Innovative Hyperspectral Data Fusion for Enhanced Mineral

Prospectivity Mapping

Roberto De La Rosa

, Michael Steffen

, Ina Storch

, Andreas Knobloch

Joana Cardoso-Fernandes

, Morgana Carvalho

, Mercedes Suárez Barrios

Juan Morales Sánchez-Migallón

, Petri Nygren

, Vaughan Williams

and Ana Cláudia Teodoro

Beak Consultants GmbH, 09599 Freiberg, Germany

Institute of Earth Sciences (ICT), Department of Geosciences, Environment and Land Planning, Faculty of Sciences,

University of Porto, 4169-007 Porto, Portugal

Department of Geology, University of Salamanca, 37008 Salamanca, Spain

Spectral Mapping Services SMAPS Oy, 20320 Turku, Finland

Aurum Exploration Ltd, A82 HK12 Kells, Ireland

Keywords: Exploration, CRM Mapping, Hyperspectral Data, Mineral Prospectivity Mapping, Deep Learning, Bayesian

Neural Networks, Self-Organizing Maps.

Abstract: To meet the European Union’s growing demand for critical raw materials in the transition to green energy,

this study presents a novel, cost-effective, and non-invasive methodology for mineral prospectivity mapping.

By integrating hyperspectral data from satellite, airborne, and ground-based sources with deep learning

techniques, we enhance mineral exploration efficiency. We employ Bayesian Neural Networks (BNNs) to

predict mineral prospective areas while providing uncertainty estimates, improving decision-making. To

address the challenge of obtaining reliable negative labels for supervised learning, Self-Organizing Maps

(SOMs) are used for unsupervised clustering, identifying barren areas through co-registration with known

mineral occurrences. We illustrate this approach in the Aramo Unit in Spain, a geologically complex region

with Cu-Co-Ni mineralized veins. Our workflow integrates local geology, mineralogy, geochemistry, and

structural data with hyperspectral data from PRISMA, airborne Specim AisaFenix, LiDAR and ground-based

spectroradiometry. By leveraging learning techniques and high-resolution remote sensing, we accelerate

exploration, reduce costs, and minimize environmental impact. This methodology supports the EU’s S34I

project by delivering high-value, unbiased datasets and promoting sustainable, cutting-edge mineral

exploration technologies.

1 INTRODUCTION

The increasing global demand for critical raw

materials (CRMs) necessary for renewable energy

technologies, consumer electronics, electric vehicles

and defence has intensified the urgency of developing

efficient, sustainable, and innovative mineral

exploration methods. The European Union (EU), in

its transition toward green energy, faces significant

challenges due to limited domestic production of

https://orcid.org/0000-0002-3004-7104

https://orcid.org/0000-0001-7515-001X

https://orcid.org/0009-0001-4332-2187

https://orcid.org/0000-0002-8043-6431

CRMs, necessitating reliance on imports. This

dependence introduces risks related to supply chain

disruptions and geopolitical instability. To address

this challenge, the EU has launched several initiatives

to promote the sustainable and responsible sourcing

of CRMs, including the Secure and Sustainable

Supply of Raw Materials for EU Industry (S34I)

project. This project aims to develop new

technologies and approaches for mineral exploration,

extraction, and processing that minimize

De La Rosa, R., Steffen, M., Storch, I., Knobloch, A., Cardoso-Fernandes, J., Carvalho, M., Barrios, M. S., Sánchez-Migallón, J. M., Nygren, P., Williams, V. and Teodoro, A. C.

Innovative Hyperspectral Data Fusion for Enhanced Mineral Prospectivity Mapping.

DOI: 10.5220/0013497900003935

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 11th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2025), pages 317-328

ISBN: 978-989-758-741-2; ISSN: 2184-500X

317

environmental impact and maximize resource

efficiency.

Mineral prospectivity mapping (MPM) is a critical

tool in addressing these challenges. MPM traditionally

uses geographic information systems (GIS) to

integrate diverse datasets as geological, geophysical,

geochemical, and remote sensing to highlight areas

with high mineralization potential. Traditional

exploration techniques, while effective, often involve

significant time, expense, and environmental

disruption. Recent advancements in technology have

revolutionized MPM, leveraging the power of GIS

platforms, machine learning (ML), and artificial

intelligence (AI) to improve the accuracy, efficiency,

and sustainability of mineral exploration (Carranza &

Hale, 2001, Carranza 2008, Nykänen et al., 2017,

2023, Yousefi et al., 2021, 2024 & Zhang et al., 2022).

MPM approaches are generally categorized into

knowledge-driven, data-driven (Yousefi & Nykänen,

2016, Torppa et al., 2019, Lawley et al., 2022 &

Nagasingha et al., 2024), and hybrid methods.

Knowledge-driven techniques rely on expert

interpretations of geological formations, making them

particularly suitable for "greenfield" exploration

regions with few known deposits. In contrast, data-

driven methods empirically model relationships

between explanatory variables and mineral

occurrences, often applying ML techniques to

established mining areas or "brownfield" regions.

Hybrid approaches combine these methodologies,

leveraging data-driven insights to enhance expert-

driven interpretations. Deep learning models,

particularly Bayesian Neural Networks (BNNs), have

demonstrated exceptional capabilities in extracting

complex patterns and relationships within large,

multidimensional datasets, improving predictive

accuracy and uncertainty quantification (Mao et al.,

2023 and Jordão et al., 2023).

Despite these advancements, key challenges

remain. One of the most significant is the scarcity and

imbalance of labelled data, where known mineral

deposits or mineralized samples (positive samples

used as training points) are scares and rare compared

to the amount of data available, which is effectively

unknown in terms of a positive or negative binary

classification for a CRM mapping. Traditionally, the

unknown areas are considered barren regions from

where the negative samples are randomly selected.

Another traditional option is an expert driven negative

sampling that requires extensive geological expertise,

a good understanding of the study area and a vast

knowledge of the parameters driving the

mineralization event, which is not always available,

possible or extremely expensive and time-consuming,

leading inevitably to an imbalance positive-negative

training samples. This imbalance can lead to biased

models and unreliable predictions (Mao et al., 2023).

Addressing this issue, our research introduces a novel

data-driven approach to negative sampling selection

by leveraging Self-Organizing Maps (SOMs). Instead

of relying on arbitrary or expert-defined barren

regions, we co-register SOM clusters with known

mineral occurrences to identify geologically

representative negative samples. This ensures that the

training dataset accurately reflects the true

background variability of the study area, leading to

improved model robustness and generalization.

The integration of BNNs into our methodology

provides another key innovation by incorporating

uncertainty quantification into mineral prospectivity

predictions. Unlike conventional neural networks that

yield deterministic outputs, BNNs estimate

probability distributions over model parameters,

allowing them to quantify prediction uncertainty

(Jordão et al., 2023). This uncertainty information is

particularly valuable for mineral exploration, as it

enables risk-aware decision-making and strategic

resource allocation (Lauzon & Gloaguen 2024 &

Zhang et al., 2024). Exploration efforts are prioritized

in areas with high predictive confidence while regions

with significant uncertainty can be flagged for further

investigation. By embedding uncertainty

quantification within the model, our approach

enhances the interpretability and transparency of the

mineral prospectivity mapping process, reducing the

risk of false positives and missed discoveries.

Furthermore, our methodology offers several

practical advantages over traditional exploration

techniques. The non-invasive nature of hyperspectral

remote sensing significantly reduces environmental

impact by minimizing the need for extensive ground

surveys. This is particularly beneficial for ecologically

sensitive or remote regions where physical access is

limited. The high spectral resolution of hyperspectral

imaging allows for the precise identification of

mineral signatures, capturing subtle spectral features

that traditional methods may overlook. Additionally,

by automating feature extraction, classification and

post-processing of the data, we reduce the need for

extensive manual interpretation, thereby increasing

efficiency and cost-effectiveness.

By aligning with the objectives of the S34I

project, our research contributes to the development

of sustainable and technologically advanced mineral

exploration methodologies. The integration of multi-

scale hyperspectral remote sensing, SOM-driven

negative sampling, and BNN-based prospectivity

mapping and uncertainty quantification represents a

S34I 2025 - Special Session on S34I - From the Sky to the Soil

318

transformative step forward in mineral prospectivity

mapping. As hyperspectral imaging technology

continues to advance and larger, higher-quality

datasets become available, the predictive accuracy

and effectiveness of this approach will further

improve. Additionally, continued innovations in

machine learning architectures, Bayesian inference,

and self-supervised learning will enhance the

capabilities of this methodology, making it an

increasingly powerful tool for mineral exploration.

Our proposed approach represents a significant

advancement in mineral exploration by providing a

scientifically rigorous, scalable, and environmentally

responsible method for identifying potential mineral

deposits. By addressing key challenges such as

negative sampling bias and uncertainty estimation,

we offer a robust framework that improves predictive

reliability and supports informed decision-making in

exploration projects. The integration of multi-scale

hyperspectral data, SOM-based negative sampling,

and deep learning via Bayesian Neural Networks not

only enhances the accuracy and efficiency of mineral

prospectivity mapping but also supports the broader

transition toward sustainable resource management

and a green energy future.

2 STUDY AREA & DATA

2.1 Geological Setting

The study area is the Aramo Unit, a thrust nappe

within the Fold and Nappe Province of the Cantabrian

Mountains in northern Spain (Figure 1). This region

comprises a diverse sequence of Paleozoic

sedimentary rocks, primarily from the Devonian and

Carboniferous periods. The stratigraphic sequence

includes Devonian shales, sandstones, and

limestones, followed by Tournaisian-Visean grey and

red nodular limestones. The Namurian succession is

characterized by black, bituminous limestone, while

the Bashkirian to Lower Moscovian sequence

consists of shales interbedded with limestones and

sandstones (Paniagua et all., 1988, 1993).

Structurally, the mineralization at the Aramo mine is

controlled by the intersection of the E-W Aramo Fault

and the Aramo Thrust front. The Aramo Fault is a

major discontinuity traversing the Aramo Unit, while

the Aramo Thrust front delineates the boundary

between the Fold and Nappe Province and the Central

Coal Basin. This structural interplay has induced

extensive dolomitization and minor silicification in

proximity to the orebody (Bruner & Smosna, 2000),

Ordóñez et al., 2005) and (Loredo & Ordóñez, 2008).

The Aramo mine hosts significant Cu-Co-Ni

mineralization, occurring as mineralized veins with

an average thickness of 25 cm. These veins are

predominantly found within the Namurian limestone,

situated along the thrust fault front. The deposit is

epigenetic and carbonate-hosted, comprising Cu-Co-

Ni sulfides and arsenides with minor precious metals

(Paniagua et al., 1988). The combination of structural

controls and lithological characteristics has played a

crucial role in the formation and localization of the

mineralization.

Figure 1: Geological and structural map of the study area.

Modified from: Aurum Exploration Ltd & Bergua et al.,

2019.

2.2 Data Acquisition

A multi-scale approach to data acquisition was

adopted for this study, integrating satellite, airborne,

and ground-based measurements. Initially,

multispectral and hyperspectral satellite data from

Sentinel-2, Lansat-9 and PRISMA sensors were

analyzed to delineate alteration zones and refine the

definition of the main study area (Carvalho et al.,

2025). These delineated regions subsequently guided

airborne data acquisition, ensuring targeted high-

resolution imaging. Furthermore, the identified

alteration areas informed and optimized ground-

based sampling strategies for geochemical analysis

and spectral validation, enhancing the overall

effectiveness of the exploration process.

Innovative Hyperspectral Data Fusion for Enhanced Mineral Prospectivity Mapping

319

2.2.1 Hyperspectral & LiDAR Data

Regional hyperspectral data were acquired by the

PRISMA satellite, providing broad coverage of the

study area. This data was processed and analysed by

the partners from the University of Porto and helped

to the definition of alteration areas, the definition of

the main study area, guided part of the ground-based

sampling and furthermore an independent component

analysis was performed (Carvalho et al., 2025) which

produced informative input layers for the machine

learning models.

Figure 2: Airborne hyperspectral data flight lines.

High-resolution hyperspectral data were acquired

using the Specim AisaFENIX sensor, flown on an

airborne platform (Figure 2). This camera implements

two sensors to cover the visible and near infrared

(VNIR 380–1000 nm) and shortwave infrared (SWIR

1000–2500 nm) regions of the electromagnetic

spectrum along 450 spectral bands. The acquisition

and pre-procesing of the data was performed by the

partners from Smaps Oy. Due to persistent harsh

weather conditions (typical of this region), it was no

possible to fly all the planned lines, and this is the

reason of the data gap at the east of the study area and

the high percentage of cloud coverage. Unfortunately,

this gap coincided with a prominent mineralization

outcrop where most of the rock samples were located

(Figure 1 and 2). The flight mission was on hold for

more than one year waiting for the appropriated

weather window. The pre-processing of the airborne

hyperspectral data was also performed by Smaps

which consisted on the orthorectification and

geometric correction, followed by the atmospheric

correction performed with the ATCORE4 software

resulting in reflectance data with ground sampling

distance of approximately 1.2 meters per pixel.

The airborne LiDAR data was acquired in 2023

by the partners from Eurosense with the waveform

processing Airborne Laser scanner Riegl VQ780

obtaining an average point density of 10pts/sqm and

resulting in a digital elevation model (DEM) with

0.5m resolution per pixel (Figure 3).

Figure 3: Airborne LiDAR flight lines and impressions on

the acquisition (Image credit from Eurosense partners).

2.2.2 Ground-Based Techniques

The study area is currently being actively explored by

Aurum Exploration Ltd., in collaboration with local

partners from the Department of Geology at the

University of Salamanca. Together, they have

conducted multiple field campaigns for sample

collection and analysis, as well as geological and

structural mapping.

Field measurements were performed using an

ASD portable spectroradiometer by the University of

Salamanca to acquire high-quality point spectral

signatures in the VNIR and SWIR regions of the

electromagnetic spectrum. These measurements were

taken from known mineral occurrences and

background lithologies to build spectral libraries for

spectral validation and the supervised processing of

hyperspectral images.

Geological maps were used to extract the host

Valdetejas Formation, and the distance to this unit

was calculated and rasterized for use as an input

S34I 2025 - Special Session on S34I - From the Sky to the Soil

320

evidence layer in machine learning models.

Additionally, the main fault structures were

categorized into two groups based on their azimuth:

E-W and N-S oriented faults. Finally, the distance to

thrust front-related faults was calculated and

incorporated as an input variable in the models.

Geochemical analyses were performed on the

collected samples, and the results were used to select

the samples that would serve as positive training

points for the supervised machine learning methods.

Specifically, samples with concentrations of Co >

0.05%, Cu > 0.4%, and Ni > 0.07% were selected,

resulting in a dataset of 32 samples.

3 METHODOLOGY

The workflow of the proposed methodology is

depicted in Figure 4. It begins with the acquisition of

airborne hyperspectral data, which is already

radiometrically calibrated and atmospherically

corrected to reflectance. This data undergoes baseline

correction and de-noising to produce corrected

hyperspectral reflectance data. This corrected data is

then processed using both unsupervised and

supervised methods.

Figure 4: General workflow for hyperspectral data fusion

with SOM and BNN for Mineral Prospectivity Mapping.

In the supervised processing stage, ground

spectroradiometer measurements are used for spectral

validation and rasterized maps containing the relevant

features extracted from the unsupervised process,

along with geochemical data, are integrated to

provide a more comprehensive understanding of the

spectral data. Additionally, the geological and

structural data are also incorporated to evidence

layers stack as single bands raster files. These diverse

data sources undergo spatial co-registration to ensure

that the spatial and spectral information aligns

correctly. The data fusion is performed within the

application of the Bayesian Neural Network (BNN),

which models the relationships within the fused data.

This approach leverages the capabilities of BNNs in

handling complex, high-dimensional data to predict

mineral distributions with higher accuracy and

reliability, ultimately producing a mineral prediction

map and the uncertainty associated to it.

Figure 5: Indices from band ratios along flight lines.

Figure 6: Automated unsupervised hyperspectral data

processing output. A: RGB. B: Cloud mask, C: PCA, D:

Zoom-in to panel C, E: Zoom-in to panel D, F: End-member

spectra, G: Minimun Wavelength Map from 1780 to 1990

nm. H: SAM for end-member of panel F.

Innovative Hyperspectral Data Fusion for Enhanced Mineral Prospectivity Mapping

321

3.1 Automated Unsupervised

Hyperspectral Data Processing

In the unsupervised processing stage, several

techniques such as Minimum Wavelength maps,

Spectral Angle Mapper, Band Ratios, Principal

Component Analysis and N-member extraction are

applied to extract meaningful features and patterns

from the hyperspectral data (Figures 5 and 6). These

processes facilitate the identification of potential

mineralogical and geochemical signatures within the

dataset. After data extraction, an automated process is

applied to balance and normalize the products for

each flight line, followed by the generation of the

final mosaic raster layer (Figure 7). All the automated

process is performed with an in-house and python-

based develop methods thanks to the availability of

publish methodologies for hyperspectral data

processing (De La Rosa et al., 2021, 2022) and open

source tools such as Spy Spectral Python library,

Mephysto (Jakob et al., 2017) and Hylite (Thiele et

al., 2020).

3.2 Self-Organizing Maps (SOM)

Self-Organizing Maps (SOM) are an unsupervised

neural network technique used to cluster input

evidence layer data while preserving its topological

structure. The training process relies on competitive

learning, where neurons compete to represent

different regions of the data space. Each neuron in the

SOM is associated with a weight vector, and when an

input vector is presented, the neuron with the closest

weight vector, known as the best matching unit

(BMU), is selected. The BMU and its neighboring

neurons are subsequently updated to better match the

input vector. Repeating this process across all input

data results in a self-organized map where similar

input vectors form distinct clusters, providing an

intuitive representation of underlying patterns in the

data (Kohonen 1990, 1997, 2001) and (Wittek et al.,

2017).

Visualization of SOM results facilitates cluster

interpretation through various techniques such as

color-coded maps and distance matrices, which

highlight similarities among spectral signatures. To

refine the clustering, k-means clustering is applied

post-SOM computation. The iterative k-means

algorithm randomly creates k centroids and assigns

the data points to the nearest centroid. Then, it

recalculates the centroids based on the mean of all

points within a particular cluster and repeats this

process until convergence. Multiple runs are

performed across a user-defined range of cluster

numbers, with the best clustering results determined

using the Davies-Bouldin index (David & Bouldin

1979). The three most optimal clustering outcomes

are displayed in the user interface and stored for

further analysis.

Figure 7: Automated mosaic raster layer for a Normalized

Carbonate Index derived from band ratios.

SOM results are visualized in geospatial and SOM

spaces (Figure 8). Furthermore, the results are plotted

and categorized into SOM space results, geo-space

results, boxplots and scatterplots. SOM space plots

include heatmaps representing the value of each

codebook vector element, the U-matrix showing

differences between neighboring SOM cells, k-means

clustering results, and data point distributions per

SOM cell. Geospace visualizations present k-means

clustering results, BMU codebook vectors, and

quantization errors in a geographical context.

Additionally, boxplots illustrate the distribution of

SOM data parameters across k-means clusters, while

scatterplots provide cross-plots of different

parameters, enabling deeper insight into data

relationships. This combination of SOM and k-means

clustering offers a powerful tool for pattern

recognition and mineral prospectivity analysis.

S34I 2025 - Special Session on S34I - From the Sky to the Soil

322

3.3 Bayesian Neural Networks (BNN)

Bayesian Neural Networks (BNNs) can be a class of

deep learning models when implementing multi-

layered network architecture. This Neural networks

integrate Bayesian inference to predict mineral

prospectivity while quantifying associated

uncertainties. Unlike traditional artificial neural

networks (ANNs), which provide deterministic point

estimates (Rosenblatt, 1958), BNNs estimate a

distribution over model parameters, enabling a

rigorous assessment of uncertainty (Jordão et al.,

2023). This capability is particularly valuable in high-

risk applications such as mineral exploration, where

uncertainty estimation enhances decision-making. In

this study, a BNN was trained using hyperspectral

derived features layers, geological and structural

derived input layers alongside positive samples

derived from the geochemical analysis and negative

labels derived from a random selection inside areas

delimited by the self-organizing maps analysis.

The BNN represents the model weights and biases

as probability distributions rather than fixed values

and through Bayesian updating, these distributions

are refined based on observed data, allowing the

model to learn while maintaining an explicit

quantification of uncertainty (Mao et al., 2023).

Variational inference is employed to approximate the

posterior distribution over model parameters,

facilitating efficient learning. The implementation of

the BNN model is develop in house as a python-based

tool and utilizes the TensorFlow Probability python

library to construct, train, and evaluate BNN

architectures.

The BNN model is still under development and

continuous improvement. This model was developed

in part through the Critical Mineral Assessments with

AI support (Critical MAAS) project. This project is a

collaboration between our company Beak

Consultants GmbH, the United States Geological

Survey (USGS) and the Defense Advanced Research

Projects Agency (DARPA). The Critical MAAS

project aims to accelerate critical mineral resource

assessments through re-design, automation, and

human-centered AI engineering. The work developed

in the frame of the project is classified as fundamental

research, and the code is open-source and available in

the following GitHub repository: https://github.com/

DARPA-CRITICALMAAS/beak-ta3.

Key Bayesian contributions in the code include

the incorporation of prior knowledge through prior

distributions, the application of variational inference

for posterior approximation, and the estimation of

predictive uncertainty.

3.4 Data-Driven Negative Sampling

To address the challenge of obtaining reliable

negative labels for training the BNN, we employed a

data-driven approach using the SOM outputs. By co-

registering the SOM clusters with known mineral

occurrences (positive labels), we identified areas

likely to be barren (negative labels) for targeted

sampling. This approach ensured that the negative

labels used for training the BNN were representative

of the true background variability in the study area.

4 RESULTS

4.1 Som Results

Figure 8: Unsupervised clustering results from SOM.

The SOM analysis successfully clustered the

hyperspectral data, the geological and structural-

based evidence layers into distinct groups

highlighting areas with similar characteristics. By co-

registering the SOM clusters with known mineral

occurrences, we identified this clusters as ‘very

likely’ to be areas showing characteristics that could

be associated with the presence of mineralized

samples and therefore, we exclude them and retain the

rest of clusters that are identified (potentially) areas

likely to be barren. These potentially barren clusters

are the ones chosen for negative labels selection for

training the BNN.

Innovative Hyperspectral Data Fusion for Enhanced Mineral Prospectivity Mapping

323

4.2 BNN Results

The BNN, trained with the hyperspectral features and

the positive and negative labels, generated a mineral

prospectivity map with associated uncertainty

estimates. The map highlighted areas with high

mineral potential, guiding future exploration efforts.

In Figure 9 and 10, the warmer colors near orange and

red represent the areas with the highest prospectivity,

where the prospectivity values are close to one. The

resulting mineral prospectivity maps can be

interpreted as follows: areas with values greater than

0.5 are the most prospective. These are the areas

where mineralization is most likely to be found. In

this case, the mineralization of interest is the Cu-Co-

Ni mineral association.

Figure 9: BNN prospectivity mapping results.

The results in Figure 9 and 10 also reveal an

interesting pattern: many of the prospective areas are

aligned with important structural features. These

include sections of the E-W Aramo fault and spatial

associations with lines representing the Aramo thrust

front. This association between prospective areas and

structural features corroborates the geological

understanding of the area, which suggests structural

control over mineralization. However, due to the

complex structural nature of the area, it is challenging

to identify this association based solely on geological

observation. These results can help guide future field

efforts to validate these findings and improve our

understanding of the factors controlling

mineralization in the area. The uncertainty estimates

provided a measure of confidence in the predictions,

allowing for more informed decision-making.

Figure 10: Zoom-in to BNN prospectivity mapping results.

5 DISCUSSION

Our research highlights the feasibility and advantages

of integrating hyperspectral data from multiple

sources with deep learning techniques for mineral

prospectivity mapping. This approach surpasses

traditional methods by offering a non-invasive, high-

resolution, cost-effective, and highly accurate

alternative for identifying potential mineral deposits.

By leveraging remote sensing and machine learning,

it minimizes environmental impact, reduces

exploration costs, and enhances predictive reliability,

making it particularly suitable for early-stage

exploration and challenging terrains.

A key innovation in our methodology is the

introduction of data-driven negative sampling, a

critical step in training BNN models for mineral

prospectivity mapping. Negative sampling is a well-

known challenge in machine learning applications, as

incorrectly labeled negative samples can significantly

degrade model performance. Traditional methods

often rely on random sampling or expert-defined

barren areas, which may not adequately capture the

true background variability. To overcome this, we

S34I 2025 - Special Session on S34I - From the Sky to the Soil

324

employed a systematic data-driven approach using

Self-Organizing Maps (SOM) to generate reliable

negative labels. By co-registering SOM clusters with

known mineral occurrences (positive labels), we

identified regions highly likely to be barren (negative

labels). This ensured that the training data more

accurately reflected the real geological variability of

the study area, improving model robustness and

reducing bias in mineral prospectivity predictions.

Another major contribution of our study is the use

of Bayesian Neural Networks (BNNs) for predictive

modeling and uncertainty quantification. Unlike

conventional artificial neural networks (ANNs),

which provide only point estimates, BNNs estimate a

probability distribution over model parameters,

allowing them to quantify the uncertainty in their

predictions. This is particularly valuable in mineral

exploration, where decision-making is inherently

uncertain and high-risk. The Bayesian framework

enables the estimation of uncertainty in model outputs

(Figure 11), offering a confidence measure for each

prospectivity prediction. This allows for more

strategic resource allocation, as exploration efforts

can be prioritized in areas with high predictive

confidence while regions with high uncertainty can be

flagged for further data acquisition. By integrating

uncertainty quantification directly into the model, our

approach provides a more transparent and

interpretable decision-support system, reducing the

risk of false positives and missed discoveries.

Figure 11: BNN Uncertainty associated to prospectivity

mapping results.

Beyond its theoretical advantages, our approach

offers several practical benefits over traditional

methods. First, its non-invasive nature minimizes

environmental impact by reducing reliance on

intrusive ground surveys such as drilling and

trenching. This is particularly important in

ecologically sensitive or remote areas where physical

access is limited. Second, the high-resolution spectral

information from hyperspectral imaging, combined

with the BNN and SOM, allows for the identification

of subtle features, leading to more accurate and

efficient exploration efforts. Third, the cost-

effectiveness of our methodology is significant; by

automating feature extraction, we reduce the need for

extensive manual interpretation from the

hyperspectral data, and the high prospective areas can

guide a more targeted oriented surface exploration,

cutting exploration costs substantially.

It is also important to highlight the significance of

the quantity and quality of training samples,

particularly the positive samples. Although the

methodology presented here offers an improved

solution as a method for data-driven negative

sampling selection, the importance of positive

samples cannot be overstated. The positive samples

are the single input data that will most significantly

affect the results of the BNN models. In real-world

scenarios, the quantity and quality of samples are not

always optimal, as exemplified by the Aramo study

case presented in this publication. The very

challenging climatic conditions characteristic of this

area in Spain resulted in several flight lines of the

planned airborne hyperspectral acquisition being not

possible to fly (this is the major data gap observed in

Figure 2, 7, 8 and 9). Unfortunately, this area

coincides with the location exhibiting the clearest

surface mineralization and where the majority of the

samples intended for training points were located. As

shown in Figure 8 and 9, most of the training data

coincides with this gap. Furthermore, a significant

percentage of the remaining acquired data was

obscured by clouds, rendering the spectral

information unusable and necessitating the

development of an automatic algorithm to mask

cloud-covered areas, further reducing the available

data. Therefore, a dataset with greater spatial

coverage and a larger number of training samples in

areas with available data would greatly enhance the

quality of the results.

In the context of the growing global demand for

critical raw materials, this research contributes to the

integration of multi-scale hyperspectral remote

sensing with BNN and SOM presenting an innovative

workflow for data fusion and prospectivity mapping

Innovative Hyperspectral Data Fusion for Enhanced Mineral Prospectivity Mapping

325

with uncertainty quantification aiming to improve

mineral exploration, offering a scalable and data-

driven solution. As hyperspectral imaging technology

advances and more high-quality datasets become

available, the accuracy and effectiveness of this

method will continue to improve. Additionally,

ongoing developments in deep learning architectures,

Bayesian inference, and self-supervised learning will

further enhance predictive capabilities and

uncertainty quantification.

Our findings emphasize the importance of data-

driven approaches in addressing key challenges in

training data selection and model interpretability. The

combination of SOM-based negative label generation

and BNN-driven uncertainty estimation provides a

novel framework for improving the reliability and

confidence of mineral prospectivity predictions. This

methodology not only enhances the accuracy of the

models but also offers a structured approach to

handling uncertainty, making it a powerful tool for

risk-aware decision-making in exploration projects.

6 CONCLUSIONS

Our innovative workflow for mineral prospectivity

mapping supports the objectives of the EU's S34I

project by providing high-value, unbiased datasets

and improving the perception of mining through the

application of cutting-edge, sustainable exploration

technologies.

In conclusion, this study represents an

advancement in mineral exploration by providing a

scientifically rigorous, scalable, and environmentally

responsible approach to identifying potential mineral

deposits. The integration of hyperspectral data, SOM-

driven negative sampling, and Bayesian Neural

Networks has proven to improve exploration

strategies, supporting a sustainable and efficient

pathway to securing critical raw materials for a green

energy future.

In the Aramo study case, the mineral prospectivity

maps reveal an interesting pattern, showing many

prospective areas aligned along important structural

features, including sections of the E-W Aramo fault

and the Aramo thrust front. This alignment

corroborates the area's geological understanding,

which suggests that mineralization is in some degree

structurally controlled. These results can guide future

fieldwork to validate these findings and enhance our

understanding of the factors controlling

mineralization in the area.

The integration of advanced deep learning and

remote sensing data not only accelerates the

exploration process but also significantly reduces

costs and environmental impact. This approach has

the potential to transform mineral exploration,

supporting the sustainable and responsible sourcing

of critical raw materials for the EU's green energy

transition.

ACKNOWLEDGEMENTS

This research has been done within the framework of

the project S34I—Secure and sustainable supply of

raw materials for EU industry¬, coordinated by Ana

C. Teodoro. This project has received funding from

the European Union’s HORIZON Research and

Innovation. Grant Agreement No. 101091616

(https://doi.org/10.3030/101091616). Portuguese

National Funds also support this work through the

FCT – Fundação para a Ciência e a Tecnologia, I.P.

(Portugal), projects UIDB/04683/2020

(https://doi.org/10.54499/UIDB/04683/2020) and

UIDP/04683/2020 (https://doi.org/10.54499/UIDP/

04683/2020).

REFERENCES

Beato Bergua, S., Poblete Piedrabuena, M.A., & Marino

Alfonso, J.L (2019). Relieve estructural y karst en la

Sierra del Aramo (Macizo Central Asturiano).

Investigaciones Geográficas, (72), 75-99.

https://doi.org/10.14198/INGEO2019.72.04.

Bruner, K. R., & Smosna, R. (2000). Stratigraphic-tectonic

relations in Spain's Cantabrian mountains: Fan delta

meets carbonate shelf. Journal of Sedimentary

Research, 70(6), 1302–1314.

Carranza, E. J. M. (2008). Geochemical anomaly and

mineral prospectivity mapping in GIS. In Handbook of

Exploration and Environmental Geochemistry (Vol.

11). Elsevier.

Carranza, E. J. M., & Hale, M. (2001). Logistic regression

for constrained mapping of gold potential, Baguio

district, Philippines. Exploration and Mining Geology,

10(3), 165–175.

Carvalho, M.; Cardoso-Fernandes, J.; González, F.J.;

Teodoro, A.C. Comparative Performance of Sentinel-2

and Landsat-9 Data for Raw Materials' Exploration

Onshore and in Coastal Areas. Remote Sens. 2025, 17,

305. https://doi.org/10.3390/rs17020305.

Davies, David L.; Bouldin, Donald W. (1979). "A Cluster

Separation Measure". IEEE Transactions on Pattern

Analysis and Machine Intelligence. PAMI-1 (2): 224–

227. doi:10.1109/TPAMI.1979.4766909.

De La Rosa, R., Khodadadzadeh, M., Tusa, L., Kirsch, M.,

Gisbert, G., Tornos, F., Tolosana-Delgado, R., &

Gloaguen, R. (2021). Mineral quantification at deposit

S34I 2025 - Special Session on S34I - From the Sky to the Soil

326

scale using drill-core hyperspectral data: A case study

in the Iberian Pyrite Belt. Ore Geology Reviews, 139 Pt

B, 104514. http://dx.doi.org/10.1016/j.oregeorev.

2021.104514.

De La Rosa, R., Tolosana-Delgado, R., Kirsch, M., &

Gloaguen, R. (2022). Automated multi-scale and

multivariate geological logging from drill-core

hyperspectral data. Remote Sensing, 14, 2676.

https://doi.org/10.3390/rs14112676.

Jakob, S., Zimmermann, R., Gloaguen, R., (2017). The

need for accurate geometric and radiometric corrections

of drone-borne hyperspectral data for mineral

exploration: MEPHySTo—a toolbox for pre-

processing drone-borne hyperspectral data. Remote

Sensing, 9(1), 88. https://doi.org/10.3390/rs9010088.

Jordão, H., Sousa, A. J., & Soares, A. (2023). Using

Bayesian neural networks for uncertainty assessment of

ore type boundaries in complex geological models.

Natural Resources Research, 32(6), 2495–2514.

https://doi.org/10.1007/s11053-023-10265-6.

Kohonen, T. (1990). The Self-Organizing Map.

Proceedings of the IEEE, 78(9), 1464–1480.

Kohonen, T. (1997). Exploration of very large databases by

Self-Organizing Maps. Proceedings of the IEEE

International Conference on Neural Networks, 4, PL1-

PL6.

Kohonen T (2001). Self-Organizing Maps. Springer-

Verlag. doi:10.1007/978-3-642-56927-2.

Lauzon, D., & Gloaguen, E. (2024). Quantifying

uncertainty and improving prospectivity mapping in

mineral belts using transfer learning and Random

Forest: A case study of copper mineralization in the

Superior Craton Province, Quebec, Canada. Ore

Geology Reviews, 105918. DOI:

https://doi.org/10.1016/j.oregeorev.2024.105918.

Lawley, C. J. M., McCafferty, A. E., Graham, G. E.,

Huston, D. L., Kelley, K. D., Czarnota, K., Paradis, S.,

Peter, J. M., Hayward, N., Barlow, M., Emsbo, P.,

Coyan, J., San Juan, C. A., & Gadd, M. G. (2022). Data-

driven prospectivity modelling of sediment–hosted Zn–

Pb mineral systems and their critical raw materials. Ore

Geology Reviews, 141, 104635.

Loredo, J., Álvarez, R., & Ordóñez, A. (2008). Mineralogy

and geochemistry of the Texeo Cu-Co mine site (NW

Spain): Screening tools for environmental assessment.

Environmental Geology, 55(6), 1299–1310.

Mao, X., Wang, J., Deng, H., Liu, Z., Chen, J., Wang, C.,

& Liu, J. (2023). Bayesian decomposition modelling:

An interpretable nonlinear approach for mineral

prospectivity mapping. Mathematical Geosciences,

55(5), 897–942. https://doi.org/10.1007/s11004-023-

10067-9.

Nagasingha, L. M. A., Bérubé, C. L., & Lawley, C. J. M.

(2024). A balanced mineral prospectivity model of

Canadian magmatic Ni (+ Cu + Co + PGE) sulphide

mineral systems using conditional variational

autoencoders. Ore Geology Reviews, 175, 106329.

DOI: https://doi.org/10.1016/j.oregeorev.2024.106329.

Nykänen, V., Niiranen, T., Molnár, F., Lahti, I., Korhonen,

K., Cook, N., & Skyttä, P. (2017). Optimizing a

knowledge-driven prospectivity model for gold

deposits within Peräpohja Belt, Northern Finland.

Natural Resources Research, 57, 571–584.

Nykänen, V., Törmänen, T., & Niiranen, T. (2023). Cobalt

prospectivity using a conceptual fuzzy logic overlay

method enhanced with the mineral systems approach.

Natural Resources Research, 1-29. https://doi.org/

10.1007/s11053-023-10255-8.

Ordóñez, A., Álvarez, R., Bros, T., & Loredo, J. (2005).

Consequences of abandoned Cu-Co mining in Northern

Spain in surface watercourses. Proceedings of the 9th

International Mine Water Congress, 611–617.

Paniagua, A., Fontboté, L., Fenoll Hach-Alí, P., Fallick, A.

E., Moreiras, D. B., & Corretgé, L. G. (1993). Tectonic

setting, mineralogical characteristics, geochemical

signatures and age dating of a new type of epithermal

carbonate-hosted, precious metal-five element deposits:

The Villamanín area (Cantabrian zone, Northern

Spain). Current Research in Geology Applied to Ore

Deposits, 531–534.

Paniagua, A., Loredo, J., & García-Iglesias, J. (1988).

Epithermal (Cu-Co-Ni) mineralization in the Aramo

mine (Cantabrian mountains, Spain): Correlation

between paragenetic and fluid inclusion data. Bulletin

de Minéralogie, 111, 383–391.

Rosenblatt, F. (1958). The perceptron: A probabilistic

model for information storage and organization in the

brain. Psychological Review, 65(6), 386-408.

Thiele, S., Lorenz, S., Kirsch, M., Gloaguen, R., (2020).

Hylite: a hyperspectral toolbox for open pit mapping.

EGU General Assembly 2020, Online, 4–8 May 2020,

p. 1. https://doi.org/10.5194/egusphere-egu2020-

13563.

Torppa, J., Nykänen, V., & Molnár, F. (2019).

Unsupervised clustering and empirical fuzzy

memberships for mineral prospectivity modelling. Ore

Geology Reviews, 107, 58-71.

https://doi.org/10.1016/j.oregeorev.2019.02.007.

Wittek, P., Gao, S. C., Lim, I. S., & Zhao, L. (2017).

somoclu: An efficient parallel library for self-

organizing maps. Journal of Statistical Software, 78(9),

1-22. doi:10.18637/jss.v078.109.

Yousefi, M., & Nykänen, V. (2016). Data-driven logistic-

based weighting of geochemical and geological

evidence layers in mineral prospectivity mapping.

Journal of Geochemical Exploration, 164, 94–106.

Yousefi, M., Carranza, E. J. M., Kreuzer, O. P., Nykänen,

V., Hronsky, J. M., & Mihalasky, M. J. (2021). Data

analysis methods for prospectivity modelling as applied

to mineral exploration targeting: State-of-the-art and

outlook. Journal of Geochemical Exploration, 229,

106839.

Yousefi, M., Lindsay, M. D., & Kreuzer, O. (2024).

Mitigating uncertainties in mineral exploration

targeting: Majority voting and confidence index

approaches in the context of an exploration information

system (EIS). Ore Geology Reviews, 165, 105930.

DOI: https://doi.org/10.1016/j.oregeorev.2024.105930.

Zhang, S. E., Lawley, C. J. M., Bourdeau, J. E., Nwaila, G.

T., & Ghorbani, Y. (2024). Workflow-induced

Innovative Hyperspectral Data Fusion for Enhanced Mineral Prospectivity Mapping

327

uncertainty in data-driven mineral prospectivity

mapping. Natural Resources Research, 1-26.

https://doi.org/10.1007/s11053-024-10322-8.

Zhang, Z., Wang, G., Carranza, E. J. M., Fan, J., Liu, X.,

Zhang, X., Dong, Y., Chang, X., & Sha, D. (2022). An

integrated framework for data-driven mineral

prospectivity mapping using bagging-based positive-

unlabeled learning and Bayesian cost-sensitive logistic

regression. Natural Resources Research, 31(6), 3041–

3060. https://doi.org/10.1007/s11053-022-10120-0.

S34I 2025 - Special Session on S34I - From the Sky to the Soil

328