remote sensing

Article
Drill-Core Mineral Abundance Estimation Using
Hyperspectral and High-Resolution
Mineralogical Data
Laura Tuşa *, Mahdi Khodadadzadeh , Cecilia Contreras , Kasra Rafiezadeh Shahi,
Margret Fuchs, Richard Gloaguen and Jens Gutzmer
Helmholtz Institute Freiberg for Resource Technology, Helmholtz-Zentrum Dresden-Rossendorf,
Chemnitzer Straße 40, 09599 Freiberg, Germany
* Correspondence: l.tusa@hzdr.de




Received: 16 March 2020; Accepted: 9 April 2020; Published: 9 April 2020


Abstract: Due to the extensive drilling performed every year in exploration campaigns for the
discovery and evaluation of ore deposits, drill-core mapping is becoming an essential step. While
valuable mineralogical information is extracted during core logging by on-site geologists, the process
is time consuming and dependent on the observer and individual background. Hyperspectral
short-wave infrared (SWIR) data is used in the mining industry as a tool to complement traditional
logging techniques and to provide a rapid and non-invasive analytical method for mineralogical
characterization. Additionally, Scanning Electron Microscopy-based image analyses using a Mineral
Liberation Analyser (SEM-MLA) provide exhaustive high-resolution mineralogical maps, but can
only be performed on small areas of the drill-cores. We propose to use machine learning algorithms to
combine the two data types and upscale the quantitative SEM-MLA mineralogical data to drill-core
scale. This way, quasi-quantitative maps over entire drill-core samples are obtained. Our upscaling
approach increases result transparency and reproducibility by employing physical-based data
acquisition (hyperspectral imaging) combined with mathematical models (machine learning). The
procedure is tested on 5 drill-core samples with varying training data using random forests, support
vector machines and neural network regression models. The obtained mineral abundance maps are
further used for the extraction of mineralogical parameters such as mineral association.

Keywords: hyperspectral imaging; drill-core; SWIR; mineral abundance mapping; mineral association;
machine learning

1. Introduction
Exploration campaigns are fundamental steps towards the discovery and evaluation of mineral
deposits required to fulfil the global demand of raw materials. Drilling is an essential part of
exploration surveys and consists of the extraction of long cylindrical core samples from underground
areas associated with relevant exploration potential. Traditionally, drill-cores are visually analyzed
by on-site geologists, who document characteristics such as mineralization type, lithology, structures
and alteration types [1]. Subsequently, core samples are used for laboratory-based geochemical and
mineralogical measurements to complement core logging results. While bulk geochemical analyses
are often available for entire boreholes, quantitative mineralogical information is usually restricted to
selected representative regions of interest. Standard quantitative analyses include X-Ray diffraction
(XRD) applied on powder samples [2] or Scanning Electron Microscopy (SEM) based image analyses
techniques [3] applied on polished thin sections prepared from areas of interest in the drill-cores.
Additionally, qualitative mineralogical analyses are performed through optical microscopy on thin

Remote Sens. 2020, 12, 1218; doi:10.3390/rs12071218 www.mdpi.com/journal/remotesensing

Remote Sens. 2020, 12, 1218 2 of 23

sections. These laboratory techniques provide valuable mineralogical information and derived
mineralogical and metallurgical parameters, but they are of small scale, highly time-consuming,
destructive, and rather expensive. This represents a challenge since thousands of meters of core are
acquired during exploration campaigns.
Hyperspectral imaging is currently being used in the mining and exploration industries as an
alternative tool to complement traditional logging techniques and to provide a rapid and non-invasive
analytical method to obtain mineralogical information [4–7]. Typical hyperspectral core imaging
systems can deliver data from a whole core tray (which holds approximately 5 m of core) in a matter of
seconds. Available sensors cover a wide range of the electromagnetic spectrum and record data in
several hundreds of contiguous spectral bands. Minerals have different spectral responses in specific
portions of the electromagnetic spectrum. These responses are influenced by the vibrational and
electronic absorption processes dependent on the bonds between atoms and electron orbitals [8].
Sensors covering the visible to near-infrared (VNIR) and short-wave infrared (SWIR) are commonly
used to identify and estimate the relative abundance of minerals such as phyllosilicates, amphiboles,
carbonates, iron oxides and hydroxides as well as sulphates [9].
Because of the increasing interest in hyperspectral data in the raw materials industry, with a wealth
of hyperspectral data becoming available, the development of methods to effectively analyze these data
is required. Traditional mapping methods include the use of spectral reference libraries (e.g., USGS
spectral library) for mineral identification and mapping on hyperspectral imagery [10,11]. Slightly
more automatic approaches, such as band ratios, or wavelength parameters such as position, depth
and width of the absorption features are also used to map the distribution and relative abundance of
specific minerals [12–14]. One of the most common procedures makes use of some of available tools
in a software called Environment for Visualizing Images (ENVI, Exelis Visual Information Solutions,
Boulder, Colorado). Such tools comprise endmember extraction, identification of the minerals using
the Spectral Analysis or Material Identification by comparison to a specific library in the software
(e.g., in ENVI) or online reference (e.g., USGS), and finally the mineral mapping task using similarity
measure algorithms or determination of partial abundances using unmixing algorithms [5,15–17].
Although these approaches may produce good results, they require continuous expert input
and thus, they tend to be time-consuming and difficult to automate for large dataset analysis. More
importantly, the performance of available unmixing algorithms highly relies on the determination of
the number of end-members and the selection of their representative spectra. In drill-core hyperspectral
data, highly mixed pixels of hardly pure mineral associations represent a challenge. Methods such
as unmixing, band ratios and minimum wavelength analysis can only provide mineral abundances
for spectrally diagnostic phases. Additionally, due to the nature of the hyperspectral data and the
spatial resolution allowed by commercially available sensors, the estimation of important mineralogical
parameters in the characterization of complex ores (e.g., mineral association), is currently challenging.
We propose a novel machine learning approach to estimate mineral quantities in drill-core
hyperspectral data. The procedure comprises four steps: 1) drill-core hyperspectral scanning (VNIR
–SWIR), 2) computing mineral abundances in a small but representative area of a drill-core by using
high-resolution mineralogical analyses (e.g., SEM-based image analyses using a Mineral Liberation
Analyser), 3) linking the mineral abundances in this small area to their corresponding spectra by
a multivariate regression model, and 4) estimating mineral abundances for the whole drill-core
hyperspectral data by using the learned model. The multivariate regression problem in the proposed
scheme is solved using three algorithms: random forest (RF), support vector machines (SVM) and
feedforward artificial neural networks (FF-ANN). The proposed procedure allows the abundance
estimation of the main mineral groups using their spectral characteristics (SWIR active) and using those
SWIR active minerals additionally as proxies for the SWIR non-active minerals or mineral groups such
as quartz, feldspar and sulphide. The obtained mineral abundance mapping results can be used for the
calculation of additional mineralogical parameters, relevant to exploration and mining projects. As an
Remote Sens. 2020, 12, 1218 3 of 23

example, the concept of mineral association at hyperspectral pixel scale based on relative abundances
is introduced in the current study.

2. Data Acquisition

2.1. Hyperspectral Data

The hyperspectral data used in this study were acquired from unpolished halves of diamond
drilling core samples with a SisuROCK drill-core scanner equipped with an AisaFENIX hyperspectral
sensor (Spectral Imaging Ltd., Oulu, Finland). The scanner is a fully automatic hyperspectral imaging
workstation which employs a tray table which carries the drill-core trays or samples under the
field-of-view of the spectrometer. The AisaFENIX camera implements two sensors to cover the VNIR
and SWIR regions of the electromagnetic spectrum. The sensor specifications and acquisition settings
are presented in Table 1.

Table 1. AisaFENIX sensor specification and setup for hyperspectral data acquisition.

VNIR 380–970 nm VNIR 15 ms

Wavelength Range Integration Time
SWIR 970–2500 nm SWIR 4 ms
VNIR 1.7 nm VNIR 2
Sampling Distance Spatial Binning
SWIR 5.7 nm SWIR 1
Number of Bands 450 Frame Rate 15 Hz
Samples 384 Scanning Speed 25.06 mm/s
Spatial Resolution 1.5 mm/pixel Field of View (FOV) 32.3◦
CMOS (VNIR) VNIR 4
Detector Spectral Binning
Stirling cooled MCT (SWIR) SWIR 1

The conversion from radiance to reflectance of the hyperspectral data was performed within
the acquisition software (LUMO Scanner version 2018-5, Spectral Imaging Ltd., Oulu, Finland) using
PTFE reference panels (>99% VNIR and >95% SWIR). To correct the sensor-specific optical distortions
(i.e., fish-eye and slit-bending effects on the images) and the spatial shift between the VNIR and
SWIR sensors, the toolbox MEPHySTo [18] was used. To avoid bands with little or no coherent
information, the data were spectrally resampled to 480—2500 nm by removing the first 30 bands. The
Savitzky–Golay filter was applied to decrease noise while preserving spectral features [19]. Principal
component analysis (PCA) [20] was performed on the hyperspectral dataset for data dimensionality
reduction and de-correlation while preserving 99.9% of the information.

2.2. Scanning Electron Microscopy-Based Mineral Liberation Analysis

Regions considered representative based on visual observations for the mineralogical variation
within the drill-core samples were cut and prepared into polished thin sections. The preparation
process consisted of grinding and polishing the sample surface followed by coating with a thin carbon
layer to avoid surface charging during data acquisition. The grinding and polishing led to the removal
of around 300 µm of material between the surface analyzed with the hyperspectral sensor and the
surface subjected to the high-resolution mineralogical analysis. Considering the sample morphology
and orientation of structural features the mineralogical variation is considered negligible for the
encountered shift.
The quantitative mineralogical data were acquired from the thin sections using an automated
approach. The analyses were carried out using Scanning Electron Microscope (SEM)-based Mineral
Liberation Analysis (MLA) [3,21]. For this, a FEI Quanta 650 F field emission SEM instrument (FEI,
Hillsboro, OR, USA), equipped with two Bruker Quantax X-Flash 5030 energy dispersive X-ray (EDX)
detectors (Bruker, Billerica, MA, USA) and the MLA Suite software package (version 3.1.4.686, FEI,
Hillsboro, OR, USA) were used. The grain-based X-ray mapping (GXMAP) mode was used to collect
the mineralogical information as follows: the MLA software collects the back-scattered electron images
Remote Sens. 2020, 12, 1218 4 of 23

(BSE) and uses them to effectively distinguish individual mineral grain boundaries based on the grey
scale variations. The grey scale values of the BSE images are proportional to the average atomic density
of the mineral grains and are used to provide a first mineralogical segmentation. The identification of
minerals is performed based on X-ray analysis by placing a closely-spaced grid on a particle in the BSE
image and collecting the X-ray data at the defined points of the grid. When dealing with fine grained
material of lower size than the placed grid, the GXMAP mode allows us to collect additional spectra
where variations in the BSE image are observed in between the measured grid points. Finally, the
mineral is determined by matching the resultant spectrum of energy peaks with a reference library
of X-ray spectra provided by the instrument company (FEI, Hillsboro, OR, USA), or from sample
extracted spectra analyzed based on peak locations and intensities [22]. Specifications of the operating
conditions used in this study are shown in Table 2.

Table 2. Operating conditions and parameters used for the acquisition of high-resolution SEM-MLA
mineralogical data.

SEM Settings MLA Settings

Acceleration voltage (kV) 25 Pixel size (µm) 3
Probe current (nA) 10 Step size (pixels) 6×6
Frame width (pixels) 1500 Acquisition time (ms) 5
Minimum grain size
BSE calibration (Au) 254 4
(pixels)

For classification, a mineral list was developed using the mineral reference editor in online mode.
The resulting mineral list contained a total of 59 entries. However, for the integration of the HSI with
SEM-MLA, further grouping was performed in this paper, such as considering all feldspars in one
class, all white micas in another or, all sulphides, sulphosalts and gold in another. Accessory minerals
were included in the final grouping labelled as “others”. As a result, ten main mineral groups are
considered: white mica (WM), biotite (Bt), chlorite (Chl), amphibole (Amp), carbonate (Cb), gypsum
(Gp), feldspar (Fsp), quartz (Qz), sulphide (Sp) and other.

3. Data Description
For testing the proposed methodology, 5 samples, labelled DC-1 to DC-5, from different locations
within the Bolcana porphyry copper-gold system [23–26] were analyzed. Hyperspectral images were
acquired on the halves cores after which thin sections were prepared from selected regions of interest
and analyzed by SEM-MLA. Each region is further labelled as a, b and/or c starting from the left-hand
side of the drill-core sample as illustrated in Figure 1. The ore minerals in the studied system are
chalcopyrite, bornite, covellite, chalcocite and gold. Gold is dominantly present as fine inclusions in
pyrite and chalcopyrite. The main encountered alteration types are potassic, sodic—calcic, phyllic
and argillic. In the studied samples the first three are present, some samples presenting a transitional
character and are described in this section. Please see Sillitoe, 2010 [27] for details on the mineralogical
characteristics of the alteration styles typically associated with porphyry Cu-Au systems.
While the summary of the results for each sample is presented in the results section, an emphasis
is made on DC-1 in order to illustrate all the potential information that can be extracted using the
proposed methodology. Therefore, a more detailed description of this sample is available in the
current section. Sample DC-1 consists of a diorite porphyry. Hydrothermal alteration in this sample
appears transitional between potassic, represented by the presence of biotite and potassic feldspar
and sodic-calcic characterized by the plagioclase-chlorite assemblage. Chlorite is more abundant than
biotite in the first two thin sections, “a” and “b”. The third thin section, though, due to the lower vein
density and implicit associated alteration presents significant amounts of biotite disseminated as well
as in clusters in the matrix. Plagioclase feldspar is dominant in all three thin sections, near the veins
however, an increase in potassic feldspar is observed.
Remote Sens. 2020, 12, 1218 5 of 23

Figure 1. RGB photograph of the analyzed drill-cores (labelled on the left-hand side from DC-1 to
DC-5) with overlain high-resolution mineral maps (labelled a, b and c) obtained by SEM-MLA.

Thin section “a” of sample DC-1 captures three main vein types: an oblique early quartz vein
which exhibits a low intensity white mica alteration halo likely associated with a younger cross-cutting
gypsum vein that has a sulphide centerline and a wide white mica-chlorite alteration halo (top). The
alteration halo here is mica-dominant in the proximity of the vein and chlorite-dominant towards its
edges. The third vein present in section “a” consists of quartz with a gypsum centerline and a spotty,
low intensity white mica alteration halo (bottom). Thin section “b” captures three main vein types as
well: two sub-vertical veins consist of variable ratios of gypsum and quartz and are surrounded by a
strong white mica low-chlorite alteration halo. Compositionally, these veins appear to be a mixture
between the first and third veins mentioned for thin section “a”; they have, however, a different
morphology. In proximity to sub-horizontal veinlets in the lower half of the thin section, an increase in
the pyrite and chlorite content is observed. The two sub-horizontal veinlets show strong similarity with
the horizontal veins in the first thin section. The alteration intensity surrounding the sub-horizontal
veinlets appears to be related to complex interactions with pre-existing veinlets in this area of section
“b”. Thin section “c” hosts several fine veinlets, of highest width, the two cross-cutting ones near the
top of the thin section. The veinlets consist of variable amounts of quartz, gypsum, pyrite and white
mica and present a white mica and chlorite alteration halo. Similar to the subvertical veins in thin
section “b” these veins appear to have a composition intermediate between the horizontal veins in
thin section “a”. Unlike the two veins in thin section “b” however, the extent of the alteration halo is
much lower.
Remote Sens. 2020, 12, 1218 6 of 23

Sample DC-2 is marked by pervasive potassic alteration characterized by the presence of K-feldspar,
biotite and minor chlorite. Two main vein types are present in this sample: veins hosting dominantly
sulphide which show a strong phyllic alteration halo caused by the late reaction of mineralizing
hydrothermal fluids with the host rock. The second vein type comprises dominantly quartz with
sulphide or with sulphide-calcium sulphate (gypsum or anhydrite) centerline. Additional veins of
varying composition are present in the sample (left-hand side as illustrated in Figure 1). They appear to
be the result of complex reopening and cross-cutting of the previously described veins. A sodic-phyllic
rock matrix hosting two main vein-types characterizes sample DC-3. The first vein comprises of
sulphide and presents a large white mica alteration halo. The second vein type consists predominantly
of quartz, calcium sulphate and sulphide. The changing symmetry and mineral association in these
latter veins indicate the reopening of an initially present quartz vein. Sample DC-4 is characterized by
the presence of intense phyllic alteration in the matrix related to the thick pyrite-quartz-gypsum vein
cross-cutting the sample. Additional fine veinlets comprising mostly quartz and pyrite are cutting the
mica-rich matrix. The matrix in sample DC-5 consists of dominantly feldspar and subordinately white
mica. Three main vein types can be observed in the samples: a sulphide dominant vein with a broad
white mica alteration halo, quartz veinlets and carbonate iron-oxide veins which show low or absent
alteration halos.
For the understanding of the modal composition of the available thin sections, the abundances
of the minerals or mineral groups for all the analyzed thin sections are illustrated in the bar charts
in Figure 2 (left). For most samples, quartz and feldspar represent the main rock-forming minerals.
There is, however, a variation in the extent of alteration of feldspar to white mica ranging from low
(DC-2a) to high (DC-4). In most of the analyzed samples, the amphibole is to a large extent altered to
chlorite and/or biotite. Biotite is only present in significant amounts in sample DC-2 and DC-1 “c”.
The variation of the quartz, carbonate and gypsum contents is related to the surface abundance of
the veins and veinlets filled mostly by these three minerals. While quartz and gypsum are present in
significant amounts in all thin sections, carbonate is mainly represented in sample DC-5. The class
“sulphide” comprises mainly pyrite, chalcopyrite, bornite, chalcocite and covellite but minor amounts
of native gold hosted as inclusions in pyrite and chalcopyrite is also considered. While pyrite is not
an ore mineral by itself, it frequently represents the host of micron-size native gold inclusions. The
sulphide content in the thin sections ranges from around 1 area % in DC-5b to almost 30 area % in
DC-4b. The main target being the quantification and understanding of the distribution of sulphide
minerals within the presented samples, their mineral association is also analyzed and presented in the
bar chart in Figure 2 (right).

Figure 2. Modal mineralogy and mineral association of analyzed thin sections illustrated through
mineral maps in Figure 1. The labels of each sample and thin section are illustrated between the two
bar charts.
Remote Sens. 2020, 12, 1218 7 of 23

While an influence of the modal mineralogy can be observed on the mineral association, a strong
increase in the white mica, chlorite, biotite, carbonate and gypsum can be seen. This is the result of
the distribution of these minerals within or surrounding the veins also hosting the bulk of sulphides.
The listed gangue minerals, unlike the sulphide, show distinct absorption features in the VNIR-SWIR
region of the electromagnetic spectrum and may, therefore, be used as proxies for the distribution of
the ore minerals.

4. Methodological Framework

4.1. HSI—SEM-MLA Data Integration

For the proposed approach, the SEM-MLA data is upscaled by adopting a re-sampling procedure.
The two-dimensional SEM-MLA mineral map with high spatial resolution is transformed to a
three-dimensional mineral abundance map with the lower spatial resolution of the hyperspectral
data [7]. The third dimension consists of the relative abundance of each mineral present in each
SEM-MLA map re-sampled to the hyperspectral pixel size (Figure 3). Note that a co-registration stage
is needed after the re-sampling of the SEM-MLA data. Following Acosta et al., 2019, the structural
features, such as veins, the mineral composition, and spectral responses are used to find suitable
tie points. As a result of the co-registration each pixel where the SEM-MLA data is available is
characterised by two vectors: the hyperspectral feature vector Xi of dimension d (i.e., the number of
bands in the hyperspectral data) or r (number of extracted features) and a mineral abundance vector Yi
containing the corresponding fractional abundances of the minerals identified by SEM-MLA.

Figure 3. Graphical illustration of the co-registration and resampling process for the SEM-MLA to
hyperspectral data. In red, the size of a hyperspectral pixel characterized by a mineral mixture in the
SEM-MLA data and a spectrum in the hyperspectral data. The color of the spectra (left) is given by the
mixture ratio of the minerals illustrated in the SEM-MLA simplified example (right).

Once the hyperspectral and SEM-MLA data are co-registered, they are divided into training and
testing. For this procedure the following approach is adopted:

• Using 50% randomly selected pixels from all thin section regions within one drill-core sample
for training, the remaining drill-core hyperspectral data for testing. The validation is performed
using the remaining 50% data points from the MLA regions.
• Using 1 thin section for training and the second for testing and validation for all drill-core samples.
• For DC-1, where 3 thin sections are available, an additional test is performed using 2 thin sections
for training and the last for testing and validation.

As can be seen from the main flowchart, shown in Figure 4, the proposed workflow is carried out
in three main phases. In the training phase, different regression models (i.e., RF, SVM and FF-ANN) are
Remote Sens. 2020, 12, 1218 8 of 23

trained following any of the three approaches mentioned before. In the prediction phase, the learned
models are used to predict the mineral abundances on the entire drill-core samples. Finally, in the
validation phase, the root mean square error (RMSE) [28] is calculated on the remaining SEM-MLA
test data to assess the performance of the abundance mapping.

Figure 4. Flowchart illustrating the three main stages of the proposed workflow.

Two analysis types are further performed on the resulting mineral abundance data. For each
validation set, the modal mineralogy is calculated based on the average abundance of each mineral
phase in each pixel and compared to the modal mineralogy data obtained from SEM-MLA. Additionally,
the concept of mineral association is adapted from the automated mineralogy field (Figure 2). There,
the mineral association is calculated by counting the neighboring pixels to a specific target mineral.
Slight changes in the approach have to be made when the spatial resolution of the hyperspectral data
is used. The association of the main target group, i.e., sulphide, is a fundamental aspect in the present
geological study. For each hyperspectral pixel the estimated mineral abundance of each mineral phase,
except of the target, is normalized by the abundance of sulphide in the respective pixel. While this
approach does not directly indicate the grain contact between the two minerals (or rather mineral
groups) it can be seen as the probability of their association and occurrence at the scale of hyperspectral
data resolution. The mineral association is calculated on the ground truth or validation data as well as
on the estimated abundances calculated with the three proposed regression models.

4.2. Random Forest Regression

Random forests (RFs) are currently one of the most popular supervised learning techniques for
classification and regression problems [29–31]. RFs are ensemble-based algorithms in which several
models (trees) are running in parallel with randomized sampling. The individual results of these trees
are then combined into the final prediction by an averaging process [32]. For regression purposes, the
trees are given numerical values as predictors whereas in classification problems they are fed class
labels. The RF technique is desirable in cases where only few training samples are available, as is
usually the case in drill-core hyperspectral imaging.

4.3. Support Vector Regression

The aim of support vector machines (SVMs) is to search for hyperplane decision boundaries to
define a linear prediction model [33,34]. To locate and orientate the hyperplane, only the samples that
are close to the hyperplane, so-called support vectors, have an influence. Therefore, SVMs perform
well when a limited number of well-chosen training samples are available [31,33,34]. This model
can be used for classification or regression tasks. SVMs were originally proposed to solve linear
problems. However, decision boundaries are often non-linear. To cope with the non-linearity problem,
Remote Sens. 2020, 12, 1218 9 of 23

the kernel-based SVMs were introduced to project the data points into a higher dimensional feature
space where the samples are linearly separable [31].

4.4. Artificial Neural Network Regression

Artificial neural networks have become some of the most popular methods in regression and
classification because of their success in capturing the non-linearity relation between independent and
dependent variables [35]. We chose a so-called “feedforward neural network” (FF-ANN) [36], as it
fits the requirements of the problem at hand. In a feedforward network, each neuron in one layer
is directly connected to neurons of the next layer with no cycle between layers. The applied neural
network consists of an input layer, one hidden layer, and an output layer. Each neuron of a layer is
computed by the product sum of the neurons of the previous layers plus a bias for the neuron [31].
A sigmoid function is applied for activation.

5. Experimental Results
In order to showcase the suitability of the proposed approach, the first drill-core sample presented
in the data section (DC-1) is used. The remaining four samples have been analyzed following the same
procedure. A summary of the results is presented in this section followed by a complete illustration
of the results in Appendix A. Additionally, all numerical results are presented in the Electronic
Supplementary Materials (Table S1).
From the entire drill-core sample (DC-1), the VNIR-SWIR hyperspectral data of size 33 by 189
pixels. The 420 spectral bands cover wavelengths from 480 nm to 2500 nm. The hyperspectral data is
subjected to PCA leading to the reduction in dimensionality to 13 principal components in the third
dimension. Moreover, the high-resolution mineralogical data obtained from representative regions
(thin sections “a”, “b” and “c”) were used. In the thin section regions of the drill-core sample, each
hyperspectral pixel covers an area of 1.5 by 1.5 mm2 , which is characterized by about 250,000 pixels
in the SEM-MLA image. The fractional abundances were computed by considering the frequency of
the identified minerals in the corresponding region of the SEM-MLA image for each hyperspectral
pixel. To have more consistent results, we considered a threshold of 250,000 pixels (i.e., a hyperspectral
pixel size) in each thin section region, for discarding minerals which have a very low frequency in
the original SEM-MLA image. Taking this factor into consideration, the following six mineral classes
remained: white mica (WM), biotite (Bt), chlorite (Chl), amphibole (Amp), gypsum (Gp), feldspar
(Fsp), quartz (Qz), sulphide including sulphosalts and native gold (SP); less abundant minerals were
grouped as “other”. Because of the low abundance of biotite and accessory minerals in thin sections
“a” and “b”, the number of mineral classes considered was decreased accordingly. The test setups
presented in the methodological framework section are used.
Cross-validation has been used to find the optimal parameters in order to train three models by
internally resampling the training data. The main tested parameter ranges for each algorithm are
presented in Table 3. The setups were chosen according to the lowest associated root-mean-square
error (RMSE) based on cross-validation within 30 averaged iterations.

Table 3. Parameters and parameter ranges for the choice in optimum setup of the three tested algorithms.

RF SVM FF-ANN
Training function – Scaled conjugate
Kernel – Radial Basis Function
gradient backpropagation
Nb. of trees – 500 : 600
Cost – 2 : 0.5 : 4 Nb. of hidden layers – 1
Sigma – 5 : 0.5 : 7 Nb. of neurons – 30 : 10 : 80

5.1. Mineral Abundance and Association Mapping

With the first experimental setup, presented in the methodological framework, 50% randomly
distributed samples of the available thin section regions were used to train the regression models the
Remote Sens. 2020, 12, 1218 10 of 23

mineral abundances estimation in the entire drill-core sample (Figure 5). Based on the visual analysis
of the core and results analysis, RF and FF-ANN show better results in estimating the abundance of
minerals with local distribution and small concentrations. With respect to matrix mineralogy, while
biotite is well estimated by SVM in comparison with RF and FF-ANN, other major components of the
matrix such as feldspar present a rather poor estimation. Similar performances of the algorithms can
be observed for vein mineral components such as gypsum and sulphide.

Figure 5. Drill-core mineral abundance maps of white mica (WM), biotite (Bt), chlorite (Chl), amphibole
(Amp), gypsum (Gp), feldspar (Fsp), quartz (Qz), sulphide (SP) and accessory minerals (Other) using
randomly distributed 50% of the available ground truth data for training for random forest (RF),
support vector machine (SVM) and feed-forward neural network (FF-ANN) regressions.

With regards to the samples DC-2 (Figure A1), DC-3 (Figure A3), DC-4 (Figure A5) and DC-5
(Figure A7), using 50% of the available ground truth data for training, RF and FF-ANN show good,
similar performances, while SVM shows limitations specifically in transitional areas between veins
and matrix. Among the SWIR-diagnostic minerals, white mica, biotite and carbonate appear well
mapped in all the samples, chlorite is slightly underestimated in samples DC-2 and DC-3 and gypsum
is overestimated in sample DC-5. Among the SWIR non-diagnostic minerals, quartz shows the highest
mapping inconsistencies between vein and matrix, particularly for samples DC-4 and DC-5. Sulphide,
however, appears to be well mapped in most areas of the samples.
The quantitative evaluation of the mineral abundance mapping through the calculation of the
RMSE supports the visual observations (Table 4). All three tested algorithms present low RMSEs and
prove suitable to be used for mineral abundance mapping purposes. RF shows the lowest overall
RMSE of 0.07, followed by FF-ANN with 0.08 and SVM with 0.1. Regarding the per class RMSE, RF
and FF-ANN show similar results with the largest error associated with quartz, which can be the
result of the lack of diagnostic absorption features in the VNIR-SWIR regions of the electromagnetic
spectrum. SVM on the other hand shows larger per class errors for feldspar together with an increase
in the error on white mica distribution. This can be explained by a misclassification between the two
mineral groups. The mineral association of the sulphide in each pixel was calculated from the results
of the mineral abundance mapping. Based on this calculation an equivalent overall performance of the
methods was obtained (Table 5). For each of the methods, the error for the association of sulphide with
feldspar is the largest.
Remote Sens. 2020, 12, 1218 11 of 23

Table 4. Evaluation of the three tested methods for the mineral abundance mapping of DC-1 through
overall RMSE and per class RMSE values.

RMSE per Class

Method RMSE
WM Bt Chl Amp Gp Fsp Qz SP Other
RF 0.07 0.06 0.06 0.06 0.06 0.06 0.05 0.08 0.06 0.05
SVM 0.10 0.12 0.03 0.05 0.01 0.12 0.21 0.12 0.04 0.02
NN 0.08 0.06 0.06 0.07 0.06 0.06 0.07 0.09 0.07 0.07

Table 5. Evaluation of the three tested methods for the mineral association mapping of DC-1 through
overall RMSE and per class RMSE values.

RMSE per Class

Method RMSE
WM Bt Chl Amp Gp Fsp Qz Other
RF 0.05 0.05 0.00 0.01 0.00 0.02 0.13 0.05 0.00
SVM 0.06 0.06 0.01 0.02 0.00 0.02 0.15 0.06 0.00
NN 0.05 0.05 0.00 0.01 0.00 0.02 0.13 0.05 0.00

To assess the importance of sampling and representativeness of the SEM-MLA regions, thin
sections “a”, “b” (Figure 6) and “a + b” (Figure 7) of sample DC-1 were used for training the models in
order to estimate the mineral abundance and association in thin section “c”.

Figure 6. White mica (WM), chlorite (Chl), amphibole (Amp), gypsum (Gp), feldspar (Fsp), quartz
(Qz) and sulphide (SP) abundance maps of TS-1c using TS-1a and TS-1b, respectively, for the training
of random forest (RF), support vector machine (SVM) and feed-forward neural network (FF-ANN)
regressions. The ground truth (GT) resized MLA data is presented for comparison.

For the three used methods, strong differences in the estimates of sample “c” mineralogy can
be observed when using thin sections “a” and “b” for training (Table 6). The use of thin section “a”
provides particularly better results for white mica and feldspar, which are confused using region “b”
that hosts distinctly lower amounts of feldspar. On the other hand, using thin section “a” for training
leads to an overestimation of the gypsum content. The use of both thin sections (“a” + “b”) for training
improves the classification leading to lower overall and per class RMSE values. As for the remaining
drill-core samples, RF outperforms SVM and FF-ANN for most training scenarios, except when using
thin section “b” for training. A similar effect of sampling on the RMSE evaluation can be seen for the
mineral association mapping of DC-1 in all the scenarios (Table 7).
Remote Sens. 2020, 12, 1218 12 of 23

Figure 7. White mica (WM), biotite (Bt), chlorite (Chl), amphibole (Amp), gypsum (Gp), feldspar (Fsp),
quartz (Qz), sulphide (SP) and accessory minerals (Other) abundance maps of TS-c using TS-a + TS-b
for the training of random forest (RF), support vector machine (SVM) and feed-forward neural network
(FF-ANN) regressions. The ground truth (GT) MLA data is presented for comparison.

Table 6. Evaluation of the three tested methods for the mineral abundance mapping of DC-1 thin
section “c” through overall RMSE and per class RMSE values using different samples for training.

Train and Valid. Overall RMSE per Class

Data RMSE WM Bt Chl Amp Gp Fsp Qz SP Other
50%–50% rand. sel 0.07 0.06 0.06 0.06 0.06 0.06 0.05 0.08 0.06 0.05
Train a—Test—c 0.10 0.08 0.06 0.01 0.06 0.18 0.13 0.03
RF

Train b—Test—c 0.12 0.15 0.05 0.01 0.05 0.24 0.09 0.03
Train a + b—Test c 0.08 0.09 0.04 0.05 0.01 0.04 0.18 0.11 0.03 0.01
50%–50% rand. sel 0.10 0.12 0.03 0.05 0.01 0.12 0.21 0.12 0.04 0.02
SVM

Train a—Test—c 0.10 0.10 0.06 0.02 0.06 0.20 0.12 0.04
Train b—Test—c 0.09 0.11 0.06 0.02 0.04 0.18 0.10 0.04
Train a + b—Test c 0.09 0.09 0.05 0.06 0.03 0.07 0.21 0.09 0.05 0.03
50”–50% rand. sel 0.08 0.06 0.06 0.07 0.06 0.06 0.07 0.09 0.07 0.07
FF-ANN

Train a—Test—c 0.17 0.12 0.07 0.02 0.14 0.28 0.27 0.06
Train b—Test—c 0.12 0.12 0.07 0.02 0.06 0.25 0.14 0.04
Train a + b—Test c 0.10 0.10 0.05 0.06 0.01 0.05 0.20 0.16 0.04 0.01

For the remaining samples, each having two regions analyzed by SEM-MLA, the mineral
abundance estimations obtained using the second setup are illustrated in Figure A2 (DC-2), Figure A4
(DC-3), Figure A6 (DC-4) and Figure A8 (DC-5).
The tested methods show similar results for mineral abundance and association mapping on the
remaining four drill-cores (Table 8). Overall, RF performs best, followed by FF-ANN and then SVM.
For samples DC-1, DC-2, DC-3 and DC-5 each method results in comparable errors where similar
amounts of training data are used. For sample DC-4 the overall RMSE values are higher, exceeding
0.2 depending on training data. For each sample the selection of the training data location plays an
important role that is reflected into the RMSE evaluation.
Remote Sens. 2020, 12, 1218 13 of 23

Table 7. Evaluation of the three tested methods for the mineral association mapping of DC-1 thin
section “c” through overall RMSE and per class RMSE values using different samples for training.

Train and Overall RMSE per Class

Validation Data RMSE WM Bt Chl Amp Gp Fsp Qz Other
50%–50% rand. sel 0.05 0.05 0.00 0.01 0.00 0.02 0.13 0.05 0.00
Train a—Test—c 0.06 0.05 0.04 0.01 0.05 0.05 0.10
RF

Train b—Test—c 0.03 0.04 0.03 0.00 0.01 0.01 0.02

Train a + b—Test c 0.02 0.01 0.01 0.04 0.01 0.00 0.01 0.06 0.00
50%–50% rand. sel 0.06 0.06 0.01 0.02 0.00 0.02 0.15 0.06 0.00
SVM

Train a—Test—c 0.05 0.05 0.02 0.02 0.07 0.06 0.06

Train b—Test—c 0.04 0.05 0.05 0.01 0.01 0.02 0.05
Train a + b—Test c 0.03 0.05 0.03 0.01 0.03 0.06 0.03 0.00 0.03
50%–50% rand. sel 0.05 0.05 0.00 0.01 0.00 0.02 0.13 0.05 0.00
FF-ANN

Train a—Test—c 0.17 0.13 0.05 0.01 0.15 0.24 0.28

Train b—Test—c 0.07 0.07 0.04 0.00 0.04 0.15 0.00
Train a + b—Test c 0.05 0.09 0.00 0.05 0.01 0.02 0.06 0.10 0.00

Table 8. Methods evaluation for the mineral abundance and association mapping of the remaining
four samples through overall RMSE and per class RMSE values using different data for training.

Sample Train and Mineral Abundance Mapping Mineral Association Mapping

ID Validation Data RF SVM FF-ANN RF SVM FF-ANN
50%–50% rand. sel 0.07 0.09 0.08 0.07 0.07 0.07
DC-2 Train a—Test—b 0.11 0.18 0.10 0.09 0.17 0.09
Train b—Test—a 0.14 0.14 0.19 0.13 0.16 0.13
50%–50% rand. sel 0.08 0.11 0.09 0.12 0.12 0.12
DC-3 Train a—Test—b 0.14 0.14 0.17 0.09 0.18 0.07
Train b—Test—a 0.11 0.14 0.14 0.10 0.11 0.09
50%–50% rand. sel 0.12 0.20 0.14 0.12 0.12 0.12
DC-4 Train a—Test—b 0.24 0.29 0.24 0.08 0.10 0.05
Train b—Test—a 0.16 0.20 0.19 0.04 0.16 0.07
50%–50% rand. sel 0.07 0.10 0.08 0.03 0.03 0.03
DC-5 Train a—Test—b 0.11 0.13 0.11 0.05 0.18 0.05
Train b—Test—a 0.13 0.13 0.15

5.2. Modal Mineralogy

The modal mineralogy in area % is calculated by averaging the mineral abundances over the entire
tested sample. To evaluate the modal mineralogy estimates sample DC-1 is used and the estimates are
compared to the ground truth, using 50% of the available SEM-MLA data for training and 50% for
testing (Table 9).

Table 9. Ground truth and estimated modal mineralogy of the SEM-MLA test regions of DC-1, using
50% randomly selected data for training.

Modal Mineralogy (Area %)

Method
WM Bt Chl Amp Gp Fsp Qz SP Other
GT 16.0% 1.1% 4.3% 1.8% 7.3% 42.3% 24.9% 1.8% 0.5%
RF 15.8% 1.1% 4.3% 1.7% 7.4% 42.2% 25.1% 2.0% 0.5%
SVM 14.2% 1.2% 3.8% 1.8% 7.8% 44.3% 24.4% 1.8% 0.6%
NN 15.8% 1.1% 4.3% 1.7% 7.3% 42.6% 24.8% 1.9% 0.5%
Remote Sens. 2020, 12, 1218 14 of 23

The estimates for all methods show good results with the highest RMSE value of 0.01 obtained
with SVM. The complete modal mineralogy results are available in Table S1. The results for all the
setups and all samples and methods are illustrated in Figure 8 by plotting the estimated values from RF
(left), SVM (centre) and FF-ANN (right) against the ground truth values known from the re-sampled
SEM-MLA data. The estimated and true values for RF and FF-ANN show overall a good correlation
with local outliers related to mineral groups such as feldspar, as these do not have distinct spectral
features in the VNIR-SWIR regions of the electromagnetic spectrum. Outliers can also be observed
for white mica where the training and testing classes were unbalanced and confusions between mica
and feldspar occurred. SVM, on the other hand, shows higher deviations from a linear correlation.
Additionally, an important factor influencing the results is the data used for sampling. All test scenarios
results are included in Figure 8 and as observed in the mineral abundance mapping results (Table 8),
sampling plays a critical role in method performance.

Figure 8. Scatter-plots of the ground truth vs. estimated mineral area % in all analyzed scenarios and
samples using A. RF, B. SVM and C. FF-ANN.

5.3. Mineral Association

The overall mineral association is calculated by averaging the sulphide association in each
classified pixel. The results for the setup consisting of 50% of the SEM-MLA regions of DC-1 for training
and 50% for testing are presented in Table 10. For each regression method, the association of sulphide
with white mica, chlorite, gypsum and quartz is underestimated, while the feldspar association is
overestimated. The same tendency is observed for the rest of the calculated mineral associations in
all samples and setups (Appendix A, Figures A1–A8). The relationship between ground truth and
estimated data is illustrated in the scatter-plots in Figure 9. The results of the mineral association are
strongly influenced by the estimation of the sulphide abundance as well as of the other mineral groups.
Therefore, the highest errors in sulphide abundance mapping are consistent with the largest errors for
sulphide association.

Table 10. Ground truth and estimated mineral association of the SEM-MLA test regions of DC-1, using
50% randomly selected data for training.

Sulphide Association
Method
WM Bt Chl Amp Gp Fsp Qz Other
GT 21.0% 0.7% 5.6% 1.5% 9.9% 30.1% 30.5% 0.6%
RF 16.2% 1.1% 4.4% 1.8% 7.6% 42.8% 25.7% 0.5%
SVM 14.5% 1.2% 3.9% 1.9% 7.8% 45.1% 24.8% 0.6%
NN 16.2% 1.1% 4.4% 1.8% 7.5% 43.2% 25.3% 0.5%
Remote Sens. 2020, 12, 1218 15 of 23

Figure 9. Scatter-plots of the ground truth vs. estimated mineral association in all analyzed scenarios
and samples using A. RF, B. SVM and C. FF-ANN.

6. Discussion
The proposed approach for data preparation and analysis illustrates the potential to arrive
at robust quantitative mineral abundance estimates from hyperspectral drill-core data—even for
those minerals that do not have diagnostic absorption features in the VNIR-SWIR regions of the
electromagnetic spectrum (e.g., feldspars, quartz, sulphides). Three regression methods were tested in
this paper for mineral abundance estimation: random forest (RF), support vector machines (SVM) and
feedforward artificial neural networks (FF-ANN). These methods were applied to quantify mineral
abundances—also of minerals devoid of characteristic HS spectral features (here sulphide minerals).
In addition, attempts were made to extract mineral association data from HS information at a lateral
resolution far below the actual size of mineral grains in the studied ore. For this purpose, the abundance
of each gangue mineral in each HS pixel is normalized to the content of ore minerals that are the
main target in the currently studied porphyry system, thus constituting a rather simple proxy for the
opportunity of two minerals or mineral groups to occur in direct contact with each other.
The abundance estimation of SWIR diagnostic mineral phases and groups is good overall,
particularly for white mica, amphibole and chlorite. For the case of gypsum, however, due to its
pervasive association with white mica in some training samples, errors in the abundance estimation
occurred. Even though it is present in minor amounts in comparison to white mica, the estimation error
can reach similar amplitudes as those of white mica. An additional reason for high errors associated
with gypsum is related to its composition. The higher the degree of hydration of anhydrite towards
gypsum the stronger and more distinct its absorption features. While SEM-MLA methods cannot
measure the amount of water in the structure of the hydrated calcium sulphate, hyperspectral sensors
are highly sensitive to these changes. Therefore, having training samples hosting mostly calcium
sulphate with low amount of water can cause miss-estimation in test samples which may have low
amounts of highly hydrated calcium sulphate. The local high errors in the estimation of biotite content
can be assigned to the low amount of training samples containing relevant amounts of biotite. Sulphide
is the main target in the current case study and this group comprises dominantly of pyrite, chalcopyrite,
bornite, covellite, chalcocite, minor sulphosalts and native gold as an inclusion in the sulphides. While
locally sulphide can be present as disseminations in the matrix, the highest fraction is present in veins.
For all methods, the abundance estimation for SWIR non-diagnostic minerals is highly dependent on
their association with the hydrothermal alteration minerals. To be able to estimate their abundance,
representative sampling is required to avoid the erroneous estimation of these minerals based on local
association with SWIR minerals that are not consistent at drill-core scale. For the analyzed samples
the highest per-class errors are obtained for feldspar and quartz, both SWIR non-diagnostic minerals.
In many cases feldspar was overestimated, particularly in samples where white mica abundance was
underestimated. As white mica is present as an alteration product of feldspar in the proximity of veins,
it can be assumed that the training samples consisted of lower alteration degrees of the feldspar to
white mica while the test samples showed contrasting composition. As a result, feldspar particularly
represented a bottleneck for the evaluation of the mineral association where their association with
Remote Sens. 2020, 12, 1218 16 of 23

sulphide was in each case overestimated. Besides the fact that this mineral group does not show
distinctive absorption features in the VNIR-SWIR regions of the electromagnetic spectrum, the spatial
resolution of the used sensor can highly influence the misclassification and the overestimation in
its association with sulphide. Feldspar is usually present in the host-rock matrix and is expected to
have a low association with sulphide, usually being altered to white mica in the proximity of the
sulphide-bearing veins. When the vein alteration halo is thinner than the spatial resolution of the
sensor (here 1.5 mm), an increase in the apparent association of sulphide with feldspar is observed.
A potential limitation resides in the removal of the mineral fractions present in low concentrations
(lower total surface abundance than the size of a hyperspectral pixel). Additionally, the compositional
variation of minerals such as white mica and chlorites is not analyzed in the current work, but could
be performed by auxiliary methods such and minimum wavelength analysis.
To evaluate the performance of the three regression methods employed in this paper, the RMSE
was calculated. In general, for the mineral abundance estimation RF performed well and derived the
lowest errors. The errors produced by FF-ANN tend to be higher than by SVMs in all the test scenarios,
except in the case when 50% of the ground truth was randomly selected as the training data. This
highlights the capabilities of SVM to perform well when a limited number of training samples are
available and of FF-ANN to achieve good results when enough training data are available. The random
selection of the training data allows for a more representative sampling per class than it is for the other
two test scenarios where one thin section is used for training and the other thin section is used for the
test. This is because certain minerals can be more abundant in one part of the core than in the other as
it was previously stated for DC-1 in the results section. Although larger per class RMSE are obtained
by minerals without diagnostic absorption features in the VNIR-SWIR, this is countered by random
sampling and errors decrease considerably. From the analysis and evaluation of the results obtained
by the utilized regression methods, the RF algorithm is the most suitable for the current dataset.
The proposed framework allows for fast evaluation of the modal mineralogy of analyzed samples
and it shows potential for further upscaling. It proves that hyperspectral drill-core scanning provides
a fast, non-invasive mineral identification and quantification if suitable training samples are available.
Domaining of the hyperspectral data before the selection of representative samples for detailed
analysis can minimize and focus the effort and amount of invasive measures related to sampling and
high-resolution mineralogical analyses. The automated character of the approach can be later used
on mine sites provided that hyperspectral drill-core scanning is available to support the geologists
in the core-logging procedure, as well as training samples characterized by high resolution methods
of mapping mineral distributions, such as SEM-based image analyses. The derived mineralogical
parameters such as modal mineralogy and mineral association can additionally prove useful past
exploration stages as they are essential in defining geometallurgical domains [37].

7. Conclusion and Remarks

Hyperspectral drill-core imaging provides fast, extensive and non-destructive mapping of certain
minerals with spectral characteristic features in the VNIR-SWIR regions of the electromagnetic spectrum.
SEM-MLA analyses allow a precise and exhaustive mineral mapping of selected small samples. We
propose to combine both analytical techniques using machine learning in order to provide mineral
abundance and association mapping over entire drill-cores. The proposed methodological framework
is illustrated on samples collected from a porphyry type deposit, but the procedure is easily adaptable
to other ore types. All tested ML algorithms deliver good results but RF is more robust to unbalanced
and sparse training sets and is recommended for further work. As a result, quasi-quantitative maps
are also produced and evaluated. The mineral abundance results can be further used to calculate
parameters such as modal mineralogy, mineral association and other mineralogical indices. Therefore,
this approach can be integrated in the standard core-logging procedure, complementing the on-site
geologists, and can serve as background for the geometallurgical analysis of numerous ore types.
Remote Sens. 2020, 12, 1218 17 of 23

Supplementary Materials: The following are available online at http://www.mdpi.com/2072-4292/12/7/1218/s1,

Table S1: Compilation of numerical results for mineral abundance and mineral association estimation.
Author Contributions: Conceptualization, L.T.; methodology, L.T. and M.K.; software, L.T., M.K. and C.C.;
validation, L.T.; formal analysis, L.T.; investigation, L.T.; writing—original draft preparation, L.T., M.K., C.C. and
K.R.S.; writing—all authors; visualization, L.T.; supervision, M.K., M.F., R.G. and J.G.; project administration, R.G.
and J.G.; All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Acknowledgments: We would like to acknowledge Deva Gold S.A., subsidiary of Eldorado Gold for sample
availability and on-site support. Robert Zimmerman (HZDR) is gratefully thanked for the support during data
acquisition and Sandra Lorenz (HZDR) for the MEPHySTo toolbox for the hyperspectral data pre-processing.
Sabine Gilbricht (TU Bergakademie Freiberg) is acknowledged for support during SEM-MLA data acquisition.
The review by Timothy Baker (Eldorado Gold) helped improve the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A
The results of mineral abundance mapping for DC-2 to DC-5 are shown in Figures A1–A8 using
all test scenarios.

Figure A1. Drill-core abundance maps of white mica (WM), biotite (Bt), chlorite (Chl), amphibole
(Amp), gypsum (Gp), feldspar (Fsp), quartz (Qz), sulphide (SP) and accessory minerals (Other) for
DC-2 using randomly distributed 50% of the available ground truth data for training for random forest
(RF), support vector machine (SVM) and feed-forward neural network (FF-ANN) regressions.
Remote Sens. 2020, 12, 1218 18 of 23

Figure A2. White mica (WM), biotite (Bt), chlorite (Chl), amphibole (Amp), gypsum (Gp), feldspar (Fsp),
quartz (Qz) and sulphide (SP) abundance maps of TS-2a using TS-2b for training and of TS-2b using
TS-2a respectively for training of random forest (RF), support vector machine (SVM) and feed-forward
neural network (FF-ANN) regressions. The ground truth (GT) represented by resized MLA data is
presented for comparison.

Figure A3. Drill-core abundance maps of white mica (WM), biotite (Bt), chlorite (Chl), amphibole
(Amp), gypsum (Gp), feldspar (Fsp), quartz (Qz), sulphide (SP) and accessory minerals (Other) for
DC-3 using randomly distributed 50% of the available ground truth data for training for random forest
(RF), support vector machine (SVM) and feed-forward neural network (FF-ANN) regressions.
Remote Sens. 2020, 12, 1218 19 of 23

Figure A4. White mica (WM), chlorite (Chl), amphibole (Amp), gypsum (Gp), feldspar (Fsp), quartz
(Qz) and sulphide (SP) abundance maps of TS-3a using TS-3b for training and of TS-3b using TS-3a
respectively for training of random forest (RF), support vector machine (SVM) and feed-forward neural
network (FF-ANN) regressions. The ground truth (GT) represented by resized MLA data is presented
for comparison.

Figure A5. Drill-core abundance maps of white mica (WM), chlorite (Chl), gypsum (Gp), feldspar
(Fsp), quartz (Qz) and sulphide (SP) for DC-4 using randomly distributed 50% of the available ground
truth data for training for random forest (RF), support vector machine (SVM) and feed-forward neural
network (FF-ANN) regressions.
Remote Sens. 2020, 12, 1218 20 of 23

Figure A6. White mica (WM), gypsum (Gp), feldspar (Fsp), quartz (Qz) and sulphide (SP) abundance
maps of TS-4a using TS-4b for training and of TS-4b using TS-4a respectively for training of random
forest (RF), support vector machine (SVM) and feed-forward neural network (FF-ANN) regressions.
The ground truth (GT) represented by resized MLA data is presented for comparison.

Figure A7. Drill-core abundance maps of white mica (WM), biotite (Bt), chlorite (Chl), amphibole (Amp),
carbonate (Cb), gypsum (Gp), feldspar (Fsp), quartz (Qz), sulphide (SP) and accessory minerals (Other)
for DC-5 using randomly distributed 50% of the available ground truth data for training for random
forest (RF), support vector machine (SVM) and feed-forward neural network (FF-ANN) regressions.
Remote Sens. 2020, 12, 1218 21 of 23

Figure A8. White mica (WM), chlorite (Chl), amphibole (Amp), carbonate (Cb), gypsum (Gp), feldspar
(Fsp), quartz (Qz) and sulphide (SP) abundance maps of TS-5a using TS-5b for training and of TS-5b
using TS-5a respectively for training of random forest (RF), support vector machine (SVM) and
feed-forward neural network (FF-ANN) regressions. The ground truth (GT) represented by resized
MLA data is presented for comparison.

References
1. Gandhi, S.M.; Sarkar, B.C.; Gandhi, S.M.; Sarkar, B.C. Chapter 8—Drilling. In Essentials of Mineral Exploration
and Evaluation; Elsevier: Amsterdam, The Netherlands, 2016.
2. Lindholm, R.C. Mineral identification using X-ray diffraction. In A Practical Approach to Sedimentology;
Springer: Dordrecht, The Netherlands, 1987; pp. 124–153.
3. Fandrich, R.; Gu, Y.; Burrows, D.; Moeller, K. Modern SEM-based mineral liberation analysis. Int. J. Miner.
Process. 2007, 84, 310–320. [CrossRef]
4. Kirsch, M.; Lorenz, S.; Zimmermann, R.; Tusa, L.; Möckel, R.; Hödl, P.; Booysen, R.; Khodadadzadeh, M.;
Gloaguen, R. Integration of terrestrial and drone-borne hyperspectral and photogrammetric sensing methods
for exploration mapping and mining monitoring. Remote Sens. 2018, 10, 1366. [CrossRef]
5. Calvin, W.M.; Pace, E.L. Mapping alteration in geothermal drill core using a field portable spectroradiometer.
Geothermics 2016, 61, 12–23. [CrossRef]
6. Kruse, F.A. Identification and mapping of minerals in drill core using hyperspectral image analysis of
infrared reflectance spectra. Int. J. Remote Sens. 1996, 17, 1623–1632. [CrossRef]
7. Acosta, I.C.C.; Khodadadzadeh, M.; Tusa, L.; Ghamisi, P.; Gloaguen, R. A Machine Learning Framework for
Drill-Core Mineral Mapping Using Hyperspectral and High-Resolution Mineralogical Data Fusion. IEEE J.
Sel. Top. Appl. Earth Obs. Remote Sens. 2019, 12, 4829–4842. [CrossRef]
8. Clark, R.N. Spectroscopy of Rocks and Minerals and Principles of Spectroscopy. Manual Remote Sens. 1999, 3,
2–52.
9. Pontual, S.; Merry, N.; Gamson, P. Spectral Interpretation Field Manual; AusSpec International Ltd.: Queenstown,
New Zealand, 1997; Volume I.
10. Mauger, A.J.; Keeling, J.L.; Huntington, J.F. Alteration mapping of the Tarcoola Goldfield (South Australia)
using a suite of hyperspectral methods. Appl. Earth Sci. 2007, 116, 2–12. [CrossRef]
11. Huntington, J.F.; Mauger, A.J.; Skirrow, R.G.; Bastrakov, E.N.; Connor, P.; Mason, P.; Keeling, J.L.;
Coward, D.; Berman, M.; Phillips, R.; et al. Automated mineralogical core logging at the Emmie Bluff iron
oxide-copper-gold Prospect. Mesa J. 2006, 41, 38–44.
12. Roache, T.J.; Walshe, J.L.; Huntington, J.F.; Quigley, M.A.; Yang, K.; Bil, B.W.; Blake, K.L.; Hyvärinen, T.
Epidote-clinozoisite as a hyperspectral tool in exploration for Archean gold. Aust. J. Earth Sci. 2011, 58,
813–822. [CrossRef]
Remote Sens. 2020, 12, 1218 22 of 23

13. Mathieu, M.; Roy, R.; Launeau, P.; Cathelineau, M.; Quirt, D. Alteration mapping on drill cores using
a HySpex SWIR-320m hyperspectral camera: Application to the exploration of an unconformity-related
uranium deposit (Saskatchewan, Canada). J. Geochemical Explor. 2017, 172, 71–88. [CrossRef]
14. Tappert, M.; Rivard, B.; Giles, D.; Tappert, R.; Mauger, A. Automated drill core logging using visible and
near-infrared reflectance spectroscopy: A case study from the Olympic Dam Iocg deposit, South Australia.
Econ. Geol. 2011, 106, 289–296. [CrossRef]
15. Littlefield, E.; Calvin, W.; Stelling, P.; Kent, T. Reflectance spectroscopy as a drill core logging technique:
An example using core from the Akutan. GRC Transaction 2012, 2011, 1281–1283.
16. Bedini, E.; van der Meer, F.; van Ruitenbeek, F. Use of HyMap imaging spectrometer data to map mineralogy
in the Rodalquilar caldera, southeast Spain. Int. J. Remote Sens. 2009, 30, 327–348. [CrossRef]
17. Kratt, C.; Calvin, W.M.; Coolbaugh, M.F. Mineral mapping in the Pyramid Lake basin: Hydrothermal
alteration, chemical precipitates and geothermal energy potential. Remote Sens. Environ. 2010, 114, 2297–2304.
[CrossRef]
18. Jakob, S.; Zimmermann, R.; Gloaguen, R. 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 Sens. 2017, 9, 88. [CrossRef]
19. Ruffin, C.; King, R.L. The analysis of hyperspectral data using Savitzky-Golay filtering-theoretical basis.
In Proceedings of the IEEE 1999 International Geoscience and Remote Sensing Symposium IGARSS 99
(Cat. No.99CH36293), Hamburg, Germany, 28 June–2 July 1999; Volume 2, pp. 756–758.
20. Rodarmel, C.; Shan, J. Principal Component Analysis for Hyperspectral Image Classification. Surv. L. Inf.
Syst. 2002, 62, 115–122.
21. Kern, M.; Möckel, R.; Krause, J.; Teichmann, J.; Gutzmer, J. Calculating the deportment of a fine-grained and
compositionally complex Sn skarn with a modified approach for automated mineralogy. Miner. Eng. 2018,
116, 213–225. [CrossRef]
22. Gu, Y. Automated scanning electron microscope based mineral liberation analysis. J. Miner. Mater. Charact.
Eng. 2003, 2, 33–41.
23. Milu, V.; Leroy, J.L.; Piantone, P. Le gisement de cuivre-or de Bolcana (monts Métallifères Roumanie) :
Premières données sur les altérations et minéralisations associées. Comptes Rendus Geosci. 2003, 335, 671–680.
[CrossRef]
24. Blannin, R.; Tusa, L.; Birtel, S.; Gutzmer, J.; Gilbricht, S.; Ivascanu, P. Metal deportment and ore variability of
the Bolcana porphyry Au-Cu system (Apuseni Mts, Romania)-Implications for ore processing. In Proceedings
of the 15th Biennial Meeting of the Society for Geology Applied to Mineral Deposits, Glasgow, UK, 27–30
August 2019.
25. Tusa, L.; Andreani, L.; Khodadadzadeh, M.; Contreras, C.; Ivascanu, P.; Gloaguen, R.; Gutzmer, J. Mineral
Mapping and Vein Detection in Hyperspectral Drill-Core Scans: Application to Porphyry-Type Mineralization.
Minerals 2019, 9, 122. [CrossRef]
26. Ivascanu, P.; Baker, T.; Lewis, P.; Kulcsar, Z.; Denes, R.; Tamas, C. Bolcana, Romania: Geology and Discovery
History of a Gold Rich Porphyry Deposit; NewGenGold: Perth, Australia, 2019.
27. Sillitoe, R.H. Porphyry Copper Systems. Econ. Geol. 2010, 105, 3–41. [CrossRef]
28. Draper, C.; Reichle, R.; de Jeu, R.; Naeimi, V.; Parinussa, R.; Wagner, W. Estimating root mean square errors
in remotely sensed soil moisture over continental scale domains. Remote Sens. Environ. 2013, 137, 288–298.
[CrossRef]
29. Waske, B.; Benediktsson, J.A.; Arnason, K.; Sveinsson, J.R. Mapping of hyperspectral AVIRIS data using
machine-learning algorithms. Can. J. Remote Sens. 2009, 35, S106–S116. [CrossRef]
30. Rodriguez-Galiano, V.; Sanchez-Castillo, M.; Chica-Olmo, M.; Chica-Rivas, M. Machine learning predictive
models for mineral prospectivity: An evaluation of neural networks, random forest, regression trees and
support vector machines. Ore Geol. Rev. 2015, 71, 804–818. [CrossRef]
31. Ghamisi, P.; Plaza, J.; Chen, Y.; Li, J.; Plaza, A.J. Advanced Spectral Classifiers for Hyperspectral Images:
A review. IEEE Geosci. Remote Sens. Mag. 2017, 5, 8–32. [CrossRef]
32. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
33. Cortes, C.; Vapnik, V. Dispensing system boosts throughput 50 percent. Assembly 2001, 44, 97.
34. Vapnik, V.N. An overview of statistical learning theory. IEEE Trans. Neural Netw. 1999, 10, 988–999. [CrossRef]
35. Specht, D.F. A General Regression Neural Network. IEEE Trans. Neural Netw. 1991, 2, 568–576. [CrossRef]
Remote Sens. 2020, 12, 1218 23 of 23

36. Rumelhart, D.E.; Hintont, G.E. Learning representations by back-propagating errors. Nature 1986, 323,
533–536. [CrossRef]
37. Van den Boogaart, K.G.; Tolosana-Delgado, R. Predictive Geometallurgy: An Interdisciplinary Key Challenge
for Mathematical Geosciences. In Handbook of Mathematical Geosciences: Fifty Years of IAMG; Daya Sagar, B.S.,
Cheng, Q., Agterberg, F., Eds.; Springer International Publishing: Cham, Switzerland, 2018; pp. 673–686.

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).