Minerals Engineering 105 (2017) 10–18

Contents lists available at ScienceDirect

Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng

Discriminating ore and waste in a porphyry copper deposit using

short-wavelength infrared (SWIR) hyperspectral imagery
M. Dalm a,⇑, M.W.N. Buxton a, F.J.A. van Ruitenbeek b
a
Resource Engineering, Faculty of Civil Engineering and Geosciences, Delft University of Technology, P.O. Box 5048, 2600 GA Delft, The Netherlands
b
Department of Earth Systems Analysis, Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente, P.O. Box 6, 7500 AA Enschede, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: A recent study by Dalm et al. (2014) showed that alteration mineralogy acquired using SWIR point spec-
Received 9 August 2016 trometry could be linked to copper grade distribution for a group of samples from a South American cop-
Revised 20 December 2016 per mine. Since it was expected that SWIR hyperspectral imagery can provide more detailed information
Accepted 21 December 2016
about the alteration mineralogy of these ores, we investigated whether this technique can be used to
improve upon the indirect characterization of copper grades. Maps showing the distributions of SWIR-
active minerals, white mica crystallinity, white mica composition, and chlorite composition were pro-
Keywords:
duced from SWIR hyperspectral images of 43 samples from the Dalm et al. (2014) study.
SWIR hyperspectral imagery
Porphyry Cu
Subsequently, a principle component analysis (PCA) was applied to the relative mineral abundances
Sensor-based sorting and the average white mica crystallinity and composition that were extracted from these maps. The
Ore characterization PCA showed that this mineralogical data could be used to discriminate a significant portion of the sam-
ples with sub-economic copper grades. Furthermore, the study showed that SWIR hyperspectral imaging
has the following advantages over SWIR point spectrometry: minerals that are present in relatively low
quantities can be detected, the SWIR-active mineralogical composition at the surface of a sample can be
quantified, and the texture of samples, such as grain sizes and cross-cutting vein structures, can be char-
acterized. However, these advantages did not improve upon the indirect characterization of copper
grades that was achieved using SWIR point spectrometry. This was attributed to the relatively small size
of the sample set and the high textural variability between samples.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction During the last decade instruments came onto the market that are
dedicated to acquiring SWIR hyperspectral images of drill core
Short-wavelength infrared (SWIR) hyperspectral sensors are samples. Examples are Corescan’s Hyperspectral Core Imager
used to capture images of objects at many different wavelengths (Corescan, 2015) and Specim’s SisuRock (Specim, 2015).
in the SWIR spectral range (1000–2500 nm). Each pixel in the The mineralization of copper in porphyry ore deposits is associ-
resulting image constitutes a spectrum of the intensity of reflected ated with the formation of certain alteration minerals (Lowell and
radiation as a function of wavelength. Certain minerals absorb Guilbert, 1970; Pirajno, 1992; Sillitoe, 2010). These alteration min-
SWIR radiation at specific wavelengths as a result of vibrating erals are often SWIR-active minerals (Sabins, 1999; Thompson
molecule bonds (Hunt, 1977; Clark, 1999). SWIR hyperspectral et al., 1999; Dalm et al., 2014). Because of this, SWIR hyperspectral
imaging can therefore be used to identify the mineralogy of rock sensors can potentially be used to indirectly characterize the cop-
samples. However, mineral identification is restricted to minerals per grade distribution of porphyry copper ores. This would allow
containing molecule bonds that produce diagnostic absorptions. these sensors to be implemented along the mining value chain to
Such minerals will be referred to as SWIR-active minerals. provide real-time information on ore value.
In geology and mineral exploration, SWIR hyperspectral sensors One specific application of the SWIR sensor is in sensor-based
are mainly known from remote sensing applications that scan the sorting. Sensor-based sorting is a technique in which singular par-
Earth’s surface from satellites and aircraft. The same sensors can ticles are mechanically separated on certain physical and/or chem-
also be applied to scan objects of sizes in the order of centimeters. ical properties after determining these properties with a sensor. If a
SWIR sensor can be used to identify waste or sub-economic ore
material, sensor-based sorting could be incorporated into mineral
⇑ Corresponding author.
processing operations in order to remove this material prior to
E-mail address: [email protected] (M. Dalm).

http://dx.doi.org/10.1016/j.mineng.2016.12.013
0892-6875/Ó 2016 Elsevier Ltd. All rights reserved.
 M. Dalm et al. / Minerals Engineering 105 (2017) 10–18 11

conventional treatment. This has potential to reduce the costs copper grades in the sample set determined by XRF analysis is pre-
associated with processing mined materials (Salter and Wyatt, sented in Fig. 1.
1991; Sivamohan and Forssberg, 1991; Buxton & Benndorf, 2013; On 20 of the pulverized samples, X-ray diffraction (XRD) analy-
Wills and Finch, 2016). sis was performed to determine the mineralogy. Since XRD is lim-
This paper investigates the applicability of SWIR hyperspectral ited by a relatively high detection limit, mineralogy was also
imagery for characterizing the copper grade of 43 porphyry copper determined by petrography for 11 out of the 43 samples. The thin
ore samples. These samples are a subset of the sample set of 150 sections for petrography were taken from the cut surface of the
samples that was used in the study described by Dalm et al. sample halves used for XRF, before these were pulverized. Petrog-
(2014). Ore minerals in the samples include chalcopyrite, chal- raphy and XRD analysis were performed on different samples. The
cocite, covellite, digenite and bornite. These are all copper sul- determined mineralogy from these analyses was used to validate
phides that do not produce diagnostic absorption features in the the occurrence of minerals that were identified from the SWIR
SWIR (Hunt et al., 1971; Clark et al., 2007). hyperspectral images.
Dalm et al. (2014) showed that an indirect relationship exists SWIR hyperspectral images of the samples were acquired with a
between the copper grade of the samples and the alteration miner- SisuCHEMA instrument (SPECIM Ltd., Oulu, Finland). The hyper-
alogy that can be determined with SWIR point spectrometry. How- spectral images were recorded in a 940–2540 nm spectral range
ever, each spectrum that is acquired with a commonly used point and at a spectral resolution of 6 nm. The images were made from
spectrometer represents a sample surface of several cm2 (e.g., ASD, the cut surface of the samples because this surface is flat, smooth
2015). Since most hyperspectral imagers acquire spectra on a pixel and uncontaminated by dust. The size of individual pixels in the
size below 1 mm2 (e.g., Corescan, 2015; Specim, 2015), these resulting images is around 0.2 by 0.2 mm.
instruments can provide more detailed information about the
abundance and spatial distribution of the SWIR-active mineralogy. 2.2. Image pre-processing
It is therefore expected that SWIR hyperspectral imagery will
reveal additional differences between samples, which might be The SWIR hyperspectral images needed pre-processing to con-
used to improve upon the indirect characterization of copper vert the recorded data to relative reflections. This was performed
grades. by using a white reference and a dark current calibration. A white
In this paper we give an example of how SWIR hyperspectral reference is a measurement of a material with almost 100% reflec-
images can be processed to produce maps that show the distribu- tion over the entire SWIR spectral range. A dark current is a mea-
tion of SWIR-active minerals. We will also show how certain min- surement that is performed when no light enters the sensor in
eral characteristics such as the white mica crystallinity can be order to capture instrumental noise. Reflection spectra were calcu-
mapped from these images. Furthermore, we present a method lated by dividing the measured spectrum by the white reference
that uses these maps as input variables in a principle component after subtracting the dark current from both these measurements.
analysis (PCA) in order to find samples that have similar miner-
alogical composition and to identify the mineralogical differences 2.3. Image processing
between sample groups. Finally, we will show how this can be
used to segment populations of ore and waste. The SWIR hyperspectral images were processed using ENVI 4.5
image processing software (ITT visual information studios, 2008).
The first step was to mask the images by removing the data of pix-
els that do not represent a sample. This was performed by manu-
2. Methods ally selecting regions of interest on the images. Subsequently, a
mineral map was produced from the masked image of each sample
2.1. Data acquisition by using the Spectral Angle Mapping (SAM) algorithm (Kruse et al.,
1993) in combination with a set of reference spectra. SAM classifies
Forty-three samples were collected from a South American each image pixel by assigning the mineral or mineral assemblage
mine that exploits a porphyry copper ore deposit. The samples to it that is represented by the reference spectrum to which the
are individual rock particles of around 5–7 cm in diameter. Prior pixel spectrum has the smallest multi-dimensional angle. A maxi-
to any measurement, all the samples were cut in half. One half of mum angle of 0.2 rad was applied as a threshold to ensure a posi-
each sample was used for acquiring SWIR hyperspectral images. tive correlation between the image pixels and reference spectra.
The other half was pulverized and used for X-ray fluorescence The SAM method was chosen for constructing mineral maps
(XRF) analysis to determine the copper grade. The distribution of because numerous studies show that it is an effective method to
classify the pixels of hyperspectral images (e.g., Girouard et al.,
2004; Rowan et al., 2005; Kruse, 2012).
The reference spectra used with SAM were selected from the
measured SWIR hyperspectral images so that the spectral range
and resolution of the used references and the image pixels are
equal. The mineral or mineral assemblage that is represented by
each reference spectrum was determined by comparing the spec-
trum with the SWIR spectra of the USGS spectral library (Clark
et al., 2007) and the spectral interpretation field manual, G-MEX
(AusSpec, 2008). This comparison was performed by inspecting
the wavelength positions and relative intensities of the absorption
features in the spectra.
An initial set of reference spectra was collected by manually
selecting spectra that represented the observed variability in the
measured SWIR hyperspectral images. SAM was applied to produce
mineral maps and the classification of randomly chosen pixels was
Fig. 1. Distribution of copper grades in the sample set determined by XRF analysis. checked by inspecting the spectra. Pixels that were incorrectly
12 M. Dalm et al. / Minerals Engineering 105 (2017) 10–18

classified or that were not classified due to the maximum angle higher variance than all the others. This was needed since the vari-
threshold were added to the set of reference spectra after the min- ables that were used in the PCA are expressed in different units.
erals that produced the spectrum were identified. This process was The PCA sample scores were compared with the XRF results to
repeated until no unclassified pixels remained and no incorrect investigate if there is a relationship between the copper grade of
classifications were observed. the ore particles and the mineralogical information that can be
Apart from the mineralogy, the following mineral characteris- derived from SWIR hyperspectral imagery. The PCA loadings were
tics were also mapped from the SWIR hyperspectral images of used to identify the minerals and/or mineral characteristics that
the samples: white mica composition, white mica crystallinity, describe this relationship.
and chlorite composition. This was based on the wavelength posi-
tions and depth ratios of specific mineral absorption features.
These spectral properties were calculated with the Hyperspectral 3. Results
Python (HypPy) toolbox (Bakker, 2012). The algorithms in this
toolbox fit a parabola to the wavelength channels around an 3.1. Minerals identified from the measured SWIR hyperspectral images
absorption feature to interpolate the wavelength position of the
absorption minimum. The absorption depth is calculated by divid- 3.1.1. White mica
ing the reflectance value of the absorption minimum by that of the White mica is the most abundant SWIR-active group of miner-
continuum around the absorption feature. Mineral characteristics als in almost all SWIR hyperspectral images of the samples. White
were only mapped for pixels that represent pure mineral spectra mica is an umbrella term that usually refers to muscovite, parago-
(i.e. spectra that only show absorption features of a single mineral). nite, celadonite, aluminoceladonite, illite, illite-smectite or any
This was done by masking the image with the resulting mineral coexisting occurrence of these minerals. These different white mica
map. species can be distinguished from each other with SWIR spectrom-
etry (AusSpec, 2008; Clark et al., 2007; Duke, 1994).
Fig. 2 presents several SWIR spectra of white mica that were
2.4. Principle component analysis used as reference spectra for producing mineral maps with SAM.
As mentioned in Section 2.3., these reference spectra originate
A principle component analysis (PCA) was applied to the miner- from the measured SWIR hyperspectral images. Spectrum 1 in
alogical information contained in the maps that were produced Fig. 2 is characteristic for muscovite and spectrum 5 is characteris-
from the SWIR hyperspectral images. PCA is a procedure in which tic for illite. The other spectra represent intermediate muscovite-
data is projected onto a k-dimensional space by a linear orthogonal illite species. The absorption features that can be observed in the
transformation (Wold et al., 1987; Jolliffe, 2002; Eriksson et al.,
2006). These k-dimensions are commonly referred to as the k num-
ber of principle components. The first principle component (PC)
represents the direction that describes the largest amount of vari-
ance in the variable space of the original data. Each subsequent PC
represents the direction of the largest remaining variance that is
orthogonal to those of the previous PCs. The projection of the min-
eralogical data on the first few PCs can therefore be used to easily
distinguish between samples that are relatively similar and those
that are more different. Furthermore, the contributions of the orig-
inal variables to each PC can be used to identify the variables for
which the samples differ most (Eriksson et al., 2006). The projected
data onto the PC space is referred to as the sample scores and the
contributions of the original variables are known as the loadings.
PCA was performed on a data matrix containing the relative
abundance of all the identified SWIR-active minerals for each sam-
ple and the sample averages of the features representing white
mica crystallinity and white mica composition. Spectral features
that describe the composition of chlorite were not included in
the PCA since not all samples contained chlorite. The relative abun-
dance of the SWIR-active minerals was calculated by dividing the
number of pixels that represent a certain mineral by the number
of pixels that represent the sample. A multiplicative zero replace-
ment followed by a centered log-ratio transformation was subse-
quently applied to this data (Martin-Fernandez et al., 2003;
Aitchison, 1986). This was required to correct for the fact that
the mineralogical compositions sum up to a constant (i.e. 100%).
This constant sum constraint can influence the outcomes of the
PCA since it places restrictions on the correlation structure of the
data (Aitchison, 1986).
The data matrix was centered and scaled before applying the
PCA by setting the mean and standard deviation of each variable
to 0 and 1 respectively. Setting the mean to 0 places the average
of the data at the origin of the variable space. This is required to
Fig. 2. SWIR spectra of white mica with decreasing crystallinity from spectrum 1–5.
ensure that the PCs will describe maximum variance (Jolliffe, Spectrum 1 is characteristic for muscovite and spectrum 5 is characteristic for illite.
2002). Setting the standard deviation of each variable to 1 prevents Spectra 2–4 represent intermediate muscovite-illite species. Spectra originate from
the PCA model being dominated by a variable that has a much the measured SWIR hyperspectral images.
 M. Dalm et al. / Minerals Engineering 105 (2017) 10–18 13

spectra are produced by vibrating molecule bonds. The diagnostic

white mica absorption occurs around 2200 nm and is produced
by an overtone combination of the OAH stretch and the AlAOH
bend vibration (Hunt, 1977; Clark et al., 1990). This absorption will
be referred to as the white mica AlAOH absorption feature.
The exact wavelength position of the minimum of the AlAOH
absorption feature correlates with the Al content of white mica
(Duke, 1994; Post and Noble, 1993). Muscovite is a common end-
member of white mica for which the AlAOH absorption minimum
is located around 2200 nm (AusSpec, 2008; Duke, 1994; Post and
Noble, 1993). Substitution of the Al by Fe and Mg in the muscovite
crystal structure through octahedral exchange (Al ? (Fe, Mg)) or
octahedral-tetrahedral Tschermak exchange (AlAl ? (Fe, Mg)Si)
shifts the minimum position of this feature to longer wavelengths
(Duke, 1994). The white mica endmembers resulting from this sub-
stitution are aluminoceladonite and celadonite. These endmem-
bers often form solid solutions with muscovite, which are
referred to as phengite (Rieder et al., 1998). The AlAOH absorption
minimum for phengite is located around 2225 nm (AusSpec, 2008).
Minimum positions of the white mica AlAOH absorption feature in
the measured SWIR hyperspectral images range from 2198 to
2220 nm. This indicates that the white mica in the samples has a
muscovitic to phengitic composition.
Fig. 2 shows that the intensity of the white mica absorption
around 1900 nm is variable. This absorption is produced by an
overtone combination of the OAH stretch and HAOAH bend vibra-
tion of molecular water (Hunt, 1977; Clark et al., 1990). It will be
referred to as the H2O absorption feature. It is possible to charac-
terize the white mica crystallinity from the intensity of this feature
(AusSpec, 2008; Dalm et al., 2014). This is because less crystalline
white mica has more interlayer sites that are occupied by water
molecules (Dalm et al., 2014). A white mica crystallinity index
can be calculated by dividing the absorption depth of the AlAOH Fig. 3. SWIR spectra of: 1; chlorite, 2; tourmaline, 3; kaolinite, 4; nontronite.
absorption by that of the H2O absorption. Muscovite and phengite Spectra originate from the measured SWIR hyperspectral images.
are white mica species with a high crystallinity index. White mica
species with similar composition, but a low crystallinity index, are
1987). When the Mg is partly substituted by Fe, the minimum of
referred to as illite and phengitic illite.
both absorption features shift to longer wavelengths. For Fe-rich
The absorption that is located around 1410 nm is an overtone of
chlorite the minimum of the FeAOH and MgAOH absorptions are
the OAH stretch vibration (Hunt, 1977; Clark et al., 1990). All min-
located around 2265 and 2365 nm (AusSpec, 2008; Mc Leod
erals that contain OH bonds will produce such an absorption fea-
et al., 1987). Minimum positions of the chlorite FeAOH and
ture. The exact minimum position of this absorption feature
MgAOH absorptions in the measured SWIR hyperspectral images
depends on the cation that the OH bond is attached to Hunt
range from 2248 to 2261 nm and from 2324 to 2360 nm respec-
(1977) and Clark et al. (1990).
tively. This indicates that the chlorite in the samples varies from
Mg-rich to Fe-rich, corresponding to clinochlore-chamosite chlo-
3.1.2. Chlorite
rite species.
Minerals from the chlorite group were identified from the SWIR
hyperspectral images of some of the samples. Chlorite is an
umbrella term that usually refers to clinochlore, chamosite, or a 3.1.3. Tourmaline, kaolinite and nontronite
solid solution between these two endmembers. One of the refer- Spectra 2–4 in Fig. 3 are examples of reference spectra that
ence spectra that was used for the SAM classification of chlorite were used for the SAM classification of tourmaline, kaolinite and
is spectrum 1 in Fig. 3. Chlorite produces diagnostic absorptions nontronite. Tourmaline produces diagnostic absorptions at 2210,
around 2250 and 2350 nm (Clark et al., 2007; AusSpec, 2008). Min- 2250, 2300 and 2360 nm (Clark et al., 2007; AusSpec, 2008). These
eral absorptions in the 2230–2295 nm range commonly result absorption features are mainly produced by the FeAOH and
from OAH stretch and FeAOH bend vibration (AusSpec, 2008; MgAOH bonds of tourmaline (AusSpec, 2008). However, some
Clark et al., 1990). Furthermore, absorptions produced by OAH authors suggest that the tourmaline absorption features can also
stretch and MgAOH bend vibration commonly occur in the be caused by overtone combination of the OAH stretch and BAOH
2300–2370 nm range (AusSpec, 2008; Clark et al., 1990). The chlo- bend vibration (Clark et al., 1990).
rite absorptions that occur around 2250 and 2350 nm will there- Spectrum 3 in Fig. 3 shows double absorption features around
fore be referred to as the chlorite FeAOH and MgAOH absorption 1400 and 2200 nm that are diagnostic for kaolinite. The absorp-
feature respectively. tions around 1400 nm are produced by the OAH bonds and the
The exact wavelength position of the minimum of the FeAOH ones around 2200 nm by the AlAOH bonds of kaolinite (Hunt,
and MgAOH absorption features are related to the relative amount 1977; Clark et al., 1990). The reason that a double absorption fea-
of Mg versus Fe ions that occur within the chlorite crystal structure ture, or doublet, is produced at both wavelength positions is
(AusSpec, 2008; Mc Leod et al., 1987). If the chlorite is rich in Mg related to the fact that kaolinite has OH groups at different loca-
the minimum position of the FeAOH and MgAOH absorptions tions in the crystal structure. OAH stretch vibration of OH groups
are around 2245 and 2320 nm (AusSpec, 2008; Mc Leod et al., located on the exterior surface of the octahedral sheets absorb
14 M. Dalm et al. / Minerals Engineering 105 (2017) 10–18

radiation at a slightly shorter wavelength than that of the OH might affect the shape and intensity of these features.
groups located in between the tetrahedral and octahedral sheets Figs. 4c and 4d show that mapping the characteristics of white
(Crowley and Vergo, 1988). The double feature around 2200 nm mica reveals additional differences between samples. Furthermore,
is produced because both types of OH groups form AlAOH bonds. these maps provide information on the texture of samples since
Spectrum 4 in Fig. 3 has a double absorption feature around vein structures containing white mica with a different composition
1420 nm and a single feature around 2290 nm that are diagnostic and/or crystallinity become visible.
for nontronite. Most nontronite spectra in spectral libraries only Fig. 5 presents two maps of the wavelength position of the chlo-
show a single absorption feature around 1420 nm (e.g., AusSpec, rite MgAOH absorption minimum. These maps correspond to sam-
2008; Clark et al., 2007). However, a double absorption feature at ple 3 and 4 presented in Fig. 4. Calculating the minimum position
this wavelength position has been reported for nontronite by of the chlorite MgAOH absorption was only performed on pixels
Clark et al. (1990). It is produced by different OAH stretching with pure chlorite spectra. Fig. 5 shows that mapping this charac-
vibrations similar to that of kaolinite. The absorption around teristic of chlorite reveals additional differences between samples.
2290 nm is produced by the FeAOH bonds of nontronite (Clark
et al., 1990; AusSpec, 2008).
3.3. Principle component analysis
3.1.4. Other minerals
Apart from the SWIR spectra presented in Figs. 2 and 3, spectra Fig. 6 presents the PCA sample scores of the first and second
with a relatively low reflection and no mineral absorption features principle component (PC). The blue dots in this figure represent
were found in the measured SWIR hyperspectral images. Minerals samples that are waste and the red dots samples that are ore.
that were not identified from the SWIR hyperspectral images, but The definition of ore and waste is based on an assumed economic
that were determined by XRD and petrography include quartz, cut-off grade of 0.40% Cu, which is representative for the mine that
feldspars (orthoclase, albite and anorthite), hematite, magnetite, the samples originate from. The cut-off grade is the minimum cop-
pyrite and the ore minerals chalcopyrite, digenite, bornite, chal- per grade that is required for economic processing of the ore
cocite and covellite. None of these minerals contain molecule material.
bonds that produce absorption features in the SWIR (Clark et al., PCA was performed on a data matrix containing the relative
2007). However, quartz and feldspars are somewhat transparent SWIR-active mineral abundances for each sample and the sample
and the spectra of the hyperspectral images will therefore show averages of the white mica crystallinity index and the wavelength
absorption features of underlying SWIR-active minerals where position of the white mica AlAOH absorption feature. The black
quartz and feldspars are located. All the other minerals that were lines in Fig. 6 represent the loadings of these variables. These were
determined by XRD and petrography are opaque minerals that all equally scaled to fit the figure and therefore do not represent
are known to have a low reflection in the SWIR (Clark et al., the actual loading values. The capital letters indicate the type of
2007; Baldridge et al., 2009). Since it is not possible to distinguish mineral that is represented by the loadings. Cryst and AlAOH refers
between the different types of opaque minerals from the SWIR to the white mica crystallinity index and the minimum position of
spectra, all spectra with a low reflection and without mineral the white mica AlAOH absorption. The chlorite MgAOH absorption
absorption features were classified as opaque minerals when pro- was not included in the PCA since not all samples contain chlorite.
ducing the mineral maps with SAM. The green rectangle in Fig. 6 indicates a region in the score plot
where mainly waste samples occur. It results that these can be dis-
3.2. Mapping SWIR-active mineralogy tinguished by using the first PC. It can be inferred from the loadings
of this PC that the samples in the green rectangle have relatively
Fig. 4 presents four examples of the results of mapping the high contents of chlorite and low contents of white mica, tourma-
SWIR-active mineralogy from the measured SWIR hyperspectral line and opaque minerals. Furthermore, these samples contain
images. The figure shows a photo of each sample, the mineral white mica with relatively low crystallinity values and with AlAOH
map resulting from the SAM classifications and maps of the AlAOH features that occur at longer wavelengths. The white mica species
feature position and the crystallinity index of white mica. of the samples in the green rectangle are therefore close to a phen-
For all samples white mica minerals were mapped. A relatively gitic illite while that of the other samples are closer to muscovite.
high abundance of either chlorite or tourmaline was common. Sev- It can be observed from Fig. 6 that the green rectangle contains
eral samples contain both chlorite and tourmaline. In that case, the one ore sample. This sample has a relatively high score on the sec-
abundance of chlorite was always higher than that of tourmaline. ond PC. It can be inferred from the loadings of this PC that this
Nontronite, kaolinite and opaque minerals only occurred in rela- relates to relatively high contents of tourmaline and opaque min-
tively low abundances. Furthermore, nontronite was not observed erals and low contents of nontronite and white mica. Furthermore,
in the mineral maps of samples with relatively high tourmaline high scores on the second PC are associated with white mica
contents. AlAOH features that occur at longer wavelengths.
The mineral maps in Fig. 4b show classifications of pure mineral Fig. 7 presents the resulting copper grade distributions when
spectra as well as mixtures with white mica. These mixtures are the samples are classified on the PCA scores of the first PC. A score
SWIR spectra in which the absorption features of white mica as of 1 was used as a threshold for the classification. Samples with
well as some other mineral can be observed. For nontronite and PCA scores above this threshold are all samples inside the green
kaolinite, the same color was assigned to the pure spectra and rectangle in Fig. 6. Although Fig. 6 shows that the second PC can
the spectral mixtures. This was done to improve the visibility of be used to also distinguish the ore sample in the green rectangle
these minerals, since the number of pixel spectra that contain from most of the waste samples, this was not performed for the
absorption features of nontronite or kaolinite is relatively low. classification result presented in Fig. 7. The reason is that this ore
Mixed spectra with opaque minerals were not classified. A pixel sample has a copper grade of 0.47%, which is only just above the
was only classified as an opaque mineral when no mineral absorp- economic cut-off grade of 0.40% Cu. Furthermore, the PCA scores
tions could be observed in the spectrum due to a low reflection. of this ore sample are relatively close to the scores of a significant
Calculating the white mica crystallinity index and the AlAOH portion of the waste samples. This means that the ore sample can-
absorption minimum was only performed on pixels with pure not be separated from these waste samples unless thresholds are
white mica spectra. This is because absorptions by other minerals used that are based on relatively small differences between the
 M. Dalm et al. / Minerals Engineering 105 (2017) 10–18 15

Fig. 4. Result of mapping mineralogy from selected SWIR hyperspectral images. 1Mapping the wavelength position of the white mica AlAOH absorption feature and the white
mica crystallinity index was only performed on pure white mica spectra (i.e. no absorption features by any other mineral).

SWIR-active mineralogy of the samples. Using such thresholds will a significant portion of the sub-economic ore. Removing these sub-
likely decrease the robustness of the resulting classification. economic ore samples from the sample set will increase the aver-
Fig. 7 shows that classifying the samples on a PCA score thresh- age copper grade from 0.45% to 0.60%.
old of 1 on the first PC distinguishes a group of 16 samples of which
15 samples are waste. This group contains 58% of all the waste
4. Discussion
samples in the sample set and has an average grade of 0.20% Cu.
Since this average copper grade is well below the economic cut-
PCA proved to be an effective method to display the systematic
off grade, it results that SWIR hyperspectral imagery in combina-
variation of the data that was extracted from mineralogy maps. It
tion with mineralogy mapping and PCA can be used to discriminate
makes it relatively easy to find samples that have similar miner-
16 M. Dalm et al. / Minerals Engineering 105 (2017) 10–18

Fig. 5. Result of mapping the wavelength position of the chlorite MgAOH absorption from selected SWIR hyperspectral images. This was only performed on pure chlorite
spectra (i.e. no absorption features by any other mineral).

Fig. 6. Sample scores and loadings of the PCA on data derived from mineralogy maps. Red dots represent ore samples. Blue dots represent waste samples. Black lines
represent the loadings. The loadings were all equally scaled to fit the figure and therefore do not represent actual loading values. C = chlorite; K = kaolinite; M = white mica;
N = nontronite; O = opaque minerals; T = tourmaline; AlAOH = average minimum position of the white mica AlAOH feature; cryst = average value of the white mica
crystallinity index. The green rectangle indicates a region where mainly waste samples occur. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

alogical composition and to identify the mineralogical differences of the absorption features in these SWIR spectra clearly indicate
between sample groups. The method can save a lot of time when the occurrence of two minerals that were not determined by
hyperspectral images of sample sets in the order of hundreds of XRD or petrography. This means that even if the identification of
samples or more need to be analyzed. kaolinite and/or nontronite was incorrect, the PCA results pre-
Petrography and XRD validated that white mica, chlorite and sented in Section 3.3 that are based on SWIR-active mineralogical
tourmaline were correctly identified from the SWIR hyperspectral differences between samples are still valid.
images. Validation of the specific white mica species was not car- The different mineral assemblages that resulted from mapping
ried out. Furthermore, kaolinite and nontronite were not deter- the mineralogy from the SWIR hyperspectral images are character-
mined with petrography or XRD. However, when these minerals istic for different zones of hydrothermal alteration within the por-
occur in any of the mineral maps they cover only a relatively small phyry system (Thompson et al., 1999; AusSpec, 2008).
surface area. It is therefore possible that they were missed in the Distinguishing low grade samples with SWIR hyperspectral ima-
thin sections on which petrography was performed or that their gery is therefore based on a relationship between copper mineral-
concentration was below the detection limit of XRD. No other min- ization and the formation of alteration minerals. This was expected
erals were identified with petrography or XRD that could be since this relationship was previously established when SWIR
responsible for the measured SWIR spectra that were classified point spectrometry was tested on the same sample set (Dalm
as kaolinite and nontronite. Furthermore, the wavelength position et al., 2014). As discussed by Dalm et al. (2014), the detection of
 M. Dalm et al. / Minerals Engineering 105 (2017) 10–18 17

Finally, SWIR hyperspectral imaging has the advantage that

the spectral information can be displayed in an image. This
provides data about the distribution of the SWIR-active mineralogy
which can be used to characterize differences in texture
including grain sizes and cross-cutting vein structures. Examples
of such textural characteristics can be seen in the different maps
of Fig. 4.
The variability in texture between the samples is relatively high
considering the size of the sample set. Because of this, many
characteristics that can be observed from the mineralogy maps
are often not shared by a sufficient amount of samples to produce
statistically meaningful classifications. An example is the vein
structure in the white mica crystallinity map of sample 2 in
Fig. 4d. Although such a characteristic provides information about
the hydrothermal alteration environment from which the samples
Fig. 7. Distribution of copper grades after classifying the samples on the PCA scores
of the first PC. A score of 1 was used as a threshold for the classification. Copper originate, it occurs on only two samples in the sample set. Both
grades were determined by XRF analysis. these samples have significant copper grades (0.62% & 0.96% Cu),
but it is uncertain if this relationship between vein structures
and copper grades is not coincidental.
alteration minerals with SWIR sensors can be used as indirect indi- The limited size of the sample set is likely the main reason that
cators for sensor-based sorting if the relationship between alter- SWIR hyperspectral imagery did not improve upon the indirect
ation mineralogy and copper mineralization persists throughout characterization of copper grades that was possible with SWIR
the entire ore deposit. Validation of the results on a sample set that point spectrometry. However, considering the advantages of SWIR
is representative of the full mineralogical variability of the deposit hyperspectral imaging, it is expected that the characterization of
is required to be able to conclude on the applicability of sensor- copper grade distribution will be improved when this technique
based sorting. The results of this study do show potential for using is tested on a sample set that is larger and more representative
SWIR sensors for sensor-based sorting of porphyry copper ores. of the full mineralogical variability of the deposit. An ideal sample
Comparing the results in Section 3.3 with those presented by set for further testing would be drill core samples.
Dalm et al. (2014) shows that using SWIR hyperspectral imaging Dalm et al. (2014) points out that the presence of dust and
did not improve upon the indirect characterization of copper grades water on the sample surface can limit the practical appliance of a
that was achieved using SWIR point spectrometry. However, SWIR SWIR sensor for sensor-based sorting. An additional limitation of
hyperspectral imaging has several advantages over SWIR point SWIR hyperspectral imaging sensors may be the time needed to
spectrometry regarding the mineralogical information that can be acquire the data at high spatial resolution. In order to reach suffi-
extracted from the data. These advantages are related to the capa- cient throughput on sensor-based sorting equipment, the SWIR
bility of SWIR hyperspectral imaging sensors to acquire SWIR data data needs to be acquired in a relatively short time period which
at a relatively high spatial resolution (i.e. small pixel size). places restrictions on the maximum spatial resolution that can
First of all, the surface area that is represented by a single SWIR be achieved. It is uncertain to what extent this limits the above dis-
spectrum is much smaller with most hyperspectral imagers than cussed advantages of SWIR hyperspectral imaging over point spec-
with commonly used point spectrometers. The spot size of SWIR trometry. For applications of SWIR hyperspectral imaging that are
point spectrometers is often in the order of several cm2 (e.g., ASD, not restricted by data acquisition time, such as drill core scanning,
2015), while SWIR hyperspectral imagers usually acquire spectra this is not an issue. Spatial resolution in SWIR hyperspectral imag-
on a pixel size below 1 mm2 (e.g., Corescan, 2015; Specim, 2015). ing is, however, an important parameter for any application since it
This relatively small pixel size allows minerals to be detected that governs the level of detail of the mineralogical information that
occur in relatively low quantities. Examples of such minerals in can be extracted from the data.
the sample set are nontronite, opaque minerals, and sometimes
kaolinite and/or tourmaline. The absorption features of these min-
erals often became completely obscured by those of the surround- 5. Conclusions
ing minerals when SWIR point spectra were measured.
The relatively small pixel size in hyperspectral imaging also The following conclusions were drawn from testing SWIR
provides spectra of pure minerals while point spectrometry fre- hyperspectral imaging on 43 porphyry copper ore samples:
quently yields mixed spectra with absorption features by two or
three minerals. The pure mineral spectra can be used to extract – Mineralogy maps produced from SWIR hyperspectral images
parameters that describe differences in the composition or crys- can be used as input variables in a PCA. This provides an effec-
tallinity of certain minerals. Although these parameters can also tive method to cluster samples that have similar mineralogical
be extracted from mixed mineral spectra, these will be less accu- composition and to identify the mineralogical differences
rate since they can be affected by the absorption features of the between sample groups.
other minerals. – SWIR hyperspectral imagery in combination with mineralogy
Another advantage of SWIR hyperspectral imaging is that a mapping and PCA could be used to discriminate 58% of the sam-
large number of spectra (±50,000 per sample) are acquired simul- ples with sub-economic copper grades. The average grade of
taneously. This allows the SWIR-active mineralogy at the surface of this group of samples is 0.20% Cu, which is well below an
samples to be quantified by classifying each spectrum on the min- assumed economic cut-off grade of 0.40% Cu.
eral and/or mineral characteristic that it represents and counting – SWIR hyperspectral imaging did not improve upon the indirect
the number of times that a certain classification occurs. This can characterization of copper grades that was achieved by Dalm
be used to distinguish between smaller compositional differences et al. (2014) using SWIR point spectrometry. The high textural
such as the occurrence of specific white mica species and the rela- variability between samples and limited size of the sample set
tive amount of chlorite or tourmaline that the samples contain. are considered to be key contributing factors.
18 M. Dalm et al. / Minerals Engineering 105 (2017) 10–18

– SWIR hyperspectral imaging has the capability to acquire SWIR Dalm, M., Buxton, M., van Ruitenbeek, J., Voncken, J., 2014. Application of near-
infrared spectroscopy to sensor based sorting of a porphyry copper ore. Miner.
data at a relatively high spatial resolution. Because of this, it has
Eng. 58, 7–16.
the following advantages over SWIR point spectrometry: Duke, E., 1994. Near infrared spectra of muscovite, tschermak substitution, and
s Provides more complete data of the mineral composition metamorphic reaction progress: implications for remote sensing. Geology 22,
of a sample including minerals that are present in relatively 621–624.
Eriksson, L., Johansson, E., Kettaneh-Wold, N., Trygg, J., Wikstrom, C., Wold, S., 2006.
low quantities. Chapter 3: PCA. In: Multi- and Megavariate Data Analysis. secon ed. Umetrics
s Enables the identification and extraction of pure mineral AB, Umea, pp. 43–70.
spectra that can be used to characterize mineral composition Girouard, G., Bannari, A., El Harti, A., Desrochers, A., 2004. Validated spectral angle
mapper algorithm for geologic mapping: comparative study between quickbird
and/or crystallinity. and landsat-TM. In: Proceedings of Geo-Imagery Bridging Continents Istanbul,
s Enables quantification of the SWIR-active mineralogical Turkey, pp. 599–604.
composition at the surface of a sample by counting the num- Hunt, G., 1977. Spectral signatures of particulate minerals in the visible and near
infrared. Geophysics 42 (3), 501–513.
ber pixels representing certain minerals and/or mineral Hunt, G., Salisbury, J., Lenhoff, C., 1971. Visible and near-infrared spectra of minerals
characteristics. and rocks: IV. Sulphides and sulphates. Modern Geol. 3, 1–14.
s Enables the characterization and evaluation of the texture ITT visual information studios, 2008. ENVI 4.5 (commercial computer software).
Jolliffe, I., 2002. Principle Component Analysis. Springer-Verlag, New York.
of samples apparent from mineral maps. Kruse, F., Lefkoff, A., Boardman, J., Heidebrecht, K., Shapiro, A., Barloon, P., Goetz, A.,
Considering the advantages of SWIR hyperspectral imaging, it is 1993. The spectral image processing system (SIPS) – interactive visualization
expected that the characterization of copper grade distribution and and analysis of imaging spectrometer data. Remote Sens. Environ. 44, 145–163.
Kruse, F., 2012. Mapping surface mineralogy using imaging spectrometry.
ore-waste designations will be improved when this technique is
Geomorphology 137, 41–56.
tested on a sample set that is larger and more representative of Lowell, J., Guilbert, J., 1970. Lateral and vertical alteration-mineralization zoning in
the full mineralogical variability of the deposit. porphyry ore deposits. Econ. Geol. 65 (4), 373–408.
Martin-Fernandez, J., Barcelo-Vidal, C., Pawlowsky-Glahn, V., 2003. Dealing with
zeros and missing values in compositional data sets using nonparametric
imputation. Math. Geol. 35 (3), 253–278.
Mc Leod, R., Gabell, A., Green, A., Gardavsky, V., 1987. Chlorite infrared spectral data
References as proximity indicators of volcanogenic massive sulphide mineralization. In:
Proceedings of the Pacific Rim Congress Gold Coast, Australia, pp. 321–324.
Aitchison, J., 1986. The Statistical Analysis of Compositional Data. Chapman and Pirajno, F., 1992. Hydrothermal Mineral Deposits, Principles and Fundamental
Hall, London. Concepts for the Exploration Geologist. Springer-Verlag, Berlin.
ASD, 2015. ASD products. (Online) Available at: <http://www.asdi.com/products> Post, J., Noble, P., 1993. The near-infrared combination band frequencies of
(accessed 2015). dioctahedral smectites, micas, and illites. Clays Clay Miner. 41 (6), 639–644.
AusSpec, 2008. G-MEX Spectral Interpretation Field Manual. AusSpec International Rieder, M. et al., 1998. Nomenclature of the micas. Can. Mineral. 36, 41–48.
Ltd.. Rowan, L., Mars, J., Simpson, C., 2005. Lithologic mapping of the mordor, NT,
Bakker, W., 2012. Hyperspectral Python (HypPy). (Online) Available at: <https:// Australia ultramafic complex by using advanced spaceborne thermal emission
www.itc.nl/personal/bakker/hyppy.html> (accessed 2012). and reflection radiometer (ASTER). Remote Sens. Environ. 99, 105–126.
Baldridge, A., Hook, S., Grove, C., Rivera, G., 2009. The ASTER spectral library version Sabins, F., 1999. Remote sensing for mineral exploration. Ore Geol. Rev. 14, 157–
2.0. Remote Sens. Environ. 113, 711–715. 183.
Buxton, M., Benndorf, J., 2013. The use of sensor derived data in optimization along Salter, J., Wyatt, N., 1991. Sorting in the minerals industry: past, present and future.
the mine-value-chain. In: Proceedings of the 15th International ISM Congress Miner. Eng. 4, 779–796.
Aachen, Germany, pp. 324–336. Sillitoe, R., 2010. Porphyry copper systems. Econ. Geol. 105 (1), 3–41.
Clark, R., King, T., Klejwa, M., Swayze, A., 1990. High spectral resolution reflectance Sivamohan, R., Forssberg, E., 1991. Electronic sorting and other preconcentration
spectroscopy of minerals. J. Geophys. Res. 95 (B8), 12653–12680. methods. Miner. Eng. 4, 797–814.
Clark, R., 1999. Spectroscopy of rocks and minerals and principles of spectroscopy. Specim, 2015. Specim products for geological applications. (Online) Available at:
In: Rencz, A. (Ed.), Manual of Remote Sensing, Remote Sensing for the Earth <http://www.specim.fi/index.php/products/geology> (accessed 2015).
Sciences, vol. 3. John Wiley and sons, New York, pp. 3–58. Thompson, A., Hauff, P., Robitaille, A., 1999. Alteration mapping in exploration:
Clark, R., Swayze, G., Wise, R., Livo, E., Hoefen, T., Kokaly, R. & Sutley, S., 2007. USGS application of short-wave infrared (SWIR) spectroscopy. Soc. Econ. Geol. Newsl.
spectral library splib06a: U.S. Geological Survey, Digital Data Series 321. 39, 15–27.
(Online) Available at: <http://speclab.cr.usgs.gov/spectral-lib.html> (accessed Wills, B., Finch, J., 2016. Chapter 14 – Sensor-based ore sorting. In: Wills’ Mineral
2014). Processing Technology. eighth ed. Butterworth-Heinemann, Boston, pp. 409–
Corescan, 2015. The corescan system. (Online) Available at: http:// 416.
www.corescan.com.au/services/the-corescan-system (accessed 2015). Wold, S., Esbensen, K., Geladi, P., 1987. Principle component analysis. Chemomet.
Crowley, J., Vergo, N., 1988. Near-infrared reflectance spectra of mixtures of kaolin- Intell. Lab. Syst. 2, 37–52.
group minerals: use in clay mineral studies. Clays Clay Miner. 36 (4), 310–316.