Predictive Mapping of Prospectivity for Gold in the Central Portion of the Tapajós Mineral Province, Brazil

Abstract

This work aims to model mineral prospectivity for intrusion–related gold deposits in the central portion of the Tapajós Mineral Province (TMP), southwestern Pará state. The scope includes experimentation and evaluation of knowledge and data-driven methods applied to multisource data to predict potential targets for gold mineralization. The radiometric data processing allowed to identify a hydrothermal alteration footprint of known gold deposits, providing information in regions with little or no field data available. The aeromagnetic data analysis prompted the identification of high magnetic zones, which are probably related to hydrothermal fluid transport. Linear features extracted from digital elevation data revealed an NNW–SSE general trend, which is consistent with the main structural control of deposits. The data were integrated through three modeling techniques—fuzzy logic (knowledge-driven), weights of evidence (WofE, data-driven), and a machine learning algorithm (SVM, data-driven)—resulting in three prospective models. In all models, the majority of indicated prospective regions coincide with the known deposits. The results obtained in the models were combined to generate an agreement map, which mapped the overlapping of their highest prospective scores, indicating new areas of prospective interest in the central portion of the TMP.

1. Introduction

The growing demand in the mineral sector for the discovery of new mineral deposits has been a relevant issue for the conception of predictive mapping models for mineral prospectivity. Such models involve the analysis and integration of evidential maps derived from multi-source geoscientific datasets to delineate and classify areas favorable to the presence of undiscovered mineral deposits [1].

For mineral prospectivity modeling, geocomputational methods guided by specialized knowledge (knowledge-driven), available data (data-driven), or hybrid methods are employed [1,2]. Knowledge-driven methods are suitable in areas with few exploration targets and limited information, as well as where there is greater knowledge of mineralization components. Data-driven methods are generally applied when there is a robust collection of data with a considerable number of known mineral occurrences [1].

Fuzzy logic (knowledge-driven) and the weights of evidence (WofE, data-driven) are commonly used methods in geological modeling (e.g., [3,4]). The use of data-driven methods based on machine learning algorithms is a growing trend. These algorithms can recognize complex spatial patterns, learn from them, and autonomously indicate the presence of these patterns in specific areas [5,6]. The simultaneous application and comparison of results obtained with different methods at exploration boundaries are also pervasive in the specialized literature [7,8,9,10,11].

Mineral prospectivity studies have a conceptual basis in the analysis of knowledge about mineral deposits, their environments, and the factors influencing their formation. The mineral systems approach [12] provides a comprehensive view of mineral deposit formation and evaluates critical factors in their formation, including source, metal transport, deposition, and post-depositional enrichment processes. Furthermore, geological processes operate at various spatial and temporal scales, encompassing common aspects of different types of deposits. Mineral systems are divided into orthomagmatic, hydrothermal, and sedimentary systems [13], which can overlap, considering the tectonic and geodynamic evolution of the region.

The Tapajós Mineral Province (TMP), the focus of this study, is one of Brazil’s most important polymetallic provinces, particularly known for gold deposits hosted in granitoid bodies. It is a vast region with high economic potential but limited geological and metallogenetic knowledge (greenfield). The TMP lacks detailed geological knowledge and analyses of mineral potential using modern data and methods to identify potential mineral targets. Indicative work on favorable zones for gold mineralization in the TMP has shown promising results [14,15,16,17]. With the increasing development of geotechnologies, coupled with the addition of new data characterizing different deposits, predictive models can be improved to indicate potential areas with greater precision.

In this context, this study aims to predict prospectivity on a semi-detailed scale, indicating areas favorable for gold mineralization related to intrusions in the central portion of the TMP, using knowledge-driven, data-driven methods, and machine learning algorithms to compare them and contribute to the definition of new strategies for mineral exploration in the region.

2. Geological Setting

The TMP is located in the central-southern part of the Amazonian Craton, corresponding to the Tapajós domain [18,19] of the Tapajós-Parima (or Ventuari-Tapajós) tectonic province [20,21] (Figure 1a,b). The region consists of Archean and Paleoproterozoic terranes that were tectonically stabilized around 1.0 Ga [22,23]. Their origin is associated with the craton’s structuring through accretion events and crustal shortening of mobile belts and magmatic arcs predominantly oriented NNW-SSE relative to an Archean protocraton [20,21,22]. In the southwest of the TMP, basement rocks outcrop, and are metamorphosed from low to high grade and with variable ductile deformation. In the southeast and central-south regions, granitoids occur with weak to moderate ductile deformation, which is related to the NW–SE shear zones. In the eastern and northeastern portions, rocks have mainly undergone brittle deformation. The gradual variation in the exposed crustal level in the province is also observed through variations in mineralization controls [14,24,25,26].

The tectonic evolution of the TMP occurred during the Paleoproterozoic. The lithological associations in the study area are illustrated in Figure 2. The basement of the province is primarily composed of rocks from the Jacareacanga Group (2.1–2.87 Ga) and the Cuiú–Cuiú Complex (2.03–2.0 Ga) [20,30], and is related to the early stages [31] of the formation of the Cuiú–Cuiú magmatic arc [14,32,33,34,35]. Among the oldest units are (i) the Comandante Arara Formation (ca. 2.0 Ga) [36], which has a more evolved calc-alkaline signature and represents the maturation of the arc in distal regions [37], and (ii) the Castelo dos Sonhos Formation, with auriferous metaconglomerates and metarenites of fluvial origin, deposited between 2.01 and 2.05 Ga [38,39], outcropping in the extreme southeast of the TMP and believed to be remnants of a rift or foreland-type basin [31,37].

Following this phase, more distal magmatic events occurred, characterized by high-potassium calc-alkaline to shoshonitic signatures, which generated rocks of the Vila Riozinho Formation and the Creporizão Suite batholiths between 2.0 and 1.95 Ga [35,37,40,41]. The São Jorge Antigo (1.98 Ga, [40]) and Pindobal (1.94 Ga, [42]) granites are correlated with this suite, whose batholiths are elongated due to emplacement along the NW-SE shear zones.

Between 1.90 and 1.89 Ga, magmatic pulses produced the granitoids of the Tropas Suite [31] and the São Jorge Jovem granite [43]. The granitoids of the Tropas suite have a less evolved calc-alkaline signature. They may represent a possible island arc related to subduction [18,30] or a late post-collisional evolution relative to the Cuiú–Cuiú orogeny [14,35].

This episode was followed by (i) significant post-orogenic granitogenesis, represented by the Parauari Suite granitoids (1.89–1.87 Ga, [20,40]), (ii) intracontinental basic magmatism with intrusions of gabbroic bodies from the Ingarana Suite and their equivalents (1.88 Ga, [31]), and (iii) late-to-post-tectonic anorogenic magmatism in an extensional setting (Uatumã magmatism—1.89–1.87 Ga), with alkaline (Maloquinha Suite), felsic (Salustiano, Aruri, and Moraes Almeida Formations), and intermediate to mafic (Bom Jardim Formation) compositions [31,44]. Most of the granitoids are controlled by regional NW–SE-trending faults, and some mafic bodies are elongated in an E–W and NW–SE direction, characterized by faulted contacts with other units [31,37].

The geological evolution of the province settles with the establishment of intracratonic estaterian sedimentary basins (Novo Progresso and Coatá Formations) associated with alkaline magmatism (Crepori Diabase, Porquinho Granite) and a mafic magmatism event around 1190 Ma, after a hiatus of about 600 Ma (Cachoeira Seca Suite), composed of olivine gabbros, olivine gabronorites, olivine diabases, and troctolites [31,45,46].

Structures, Metallogeny, and Gold Mineralization

The structural framework of the TMP suggests mutually compressive and extensional deformation, occurring in both ductile and ductile–brittle to brittle regimes. The regional structural orientation is primarily NNW–SSE to NW–SE [47]. Santos [48] identified three deformation and thermotectonic events involving the mobilization and accretion of magmatic arcs on a regional scale during the Paleoproterozoic.

The first event (2005–1997 Ma [47]) generated thrust faults with a NNE–SSW orientation and NW–SE, N–S, and NE–SW lineaments, which were overprinted by the subsequent event. The second event (1970–1950 Ma [47]) produced brittle shear zones with primary orientations of NW–SE and NNW–SSE, which are the most prominent in the TMP. The third event (ca. 1900 Ma [30]) created ductile–brittle shear zones with a NNE–SSW orientation, accompanied by auriferous quartz veins. Progressive deformation altered and reactivated NW–SE and WNW–ESE shear zones. A younger extensional brittle event, related to NW-SE shearing [26,48], is manifested as extensive E–W and ENE–WSW lineaments in the central-eastern part of the TMP. Subsequent deformation events produced brittle structures and/or were associated with reactivations of older structures, favoring the intrusion of Cambrian mafic dikes [47].

The TMP encompasses numerous magmatic–hydrothermal deposits of gold associated with Ag, Cu, Pb, Zn, Bi, and Mo, with mineralizations of various types and styles, representing distal, proximal, or dome portions, which are deep or shallow parts of a structurally controlled mineral system. The primary mineralization occurs in quartz veins and, secondarily, in stockwork and disseminations, with all three forms eventually occurring simultaneously (e.g., São Jorge and Cantagalo). Gold is associated with quartz and sulfides (mainly pyrite and chalcopyrite, and subordinately sphalerite, galena, and pyrrhotite), either as sub-microscopic inclusions or in microfractures (e.g., Tocantinzinho, Batalha, and São Jorge). The deposits typically have well-developed hydrothermal alteration halos (e.g., São Jorge, Palito), with fissure alteration (e.g., Cuiú–Cuiú, Batalha), and sericitic alteration zones are most directly associated with mineralization [47]. Ore bodies are primarily located within structures such as shear zones and have a strong spatial correlation with the emplacement of mafic dikes along these structures at contact zones.

3. Materials and Methods

To assess methods for predictive mineral prospectivity mapping in the TMP, this study involved the following stages: (i) selection of the dataset, (ii) data processing and generation of evidential themes, (iii) prospectivity modeling using various integration techniques, and (iv) model validation. Figure 3 illustrates the details of these stages.

The approach to characterizing mineral deposits was developed following the concept of a mineral system, which evaluates the significance of numerous geological processes for the formation and preservation of deposits, across all scales of observation [12,49,50,51,52].

3.1. Dataset

The data primarily come from the Geological Survey of Brazil and include geological and mineral resource maps, aerogeophysical data (radiometrics and magnetometry), and remote sensing data (digital elevation model—DEM). A total of 21 primary gold occurrence points were selected for training and testing the prospectivity models, with the inclusion of 21 non-deposit points generated randomly. These non-deposit points were distributed at a minimum distance of 12 km from any known deposit point and 7 km from each other. Such control criteria are employed to ensure that the marked points are not coincident or too close to known deposit locations. This is based on the assumption that locations near existing mineral deposits are likely to have similar multivariate spatial data signatures as deposit sites, which could jeopardize the results.

Table 1. Data summary for the main gold prospects in the central portion of the TMP.

Occurrence	Mineralization Style	Alteration	Structural Control	References
Abacate	qz veins	sil, tur	N80E/70SE	[32]
Água Branca	qz veins	sil, carb, sulf, ser, cl	88/125, 35/290, 65/315	[58,59]
Asa Branca	qz veins	sulf, epid	88/175, 88/125	[14,58]
Batalha	stockwork, qz veins	ser, potas, alb, epid, sil, cl	70/315, 90/300	[32,60,61]
Boa Esperança	qz veins	arg, ser, cl, sulf	220/75SW, 80/80SE	[14]
Cantagalo	qz veins, stockwork, disseminations	sulf, lim, sil, kaol, arg, ser	85/060, 90/055, 80/350	[32,34,62]
Cuiú–Cuiú/Central	qz veins, hydrothermal breccia	ser, cl, carb	75/075	[58,63]
Cuiú–Cuiú/Moreira Gomes	stockwork, veins	ser, sulf, carb, cl, sil, epid	31/305, 10/341	[64,65]
Carneirinho	stockwork	sulf, sil, ser, epid, potas	N80W/85NE, N15/85SE	[14,62]
Davi	qz veins	ser, carb, sulf, epid, cl, potas	90/135, 90/305	[32,62]
Mamoal	mafic dykes disseminations	sulf, potas	88/020	[58,62]
Ouro Roxo	disseminations, py and qz-py veins	ser, cl, carb	65/090, 35/095	[32,34,66]
Palito	veins	ser, sulf, epid, cl, carb, potas	88/045	[58]
Pimenteiras	veins	sulf	90/095	[34]
São Chico	veins	ser, cl, sulf, epid, musc, kaol, sil	50/170, 88/170	[58]
São Domingos/Fofoca	qz-sulphide veins, stockwork	ser, sil, sulf, epid	88/165, 72/320	[58,67]
São Domingos/Tucano	stockwork, qz-sulphide veins	ser, sil, sulf, epid	60/75NW	[58,68]
São João	qz veins	epid, arg, sulf	N45–65E, N30E/75SE	[37]
São Jorge	qz veins, stockwork disseminations	sulf, musc, cl, carb	80/160	[58,69]
Sucuba	qz veins	sil, cl	EW	[37]
Tocantinzinho	qz veins, stockwork	mic, cl, ser, sil, carb	80/125, 80/345	[58,70]

py = pyrite; qz = quartz; sph = sphalerite; alb = albitization; arg = argilization; carb = carbonation; cl = chloritization; epid = epidotization; kaol = kaolinization; lim = limonitization; mic = microclinization; musc = muscovitization; potas = potassification; ser = sericitization; sil = silicification; sulf = sulfidation; tur = tourmalization.

shows the characteristics of the main prospects in the study area, which are used here as training points. Although the number of training points is small relative to the analyzed area, balancing the dataset improves prediction accuracy, especially in machine learning-based classifiers, compared to imbalanced datasets. Carranza and Laborte [53] demonstrated the successful applicability of various mineral potential modeling tools, including ML methods that require a large amount of training data, using few deposit points (e.g., [3,54,55,56,57]).

3.2. Data Processing

The availability of data and the geological knowledge of the region are heterogeneous. Most of the known deposits are concentrated in the eastern part, while the central part is less explored. To enhance information and reduce subjectivity in the identification of new features important for the mineralization context, statistical and semi-automatic methods were employed.

Aerogeophysical data were integrated into a single database, and grids were interpolated using the minimum curvature method [71] into cells of 125 m. Radiometric channels were verified, and variables with inconsistent physical behavior for natural radiation sources were removed to reduce the impact of these values in mathematical operations between channels.

To map hydrothermal alteration zones, parameters (such as F), anomalous potassium (K_d), anomalous uranium (U_d), and the eTh/K ratio were calculated. The joint analysis of these parameters helps identify alteration zones associated with the presence of K-feldspar and/or sericite/muscovite, such as potassic and phyllic alteration. The F parameter (K × (eU/eTh)) highlights the simultaneous abundance of potassium and uranium in relation to thorium, and anomalous values are good indicators of altered rocks (Efimov cited in Gnojek and Přichystal [72]). This parameter allows the distinction between zones of hydrothermal alteration in strongly weathered areas and has been widely used in identifying mineralization related to these zones (e.g., [73,74,75,76]). The K_d and U_d parameters indicate anomalous concentrations obtained by normalizing K and eU with eTh, calculating the ideal values K_i ((K_mean/eTh_mean) × eTh) and U_i ((U_mean/eTh_mean) × eTh), and computing the deviation of measured values from these ideals (K_d = ((K − K_i)/K_i); U_d = ((U − U_i)/U_i)) [77,78]. The data are previously classified into gamma spectrometric domains with relatively homogeneous signatures. In heterogeneous areas, the calculation of K_i and U_i results in false enhancements. The eTh/K ratio has been used in mapping hydrothermal alteration zones (e.g., [79,80]) due to the observation that eTh mobility is abnormally low compared to K and eU in hydrothermal processes. It is a valuable exploration vector in magmatic–hydrothermal deposit settings [81], providing a more accurate indication of the extent of alteration and minimizing the influence of naturally K-rich felsic intrusions that can mask deposit signatures ([82], among others).

Magnetometric data were reduced from the International Geomagnetic Reference Field (IGRF), micro-leveled, and interpolated to generate the anomalous magnetic field (AMF) map. The reduction to the pole (RTP—[83]) filter was applied to the AMF map to reduce the influence of induced magnetic fields and to centralize anomalies over their sources. Due to instability in low magnetic latitudes [84], the RTP filter was first tested on synthetic 3D models (as proposed by Bongiolo [85]) and generated using the open source MATLAB-based GRAV MAG PRISM program [86], following the geometric and magnetic parameters shown in Figure 4. The evaluated prisms represent magnetic bodies at different depths in the crust, with geographic positioning and magnetic field conditions (i) in the study area, (ii) reduced to the pole in the study area, and (iii) at the Earth’s magnetic pole.

To verify the continuity of structures in depth, attenuate high spatial frequency signals, and minimize the influence of shallow sources, upward continuation filters were applied [87,88]. The following enhancement filters were applied and tested on the continued data: (i) first vertical derivative (1Dz—[89,90]); (ii) analytic signal amplitude (ASA—[91,92,93,94]); (iii) total horizontal gradient (GHT or THDR—[95,96]); (iv) tilt angle (ISA or TDR—[97]); (v) tilt angle of the total horizontal gradient (TAHG—[98]). These filters improve magnetic responses and highlight subtle anomalies more accurately. Additionally, magnetic worm products were generated on upward continuations to separate shallow high-frequency data from deeper low-frequency data, highlighting edge features. This product helps reduce ambiguity in interpretation and provides information about the overall shape and relative depth of edges, such as faults and geological contacts [99].

The DEM data were processed for the automatic delineation of linear features. Shaded relief images were generated at illumination azimuths of 0°, 45°, 90°, and 135°, which were subsequently overlaid on a multi-directional illumination map, enabling a three-dimensional perception, and minimizing false trends from artificial lighting. The “LINE” algorithm, part of the PCI Geomatica platform (2018, PCI Geomatics Enterprises Inc., Markham, Ontario, Canada), performs processing in three steps: (i) edge detection; (ii) thresholding; (iii) curve extraction. These steps result in the detection of linear patterns in an image and their automatic conversion into vector segments [100].

3.3. Mineral Prospectivity Modelling

The conception of prospectivity maps requires numerous classes of geoscientific information to be spatialized in a GIS environment and combined with various mathematical and logical operations. Mappable criteria indicate the critical parameters (geological processes) of the studied mineral system, which are translated into layers of evidence that result in the classification of promising target areas for future exploration [1,2].

Based on the critical components discussed by Hagemann et al. [13], three critical elements were adopted for mapping the footprints of the mineral system in the TMP: (i) the source of energy, metals, and fluids; (ii) fluid migration pathways; (iii) concentration and deposition processes. From these elements, and based on the theoretical understanding of the TMP’s gold magmatic–hydrothermal system, geological processes and their mappable criteria were established. Spatial proxies (evidence layers) were obtained and pre-processed as described in Section 3.2.

Mineral potential modeling can be achieved by an empirical (data–driven), conceptual (knowledge–driven), or hybrid approach [1,2]. Empirical approaches are applied in areas with limited and poorly distributed data and few known deposits, relying on prior geological knowledge of the study area. Conceptual approaches are used in well-explored areas with robust and well-distributed data, using known mineralized sites as control points. In this study, to examine and compare the results obtained from each approach, the evidence layers were combined using three different integration methods: a knowledge-driven (fuzzy logic) approach and two data-driven approaches, one probabilistic (weights of evidence) and one based on a machine learning algorithm (SVM).

Fuzzy logic [2,3,4,101,102,103,104] is a mathematical tool that converts the probability of a variable into a continuous series of possibilities using fuzzy membership functions, where values are distributed from a midpoint (membership) and spread from there. For the integration of evidence layers, fuzzy OR (logical union) and GAMMA (a combination of the fuzzy PRODUCT and SUM operators) operators were used. Details on fuzzy functions and operators can be found in [105,106].

The weights of evidence (WofE) method, or Bayesian logic, works with parameters estimated from the statistics of the data. Following the procedure described by [2], the method was used to define weights W+ and W− for each of the classes in the evidence layers and the contrast C, defined as the difference between the weights. The evidence layers with relevant weights concerning the deposits are combined, considering the interaction between the measured probabilities in the areas of the layers and the known deposits, both a priori and a posteriori. The method assumes conditional independence between events, and the probability of finding new ore bodies generally increases with the inclusion of new information and the calculation of weights.

The support vector machines (SVM) method [7,107] consists of a set of supervised learning algorithms based on the statistical learning theory [108]. The dataset used is labeled with known class labels, and the algorithm is trained to project an ideal linear hyperplane that optimizes the distance between the two closest sample points (support vectors) to separate multiple classes. The larger the distance between the support vectors, the lower the classification error [109,110].

To explain the method, consider a training set with two separable classes (e.g., mineralized and non-mineralized zones), with $l$ samples represented as $\left(x_1,y_1\right),\cdots ,\left(x_l,y_l\right)$, where $x$ belongs to an n-dimensional space, and $y$ is a class label equal to 1 for one class and −1 for the other class (i.e., $y\in \left\{-\mathrm{1,1}\right\}$) [111]. The classification hyperplanes satisfy the equation $y_i\left(w{x}_i+b\right)\ge 1$, where $i=\left[\mathrm{1,2},\cdots ,n\right]$; $w$ and $b$ are parameters of the hyperplane’s decision function [112]. The decision function is given by $f\left(x\right)=sgn\left(wx+b\right)$, where $sgn$ is a function equal to $1$, if $x>0$; $0$, se $x=0$; and $-1$, if $x<0$, defining margin hyperplanes ($1$ and $-1$) and the central hyperplane ($0$). The parameters $w$ e $b$ are obtained with $maximize f\left(w,b\right)=1/2{‖w‖}^2$, which is subject to $y_i\left(w{x}_i+b\right)\ge 1$. This optimization is constrained, and the solution is given by Lagrange multipliers, as per the equation ${L\left(w,b,\alpha \right)=1/2{‖w‖}^2-\sum _{i=1}^l{\alpha }_i(y}_i\left(\left(w{x}_i\right)+b\right)-1)$. This function is minimized with respect to $w$ e $b$ and maximized with respect to ${\alpha }_i>0$. The ${\alpha }_i$ multiplier is determined by the equation $maximize{\sum }_{i=1}^l{\alpha }_i-1/2{\sum }_{i,j=1}^l{\alpha }_i{\alpha }_j{y}_i{y}_j\left(x_i{x}_j\right)$, which is subject to ${\alpha }_i>0$ e ${\sum }_{i=1}^l{\alpha }_i{y}_i=0$. Thus, the rule for classifying based on the ideal hyperplane is determined by the function $f\left(x\right)=sgn\left({\sum }_{i=1}^l{y}_i{\alpha }_i\left(x{x}_i\right)+b\right)$.

Non-linear datasets are converted into linear ones by transferring input data to a higher-dimensional feature space using kernel functions. In this work, the radial basis function (rbf) kernel function was selected because it has lower error rates compared to others [107], and it requires simple parameters for geoscientific data applications, namely the parameter C (balancing errors, margin width, and the number of support vectors) and the parameter γ (related to the width of the distribution and optimized for better results).

To properly assess the prospective model’s performance, the F1 score [113] was used, considering both recall (R) and precision (P). R is the percentage of mineralized points classified as true positives (tp) relative to the total number of samples in the positive class (tp and false negatives—fn). P is the ratio of tp to the total number of samples classified as positive (i.e., tp and false positives—fp). The score is obtained by calculating the harmonic mean between R and P. The optimization of the learning of SVM algorithms is achieved by choosing the kernel function (in this case, rbf) to determine the ideal hyperparameters C and γ.

The resulting models from the three approaches were validated through graphs for visualization and organization of classifiers based on their performance (Figure 3). The receiver operator characteristic (ROC) curve presents the probability that favorable and unfavorable areas are correctly classified. The area under the curve (AUC) measures the model’s efficiency in correctly classifying training points [114,115]. The WofE model was tested with the conditional independence (CI) ratio and test [54], given by the ratio between the number of observed known deposits (n) and the predicted deposits (T), where the hypothesis that the difference T-n is null is admitted. High values indicate that the hypothesis of the data independence condition was not met, i.e., the smaller the values of this ratio, the better the model [54]. Additionally, success–rate and prediction–rate curves were used to indicate the cumulative percentage of deposits relative to the cumulative percentage of prospective areas.

4. Results

4.1. Hydrothermal Alteration Mapping with Gamma-ray Spectrometry

The K, eTh, and eU channel maps were used to individualize gamma spectrometric patterns. Each channel was statistically reclassified into concentration classes with thresholds defined by the difference between the mean and half of the standard deviation of the values (

Table 2. Classification parameters for gamma-ray spectrometric channel concentrations.

Parameters		K (%)	eTh (ppm)	eU (ppm)
Mean (M)		0.324491	19.620647	2.358534
Standard Deviation (σ)		0.166002	10.636025	0.932671
Class	Interval
Low	<M − ½σ	<0.24	<14.30	<1.89
Intermediate	>M − ½σ and <M + ½σ	0.24–0.41	14.30–24.94	1.89–2.83
High	>M + ½σ	>0.41	>24.94	>2.83

). The combination of these data resulted in a map containing 27 concentration classes, which provided better separation between adjacent regions with similar gamma spectrometric signatures. The lithogeophysical map in Figure 5 was generated from the joint analysis of the class map with the ternary K/eTh/eU map and the mapped geological units.

From this product, the parameters F, K_d, U_d and the eTh/K ratio were calculated, and were processed in each domain to prevent the masking of ideal K_i and U_i values with adjacent lithologies naturally rich in K. The resulting maps were integrated into mosaics (Figure 6a–d) using the stitching technique [116]. The F, K_d, and eTh/K maps served as the inputs for the ternary composition F-K_d-eTh/K in Figure 6e. Areas coinciding with mineralized sites exhibited signatures with high F and K_d values, high to medium U_d values, and low eTh/K ratios. Therefore, it was considered that zones identified with this pattern were potentially hosting a hydrothermal alteration.

4.2. Interpretation of Magnetic Lineaments

Considering the magnetic and geographic parameters applied to synthetic models, the following maps were generated: (i) AMF (Figure 7a), (ii) AMF reduced to the pole (Figure 7b), and (iii) the Earth’s magnetic pole (Figure 7c). These maps were subjected to enhancement filters, such as ASA (Figure 7d–f), THDR (Figure 7g–i), TDR (Figure 7j–l), and TAHG (Figure 7m–o). By comparing the prisms of the anomalous field in the study area and at the pole, it was observed that the positioning of anomalies did not undergo significant displacement, allowing the reduction to the pole to be applied to real data.

The reduction to the pole of real magnetic data generated artificial trends coinciding with magnetic declination, which was corrected with a centered cosine directional filter at 345° and degree 1. Upward continuation filters with distances of 500, 1000, and 2000 m were also applied to eliminate noise caused by the influence of shallow sources. The map continued to 1000 m and was selected for the application of the other filters.

From the AMF reduced to the pole, the magnetic worm map was produced, with delineating features at a set of upward continuations from 1000 m to 10,000 m, with an emphasis on higher-order features. The map shows linear features primarily in the NE–SW, WNW–ESE, and E–W directions, extending in depth, sometimes contouring batholithic bodies.

The magnetic lineaments interpreted on the TAHG reduced to the pole, along with the analysis of other enhancement products (Figure 8), reflect intermediate structures with a large lateral extent and spatial correlation in depth, indicating that these structures may have served as conduits for the ascent and percolation of hydrothermal fluids from deeper crustal zones, favoring the deposition of sulfides and the formation of gold deposits. The signal of NE–SW trends is well-marked at deeper levels, suggesting that this is the structural framework of the province’s basement. The WNW–ESE to NW–SE pattern is more prevalent at intermediate depths and is related to truncation with basement structures. Lineaments in the N–S and NNW–SSE directions are less distinguishable at the analyzed depths and are concentrated to the west of the area, with truncation relationships due to breaks in the magnetic signal.

4.3. Semi-Automatic Surface Linear Features

The multidirectional illumination map (Figure 9a) served as the input for the linear feature detection algorithm. The threshold values used were determined after various combinations to achieve a regular distribution of features, with dimensions and densities proportional to the working scale and suitable for the geological characteristics of the area. The result shows a higher density of linear features in the west and center of the area, with scattered features to the east (Figure 9b). The lineaments are predominantly distributed in the approximate NNW–SSE to NW–SE direction, consistent with regional structuring, confirming the possibility of a higher incidence of structuring in these directions at intermediate to shallow levels.

4.4. Data Integration

The data resulting from the processing described in the previous sections were converted into images using data conversion, reclassification, and presentation tools. The interpreted magnetic lineaments and intrusive contacts were scaled by their Euclidean distances, while surface lineaments were scaled by their density. The selected evidence for modeling included potentially hosting geological units, proximity to intrusive contacts, structural framework, and potential deposition sites (hydrothermal alteration zones).

The maps were normalized to the 0 to 1 range to proportionally assess the values of each evidence map at the pixel where the training points are located. The cumulative value of the evidence ranges from 49% to 65% of the total sum, with an average of 54%.

4.4.1. Fuzzy Model

For the modeling of the fuzzy prospectivity map, the evidence maps were combined using mathematical operators. The geological map was classified with respect to rocks most favorable for hosting mineralizations, such as the Creporizão suite and alike, the Tropas Suite (a batholith cut by a shear zone with hydrothermal alteration), and a granitic host of mineralization related to this suite. Contacts of intrusive bodies with ages between 1.95 and 1.88 Ga are representative of the plutonovolcanic events that drove magmatic activity and provided energy to the mineralization system [117]. The intrusion of mafic bodies from the Ingarana suite may have been responsible for transporting fluids, ligands, and metals from deep portions of the crust, with occurrences of deposits near these contacts. Distances up to 5000 m from these contacts were considered prospective.

Interpreted magnetic lineaments were organized into four groups. The first group (NW–SE trend) is related to a brittle/ductile deformation phase (1970–1950 Ma) that affected the rocks of the province and hosts a large portion of mineralizations in shear zones. The second (NE–SW trend) represents directions of secondary structures with ore deposition. The third (E–W trend) corresponds to a predominantly brittle late phase with extensional tectonics, marked with late mineralized veins. The fourth group (N–S trend) is restricted to the western part and denotes structures and shear zones that were important ore deposition sites for various known deposits and occurrences (prospective distances in

Table 3. Summary of parameters used in the fuzzy logic spatial modeling process.

Data Source	Evidence Map	Classifying Process	Prospective Thresholds	Fuzzy Operators	Integrative Operator
Geologic map, scale 1:100,000	Favorable host rocks	Data classified according to the frequency of occurrences and geochemical signature.	-	-	GAMMA, index 0.75
Geologic map, scale 1:100,000	Intrusive contacts	Proximity analysis of deposits with the contacts of intrusive bodies.	5000 m	-
Airborne magnetometry	Magnetic lineaments	Proximity to lineaments interpreted from magnetometry, which represent the magnetic signature of mineralized shear zones, secondary deposit controls, and mineralized late veins. Euclidean distance.	2000 m (NW) 1000 m (NE and EW) 2500 m (NS)	OR
Radar image	Density of surface traces	Density of lineaments extracted from DEM.	Densities > 1
Airborne radiometry	F parameter	Statistical classification of the values of parameters F, Kd, and Ud and the eTh/K ratio, which represent the signature of The hydrothermal alteration with positive a correlation with mineralization.	Values > M + 1σ	GAMMA, index 0.75
	Kd parameter		Values > M + 1σ
	Ud parameter		Values < M − ½σ
	eTh/K ratio		Values < 0.7

).

The Kernel density was calculated for the interpreted DEM lineaments to highlight regions with a cumulative influence of features. These were considered favorable and fuzzified using the fuzzy LARGE function.

Maps of hydrothermal alteration zones related to gold deposition were statistically classified, where the parameters F, K_d, and U_d have a direct relationship with the values and were fuzzified with the fuzzy MSLARGE function. The eTh/K ratio has an inverse prospective relationship, and the values were inverted using the fuzzy LARGE function for classification.

These themes, which represent the critical factors of the intrusion-related gold system, were integrated using fuzzy operators, resulting in intermediate themes that were once again combined to generate the fuzzy prospectivity model shown in Figure 10 (processing steps and classification criteria are summarized in Table 3).

4.4.2. WofE Model

The WofE modeling method was applied to assess prospective response based solely on data. The a priori probability was 0.000016, with an area unit of 0.015625 km² in a 125 × 125 m² cell. The following procedures were adopted in preparing the themes: (i) reclassification of evidence maps; (ii) principal component analysis (PCA) to analyze the evidence maps for potential correlations; (iii) calculation of the weights of spatial association between the classes of maps and deposits; (iv) analysis of the weight distribution in each map and selection of the maps; (v) integration of the evidence maps.

The classes in the evidence maps were reduced to improve the estimation of weights for each class, which may not be robust when the number of deposits is small [118].

The evidence maps derived from radiometric parameters underwent PCA to eliminate possible correlations between them. PC1, which explained about 85% of the data variance, was selected as the input for the model, representing the mapped hydrothermal alteration (

Table 4. PCA eigenvector matrix over the radiometric parameters. The values highlighted in blue indicate the evidence map with the greatest contribution in each PC.

	PC1	PC2	PC3	PC4
F	0.34639	−0.89043	0.29401	0.02670
Kd	0.20787	−0.17201	−0.70639	−0.65438
Ud	0.17715	−0.11325	−0.62031	0.75565
eTh/K	0.89745	0.40587	0.17258	−0.00790

) and was reclassified into a binary component.

In the next step, the evidence maps weights were calculated with a confidence level of 2 for associations classified as positive. The classes that achieved the best contrast values were selected to compose the prospectivity map, considering the consistency of weights for the intrusion-related gold model. From these results, 7 predictor variables were selected, and the posterior probability map was generated (Figure 11). The model mapped the main deposit sites and indicated prospective zones aligned with structural trends in the NW–SE direction, as well as several other sites indicated as potential zones, especially in the central-southern and northwestern portions of the area.

4.4.3. Support Vector Machine (SVM) Model

For the application of the SVM method, 11 key information layers previously defined and classified according to a priori knowledge were considered. The modeling consisted of five main steps: (i) composition of a dataset containing all predictor variables, represented by bands; (ii) automatic and random selection of deposits, separated into 75% for training and 25% for testing; (iii) conditioned generation of random non-deposit points, adhering to a distance of more than 15 km from any known deposit and a minimum distance of 7 km between them; (iv) training and parameterization of the algorithm; and (v) conversion of the output result into a map.

The grid search algorithm, available in the scikit-learn library [119], was used to select the best parameters for C and γ, with the following ranges chosen: (i) C [2, 2.1, 2.2, 2.3, 2.4, 2.5, 3, 3.5, 4]; (ii) γ [100, 10, 1, 0.1, 0.01]. The algorithm selected the best values based on the F1 score from a 5-fold stratified cross-validation. The values obtained were C = 2.3 and γ = 1.

Stratified k-fold cross-validation [120] randomly partitions the training point set into k subsets of approximately equal sizes. In the stratified variation, the k parts preserve the original percentage of deposit and non-deposit points. Then, the model is trained k times, using points in k-1 training parts each time. The points from the remaining part are set aside for model validation. The trained model was applied to the set of evidential variables, resulting in the map shown in Figure 12.

4.5. Validation

The models were tested using the ROC curve, which uses training points as actual positive and negative evidence that needs to be verified in the model. The AUC (area under the curve) is calculated, where a value of 0.5 indicates random model predictions, and a value of 1 indicates perfect classification. The steeper the ROC curve, the better the model’s performance in mapping real deposits and predicting potential zones. The fuzzy model had an AUC of 0.980 (Figure 13a), the WofE model had an AUC of 0.948 (Figure 13b), and the SVM model achieved an AUC value of 0.969 (Figure 13c), demonstrating high model efficiency.

The WofE and SVM models were also validated using success–rate and prediction–rate curves, which indicate the cumulative percentage of deposits relative to the cumulative percentage of prospective areas [121]. The WofE model correctly classified about 94% of training points within the top 15% of the highest posterior probability classes. In relation to test points, they all were mapped within about 5% of the area (Figure 13b). The SVM model correctly classified about 94% of training points within the top 10% of the area. A total of 100% of test points were mapped within less than 5% of the area (Figure 13c).

The assumption of conditional independence between predictor variables is often violated when producing a prospectivity map with the WofE method, although the degree of violation depends on the choice and number of maps used as predictors. To evaluate this condition, the CI ratio and the CI test were calculated [54]. The CI ratio obtained was 0.93, indicating a very low degree of violation, as values between 1 and 0.85 are acceptable for modeling geological scenarios [2]. The CI test indicated subtle conditional dependence, with a probability of conditional independence of 37.5%.

5. Discussions

The processing of aerogeophysical data to generate spatial proxies revealed geological features in regions with little or no field information, highlighting valuable information for interpretation. The maps of the parameters F, K_d, and U_d and the eTh/K ratio showed signatures of K enrichment in known hydrothermalized areas. The association with high values of F and K_d indicates potassium enrichment in areas with pervasive sericitic alteration. The low eTh/K ratios corroborate the low mobility of thorium in hydrothermal processes [122], accentuated by potassium enrichment in these altered zones. Favorable lithological units for hosting mineral deposits guided the hydrothermal alteration responses. The magnetic–structural framework allowed the identification of important features related to compressive and extensional deformational events that affected the region and led to the formation of deposits. The density of lineaments extracted from the DEM demonstrated a positive spatial association with regional lineaments and known mineralized sites.

The three approaches used for prospectivity modeling in the central portion of the TMP allow a mutual comparison, with highlights provided in

Table 5. Comparison between the integration methods used in this work.

	Fuzzy	WofE	SVM
Advantages	(1) Easy implementation and operator knowledge-based control.	(1) Allows analysis of the weight of each evidence map and selection of the best ones for integration.	(1) Greater robustness, and handles both linear and non-linear data; (2) generalization ability, even with few training and testing points; (3) effective separation of data with a maximum margin; (4) increases decision-making resolution.
Disadvantages	(1) Less precise propositions and data inputs being more susceptible to bias; (2) Many steps and parameters to define (e.g., membership functions, data classification rules, integration operators).	(1) Need to reduce the number of classes (reclassification of evidential maps); (2) the assumption of conditional independence burdened and invalidated the modeling numerous times, a problem which can be minimized with the use of PCA.	(1) Requires extensive knowledge of ML architectures and programming languages.

. The prospective regions correspond to the major deposits in the region. Further, in a few deposits, the indication of potential zones was of low value or absent. This may have been caused by (i) the relationship of deposits with other mineralization conditions or structures not considered in the models; and (ii) biased prior classification of any of the evidence maps, excluding the region where the deposit is located.

It is important to note that the prospective responses of the three methods are possible and subject to errors and biases. Therefore, they should be treated and used as the supporting materials in decision making.

In a final analysis, an agreement map (Figure 14) was generated, which considers the best results from the three approaches, combining the areas corresponding to the top 5% scores from each model. Regions with a higher density of overlapping areas were delineated as prospective areas of interest. The areas in the western part of the map (Figure 14a) are elongated and spatially aligned with ductile–brittle shear zones, ranging from N–S to NNW–SSE, which host gold deposits and control granitic and tonalitic bodies. In the central part of the area (Figure 14b), where geological mapping is scarce and primarily based on geophysical products, four areas of prospective interest stand out: one elongated and coinciding with an E–W structure, and three others aligned approximately in the NE direction. Although there is very little information to infer the cause of the high potential of these areas, it is presumed that they reflect the presence of mineral occurrences embedded in structures or reactivated shear zones with essentially brittle deformation, characteristic of a younger deformation event. In the eastern portion of the area (Figure 14c), the prospective areas are smaller and therefore more limited, with most being elongated in the NW–SE direction, coinciding with the main regional structuring of essentially brittle shear zones that control the emplacement of most granitic bodies in that region. Other areas indicated in the overlap of models adjacent to known deposits and mineralized fields are also of great importance and can be understood as extensions of the already discovered mineralized bodies or zones.

6. Conclusions

In this study, various methods to produce prospectivity maps were evaluated and compared. The results can be summarized as follows:

• Radiometric data enhancement is effective for mapping K-enrichment and identifying gold-related hydrothermal alteration zones based on higher intensities of the parameters examined.
• Magnetic data processing with enhancement filters produced satisfactory results for interpreting the structural framework, despite challenges in calculating RTP at low magnetic latitudes.
• Semi-automated extraction of linear features from DEM provided valuable information on morphostructural lineaments in the study area, which is difficult to access in a tropical zone with high vegetation and cloud cover. The distribution of features is consistent with the regional tectonic framework of the area, and the density reflects the higher incidence of structuring at intermediate to shallow crustal levels.
• Three different methods for data integration were used for prospectivity modeling in the central portion of the TMP, resulting in prospectivity sites indicative of the main deposits in the region.
• The fuzzy model (i) effectively identified potential targets, especially in the eastern portion, reflecting the data availability panorama; and (ii) mapped known mineralization sites reasonably, although some deposits had low or zero prospectivity scores. The model validation with an ROC/AUC curve of 0.980 demonstrates high confidence in the degree of randomness explored to map the mineralizing event.
• The WofE method (i) indicated elongated potential zones aligned with prospective structural trends; and (ii) mapped most of the known deposits in areas of higher probability and performed well with points not used in the modeling, attested by an AUC of 0.948 and an ROC curve demonstrating excellent model efficiency in mapping known deposits and predicting new potential targets.
• The ML algorithm (SVM) (i) presented better-defined prospectivity sites and (ii) mapped nearly all known deposits in areas of higher scores and performed even better with omitted points, achieving an AUC of 0.969—the closest to the best classification value. Although the training data set is limited and not ideal, satisfactory results using a limited training data set (<20) can be achieved (e.g., [8,53,123]). The advancement in the use of more sophisticated machine learning techniques, as well as the use of classical methods guided by knowledge or data, can substantially contribute to the risk reduction in mineral exploration and enable decision making through indirect spatial information complementary to field data.
• An agreement map combining the top 5% scores from each model pinpointed the best prospective interest areas, offering valuable insights for future exploration.