Geological mapping of copper deposits in the democratic republic of Congo through remote sensing data and machine learning

Tshanga Matthieu, Matthieu; Ncube, Lindani; Thamaga, Kgabo Humphrey

doi:10.3389/frsen.2025.1678991

ORIGINAL RESEARCH article

Front. Remote Sens., 14 January 2026

Sec. Image Analysis and Classification

Volume 6 - 2025 | https://doi.org/10.3389/frsen.2025.1678991

Geological mapping of copper deposits in the democratic republic of Congo through remote sensing data and machine learning

Matthieu Tshanga Matthieu¹*

Lindani Ncube¹

Kgabo Humphrey Thamaga²^†

¹Department of Environmental Sciences, College of Agriculture and Environmental Sciences, University of South Africa, Pretoria, South Africa
²Department of Geography, University of South Africa, Pretoria, South Africa

Introduction: This study aimed to identify hydrothermal alteration zones and structural features associated with copper mineralisation in the Musonoi region, Lualaba Province, Democratic Republic of the Congo (DRC), within the Central African Copperbelt, using remote sensing and machine learning (ML). The study responds to the need for cost-effective and scalable exploration approaches in structurally complex tropical terrains.

Methods: Multispectral satellite data from ASTER and Landsat 8 OLI, integrated with field observations and borehole information, supported the development of a predictive model. Principal Component Analysis (PCA), band ratios, and both manual and automated lineament extraction were used to enhance spectral and structural features. A lineament density map and hydrothermal alteration indices were produced and integrated with geological field data to verify the relationship between surface anomalies and subsurface mineralisation.

Results: Random Forest classification indicated strong mineralisation in zones with high lineament density, fault intersections, and chlorite and kaolinite alteration. The main controlling variables were lineament density at 33.6%, fault proximity at 31.0%, and hydrothermal alteration index at 26.3%. Siliceous laminated rocks and basal dolomitic shale hosts mineralised units. Field validation confrmed that the model reflects known deposits, showing the strength of remote sensing and machine learning for exploration in complex tropical terrains.

Discussion: The study highlights the novelty of integrating Random Forest with multi source geospatial information in a structurally complex tropical terrain, and shows that this approach provides a cost effective and scalable tool for exploration in the Central African Copperbelt and similar geological provinces. Limitations related to spatial resolution and training data coverage remain and should be addressed in future work.

1 Introduction

Copper deposits around the world are currently estimated to be 720 million tons, while the remaining resources are anticipated to be worth roughly 5.6 million tons (Ali et al., 2024). Chile produces 27% of global copper, followed by the DRC, which has led Africa since 2012 (Mudd and Jowitt, 2018; Shengo et al., 2019). The Musonoi area in Lualaba Province is one of the DRC’s most promising copper-cobalt zones, with much mineralisation concealed below the surface, making it ideal for improving exploration strategies.

Copper’s corrosion resistance, conductivity, and ductility make it essential across sectors. Consumers use ∼40% for electricity generation, 13% in electronics (Sekandari et al., 2020), 13% in transport, and 25% in construction (Kenzhaliyev et al., 2025). The remaining 9% supports manufactured goods, coins, jewelry, instruments, and art (Liu et al., 2023). Rising demand increases the need for effective exploration, with Machine Learning—especially Random Forest—integrating geological and remote sensing data to identify copper-rich zones.

Copper occurs in structurally complex settings (Tshanga et al., 2024), which makes traditional mapping and geophysical work costly, slow and difficult in remote areas (Sang et al., 2020). Yet precise mapping remains essential for locating economic deposits (Agrawal et al., 2024) and for understanding structural controls on mineralisation (Parsa et al., 2018).

Remote sensing supports geological surveying, mapping, and interpretation (Shirmard et al., 2022). It identifies ore-related features and mineralisation processes (Takodjou Wambo et al., 2024) and overcomes terrain and access challenges using multi-resolution satellite imagery (Wu et al., 2023).

Machine learning techniques such as Random Forest, Support Vector Machines, and Convolutional Neural Networks are increasingly becoming central to mineral exploration workflows (Takodjou Wambo et al., 2024; Mahnaz et al., 2023). These approaches are well suited for the analysis of extensive satellite datasets such as ASTER and Landsat 8 and for the identification of patterns that connect surface features with deeper faults and fractures. Their strength comes from their capacity to handle complex, multidimensional information and to capture the nonlinear relationships that shape mineral formation processes. This analytical capability improves exploration accuracy and contributes to more sustainable mineral development practices (Mahboob et al., 2024; de Oliveira and Bertossi, 2022; Takodjou Wambo et al., 2020).

Significant copper and cobalt resources have been confirmed in the region, indicating strong long-term mining potential (de Oliveira and Bertossi, 2022; Pour et al., 2023). However, despite this confirmation, it was also found that there is still limited reliable data available on this major copper deposit, and this lack of sufficient data is restricting the understanding of how the surface features and the subsurface features are interacting together to control the processes of copper mineralisation (Williams and Bornhorst, 2023). The novelty of this present study is that it is combining ASTER-derived hydrothermal alteration indices together with Landsat-derived lineament density, and both of these are then integrated with Random Forest modelling in order to jointly evaluate the roles of structural lineaments, alteration zones, and also structural geology parameters. While some similar approaches have been applied in other regions, it was observed that these kinds of studies are still very rare in structurally complex tropical regions, and moreover, such research has never yet been carried out in the Democratic Republic of Congo. This fact is making the contribution of the present study unique and original in its context.

The study aims to develop a predictive geological model for copper mineralisation in Musonoi using remote sensing and machine learning, improving targeting in Musonoi and supporting more effective exploration across the Copperbelt.

2 Geological setting

The study area lies between 10°38′50″–10°4220″S and 25°24′10″–25°29′50″E (Figure 1), with elevations from 1,138 to 1,562 m. It consists of low hills underlain by silicified, weather-resistant dolomites. Erosion of dolomites and siltstones has produced gentle slopes and shallow valleys that drain to the Musonoi and Dilala rivers. To the northwest, the Dilala joins the Musonoi, which then flows into the Lualaba River (Cheyns et al., 2014).

Figure 1

Map showing territories in a region with major towns marked in red, including Kolwezi, Fungurume, Likasi, and Lubumbashi. Green areas indicate protected zones. An inset highlights the region within the Democratic Republic of the Congo. Roads and territorial boundaries are depicted. Scale and coordinates are provided.

Figure 1. Locality of Musonoi copper deposit (MCD).

2.1 Stratigraphy

The Musonoi copper deposit lies in the Neoproterozoic Katangan Belt within the ∼10 km-thick Katanga Supergroup extending across northwest Zambia and southeast DRC (Kampunzu et al., 2009; Cailteux and De Putter, 2019). This sequence rests unconformably on basement rocks of the Ubendian Belt—high-grade gneisses and metasediments—and the Kibaran Belt, which contains low-grade siliciclastics and pelitic units intruded by granites and mineralised pegmatites (Mambwe et al., 2023; Yantambwe and Cailteux, 2019). The Katanga Supergroup comprises the Roan, Nguba, and Kundelungu Groups, separated by two major glacial deposits: the Grand Conglomérat at the base of the Nguba Group and the Petit Conglomérat (Kiandamu Formation) at the base of the Kundelungu Group (Turner et al., 2023; Cailteux et al., 2005) (Figures 2, 3).

Figure 2

Geological map illustrating the northwest Congolese Copperbelt, highlighting key features such as major and minor copper deposits, faults, and various rock groups. Notable areas include the towns of Kolwezi, Fungurume, and Tenke. The map shows different geological formations, with a study area marked by a red rectangle. A legend explains symbols for rock types, deposits, and geological structures. The map includes a scale bar and directional arrow, covering the region near Zambia.

Figure 2. Geological Map of the Central African Copperbelt showing the location of the Musonoi deposits. After François (1974), Lepersonne (1974), Kampunzu and Cailteux (1999), Kipata et al. (2013), Dupin et al. (2013), and Gecamines François (1990).

Figure 3

Stratigraphic summary chart of the Central African Copperbelt, detailing various geological subgroups and age groupings from the Archaean to Karoo periods. The chart distinguishes regions like the Zambian Copperbelt, Domes Region, and Congolese Copperbelt, with remarks on geological events. Certain units with copper mineralization are highlighted in red, while transgressive lead-zinc-copper regions are marked with blue. Geological formations and subgroups are organized chronologically, with significant events such as the Lufilian Orogeny and Irimide Granites noted.

Figure 3. Lithostratigraphy of the katangan supergroup (Cailteux and De Putter, 2019).

The Roan Group within the Katanga Supergroup is mainly dolomitic and was deposited in a mix of supratidal, intertidal, and lagoonal environments. As a result, its lithology is quite varied, ranging from conglomerates, arkoses, and quartzites to dolomitic shales, siltstones, and dolostones, the latter often occurring alongside anhydrite (Sośnicka et al., 2019; Hendrickson et al., 2015; Cailteux and De Putter, 2019).

Geologists recognise four subgroups within the Roan Group: the Mwashya, Fugurume (formerly Dipeta), Mines, and R.A.T. (Roches Argilo-Talqueuses) subgroups (Dewaele et al., 2006).

2.2 Local geology

The Musonoi copper–cobalt deposit lies within the Kolwezi Nappe, a northeast-trending synclinal basin with minor and major axes of about 10 km and 20 km. Copper-bearing Series des Mines lithologies occur as structurally dismembered fragments that locally preserve complete successions, with variable dips and strikes both within and between rafts. These rafts are tightly folded, showing complex structural patterns and abrupt changes in orientation. Tectonic deformation is mainly distributed in a near-EW direction, and fracture planes typically dip southward at 25°–45° (Von der Heyden et al., 2023; Mambwe et al., 2023) (Figure 4).

Figure 4

Geological map showing the orebody distribution of the Kansanshi mine area. Various colors represent geological formations: yellow for R.A.T. Subgroup, blue for Mines Subgroup, pink for Kamoto Formation, and others. Black lines indicate thrust and fault lines. The map includes labels for regions such as KTO, Mashamba, and Kananga. A scale bar denotes distances in kilometers.

Figure 4. Geology and mineral deposits of the Kolwezi Klippe, Katanga Province, DRC (After: Jackson et al., 2003; François, 1990).

3 Materials and methods

3.1 Field data collection

A stratified random sampling approach ensured representation of major lithological units and alteration zones. In total, 187 sampling points were recorded using a Garmin eTrex GPS (<3 m accuracy). Seventy geological features were mapped at surface and 117 from underground borehole data, then consolidated into 80 ground control points to avoid redundancy. Field observations described local geology, mineral composition, structures, and alteration (Sekandari et al., 2020). Structural analyses were used to verify whether subsurface features experienced the same tectonic processes identified by remote sensing, with X–Y coordinates collected at each point (Togaev et al., 2022).

Borehole data were standardised through lithological coding and core logging, then merged with surface mapping by coordinate matching and cross-sectional correlation. Copper grades ≥1.0% were labelled as mineralised, <0.5% as non-mineralised, and intermediate values excluded. The dataset was split 70% for training and 30% for validation (Guha et al., 2024; Bahrami et al., 2021).

3.2 Image acquisition and preprocessing

ASTER and Landsat-8 OLI imagery (Figure 5) from the USGS portal supported field data (Table 1) (Zaman et al., 2024; Tobi et al., 2022). ASTER VNIR (15 m) aided alteration mapping, while Landsat-8 provided regional lithological and structural information (Khaleghi et al., 2020; Liu et al., 2023; Zhang et al., 2016). A cloud-free ASTER scene (9 July 2006) and a Landsat-8 image (20 May 2020, 1.09% cloud cover) were selected to minimise vegetation effects (Kulkarni et al., 2023; U.S. Geological Survey, 2022; U.S. Geological Survey, 2023).

Figure 5

Graph depicting atmospheric transmission percentages across wavelengths, labeled with ASTER, Landsat 7, ETM+, OLI, and TIRS sensor bands. VIS, SWIR, and TIR regions are marked with respective spatial resolutions, showing transmission variability at different spectral bands.

Figure 5. Electromagnetic spectrum ASTER and Landsat OLI/TIRS bands.

Table 1

Table 1. Details of acquired satellite images.

Landsat-8 supplies nine spectral bands at 30 m, thermal at 100 m, and a 15 m panchromatic band (Ourhzif et al., 2019; Ngassam Mbianya et al., 2021), with a long-term archive valuable for geology (Anderson et al., 2002; Yuan and Niu, 2008). To harmonise datasets, only overlapping spectral ranges (ASTER 1–9; Landsat-8 2–7) were used. Spectral similarity was checked using principal components and cosine similarity; band adjustment factors and cross-calibration reduced differences to <5% (Table 2) (Yuan et al., 2025; Anderson et al., 2002).

Table 2

Table 2. Description of the Aster and Landsat 8 Oli sensors.

Pre-processing used the FLAASH algorithm in ENVI for atmospheric and radiometric correction. Terrain-specific corrections (C-correction, Minnaert, slope–aspect) were not applied; this was acceptable due to low cloud cover, though some terrain effects remained (Zaman et al., 2024; Thamaga et al., 2022). Slope and aspect from the ASTER GDEM were added as Random Forest predictor variables to indirectly account for topography. Future work will apply explicit DEM-based corrections (Yu et al., 2024; Liu et al., 2023; Zhang et al., 2016).

Principal Component Analysis (PCA) was used to reduce multispectral data to its most informative components, helping highlight structural features such as lineaments that may indicate faults or fractures. Hydrothermal alteration zones linked to copper mineralisation were mapped using ASTER imagery due to its strong spectral sensitivity to alteration minerals (Estornell et al., 2013; Liu et al., 2023).

3.2.1 PCA application

Principal Component Analysis (PCA) reduces multispectral data by transforming original spectral bands into a smaller set of components that retain most geological information (Safari et al., 2017; El-Omairi et al., 2024). ASTER and Landsat images were radiometrically adjusted, cloud-masked, and standardised using z-score normalisation in ArcGIS 10.8 before PCA (Pietrzyk and Tora, 2018; Sekandari et al., 2020). Eigenvector loadings were evaluated to determine which PCs highlight target geological features (Mahboob et al., 2024; Takodjou Wambo et al., 2020).

3.2.1.1 Landsat imagery

PCA was applied to Bands 2–7 for their geological relevance, while Band 1 was excluded due to atmospheric sensitivity, Band 8 for mismatched 15 m resolution, and Band 9 for limited geological use (Masoumi et al., 2017; El Atillah et al., 2018). PCA composites helped distinguish lithologies and visualise faults, joints, and structural patterns (Duz et al., 2023; Wu et al., 2023; Sothe et al., 2017) (Table 3).

Table 3

Table 3. Basic statistic and PCA results of Landsat image bands.

3.2.1.2 ASTER data

PCA focused on VNIR and SWIR Bands 1–9 for detecting alteration minerals such as chlorite, ferrous silicates, and kaolinite linked to copper mineralisation, while TIR Bands 10–14 were excluded due to coarse resolution and poor alteration performance in tropical terrains (Rezaei et al., 2020; Sekandari et al., 2020; Gojiya et al., 2023) (Table 4).

Table 4

Table 4. Basic statistic and PCA results of Aster image bands.

To compare PCs from both sensors, ASTER Bands 1–9 were matched to Landsat Bands 2–7. A cross-similarity matrix using cosine similarity measured alignment between sensors (Figure 6).

Figure 6

Cosine similarity matrix comparing LANDSAT and ASTER principal components (PCs). The matrix shows color-coded values ranging from red for higher similarity to blue for lower. PC1_LAN vs. PC1_AST has the highest similarity at 0.93, while PC5_LAN vs. PC5_AST has the lowest at -0.86. A color bar on the right provides a gradient scale from -0.8 to 0.8.

Figure 6. Cross-similarity matrix.

PCA was performed on the six Landsat-8 bands to reduce data complexity and extract the most meaningful principal components (PCs). The eigenvectors (Table 3) show that PC1 mainly reflects overall image brightness, while PC2 separates vegetation from geological features and was therefore not central to this study (Baisantry and Sao, 2019; Ourhzif et al., 2019). PC3 highlight structural features like lineaments, faults, and fractures, showing strong negative loading in band 4 (red) and positive loadings in bands 6 and 7 (SWIR1, SWIR2), which are sensitive to lithological variations. PC6 shows slight changes in surface materials, marked by a strong negative value in band 4 (red) and positive values in bands 5 (NIR) and 6 (SWIR1), which can help detect moisture levels and specific rock types (Ali et al., 2024; Pietrzyk and Tora, 2018; Shi et al., 2015).

PCA was applied to six Landsat-8 bands to reduce data complexity and highlight key principal components (PCs). Eigenvectors (Table 4) indicate that PC1 reflects overall brightness, while PC2 separates vegetation from geology and was not central to the analysis (Baisantry and Sao, 2019; Ourhzif et al., 2019). With negative Red and positive SWIR loadings representing lithological changes, PC3 was best for spotting lineaments, faults, and fractures. PC6 showed moisture and rock type-related surface variations (Ali et al., 2024; Pietrzyk and Tora, 2018; Shi et al., 2015).

The cosine-similarity matrix comparing Landsat and ASTER PCs in Figure 6 highlights spectral relationships between structural features and alteration zones. PC1 from both sensors shows very high similarity (0.9286), indicating shared lithological patterns. Strong negative correlations, such as PC1_LAN vs. PC4_AST (−0.7750) and PC5_LAN vs. PC5_AST (−0.8615), suggest each sensor emphasises different alteration signatures. A similarity of 0.7939 between PC6_LAN and PC6_AST shows consistency in detecting weaker signals linked to secondary alteration minerals. Overall, complementary matches and contrasts indicate that combining Landsat and ASTER improves structural and alteration interpretation.

3.2.2 Band ratios

Band ratioing (BR) is one of the most powerful, simple and common band math procedures for mapping hydrothermal alteration zones. The approach works by enhancing or emphasising the anomalies of target objects. BR minimises the effect of topography, and therefore, augmentation of the differences in spectral responses of each band (Zhang et al., 2016; Liu et al., 2023; Shirmard et al., 2022). Different surface materials have been mapped using ASTER data band ratios, which reflect their distinct absorption and reflection properties in the electromagnetic spectrum (Guha et al., 2024). The technique was used to identify specific mineral spectral signatures. The selected bands for the band ratio classification need to show contrasting responses within the absorption features of the target mineral or mineral group’s reflectance curve (Bahrami et al., 2021; Ngassam Mbianya et al., 2021).

In this research, ASTER band ratios were applied to map hydrothermal alteration minerals associated with the copper deposit. The ratios relied on mineral absorption properties in ASTER and USGS spectral libraries and field knowledge of the Katanga Copperbelt, especially the Musonoi deposit. No field spectrum was acquired due to logistical limits, but reference data and regional geology were sufficient (Kokaly et al., 2014; Abdelkareem et al., 2024). The following ASTER band ratios were used based on diagnostic strength:

• Band 5 over 4 highlights ferrous iron content related to hematite and goethite, which are typical oxidation products near hydrothermal zones

• Band 7 over 5 enhances kaolinite-rich zones as Band 7 captures absorption near 220 μm, where kaolinite strongly contrasts with Band 5

• Band 7 plus Band 9 over Band 8 shows Mg OH minerals like chlorite and epidote found in propylitic alteration

Band ratio images Improve contrast between features via adjusting brightness at peaks and troughs on the reflectance curve after atmospheric correction. Spectral band ratioing improves mineral composition detection while reducing other surface information (El-Omairi et al., 2024; Wang et al., 2022).

3.2.3 Lineament extraction with PCI geomatica

Lineaments from Landsat imagery were extracted using PCI Geomatica because faults and fractures indicate structural controls (Saed et al., 2022; Pandey and Sharma, 2019). A two-stage workflow combined manual digitisation of PCA and SWIR composites with automated extraction (Adhab, 2019; Sedrette and Rebai, 2020). Edge detection used directional filters (0°, 45°, 90°, 135°) (Agbebia and Egesi, 2020), and GIS mapping visualised structural relationships. In total, 89 lineaments were digitised and 203 extracted automatically (Figure 7). Accuracy was validated against published geology and field observations (Figure 8). Shaded relief with various light angles minimized directional bias, and lineaments were verified against features from DEM and multispectral data. Final orientations matched regional tectonic trends (Cailteux and De Putter, 2019). The resulting lineament density map served as input for the Random Forest model to link structure and alteration (Ganguly and Mitran, 2023).

Figure 7

Flowchart illustrating the workflow for creating a predicted mineralisation map using input data and machine learning. It begins with input data, including field data collection and satellite image acquisition from USGS and NASA. The data undergoes preprocessing, such as radiometric calibration and geometric correction, leading to color composite creation. Subsequent steps include principal component analysis, lineament mapping, mineral indices, and alteration zones mapping. These processes contribute to creating a fracturation density map and structural geological analysis. The final output is a predicted mineralisation map, generated using a Random Forest Machine, validated with training and validation data.

Figure 7. Flowchart of the methodology.

Figure 8

Geological map of a region highlighting various geological features such as fold sense with anticlines, synclines, major structures, and bedding. The map shows mines or deposits, with labels like K.O.V., Kamoa, Mutanda, and Fungurume. Color coding indicates different geological units and rock types. The map is marked with coordinates and includes a legend for reference.

Figure 8. The Kolwezi subbasin portrayed as a thrust-bounded outlier rising above younger Kundelungu Group sedimentary layers and carrying Roan breccia (After François 1974).

3.2.4 Statistical computation

Optimum Index Factor statistical analysis was conducted to refine the interpretation of the geological structures. The Optimum Index Factor is a statistical method made to increase the information content (sum of standard deviations) while reducing overlap (correlation between bands) (Table 5) (Van der Werff and van der Meer, 2016). It was used to select the best combination of three Landsat bands for visual analysis.

The formula for OIF is : O I F = \frac{\sum_{k = 1}^{3} S K}{\sum_{j = 1}^{3} |r j|}

where k denotes the standard deviations of the combinations of the band, and $r$ represents the sum of the absolute values of the correlation coefficients.

Table 5

Table 5. Results of the OIF analysis.

In the Optimal Index Factor (OIF), the numerator is the sum of standard deviations of three selected bands, and the denominator is the sum of their absolute correlation coefficients, ensuring maximum information with minimal redundancy (Wu et al., 2023; Lin et al., 2021). Instead of using OIF results alone, selected band combinations were integrated with PCA layers for later modelling.

OIF-enhanced RGB composites highlighted alteration zones and lineament patterns, which were validated and used in the Random Forest model. Interpretation involved identifying colour anomalies from consistently behaving bands, guided by OIF calculations (Courba et al., 2023).

Table 5 shows the calculated OIF values for different spectral ranges (VISIBLE, NIR, and VNIR + SWIR) using Landsat TM bands 1, 2, 3, 4, 5, and 7, excluding the thermal band. The results indicate that the band combination 6-5-1 (band 6: SWIR1, band 5: SWIR, and band 1: Blue) has the highest OIF value, at 2009.93. This combination reveals fractures and lineaments more clearly than the surrounding areas. Compositions 7-6-5 and 6-5-1, representing bands 7, 6, and 5, and 6, 5, and 1, follow. These bands represent red, green, and blue channels (Lin et al., 2021; Ngassam Mbianya et al., 2021).

3.3 Machine learning (Random Forest)

Random Forest (RF), introduced by Breiman (2001), is an ensemble learning method that builds multiple decision trees from random subsets of data and combines their outputs, reducing overfitting and improving robustness to noise (Mahnaz et al., 2023). RF performs regression by averaging trees and classification by majority voting (Albon, 2018). In this study, RF classified structural geological features and remote sensing variables (Table 6), effectively handling complex, multi-source datasets for mineralisation mapping in Musonoi (Figure 9) (Nair et al., 2023).

Table 6

Table 6. Random forest feature, importance and geological relevance.

Figure 9

Bar chart showing the importance percentage of various geological factors. Alteration Index (Hydrothermal), Fault Proximity, and Lineament Density have the highest percentages, each over 30%. Lithology types such as Roches argileuses talqueuses, Shale dolomitique de base, and Roches siliceuses feuilletées exhibit much lower importance, all below 10%.

Figure 9. Feature importance ranking.

RF was chosen over k-Nearest Neighbors and Naïve Bayes because it handles high-dimensional, mixed-type datasets, resists overfitting, and needs minimal tuning. The model used 70% of data for training and 30% for testing, stratified by mineral presence, with 5-fold spatial cross-validation to limit spatial autocorrelation (Davies et al., 2025). Hyperparameters were tuned using a randomised grid search across number of trees (100–500), depth (10–30), minimum samples per split (2–10), and max_features (‘sqrt’ or ‘log2’).

Input data included ASTER alteration indices, Landsat lineament density, lithological classes, and interpreted structural lineaments (Singh et al., 2023). All rasters were resampled to 30 m, aligned, and clipped; lithologies were one-hot encoded and continuous variables scaled using Min–Max normalisation. Ground-truth samples were extracted via zonal statistics.

RF outputs a probability map (0–1), with a 0.6 threshold from ROC analysis defining high-potential copper zones (Kenzhaliyev et al., 2025). Alteration indices and proximity to faults were most influential, and SHAP values clarified variable importance. The final potential map in QGIS shows low, moderate, and high zones to guide field exploration (Taha et al., 2023).

4 Results and discussion

4.1 Mineral mapping

ASTER band ratio combinations (7 + 9)/8, 7/5, and 5/4 were applied to delineate chlorite, kaolinite, and ferrous silicate alteration zones, respectively, enabling their distinction based on spectral characteristics (Figures 10–12) (Yohanna et al., 2023; Canbaz and Ünal Çakir, 2022).

Figure 10

Chlorite ratio map showing grayscale variations representing concentrations over a specific geographical area. The map includes a scale in kilometers and latitude and longitude coordinates along the edges. A north directional arrow is present.

Figure 10. Chlorite ratio (7 + 9)/8 showing chlorite-rich zones in bright pixels.

Figure 10 presents a greyscale chlorite alteration map from the ASTER ratio (Band 7 + Band 9)/Band 8, highlighting chlorite-rich hydrothermal zones (Rowan et al., 2006; Saed et al., 2022). Bright areas in the central and eastern sections align with linear and curvilinear structures, indicating faults and folds that likely controlled fluid flow (Maarifa et al., 2024; Yuan et al., 2025). A strong chlorite anomaly in the north-central zone may mark a major alteration corridor, while darker areas indicate unaltered rock. Because chlorite is common in sediment-hosted copper systems and linked to propylitic alteration, these zones form key targets for field verification (Anwar et al., 2023; Samir et al., 2023).

The 7/5 band ratio map highlights kaolinite- or alunite-rich zones, where bright pixels show strong Al–OH absorption in ASTER Band 7 (2.26 μm) relative to Band 5 (2.17 μm) (Hosseini-Nasab and Agah, 2023; Liu et al., 2023). Bright areas in the southeast and central-east indicate advanced argillic alteration linked to acidic hydrothermal activity, often associated with epithermal systems (Madani and Emam, 2011; Orynbassarova et al., 2025). Linear bright zones may reflect structural controls such as faults that channel acidic fluids and alter feldspar to kaolinite (Alkashghari et al., 2020). Dark regions in the northwest and south-central areas likely contain unaltered minerals or propylitic alteration phases like chlorite. These anomalies may indicate shallow mineralisation and highlight targets for exploration, especially where structural features are present (El-Desoky et al., 2022; Yuan et al., 2025).

Figure 12 shows a grayscale image derived from the ASTER 5/4 band ratio, useful for detecting ferrous silicates and iron oxides linked to hydrothermal alteration in Cu–Co systems (Zaman et al., 2024; Deshpande et al., 2019). Bright areas in the southeast and east indicate high Fe²⁺, likely from oxidised alteration zones and sulfide breakdown associated with structurally controlled mineralised veins (Tobi et al., 2022; Hosseini-Nasab and Agah, 2023). Curved or patchy bright patterns suggest fluids moved along faults or fractures (Khaleghi et al., 2020; Ouhoussa et al., 2023). Dark zones in the northwest and west-central areas likely reflect minimal iron alteration, vegetation cover, or hard rock with weak signals. Combined with the chlorite (Figure 10) and kaolinite (Figure 11) maps, the 5/4 ratio supports a hydrothermal zoning pattern from chlorite-rich margins to iron-rich zones and kaolinite cores, typical of sediment-hosted Cu–Co deposits.

Figure 11

Kaolinite ratio map showing a grayscale distribution with varying intensity of kaolinite presence. The map includes coordinates from 25°15'0

Figure 11. The band ratio (7/5) highlights kaolinite-rich zones, appearing as bright pixels.

Figure 12

Map showing an Fe oxide Cu-Co alteration ratio with geographical coordinates marked on the borders. A scale in kilometers is provided and a directional arrow points north.

Figure 12. Highlighting ferrous silicate areas in bright pixels with the Fe oxide Cu-Co alteration ration (5/4).

Figure 13 shows a false-colour PCA image from ASTER data, which reduces redundancy and enhances alteration spectral contrasts (Zhang et al., 2016; Karifene et al., 2024). Blue to cyan tones in the south-central and northeast likely indicate chlorite, kaolinite, or iron oxides. These zones align with structural features such as faults and shear zones shown in Figures 10–13, confirming strong structural control on alteration (Takodjou Wambo et al., 2020; Saed et al., 2022; Gojiya et al., 2023). Cyan patches in the southwest may reflect iron-rich surfaces, gossans, or man-made features such as tailings (Ewais et al., 2022; Son et al., 2021). Dark areas in the south likely represent unaltered rock, vegetation, water, or shadow. Combining this PCA image with band-ratio maps helps separate altered and unaltered zones, improving target identification in structurally complex areas (Ouhoussa et al., 2023; Yuan et al., 2025).

Figure 13

Colorful hydrothermal map with coordinates ranging from 25°15'0

Figure 13. Map based on PCA technique highlighting hydrothermally altered areas in blue-cyan tones.

4.2 Lineament analysis

Lineaments represent faulting and fracturing related to mineralisation. A Landsat false-colour composite (bands 4, 6, 7) highlighted structural features, which correspond to faults, joints, shear zones, or lithological boundaries (Yamusa et al., 2018; Ibraheem et al., 2022; Pour et al., 2023). The lineament density map (Figure 14) was classified into nine density classes from very low (0–0.3539 km/km²) to peak density (2.8313–3.1852 km/km²).

Figure 14

Lineament density map showing various densities in a terrain, represented by colors from dark green to red. A legend at the bottom indicates density ranges, with lower densities in dark green and higher densities in red. Coordinates and a north arrow are included for orientation.

Figure 14. Lineament density map.

Landsat-derived orientations show dominant NE–SW and NW–SE trends, consistent with regional tectonic influences such as the Pan-African orogeny (Andarawus et al., 2023; Mouzoun et al., 2025). Figure 14 illustrates the structural geology of the study area, which controls subsurface features and copper distribution.

The map’s central and northeastern regions show high lineament concentrations, likely marking fault intersections, shear zones, or intense fracturing. Such zones commonly host ore bodies, veins, and altered rock formed by hydrothermal fluid flow (Mat Akhir, 2004; Liu et al., 2023). These structurally complex areas are strong targets for copper–cobalt and other sulfide deposits. Moderate lineament density in the western, southeastern, and central regions may indicate secondary structures that control local fluid pathways or satellite mineralisation (Ngassam Mbianya et al., 2021; Nforba et al., 2019). Low-density areas in the southwest and far northeast likely represent stable zones or post-mineral cover, where exploration interest is low unless geochemical or geophysical anomalies are present (Ourhzif et al., 2019). Lineament density is a key indicator of structural targets, with high-density zones often matching mineralised fluid channels (Ibraheem et al., 2022).

4.3 Structural analysis

A structural geological analysis confirmed remote sensing results and linked surface features to subsurface copper-bearing structures. Satellite interpretations were validated through fieldwork (Turner et al., 2023). A total of 338 structural measurements were collected using a compass-clinometer, including bedding, joints, reverse and normal faults, and slickensides. Strike and dip data were processed using stereonet projection and rose diagrams to determine structural trends and fracture intensity (Yamusa et al., 2018; Abolins, 2019).

4.3.1 Bedding planes (S₀)

In total, 187 bedding planes were measured in RAT, DSTRAT, RSF, SD, and CMN formations (Cailteux and De Putter, 2019). RAT consists of red terrigenous-dolomitic rocks with hematite, quartz, dolomite, and tourmaline; it is typically barren in Musonoi but locally mineralised elsewhere (Yantambwe and Cailteux, 2019). DSTRAT comprises layered argillaceous dolomite, 6–18 m thick, with fractures and moderate to high sulphide concentrations (Mambwe et al., 2023). RSF is layered silicified dolomite (1–8 m thick) with high bornite, chalcocite, and carrollite (Turner et al., 2023; Cailteux et al., 2005). SD consists of dolomitic shale (60–120 m), strongly fractured and sulphide-bearing (chalcopyrite, chalcocite, bornite), subdivided into basal, black ore, and carbonaceous units (Cailteux and De Putter, 2019; Hendrickson et al., 2015). CMN exceeds 100 m, consisting of laminated dolomite and carbonaceous shale, but is barren at Musonoi (Cailteux and De Putter, 2019; Cailteux et al., 2005; Turner et al., 2023).

The rose diagram (Figure 15a) shows dominant NW–SE bedding orientations with a secondary NE–SW set, likely related to shearing at fold bases. Dip histogram (Figure 15b) reveals steep dips (45°–75°) (López Isaza et al., 2021; Jannah et al., 2017). Pole plots (Figure 15c) indicate two families: main dips to the NE and secondary dips to the SE, consistent with an anticlinal fold. Bedding isodensity (Figure 15e) identified two co-zonal planes (40/196 and 44/302) used to model a NW–SE compressive stress tensor (σ1). The fold was characterised using a best-fit zonal plane (Figure 15d), yielding an axial plane striking N106°, dipping 76° SSW, and a fold axis N106/20° ESE, indicating an inclined fold (Cui et al., 2024).

Figure 15

A composite image contains diagrams related to geological data analysis. Diagram (a) is a circular plot with a green bar extending diagonally, labeled with cardinal directions. Diagram (b) is a qualitative dip chart with a histogram displaying dip counts, peaking at the last bin. Diagram (c) features a circular density plot with multicolored contours and a web of black lines. Diagram (d) is similar to (c) but includes labeled axes and annotations. Diagram (e) contains PBT axes on a yellow background with marked symbols and arrows indicating directions, along with a key for interpretation.

Figure 15. (a) Rose diagram of stratification. (b) Frequency histograms of S₀. (c) Polar plots and cyclograms of the S₀ planes. (d) Zonal plane, axial plane, and axis of the major fold. (e) PBT diagram (P = contraction-axis, B = neutral-axis, T = extension-axis) of the stress tensor S₀.

4.3.2 Directional joints

Directional joints observed in the field are generally parallel to bedding. Some are mineralised with malachite, forming bedding-parallel veins or sills; others contain quartz, black oxides, or no infill. Malachite-bearing joints suggest fluid pathways during hydrothermal circulation, though it is unclear whether all mineralised veins formed with joint development or from later fluid events.

The joint set was analysed structurally. The rose diagram (Figure 16a) shows a dominant NW–SE orientation, consistent with bedding. The frequency histogram (Figure 16b) indicates steep dips around 70°. Pole plots (Figure 16c) reveal two main joint families dipping NE and SE. Density contours (Figure 16d) define two mean planes at 65/71 and 44/131. The compressive stress tensor diagram (Figure 16e) shows a maximum principal stress (σ1) oriented NNE–SSW (Shakir, 2020; Djezzar et al., 2022).

Figure 16

Five panels showing geological data:a. A polar plot with green shaded areas and directional indicators.b. A bar graph titled

Figure 16. (a) Rose diagram of directional joint frequencies. (b) Frequency histograms of directional joints. (c) Polar plots of directional joints. (d) Coaxial planes of directional joints. (e) Stress tensor of directional joints.

4.3.3 Cross-joints (joint perpendicular to S₀)

Cross-joints are discontinuities perpendicular to bedding (S₀). Twenty-six planes of this type were recorded. The rose diagram (Figure 17a) shows a dominant NE–SW orientation (N20°E–N30°E), perpendicular to the main bedding strike. Pole plots (Figure 17b) indicate steep dips of ∼70° toward the SE, showing sharp, near-vertical fractures. Most cross-joints lack mineralisation or hydrothermal halos, suggesting limited fluid circulation and no copper infill (Harris and Holcombe, 2014; Nkodia et al., 2024). Density contours (Figure 17c) define two co-zonal planes at 30/13 and 30/262. The compressive stress tensor (Figure 17d) shows a maximum principal stress (σ1) oriented N48°, confirming a NE–SW compression regime (Bah et al., 2023; Rani et al., 2020).

Figure 17

Four stereographic projections labeled a, b, c, and d illustrate geological data. a: Features green shaded regions along principal orientations. b: Displays density with a color gradient centered around two clusters. c: Depicts data distribution with density contours and principal component markers (PC1, PC2). d: Shows PBT axes on a yellow background with red and blue markers, indicating stress directions. Accompanying tables provide density values.

Figure 17. (a) Rose diagram of cross-joint. (b) Polar plots of cross-joints. (c) Cozonal cross-joint planes. (d) Local deformation regime stress tensor.

4.3.4 Reverse faults

Several reverse faults were identified, likely formed by reactivation of pre-existing joints. The rose diagram (Figure 18a) shows a dominant NE–SW trend, parallel to joint orientations. The frequency histogram (Figure 18b) indicates steep dips above 60°, typical of strong compression. Pole plots (Figure 18c) reveal three fault families dipping SE, S, and NW, indicating complex deformation. Two co-zonal planes (11/335 and 36/251) were extracted from density contours (Figure 18d).

Figure 18

Composite image with five panels: (a) Stereogram showing directional variance with green shading and a legend detailing data points.(b) Bar chart titled

Figure 18. (a) Rose diagram of reverse fault frequencies. (b) Dip histograms of reverse faults. (c) Polar plots of reverse faults. (d) Cozonal reverse fault planes. (e) Geodynamic regime of reverse faults, Paleo-stress PBT module.

These faults accommodated compressive strain and thickened stratified units such as DSTRAT, RSF, SD, and CMN. No malachite or sulphides were observed, so the faults are structurally significant but mineralogically sterile (Willner et al., 2009; Hamimi et al., 2023). Cross-cutting relationships in field outcrops (Figures 21, 22) suggest exploitation of pre-existing anisotropies. The stress tensor (Figure 18e) indicates a compressive regime with σ1 oriented NE–SW, supporting fault reactivation during late compression (Kazadi Banza, 2012; Nkodia et al., 2024).

4.3.5 Normal faults

Normal faults in the study area show clear structural organisation. The rose diagram (Figure 19a) indicates a dominant N–S trend, oblique to bedding (S₀), suggesting reactivation of cross-cutting joints. Frequency histograms (Figure 19b) show steep dips (42°–80°), and pole plots (Figure 19c) indicate most planes dip NE.

Figure 19

a. Rose diagram with a table showing strike orientations and percentage values. b. Bar chart labeled

Figure 19. (a) Rose diagram of normal faults. (b) Frequency histograms of normal faults. (c) Polar plots and cyclograms of normal faults. (d) Cozonal planes of normal faults. (e) Local deformation regime (PBT diagram).

Isodensity contours (Figure 19d) define two co-zonal planes at 45/231 and 47/298. The associated stress tensor (Figure 19e) indicates a local extensional regime with σ1 oriented NNW–SSE (N177°/33° NNW), consistent with post-compressional normal faulting (Nkodia et al., 2021). PBT axes show σ1 nearly vertical, σ3 horizontal east–west, and σ2 intermediate.

These normal faults post-date compressional structures, truncating older reverse faults and folded strata in the field. Their steep NE dips may reflect gravitational collapse after crustal thickening. Their influence on mineralisation remains unclear, and further work is needed to determine links with copper-bearing zones, especially along fault damage zones where secondary minerals may occur (Harris and Holcombe, 2014; Cailteux and De Putter, 2019).

4.3.6 Fault mirrors

Normal faults in the study area show clear structural organisation. The rose diagram (Figure 20a) indicates a dominant N–S trend, oblique to bedding (S₀), suggesting reactivation of cross-cutting joints. Frequency histograms (Figure 20b) show steep dips (42°–80°), and pole plots (Figure 20c) indicate most planes dip NE. Isodensity contours (Figure 20d) define two co-zonal planes at 45/231 and 47/298. The associated stress tensor (Figure 20e) reveals an extensional regime with σ1 oriented NNW–SSE (N177°/33° NNW), consistent with post-compressional normal faulting (Nkodia et al., 2021). These faults post-date compressional structures, truncating older reverse faults and folds. Their steep NE dips may reflect gravitational collapse after crustal thickening. Links with copper-bearing zones remain uncertain and need further investigation, especially in fault damage zones where secondary minerals may occur (Harris and Holcombe, 2014; Cailteux and De Putter, 2019).

Figure 20

Four stereonet plots labeled a, b, c, and d. Plot a has green sectors with a table of results. Plot b shows contour lines and two hot spots with a data table. Plot c highlights two clustered areas with labeled data points and a table. Plot d displays PBT axes on a yellow background with symbols and values, indicating orientation and deviation.

Figure 20. (a) Rose diagram of fault mirror frequencies. (b) Polar plots of fault mirrors. (c) Cozonal fault mirror planes. (d) Local deformation regime stress tensor.

This section describes field verification used to confirm remote sensing interpretations. Structural data focused on bedding, joints, and faults, with surface observations guiding lineament interpretation (Figure 21) (Sedrette and Rebai, 2020; Canbaz and Ünal Çakir, 2022). Bedding measurements show dominant NW–SE and secondary NE–SW orientations, with dips toward the NE, SE, and NW. This indicates an inverted syncline followed by an anticline within the Kolwezi Nappe syncline, supporting fracture zones identified from satellite analysis.

Figure 21

Geological map of a study area with various lithology types marked in different colors: green for limestones, quartzites, and shalestones; pink for quartzites, shales, and limestones; blue for shales, sandstones, dolomites, and conglomerates; orange for sands and gravels; with black lines indicating extraction lineaments. Includes a scale bar and north arrow with coordinates labeled.

Figure 21. Geology map.

Directional joints align with bedding, while cross joints run perpendicular. Reverse faults generally follow cross-joint directions, suggesting reactivation of earlier structures; oblique normal faults likely formed from transverse joint reactivation. Fault mirrors show reverse movement features (Abolins, 2019; Callahan et al., 2020; Kazadi Banza, 2012). A NE–SW principal stress axis (σ₁) implies a dominant compressive phase later modified by post-compressional shearing.

Hydrothermal alteration along lineaments, especially chlorite and kaolinite, provided additional structural evidence. Surface goethite and hematite indicated underlying mineralisation, validated by drilling that intersected malachite, chalcocite, and bornite (Figures 22a–g, 23a–e) (El-Omairi et al., 2024; Yohanna et al., 2023). Structural discontinuities, alteration patterns, and ore zones show strong structural control.

Figure 22

a) Close-up of rock layers with a hand holding a pen for scale. Orange dashed lines indicate fracture directions. b) Detailed view of stratified rocks marked with white lines and yellow arrows. c) Finger pointing to a linear feature on rocks with dashed lines. d) Compass is used to measure rock planes with white markings. e) Core samples aligned in trays with red markings and a pencil for size reference. f) Red rock face with yellow arrows; compass held in the foreground. g) Landscape view of a mining area with exposed rock formations under the sky.

Figure 22. Field photographs: (a) Cross-joint. (b) Cross-joint reactivated into inverse fault set in iron-rich argilo-talcose rock (RAT lilas). (c) Quartz vein deformed by a dextral strike-slip fault, showing folded black minerals aligned with the fault, indicating reactivation as a shear zone. (d) Fault mirror with marked slip striations. (e) Drill core sample showing chlorite alteration and silicified dolomite with disseminated chalcocite in the rock mass and fractures. (f) Fault mirror with marked slip striations. (g) Overview of a site within the study area showing an abandoned open pit (T17), currently exploited by artisanal miners.

Figure 23

Geological analysis images showing various rock formations and measurements. Panel (a) displays rock layers marked with lines and a compass. Panel (b) highlights a fault with annotations. Panel (c) depicts a rock face with lines indicating directions. Panel (d) shows a person standing beside a rock formation with curved lines. Panel (e) illustrates cross-joints with a compass for orientation. Each section is labeled with directional markers like NW, SE, W, and E.

Figure 23. Field photographs: (a) Joint parallel to Stratification (S₀), observed in iron-rich argilo-talcose rocks (RAT lilas), showing hematite-rich hydrothermal alteration. (b) Small-scale dextral fault (F) with a horizontal displacement of approximately 15 cm, offsetting the bedding (S₀). (c) Cross-joint sets, one of which has been reactivated and thus neoformed as a fault, with the rock showing visible kaolinite alteration. (d) Outcrop-scale fold. (e) Stratification (S₀) intersected by cross-joints.

Most copper occurs in the DSTRAT, RSF, and SD formations, coinciding with high hydrothermal anomalies. Drilling confirms high Cu–Co grades in southern high-lineament areas, while RAT, RSC, and CMN remain barren. Lineament density correlates with copper grade, demonstrating structural control on mineral pathways.

A key verification site in Neoproterozoic schist and dolomite hosts N–S mineralised structures controlled by an asymmetric syncline. The southern limb is cut by two major faults, forming fracture zones that conceal mineralisation at ∼50 m depth. Combined surface mapping and drilling confirmed remote sensing interpretations and improved understanding of tectonic controls and mineral potential at Musonoi.

4.4 Random Forest

Integrating remote sensing and structural geology in a Machine Learning model, specifically Random Forest, significantly improves copper exploration. ASTER and Landsat imagery provided key inputs—lineaments, alteration zones, and mineralisation indicators—for the model, helping identify structural patterns linked to copper (Figure 23) (Davies et al., 2025). Model accuracy in predicting mineralised zones was evaluated using multiple statistical metrics (Tables 7, 8). Field structural data on faults, bedding, and joints supplied ground-truth information, validating remote sensing results and ensuring the model was trained on verified geological features (Kulkarni et al., 2023).

Table 7

Table 7. Performance metrics.

Table 8

Table 8. Kappa statistic calculation.

Table 9

Table 9. Confusion matrix.

The RF classifier identified the patterns with 85% accuracy and a Kappa value of 0.68, showing strong agreement between predicted and actual structures in Figure 23. This confirms the method’s reliability for geological interpretation. Remote sensing has shown to be a useful and trusted tool for detecting subsurface lineaments, supporting the study’s goal (Singh et al., 2023).

The Predicted Mineralisation Map is a useful guide for future exploration (Figure 24). Mineralisation patterns suggest links between surface geology and possible subsurface structures, though these interpretations remain tentative due to limited borehole data and restricted underground access at Musonoi. As a result, subsurface conclusions should be viewed as indicative rather than definitive. Nevertheless, using structural patterns and satellite data, the model identifies promising, low-cost exploration targets. This demonstrates the value of combining machine learning, remote sensing, and structural geology in copper exploration (Taha et al., 2023; Kenzhaliyev et al., 2025). Future work should include geophysics, systematic drilling, and greater underground access to refine these interpretations.

Figure 24

A scatter plot showing mineral potential mapping (MPM) with coordinates ranging from 25°16'20

Figure 24. Predicted mineralisation map.

4.5 Validation of mineralised potential zone map

ROC analysis (Figure 25) shows an AUC of ∼0.83, indicating ∼85% prediction accuracy. This demonstrates strong performance of the Random Forest model in identifying mineralised zones at Musonoi. For spatial validation, 80 points were selected using stratified random sampling to represent different geological settings. Although clustering was reduced, some remained and may slightly affect robustness.

Figure 25

Receiver Operating Characteristic (ROC) curve showing a blue line for the RF Model with an Area Under Curve (AUC) of 0.833. The curve significantly outperforms the gray dashed line of the Random Classifier with an AUC of 0.5, indicating strong model performance. A red point highlights a specific threshold.

Figure 25. ROC curve for the MPZ map.

Previous studies also show high RF accuracy for mineral prospectivity. Zhang et al. (2018) reported strong AUC performance for gold mapping in and Lachaud et al. (2023) demonstrated RF effectiveness in evaluating epithermal gold potential.

Confusion matrices further assessed model errors. Misclassification mainly occurred along mineralised zone margins due to transitional lithologies and limited spectral separability. Errors were far less frequent in core ore zones, showing the model is most reliable within well-defined structural and alteration areas.

4.6 Strengths and limitations of the present study

The methods used provide detailed geological insights while reducing challenges from vegetation cover. PCA and machine learning enabled the detection of complex patterns useful for mineral exploration. However, atmospheric effects, sensor limitations, and mixed pixels can still cause misinterpretation of mineral signatures (Shirmard et al., 2022; Agrawal et al., 2024). Despite FLAASH corrections, spectral mixing remains possible, especially with 30 m Landsat OLI data, and confusion between alteration minerals, iron oxides, and soil moisture may occur. Hyperspectral sensors (AVIRIS, Hyperion) or UAV-mounted high-resolution systems would improve mineral identification (Pour et al., 2023; Yuan et al., 2025).

Model reliability was limited by a small training dataset (80 control points). Increasing samples using borehole data, geochemistry, or direct spectral measurements from drill cores would improve statistical robustness. Only single-date imagery was used, preventing detection of seasonal spectral changes. Multi-temporal images would minimise time-related errors and highlight stable geological features (Behera et al., 2025; Ourhzif et al., 2019; Kulkarni et al., 2023).

Findings align with previous remote sensing studies. Random Forest has proven effective for mineral potential mapping (Taha et al., 2023; Kenzhaliyev et al., 2025). This approach is suitable for Musonoi’s complex Neoproterozoic structures and supergene enrichment, supporting targeted exploration. Future work should incorporate structural restoration, petrography, and isotopic dating to refine mineralisation timing (Nair et al., 2023; Mahnaz et al., 2023; Mahboob et al., 2024). Outputs should also support decision-making, including prospectivity maps, risk modelling, and sustainability indicators. Broader seasonal field checks may reveal changing surface mineral signatures.

5 Conclusion

Remote sensing, structural analysis, and Random Forest modelling confirm strong copper potential at Musonoi. Spatial correlations above 75% between high lineament density and strong PCA and band-ratio anomalies demonstrate reliable detection of alteration linked to known mineralisation. Structural convergence with altered zones, especially in the southeast, marks priority targets for follow-up work. Limitations include satellite resolution, possible Random Forest training bias, and incomplete field validation; thus, results should be viewed as probabilistic guides rather than precise mineral boundaries.

The workflow can be applied to other Central African Copperbelt areas with local calibration. Future research should test alternative ML models such as CNNs or Graph Neural Networks, and integrate geophysics (induced polarisation, gradient magnetometry) for buried structure detection.

Finally, this method supports sustainable exploration. Remote sensing and predictive modelling reduce invasive surveys and environmental disturbance. With rising copper demand for renewable energy, this integrated, data-driven workflow offers efficient, selective, and environmentally responsible exploration strategies.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

Ethics statement

Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

Author contributions

MT: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review and editing. LN: Conceptualization, Data curation, Formal Analysis, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing – review and editing. KT: Conceptualization, Investigation, Resources, Validation, Visualization, Writing – review and editing.

Funding

The author(s) declared that financial support was not received for this work and/or its publication.

Acknowledgements

I would like to express my sincere gratitude to Prof. Mashala for his assistance whenever needed during this research. My deepest gratitude goes specifically to my wife, Ngombe Melie Daniella Tshanga, for her patience, support, and courage, as well as understanding during a lengthy period of absence. This achievement belongs to you just as much as it does to me. Each day, my sons, Tshanga Matthieu Ephphatha and Tshanga Guevis Eli Emeth, and my daughter, Miradi Tshanga, bring purpose and inspiration that fuel my dedication. Gratitude is also due to researchers and professionals working in environmental science and geology, whose ongoing efforts continue to shape and move this field forward.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Abdelkareem, M., Abdalla, F., and Yousef, S. (2024). The use of ASTER and landsat remote sensing data for exploring hydrothermal mineral in arabian-nubian shield. Kuwait J. Sci. 51 (3), 100240. doi:10.1016/j.kjs.2024.100240

Geological mapping of copper deposits in the democratic republic of Congo through remote sensing data and machine learning

1 Introduction

2 Geological setting

2.1 Stratigraphy

2.2 Local geology

3 Materials and methods

3.1 Field data collection

3.2 Image acquisition and preprocessing

3.2.1 PCA application

3.2.1.1 Landsat imagery

3.2.1.2 ASTER data

3.2.2 Band ratios

3.2.3 Lineament extraction with PCI geomatica

3.2.4 Statistical computation

3.3 Machine learning (Random Forest)

4 Results and discussion

4.1 Mineral mapping

4.2 Lineament analysis

4.3 Structural analysis

4.3.1 Bedding planes (S0)

4.3.2 Directional joints

4.3.3 Cross-joints (joint perpendicular to S0)

4.3.4 Reverse faults

4.3.5 Normal faults

4.3.6 Fault mirrors

4.4 Random Forest

4.5 Validation of mineralised potential zone map

4.6 Strengths and limitations of the present study

5 Conclusion

Data availability statement

Ethics statement

Author contributions

Funding

Acknowledgements

Conflict of interest

Generative AI statement

Publisher’s note

References

4.3.1 Bedding planes (S₀)

4.3.3 Cross-joints (joint perpendicular to S₀)