Sage Journals: Discover world-class research

Abstract

This study determines which predictors derived from geophysics or remote sensing data best generate a mineral prospectivity model (MPM) over Ghana's southern Kibi-Winneba belt in a scenario-based modeling case using Random Forest (RF) algorithm. Ten geophysically-derived predictors and six-remote sensing derived predictors were used as inputs in the first and second scenarios respectively. In the third case, the sixteen predictors derived from these afore-mentioned geoscientific datasets were used as inputs. Thus, three binary RF-based MPM were generated, and compared accordingly. The predictive performance in all three scenario-based RF-derived MPM produced was determined using the area under the receiver operating characteristic curve (AUC). AUC scores of 0.840, 0.785 and 0.809 respectively, were obtained for the first, second and third scenarios. The AUC scores obtained further indicates that, MPM developed based on using only the geophysics-sourced layers as inputs performed better in comparison with the MPMs generated in second and third scenarios.

Keywords

random forest geophysical data remote sensing mineral potential mapping southern Kibi-Winneba belt

Introduction

Mineral prospectivity modelling (MPM) is a crucial procedure that can assist mineral exploration geoscientists in defining regions that are significantly endowed with the potential of hosting a specific mineral deposit.^1,2 However, it is noteworthy that uncertainties related to the geological and mathematical frameworks applied, as well as uncertainties pertaining to the geoscientific and geospatial datasets employed, are the main factors impeding the effectiveness of the use of the MPM technique.^3–5 Uncertainties in the geoscientific and geospatial datasets used may result from the low resolution of the data as well as mistakes made during data reduction and cleaning procedures.^6,7 Usually, these inefficiencies have an impact on the quality of the geoscientific thematic layers obtained. Furthermore, the use of thematic layers that have no geoscientific relevance to the sought-after mineral reduces the efficiency of the MPM generated. Uncertainties in the geological framework arise from a lack of a complete understanding and appreciation of the processes that led to the formation of the mineral deposit of interest within a specific geologic environment.^2,8 Uncertainty in mathematical framework comes from the use of a non-robust mathematical approach to determine weights of various geoscientific layers that are to be synthesised into a prospectivity model.² The integration of various geoscientific thematic layers to produce an MPM based on a mathematical framework generally takes two forms: knowledge- and data-driven mathematical frameworks.¹ The former is used to generate MPM in areas where known mineral occurrences are rare or absent; in this case, mineral exploration geoscientists’ opinions are incorporated in determining the relative importance of each of the thematic or evidential layers in relation to the mineral deposit to be explored.^9,10 However, one major disadvantage of using knowledge-driven mathematical frameworks is the extreme subjectiveness involved in assigning weights to the thematic layers that would be synthesised to produce the MPM.¹¹ For the data-driven frameworks (which could be bivariate or multivariate), regions of known mineral deposit occurrences are used in the production of an MPM.^11,12,13 It is noteworthy that several multivariate-based statistical approaches such as machine learning (ML) algorithms are included in the mathematical frameworks used in data-driven-based MPM.¹⁴ Several studies have demonstrated the efficacy of ML algorithms (multivariate data-driven methods) in MPM and have subsequently lauded their superiority over the bivariate data-driven methods in a variety of geologic environments.^2,15–23 Although many ML-based supervised classification algorithms have been employed in the MPM over several geological terranes, the random forests (RFs) are commonly used. This could be attributed to the premise that the RF classifier is more robust than the other ML algorithms in terms of its ability to reduce the effect of overfitting.^{2,17–20,24,25}

Of all the gold belts in Ghana, it is the Kibi-Winneba belt and particularly, its southern part, otherwise known as the southern Kibi-Winneba belt (SKWB) that is the most under-explored, despite its geological terrane being thought to be analogous to the well-known Ashanti belt.²⁶ Efforts to understand the mineral potential of the SKWB have resulted in the use of both knowledge-driven and bivariate data-driven methods to generate an MPM.^11,27 To highlight the mineral prospects as well as target zones that can guide future exploration programmes over the SKWB, this study determines which category of geospatial datasets whose thematic layers can produce the best accuracy when employed in delineating prospective zones of mineral occurrences. Also, by taking into cognisance the superiority of ML algorithms over traditional data-driven methods, and in order to contribute to narrowing down to more favourable zones for the delineation of gold mineral deposits, this study employs the RF algorithm to develop MPM over the SKWB based on three scenarios: thematic layers (a) derived from geophysical datasets comprising magnetic, radiometric and gravity, (b) derived from remote sensing data (Landsat 8 imagery) and (c) derived from geophysical and remote sensing datasets. The performance of the MPMs generated would be assessed using the receiver operating characteristics (ROC) curve to determine which scenario-based dataset is best for outlining prospective zones of mineral occurrences within the SKWB.

Study area and geological setting

Study area

The study area spans three municipalities (thus Efutu, Agona West and Mfantsiman municipalities) and five districts (which include the Awutu Senya, Gomoa East, Gomoa West, Ajumako Enyan Esiam and Asikuma Adoben Brakwa) in the Central Region of Ghana and covers an area of 1902 km². The area of focus is defined by the longitude/latitude World Geodetic System ellipsoid coordinates 5.63^∘N, 5.18^∘N and 0.98^∘W and 0.53^∘W (Figure 1(a)). The study area is 65.5 km southwest of the nation’s capital, accessible via a sealed bitumen road. The study area’s topography generally ranges from 6 to 255 m.²⁸ Lower elevations are closer to the coast, while higher elevations are closer to the northern half of the study area (Figure 1(b)).

Figure 1.

Map of (a) the central region of Ghana showing various administrative districts. (b) The digital elevation model of the study area (marked in red).

Regional geological setting

The northeast-southwest oriented Eburnean tectono-magmatic and metamorphic Kibi-Winneba greenstone belt cropping out in Ghana is part of the Birimian or Boulé-Moussi domain of the West Africa Craton. The belt (Figure 2) is bounded to the east by the Suhum basin, to the west by the Cape-Coast basin, to the south by the Gulf of Guinea’s Atlantic equatorial coast, and to the north by the Volta Basin.²⁹ The Kibi-Winneba greenstone belt is a belt divided into two halves (thus north and southern halves) by the granitic intrusions composed of the 2116 $\pm$ 2 Ma (from U-Pb dating of zircons by Hirdes et al.³⁰ to 2102 $\pm$ 4 Ma biotite ( $\pm$ hornblende $\pm$ muscovite) granitoid of the Cape Coast basin and 2134 $\pm$ 4 Ma biotite gneiss of the Suhum basin (Figure 2; Agyei Duodu et al.²⁹). The southern half of the belt which straddles towns such as Otuam, Apam, Winneba and Agona Swedru and is composed of syn-volcanic and plutonic rocks of the 2179 $\pm$ 2 Ma³⁰ hornblende-biotite granitoid, 2113 $\pm$ 2 Ma³¹ biotite granite with Archean continental signatures, 2159 $\pm$ 4 Ma volcanoclastic sediments (Figure 2; Hirdes et al.³²), volcanic flows, amphibolite, and quartz-sericite schist with $\pm$ straoulite and garnet. The northern half of the belt is composed of basaltic flows, undifferentiated volcaniclastic rocks and dacitic to rhyolitic flow.^33,34 These rocks have been overlain by the late Paleoproterozoic Tarkwaian molassic sediments composed of detrital sediments (sandstones and conglomerates). Juxtaposing the Kibi-Winneba belt to the east is the Suhum basin. This basin is underlain by a granitic complex composed of granitoids with migmatitic and gneissic components. According to Agyei Duodu et al.,²⁹ the granitic complex comprises four types of granitoids. The first two types (thus Types 1 and 2) are composed of 2187 $\pm$ 1 Ma biotite gneisses with minor schist and 2134 Ma granitoid termed as the ‘Tamnean protolith’, respectively. The first two types have been intruded by Types 3 and 4 composed of 2106 $\pm$ 1 Ma K-feldspar rich granitoid and 2088 $\pm$ 1 Ma³⁰ two mica granite with minor leucogranites. The west of the Kibi-Winneba belt is the Cape Coast basin. The basin can be divided into two halves. The southern half is composed of a granitoid complex of 2102 $\pm$ 1 Ma biotite granitoid all intruded by 2088 $\pm$ 1 Ma two mica granitoid. The north half is composed of volcaniclastic sediments, 2187 $\pm$ 1 Ma biotite gneiss and ultramafics composed of peridotite and pyroxenites (Figure 2).

Figure 2.

Geological map of the study area.²⁹

The tectonic evolution of southern Ghana, which includes the Kibi-Winneba belt, is based primarily on the work of Feybesse et al.³⁵ These authors invoked a polyphase tectonic framework denoted as D1 to D4 deformation event. D1 is composed of northeast-southwest shortening defined by thrust faulting around 2107 to 2097. This finite strain affected all the magmatic and sedimentary rocks deposited before that period. This period also corresponds to the deposition of the Tarkwaian sediments on the belts in southern Ghana. The rocks affected by this deformation event were metamorphosed up to a temperature between 550^∘C and $650^{\circ}$ C and a pressure of 5 kbar³⁶ D2/D3 is evident by an northwest-southeast shortening manifesting by northeast-southwest strain which includes shear zones and associated strike-slip faults, folding of Tarkwaian sediments and the precipitation of hydrothermal gold mineralisation. This event is associated with greenschist metamorphism with temperatures between 200 and 300^∘C and pressures of 2 to 3 kbar. The last event is associated with late plutonism.

Mineralisation style of the area

The primary lode gold mineralisation types associated with the rocks in the Birimian are:

The steeply (Konongo, Ashanti mines by Oberthür et al.³⁷ and Bepkong deposit³⁸ and shallow dipping (Julie deposit; Amponsah et al.³⁹) quartz vein in shear zones).

Disseminated sulphides (up to 5% sulphide intensity and comprises of pyrite, chalcopyrite, pyrrhotite and sphalerite) spatially associated with quartz vein in metasedimentary rocks with arsenopyrite as the host.^37,38

Disseminated and stockwork mineralisation associated with granitoid.

Gold mineralisation is associated with the Tarkwaian sediments.

The hydrothermal and mesothermal gold type described in 1–3 usually occurs as a micron size gold inclusion or larger grains within the fractures and the edges of the sulphides or as free gold in the fractures associated with the sulphide and usually sourced from up to 10 km deep.^37,38,40 Apart from the Tarkwaian deposit which is a paleoplacer or molassic deposit mineralisation, all the other three are associated with the late Eburnean events that affected the rocks of the Birimian.^41,42 Indicator minerals associated with these gold mineralisation types are galena, sphalerite, chalcopyrite, pyrrhotite, marcasite, pyrite, gahnite, jarosite, barite, quartz, k-feldspar and tourmaline, bournonite, lollingite and xenotime, chlorite and ankerite.^37,38,39,43 The pathfinder elements associated with this mineralisation type associated with the metasediments are copper (Cu), zinc (Zn), lead (Pb), arsenic (As), stibnite (Sb) and the ones associated with granitoid are tellurium (Te), bismuth (Bi) and wolframite (W). By way of ore geometry, these hydrothermal quartz veins occur as a series of quartz lodes of varying thickness ranging from 5 to 50 m, which are mostly parallel to the shear zone associated with the late Eburnean deformation^42,44 or strata bound gold mineralisation associated with conglomerates in the Tarkwaian. The Fluid system studied across various deposits in the Birimian in Ghana shows that gold precipitate in CO₂–CH₄–N₂ and CO₂–H₂O fluids with salinities ranging from 0 to 8 wt. eq. NaCl at a temperature of 200^∘C to 370^∘C.^{38,39,40,45,46}

Materials and methods

Data

Sources of data

The airborne geophysical datasets comprising radiometrics and magnetics used in this study were sourced from the Ghana Geological Survey Authority, satellite gravity data from the GFZ German Research Centre for Geoscience (www.gfz-potsdam.de), Landsat 8 OLI data was sourced from the United States Geological Survey Earth Resources Observation and Science Center) and the Southern Kibi-Winneba belt 1:100,000 digital geological map was compiled by Agyei Duodu et al.²⁹ The Landsat 8 OLI is made up of nine bands with Level 1 processing applied. The airborne radiometric data was supplied with the channels of the three naturally occurring radioactive elements comprising uranium (eU in ppm), thorium (eTh in ppm) and potassium (K in %). The total magnetic intensity (TMI) grid (in nT) and Bouguer anomaly grid (in mGal) were derived from the airborne magnetic data and satellite gravity data, respectively.

Data preprocessing

Prior to employing the RF classifier to execute ML-based training and evaluation, the input data (predictors and target labels) were projected to WGS84 UTM Zone 30 N. Six thematic layers comprising B4/B2, B4/B3, B4/B7, B6/B2, B6/B7 and B6/B5 $\times$ B4/B5 were derived from the Landsat 8 data and used as predictors. As part of the preprocessing procedure leading to the generation of predictors from the Landsat 8 dataset, supervised and unsupervised classification approaches were applied to bands 2, 3, 4, 5, 6 and 7, so as to overcome the probable overlap between mineral spectrum reflectances and vegetation-associated reflectances.⁴⁷ Thus, this approach was vital in distinguishing bare soils and rocks from areas covered with vegetation. In the case of the airborne magnetic data, the TMI data obtained were transformed using the reduction to pole (RTP) technique in geosoft oasis montaj software to generate the RTP grid. Transforming the magnetic data using the RTP technique removes asymmetric effects observed in magnetic responses in low magnetic latitude regions such as the study area. Thus, transforming the magnetic data made the magnetic responses observed more meaningful to understand with respect to various geological units within the terrane. By using the analytical signal filter on the TMI, the analytical signal predictor was also generated. The first vertical derivative (FVD) grid was derived by applying the FVD filter on the RTP grid. To create the lineament density grid, the RTP grid was subjected to the Center for Exploration Targeting (CET) grid analysis.⁴⁸ The Bouguer anomaly data were subjected to topographic corrections to generate the complete Bouguer anomaly grid which was further processed to generate the residual gravity (RG) response grid by carrying out the regional–residual separation technique. For the airborne radiometric data, the individual channels were gridded to assist in the generation of channel ratios. In all, 16 predictors comprising RG, RTP, analytic signal (AS), FVD, fracture density (Frac_Den), B4/B2, B4/B3, B4/B7, B6/B2, B6/B7, B6/B5 $\times$ B4/B5, K, eTh, eU, K/eTh ratio and eU/eTh ratio were derived for this study. These aforementioned predictors derived from geophysical and remote sensing layers were necessary for the following reasons:

The RTP and AS predictors are capable of delineating highly magnetic anomalous responses, which could be due to indicator minerals such as arsenopyrite, pyrite and magnetite.^9,27

FVD and Frac_Den predictors are capable of outlining various structures such as faults, joints, folds and contacts that are likely to host mineralisations.⁴⁹

In the determination of bulk mineral deposits which have inclusions of arsenopyrite, pyrite and chalcopyrite, the gravity predictor serves a good purpose in that regard.⁵⁰

Radiometric responses are essentially dependent on the occurrence of mineral species that contain one or more of K, eTh and eU. Thus, eU and eTh are observed in quartz and other rock-forming minerals.⁵¹ For the ratios such as K/eTh and eU/eTh, their significance is observed in the delineation of zones of hydrothermal alteration occurrence, with relevance towards mineralisation within the study area.

The remote sensing generated bands comprising B4/B2, B4/B3, B4/B7, B6/B5 $\times$ B4/B5, B6/B2, B6/B7, are useful in mapping ferric iron minerals, carbonate minerals, iron minerals, metamorphic, volcanic rocks and hydroxyl bearing minerals consisting of sericite, chalcopyrite and argillite associated minerals that are essential indicator minerals to gold mineralisation within the study area.^40,52

The predictors (derived from the geophysical, remote sensing and geological datasets) were cropped to the desirable study area extent and were subsequently resampled to a cell resolution of 100 m (resulting in a grid of size 491

\times

554) by employing the bilinear interpolation in R programme. Images produced based on the bilinear interpolation technique are smoothly interpolated and devoid of erroneous substructures that may arise during interpolation.^53,54 Each of these 16 predictors was normalised to achieve a zero mean and unit variance in order to make them suitable for the generation and prediction of a classification model. The prospects of these predictors for delineating prospective zones of mineralisation occurrences within the study area were evaluated using the following scenarios based on the RF classifier:

First scenario: Classification using layers derived from geophysical datasets (AS, FVD, Frac_Den, RTP, RG, K, eTh, eU, K/eTh ratio and eU/eTh ratio).

Second scenario: Classification using layers derived from remote sensing datasets (B4/B2, B4/B3, B4/B7, B6/B2, B6/B7 and B6/B5 $\times$ B4/B5).

Third scenario: Classification using all 16 layers obtained from geophysical and remote sensing layers.

The essence of generating an MPM based on each of these aforementioned scenarios is to determine which dataset’s predictors perform better in delineating prospective zones of mineral occurrences within the study area.

Preparation of target labels

When training and testing a predictive model, target labels describing the occurrence and non-occurrence of a certain mineral deposit are given values 1 and 0, respectively. This is because the occurrence of a mineral deposit over a region of interest is dichotomous.⁵⁵ In the Southern Kibi-Winneba belt, 78 known locations of the gold mineral deposit were reported based on information obtained by Newmont Ghana Limited and Geodita Resource Limited.^56–60 In this study, the point pattern analysis sampling technique was employed to determine the optimal distance beyond which the locations of non-gold (sterile) occurrences should be situated. This technique was employed in this study due to its usefulness in visualising and interpreting the spatial distribution of the point patterns of the known locations of gold occurrences. Based on the point pattern analysis sampling procedure which has also been employed by Carranza et al.¹² and Zu and Carranza,¹⁶ the various locations that are sterile to gold occurrence, which characterise locations of non-gold deposits, were chosen for this investigation. According to this procedure, the number of sterile gold occurrence locations selected should be the same as the number of gold occurrence locations. This eliminates potential problems with data imbalance that can have an impact on a predictive model’s training and testing outcomes. Additionally, it is believed that areas adjacent to proven gold deposit locations have a higher likelihood of developing the same mineralisation style or circumstances. In this regard, it is important to select these gold-sterile sites well away from proven gold deposits. In this study, the use of the point pattern analysis technique was crucial in determining the ideal distance from which different locations of gold-sterile deposits may be formed. Due to this, the maximum distance between any two locations of confirmed gold occurrences was statistically determined using the closest neighbour analysis technique. A value of 2262 m was obtained for the maximum distance, indicating that there is an optimal likelihood of defining gold occurrences within its 2262 m radius for any of the gold deposit locations. It is important to note that, within the statistically computed maximum distance, only a few locations can be selected; hence, a buffer distance of 1893 m indicates a region where there is 83.7% likelihood of identifying another gold deposit that is in a close range in relation to a particular known gold occurrence was selected. The final step in this procedure suggested by Carranza et al.¹² and Zuo and Carranza¹⁶ stipulate that the occurrence of a mineral within a given region is spatially distributed in a clustered fashion owing to the premise that mineral occurrence is scarce and characteristically bonded to a systematic ore-forming process. In contrast to this, various locations of gold-sterile deposits selected ought to be spatially distributed in an unsystematic manner as they ensued from common geological processes. From the aforementioned procedure outlined, non-deposit (gold-sterile) locations within the study area were selected randomly as part of the target labels. In this study, 80% of the target labels (with an equal number of gold occurrences and non-occurrences) were used to train the RF model whereas the remaining 20% was used to test the performance of the model produced. The splitting of the target labels into training and testing labels was carried out using the splitting procedure outline by Le Rest et al.,⁶¹ which is referred to as the spatial leave-one-out cross-validation approach. By employing this procedure, issues that pertain to spatial autocorrelation are taken care of.

RF algorithm

The RF algorithm, which was first proposed by Ho⁶² and later developed by Breiman⁶³ is a reliable and robust ML algorithm for carrying out classifications. It has numerous applications in geospatial studies for geological modelling^64,65 as well as predictions for mineral^2,66 and groundwater^67,68 The fundamental principle underlying the RF approach has been extensively documented in various works of literature^69–72 The RF algorithm is a conglomerate of several decision trees; each tree is a collection of multiple bootstrap samples that are created by original samples (referred to as bagging). During bagging, many new values of training data area created by resampling the original data in a random fashion. A bootstrap sample is employed to construct each individual tree within the decision tree, whereby a subset of various attributes are selected at each node of a tree.⁷³ In this study, the Gini index was deemed as the optimal parameter during the implementation of the RF algorithm in python programming language to determine the impurity of an attribute with respect to various prospectivity (target) classes involved based on the grid-search cross-validation approach.

Training and evaluating RF classifier

A training process was carried out to generate the RF-based ML model after preparing the input data, which included predictors and target labels. A critical stage in training an ML model is determining the key parameters that will be used to generate a model with high accuracy. However, it is important to note that in data-driven modelling, the specification of the aforementioned parameter based on a suitably deduced configuration is extremely difficult due to the premise that obtaining optimal parameters for ML models changes depending on the nature of the input data employed. This is sufficient to show that there is no universally recognised empirical procedure that must be followed in order to arrive at appropriate parameters that best fit a predictive model developed. In this study, the grid search cross-validation (cv) approach otherwise known as grid search cv was implemented to determine the optimal parameters (shown in Table 1) for each scenario based on a particular range of values chosen based on literature (shown in Table 2). These optimal parameters were subsequently used to generate the predictive models in each of the aforementioned scenarios. A 10-fold cross-validation method using the best parameters found using the grid search cv technique was used to evaluate the effectiveness of the classification performed using the RF classifier.

Table 1.

Best or optimum parameters for the random forest classifier training in each scenario using the grid search cross-validation (cv).

Parameter	Scenario 1 optimal parameters	Scenario 2 optimal parameters	Scenario 3 optimal parameters
Bootstrap	True	True	True
Criterion	gini	gini	gini
Minimum depth	46	26	6
Minimum samples leaf	2	2	6
Minimum samples split	27	17	27
Number of estimators	26	11	11

Table 2.

Parameter ranges for training the random forest classifier.

Parameter	Parameter description	Chosen range
Bootstrap	This indicates whether bootstrap samples are to be used when building trees.	True
Criterion	The characterise the function that measures the quality of a split.	gini, entropy, log_loss
Maximum depth	The maximum depth of the tree.	1–100
Minimum samples leaf	The minimum number of samples required to be at a leaf node.	1–40
Minimum samples split	The minimum number of samples required to split an internal node.	2–15
Number of estimators	Contains the number of trees the random forest classifier builds before averaging the predictions.	1–1000

In carrying out predictive modelling, it is also important to ascertain the influence of each of the predictors used towards the model created (known as feature importance). In this study, the determination of the importance of each of the predictors used was determined based on the optimum (best) parameters determined by the grid search cv technique. The feature importance depicts the contribution of each predictor towards the predictive model generated. The feature importance ( $f i$ ) for a particular predictor $i$ is computed based on the expression in equations (1), (2) and (3).

f i_{i} = \frac{\sum_{j : node j splits o n feature i} n i_{j}}{\sum_{k \in a l l nodes} n i_{k}}

(1)

f i_{i, :, norm} = \frac{f i_{i}}{\sum_{j \in all features} f i_{j}}

(2)

where

f i_{i}

and

n i_{j}

represent, respectively, the importances of feature

i

and node

j

f i_{i, norm}

depicts the normalised

f i

for predictor

i

in tree

j

with values ranging between 0 and 1.

Based on the RF approach, the final feature importance is computed based on equation (3):

f i_{i, RF} = \frac{\sum_{j \in all trees} norm f i_{j}}{T}

(3)

where

f i_{i, RF}

represents the

f i

for a predictor

i

that is computed based on all trees in the RF model and

T

is the total number of trees. Thus, the computation of the feature importance for each of the predictors used was done using the RandomForestClassifier().feature_importances_ tool in the scikit-learn library in python programming language. It is noteworthy that, from the expression in equation (3),

f i_{i, RF}

could have a maximum value >1 or

< 1

. Thus, for a set of predictors that are to be synthesised to generate a model, the one with the highest

f i_{i, RF}

value is deemed as the predictor

i

that is the most important to the predictive model generated whereas the one with

f i_{i, RF}

close to zero is said to have the least influence towards the predictive model generated.

In geospatial modelling, predictive models generated should be evaluated to convey confidence in the outputs produced.⁷⁴ In this regard, the efficacy of the MPM generated based on the RF algorithm was evaluated using the receiver operating characteristics (ROC) curve to ascertain how the known gold occurrences and the mineral predictive model generated spatially correlate. An ROC curve visualises and measures the performance of a classification model based on its rate or correct and incorrect classifications. The precise trade-off between the false positive rate and the true positive rate is also determined by the ROC curve. For well-balanced target labels, the use of the ROC curve is more appropriate. A typical ROC curve features the true positive rate, which is analogous to $s e n s i t i v i t y$ along the ordinate axis and the false positive rate (synonymous to $1 - s p e c i f i c i t y$ ) along the abscissa. The false positive rate and the true positive rate are mathematically expressed in equation (4) and equation (5), respectively. The area under the ROC curve (AUC) produces a score, which outlines the spatial correlation between the predictive model generated and the testing datasets. It is noteworthy that, high AUC score values signal a situation of a strong spatial correlation between the predictive model and the testing dataset within the study. A high AUC score also indicates a better model and vice versa.

FPR = \frac{FP}{TN + FP}

(4)

TPR = \frac{TP}{FN + TP}

(5)

From equations (4) and (5), FP, TN, TP and FN characterise, respectively, the false positive, true negative, true positive and false negative.

Results and discussion

Importance of the thematic layers towards the generation of the MPMs

Feature importance constitutes how much a particular geoscientific conditioning parameter is essential when carrying out mineral potential zonation. To assess the contribution of each of the features or thematic layers (Figures 3(a) to (d), 4(a) to (f), and 5(a) to (f)) towards the MPMs produced in each of the scenarios. As suggested by Breiman,⁶³ the relevance of a thematic layer is obtained by permuting its corresponding values after training it based on an RF classifier. Figure 6 illustrates the importance of each of the geophysically derived thematic layers that served as input for the generation of the MPM in the first scenario. By using only the aforementioned geophysically-derived layers as input features for the RF-based MPM, the K/eTh ratio layer was the feature with the most influence; followed by analytic signal, first vertical derivative, uranium concentration, eU/eTh ratio, thorium concentration, fracture density, RG, potassium concentration and RTP layer in a decreasing order of influence. In the case of the second scenario, where remote sensing layers were used as input features to produce an MPM, band ratio B6/B2 had the highest feature importance and thus its contribution towards the MPM produced is the highest as shown in Figure 7. Conversely, the feature importance of the B4/B3 ratio was the least, an indication of its influence towards the predictive model being the lowest. In the third scenario (shown in Figure 8), where both the geophysical and remote sensing layers were integrated to produce the MPM, the potassium/thorium ratio layer was observed as the feature with the highest importance towards the MPM, whereas the band ratio B4/B7 had the least influence on the model generated. K/eTh ratio being the feature with the highest relevance towards the MPMs produced in this instance corroborates with literature assertion that regions with K increasing and eTh decreasing indicate the prevalence of alteration in mineral ore deposits within the southern Kibi-Winneba belt⁵⁰. Thus, regions delineated as prospective zones of gold mineralisation occurrence within the southern Kibi-Winneba belt in the first and third scenarios, where K/eTh was used as an input feature generally exhibited high values of K/eTh ratios.

Figure 3.

Normalised image of (a) analytic signal layer, (b) RTP layer, (c) FVD layer and (d) fracture density layer. RTP: reduction to pole; FVD: first vertical derivative.

Figure 4.

Normalised image of (a) eTh concentration layer, (b) eU/eTh ratio layer, (c) eU concentration layer, (d) K concentration layer, (e) K/eTh ratio layer and (f) residual gravity layer.

Figure 5.

Normalised image of (a) Band 4/Band 2 layer, (b) Band 4/Band 3 layer, (c) Band 4/Band 7 layer, (d) Band 6/Band 2 layer, (e) Band 6/Band 5 $\times$ Band 4/Band 5 layer and (f) Band 6/Band 7 layer.

Figure 6.

Feature importance of the predictors in the first scenario.

Figure 7.

Feature importance of the predictors in the second scenario.

Figure 8.

Feature importance of the predictors in the third scenario.

Mineral prospectivity models

In order to create a predictive model that outlines the prospective and non-prospective classes over the southern Kibi-Winneba belt, the RF classifier was employed by using geoscientific thematic layers in the aforementioned three scenarios (based on only geophysical (GP) layers, only remote sensing (RS) layers and both GP and RS layers). Figure 9(a) to (c) depicts the MPMs generated by the RF classifier using only the geophysically derived layers, only remote sensing-derived layers and layers derived from both geophysical and remote sensing layers respectively. In the case of the first scenario (Figures 9(a)), where only geophysical layers were used for the classification, 20.17% representing an area size of 386.93 km² was delineated as the prospective class of gold mineralisation over the study area as shown in Table 3. The regions delineated as prospective classes were observed to be predominant over the southeastern and the northern northwestern portions of the study area. For the MPM produced based on only the remote sensing layers (Figures 9(b)), the areal extent of the regions delineated as prospective and non-prospective were, respectively, 540.03 km² and 1378.30 km², representing respectively 28.15% and 71.85% of the total areal size of the study area. For this remote sensing scenario, prospective classes of gold mineralisation were observed to dominate the central and eastern portions of the study area. For the third scenario where the MPM was produced based on the synthesising of all layers derived from geophysical and remote sensing datasets (Figures 9(c)), it was observed (as shown in Table 3) that 22.11% of the study area, which is analogous to an area extent of 424.14 km² was delineated as the prospective class with gold mineralisation. Regions delineated as prospective of gold mineralisation based on the RS layers were very profound in the eastern and north-western portions of the study area.

Figure 9.

Mineral prospectivity model (MPM) produced in the (a) first scenario, (b) second scenario and (c) third scenario.

Table 3.

Area extent and percentage of mineral prospectivity classes.

MPM	Scenario 1		Scenario 2		Scenario 3
Class	Area of class (km²)	Percentage (%)	Area of class (km²)	Percentage (%)	Area of class (km²)	Percentage (%)
Non-prospective	1531.40	79.83	1378.30	71.85	1494.19	77.89
Prospective	386.93	20.17	540.03	35.97	424.14	22.11

MPM: mineral prospectivity model.

Evaluation of the MPMs

Predictive models produced are only worthy for consideration of any meaningful decision-making when they have been evaluated to assess their efficacy, which correspondingly builds confidence in users of the predictive models.⁷⁵ In view of this, the performance of the predictive models generated based on the RF classifier in the aforementioned three scenarios was assessed using the receiver operating characteristics (ROC) curve. The performance assessment of the MPMs produced in these scenarios indicates the predictive efficacy of the outputs generated in each of the three scenarios. The area under the ROC curve scores obtained for the MPM results in the first, second and third scenarios are, respectively, 0.840, 0.785 and 0.809 as shown in Figure 10. The AUC scores obtained in the scenarios indicate that the MPM produced by using only the geophysically derived thematic layers as input showed a better performance in comparison with the MPMs produced based on only the remote sensing layers as well as all the sixteen layers (both remote sensing and geophysics derived layers). This suffices to show that, the MPM generated based on the only geophysical layers obtained the highest accuracy while predicting prospective zones of gold mineralisation occurrence within the study area.

Figure 10.

ROC curve for the scenario-by-scenario based MPM generated. ROC: receiver operating characteristics; MPM: mineral prospectivity model.

Discussion

The contribution of the mineral sector towards the economic gains of various countries with natural resources cannot be overemphasised.⁷⁶ Hence, the search for prospective zones of mineral occurrence is on the ascendancy in most mineral-rich countries including Ghana. Ghana’s gold resources are primarily known to be hosted in six main belts comprising the Ashanti, Bole-Nangodi, Bui, Kibi-Winneba, Lawra and Sefwi belts.²⁹ Among these aforementioned belts, it is the Kibi-Winneba belt, particularly its southern part that is being under-explored. In recent years, defining exploration targets for mineral prospecting, otherwise known as mineral prospectivity modelling, various predictors sourced from various geoscientific datasets are synthesised using ML approaches. As an initial stage in the mineral exploration programme, one task that is missing in the literature is to identify which category of geoscientific datasets can best guide the delineation of prospective zones of a sought-after mineral over a region of interest. In view of this, this study generated and compared MPM generated based on the RF approach using only geophysically sourced layers, only layers sourced from remote sensing datasets as well as layers sourced from both geophysical and remote sensing datasets. The development of MPMs whose prospective classes can serve as essential exploration targets requires that issues of target label imbalance, as well as spatial autocorrelation, are addressed. By employing the point pattern analysis technique the optimal distances beyond which non-occurrence data points (of equal number to that of the known occurrences) were created randomly to avoid data imbalance¹⁶. The random creation of these datasets was essential in minimising the biases that may arise if non-occurrence points were done manually. Spatial autocorrelation which arises due during train/test split was addressed by applying the splitting approach proposed by Le Rest et al.⁶¹ In the first and third scenarios, where MPM was generated based on only geophysically sourced layers as well as layers from both geophysics and remote sensing datasets, the K/eTh predictor was deemed to be the most important predictor. This corroborates with literature as the K/eTh ratio predictor is capable of mapping hydrothermal alteration zones that consequentially essential targets for mineral exploration within the study area^77,50. For the MPM generated based on only the predictors sourced from the remote sensing datasets, band ratios B4/B2 were deemed to have the most influence followed by B6/B7. This also corroborates with the literature as these two band ratios are capable of delineating respectively iron and hydroxyl-associated minerals such as pyrite, arsenopyrite, chalcopyrite, and sericite, which are essential indicator minerals to gold occurrence within the study area^40,52. For a study that used a well-balanced target labels as carried out in this study, the determination of the performance of the MPMs generated can be appropriately done using the ROC curve. Through the use of the ROC curve, a score is produced for the area under it, otherwise known as the AUC score. MPM with a higher AUC score indicates a higher performance. Thus, MPM produced based on the geophysically sourced layers has a better performance (with an AUC score of 0.840) than MPM produced based on both geophysics and remote sensing predictors (AUC score of 0.809) as well as only remote sensing predictors (AUC score of 0.785). Also, the outputs from the geophysics-based MPM corroborate enormously with the literature such that the delineated prospective classes are predominantly observed within the metavolcanic (central to the southeastern part of the study area) as well as the northwestern portion of the study area, where gold mineralisation is generally deemed to occur due to the metavolcanic as well as the prevalence of potassium metasomatism due to hydrothermal leaching of quartz within the granitoids^78,50. The outputs of this study further indicate that, in carrying out an initial exploration programme in terranes synonymous to the SKWB, the use of geophysically sourced predictors for MPM can give you optimal results.

Conclusion

Various machine-learning approaches have been employed for regional mineral prospectivity delineation worldwide. In geospatial classification, the RF remains the ML approach, which is robust and commonly used. The main objective of this study was to determine which predictors among geophysical and remote sensing datasets can give optimal results in the generation of an MPM over the southern Kibi-Winneba belt of Ghana using the RF algorithm. The RF-based mineral potential classification in this study was carried out in three scenarios: (a) by using only the geophysical layers as input layers, (b) by employing the layers derived from remote sensing data as inputs and (c) by employing all the layers sourced from both geophysical and remote sensing datasets. Results obtained from the MPM produced indicate that the percentage of the study area delineated as prospective in the first, second and third scenario were, respectively, 20.17%, 28.15% and 22.11%, which is, respectively, analogous to an areal extent of 386.93 km², 540.03 km² and 424.14 km². In order to assess and evaluate the efficacy of the delineated classes based on the RF algorithm in each of the three classification scenarios, the area under the receiver operating characteristics curve was applied by using the test data. According to the AUC scores obtained, the RF-based mineral prospectivity classification in the first, second and third scenarios were 0.840, 0.785 and 0.809. This indicates that the use of geophysics layers as thematic layers for MPM over the southern Kibi-Winneba belt gives a better performance than remote sensing layers. These AUC scores observed are all >0.70, thus the models generated in these three scenarios can be useful for mineral exploration geoscientists in carrying out detailed mineral exploration programmes within the study area.

Footnotes

Acknowledgements

The authors are grateful to the University of Ghana-Carnegie Corporation and Building a New Generation Africa (BaNGA-Africa) for their immense support in making this study a success by funding the first author's postdoctoral fellowship. The authors also wish to thank the United States Geological Survey Earth Resources Observation and Science Center, Geodita Resources Limited and GFZ German Research Centre for Geoscience (Potsdam-Germany) for making data available for use in this study.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

References

Carranza

EJM

. Geochemical anomaly and mineral prospectivity mapping in GIS. Amsterdam: Elsevier, 2008.

Parsa

Maghsoudi

. Assessing the effects of mineral systems-derived exploration targeting criteria for random forests-based predictive mapping of mineral prospectivity in Ahar-Arasbaran area, Iran. Ore Geol Rev 2021; 138: 104399.

Zuo

Zhang

, et al. Evaluation of uncertainty in mineral prospectivity mapping due to missing evidence: A case study with skarn-type Fe deposits in southwestern Fujian province, China. Ore Geol Rev 2015; 71: 502–515.

Burkin

Lindsay

Occhipinti

, et al. Incorporating conceptual and interpretation uncertainty to mineral prospectivity modelling. Geosci Front 2019; 10: 1383–1396.

Wang

Yin

Caers

, et al. A Monte Carlo-based framework for risk-return analysis in mineral prospectivity mapping. Geosci Front 2020; 11: 2297–2308.

Buckley

Howell

Enge

, et al. Terrestrial laser scanning in geology: Data acquisition, processing and accuracy considerations. J Geol Soc 2008; 165: 625–638.

Parsa

Maghsoudi

Ghezelbash

. Decomposition of anomaly patterns of multi-element geochemical signatures in Ahar area, NW Iran: A comparison of u-spatial statistics and fractal models. Arab J Geosci 2016; 9: 1–16.

Ford

McCuaig

. The effect of map scale on geological complexity for computer-aided exploration targeting. Ore Geol Rev 2010; 38: 156–167.

Forson

Menyeh

Wemegah

, et al. Mesothermal gold prospectivity mapping of the southern Kibi-Winneba belt of Ghana based on fuzzy analytical hierarchy process, concentration-area (c-a) fractal model and prediction-area (p-a) plot. J Geophys 2020; 174: 103971.

10.

Amponsah

Danuor

Wemegah

, et al. Groundwater potential characterisation over the Voltaian basin using geophysical, geological, hydrological and topographical datasets. J Afr Earth Sci 2022; 192: 104558.

11.

Forson

Wemegah

Hagan

, et al. Data-driven multi-index overlay gold prospectivity mapping using geophysical and remote sensing datasets. J Afr Earth Sci 2022; 190: 104504.

12.

Carranza

Hale

Faassen

. Selection of coherent deposit-type locations and their application in data-driven mineral prospectivity mapping. Ore Geol Rev 2008; 33: 536–558.

13.

Yousefi

Nykänen

. Data-driven logistic-based weighting of geochemical and geological evidence layers in mineral prospectivity mapping. J Geochem Explor 2016; 164: 94–106.

14.

Chen

Guestrin

. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016, pp.785–794.

15.

Porwal

Carranza

EJM

Hale

. Bayesian network classifiers for mineral potential mapping. Comput Geosci 2006; 32: 1–16.

16.

Zuo

Carranza

EJM

. Support vector machine: A tool for mapping mineral prospectivity. Comput Geosci 2011; 37: 1967–1975.

17.

Rodriguez-Galiano

Sanchez-Castillo

Chica-Olmo

, et al. Machine learning predictive models for mineral prospectivity: An evaluation of neural networks, random forest, regression trees and support vector machines. Ore Geol Rev 2015; 71: 804–818.

18.

Zhang

Zuo

Xiong

. A comparative study of fuzzy weights of evidence and random forests for mapping mineral prospectivity for skarn-type fe deposits in the southwestern Fujian metallogenic belt, China. Sci China Earth Sci 2016; 59: 556–572.

19.

Sun

Chen

Zhong

, et al. Gis-based mineral prospectivity mapping using machine learning methods: A case study from tongling ore district, eastern china. Ore Geol Rev 2019; 109: 26–49.

20.

Sun

, et al. Data-driven predictive modelling of mineral prospectivity using machine learning and deep learning methods: A case study from southern Jiangxi province, china. Minerals 2020; 10: 102.

21.

Shirmard

Farahbakhsh

Müller

, et al. A review of machine learning in processing remote sensing data for mineral exploration. Remote Sens Environ 2022; 268: 112750.

22.

Amponsah

Forson

. Geospatial modelling of mineral potential zones using data-driven based weighting factor and statistical index techniques. J Afr Earth Sci 2023; 206: 105020.

23.

Forson

Amponsah

. Mineral prospectivity mapping over the Gomoa area of Ghana’s southern Kibi-Winneba belt using support vector machine and Naive Bayes. J Afr Earth Sci 2023; 206: 105024.

24.

Polikar

. Ensemble based systems in decision making. IEEE Circuits Syst Mag 2006; 6: 21–45.

25.

McKay

Harris

. Comparison of the data-driven random forests model and a knowledge-driven method for mineral prospectivity mapping: A case study for gold deposits around the Huritz group and Nueltin suite, Nunavut, Canada. Nat Resour Res 2016; 25: 125–143.

26.

Eisenlohr

Hirdes

. The structural development of the early Proterozoic Birimian and tarkwaian rocks of southwest Ghana, West Africa. J Afr Earth Sci (Middle East) 1992; 14: 313–325.

27.

Forson

Menyeh

. Best worst method-based mineral prospectivity modeling over the central part of the southern Kibi-Winneba belt of Ghana. Earth Sci Inform 2023; 16: 1657–1676.

28.

TessaDEM. 2022. https://en-gb.topographic-map.com/map-mj6114/Kibi/?center=5.44772%2C-0.74803&zoom=119 (accessed 1 November 2022).

29.

Agyei Duodu

Loh

Boamah

, et al. Geological map of Ghana 1: 1 000 000. Geological Survey Department, 2009.

30.

Hirdes

Davis

Eisenlohr

. Reassessment of proterozoic granitoid ages in Ghana on the basis of u/pb zircon and monazite dating. Precambrian Res 1992; 56: 89–96.

31.

Davis

Hirdes

Schaltegger

, et al. U-Pb age constraints on deposition and provenance of Birimian and gold-bearing Tarkwaian sediments in Ghana, west Africa. Precambrian Res 1994; 67: 89–107.

32.

Hirdes

Konan

N’Da

, et al. Geology of the northern portion of the Oboisso Area, Côte d’Ivoire. Sheets 4A, 4B, 4B BIS, 4. Hanover, Germany: Direction de la Géologie, Abidjan, Côte d’Ivoire and Bundesanstalt für Geowissenschaften und Rohstoffe, 2007.

33.

Yao

Robb

. The Birimian granitoids of Ghana: A review. Information circular 322. University of the Witwatersrand, Economic Geology Research Unit, Johannesburg, 1998.

34.

Anum

Sakyi

B-X

, et al. Geochemistry and geochronology of granitoids in the kibi-asamankese area of the Kibi-Winneba volcanic belt, southern Ghana. J Afr Earth Sci 2015; 102: 166–179.

35.

Feybesse

J-L

Billa

Guerrot

, et al. The Paleoproterozoic Ghanaian province: Geodynamic model and ore controls, including regional stress modeling. Precambrian Res 2006; 149: 149–196.

36.

Perrouty

Aillères

Jessell

, et al. Revised Eburnean geodynamic evolution of the gold-rich southern Ashanti Belt, Ghana, with new field and geophysical evidence of pre-Tarkwaian deformations. Precambrian Res 2012; 204: 12–39.

37.

Oberthür

Hirdes

Höhndorf

, et al. A review of gold mineralisation in the Ashanti Belt of Ghana and its relation to the crustal evolution of the terrane. Commun Geol Surv Namibia 1995; 10: 121–127.

38.

Amponsah

Salvi

Didier

, et al. Multistage gold mineralization in the WA-Lawra greenstone belt, NW Ghana: The Bepkong deposit. J Afr Earth Sci 2016; 120: 220–237.

39.

Amponsah

Salvi

Béziat

, et al. Geology and geochemistry of the shear-hosted Julie gold deposit, NW Ghana. J Afr Earth Sci 2015; 112: 505–523.

40.

Klemd

Hirdes

Olesch

, et al. Fluid inclusions in quartz-pebbles of the gold-bearing Tarkwaian conglomerates of Ghana as guides to their provenance area. Miner Deposita 1993; 28: 334–343.

41.

Baratoux

Metelka

Naba

, et al. Juvenile Paleoproterozoic Crust Evolution during the Eburnean Orogeny (2.2–2.0 ga), Western Burkina Faso. Precambrian Res 2011; 191: 18–45.

42.

Block

Jessell

Aillères

, et al. Lower crust exhumation during Paleoproterozoic (Eburnean) orogeny, NW Ghana, West African Craton: Interplay of coeval contractional deformation and extensional gravitational collapse. Precambrian Res 2016; 274: 82–109.

43.

Hirdes

Nunoo

. The proterozoic paleoplacers at Tarkwa gold mine, SW Ghana: Sedimentology, mineralogy, and precise age dating of the main reef and west reef, and bearing of the investigations on source area aspects. Geol Jahrb D 1994; 100: 247–311.

44.

Nunoo

Hofmann

Kramers

. Geology, zircon u–pb dating and

ε

hf data for the Julie Greenstone belt and associated rocks in NW Ghana: Implications for Birimian-to-Tarkwaian correlation and crustal evolution. J Afr Earth Sci 2022; 186: 104444.

45.

Yao

Murphy

Robb

. Fluid characteristics of granitoid-hosted gold deposits in the Birimian Terrane of Ghana: A fluid inclusion microthermometric and Raman spectroscopic study. Econ Geol 2001; 96: 1611–1643.

46.

White

Waters

Robb

. Exhumation-driven devolatilization as a fluid source for orogenic gold mineralization at the Damang deposit, Ghana. Econ Geol 2015; 110: 1009–1025.

47.

Mshiu

. Landsat remote sensing data as an alternative approach for geological mapping in Tanzania: A case study in the Rungwe volcanic province, south-western Tanzania. Tanzania J Sci 2011; 37.

48.

Amponsah

Wemegah

Danuor

, et al. Depth-based correlation analysis between the density of lineaments in the crystalline basement’s weathered zones and groundwater occurrences within the Voltaian basin, Ghana. Geophys Prospect Epub ahead of print 2023.

49.

Holden

E-J

Wong

Kovesi

, et al. Identifying structural complexity in aeromagnetic data: An image analysis approach to greenfields gold exploration. Ore Geol Rev 2012; 46: 47–59.

50.

Forson

Menyeh

Wemegah

. Mapping lithological units, structural lineaments and alteration zones in the southern Kibi-Winneba belt of Ghana using integrated geophysical and remote sensing datasets. Ore Geol Rev 2021; 137: 104271.

51.

Dentith

Mudge

. Geophysics for the mineral exploration geoscientist. Cambridge: Cambridge University Press, 2014.

52.

Ourhzif

Algouti

Hadach

. Lithological mapping using Landsat 8 OLI and ASTER multispectral data in imini-OUNILLA district south high atlas of marrakech. Int Arch Photogramm Remote Sens Spat Inform Sci 2019; 42: 1255–1262.

53.

Smith

. Bilinear interpolation of digital images. Ultramicroscopy 1981; 6: 201–204.

54.

Liu

. Bilinear interpolation method for quantum images based on quantum Fourier transform. Int J Quant Inform 2018; 16: 1850031.

55.

Carranza

EJM

Laborte

. Data-driven predictive modeling of mineral prospectivity using random forests: A case study in Catanduanes island (Philippines). Nat Resour Res 2016; 25: 35–50.

56.

Dove

. Geology of the

I = 4

_field sheets nos. 29 and 31, Winneba SW and NW. Archive Report 25. Ghana Geological Survey Department, Accra, 1991.

57.

Newmont Ghana Limited. Terminal report on gomoa prospecting license (pl3/66). Unpublished internal report. 2006.

58.

Geodita Resources Ltd. Exploration summary on the gomoa mangoase licence. Unpublished internal report, 2007.

59.

Geodita Resources Ltd. Trench report on gomoa mangoase concession. Unpublished internal report, 2012.

60.

Geodita Resources Ltd. Drill report on gomoa mangoase concession. Unpublished internal report, 2013.

61.

Le Rest

Pinaud

Monestiez

, et al. Spatial leave-one-out cross-validation for variable selection in the presence of spatial autocorrelation. Glob Ecol Biogeogr 2014; 23: 811–820.

62.

. The random subspace method for constructing decision forests. IEEE Trans Pattern Anal Mach Intell 1998; 20: 832–844.

63.

Breiman

. Random forests. Mach Learn 2001; 45: 5–32.

64.

Kuhn

Cracknell

Reading

. Lithological mapping in the central African copper belt using random forests and clustering: Strategies for optimised results. Ore Geol Rev 2019; 112: 103015.

65.

Wang

Zuo

Jing

. Fusion of geochemical and remote-sensing data for lithological mapping using random forest metric learning. Math Geosci 2021; 53: 1125–1145.

66.

Yang

Zhang

Yang

, et al. Mineral prospectivity prediction by integration of convolutional autoencoder network and random forest. Nat Resour Res 2022; 31: 1103–1119.

67.

Rahmati

Pourghasemi

Melesse

. Application of GIS-based data driven random forest and maximum entropy models for groundwater potential mapping: A case study at Mehran region, Iran. Catena 2016; 137: 360–372.

68.

Sachdeva

Kumar

. Comparison of gradient boosted decision trees and random forest for groundwater potential mapping in Dholpur (Rajasthan), India. Stoch Environ Res Risk Assess 2021; 35: 287–306.

69.

Pal

. Random forest classifier for remote sensing classification. Int J Remote Sens 2005; 26: 217–222.

70.

Denil

Matheson

De Freitas

. Narrowing the gap: Random forests in theory and in practice. In International conference on machine learning, 2014, pp.665–673. PMLR.

71.

Wager

. Asymptotic theory for random forests. arXiv preprint arXiv:1405.0352, 2014.

72.

Biau

Scornet

. A random forest guided tour. Test 2016; 25: 197–227.

73.

Miraki

Zanganeh

Chapi

, et al. Mapping groundwater potential using a novel hybrid intelligence approach. Water Resour Manage 2019; 33: 281–302.

74.

Amponsah

Forson

Sungzie

, et al. Groundwater prospectivity modeling over the Akatsi districts in the Volta region of Ghana using the frequency ratio technique. Model Earth Syst Environ 2023; 9: 1–19.

75.

Forson

Amponsah

Hagan

, et al. Frequency ratio-based flood vulnerability modeling over the Greater Accra region of Ghana. Model Earth Syst Environ 2023; 9: 1–20.

76.

Amponsah-Tawiah

Dartey-Baah

. The mining industry in Ghana: A blessing or a curse. Int J Bus Soc Sci 2011; 2: 62–69.

77.

Elkhateeb

Abdellatif

MAG

. Delineation potential gold mineralization zones in a part of central eastern desert, Egypt using airborne magnetic and radiometric data. NRIAG J Astron Geophys 2018; 7: 361–376.

78.

Cathelineau

. The hydrothermal alkali metasomatism effects on granitic rocks: Quartz dissolution and related subsolidus changes. J Petrol 1986; 27: 945–965.

Random forest-based mineral prospectivity modelling over the Southern Kibi–Winneba belt of Ghana using geophysical and remote sensing techniques

Abstract

Keywords

Introduction

Study area and geological setting

Study area

Regional geological setting

Mineralisation style of the area

Materials and methods

Data

Sources of data

Data preprocessing

Preparation of target labels

RF algorithm

Training and evaluating RF classifier

Results and discussion

Importance of the thematic layers towards the generation of the MPMs

Mineral prospectivity models

Evaluation of the MPMs

Discussion

Conclusion

Footnotes

Acknowledgements

Declaration of conflicting interests

Funding

References