PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Hojat Shirmard, Ehsan Farahbakhsh, R. Dietmar Muller, Rohitash Chandra

arXiv:2103.07678v2cs.LGcs.CVstat.AP

TL;DR

Mineral exploration needs more efficient ways to map geological features amid declining discoveries, rising mineral demand, and constraints on field-based mapping. This review categorizes machine learning methods and examines their use with remote sensing data for geological-feature and ore-deposit detection. It concludes that the combination has strong capability for mapping exploration features, while newer methods and data offer scope for improved mineral prospectivity maps.

  • Problem

    Mineral exploration requires efficient processing of diverse data to map geological features, while field data alone face common mapping problems.

  • Method

    The paper reviews and categorizes machine learning methods for processing satellite, airborne, and other remote sensing data in mineral exploration.

  • Results

    The review finds high capability in combining remote sensing and machine learning to map geological features used for mineral prospectivity mapping.

  • Takeaways & Limitations

    Advanced methods such as graph deep learning, Bayesian deep learning, variational autoencoders, and GANs provide scope for processing new-generation remote sensing data.

  • Takeaways & Limitations

    Deep learning requires integration of multiple knowledge and data sources to develop robust models.

Abstract

from arXiv · show

The decline of the number of newly discovered mineral deposits and increase in demand for different minerals in recent years has led exploration geologists to look for more efficient and innovative methods for processing different data types at each stage of mineral exploration. As a primary step, various features, such as lithological units, alteration types, structures, and indicator minerals, are mapped to aid decision-making in targeting ore deposits. Different types of remote sensing datasets, such as satellite and airborne data, make it possible to overcome common problems associated with mapping geological features. The rapid increase in the volume of remote sensing data obtained from different platforms has encouraged scientists to develop advanced, innovative, and robust data processing methodologies. Machine learning methods can help process a wide range of remote sensing datasets and determine the relationship between components such as the reflectance continuum and features of interest. These methods are robust in processing spectral and ground truth measurements against noise and uncertainties. In recent years, many studies have been carried out by supplementing geological surveys with remote sensing datasets, which is now prominent in geoscience research. This paper provides a comprehensive review of the implementation and adaptation of some popular and recently established machine learning methods for processing different types of remote sensing data and investigates their applications for detecting various ore deposit types. We demonstrate the high capability of combining remote sensing data and machine learning methods for mapping different geological features that are critical for providing potential maps. Moreover, we find there is scope for advanced methods to process the new generation of remote sensing data for creating improved mineral prospectivity maps.

1. Introduction

Mineral exploration depends on mapping geological features, but field-based mapping is constrained by accessibility and environmental conditions. Remote sensing combined with machine learning is presented as an increasingly important way to process geological data and support exploration.

  • Geological maps localize lithological units, alteration types, structures, and indicator minerals related to target mineralization.
  • Traditional mapping is difficult in inaccessible areas and is affected by climate, topography, field expertise, and operating approaches.
  • Remote sensing instruments provide data across different spatial, spectral, and temporal resolutions for geological mapping.
  • The review categorizes machine learning methods into dimensionality reduction, classification, clustering, regression, and deep learning.
  • It surveys remote sensing applications for mapping geological target features and discusses newer approaches including graph deep learning, Bayesian deep learning, and variational autoencoders.

2. Methodology

The methodology combines a categorization scheme with a literature review of machine learning applications to remote sensing-based detection of geological target features. Publication trends are examined using searches of Google Scholar and Scopus, with mineral-exploration applications showing increasing attention but a smaller literature base.

  • The review categorizes machine learning methods and surveys their use in remote sensing data processing for mineral exploration.
  • Searches used Google Scholar and Scopus with keywords covering machine learning, remote sensing, mineral exploration, and geological target features.
  • Publication counts were plotted for keyword combinations covering general remote sensing, minerals, and mineral exploration.
  • The mineral-exploration literature is smaller than the broader machine-learning and remote-sensing literature, but both show an increasing trend.

3. Remote sensing data

The review covers remote sensing data from satellite, airborne, and ground-based platforms used to map geological features. It distinguishes passive optical and active radar systems and summarizes their spectral, spatial, temporal, and coverage characteristics.

  • Remote sensing data used in mineral exploration are acquired by satellite, airborne, and ground-based instruments.
  • Satellite systems include passive optical sensors and active radar systems, which constitute major data types for geological mapping.
  • Landsat, Sentinel, SPOT, and Google Earth provide optical imagery used for geological observation and mapping.
  • Table 1 organizes sensor characteristics such as spectral range, ground resolution, swath width, and launch or operation year.
  • Airborne platforms provide multispectral and hyperspectral data through sensors including AMSS and AVIRIS-NG.
  • Ground-based hyperspectral sensing commonly targets VNIR and SWIR wavelengths and can provide approximately 10 nm spectral resolution for rapid characterization.

4. Target features

Remote sensing-based mineral exploration targets lithology, structure, alteration, and mineralization-related indicator features. These targets reflect geological controls on ore deposits and can be mapped using spectral, image-processing, and classification methods.

  • Mineral exploration image processing commonly targets lithological, structural, alteration, and mineralization zones.
  • Porphyry copper models link hydrothermal alteration types with associated ores and interpreted exposure levels in ASTER-mapped Iranian sites.
  • Identifying and discriminating rock units is fundamental because deposit types occur in distinct geological settings and host rocks.
  • Alteration patterns vary among deposit systems, making altered rocks and alteration halos important mapping targets.
  • Faults, veins, shear zones, and lineaments support exploration of structurally controlled mineralization and can be extracted from radar and optical data.
  • Spectral analyses and supervised classifiers are used to identify indicator minerals and mineralization zones associated with porphyry copper, epithermal gold, and VMS deposits.

5. Machine learning

Machine learning methods provide tools for mapping geological features from remote sensing data, including approaches that learn without labels, model complex patterns, or combine multiple models.

  • Remote sensing and machine learning can map lithological units, alteration zones, structures, and indicator minerals associated with mineral deposits.
  • Unsupervised learning recognizes patterns without target labels, using methods such as clustering and principal component analysis.
  • Deep learning models high-level data abstractions through multiple processing layers and includes CNNs, RNNs, autoencoders, and related architectures.
  • Ensemble methods integrate multiple models for classification or regression and typically outperform standalone methods.
  • Table 2 compiles sample studies using machine learning and remote sensing to map potential mineralization zones.

5.1. Dimensionality reduction techniques

Dimensionality reduction transforms correlated remote sensing variables into uncorrelated or independent components and has been applied across diverse spectral imaging datasets.

  • PCA, ICA, and MNF transform correlated input variables into uncorrelated or independent components for remote sensing processing.
  • PCA and MNF applied to Landsat 8 imagery supported lithological and alteration mapping for alluvial gold exploration.
  • Recent datasets included UAS-borne, terrestrial, and other spectral imaging acquired across VNIR, SWIR, and LWIR resolutions.

5.2. Classification

Classification methods are used to map geological features from remote sensing data, with reviewed applications spanning conventional, margin-based, neural, and ensemble approaches.

  • The review focuses on machine learning methods used for classification problems in remote sensing mineral exploration.
  • 5.2.1. Minimum distance classification: Minimum distance classification assigns pixels according to their Euclidean distance from each class’s mean vector in multidimensional space.
  • 5.2.2. Support vector machines: SVM projects data into a lower-dimensional feature space to simplify classification and has been applied to mineral exploration.
  • 5.2.2. Support vector machines: Jointly processing reflectance, at-sensor temperature, texture, and geomorphic parameters produced high-precision classification layers and identified a new 0.3 km^2 chromite-bearing site.
  • 5.2.2. Support vector machines: SVM-based studies reliably delineated alteration zones and outlined favorable gold metallogenic areas using remote sensing and field inspection.
  • Neural networks learn complex patterns and nonlinear relationships between explanatory and dependent variables in remote sensing analysis.
  • Decision trees handle large, complicated datasets without imposing a complicated parametric structure, but can overfit and generalize poorly.
  • Random forests have been applied to classify rock units using airborne polarimetric and geophysical data, while SAR can assist mapping beneath vegetation canopies.

5.3. Clustering

Clustering methods group remote sensing data without requiring labeled targets, but mixed pixels can complicate conventional clustering and motivate improved algorithms.

  • K-means assigns data points to the cluster with the smallest discrepancy from its mean and can process high-dimensional datasets.
  • Mixed pixels can complicate alignment between data points and cluster centers in standard K-means clustering.
  • An improved K-means algorithm used multiple initialization strategies and spectral-information similarity instead of Euclidean distance.
  • The enhanced algorithm produced stronger clustering results and greater mineral mapping precision than the conventional algorithm.
  • Spectral clustering identifies subgraphs from node connections, while ISODATA is an unsupervised classification approach applied to satellite imagery.

5.4. Regression analysis

Regression methods support statistical prediction and mapping of mineralization-related targets from remote sensing and integrated geological data. Applications include alteration, iron outcrop, and skarn deposit mapping, while uncertainty quantification remains limited.

  • Regression models statistically relate predictor variables to target values and can predict mineralized zones from remote sensing data.
  • Multilinear regression mapped alteration trends and distinguished potassic, phyllic, propylitic, argillic, and silicification zones.The mapped alteration patterns also identified buried faults that may indicate alteration sources.
  • Multivariate regression was used to construct statistical models for mapping iron outcrops and other geological targets.
  • Conjugate gradient logistic regression integrated geological, remote sensing, and geochemical evidential variables to predict skarn deposits.
  • Prediction uncertainty arises from model parameters, noise, sparse datasets, and sensor limitations, but few studies quantify it.The review identifies Bayesian regression as a potential direction for uncertainty analysis.

5.5. Deep learning

Deep learning methods can model complex remote sensing datasets, with CNNs especially prominent for image-based mineral exploration. Their applications include automatic feature extraction, spatial-spectral fusion, and transfer across fields, while recurrent approaches remain less developed.

  • Only a few studies have applied deep learning to remote sensing data processing for mineral exploration.
  • Deep learning supports supervised, unsupervised, and semi-supervised learning while modeling complex and large datasets.Semi-supervised learning uses a small labeled subset alongside many unlabeled examples.
  • CNNs are prominent because convolutional and pooling layers automatically extract image features.
  • A CNN model combined local spatial characteristics with spectral data and was extended to fields excluded from training.The study used aerial imagery, Landsat reflectance, and high-resolution digital elevation data across five areas.
  • LSTM networks model temporal data and dynamical systems, but their use in mineral exploration remains limited.The review also notes combinations such as LSTM-CNN for spatiotemporal datasets.

6. Discussion: challenges and future prospective

Mineral exploration is costly and risky, motivating the use of remote sensing and machine learning to improve geological targeting. The review highlights uncertainty, validation, data integration, class imbalance, missing regions, and advanced deep-learning methods as continuing challenges and opportunities.

  • Declining discovery success and increasing demand for critical metals motivate new data types and machine-learning approaches in mineral exploration.Exploration seeks economically productive deposits at low cost and within a short period, despite low success rates and returns on investment.
  • Remote sensing and machine learning can support identification of target features including lithological units, alteration types, structures, and indicator minerals.
  • Uncertainty quantification remains a gap because remote-sensing model validation commonly relies on ground truth or alternative information presumed to represent it.
  • Autoencoders may compress and reconstruct remote sensing data, while GANs may address class imbalance or reconstruct missing regions.
  • Integrating diverse knowledge and data sources is identified as necessary for developing robust models.

7. Conclusions

The review finds that combining remote sensing with machine learning, particularly advanced deep learning, can support geological-feature mapping and mineral discovery. It also identifies scope for processing new-generation remote sensing data to improve mineral prospectivity maps.

  • The review covers popular and recently established machine-learning methods for remote sensing processing and their applications to different ore deposits.
  • Remote sensing provides an additional data resource for overcoming problems in mapping geological features from field data alone.
  • Dimensionality reduction can transform high-dimensional remote sensing problems into lower-dimensional spaces and potentially extract useful mineral-exploration features.
  • Advanced deep learning methods can process large and complex remote sensing data involving spectral and ground truth measurements against noise and uncertainties.
  • The review identifies scope for advanced methods to process new-generation remote sensing data for improved mineral prospectivity maps.
Loading 2103.07678v2…