©2024 Society of Economic Geologists, Inc.

Economic Geology, v. XXX, no. XX, pp. X–X

Mapping Ni-Cu-Platinum Group Element-Hosting, Small-Sized,

Mafic-Ultramafic Rocks Using WorldView-3 Images and a Spatial-Spectral
Transformer Deep Learning Method
Chuntao Yin,1,* Yaqian Long,2,* Lei Liu,1,3,† Yasir Shaheen Khalil,1,4 and Songxing Ye2
1School
of Earth Science and Resources, Chang’an University, Xi’an 710054, China
2College
of Computer Science and Software Engineering, Shenzhen University, Shenzhen 518000, China
3Key Laboratory of Strategic Mineral Resources of the Upper Yellow River, Ministry of Natural Resources,

Lanzhou 730046, China

4Geological Survey of Pakistan, Peshawar 25100, Pakistan

Abstract
Mafic-ultramafic intrusive complexes possess a considerable capacity for hosting Ni-Cu-platinum group ele-
ment (PGE) sulfide deposits. However, the mapping of small outcrops over large areas by field surveys is time-
consuming. In this study, WorldView-3 (WV-3) data with moderate spectral and very high spatial resolution
were employed for mapping mafic-ultramafic units using spectral indices and the spatial-spectral transformer
(SSTF) deep learning method in the Luotuoshan area of Beishan, Gansu Province, China. Based on represen-
tative reflectance signatures extracted from imagery of known locations, false-color composites of three-band
ratios and a newly proposed shortwave infrared (SWIR) spectral index provided reasonable delineation of mafic-
ultramafic rocks. The SSTF method facilitated mapping the occurrence of small mafic-ultramafic outcrops and
defining much clearer boundaries, particularly for tiny units at meter scale. Moreover, the SSTF method is not
sensitive to the occurrence of carbonate lenses that may affect the reflectance of outcrops. Field investigation
and laboratory sample analyses confirmed the occurrence of mafic and ultramafic rocks with substantial metal-
lic mineral potential in this area. Seven prospects were confirmed to be related to mafic-ultramafic intrusions
during field validation, four of which contained metallic minerals such as chalcopyrite, pentlandite, pyrite,
and chromite in the samples observed by scanning electron microscopy and energy dispersive spectrometry.
This study proves that the spectral indices and SSTF deep learning method applied on WV-3 multispectral
data are useful for discriminating small-sized mafic-ultramafic intrusive rocks (<100 m) for prospecting of
local mineralization.

Introduction only 1.3 and 0.07 km2, respectively (Tang et al., 2006; Gao et
Magmatic nickel (Ni), copper (Cu), and platinum group el- al., 2012; Lu et al., 2019). The ore-bearing intrusions typically
ement (PGE) sulfide deposits typically occur in mafic-ultra- occur in clusters of veins, dikes, and sills, with lithology docu-
mafic suites. These deposits contribute significantly to the mented as peridotite, lherzolite, gabbro, and pyroxenite (Tang
global mining industry with approximately 56% of Ni, 5.5% et al., 2006, 2012). It is worth noting that the small-sized ore-
of Cu, and 96% of PGEs (Naldrett, 1997; Mungall and Nal- bearing mafic-ultramafic outcrops may be overlooked during
drett, 2008; Mudd and Jowitti, 2014; Barnes et al., 2017). The field geologic surveys, as they may be much smaller than the
association between magmatic sulfide deposits and mafic or mappable unit at a given scale during regional geologic map-
ultramafic rocks is well documented (Naldrett, 1997). The ping. Therefore, mapping the mafic-ultramafic rock outcrops
formation of magmatic Ni-Cu-PGE sulfide deposits is attrib- of all sizes is essential for the exploration of Cu, Ni, and PGE
uted to the segregation and concentration of liquid sulfide mineralization (Barnes and Lightfoot, 2005; Liu et al., 2014;
droplets from mafic or ultramafic magma and the partitioning Rogge et al., 2014).
of chalcophile elements into these from a silicate melt (Chai Multispectral and hyperspectral satellites provide feasible
and Naldrett, 1992; Arndt et al., 2005; Lightfoot et al., 2012a). alternatives to traditional field surveys for lithologic map-
Therefore, mafic-ultramafic rocks are considered a key indica- ping, providing significant advantages in terms of efficiency
tor for the possible presence of Ni-Cu-PGE sulfide deposits and cost-effectiveness (Rowan et al., 1977; Goetz et al., 1983;
(Naldrett, 1997; Maier, 2005; Barnes et al., 2016; Graham et Kruse et al., 2003; Ninomiya et al., 2005; Nair and Mathew,
al., 2017). 2012). Satellite images from the Thematic Mapper (TM),
The majority of ore-bearing mafic-ultramafic intrusions are the Enhanced Thematic Mapper (ETM), and the Advanced
characterized by areas that are less than a few square kilo- Spaceborne Thermal Emission and Reflection Radiometer
meters, with notable examples including the Norilsk-Talnakh (ASTER) have been widely used in lithological mapping and
giant Cu-Ni-PGE deposit (Russia) and Voisey’s Bay giant Ni- mineral delineation (Di Tommaso and Rubinstein, 2007; Pour
Cu-Co deposits (Canada), which are only 3 to 6 km2 in out- and Hashim, 2012; Zadeh et al., 2014; Pour et al., 2018). Spe-
crop area (Amelin et al., 1999; Naldrett, 1999; Lightfoot et cifically, TM, ASTER, and simulated EnMAP data have been
al., 2012b; Porter, 2016). Similarly, the outcrop areas of the successfully used to discriminate ophiolite complexes and
Jinchuan and Kalatongke giant Ni-Cu deposits (China) are mafic-ultramafic rock units (Abrams et al., 1988; Rowan and
Mars, 2003; Ninomiya et al., 2005; Rogge et al., 2014; Qasim
†Corresponding author: e-mail, liul@chd.edu.cn et al., 2022). However, the low spatial resolution of these sen-
*Chuntao Yin and Yaqian Long contributed equally to this work.
ISSN 0361-0128; doi:10.5382/econgeo.5056; 16 p. 1 Submitted: April 6, 2023/Accepted: September 12, 2023

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 2 YIN ET AL.

sors (30 m) presents a limitation to the accurate mapping of Study Area and Data Collection
small mafic-ultramafic rock outcrops.
WorldView-3 (WV-3) is the solitary commercially available Work area and geology
imaging satellite that furnishes high spatial resolution multi- Beishan is located in northwest China and spans Xinjiang,
spectral images over both visible near-infrared (VNIR) and Gansu, and Inner Mongolia (Fig. 1a). This semiarid area is
short-wave infrared (SWIR) spectral regions, which contain suitable for the growth of sparse vegetation such as grasses
lithological and mineralogical information. WV-3 has eight and shrubs. The diverse geology, tectonics, and metallogenic
VNIR bands with 1.24-m spatial resolution and eight SWIR environment make the Beishan area one of the largest min-
bands with spatial resolution of 3.7 m (Mars, 2018). In addi- eral resource belts in China. More than 90 ore deposits (pri-
tion, it also has a panchromatic band with 0.3-m resolution. marily gold, copper-nickel, and lead-zinc) have been discov-
This high spatial resolution is particularly valuable in the ered in this belt. More than 20 deposits are medium to large
accurate identification of small rock units that might not be sized (Fig. 1b) (Yang et al., 2006). Mafic-ultramafic intru-
easily discernible with other satellite images. WV-3 has been sions or complexes host Ni-Cu-Co deposits that are preva-
used to accurately identify small rock units, ferric and ferrous lent throughout the Beishan area (Yang et al., 2012). Over
minerals, hydroxyl group minerals, and carbonates (Mars, 70 such complexes have been mapped in the Beishan belt in
2018; Ninomiya and Fu, 2019). Kruse and Perry (2013) ef- the Gansu Province (Yang et al., 2006). A recent discovery
fectively mapped various minerals including kaolinite, alu- is the Pobei superlarge Ni-Cu-Co deposit hosted by several
nite, buddingtonite, muscovite, calcite, and hydrothermal co­genetic mafic-ultramafic intrusions in the Beishan belt
silica at Cuprite, Nevada, using simulated WV-3 SWIR imag- (Yang et al., 2014; Ye et al., 2017). Some other well-known
ery. A subsequent study by Kruse et al. (2015) confirmed the ore deposits in the belt include the Hongshishan chromium
mapping potential of WV-3 data using on-orbit WV-3 SWIR ore deposit and the Heishan Ni-Cu ore deposit (Yang et al.,
data at the same site. Mars (2018) successfully mapped dif- 2006). Therefore, the Beishan belt is recognized as an im-
ferent lithologic units at Mountain Pass, California, including portant mineral resource zone related to mafic-ultramafic
muscovite-rich granitic gneiss, chlorite-epidote–rich migma- intrusions. There exist some known mafic-ultramafic intru-
tite, schist, gneiss, limestone, and dolostone using WV-3 data sions around the Luotuoshan area in southwestern Beishan,
with band ratios, false-color composites, and logical operator Gansu Province (Fig. 1b). However, limited field surveys
algorithms. Sekandari et al. (2020) demonstrated that VNIR have been conducted in this region to date. It is imperative
to quickly map these mafic-ultramafic intrusions and assess
bands of WV-3 were useful in discriminating ferrous and fer-
their mineralization potential with the aid of high spatial
ric iron-bearing minerals in the Pb-Zn mineralization district
resolution imagery.
of Kerman-Kashmar, Iran. Similarly, more recent studies have
The study area (Fig. 1c) predominantly comprises five Pa-
substantiated the potential of the VNIR and SWIR bands of
leozoic units, namely the Permian Shuangbaotang Formation
WV-3 for lithologic and alteration mineral mapping (Sun et
(P1s), the Carboniferous Shibanshan Formation (C2s), the
al., 2017; Pour et al., 2019; Son et al., 2021; Zhao et al., 2021). Carboniferous Hongliuyuan Formation (C1h), the Devonian
Nevertheless, no study has examined the use of WV-3 for Dundunshan Group (D3dn), and the Ordovician Huaniushan
mapping mafic-ultramafic rocks, which present a challenge Group (O1-2hnc). The Precambrian crystalline basement,
due to their low reflectance and weak characteristic absorp- which mainly consists of metamorphic rocks, is sporadically
tions in the VNIR-SWIR region. exposed in the study area. Unconformable overlying strata
Previous studies have proposed several approaches to de- ranging from the Cambrian to Permian periods are com-
tect mafic-ultramafic rocks using hyperspectral data, such as monly lower-grade metasedimentary rocks dominated by
band ratios, neural network classifiers, and linear spectral un- limestones, slates, phyllites, sandstones, and quartz schists
mixing algorithms with different data sets (Rogge et al., 2010; (Ruan et al., 2021). Extensive tectonics and magmatic activity
Leverington and Moon, 2012; Rogge et al., 2014). The recent occurred during the Carboniferous and Permian periods (Xia
development of deep learning techniques has demonstrated et al., 2013), resulting in the emplacement of mafic-ultramafic
significant potential in lithologic classification from multispec- complexes with similar rock types and early Permian ages (Su
tral remote sensing images. Owing to its robust modeling abil- et al., 2012; Liu et al., 2017; Ruan et al., 2020). The complexes
ity of combining spectral and spatial information, the deep discovered so far are primarily distributed in the western part
learning method can deliver more accurate mapping results of the Beishan region and intrude the Proterozoic and Car-
with continuous spatial patterns and well-defined edges. Ac- boniferous strata (Jiang et al., 2006; Su et al., 2009, 2011).
cordingly, this study aims to investigate the transformer meth- They are controlled by a complex fracture system with north-
od and evaluate its performance in classifying rock types for east, northwest, and nearly east-west trends. The orientation
mapping occurrence of small mafic-ultramafic outcrops. The of the northeast and east-west fractures is nearly identical to
main objectives are (1) to differentiate mafic-ultramafic rocks tectonic lineations in the region, which are considered to be
(e.g., gabbro, pyroxenite, and peridotite) from other bedrocks part of the regional faults.
(e.g., carbonates, granite, and sandstone), which would assist
in mineral exploration over a large area (100 km2) in Beis- Data and measurements
han, Northwest China, (2) to assess the ability of WV-3 data WorldView-3 multispectral data: Two WV-3 data sets (stan-
in mapping small-sized outcrops with low reflectance, and (3) dard-level 2A product) with no cloud or snow and very little
to investigate the feasibility of the deep learning method in vegetation cover (<1%) were acquired for the study area on
lithologic mapping with WV-3 high spatial resolution imagery. April 30, 2019. The data sets contain 16 bands ranging from

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 DEEP LEARNING APPLIED TO MAFIC-ULTRAMAFIC ROCKS 3

90°0'E 120°0'E 95°0'E 96°0'E 97°0'E

R U S S I A

MONGOLIA
± < Huangshan
<
<
Hongshishan

_
^
- <
< <
a n
g
.
Jijitaizi

42°0'N
j i xin g x ia
n Xing
30°0'N

C H I N A
X i < < G a n s u
Mazongshan
o
INDIA Yingmaotu <
Yushishan Niuquanzi

Dashantou <
Heishan .
>
10°0'N

shan

41°0'N
Huitong Daqishan
0 25
50
_
^ Study Area Km
0 1,000
National Boundaries Km Study
Fault Provincial Mafic Deposit Type
Area Boundary Rocks . Fe >V < Cu-Ni
95°15'E 95°20'E

± γ42c O1-2hnc γ51 γ51 Q4al-pl

O1-2hnc
41°8'N

2b
δ4 D3dn
D3dn D3dn δο33
2c
γ4
γ42c ν42a
O1-2hnc βμ42a
πγ42c O1-2hnc O1-2hnc
γβ51
P1s1 O1-2hnc
δ42b Q4al-pl
γβ51 ν42a
O1-2hnc
γ42c ν42a
41°6'N

D3dn ν42a

O1-2hnc D3dn
ν42a P1s2 P1s2 P1s3
Q4 al-pl C 1 h
0 1 2
πγ42c γβ51 Q3al-pl δ33 C2s Km
Q4al-plAlluvial-diluvial
Q4al-pl Alluvial-diluvial Q al-pl Alluvial-diluvial
Q33al-pl Alluvial-diluvial P1sP s3 Hard
3 Hardy Sandy
sandy P Ps2 2 Siltstone
s with
Siltstone with PP s s Sandstone
1 1 Sandstone
C Cs s Sandstone
Sandstone
2 Limestone
with Fault Fault
Sand
sandand
and Gravel
gravel Gravel
gravel Layer
layer
1 Sandstone
sandstone 1 1 Limestone
limestone
11
andglutenite
and Glutenite 2
with limestone
Sandstone with Felsic and Mafic c Schist with Marble Brick Red 1 Brick Red 2c Soil Yellow, Flesh Red
mafic OOhn chn
1
C1Ch1hLimestone
Sandstone Lens
with DD33dn Felsic and Rock Schist with marble γ
1 Brick red 1 γβBrick red πγyellow,
2c Soil flesh Porphyry
4 Granite
dn Volcanic 1-2
1-2 and Quartzite γ5 5 Granite γβ5 5 Biotite Granite
πγ4
limestone lens volcanic rock and quartzite granite biotite granite red granite porphyry
2c Soil Yellow, Flesh 2b Gray and Dark 2a Dark Green 2a Dark Green 3 Light Gray 3 Dark Gray Granite Vein
γ4 Red Granite δ4 Gray Diorite ν4 Gabbro βμ4 2aDiabase δο3 3 Quartz Diorite δ3 3 Diorite
γ 2c Soil yellow,
4
2b Gray and dark
δ4 ν 2a Dark green
4 βμ Dark green
4 δο Light gray 3 δ Dark gray
3
Granite
flesh red granite gray diorite gabbro diabase quartz diorite diorite vein

Fig. 1. Study area and geologic map. (a) Location of study area. (b) Regional map of known deposits in Beishan (Yang et al.,
2006). (c) 1:50,000 geologic map of the study area (modified based on Gansu Bureau of Geology and Mineral Resources,
unpub. maps, 1985, 2006).

VNIR to SWIR. The band positions are presented in Figure The VNIR spectra of rock samples were acquired using the
2. The SWIR bands were provided with 7.5-m spatial resolu- SR-3500 portable spectrometer from Spectral Evolution Inc.,
tion, whereas the VNIR bands were provided at 2.0-m spatial Lawrence, Massachusetts, USA. The spectrometer acquired
resolution. The SWIR data were spatially resampled to 2.0 m high-resolution spectra (3 nm from 350 to 1,000 nm; 8 nm
and coregistered to the VNIR data. The high spatial and mod- from 1,000 to 1,900 nm; and 6 nm from 1,900 to 2,500 nm)
erate spectral resolution ensures the accurate extraction and within the wavelength range of 350 to 2,500 nm. Spectral
identification of small outcrops at a scale of a few meters. measurements were conducted on weathered surfaces and
Samples and laboratory measurements: In the study area, fresh surfaces separately, with 1 to 5 spot spectra collected per
a total of 37 hand samples were collected, including gabbro surface type. To ensure consistent illumination conditions, a
and pyroxenite. All samples had surface areas greater than contact probe (field of view of 25°) with an internal illumina-
10 × 10 cm, with at least one weathered surface and one fresh tion source was used. Radiance spectra were calibrated and
surface for each sample. converted to reflectance using a white panel made of BaSO4.

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 4 YIN ET AL.

datum and coregistered with preexisting 1:10,000-scale digital

topographic maps of the study area.
Spectral characteristics of rock units
Spectra from the Johns Hopkins University (JHU) spectral
library were employed as preliminary reference (Fig. 2) to
distinguish various types of mafic-ultramafic rocks from back-
ground. According to the local geologic settings, five mafic-
ultramafic (diabase, basalt, gabbro, dunite, and picrite) and
four nonmafic rock types (quartz monzonite, granite, grano-
diorite, and diorite) were selected for identification of poten-
tial spectral features. Mafic rocks, which contain iron oxide,
pyroxene, and amphibole, exhibit absorption features from
ferrous and ferric ions in the vicinity of 0.7 and 1.0 μm (Hunt,
1977; Gupta, 2003). Some mafic rocks, such as pyroxenite and
gabbro, present Fe-OH and Mg-OH absorption features of
biotite and hornblende at 2.32 and 2.38 μm (Rowan et al.,
2005). Diorite has a broad absorption feature from 0.7 to
1.2 μm and another prominent feature near 2.33 μm. Rocks
with hornblende and epidote show absorption features at
2.25, 2.32, and 2.39 μm. Additionally, the weathered surfaces
of certain rock types (e.g., granite and quartz monzonite) ex-
hibit diagnostic absorption features of Al-OH at 2.20 μm and
Fe3+ features at 0.49 and 0.87 μm (Hunt and Ashley, 1979;
Gaffey, 1986).
Even though WV-3 data only has 16 bands from VNIR to
SWIR, it has been demonstrated to be effective in detecting
minerals and mapping lithological units over areas at meter
scale (Mars, 2018). WV-3 data have been utilized for the iden-
tification of typical alteration minerals including hydroxyl-
bearing, iron oxide-bearing, and carbonate minerals (Kruse
and Perry, 2013; Sun et al., 2017). The VNIR bands of WV-3
have proven to be useful for mapping lithological units by dif-
ferentiating between ferrous and ferric iron-bearing miner-
als (Ye et al., 2017; Salehi and Tangestani, 2018; Sekandari et
Fig. 2. Johns Hopkins University library spectra (solid lines) of typical mag-
matic rocks and their resampled spectra (dashed lines) to WorldView-3 band-
al., 2020). Specifically, prominent spectral absorption features
passes. SWIR = short-wave infrared, VNIR = visible near infrared. in ferrous iron minerals such as epidote and chlorite can be
identified by the obvious decrease in reflectance from VNIR
band 3 to SWIR band 4. Ferric iron minerals show a decrease
An average of 10 consecutive measurements was made for from VNIR band 2 to VNIR band 4 and from VNIR band 7
each saved spectrum. to SWIR band 1. The SWIR bands can be used to detect the
In order to characterize the mineralogy of the rock samples, absorption features of Al-OH (2.20 μm), Fe-OH and Mg-OH
a total of 19 thin sections and six polished sections were pre- (2.30+ μm), and CO32– (2.16, 2.32–2.35 μm) found in clays, ep-
pared and investigated with a Nikon Labophot2 polarizing- idote, chlorite, and calcite/dolomite (Clark et al., 1993; Mars,
light microscope. Additionally, the polished sections were 2018). This study aims to identify appropriate VNIR or SWIR
sent for scanning electron microscopy (SEM) and energy bands and spectral indices for mapping mafic-ultramafic rocks
dispersive spectrometry (EDS) analysis using an FEI Quanta in the work area.
650 FEG SEM instrument at 20 kV to identify metallic min-
erals such as chalcopyrite, pentlandite, pyrite, and chromite. Mapping mafic-ultramafic units
Spectral indices: Spectral indices that could enhance the spec-
Method tral features of minerals (e.g., iron oxides and hydroxyl-bear-
ing minerals) were used to highlight the spatial distribution
Preprocessing of WV-3 images of mafic and ultramafic rocks. The bands utilized for calculat-
The WV-3 radiance data were converted to surface reflec- ing spectral indices were based on previous studies (Rowan et
tance using the Fast Line-of-Sight Atmospheric Analysis of al., 2005; Amer et al., 2010; Nair and Mathew, 2012), as well
Spectral Hypercubes (FLAASH) algorithm (Berk et al., 1998) as on spectroscopic analysis of minerals available in the JHU
with parameters of 617-km sensor altitude, 1,900-m ground spectral library.
elevation, subarctic summer atmospheric model, rural aerosol Based on an existing geologic map (Fig. 1c), this study se-
model, 40-km initial visibility, and no spectral polishing. The lected 11 dominant rock types from which representative
reflectance data were subsequently georeferenced to a local spectra were extracted for image spectral analysis. The spec-

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 DEEP LEARNING APPLIED TO MAFIC-ULTRAMAFIC ROCKS 5

tral signatures of the selected rock types displayed significant

differences across the 16 bands of the WV-3 (Fig. 3). Based
on the spectral features illustrated in Figure 2, three band
ratios (SWIR bands 6/8, SWIR bands 5/3, and VNIR bands
3/2) were calculated and then merged into a false-color com-
posite image to enhance the differences among rock types.
The three band ratios were used to identify the three compo-
sitional indicators from local rock types: (1) Fe-OH and Mg-
OH in mafic rocks such as pyroxenite and gabbro, (2) ferrous
iron in diorite and peridotite, and (3) ferric iron in sandstone,
schist, and granite.
To discriminate the mafic rocks (e.g., gabbro, diabase, and
pyroxenite) from nonmafic rocks (e.g., diorite and granite),
spectral index (SWIR band 4 + SWIR band 6)/(2 × SWIR
band 5) was proposed. This index was expected to enhance
the spectral difference between the mafic and nonmafic
rocks, as the spectral slope from SWIR band 4 to SWIR band
6 is steeper for nonmafic rocks compared to the relatively flat
slope observed in mafic rocks (Fig. 3).
Classification of rock units using deep learning: A spatial-
spectral transformer (SSTF) was proposed for the final litho-
logic classification task (Fig. 4). The SSTF is a type of deep
learning method called “transformer” that has demonstrated
robust performance in various applications (Vaswani et al.,
2017; Hong et al., 2022; Yu et al., 2022; Zhang et al., 2023).
The transformer method originated from the natural language
processing (NLP) and computer vision (CV) image analyses
and has powerful modeling ability. The transformer network
uses encoder and decoder architecture and relies solely on
a self-attention mechanism that differs greatly from convo-
lutional and recurrence neural networks. It treats the image
as sequence data and builds encoder and decoder layers with
the same architecture. The encoder transforms the input data
Fig. 3. WorldView-3 image spectra from the study area. SWIR = short-wave
to sequence representations with position information, and infrared, VNIR = visible near infrared.
then the decoder converts the transformed representation to
an output sequence. The encoder and decoder are connected
by an attention mechanism that automatically adjusts the rela- (CAF) (which is suitable for multispectral data) (Fig. 4).
tionship between different locations and focuses on the areas To make the network more adaptive to the high-resolution
that require attention. The essence of the attention mecha- WV-3 data, we modified the transformer by adding a block
nism is to adapt the weights for classifying pixels, through mechanism that improves the extraction of spectral and spa-
which the network can form a feature map for each class. tial information (Fig. 4). The SSTF adopts two blocks of GSE
To represent the long-range dependence between input and CAF proposed by Hong et al. (2022). While the ordinary
and output, an attention matrix was computed by a function transformer method often applies a bandwise classification to
of query (Q), keys (K), and values (V) in the form of vectors a spectral image, the SSTF uses the block of GSE that can
as below: learn pixelwise representation and address the spectral corre-
Attention(Q, K, V) = softmax
QKT
V, ( ) (1)
√dk
lation between adjacent bands (Hong et al., 2022). Therefore,
this strategy is more suitable for capturing spectral features
such as spectral peaks and troughs of the WV-3 data. The
where softmax is a normalized function that transforms a vec- block of CAF aims to enhance the connectivity across long-
tor into values in the range of 0 to 1, and dk represents the range network layers. Additionally, SSTF includes a feature
dimension of keys and is a scaling factor related to K. Fur- extraction block of a three-dimensional Gabor filter (Shen and
thermore, the transformer employs a multihead attention Jia, 2011) to detect both spectral and spatial variance of the
mechanism that allows parallel processing and speeds up the data. In this Gabor block, varying frequencies (Rivard et al.,
classification. 2008) and directions are defined. The Gabor filtering based
In this study, the WV-3 data were fed into the neural net- on Gaussian kernel functions assists us in capturing more rep-
works in two ways. One is a spatial transformation using 3-D resentative information (e.g., well-defined texture) from the
Gabor filtering based on a 7 × 7 spatial window, and the other image. The merit of Gabor features improves the classifica-
is per-pixel spectral embedding using a visual transformer tion performance that has been demonstrated in a series of
(ViT) network modified by two modules of Groupwise Spec- studies (Hu et al., 2018; Huang et al., 2022). Given the task of
tral Embedding (GSE) and Cross-Layer Adaptive Fusion this study, we used a framework for multiclass classification.

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 6 YIN ET AL.

Fig. 4. Flowchart of the spatial-spectral transformer method.

The classifier was conducted using the reflectance data with- tral properties of these samples were assessed by measuring
out any spectral transformation. the reflectance of both fresh and weathered sample surfac-
The number of pixels selected for each training class (rock es. A total of 37 rock samples, including gabbro, pyroxenite,
type) is listed in Table 1. While the original geologic map fea- peridotite, and olivine gabbro, were collected from 22 sites.
tured 18 rock types (Fig. 1c), similar rock types were merged ­Twenty-five (out of 37) samples containing sulfides were se-
manually to address only compositional variations. Thus, a to- lected for further analysis based on rock types and visual as-
tal of 11 classes were selected as regions of interest. sessment of hand samples. Thin sections or polished sections
The classification accuracy of each rock type was calcu- were prepared for petrographic analysis.
lated based on cross-validation, and a confusion matrix was
constructed to describe the false alarms among different rock Results
types. For the accuracy evaluation, pixels within each of the
Detection of mafic-ultramafic outcrops from
labeled regions of interest were randomly split into a training
spectral indices
set (80%) and a testing set (20%).
False-color composites of the three band ratios (red = SWIR
Field validation bands 6/8 for Fe-OH, Mg-OH; green = SWIR bands 5/3 for
The accuracies of the two mapping methods were validated Fe2+; blue = VNIR bands 3/2 for Fe3+) were constructed (Fig.
by a field campaign and laboratory analyses of samples. Spec- 5b). Mafic-ultramafic units such as gabbro and pyroxenite are
represented by red to dark-red colors owing to the presence
Table 1. Training Data of the 11 Lithologic Units of the Study Area of pyroxene, chlorite, and epidote. Diabase is indicated by
light- to dark-green colors owing to the lack of a strong Fe2+
Class name Color Pixels absorption that results in a relatively high SWIR band 5/3 val-
Brick-red granite Red 14,773 ue. Additionally, granite, which exhibits a strong Fe3+ absorp-
Dark-gray diorite Green 2,783
Diabase Blue 21,247
tion feature, is shown as blue owing to its high VNIR band 3/2
Gabbro Yellow 11,240 value. Compared to the true-color image (Fig. 5a), the false-
Gray diorite Cyan 5,952 color composite image of band ratios significantly enhanced
Sand and gravel Aquamarine 13,573 rock types and highlighted the occurrence of mafic-ultramafic
Sandstone Magenta 7,199 outcrops, particularly in the northeast part of the study area.
Schist Maroon 18,285
Siltstone and limestone Sea green 7,350 Furthermore, small gabbro outcrops that were not visible on
Flesh-red granite Purple 27,844 the natural color image and not mapped on the traditional
Granite porphyry Coral 10,301 geologic map were differentiated from large diabase sills in
the false-color band ratio composite image (Fig. 5b).

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 DEEP LEARNING APPLIED TO MAFIC-ULTRAMAFIC ROCKS 7

Fig. 5. A subset of 1:50,000 geologic boundaries overlaid on (a) WorldView-3 true-color composite image (VNIR band 5 =
red, VNIR band 3 = green, VNIR band 2 = blue), (b) band ratios composite (SWIR band 6/band 8 = red, SWIR band 5/band
3 = green, VNIR band 3/band 2 = blue), and (c) band index image [(SWIR band 4 + SWIR band 6)/(2 × SWIR band 5)]. See
legend in Figure 1 for lithology abbreviations. SWIR = short-wave infrared, VNIR = visible near infrared.

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 8 YIN ET AL.

Results of the SWIR spectral index (SWIR band 4 + SWIR od. Figure 6b shows only the distribution of mafic-ultramafic
band 6)/(2 × SWIR band 5) conveyed similar information, rocks (diabase and gabbro) overlaid on the true-color WV-3
except the mafic-ultramafic rocks were represented as bright image. The result shows more precise lithological boundary
areas on the image (Fig. 5c), whereas the other rock types, delineation and the detection of small outcrops at a scale of
including granite and diabase, demonstrated relatively low in- meters. Compared with the color composite of band ratios in
dex values. This spectral index provides a convenient tool for Figure 5b, the deep learning classification result is able to map
quickly mapping the occurrence of mafic-ultramafic outcrops. the mafic-ultramafic intrusions with clearly defined boundar-
Nonetheless, false positives occur when lenses of carbonate ies. Compared with the 1:50,000 geologic map in Figure 1c,
are present. These are mainly distributed in the southeast part the SSTF generated a very similar mapping of granite, sand-
of the study area (yellow arrows in Fig. 5c). In contrast, the stone, and diorite. For units that were labeled as gabbro, dia-
false-color composite image (Fig. 5b) shows that the lenses base, schist, siltstone, and limestone on the geologic map, the
are not mafic-ultramafic rocks. SSTF classification showed much more intraclass variability,
thereby enabling the discrimination of compositional details.
Lithologic mapping using the SSTF classification method Large intrusions (D8 and D9 on Fig. 6), labeled as gabbro on
Figure 6a shows the color-coded classification result of 11 rock the 1:50,000 geologic map, were adequately mapped in the
types that was obtained using the SSTF deep learning meth- classification result, providing additional support for the valid-

Fig. 6. Spatial-spectral transformer (SSTF) mapping results. (a) Lithological classification results from the SSTF.
(b) Mafic and ultramafic outcrops mapped by the SSTF and overlaid on a WorldView-3 true-color image. Arrows indicate
field survey sites.

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 DEEP LEARNING APPLIED TO MAFIC-ULTRAMAFIC ROCKS 9

ity of the SSTF approach. More importantly, the SSTF clas- Table 2. Accuracy for the Lithological Map Produced
sification revealed many small mafic-ultramafic intrusions that by the Transformer Algorithm
were not previously shown on the geologic map. For instance, Class name Color Accuracy (%)
large areas of diabase and schist (Fig. 1) located in the eastern
Brick-red granite Red 97.06
part were mapped as distinct mafic units (D4 and D5 in Fig. Dark-gray diorite Green 99.96
6a). Small intrusions of mafic-ultramafic rocks (D6 and D7 in Diabase Blue 94.32
Fig. 6) in the northeastern part were separated and accurately Gabbro Yellow 98.12
classified as gabbro, which represents a significant improve- Gray diorite Cyan 83.20
Sand and gravel Aquamarine 99.57
ment over the prior geologic map wherein they were merged Sandstone Magenta 99.62
with adjacent units. This improved discrimination is impor- Schist Maroon 94.14
tant as it allows for the effective delineation of clustered, small Siltstone and limestone Sea green 98.28
outcrops. Furthermore, the SSTF mapped small outcrops of Flesh-red granite Purple 96.30
mafic rocks (D1, D2, and D3), all less than 100 m across, in Granite porphyry Coral 93.72
Overall accuracy 95.86
the western part of the study area. These were misidentified
as granite and schist in the 1:50,000 geologic map. Most im-
portantly, the carbonate lenses (yellow arrows in Fig. 5) that
affected the spectral index map had no influence on the SSTF chlorite and epidote) at 0.60 to 1.21 μm (gray shadow in Fig.
classification results. 3), whereas the absorption features of Al-OH, Fe-OH, and
In general, results showed that SSTF outperformed the Mg-OH were not obvious at 2.16 to 2.33 μm (Fig. 3).
band ratio and spectral index by providing a more detailed
and accurate map with clear lithologic boundaries. While all Field validation of mapping results
three methods were able to identify mafic-ultramafic rocks, Field inspections were conducted to assess the accuracy of
the spectral index method had more limitations. First, it was mapping results obtained by the false-color composite and
not effective in defining clear boundaries for different rock SSTF methods. A total of seven small-sized mafic-ultramaf-
types, particularly for small outcrops. Additionally, it incor- ic intrusions were examined (Fig. 7). None of them were
rectly identified certain carbonate lenses as mafic-ultramafic mapped on traditional geologic maps. These intrusions, rang-
in the southeast part of the study area (denoted by yellow ing in size from 1,000 to 70,000 m2, exhibit varying rock types
arrows in Fig. 5c). Finally, the spectral index method lacked including gabbro, pyroxenite, pyroxene peridotite, and olivine
sensitivity in identifying all lithological units except mafic- gabbro with strong weathering (Fig. 7). Four of the sites show
ultramafic rocks, indicating its limited capability as a single significant mineralization with the obvious presence of chal-
band ratio technique. copyrite on fresh rock surfaces. Further laboratory analysis
The SSTF image and false-color composite image of band was conducted to determine the mineralogical, spectral, and
ratios both successfully delineated the locations of mafic- alteration properties of these intrusions.
ultramafic intrusions with consistent results. Some litholog- Three small mafic-ultramafic rocks (D1, D2, D3) were
ical boundaries of more felsic rock units in the SSTF image found in the southwest portion of the study area. D1 (80 m
were more distinct than those in the false-color composite in diam) represented a newly identified mafic-ultramafic
image, indicating that the deep learning supervised classifica- unit, which was confirmed by thin-section investigation to
tion method works better in defining boundaries between dif- be strongly altered, serpentinized pyroxenite. Its surface is
ferent rock types. The effectiveness of the SSTF in discrimi- weathered to a dark-maroon color (Fig. 7a), whereas the fresh
nating lithological units with subtle spectral differences can surface is black. D2 (60 m in diam) and D3 (30 m in diam)
also be illustrated by its ability to distinguish granite porphyry are both gabbroic units (Fig. 7b), with melanocratic minerals
from granite in the western part (inside the green line in Fig. faded owing to weathering. It was observed that D1 and D3
6a). The granitic rocks appear as similar green-blue colors were mapped as granite on the geologic map, whereas D2 was
on the false-color composite image of band ratios. The SSTF mapped as schist (Fig. 1c). All three intrusions were omitted
output image looks somewhat noisy when the size of outcrops by previous work done by traditional geologic surveys, most
is small. For example, the diorite outcrops (inside the white likely because of their small sizes.
polygon in Fig. 6a) were identified as clusters of scattered Two mafic-ultramafic intrusions (D4 and D5) were discov-
pixels of diorite and diabase. The small clusters represent the ered in the eastern part of the study area. The D4 intrusion
widespread occurrence of diabase dikes in this area. was identified as olivine gabbro with 35% olivine, whereas D5
Cross-validation of the SSTF method with 20% training intrusion is gabbro with heterogeneous internal lithofacies
pixels produced an overall accuracy of 95.86% and a Kappa and a gray weathered surface. As illustrated in Figure 8a, the
coefficient of 0.95 (Table 2). The accuracy for seven classes fresh surface spectrum of D4 displays an absorption feature
(including brick-red granite, dark-gray diorite, gabbro, sand near 0.91 μm indicative of hypersthene (Clark et al., 1993;
and gravel, sandstone, siltstone and limestone, and flesh-red Yan et al., 2012), which was consistent with microscopic ob-
granite) was greater than 95%. The accuracy of diabase, schist, servation. In contrast, the weathered surface of D4 exhibits an
and granite porphyry ranged from 93.72 to 94.32%, whereas absorption near 2.32 μm characteristic of serpentine. The ser-
the accuracy of gray diorite was the lowest at 83.20%. The pentine formed along cracks and edges of olivine where the
confusion between diabase and gray diorite was the highest free iron formed magnetite. Absorption features near 2.35 μm
because of similar mineralogy. Diabase and gray diorite show indicative of epidote were observed in spectra of both fresh
similar spectral features of Fe2+/Fe3+-bearing minerals (e.g., and weathered surfaces of D5. D4 and D5 intrusions were not

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 10 YIN ET AL.

a b
D1 (pyroxenite)

D2 (gabbro) D3 (gabbro)

c D4 (olivine gabbro) d
D5 (gabbro)

chalcopyrite
chalcopyrite

e f
D6 (pyroxene peridotite)

D7 (olivine gabbro)

Fig. 7. Field photos and hand samples from different validation sites.

shown in the geologic map likely because of their small sizes. visible chalcopyrite. In thin section, it can be classified as ol-
The D4 intrusion was mapped as diabase covering a large ivine gabbro (Fig. 9b). As illustrated in Figure 9a, there is a
area, and the D5 intrusion was mapped as schist (Fig. 1c). narrow and shallow absorption feature near 2.23 μm due to
Two newly discovered intrusions (D6 and D7) have been Mg-OH vibrational overtones (Calvin and King, 1997; Ehl-
found in the northeast part of the study area. The D6 intru- mann et al., 2010). This is indicative of chlorite. Also, there is
sion, which is less than 100 m in diameter (Fig. 7e), is iden- a deep absorption near 2.35 μm characteristic of epidote and
tified as a pyroxene peridotite. The fresh surface of the D6 zoisite. Both of these two intrusions were mapped as part of
sample is dark green with widespread chalcopyrite. The D7 a large area of diabase. Neither of them is found as separate
intrusion (about 55 m in diam) is a dark gabbro (Fig. 7f) with intrusion on the 1:50,000 geologic map (Fig. 1c).

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 DEEP LEARNING APPLIED TO MAFIC-ULTRAMAFIC ROCKS 11

a b
20 D5-Weathered

Chr
Reflectance, %

15
D4-Weathered Ccp

10 D5-Fresh
Po
D4-Fresh

5
500 1000 1500 2000 2500
Wavelength, nm 200μm

c d

Ccp Py
Pn
Ccp
Pn
Py Pn
Pn

Py
Pn

200μm 1mm

S e S f

Cu
Ni
Fe
Fe
Fe Fe
Cu Ni
Fe Cu Fe Ni
0 2 4 6 8 0 2 4 6 8
FS: 781 cts CUR: 20.250 (0 cts) KeV FS: 904 cts CUR: 20.250 (0 cts) KeV
Fig. 8. Spectral and petrographic characteristics of samples from sites D4 and D5. (a) Reflectance spectra. (b) Reflected-light
photomicrograph for D4. (c) Reflected-light photomicrograph for D5. (d) Backscattered electron image for metallic minerals
in D5. (e) Energy dispersive spectrometry (EDS) spectra of chalcopyrite in a D5 sample. (f) EDS spectra of pentlandite in a
D5 sample. Ccp = chalcopyrite, Chr = chromite, CUR = cursor, FS = full scale, Pn = pentlandite, Po = pyrrhotite, Py = pyrite.

Discussion 7.55–12.5 μm, for example) is not effective in detecting small

outcrops owing to the required lower spatial resolution.
Effectiveness of WV-3 imagery for lithologic identification
Traditional geologic surveys are limited by their high cost, Relationships between laboratory measured spectra and
and they are not always effective in mapping small outcrops, image spectra
as illustrated by the omission of some outcrops (D1-D7) in Knowledge of rock mineralogy and spectroscopy is crucial
the 1:50,000-scale geologic map. Therefore, it is necessary to in choosing the appropriate method for lithological mapping
use high-resolution remote sensing imagery to guide further using satellite data (Karimzadeh and Tangestani, 2021). It is
exploration. The high spatial resolution of VNIR and SWIR useful to collect spectra from field samples that can be com-
WV-3 data allows for the mapping of relatively small rock pared to spectra from spaceborne sensors. Some rock types
outcrops. Thermal infrared data (with a wavelength range of have low reflectance (<20%) (Figs. 8a, 9a) that can make it

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 12 YIN ET AL.

a b
20
D6-Weathered Ser
Cpx
Reflectance, %

15

10 D7-Weathered
D7-Fresh
5 D6-Fresh Ol
500 1000 1500 2000 2500
Wavelength, nm 200μm
c d

Po Ccp Pn Ccp
Py
Mag

Po
Mag
Pn
Ccp
Ccp

200μm 100μm
e f

Ccp
Mag
Pn
Mag Pn
Py Chr
Po

Al-Spl
1mm 50μm
Ti g Fe h
Fe Cr
Cr O
O Al

Fe
Mg
Al Cr Cr Fe
Ti Fe
Ti Cr Cr Fe
0 2 4 6 8 0 2 4 6 8
FS: 750 cts CUR: 20.250 (0 cts) KeV FS: 985 cts CUR: 20.250 (0 cts) KeV
Fig. 9. Spectral and petrographic characteristics of samples from sites D6 and D7. (a) Reflectance spectra. (b) Transmitted-
light photomicrograph for D7 (crossed polarizer, 10×). (c) Reflected-light photomicrograph for D6. (d) Reflected-light pho-
tomicrograph for D7. (e) Backscattered electron image for metallic minerals in D6. (f) Backscattered electron image for
chromite in D6. (g) Energy dispersive spectrometry (EDS) spectra of titaniferous chromite from site D6. (h) EDS spectra
of magnesium-aluminum spinel from site D6. Al-Spl = aluminum spinel, Ccp = chalcopyrite, Chr = chromite, Cpx = clino-
pyroxene, CUR = cursor, FS = full scale, Mag = magnetite, Ol = olivine, Pn = pentlandite, Po = pyrrhotite, Py = pyrite,
Ser = sericite.

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 DEEP LEARNING APPLIED TO MAFIC-ULTRAMAFIC ROCKS 13

more difficult to match spectra. Although matching of the im- tite, pyrite, and chalcopyrite (Fig. 8c). Pyrite exhibits an id-
age spectra to laboratory spectra could be affected by atmo- iomorphic crystalline-granular texture, whereas chalcopyrite
spheric effects, instrument effects, spatial resolution, mixed developed along the edges of pyrite (Fig. 8d). Pentlandite,
pixels, and weathering (Yu et al., 2012), the shape of lab spec- which has a xenomorphic granular texture, generally occurs in
tra for weathered gabbro for example is comparable to WV-3 pyrite (Fig. 8c). The metallic minerals found in D6 are mainly
gabbro spectra (Fig. 3). This demonstrates that the FLAASH pentlandite, pyrrhotite, pyrite, chalcopyrite, magnetite, and
algorithm can accurately convert WV-3 radiance data into re- some chromite (Fig. 9c). Pentlandite and chalcopyrite exhibit
flectance (Berk et al., 1998; Mars, 2018). an idiomorphic to subidiomorphic granular texture, while
magnetite and pyrrhotite crystallized later. Spinel in D6 sam-
Influencing factors of deep learning and traditional ples developed an obvious inverse zonal texture in the back-
spectral indices scattered images (Fig. 9f). According to the EDS, the brighter
This study shows the efficiency of the spectral index approach region in the core is ferrochrome spinel, namely chromite,
to map large mafic-ultramafic intrusions, albeit with some which is rich in Fe-Cr with a small amount of titanium (Fig.
limitations. Specifically, the presence of carbonate units was 9g). The darker zone on the edge is rich in Mg-Al, which is
found to interfere with the spectral index results. The WV- magnesium-aluminum spinel with less chromium and iron
3 spectra of the carbonate lenses display high reflectance at (Fig. 9h). In contrast, mineralization in D7 samples is less in-
SWIR band 4 and low reflectance at SWIR band 8, which af- tense, with pyrite, chalcopyrite, and magnetite being the main
fected the spectral index (SWIR band 4 + SWIR band 6)/(2 metallic minerals present (Fig. 9d).
× SWIR band 5). Further work is needed to minimize the This study shows that the spectral indices and deep learn-
impact of carbonates on the spectral index results. ing methods are useful for prospecting for Ni-Cu-PGE–bear-
The present study demonstrates the potential of using a ing mafic-ultramafic intrusions with WV-3 data in this area of
false-color composite image of band ratios to detect mafic- Beishan, Northwest China. Mafic-ultramafic intrusions that
ultramafic rocks and to map most of the lithologic boundar- are associated with Cu-Ni mineralization in the area are typi-
ies. However, the SSTF approach is a more advanced, super- cally smaller than 1 km2 (Tang et al., 2012). It is important to
vised method that relies on the accuracy of training samples note that not all mafic-ultramafic intrusions in this region con-
for mapping. It is noteworthy that the carbonate lenses were tain economic Ni-Cu-PGE mineralization (Guo et al., 2017).
not defined as mafic-ultramafic rocks either in the false-color
composite image or in the SSTF results. Conclusions
The high accuracy of the SSTF classification result shows This study demonstrates that relatively small mafic-ultramafic
the reliability of the deep learning method for lithological rock outcrops can be detected with WV-3 VNIR and SWIR
mapping. Compared to the other two methods employed in multispectral data using spectral indices and the spatial-spec-
the study, the SSTF classification approach allowed more de- tral transformer (SSTF) deep learning method. WV-3 data
tailed identification of lithological units with greater intraclass has sufficient spectral resolution for discriminating mafic-ul-
variability. However, the classification accuracy could be af- tramafic rocks from other rock types in the study area, and its
fected by rock types of similar mineralogy. For example, the high spatial resolution helps to map smaller intrusions.
gray diorite in the northwest (inside the white polygon in Fig. The composite of the three band ratios (SWIR band
6a) was mostly confused with diabase owing to the similar 6/band 8 in red for Fe-OH and Mg-OH; SWIR band 5/
spectral features of Fe2+/Fe3+ in VNIR wavelengths and weak band 3 in green for Fe2+; VNIR band 3/band 2 in blue for
spectral features of Fe-OH and Mg-OH. It was determined Fe3+) and the newly proposed SWIR spectral index (SWIR
that the accuracy of the classification map generated by the band 4 + SWIR band 6)/(2 × SWIR band 5) provide simple
deep learning algorithm is highly dependent on the quality of but effective methods for detecting mafic-ultramafic rocks.
the samples used for training. In this study, training samples However, the deep learning method proves to be more pow-
were collected based on the 1:50,000 geologic map and previ- erful and enables a more detailed discrimination of different
ous geologic knowledge. Therefore, there could be some level rock types than the spectral indices-derived images. Most im-
of uncertainty about the results. portantly, the SSTF classification method was not affected by
the carbonate lenses that impacted the spectral index map.
Mineralization potential The contacts of rock types were also more clearly defined by
This study has successfully identified several mafic-ultramafic the SSTF classification. This study shows that spectral indi-
occurrences that are absent from the geologic map. This has ces and a deep learning method provide practical methods for
led to the discovery of four mafic-ultramafic intrusions with detecting mafic-ultramafic outcrops. However, the attention-
Cu-Ni-PGE mineralization. These are chalcopyrite-bearing based SSTF method is more effective and results in higher
intrusions of various rock types including gabbro, pyroxenite, classification accuracy.
and pyroxene peridotite. The presence of metallic minerals
was confirmed by the analysis of SEM and EDS data. Acknowledgments
Chalcopyrite, a copper iron sulfide mineral, was found in This work was supported by the Key Laboratory of Strate-
rock samples from all the D4, D5, D6, and D7 intrusions gic Mineral Resources of the Upper Yellow River, Ministry
(Figs. 7–9). The metallic minerals in D4 are mainly pyrrho- of Natural Resources (grant YSMRKF202203), the Natural
tite, chalcopyrite, and a small quantity of chromite, which Science Basic Research Program of Shaanxi Province (grant
developed along chalcopyrite fractures (Fig. 8b). D5 shows 2023-JC-ZD-18), the National Natural Science Foundation of
more evident mineralization, including pentlandite, pyrrho- China (grant 62001303), the Guangdong Basic and Applied

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 14 YIN ET AL.

Basic Research Foundation (grant 2023A1515012053), and with transformers: Institute of Electrical and Electronics Engineers, IEEE
the National Key Research and Development Program of Transactions on Geoscience and Remote Sensing, v. 60, p. 1–15.
Hu, J., He, Z., Li, J., He, L., and Wang, Y.W., 2018, 3D-Gabor inspired mul-
China (2022YFB3903702). tiview active learning for spectral-spatial hyperspectral image classification:
Remote Sensing, v. 10, article 1070.
Statements and Declarations Huang, K.K., Ren, C.X., Liu, H., Lai, Z.R., Yu, Y.F., and Dai, D.Q., 2022,
The authors declare that they have no known competing fi- Hyperspectral image classification via discriminant Gabor ensemble filter:
nancial interests or personal relationships that could have ap- Institute of Electrical and Electronics Engineers, IEEE Transactions on
Cybernetics, v. 52, p. 8352–8365.
peared to influence the work reported in this paper. Hunt, G.R., 1977, Spectral signatures of particulate minerals in visible and
near irfrared: Geophysics, v. 42, p. 501–513.
REFERENCES Hunt, G.R., and Ashley, R.P., 1979, Spectra of altered rocks in the visible and
Abrams, M.J., Rothery, D.A., and Pontual, A., 1988, Mapping in the Oman near infrared: Economic Geology, v. 74, p. 1613–1629.
ophiolite using enhanced Landsat thematic mapper images: Tectonophys- Jiang, C.Y., Cheng, S.L., Ye, S.F., Xia, M.Z., Jiang, H.B., and Dai, Y.C., 2006,
ics, v. 151, p. 387–401. Lithogeochemistry and petrogenesis of Zhongposhanbei mafic rock body, at
Amelin, Y., Li, C., and Naldrett, A.J., 1999, Geochronology of the Voisey’s Beishan region, Xinjiang: Acta Petrologica Sinica, v. 22, p. 115–126.
Bay intrusion, Labrador, Canada, by precise U-Pb dating of coexisting bad- Karimzadeh, S., and Tangestani, M.H., 2021, Evaluating the VNIR-SWIR
deleyite, zircon, and apatite: Lithos, v. 47, p. 33–51. datasets of WorldView-3 for lithological mapping of a metamorphic-igneous
Amer, R., Kusky, T., and Ghulam, A., 2010, Lithological mapping in the Cen- terrain using support vector machine algorithm: A case study of Central
tral Eastern Desert of Egypt using ASTER data: Journal of African Earth Iran: Advances in Space Research, v. 68, p. 2421–2440.
Sciences, v. 56, p. 75–82. Kruse, F.A., and Perry, S.L., 2013, Mineral mapping using simulated World-
Arndt, N.T., Lesher, C.M., and Czamanske, G.K., 2005, Mantle-derived mag- view-3 short-wave-infrared imagery: Remote Sensing, v. 5, p. 2688–2703.
mas and magmatic Ni-Cu-(PGE) deposits: Economic Geology 100th Anni- Kruse, F.A., Boardman, J.W., and Huntington, J.F., 2003, Comparison of
versary Volume, p. 5–24. airborne hyperspectral data and EO-1 Hyperion for mineral mapping:
Barnes, S.J., and Lightfoot, P.C., 2005, Formation of magmatic nickel sulfide Institute of Electrical and Electronics Engineers, IEEE Transactions on
deposits and processes affecting their copper and platinum group element Geoscience and Remote Sensing, v. 41, p. 1388–1400.
contents: Economic Geology 100th Anniversary Volume, p. 179–214. Kruse, F.A., Baugh, W.M., and Perry, S.L., 2015, Validation of DigitalGlobe
Barnes, S.J., Cruden, A.R., Arndt, N., and Saumur, B.M., 2016, The mineral WorldView-3 Earth imaging satellite shortwave infrared bands for mineral
system approach applied to magmatic Ni-Cu-PGE sulphide deposits: Ore mapping: Journal of Applied Remote Sensing, v. 9, article 096044.
Geology Reviews, v. 76, p. 296–316. Leverington, D.W., and Moon, W.M., 2012, Landsat-TM-based discrimina-
Barnes, S.J., Le Vaillant, M., and Lightfoot, P.C., 2017, Textural development tion of lithological units associated with the Purtuniq ophiolite, Quebec,
in sulfide-matrix ore breccias in the Voisey’s Bay Ni-Cu-Co deposit, Labra- Canada: Remote Sensing, v. 4, p. 1208–1231.
dor, Canada: Ore Geology Reviews, v. 90, p. 414–438. Lightfoot, P.C., Evans-Lamswood, D., and Wheeler, R., 2012a, The Voisey’s
Berk, A., Bernstein, L.S., Anderson, G.P., Acharya, P.K., Robertson, D.C., Bay Ni-Cu-Co sulfide deposit, Labrador, Canada: Emplacement of silicate
Chetwynd, J.H., and Adler-Golden, S.M., 1998, MODTRAN cloud and and sulfide-laden magmas into spaces created within a structural corridor:
multiple scattering upgrades with application to AVIRIS: Remote Sensing Northwestern Geology, v. 45, p. 17–38.
of Environment, v. 65, p. 367–375. Lightfoot, P.C., Keays, R.R., Evans-Lamswood, D., and Wheeler, R., 2012b,
Calvin, W.M., and King, T.V.V., 1997, Spectral characteristics of iron-bear- S saturation history of Nain Plutonic Suite mafic intrusions: Origin of the
ing phyllosilicates: Comparison to Orgueil (CI1), Murchison and Murray Voisey’s Bay Ni-Cu-Co sulfide deposit, Labrador, Canada: Mineralium
(CM2): Meteoritics and Planetary Science, v. 32, p. 693–701. Deposita, v. 47, p. 23–50.
Chai, G., and Naldrett, A.J., 1992, Characteristics of Ni-Cu-PGE mineral- Liu, L., Zhou, J., Jiang, D., Zhuang, D.F., and Mansaray, L.R., 2014, Litho-
ization and genesis of the Jinchuan deposit, Northwest China: Economic logical discrimination of the mafic-ultramafic complex, Huitongshan, Beis-
Geology, v. 87, p. 1475–1495. han, China: Using ASTER data: Journal of Earth Science, v. 25, p. 529–536.
Clark, R.N., Swayze, G.A., Gallagher, A.J., King, T.V.V., and Calvin, W.M., Liu, Y.G., Li, W.Y., Lu, X.B., Liu, Y.R., Ruan, B.X., and Liu, X., 2017, Sulfide
1993, Digital spectral library: Version 1: 0.2 to 3.0 μm: U.S. Geological Sur- saturation mechanism of the Poyi magmatic Cu-Ni sulfide deposit in Beis-
vey, Open File Report 93-592, p. 93–592. han, Xinjiang, Northwest China: Ore Geology Reviews, v. 91, p. 419–431.
Di Tommaso, I., and Rubinstein, N., 2007, Hydrothermal alteration mapping Lu, Y.G., Lesher, C.M., and Deng, J., 2019, Geochemistry and genesis of
using ASTER data in the Infiernillo porphyry deposit, Argentina: Ore Geol- magmatic Ni-Cu-(PGE) and PGE-(Cu)-(Ni) deposits in China: Ore Geol-
ogy Reviews, v. 32, p. 275–290. ogy Reviews, v. 107, p. 863–887.
Ehlmann, B.L., Mustard, J.F., and Murchie, S.L., 2010, Geologic setting of Maier, W.D., 2005, Platinum-group element (PGE) deposits and occur-
serpentine deposits on Mars: Geophysical Research Letters, v. 37, article rences: Mineralization styles, genetic concepts, and exploration criteria:
L06201. Journal of African Earth Sciences, v. 41, p. 165–191.
Gaffey, S.J., 1986, Spectral reflectance of carbonate minerals in the visible Mars, J.C., 2018, Mineral and lithologic mapping capability of WorldView
and near infrared (0.35–2.55 microns); calcite, aragonite, and dolomite: 3 data at Mountain Pass, California, using true- and false-color composite
American Mineralogist, v. 71, p. 151–162. images, band ratios, and logical operator algorithms: Economic Geology, v.
Gao, J.F., Zhou, M.F., Lightfoot, P.C., Wang, C.Y., and Qi, L., 2012, Origin 113, p. 1587–1601.
of PGE-poor and Cu-rich magmatic sulfides from the Kalatongke deposit, Mudd, G.M., and Jowitt, S.M., 2014, A detailed assessment of global nickel
Xinjiang, Northwest China: Economic Geology, v. 107, p. 481–506. resource trends and endowments: Economic Geology, v. 109, p. 1813–1841.
Goetz, A.F.H., Rock, B.N., and Rowan, L.C., 1983, Remote sensing for explo- Mungall, J.E., and Naldrett, A.J., 2008, Ore deposits of the platinum-group
ration: An overview: Economic Geology, v. 78, p. 573–590. elements: Elements, v. 4, p. 253–258.
Graham, S.D., Holwell, D.A., McDonald, I., Jenkin, G.R.T., Hill, N.J., Nair, A., and Mathew, G., 2012, Lithological discrimination of the Phenaimata
Boyce, A.J., Smith, J., and Sangster, C., 2017, Magmatic Cu-Ni-PGE-Au felsic-mafic complex, Gujarat, India, using the Advanced Spaceborne Ther-
sulfide mineralisation in alkaline igneous systems: An example from the mal Emission and Reflection Radiometer (ASTER): International Journal
Sron Garbh intrusion, Tyndrum, Scotland: Ore Geology Reviews, v. 80, of Remote Sensing, v. 33, p. 198–219.
p. 961–984. Naldrett, A.J., 1997, Key factors in the genesis of Noril’sk, Sudbury, Jinchuan,
Guo, Y.Z., Zhang, Z.H., Zhang, J., and Li, L., 2017, The petrophysical Voisey’s Bay and other world-class Ni-Cu-PGE deposits: Implications for
characteristics of the Heishan copper-nickel sulfide ore district in the Beis- exploration: Australian Journal of Earth Sciences, v. 44, p. 283–315.
han metallogenic belt and their prospecting significance: Geophysical and ——1999, World-class Ni-Cu-PGE deposits: Key factors in their genesis:
Geochemical Exploration, v. 41, p. 662–666 (in Chinese with English abs.). Mineralium Deposita, v. 34, p. 227–240.
Gupta, R.P., 2003, Remote sensing geology: Berlin, Springer, 655 p. Ninomiya, Y., and Fu, B.H., 2019, Thermal infrared multispectral remote
Hong, D.F., Han, Z., Yao, J., Gao, L.R., Zhang, B., Plaza, A., and Chanus- sensing of lithology and mineralogy based on spectral properties of materi-
sot, J., 2022, SpectralFormer: Rethinking hyperspectral image classification als: Ore Geology Reviews, v. 108, p. 54–72.

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 DEEP LEARNING APPLIED TO MAFIC-ULTRAMAFIC ROCKS 15

Ninomiya, Y., Fu, B.H., and Cudahy, T.J., 2005, Detecting lithology with Mongolia, based on the WorldView-3 and field shortwave-infrared spec-
Advanced Spaceborne Thermal Emission and Reflection Radiometer troscopy analyses: Remote Sensing, v. 13, article 914.
(ASTER) multispectral thermal infrared “radiance-at-sensor” data: Remote Su, B.X., Qin, K.Z., Sun, H., Tang, D.M., Xiao, Q.H., and Cao, M.J., 2009,
Sensing of Environment, v. 99, p. 127–139. Petrological and mineralogical characteristics of Hongshishan mafic-ultra-
Porter, T.M., 2016, Regional tectonics, geology, magma chamber processes mafic complex in Beishan area, Xinjiang: Implications for assimilation and
and mineralisation of the Jinchuan nickel-copper-PGE deposit, Gansu fractional crystallization: Acta Petrologica Sinica, v. 25, p. 873–887 (in Chi-
Province, China: A review: Geoscience Frontiers, v. 7, p. 431–451. nese with English abs.).
Pour, A.B., and Hashim, M., 2012, The application of ASTER remote sens- Su, B.X., Qin, K.Z., Sakyi, P.A., Liu, P.P., Tang, D.M., Malaviarachchi, S.P.K.,
ing data to porphyry copper and epithermal gold deposits: Ore Geology Xiao, Q.H., Sun, H., Dai, Y.C., and Hu, Y., 2011, Geochemistry and geo-
Reviews, v. 44, p. 1–9. chronology of acidic rocks in the Beishan region, NW China: Petrogenesis
Pour, A.B., Hashim, M., Park, Y., and Hong, J.K., 2018, Mapping alteration and tectonic implications: Journal of Asian Earth Sciences, v. 41, p. 31–43.
mineral zones and lithological units in Antarctic regions using spectral Su, B.X., Qin, K.Z., Sakyi, P.A., Tang, D.M., Liu, P.P., Malaviarachchi, S.P.K.,
bands of ASTER remote sensing data: Geocarto International, v. 33, p. Xiao, Q.H., and Sun, H., 2012, Geochronologic-petrochemical studies of
1281–1306. the Hongshishan mafic-ultramafic intrusion, Beishan area, Xinjiang (NW
Pour, A.B., Park, T.Y.S., Park, Y., Hong, J.K., Muslim, A.M., Läufer, A., China): Petrogenesis and tectonic implications: International Geology
Crispini, L., Pradhan, B., Zoheir, B., Rahmani, O., Hashim, M., and Hos- Review, v. 54, p. 270–289.
sain, M.S., 2019, Landsat-8, Advanced Spaceborne Thermal Emission and Sun, Y.Q., Tian, S.F., and Di, B.G., 2017, Extracting mineral alteration infor-
Reflection Radiometer, and WorldView-3 multispectral satellite imagery mation using WorldView-3 data: Geoscience Frontiers, v. 8, p. 1051–1062.
for prospecting copper-gold mineralization in the northeastern Inglefield Tang, Z.L., Yan, H.Q., Jiao, J.G., and Li, X.H., 2006, New classification of
mobile belt (IMB), northwest Greenland: Remote Sensing, v. 11, article magmatic sulfide deposits in China and ore-forming processes of small
2430. intrusive bodies: Mineral Deposits, v. 25, p. 1–9 (in Chinese with English
Qasim, M., Khan, S.D., Haider, R., and Rasheed, M.u., 2022, Integration of abs.).
multispectral and hyperspectral remote sensing data for lithological map- Tang, Z.L., Xu, G., Wang, Y.L., Qiu, G.L., and Dai, J.F., 2012, The new explo-
ping in Zhob ophiolite, western Pakistan: Arabian Journal of Geosciences, ration of magmatic mineralization: Small intrusion mineralization and geo-
v. 15, article 599. logical prospecting breakthrough: Northwestern Geology, v. 45, p. 1–16 (in
Rivard, B., Feng, J., Gallie, A., and Sanchez-Azofeifa, A., 2008, Continuous Chinese with English abs.).
wavelets for the improved use of spectral libraries and hyperspectral data: Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N.,
Remote Sensing of Environment, v. 112, p. 2850–2862. Kaiser, L., and Polosukhin, I., 2017, Attention is all you need: Conference
Rogge, D., Rivard, B., Grant, B., and Pardy, J., 2010, Mapping Ni-Cu (PGE) on Neural Information Processing Systems (NIPS 2017), 31st, Long Beach,
bearing ultramafic rocks with hyperspectral imagery, Nunavik, Canada: California, USA, 2017, Proceedings, p. 5998–6008.
Workshop on Hyperspectral Image and Signal Processing: Evolution in Xia, M.Z., Jiang, C.Y., Li, C.S., and Xia, Z.D., 2013, Characteristics of a newly
Remote Sensing (WHISPERS), 2nd, Reykjavik, Iceland, 2010, Proceedings, discovered Ni-Cu sulfide deposit hosted in the Poyi ultramafic intrusion,
p. 1–4. Tarim craton, NW China: Economic Geology, v. 108, p. 1865–1878.
Rogge, D., Rivard, B., Segl, K., Grant, B., and Feng, J., 2014, Mapping of Yan, B.K., Li, J.Z., Gan, F.P., Yang, S.M., and Wang, R.S., 2012, A spectral
NiCu-PGE ore hosting ultramafic rocks using airborne and simulated unmixing method of estimating main minerals abundance of lunar soils:
EnMAP hyperspectral imagery, Nunavik, Canada: Remote Sensing of Envi- Spectroscopy and Spectral Analysis, v. 32, p. 3335–3340.
ronment, v. 152, p. 302–317. Yang, H.Q., Li, Y., Yang, J.G., Li, W.M., Yang, L.H., Zhao, G.B., Ye, D.J.,
Rowan, L.C., and Mars, J.C., 2003, Lithologic mapping in the Mountain Pass, Zhao, Y.Q., Zhao, J.G., Shen, C.L., Wang, X.L., and Su, X.X., 2006, Main
California area using Advanced Spaceborne Thermal Emission and Reflec- metallogenic characteristics in the Beishan orogen: Northwestern Geology,
tion Radiometer (ASTER) data: Remote Sensing of Environment, v. 84, p. v. 39, p. 78–95 (in Chinese with English abs.).
350–366. Yang, J.G., Xie, C.L., Wang, X.H., Xie, X., Wang, L., and Wang, J.P., 2012,
Rowan, L.C., Goetz, A.F.H., and Ashley, R.P., 1977, Discrimination of hydro- Basic tectonic framework and features of metallogenic series in Beishan
thermal altered and unaltered rocks in visible and near infrared multispec- area, Gansu Province: Geological Bulletin of China, v. 31, p. 422–438 (in
tral images: Geophysics, v. 42, p. 522–535. Chinese with English abs.).
Rowan, L.C., Mars, J.C., and Simpson, C.J., 2005, Lithologic mapping of the Yang, S.H., Zhou, M.F., Lightfoot, P.C., Xu, J.F., Wang, C.Y., Jiang, C.Y., and
Mordor, NT, Australia ultramafic complex by using the Advanced Space- Qu, W.J., 2014, Re-Os isotope and platinum-group element geochemistry
borne Thermal Emission and Reflection Radiometer (ASTER): Remote of the Pobei Ni-Cu sulfide-bearing mafic-ultramafic complex in the north-
Sensing of Environment, v. 99, p. 105–126. eastern part of the Tarim craton: Mineralium Deposita, v. 49, p. 381–397.
Ruan, B.X., Yu, Y.M., Lv, X.B., Wu, C.M., and Liu, X., 2020, Sulfide segrega- Ye, B., Tian, S.F., Ge, J., and Sun, Y.Q., 2017, Assessment of WorldView-3
tion mechanism of magmatic Ni mineralization in western Beishan region, data for lithological mapping: Remote Sensing, v. 9, article 1132.
Xinjiang, NW China: Case study of the Hongshishan mafic-ultramafic com- Yu, B., Chen, F., Wang, N., Yang, L., Yang, H.P., and Wang, L., 2022,
plex: Ore Geology Reviews, v. 122, article 103503. MSFTrans: A multi-task frequency-spatial learning transformer for building
Ruan, B.X., Wei, W., Yu, Y.M., and Lv, X.B., 2021, Geology, geochronology, extraction from high spatial resolution remote sensing images: GIScience
mineral chemistry and geochemistry of the Hongnieshan mafic-ultramafic and Remote Sensing, v. 59, p. 1978–1996.
complex in the Beishan area, southern Central Asian orogenic belt, NW Yu, L., Porwal, A., Holden, E.J., and Dentith, M.C., 2012, Towards automatic
China: Implications for petrogenesis and regional Ni mineralization: Ore lithological classification from remote sensing data using support vector
Geology Reviews, v. 139, article 104423. machines: Computers and Geosciences, v. 45, p. 229–239.
Salehi, T., and Tangestani, M.H., 2018, Large-scale mapping of iron oxide Zadeh, M.H., Tangestani, M.H., Roldan, F.V., and Yusta, I., 2014, Mineral
and hydroxide minerals of Zefreh porphyry copper deposit, using World- exploration and alteration zone mapping using mixture tuned matched fil-
view-3 VNIR data in the Northeastern Isfahan, Iran: International Journal tering approach on ASTER data at the central part of Dehaj-Sarduiyeh cop-
of Applied Earth Observation and Geoinformation, v. 73, p. 156–169. per belt, SE Iran: Institute of Electrical and Electronics Engineers, IEEE
Sekandari, M., Masoumi, I., Pour, A.B., Muslim, A.M., Rahmani, O., Hashim, Journal of Selected Topics in Applied Earth Observations and Remote
M., Zoheir, B., Pradhan, B., Misra, A., and Aminpour, S.M., 2020, Applica- Sensing, v. 7, p. 284–289.
tion of Landsat-8, Sentinel-2, ASTER and WorldView-3 spectral imagery Zhang, B., Zhang, Y.J., Li, Y.S., Wan, Y., and Yao, Y.X., 2023, CloudViT: A
for exploration of carbonate-hosted Pb-Zn deposits in the Central Iranian lightweight vision transformer network for remote sensing cloud detection:
terrane (CIT): Remote Sensing, v. 12, article 1239. Institute of Electrical and Electronics Engineers, IEEE Geoscience and
Shen, L.L., and Jia, S., 2011, Three-dimensional Gabor wavelets for pixel- Remote Sensing Letters, v. 20, article 5000405.
based hyperspectral imagery classification: Institute of Electrical and Elec- Zhao, Z.F., Zhou, J.X., Lu, Y.X., Chen, Q., Cao, X.M., He, X.H., Fu, X.H.,
tronics Engineers, IEEE Transactions on Geoscience and Remote Sensing, Zeng, S.H., and Feng, W.J., 2021, Mapping alteration minerals in the
v. 49, p. 5039–5046. Pulang porphyry copper ore district, SW China, using ASTER and World-
Son, Y.S., You, B.W., Bang, E.S., Cho, S.J., Kim, K.E., Baik, H., and Nam, View-3 data: Implications for exploration targeting: Ore Geology Reviews,
H.T., 2021, Mapping alteration mineralogy in Eastern Tsogttsetsii, v. 134, article 104171.

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana
 16 YIN ET AL.

Chuntao Yin is a current geology Ph.D. candi- Lei Liu received B.S and M.S. degrees from
date at the School of Earth Science and Resources, Chang’an University and a Ph.D. degree from the
Chang’an University, China, with expectations to Chinese Academy of Sciences in 2004, 2007, and
graduate in June 2024. She specializes in economic 2011, respectively. Lei Liu is currently a profes-
geology with a special interest in the application of sor in the School of Earth Science and Resources,
remote sensing to the study of lithologic interpreta- Chang’an University, China. He specializes in litho-
tion and mineral resource exploration. Her research logic and mineral mapping and mineral resource
interest is in the application of multispectral and hyperspectral imagery and prospecting using multispectral and hyperspectral data.
spectral information to the extraction of mineralization information and
inversion of quantitative elements, with research projects involving a number
of image processing methods including false-color composite images, band
ratios, logical operator algorithms, and deep learning techniques.

Downloaded from http://pubs.geoscienceworld.org/segweb/economicgeology/article-pdf/doi/10.5382/econgeo.5056/6236524/5056_yin+et+al.pdf

by Edilber A. Chipana