A review of machine learning in processing remote sensing data for mineral exploration

Introduction

• Mapping lithological units, alteration types, structures, and indicator minerals is a fundamental step in locating geological features associated with mineralization.The introduction presents remote sensing and machine learning as an increasingly prominent combination for geological mapping and decision-making.Source: §1 · p. 1
• Remote sensing data from multiple platforms and resolutions helps address field-mapping challenges, including access to remote areas and effects of climate, topography, and fieldwork conditions.The review describes satellite, airborne, drone, and ground-based data as relevant sources for exploration mapping.Source: §1 · pp. 1–2
• Processing spectral imagery supports mapping lithological units, alteration, structures, rocks, and minerals, while machine learning can manage high-dimensional data and map complex features.The paper frames combining remote sensing and machine learning as a means of facilitating and improving mineral exploration.Source: §1 · p. 2
• The review organizes remote sensing applications for mineral exploration around data characteristics, mapping key geological features, machine-learning methods, and challenges and future work.It groups methods into dimensionality reduction, classification, clustering, regression, and deep learning, and highlights emerging approaches including graph and Bayesian deep learning, variational autoencoders, and transformer recurrent neural networks.Source: §1 · p. 2

2. Methodology

• The review categorizes machine-learning methods and surveys their use in remote-sensing detection of lithology, alteration, structures, and indicator minerals.These geological target features are identified as mappable elements relevant to mineral exploration.Source: §2 · p. 2
• Searches of Google Scholar and Scopus used expert- and prior-study-informed keywords, including “machine learning,” “remote sensing,” and “mineral exploration.”The review also searched for studies applying machine learning to delineate the specified geological target features.Source: §2 · p. 2
• Scopus publication counts show growing research on machine learning and remote sensing, but mineral-exploration applications remain a small share and reveal an application gap.The review compares keyword searches spanning machine learning and remote sensing, minerals, and mineral exploration.Source: §2 · pp. 2–3, Figure 1 · p. 3

3.1. Satellite data

• Satellite optical and radar data are the principal types used for geological mapping, relying respectively on ambient energy and self-generated energy.Optical sensors measure electromagnetic intensity across wavelength channels, while radar operates in the microwave range of 1 mm to 1 m.Source: §3.1 · p. 3

3.1.1. Optical data

• Optical satellite datasets range from multispectral Landsat and Sentinel-2 imagery to commercial high-resolution and spaceborne hyperspectral systems.Landsat 7’s SWIR bands are sensitive to soil and rock changes and can help differentiate basic rock types; Sentinel-2 provides 13 VNIR/SWIR bands at 10, 20, and 60 m resolutions. WorldView-3 combines eight VNIR bands at 1.2 m with eight SWIR bands at 3.7 m, while Hyperion supplies 242 VNIR/SWIR bands at 30 m. Google Earth imagery supports outcrop identification during pre-field planning and reconnaissance.Source: §3.1.1 · pp. 3, 7

3.1.2. Radar data

• Satellite SAR data are used to map geological structures, with RADARSAT, Sentinel-1, and PALSAR offering C- or L-band observations and varied modes and resolutions.Sentinel-1 is an all-weather, day-and-night C-band mission with dual polarization and 5×20 m spatial resolution.Source: §3.1.2 · p. 7

3.2. Airborne data

• Airborne sensors provide multispectral and hyperspectral observations, while drones can survey difficult terrain quickly and safely with limited on-site personnel.AVIRIS-NG provides 427 contiguous 5 nm bands from 0.38 to 2.51 µm at 8.1 m spatial resolution; UAV sensors can achieve millimeter-to-centimeter spatial resolutions. A positioning system not dependent on GPS enables precise mapping where GPS reception is low, such as mine tunnels and narrow valleys.Source: §3.2 · p. 8

3.3. Ground-based data

• Ground-based hyperspectral sensing supports alteration-mineral mapping and high-resolution characterization of rocks, minerals, and soil.In the Gongchangling iron deposit, chlorite and garnet wavelengths grew longer and absorption depths shrank from proximal to distal zones. Portable HySpex and snapshot cameras enable detailed capture; the UHD 285 records over 15 spectral cubes per second.Source: §3.3 · p. 8

4. Target features

• Target mapping spans lithology and structure, hydrothermal alteration, and mineralization zones, with Figure 2 illustrating porphyry copper alteration and ore zones.Figure 2 also maps interpreted exposure levels for Iranian alteration sites using ASTER-mapped alteration units.Source: §4 · pp. 8–9, Figure 2 · p. 9

4.1. Lithology

• SVM classification of ASTER imagery, DEM, and aeromagnetic data had better precision than MLC on independent validation samples and resembled the available bedrock database.Remote-sensing lithology mapping also uses PCA, ICA, MNF, SAM, and MLE; high-resolution WV-3 imagery has been evaluated alongside ASTER and OLI for large-scale mapping.Source: §4.1 · p. 9

4.2. Alteration types

• Alteration minerals produce diagnostic spectral patterns, enabling remote sensing to distinguish altered zones; PCA, SAM, SID, selective dimensionality reduction, and AVIRIS-NG hyperspectral data have been applied.AVIRIS-NG data at 5 nm spectral resolution enabled identification of altered and weathered clay groups; mapped alteration types include propylitic, phyllic, argillic, and advanced argillic.Source: §4.2 · pp. 9–10

4.3. Structures

• Satellite optical and radar data support mapping faults, dykes, and other lineaments associated with hydrothermal mineralization, using filtering, edge detection, spectral methods, classification, or manual extraction.SAR processing has also been used to evaluate regional structural control of scattered gold anomalies and reveal lithotectonic patterns in Egypt.Source: §4.3 · p. 10

4.4. Minerals

• Hyperspectral spectral matching and unmixing map indicator minerals and alteration facies, while PCA, ICA, and spectral analyses of ASTER, ALI, and Hyperion data support mineral extraction.Mapped examples include clay, sulfate, carbonate, iron-oxide, and silica minerals; minimum-distance classification extracted minerals from AVIRIS, ASTER, and WorldView-3, and drone hyperspectral data mapped REEs.Source: §4.4 · pp. 10–11

Machine learning

• Machine learning with remote sensing can map lithological units, alteration zones, structures, and indicator minerals, learning input–output relationships while remaining robust to noise and uncertainty.The paper describes this combined approach as easy and inexpensive for mineral exploration.Source: §5 · p. 11
• Supervised methods use labeled data for regression and classification, but performance varies by problem and dataset, so candidate methods should be examined for each task.Source: §5 · p. 11
• Unsupervised methods identify patterns without target labels; clustering groups similar instances, while dimensionality reduction can lower complexity, remove noise or outliers, and shorten training.Simplified models may be more robust on smaller datasets.Source: §5 · p. 11
• The review covers dimensionality reduction, classification, clustering, regression, and deep learning, including neural-network and ensemble approaches, and presents a workflow for evidential maps in prospectivity mapping.Table 2 lists sample studies mapping potential mineralization zones, and Figure 3 presents the workflow combining remote sensing data and machine learning.Source: §5 · p. 11, Table 2 · p. 12

5.1. Dimensionality reduction techniques

• Dimensionality reduction methods including PCA, ICA, and MNF transform correlated input variables into uncorrelated or independent components and are popular in remote-sensing processing.Source: §5.1 · p. 13
• Landsat 8 PCA and MNF mapped lithological units and alteration minerals for alluvial gold exploration, while Sentinel-2 PCA supported IOCG exploration.Source: §5.1 · p. 13
• PCA and ICA extracted spectral details of vegetation, iron oxides and hydroxides, alteration minerals, carbonates, and silicification regionally (Takodjou Wambo et al., 2020); PCA was also used to prospect for Zn-Pb.Source: §5.1 · p. 13
• Selective PCA maps of hydrothermal alteration, validated with fieldwork, XRD, petrography, and spectroscopy, indicated high potential for epithermal polymetallic veins.Source: §5.1 · p. 13
• PCA applied to ASTER data mapped lithological units and structures in Morocco’s Bou Azzer-El Graara inlier, with 74% performance in lineament mapping, especially after removing paths, borders, geological boundaries, hill summits, and shadows (El Atillah et al., 2019).PCA and MNF were considered across three real-life geoscience and mineral-exploration datasets using UAS-borne or small-angle terrestrial imaging at VNIR, SWIR, and LWIR resolutions.Source: §5.1 · p. 13

5.2.1. Minimum distance classification

• Minimum distance classification measures each pixel’s Euclidean distance from class mean vectors and has been used for mineral, lithological, land-use, and pegmatite mapping, though accuracy can be poor.Wang and Jiang (2019) reported that a weighted minimum-distance classifier based on relative offset boosted accuracy.Source: §5.2.1 · p. 13

5.2.2. Support vector machines

• SVMs classify remote-sensing data for lithological, alteration, and mineral mapping, with studies reporting new targets and reliable delineation but reduced accuracy when classes are imbalanced.Othman and Gloaguen (2014) reported a new chromite-bearing site; other studies mapped alteration zones and lithium-bearing pegmatites, while class inequality caused some known Li-pegmatites to go undetected.Source: §5.2.2 · pp. 13, 15

5.2.3. Artificial neural networks

• Neural networks model complex, nonlinear relationships and have been used to map gold potential, integrate mineral anomalies, discriminate lithologies, and identify lineaments.Rigol-Sanchez et al. (2003) produced a gold-potential map indicating historically known and previously unknown potentially mineralized areas; Wang et al. (2010) integrated geological, geophysical, geochemical, and remote-sensing anomaly information.Source: §5.2.3 · p. 15

5.2.4. Random forest

• Random forests ensemble decision trees to address overfitting and have been applied to classify minerals and lithologies, including Li-bearing pegmatites correctly identified in three open-pit mines.Applications include selecting representative hyperspectral bands for magnesite and gangue minerals and lithological mapping with Sentinel-2 imagery.Source: §5.2.4 · p. 16

5.3. Clustering

• Ren et al. (2019) improved KMC initialization and similarity measures, achieving stronger clustering and more precise mineral mapping than conventional KMC in Cuprite.They compared the results with the area's mineral distribution map and the USGS mineral spectral library.Source: §5.3 · p. 16
• KMC produced lithological and mineral maps consistent with field data and geological maps across multisensor satellite and airborne hyperspectral studies.El Atillah et al. (2019) used ASTER, Landsat 7, Landsat 8, and Sentinel-2; Ishidoshiro et al. (2016) combined KMC with spectral unmixing and full-pixel classification on HyMap imagery.Source: §5.3 · p. 16
• AGNES hierarchically merges singleton clusters into a dendrogram, DBSCAN finds arbitrary-shaped clusters through local density, and spectral clustering partitions graphs into communities.The section presents these methods as progress that could improve KMC applications on large, complex datasets.Source: §5.3 · pp. 16–17
• ISODATA was used for lithological cartography and iron-ore identification with multispectral, hyperspectral, topographic, and geophysical data.El Atillah et al. (2019) assessed lithological, structural, and hydrothermal-alteration findings against field data and geological maps; Ducart et al. (2016) classified Hyperion, LIDAR-derived terrain, gamma-ray, gravimetry, and OLI-derived NDVI data selected for their importance to outcropping iron-ore identification.Source: §5.3 · p. 17

5.4. Regression analysis

• Yetkin et al. (2004) used multi-linear regression to map alteration trends and found buried faults associated with alteration zones.The study mapped potassic, phyllic, propylitic, argillic, and silicification zones; Hasandag was poorly altered, while Keciboyduran, Melendiz, and Tepek¨oy were strongly altered.Source: §5.4 · p. 17
• Mansouri et al. (2018) used multivariate regression to map iron outcrops, verifying model accuracy against an outcrop map and geological exploration.Field observations placed iron mineralization in contact with limestone and intrusive rocks.Source: §5.4 · p. 17
• Hoang and Koike (2017) used ETM+ band-reflectance correlations to generate pseudo-Hyperion imagery for mapping metallic minerals, including gold, silver, and copper.Source: §5.4 · p. 17
• Lin et al. (2020) integrated geological, remote-sensing, and geochemical evidence in a conjugate-gradient logistic regression model to predict skarn deposits.The authors report that the model efficiently integrated the three data types.Source: §5.4 · p. 17
• Li et al. (2018) used hyperspectral imaging in a multivariate regression model to quantitatively analyze and predict device capability and optimize mineral-detection parameters.Source: §5.4 · p. 17
• Uncertainty from model parameters and data collection—including noise, sparse datasets, and sensor limitations—remains underexplored, with Bayesian regression proposed as a potential direction.The section cites Bayesian linear and logistic regression as possibilities for remote sensing and mineral exploration.Source: §5.4 · p. 17

• Latifovic et al. (2018) reviewed CNN mapping of surficial geological classes using field-specific samples or transfer to areas excluded from training.The study covered five areas with aerial imagery, Landsat reflectance, and high-resolution digital elevation data; its fusion spatial-spectral function stacked CNN-derived local spatial characteristics with spectral data. CNNs can automatically extract image features through convolutional and pooling layers, whose optimal number is often experimentally determined.Source: §5.5 · pp. 17–18
• 94.70 percent accuracy was achieved by Zhao et al. (2020)’s HSS-LSTM classification of AVIRIS hyperspectral data from Nevada, exceeding the most widely used methods.Source: §5.5 · p. 18
• LSTMs model temporal data and dynamical systems, but have seen little use in mineral exploration; LSTM-CNN and convolutional LSTM combinations offer potential for spatiotemporal remote sensing.The passage notes prior remote-sensing applications of convolutional LSTM and identifies these combinations as having potential in mineral exploration.Source: §5.5 · p. 18
• Sang et al. (2020) combined UAV imagery and ground study in a SLIC-CNN framework that classified image structure and outlined rock-mass boundaries for lithological mapping.CNN validated lithological distribution, while SLIC outlined boundaries; fusion and mapping results were explained through mode and expert decision processes.Source: §5.5 · p. 18
• Li et al. (2020b)’s CNN-GNN framework exploited scene-graph spatiotopological relationships and showed superior efficiency on two public multilabel scene-classification datasets; Gao et al. (2021)’s H-GCN showed viability for remote-sensing scene classification.Source: §5.5 · p. 18

Challenges and future prospective

• Mineral exploration faces low discovery success and rising demand, while remote-sensing big data challenge acquisition, preparation, analysis, and interpretation across multisource, multi-scale, high-dimensional, dynamic, and nonlinear data.Exploration methods also depend on deposit type, location, infrastructure, and existing geological knowledge.Source: §6 · p. 18
• CNNs can automate image-feature extraction and exploit spectral and spatial information; Ding et al. (2017) reported improved HSI classification for geological targets relevant to mineral exploration.Their fully convolutional kernels are learned automatically through clustering without prespecifying the number of clusters.Source: §6 · pp. 18–19
• Uncertainty quantification remains a gap in remote-sensing applications, while Bayesian inference could represent uncertainty in model parameters and data, including noise, sparse observations, and missing data.Bayesian neural networks and deep learning have seen progress but have not been much used in remote sensing, and their mineral-exploration application is absent; MCMC-based approaches could integrate multiple data sources.Source: §6 · p. 19
• Transfer learning, autoencoders, and GANs are potential approaches for remote-sensing mineral exploration, including data compression, class imbalance, and reconstruction of missing regions.Deep autoencoders with CNN features may support compression (Pu et al., 2016), while GANs may address class imbalance or missing regions.Source: §6 · p. 19
• Spatio-temporal tectonic data can augment mineral-exploration models by placing deposits in plate-tectonic evolution contexts; Diaz-Rodriguez et al. (2021) modeled subduction-zone evolution for porphyry Cu systems over 100 million years.Such approaches may improve understanding of the geological niche environments associated with particular deposits through space and time.Source: §6 · p. 19
• Machine-learning use must account for rapidly rising computational costs, accuracy-effort trade-offs, overfitting, problem dependence, outliers, and limited interpretability; no single technique suits all applications.Method selection depends on the specific application and experimental conditions, and geologists need stronger understanding of exploration problems and model settings.Source: §6 · p. 19

7. Conclusions

• Remote sensing datasets provide a resource for overcoming geological-mapping problems associated with relying on field data alone.The review presents remote sensing and machine learning together as useful for mapping features relevant to mineral exploration.Source: §7 · p. 20
• Deep learning can process large, complex remote sensing data, handle noise and uncertainty, and help identify target features and minerals.The review says advanced deep learning may improve target-feature mapping in both small- and large-scale studies.Source: §7 · p. 20
• Neural networks are widely used for remote sensing classification and prediction, while dimensionality reduction can extract useful features from high-dimensional data.The review describes these methods as applicable to mineral exploration and remote sensing processing.Source: §7 · p. 20
• Limitations and specific requirements of machine learning methods remain key obstacles for exploration geologists.The review also notes that advanced analytics may support mining sustainability by reducing prospecting’s ecological and environmental impacts through efficient, cost-cutting solutions.Source: §7 · p. 20