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

In plain words

Mineral explorers look for ore by mapping the rock types, the chemical alteration of rock, the faults and the surface minerals of a region. Satellites, aircraft, drones and ground-based instruments record these features as spectra and images, and the amount of such data keeps growing. This review asks how machine learning, which learns patterns from data and makes predictions from them, can turn that data into maps. It sorts the common methods into five groups, shows where each has been tried for ore-related mapping, and points to newer deep learning approaches that the authors think are underused in this field. It is a review, so it runs no new experiments.

TL;DR

• A review, not a method: it runs no experiments and gives no equations, algorithms or training settings §1, §6.
• Covers optical (Landsat, ASTER, Sentinel-2, WorldView, Hyperion), radar (RADARSAT, Sentinel-1, PALSAR), airborne (AVIRIS-NG, HyMap, UAV) and ground-based hyperspectral data §3.
• Five method groups: dimensionality reduction, classification, clustering, regression and deep learning §5.
• Use it as a map of which sensors and methods were tried for lithology, alteration, structures and minerals, and which gaps the authors name §5, §6. Do not use it as an implementation guide.

Problem

Mapping rock units, alteration zones, structures and minerals in the field is slow, hard in remote areas, affected by climate and terrain, and subject to the judgement of the surveyor §1. Remote sensing covers larger areas, but its volume and complexity call for new processing methods §1, §6. The authors motivate the review with a falling rate of new discoveries and rising demand for critical metals §6.

The review assumes that the best method depends on the data and the problem. It notes that no single approach solves every case, and that the choice must be tested per dataset §5, §6.

Method

This is a literature review. The authors searched Google Scholar and Scopus with the keywords "machine learning", "remote sensing" and "mineral exploration", plus terms for lithological units, alteration types, structures and indicator minerals §2. The number of papers found, the date range and the inclusion rules are not stated §2. Figure 1 plots Scopus publication counts over the last decade for three keyword sets; the most specific set is far below the broadest one. The counts appear only in the figure §2, fig. 1.

The paper gives no algorithm. The sketch below restates the workflow that the sections and Figure 3 describe §3–§6, fig. 3. It is our summary, not the authors' pseudocode.

# Mineral-mapping workflow as the review describes it  [§3-§6, fig. 3]
# Input:  remote sensing scenes (satellite, airborne or ground); field samples
# Output: a map of a target feature, e.g. a prospectivity map

def map_targets(scenes, field_samples):
    # 1. Choose the platform, resolution and bands  [§3]
    data = select_bands(scenes)
    # 2. Reduce correlated bands, optionally: PCA, ICA or MNF  [§5.1]
    data = reduce_dimensions(data)
    # 3. Fit one method on the samples; labels optional for clustering  [§5]
    model = fit(method, data, field_samples)
    # 4. Predict the target feature for each pixel  [§4, §5]
    target_map = predict(model, data)
    # 5. Compare the map with ground truth or known geology  [§6]
    validate(target_map, field_samples)
    return target_map

Remote sensing data §3. Each platform is described by its sensors, spectral bands and resolutions. Table 1 lists bands, ground resolution, swath width and launch year for the main sensors tab. 1.

1. Satellite optical §3.1.1:
   • Landsat 5 (TM, 1984), Landsat 7 (ETM+, 1999) and Landsat 8 (OLI and TIRS, 2013). Landsat 7 has 30 m VNIR and SWIR bands, a 15 m panchromatic band and a 60 m thermal band §3.1.1, tab. 1.
   • ASTER on Terra: 15 m VNIR bands, 30 m SWIR bands and 90 m thermal bands in Table 1; the text gives 15 m and 30 m only §3.1.1, tab. 1.
   • Sentinel-2: 10 m, 20 m and 60 m bands; 13 VNIR and SWIR bands §3.1.1.
   • Commercial very-high-resolution sensors: WorldView-1 (50 cm panchromatic), WorldView-2 (46 cm panchromatic, eight multispectral bands) and WorldView-3 (16 bands, with eight SWIR bands at 3.7 m and eight VNIR bands at 1.2 m) §3.1.1, tab. 1.
   • Hyperion: 242 bands from 0.36 to 2.58 µm at about 10 nm spectral and 30 m spatial resolution §3.1.1.
2. Satellite radar §3.1.2:
   • Sentinel-1 C-band SAR with dual polarisation and 5 × 20 m resolution §3.1.2.
   • PALSAR L-band SAR with fine, polarimetric and ScanSAR modes at 10, 30 and 100 m §3.1.2.
3. Airborne §3.2:
   • AVIRIS-NG: 427 contiguous 5 nm bands from 0.38 to 2.51 µm, 8.1 m spatial resolution §3.2.
   • HyMap: 126 bands from 0.45 to 2.48 µm, 2 to 10 m resolution §3.2.
   • UAV-borne visible, NIR, thermal, multispectral and hyperspectral cameras, laser scanners and SAR; UAV hyperspectral imaging can reach millimetre to tens-of-centimetre resolution §3.2.
4. Ground-based §3.3:
   • HySpex: 160 bands from 410 to 990 nm at 3.7 nm sampling, 12-bit §3.3.
   • UHD 285 snapshot camera: 450 to 950 nm, 1 ms integration per cube §3.3.

Target features §4. The review names four feature types in mineral exploration:

• Lithology: rock units tied to ore systems such as VMS, sediment-hosted lead-zinc and porphyry and epithermal Cu, Mo and Au deposits §4.1.
• Alteration: argillic, phyllic, propylitic, advanced argillic and silicic zones, which indicate porphyry and epithermal deposits §4.2.
• Structures: faults, dykes and lineaments, mapped from optical and radar data §4.3.
• Minerals: clays, sulfates, carbonates, iron oxides and silica, which mark alteration facies §4.4.

Figure 2 is a schematic porphyry-copper deposit model with alteration zones, ore associations and an Iranian exposure scale fig. 2.

Machine learning methods §5. Table 2 lists sample studies by method, platform, target feature and commodity tab. 2. The selection rule for these samples is not stated. The table is summarised below.

Summary of tab. 2.

Per group, as described in the text:

1. Dimensionality reduction §5.1. PCA, ICA and MNF transform correlated bands into uncorrelated or independent components. The authors report using them for lithology, alteration and structures, and show a classified map with known occurrences overlaid §5.1, fig. 4.
2. Minimum distance §5.2.1. Each pixel is assigned to the class whose mean vector is nearest in Euclidean distance §5.2.1. The authors say it is seldom used in this field and has poor accuracy in general; the source of the accuracy claim is not stated §5.2.1.
3. Support vector machines §5.2.2. Described as a robust supervised classifier for classification and regression, applied to ASTER, Hyperion and Sentinel-2 data for chromite, gold, IOCG and lithium mapping §5.2.2.
4. Neural networks §5.2.3. Multi-layer perceptrons applied to gold potential, Mo and Pb-Zn-Ag potential and lineament mapping §5.2.3.
5. Random forests §5.2.4. Bagging ensemble of decision trees. The vote gives the class for classification; the mean of the trees gives the value for regression §5.2.4. The text says the method suits large, complicated datasets without a parametric form §5.2.4.
6. Clustering §5.3. K-means assigns each point to the nearest cluster centre §5.3. The review also covers agglomerative clustering, DBSCAN, spectral clustering and ISODATA §5.3.
7. Regression §5.4. Linear, multinomial, logistic and boosted regression trees are applied to alteration and mineral mapping. The authors say Bayesian regression has seen little use so far and see scope for it §5.4.
8. Deep learning §5.5. CNNs for image classification and lithology mapping, RNNs and LSTMs for spatiotemporal data, ConvLSTM, and GNN and graph-convolution models for scene classification. The authors state that few studies apply deep learning to mineral exploration §5.5.

Discussion §6. The authors name five open issues: uncertainty quantification, which is rarely done; the cost of high-dimensional data; over-fitting and problem dependence; a lack of guidelines for model settings and interpretation; and the open question of how transfer learning, autoencoders, GANs and Bayesian deep learning can serve remote sensing in mineral exploration. The authors call these promising, and say Bayesian deep learning has not been used for remote sensing in this field §6. They also propose spatio-temporal tectonic data as a way to add context §6.

Hyperparameters

Not applicable: the paper trains no model. No settings are stated for the reviewed studies, including kernel choice, tree count, cluster count K, network depth or train/test split §5. Table 2 lists the methods but no parameters tab. 2.

Results

This review reports no new results. The few quantitative figures it repeats come from the cited studies and are author-reported there:

Other comparisons are qualitative. For example, an SVM was reported to have better precision than maximum likelihood on lithology, with no number given §4.1. The Figure 1 publication counts are shown only in the figure §2, fig. 1.

Paper vs code

No official code. The paper names no repository, no code link and no project page, and the arXiv metadata carries no code link. A GitHub repository search for "remote sensing mineral exploration machine learning" returned no results, and a search for "mineral exploration remote sensing" returned eight repositories. None is linked to the authors, and none is cited here.

Pitfalls

• Search protocol missing. The database search gives no date range, number of hits, screening steps or inclusion rules §2. The sample list in Table 2 has no selection criterion tab. 2. Do not read it as exhaustive or as a ranking.
• Numbers are not comparable. The two figures in Results come from different sensors, metrics and baselines, and the paper does not define the 74% metric §5.1, §5.5. Do not compare them with each other or with other cards.
• Resolution mismatch for WorldView-2. The text gives 1.85 m for the VNIR bands §3.1.1; Table 1 gives 1.84 m tab. 1.
• Category boundaries are the authors' choice. Logistic regression appears under regression §5.4, tab. 2, although it is usually a classifier, and the text says the classification section focuses on classification while neural networks and random forests also regress §5.2.
• Mismatch for Cracknell et al. §5.2.4 says Cracknell2014 compared five classifiers, but Table 2 lists that study only under random forest §5.2.4, tab. 2.
• SVM description (our note, not from the paper). The text says the SVM projects data into a space of smaller dimension §5.2.2. Kernel SVMs usually map to a higher-dimensional space. The paper gives no kernel, regularisation or class-weight settings.
• Opinion versus evidence. Statements such as "poor classification accuracy" for minimum distance §5.2.1 and "few studies" for deep learning §5.5 are the authors' view. No count or source supports them in the text.
• Uncertainty and Bayesian methods are described as rarely used in this field §5.4, §6. No study list supports that claim in the paper.
• Figures are not published here. Figure 2 is modified from a third-party deposit model fig. 2. Figure 4 reproduces a classified map cited to Shirmard2020 fig. 4. Neither is used on this card, per the licensing rules; the text describes them §4, §5.1.

Lineage

• Builds on: none; this is a review.
• Cites Shirmard2020, a study on PCA, ICA and MNF for alteration mapping that is the source of Figure 4 §5.1, fig. 4. That study is not on this wiki.
• Compared against: not applicable.
• Superseded by: not stated.