parameters for a priori data. On a typical ROC curve, the abscissa characterises the false positive 
fitting configuration is very strenuous because the attainment of optimal rate (FPR) (analogous to 1 specificity) as expressed in equation (3) (Sun 
parameters for various machine learning models differ with respect to et al., 2020). Along the ordinate is the true positive rate (TPR), which is 
the input dataset used. This presupposes that the determination of these analogously referred to as sensitivity and is expressed in equation (3). A 
aforementioned parameters does not rely on any empirical rule that is determination of the spatial relationship between the predictive models 
universal in nature, and thus, a less subjective trial-and-error procedure generated and the testing datasets was carried out based on the area 
is required to invariably attain optimal parameters. In this study, the under the ROC curves (AUC) obtained. Higher values of AUC depict a 
trial-and-error procedure was carried out by applying the grid search strong spatial association between the predictive models and the testing 
cross-validation (cv) technique with reference to previous studies target labels within the study area. 
(Rodriguez-Galiano et al., 2015; Bérubé et al., 2018; Cardoso-Fernandes 
et al., 2020; Pham et al., 2021; Bédard et al., 2022), as summarized in FPFPR= (3) 
Table 1. By carrying out this technique, the optimal parameters that TN + FP
were attained (which have been summarized in Table 2) were used in 
6
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
TP concentration predictors, as shown in Fig. 6(b). In the SVM-based MPM, 
TPR= (4) 
FN + TP the eTh concentration predictor is the feature with the least contribution 
towards the predictive model generated. For the NB-based MPM (Fig. 7 
From equations (3) and (4), FP, TN, TP, and FN characterises, (b)), it is the geology feature that made the optimum contribution to-
respectively, the false positives, true negatives, true positives, and false wards the predictive model generated, followed by RG, RTP, LD, K/eTh, 
negatives. Furthermore, model evaluation indices comprising precision AS, eU concentration, OH concentration, eU/eTh ratio, Fe concentra-
(a quality determinant of positive predictions made by a model is tion, FVD, K concentration, and eTh concentration features, as shown in 
expressed in equation (5)), recall (shown in equation (6) represents the Fig. 6. For both mineral prospectivity models produced, thematic fea-
fraction of data samples that are correctly identified by a predictive tures comprising AS, K/eTh, geology, OH, lineament density, residual 
model for a given class), accuracy (expressed in equation (7) depicts the gravity, uranium concentration, and RTP were generally observed to 
percentage of correct predictions made by a model), and F1 score (which impose enormous influence on the predictive models. This corroborates 
characterises a predictive model’s accuracy and it is expressed in the literature assertion that the mineralisation within the study area is 
equation (8)) were also used to evaluate the performance of the mineral associated with magnetically high indicator minerals such as magnetite 
prospectivity models generated. and thus is influenced by the RTP and analytic signal magnetic features 
TP (Forson et al., 2021). These magnetite ores often contain sulphide 
Precision= (5)  
TP+ FP minerals comprising arsenopyrite, pyrite, and chalcopyrite. In the upper 
greenschist metamorphic zones within the granitoids, pyrite to pyrrho-
TP
Recall (6)  tite as well as amphibole-bearing minerals exhibit high magnetic re-=
TP+ FN sponses, making the contribution of the RTP and AS very essential to the 
outputs generated (Perrouty et al., 2012; Forson et al., 2022b). Also, the 
TP+ TN
Accuracy= (7)  residual gravity feature’s influence could be due to the premise that 
TP+ TN + FP+ FN these sulphide indicator minerals, comprising arsenopyrite, chalcopy-
[ ]
Precision x Recall rite, pyrite, and pyrrhotite are iron-predominated and are most likely to 
F1 Score= 2 x (8)  
Precision Recall be associated with high specific gravity (Klemd et al., 2002; Forson +
et al., 2021). The significant influence observed by the eU concentration 
4. Results and discussion affirms the literature findings that anomalously high regions of uranium concentration characterise regions where gold mineralisation is likely to 
4.1. Relevance of the mineral potential factors occur (Forson et al., 2021). The potassium-thorium (K/eTh) layer, which was also observed to have a significant influence in the models gener-
The predictive modeling results obtained based on the use of the ated, could be due to potassium-thorium antagonism (due to a rise in 
support vector machine and the naive bayes classifiers were interpreted potassium and a decline in thorium concentration), resulting in the 
by carrying out a quantitative computation of the relative importance of occurrence of alteration zones that are essential targets for mineralisa-
each of the mineral potential factors (Fig. 3(a), (b), 3(c), 3(d), 3(e), 3(f), tion occurrence (Wemegah et al., 2015; Forson et al., 2021). The sig-
4(a), 4(b), 4(c), 4(d), 4(e), 4(f) and 5) towards the generation of the nificant contribution by the lineament density feature towards the 
mineral prospectivity models. By carrying out permutation feature predictive models is also expected because, in the study area, quartz vein 
importance on each of the two predictive models generated, the mineralisations have been association with geological structures such as 
contribution and influence of each of the thematic features towards the faults, folds and dykes (Klemd et al., 2002; Forson et al., 2021). The 
MPMs generated were determined (shown in Fig. 6). In the case of the relevance of geology in mineralisation occurrences cannot be under-
mineral prospectivity model produced based on the SVM classifier estimated, and thus the observation that it is highly significant towards 
(Fig. 7(a)), the lineament density and K/eTh ratio predictors make the the SVM-based MPM and the NB-based MPM is expected. Within the 
most contribution towards the SVM-based MPM (shown in Fig. 6). This study area, hydroxyl-bearing minerals such as illite, goethite, limonite, 
is orderly preceded by residual gravity, analytic signal, geology, RTP, and sericite are observed within adjacently lying country rocks with 
eU/eTh ratio, eU concentration, first vertical derivative (FVD), hydroxyl carbonate-chalcopyrite-arsenopyrite-gold-tourmaline-sericite accumu-
(OH) concentration, iron (Fe) concentration, and potassium (K) lations (Klemd et al., 2002; Dzigbodi-Adjimah, 2004). Thus, the signif-icant contribution of the hydroxyl (OH) feature towards the predictive 
models generated is expected. In both predictive models generated, the 
potassium concentration feature was observed to exhibit low signifi-
cance. This corroborates with a study by Forson et al. (2022b), which 
observed the potassium concentration feature as a non-indicator feature 
to gold mineralisation over the southern Kibi-Winneba belt. The rele-
vance of these thematic features towards gold mineralisation prospects 
may facilitate future gold prospecting and provide insights into the 
model that characterises gold mineralisation within the study area. 
4.2. Mineral prospectivity models (MPMs) 
To assist in the production of a predictive model that delineates 
various regions within the study area that are prospective and non- 
prospective to gold mineralisation, the support vector machine and 
the naive bayes classifiers were employed on the thirteen thematic 
features derived from geological, remote sensing, and geophysical 
layers. Output for each of the predictive models was discretised into two 
classes comprising prospective (regions with mineral potential capacity) 
and non-prospective (regions deemed to exhibit minimal or no potential 
Fig. 6. Feature importance of thematic layers towards SVM-based MPM and of the sought-after mineral deposit) areas. For the mineral prospectivity 
NB-based MPM produced. model produced based on the SVM classifier (shown in Fig. 7(b)), 
7
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
Fig. 7. Mineral prospectivity models based on (a) support vector Machine classifier (b) naive bayes classifier.  
regions delineated as prospective cover an areal extent of 181.62 km2 
(analogous to 22.07% of the total size of the study area), whereas the 
non-prospective class was observed to cover an area of 641.41 km2, 
representing 77.93% of the total study area size. The prospective regions 
within the SVM-based MPM were predominantly observed within the 
southeastern part of the study area. With respect to the naive bayes- 
based mineral prospectivity model (Fig. 7(a)), 35.97% (representing 
296.02 km2) of the total study area extent has been delineated to 
comprise the prospective class, whereas the remaining portion of the 
study area (made up of 527.01 km2) characterises regions delineated as 
non-prospective to gold mineralisation. It can be observed from the NB- 
based MPM that the prospective class generally characterises the 
eastern, western, southeastern and the central portions of the study area. 
In general, the southeastern portion, which was delineated as prospec-
tive on both predictive models, falls within the Birimian metavolcanics 
terrane, known to be associated with high prospects for gold minerali-
sation occurrence within the study area (Klemd et al., 2002). 
Fig. 8. Receiver operating characteristics (ROC) curve for the predic-
4.3. Evaluation of the mineral prospectivity models tive models. 
To make the predictive models generated essentially reliable for any Table 3 
further deductions, interpretations, and usage by various geoscientists, Predictive performance of machine learning-based mineral prospectivity 
they ought to be evaluated to determine their efficacy and accuracy. models.  
Evaluating predictive models makes them worthy for decision-making Evaluation Metrics SVM-based MPM NB-based MPM 
and builds confidence in users of those models (Forson et al., 2022a). 
In this regard, the predictive efficacy of the mineral prospectivity models Precision 83.33% 81.25% 
Recall 83.33% 72.20% 
generated based on the SVM and NB classifiers was evaluated using the Accuracy 83.33% 76.50% 
receiver operating characteristics (ROC) curve. The scores obtained for F1 Score 83.33% 77.80%  
the area under the ROC curve (AUC) for the MPM results produced by 
the SVM classifier and NB classifier are, respectively, 0.90 and 0.83 (as 
shown in Fig. 8). Results obtained for the AUC scores explicitly indicate complementing efforts geared towards creating sustainable jobs for the 
that the MPM results produced by the support vector machine classifier unemployed, improving life expectancy, and providing essential infra-
perform better than the predictive model generated by the naive bayes structure in the health and education sectors (Walser, 2002; Hossein-
classifier. This analogously stipulates that, the mineral prospectivity pour et al., 2022). The delineation of prospectively new areas of 
models created by employing the SVM classifier obtained the highest mineralisation occurrences (known as mineral prospectivity mapping) is 
accuracy while predicting the prospective zones within the study area achieved by collecting, analysing, and synthesising several layers 
that possess gold mineralisation occurrences. Results obtained for the sourced from geological, geophysical, and remote sensing datasets over 
performance metrics (precision, recall, accuracy, and the f1 score) a region of interest. Synthesising layers for MPM has been realised 
indicate a generally enhanced performance for the SVM-based MPM through the use of heuristic or knowledge driven (Abedi et al., 2012b; 
with respect to the NB-based MPM, as shown in Table 3. Du et al., 2016; Amponsah et al., 2022b; Forson and Menyeh, 2023), 
bivariate (Zhang et al., 2014; Harris et al., 2015), and multivariate 
(Xiong et al., 2018; Xu et al., 2021) methods. 
4.4. Discussion In this study, machine learning (multivariate) methods comprising 
the SVM and the NB were adopted for the integration of thirteen 
In many countries, the mining sector has been an essential contrib- geoscientifically-sourced thematic layers for the onward generation of 
utor towards their economic gains. Thus, the delineation of prospec- predictive models that depict regions of potential mineral occurrences 
tively new regions of mineral occurrence is very important in 
8
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
over the Gomoa Area of Ghana’s Southern Kibi-Winneba belt based on a interests or personal relationships that could have appeared to influence 
number of reasons: (a) machine learning methods are statistically the work reported in this paper. 
objective, reproducible, and are able to analyse the influence of the 
thematic layers towards mineralisation occurrence in a quantitative Data availability 
manner; (b) machine learning-based predictive modeling is less time 
consuming and relatively easy to carry out in comparison with the Data will be made available on request. 
conventional methods (heuristic and bivariate methods) (Youssef and 
Pourghasemi, 2021); (c) the accuracy of machine learning-based pre- Acknowledgements 
dictive models generated are better than the conventional methods; and 
(d) the use of machine learning models in mineral prospectivity map- Authors are grateful to the University of Ghana-Carnegie Corpora-
ping within Ghana and West Africa is rare in literature. It is noteworthy tion and Building a New Generation Africa (BaNGA-Africa) for their 
that, the SVM and NB are not without defects when employed in pre- immense support in making this study a success. Authors also wish to 
dictive modeling. In the case of the SVM classifier, its execution capa- thank the United States Geological Survey Earth Resources Observation 
bility is flawed in situations where the target classes overlap. SVM would and Science Centre, Geodita Resources Limited, and GFZ German 
also underperform in a situation where a large amount of data is Research Centre for Geoscience (Potsdam-Germany) for making data 
involved. It also performs poorly when there is an imbalance between available for use in this study. Authors are also thankful for the three 
the training datasets (Deng et al., 2017). The reliability of the Naive anonymous reviewers for reviewing our paper. Their comments were 
Bayes classifier is greatly affected in a situation where the distribution of vital to improving the manuscript. 
the thematic layers and the training datasets differ significantly, and 
thus its functionality requires the attributes (layers) to be independent References 
(Tien Bui et al., 2012). This makes the NB classifier susceptible to the 
embedment of elicit subtle patterns in geoscientific datasets (Naghibi Abedi, M., Norouzi, G.-H., Bahroudi, A., 2012a. Support vector machine for multi- 
et al., 2017). In view of this, this study was further set out to compare the classification of mineral prospectivity areas. Comput. Geosci. 46, 272–283. Abedi, M., Torabi, S.A., Norouzi, G.-H., Hamzeh, M., Elyasi, G.-R., 2012b. Promethee ii: a 
results of MPMs produced based on the SVM and NB as well as their knowledge-driven method for copper exploration. Comput. Geosci. 46, 255–263. 
predictive performance. Results obtained for the model evaluation Agyei Duodu, J., Loh, G., Boamah, K., Baba, M., Hirdes, W., Toloczyki, M., Davis, D., 
metrics comprising AUC score, precision, recall, accuracy, and F1 score 2009. Geological Map of Ghana 1: 1 000 000. Geological Survey Department. 
Al-Kindi, K.M., Janizadeh, S., 2022. Machine learning and hyperparameters algorithms 
(Table 3 and Fig. 8) indicate that, although both the SVM and the NB for identifying groundwater aflaj potential mapping in semi-arid ecosystems using 
models showed good performance, the SVM-based MPM was observed lidar, sentinel 2, GIS data, and analysis. Rem. Sens. 14 (21), 5425. 
to exhibit an improved performance over the NB-based MPM. This Ama Salah, I., Liégeois, J.-P., Pouclet, A., 1996. Evolution d’un arc insulaire océanique 
further suggests that the delineated prospective zones (Fig. 5(a) and (b)) birimien précoce au liptako nogérien (sirba): géologie, géochronologie et géochimie. J. Afr. Earth Sci. 22, 235–254. 
generally contained a high proportion of the number of known Au oc- Amponsah, P.O., Kwayisi, D., Awunyo, E.K., Sapah, M.S., Sakyi, P.A., Su, B.X., Lu, Y., 
currences within the target labels, whereas the non-prospective zones Nude, P.M., 2023. New evidence for crustal reworking and juvenile arc-magmatism 
were characterized by a greater proportion of non-Au occurrences. during the Palaeoproterozoic Eburnean events in the Suhum Basin, South-east Ghana. Geol. J. In Press. 
Amponsah, P.O., Salvi, S., Béziat, D., Siebenaller, L., Baratoux, L., Jessell, M.W., 2015. 
5. Conclusion Geology and geochemistry of the shear-hosted julie gold deposit, nw Ghana. J. Afr. 
Earth Sci. 112, 505–523. 
Amponsah, P.O., Salvi, S., Didier, B., Baratoux, L., Siebenaller, L., Jessell, M., Nude, P.M., 
As a preliminary step in mineral exploration programmes, it is Gyawu, E.A., 2016. Multistage gold mineralization in the wa-lawra greenstone belt, 
desirable to efficiently evaluate the mineral prospect over an area of nw Ghana: the bepkong deposit. J. Afr. Earth Sci. 120, 220–237. 
interest for its sufficient availability and optimum utilisation by adopt- Amponsah, P.O., Forson, E.D., 2023. Geospatial modeling of mineral potential zones using data-driven based weighting factor and statistical index techniques. J. Afr. 
ing efficient machine learning approaches. Thus, in this study, two Earth Sci., 105020 
machine learning classifiers comprising support vector machines and Amponsah, P.O., Forson, E.D., Sungzie, P.S., Loh, Y.S.A., 2022a. Groundwater 
naive bayes have been employed on thirteen input layers derived from Prospectivity Modeling over the Akatsi Districts in the Volta Region of ghana Using 
the Frequency Ratio Technique. Modeling Earth Systems and Environment, pp. 1–19. 
geophysical, remote sensing, and geological datasets over the Gomoa Amponsah, T.Y., Danuor, S.K., Wemegah, D.D., Forson, E.D., 2022b. Groundwater 
Area of the Kibi-Winneba belt of Ghana to generate models that classify potential characterisation over the voltaian basin using geophysical, geological, 
the various mineral prospective zones within the study area. The results hydrological and topographical datasets. J. Afr. Earth Sci. 192, 104558. 
Anum, S., Sakyi, P.A., Su, B.-X., Nude, P.M., Nyame, F., Asiedu, D., Kwayisi, D., 2015. 
obtained for the mineral prospectivity model produced based on the Geochemistry and geochronology of granitoids in the kibi-asamankese area of the 
naive nayes classifier indicate that 35.97% of the total study area size kibi-winneba volcanic belt, southern Ghana. J. Afr. Earth Sci. 102, 166–179. 
(representing an areal extent of 296.02 km2) was delineated as pro- Asiedu, D.K., Agoe, M., Amponsah, P.O., Nude, P.M., Anani, C.Y., 2019. Geochemical 
spective, whereas in the case of the SVM-based mineral prospectivity constraints on provenance and source area weathering of metasedimentary rocks 
2 from the paleoproterozoic (~ 2.1 ga) wa-lawra belt, southeastern margin of the west model, 22.07% of the study area (with an area size of 181.62 km ) was african craton. Geodin. Acta 31 (1), 27–39. 
deemed prospective. Assessing the performance of a predictive model Bachri, I., Hakdaoui, M., Raji, M., Teodoro, A.C., Benbouziane, A., 2019. Machine 
created is essential, as it emphasises confidence in the output produced. learning algorithms for automatic lithological mapping using remote sensing data: a case study from souk arbaa sahel, sidi ifni inlier, western anti-atlas, Morocco. ISPRS 
Hence, the area under the receiver operating characteristics curve, Int. J. Geo-Inf. 8 (6), 248. 
which is a standard statistical evaluation technique, was employed on Baratoux, L., Metelka, V., Naba, S., Jessell, M.W., Grégoire, M., Ganne, J., 2011. Juvenile 
the two MPMs generated. AUC scores obtained for the predictive models Paleoproterozoic crust evolution during the eburnean orogeny (2.2–2.0 ga), western Burkina Faso. Precambrian Res. 191 (1–2), 18–45. 
in this study indicate that the SVM-based MPM and NB-based MPM Bédard, É., de Vazelhes, V.D.B., Beaudoin, G., 2022. Performance of predictive 
produced showed good prediction ability, with AUC scores of 0.90 and supervised classification models of trace elements in magnetite for mineral 
0.83, respectively. However, the SVM-based MPM is the best at accurate exploration. J. Geochem. Explor. 236, 106959. 
Bérubé, C.L., Olivo, G.R., Chouteau, M., Perrouty, S., Shamsipour, P., Enkin, R.J., 
prediction of mineral prospectivity zones within the study area. Thus, Morris, W.A., Feltrin, L., Thiémonge, R., 2018. Predicting rock type and detecting 
the use of the support vector machine and the naive bayes classifiers can hydrothermal alteration using machine learning and petrophysical properties of the 
be applied in other geological terranes for the accurate prediction of canadian malartic ore and host rocks, pontiac subprovince, québec, Canada. Ore 
Geol. Rev. 96, 130–145. 
mineral prospective zones. Béziat, D., Bourges, F., Debat, P., Lompo, M., Martin, F., Tollon, F., 2000. 
A paleoproterozoic ultramafic-mafic assemblage and associated volcanic rocks of the 
Declaration of competing interest boromo greenstone belt: fractionates originating from island-arc volcanic activity in 
the west african craton. Precambrian Res. 101 (1), 25–47. 
Bonhomme, M., 1962. Contribution à l’étude géochronologique de la plate-forme de 
The authors declare that they have no known competing financial l’Ouest Africain. Imprimerie Louis-Jean. 
9
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
Cao, J., Zhang, Z., Du, J., Zhang, L., Song, Y., Sun, G., 2020. Multi-geohazards Hosseinpour, M., Osanloo, M., Azimi, Y., 2022. Evaluation of positive and negative 
susceptibility mapping based on machine learning?a case study in jiuzhaigou, China. impacts of mining on sustainable development by a semi-quantitative method. 
Nat. Hazards 102, 851–871. J. Clean. Prod. 366, 132955. 
Cardoso-Fernandes, J., Teodoro, A., Lima, A., Roda-Robles, E., 2019. Evaluating the Jessell, M.W., Amponsah, P.O., Baratoux, L., Asiedu, D.K., Loh, G.K., Ganne, J., 2012. 
performance of support vector machines (svms) and random forest (rf) in li- Crustal-scale transcurrent shearing in the paleoproterozoic Sefwi-Sunyani-Comoe 
pegmatite mapping: preliminary results. In: Earth Resources and Environmental region, West Africa. Precambrian Res. 212, 155–168. 
Remote Sensing/GIS Applications X, vol. 11156. SPIE, pp. 146–157. Klemd, R., Hünken, U., Olesch, M., 2002. Metamorphism of the country rocks hosting 
Cardoso-Fernandes, J., Teodoro, A.C., Lima, A., Roda-Robles, E., 2020. Semi- gold–sulfide-bearing quartz veins in the paleoproterozoic southern kibi-winneba belt 
automatization of support vector machines to map lithium (li) bearing pegmatites. (se-Ghana). J. Afr. Earth Sci. 35 (2), 199–211. 
Rem. Sens. 12 (14), 2319. Ledru, P., Milési, J., Vinchon, C., Ankrah, P., Johan, V., Marcoux, E., 1988. Geology of 
Carranza, E.J.M., 2008. Geochemical Anomaly and Mineral Prospectivity Mapping in the birimian series of Ghana. In: International Conference and Workshop on the 
GIS. Elsevier. Geology of Ghana with Special Emphasis on Gold Programme and Abstracts. Accra, 
Carranza, E.J.M., Laborte, A.G., 2016. Data-driven predictive modeling of mineral Ghana, pp. 26–27. 
prospectivity using random forests: a case study in catanduanes island (Philippines). Leube, A., Hirdes, W., Mauer, R., Kesse, G.O., 1990. The early proterozoic birimian 
Nat. Resour. Res. 25 (1), 35–50. supergroup of Ghana and some aspects of its associated gold mineralization. 
Carranza, E., Hale, M., Faassen, C., 2008. Selection of coherent deposit-type locations Precambrian Res. 46 (1–2), 139–165. 
and their application in data-driven mineral prospectivity mapping. Ore Geol. Rev. Li, X., Sun, Y., 2020. Stock intelligent investment strategy based on support vector 
33 (3–4), 536–558. machine parameter optimization algorithm. Neural Comput. Appl. 32, 1765–1775. 
Chen, Y., Wu, W., 2017. Application of one-class support vector machine to quickly Liu, Z., Xu, J., Liu, M., Yin, Z., Liu, X., Yin, L., Zheng, W., 2023. Remote Sensing and 
identify multivariate anomalies from geochemical exploration data. Geochem. Geostatistics in Urban Water-Resource Monitoring: A Review. Marine and 
Explor. Environ. Anal. 17 (3), 231–238. Freshwater Research. 
Cheng, Y., Fu, L.-Y., 2022. Nonlinear seismic inversion by physics-informed caianiello Lüdtke, G., Hirdes, W., Konan, K., Koné, Y., N’Da, D., Traoré, Y., Zamblé, Z., de la 
convolutional neural networks for overpressure prediction of source rocks in the o Géologie, D., et al., 2000. Géologie de la region haute comoe sud/geology of the 
shore xihu depression, east China. J. Petrol. Sci. Eng. 215, 110654. haute-comoe south area. 
Davis, D., Hirdes, W., Schaltegger, U., Nunoo, E., 1994. U-pb age constraints on Ma, J., Xia, D., Wang, Y., Niu, X., Jiang, S., Liu, Z., Guo, H., 2022. A comprehensive 
deposition and provenance of birimian and gold-bearing tarkwaian sediments in comparison among metaheuristics (mhs) for geohazard modeling using machine 
Ghana, west africa. Precambrian Res. 67 (1–2), 89–107. learning: insights from a case study of landslide displacement prediction. Eng. Appl. 
De Boissieu, F., Sevin, B., Cudahy, T., Mangeas, M., Chevrel, S., Ong, C., Rodger, A., Artif. Intell. 114, 105150. 
Maurizot, P., Laukamp, C., Lau, I., et al., 2018. Regolith-geology mapping with McKay, G., Harris, J., 2016. Comparison of the data-driven random forests model and a 
support vector machine: a case study over weathered Ni-bearing peridotites, New knowledge-driven method for mineral prospectivity mapping: a case study for gold 
Caledonia. Int. J. Appl. Earth Obs. Geoinf. 64, 377–385. deposits around the huritz group and nueltin suite, nunavut, Canada. Nat. Resour. 
Deng, C., Pan, H., Fang, S., Konaté, A.A., Qin, R., 2017. Support vector machine as an Res. 25 (2), 125–143. 
alternative method for lithology classification of crystalline rocks. J. Geophys. Eng. Milési, J.-P., Ledru, P., Feybesse, J.-L., Dommanget, A., Marcoux, E., 1992. Early 
14 (2), 341–349. proterozoic ore deposits and tectonics of the birimian orogenic belt, west Africa. 
Diatta, F., Ndiaye, P.M., Diène, M., Amponsah, P.O., Ganne, J., 2017. The structural Precambrian Res. 58 (1–4), 305–344. 
evolution of the Dialé-Daléma basin, Kédougou-Kéniéba Inlier, eastern Senegal. Naghibi, S.A., Moghaddam, D.D., Kalantar, B., Pradhan, B., Kisi, O., 2017. A comparative 
J. Afr. Earth Sci. 129, 923–933. assessment of GIS-based data mining models and a novel ensemble model in 
Dove, K., 1991. Geology of the I = 4_field Sheets Nos. 29 and 31, Winneba SW and NW. groundwater well potential mapping. J. Hydrol. 548, 471–483. 
Ghana Geological Survey Department, Accra. Archive Report 25.  Noble, W.S., 2006. What is a support vector machine? Nat. Biotechnol. 24 (12), 
Du, X., Zhou, K., Cui, Y., Wang, J., Zhang, N., Sun, W., 2016. Application of fuzzy 1565–1567. 
analytical hierarchy process (ahp) and prediction-area (pa) plot for mineral Nunoo, S., Hofmann, A., Kramers, J., 2022. Geology, zircon U–Pb dating and εHf data for 
prospectivity mapping: a case study from the dananhu metallogenic belt, xinjiang, the Julie greenstone belt and associated rocks in NW Ghana: implications for 
nw China. Arabian J. Geosci. 9, 1–15. Birimian-to-Tarkwaian correlation and crustal evolution. J. Afr. Earth Sci. 186, 
Dzigbodi-Adjimah, K., 2004. The mineralogy and petrography of the ferruginous 104444. 
manganese rocks at mankwadzi, Ghana. J. Afr. Earth Sci. 38 (3), 293–315. Opare-Addo, E., Browning, P., John, B., 1993. Pressure-temperature constraints on the 
Feng, X., Wang, E., Ganne, J., Amponsah, P., Martin, R., 2018. Role of volcano- evolution of an early Proterozoic plutonic suite in southern Ghana, west africa. 
sedimentary basins in the formation of greenstone-granitoid belts in the west african J. Afr. Earth Sci. 17 (1), 13–22. 
craton: a numerical model. Minerals 8 (2), 73. Parsa, M., Maghsoudi, A., 2021. Assessing the e ects of mineral systems-derived 
Feng, X., Wang, E., Amponsah, P.O., Ganne, J., Martin, R., Jessell, M.W., 2019. Effect of exploration targeting criteria for random forests-based predictive mapping of 
pre-existing faults on the distribution of lower crust exhumation under extension: mineral prospectivity in Ahar-Arasbaran area, Iran. Ore Geol. Rev. 138, 104399. 
numerical modelling and implications for NW Ghana. Geosci. J. 23, 961–975. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., 
Feybesse, J.-L., Billa, M., Guerrot, C., Duguey, E., Lescuyer, J.-L., Milesi, J.-P., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., 
Bouchot, V., 2006. The paleoproterozoic ghanaian province: geodynamic model and Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E., 2011. Scikit-learn: machine 
ore controls, including regional stress modeling. Precambrian Res. 149 (3–4), learning in Python. J. Mach. Learn. Res. 12, 2825–2830. 
149–196. Perrouty, S., Aillères, L., Jessell, M.W., Baratoux, L., Bourassa, Y., Crawford, B., 2012. 
Forson, E.D., Menyeh, A., 2023. Best worst method-based mineral prospectivity Revised Eburnean geodynamic evolution of the gold-rich southern ashanti belt, 
modeling over the central part of the southern kibi-winneba belt of Ghana. Earth Ghana, with new field and geophysical evidence of pre-Tarkwaian deformations. 
Science Informatics 1–20. Precambrian Res. 204, 12–39. 
Forson, E.D., Menyeh, A., Wemegah, D.D., Danuor, S.K., Adjovu, I., Appiah, I., 2020. Pham, B.T., Jaafari, A., Van Phong, T., Mafi-Gholami, D., Amiri, M., Van Tao, N., 
Mesothermal gold prospectivity mapping of the southern kibi-winneba belt of Ghana Duong, V.-H., Prakash, I., 2021. Naïve bayes ensemble models for groundwater 
based on fuzzy analytical hierarchy process, concentration-area (c-a) fractal model potential mapping. Ecol. Inf. 64, 101389. 
and prediction-area (p-a) plot. J. Geophys. 174, 103971. Porwal, A., Carranza, E.J.M., Hale, M., 2006. Bayesian network classifiers for mineral 
Forson, E.D., Menyeh, A., Wemegah, D.D., 2021. Mapping lithological units, structural potential mapping. Comput. Geosci. 32 (1), 1–16. 
lineaments and alteration zones in the southern kibi-winneba belt of Ghana using Rodriguez-Galiano, V., Sanchez-Castillo, M., Chica-Olmo, M., Chica-Rivas, M., 2015. 
integrated geophysical and remote sensing datasets. Ore Geol. Rev. 137, 104271. Machine learning predictive models for mineral prospectivity: an evaluation of 
Forson, E.D., Amponsah, P.O., Hagan, G.B., Sapah, M.S., 2022a. Frequency Ratio-Based neural networks, random forest, regression trees and support vector machines. Ore 
Flood Vulnerability Modeling over the Greater Accra Region of ghana. Modeling Geol. Rev. 71, 804–818. 
Earth Systems and Environment, pp. 1–20. Saberioon, M., Císar, P., Labbé, L., Soucek, P., Pelissier, P., Kerneis, T., 2018. 
Forson, E.D., Wemegah, D.D., Hagan, G.B., Appiah, D., Addo-Wuver, F., Adjovu, I., Comparative performance analysis of support vector machine, random forest, 
Otchere, F.O., Mateso, S., Menyeh, A., Amponsah, T., 2022b. Data-driven multi- logistic regression and k-nearest neighbours in rainbow trout (oncorhynchus mykiss) 
index overlay gold prospectivity mapping using geophysical and remote sensing classification using image-based features. Sensors 18 (4), 1027. 
datasets. J. Afr. Earth Sci. 190, 104504. Salvi, S., Amponsah, P.O., Siebenaller, L., Béziat, D., Baratoux, L., Jessell, M., 2016. 
Geodita Resources Ltd, 2007. Exploration Summary on the Gomoa Mangoase Licence Shear-related gold mineralization in Northwest Ghana: the Julie deposit. Ore Geol. 
(Unpublished internal report).  Rev. 78, 712–717. 
Geodita Resources Ltd, 2012. Trench Report on Gomoa Mangoase Concession Sapah, M.S., Agbetsoamedo, J.E., Amponsah, P.O., Dampare, S.B., Asiedu, D.K., 2021. 
(Unpublished internal report).  Neodymium isotope composition of palaeoproterozoic birimian shales from the wa- 
Geodita Resources Ltd, 2013. Drill Report on Gomoa Mangoase Concession (Unpublished lawra belt, north-west Ghana: constraints on provenance. Geol. J. 56 (4), 
internal report).  2072–2081. 
Ghana Limited, Newmont, 2006. Terminal Report on Gomoa Prospecting License (Pl3/66 Shabankareh, M., Hezarkhani, A., 2017. Application of support vector machines for 
(Unpublished internal report).  copper potential mapping in kerman region, Iran. J. Afr. Earth Sci. 128, 116–126. 
Harris, J., Grunsky, E., Behnia, P., Corrigan, D., 2015. Data-and knowledge-driven Shirmard, H., Farahbakhsh, E., Müller, R.D., Chandra, R., 2022. A review of machine 
mineral prospectivity maps for Canada’s north. Ore Geol. Rev. 71, 788–803. learning in processing remote sensing data for mineral exploration. Rem. Sens. 
Hirdes, W., Davis, D., Eisenlohr, B., 1992. Reassessment of proterozoic granitoid ages in Environ. 268, 112750. 
Ghana on the basis of u/pb zircon and monazite dating. Precambrian Res. 56 (1–2), Sun, T., Chen, F., Zhong, L., Liu, W., Wang, Y., 2019. Gis-based mineral prospectivity 
89–96. mapping using machine learning methods: a case study from tongling ore district, 
eastern China. Ore Geol. Rev. 109, 26–49. 
10
E.D. Forson and P.O. Amponsah                                                                                                                                                          J  o u  r n  a  l  o  f  A  f r i c  a  n  E  a  r t h   S  c i e  n  c e  s  2 06 (2023) 105024
Sun, T., Li, H., Wu, K., Chen, F., Zhu, Z., Hu, Z., 2020. Data-driven predictive modelling Zeghouane, H., Allek, K., Kesraoui, M., 2016. Gis-based weights of evidence modeling 
of mineral prospectivity using machine learning and deep learning methods: a case applied to mineral prospectivity mapping of sn-w and rare metals in laouni area, 
study from southern Jiangxi Province, China. Minerals 10 (2), 102. central hoggar, Algeria. Arabian J. Geosci. 9 (5), 1–13. 
Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., 2012. Landslide susceptibility Zhan, C., Dai, Z., Yang, Z., Zhang, X., Ma, Z., Thanh, H.V., Soltanian, R.M., 2023. 
assessment in vietnam using support vector machines, decision tree, and naive bayes Subsurface sedimentary structure identification using deep learning: a review. Earth 
models. Math. Probl. Eng. 20 (12), 2012.  Sci. Rev. 239, 104370. 
Vapnik, V., 1999. The Nature of Statistical Learning Theory. Springer Science & Business Zhang, Y., Liu, L., 2019. Temporal point pattern analysis of human activities using gis 
Media. methods: a case study of library visiting activities in chengdu city, China. Prof. 
Wemegah, D.D., Preko, K., Noye, R.M., Boadi, B., Menyeh, A., Danuor, S.K., Geogr. 71 (4), 738–750. 
Amenyoh, T., et al., 2015. Geophysical interpretation of possible gold mineralization Zhang, D., Agterberg, F., Cheng, Q., Zuo, R., 2014. A comparison of modified fuzzy 
zones in kyerano, south-western Ghana using aeromagnetic and radiometric weights of evidence, fuzzy weights of evidence, and logistic regression for mapping 
datasets. J. Geosci. Environ. Protect. 3 (4), 67. mineral prospectivity. Math. Geosci. 46, 869–885. 
Xiong, Y., Zuo, R., Carranza, E.J.M., 2018. Mapping mineral prospectivity through big Zhang, L., Huang, M., Li, M., Lu, S., Yuan, X., Li, J., 2021. Experimental study on 
data analytics and a deep learning algorithm. Ore Geol. Rev. 102, 811–817. evolution of fracture network and permeability characteristics of bituminous coal 
Xtra-Gold Resource Corporation, 2021. Kibi Gold Project. Exploration report. under repeated mining e ect. Natural Resources Research, pp. 1–24. 
Xu, Y., Li, Z., Xie, Z., Cai, H., Niu, P., Liu, H., 2021. Mineral prospectivity mapping by Zuo, R., Carranza, E.J.M., 2011. Support vector machine: a tool for mapping mineral 
deep learning method in yawan-daqiao area, gansu. Ore Geol. Rev. 138, 104316. prospectivity. Comput. Geosci. 37 (12), 1967–1975. 
Youssef, A.M., Pourghasemi, H.R., 2021. Landslide susceptibility mapping using machine Zuo, R., Carranza, E.J.M., Wang, J., 2016. Spatial analysis and visualization of 
learning algorithms and comparison of their performance at abha basin, asir region, exploration geochemical data. Earth Sci. Rev. 158, 9–1.  
Saudi Arabia. Geosci. Front. 12 (2), 639–655. 
11