Turkish Journal of Earth Sciences Turkish J Earth Sci

(2014) 23: 627-644

http://journals.tubitak.gov.tr/earth/
© TÜBİTAK
Research Article doi:10.3906/yer-1401-9

Alteration mineral mapping with ASTER data by integration of coded spectral ratio
imaging and SOM neural network model
1, 1 2
Mohammad Hassan TAYEBI *, Majid HASHEMI TANGESTANI , Robert Keller VINCENT
1
Department of Earth Sciences, Faculty of Sciences, Shiraz University, Shiraz, Iran
2
Department of Geology, Bowling Green State University, USA

Received: 08.01.2014 Accepted: 25.09.2014 Published Online: 03.11.2014 Printed: 28.11.2014

Abstract: This study applied coded spectral ratio imaging and a self-organizing map (SOM) neural network model for subpixel
classification of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) data to map alteration minerals of the
Masahim volcano, SE Iran. Spectra of samples were measured using a FieldSpec spectroradiometer, and were interpreted for finding
key alteration minerals. Key minerals including sericite and kaolinite were selected. Thirty-six nonreciprocal spectral ratios of 9 ASTER
reflective bands were created. Each spectral parameter (of 9 spectral bands and 36 spectral ratios) for the field and library mineral
spectra was divided into deciles. Each decile of spectral bands or ratios was labeled from 9 for the highest down to 0 for the lowest decile.
Spectral ratio code signature patterns of field samples were then compared to the US Geological Survey (USGS) spectral library to find
3 spectral ratio images for each key mineral. Based on these 3 ratio images, color ratio images were created and training areas were
selected by a geographic information systems (GIS) overlay approach and were used to train the SOM model. This model was applied
to the 5, 15, and 36 ratio image data sets as chemical composition images, and the results were compared using confusion matrices
and coefficient of determination (R2). The results of SOM applied to 15 and 36 ratio image data sets showed good identification of key
minerals. We achieved overall accuracies of 87% for SOM applied to the 36 ratio images, 86% for 15, and 85% for 5 ratio image data sets.
The R2 produced from 15 ratio image data sets for sericite and kaolinite were 0.35 and 0.52, whereas they were 0.64 and 0.68, for 36 ratio
image data sets. It was concluded that using SOM applied to ratio images could, therefore, be useful in alteration mapping of a volcano.

Key words: Alteration mapping, Masahim volcano, SOM neural network model, coded spectral ratio imaging, GIS

1. Introduction and Tangestani, 2011). Hence, data from such sensors

The study of volcanoes by earth observation technology,
imaging spectrometry, and geographic information system
(GIS) is important endeavor for a volcanologist in earth
and planetary exploration. It provides the opportunity to
map lithology, mineralogy, altered rocks, debris flows, and
lahar paths (Kruse, 1988; Crowley and Zimbelman, 1997;
Crowley et al., 2003; Kerle et al., 2003; Terunuma et al.,
2005; Hubbard et al., 2007; Joyce et al., 2009) and to assess
hazards related to volcanoes.
Multispectral imaging sensors with discrete spectral
bands have proven useful for discrimination and
classification of minerals and lithological units (Cudahy
and Hewson, 2002; Rowan and Mars, 2003; Yamaguchi
and Naito, 2003; Hellman and Ramsey, 2004; Gomez
et al., 2005; Zhang et al., 2007; Perry and Vincent, 2009;
Bedini, 2009; Mars and Rowan, 2010; Rajendran et al.,
2011), and also for mapping altered rocks (Crosta et al.,
2003; Rowan et al., 2003; Souza Filho et al., 2003; Mars
and Rowan, 2006; Tangestani et al., 2008; Hosseinjani
* Correspondence: mhtayebi@shirazu.ac.ir
are capable of distinguishing many more differences in
chemical composition than could ever be distinguished
from visible reflected sunlight alone (Perry and Vincent,
2009). However, many factors can influence the accuracy
of chemical composition maps derived from multispectral
remote sensing data, such as the number of spectral bands
employed (Perry and Vincent, 2009), and variations in
slope, solar illumination, and atmospheric haze (Vincent,
1972, 1997).
There are 2 possible approaches one can take to map
and detect minerals and lithology on the basis of chemical
composition using multispectral imaging. One approach is
the theoretical approach of the modeler, where all
extraneous information (atmospheric attenuation, sensor
electronic gain and offset, etc.) is subtracted off and
divided out, and laboratory (or field) spectra of materials
are used as training sets to map features in the scene of
similar composition. This approach is very difficult to keep
accurate because of variations in the atmosphere, solar
627
 TAYEBI et al. / Turkish J Earth Sci
illumination, and sensor electronic gain and offset that are
not always perfectly recorded or accounted for (Vincent,
1997). A second approach is an empirical approach,
whereby the corrections for atmospheric absorption and
scattering, solar illumination variations with topographic
slope, and sensor electronic gain and offset are made
from the image data itself, and then selected spectral
parameters of each pixel are compared to the reflectance
spectrum of laboratory and field spectra of user-selected
objects to search the image for objects that have a similar
chemical composition. Dark-object–corrected spectral
ratio imaging is such an empirical approach, in which the
shadow/slope factor and additive factor (the dark object,
or DO) from atmospheric haze and electronic offset of the
sensor, almost constant for the entire data set, are removed
(Vincent, 1997). This becomes less accurate if the elevation
differences within the scene of interest are very large, but
this problem also occurs in the theoretical approach.
It has been shown that dark-object–corrected spectral
ratio images enhance chemical compositional differences,
while suppressing variations in slope, solar illumination,
and atmospheric haze (Vincent, 1972, 1997; Rowan et al.,
1974; Perry and Vincent, 2009). Triplets of spectral ratios
displayed in colors of red, green, and blue (as a ratio color
composite) are commonly used to enhance the appearance
of specific minerals of interest, where they are exposed
on the ground (Perry and Vincent, 2009). However, the
problem is that the color ratio images cannot show the
fractions of desired minerals.
Detection and quantification of fractions of the target
materials for each pixel can be important for dark-object–
corrected spectral ratio mages. Therefore, evaluating
a model like neural network to map and estimate
abundance  fractions  of minerals present in ratio images
can be useful for the mapping of volcanoes. Artificial
neural networks (ANNs) were originally designed
as pattern-recognition (DuBose and Klimasauskas,
1989) and data analysis tools that mimic the neural
storage and analytical operations of biological neural
networks (DuBose and Klimasauskas, 1989; Villmann
et al., 2003). ANN approaches have a distinct advantage
over statistical classification methods in that they are
nonparametric, require little or no a priori knowledge of
the distribution model of input data and no assumption
about the probabilistic models of data, are robust in noisy
environments, and have the ability to learn complex
patterns (Ji, 2000) and perform generalizations (Zhuang et
al., 1991; Geaga et al., 1992; Emaruchi et al., 1997).
Among the ANN techniques, the self-organizing map
(SOM) neural network model has been used in different
fields such as land-use classification (Ji, 2000; Bagan et
al., 2005; Jianwen and Bagan, 2005), lithological mapping
(Mather et al., 1998; Grebby et al., 2010; Bedini, 2012),
628
and geomorphometric feature analysis (Ehsani and
Quiel, 2008). The SOM is a 2-layer structure, including
1 input layer and 1 output layer. Therefore, the dilemma
of determining the hidden layer size is avoided (Hu and
Weng, 2009). It is capable of coping with presence-only
data (Li and Eastman, 2006), it is not affected by the local
minima problem and it is insensitive to the structure of the
codebook vector map (Lee and Lathrop, 2006).
There were several questions raised at the onset of this
study. Is it efficient for us to use the SOM model for subpixel
classification of ratio images as chemical composition
images? How many ratio images do we need to input to the
SOM to get good subpixel classification results? The study
area is the Masahim composite volcano, which is situated
in western Kerman, Iran. In this study, a coded spectral
ratio imaging method was used to find suitable training
areas and a SOM model was applied to the different ratio
image data sets, including 5 and 15 short wave infrared
(SWIR) and 36 visible near infrared and short wave
infrared (VNIR+SWIR) ratio images, respectively.
2. Geological setting
The Masahim composite volcano (55°07ʹ–55°30ʹE;
30°10ʹ–30°30ʹN) lies 265 km west of Kerman, SE Iran
(Figure 1); it is situated in the Dehaj–Sarduiyeh district,
which is a part of the Urumieh-Dokhtar Magmatic Arc
(UDMA). The Masahim volcano was formed between 6.8
± 0.4 (older stratigraphic level) and 6.3 ± 0.9 (younger
stratigraphic level) Ma ago (Hassanzadeh, 1993). Its
eruptions have been mainly phreato-magmatic (including
interaction of water and magma) to phreato-plinian (same
interaction with higher plumes), including lava flows,
pyroclastic surges, pyroclastic falls, and pyroclastic flows
(Ahmadipour, 1994) as the main components.
Ahmadipour (1994) divided the Masahim volcano
according to petrographic and structural characteristics,
as well as distances to their origins, into 4 main facies:
central (caldera), proximal, medial, and distal, with several
units in each. The central facies mainly consists of lava
breccias, which are rhyodacitic-dacitic in composition and
in some part display intensive hydrothermal alteration, as
well as pre- and postcaldera stage intrusives. Precaldera
stage intrusive rocks generally contain plagioclase, biotite,
hornblende, and quartz (andesitic-dacitic) that have been
exposed by collapse and erosion, and have undergone
intensive alteration (Ahmadipour, 1994). At the central
facies mafic constitutents are completely altered, as
well as feldspar, and silicification represents the final
process (Dimitrijevic, 1973). However, postcaldera stage
intrusives (plugs and dikes) lack hydrothermal alteration
(Dimitrijevic, 1973), and only display evidence for surficial
weathering. Alluvial and colluvial deposits cover portions
of the central facies, and are especially concentrated
 TAYEBI et al. / Turkish J Earth Sci

55°10'0"E 55°20'0"E
dc-a
Ngvc Ec

Etab Iran
dc-a
El

Study area
a
Kerman
Et
a
Pe

Ngp
rs
ia
30°20'0"N

30°20'0"N
G
ul

Ettb
f

dc-a
Urumieh Dokhtar Magmatic Belt

Ngar d Ev
d
Ngpy

Qal
Ngp
dc-a
Meymand
a
Ngvc Ngp
gd Q
MF
Qt2
Sf Q DF

0 1.5 3 6 9
Km

55°10'0"E 55°20'0"E DF: Distal facies

Sf Sand flats Trachyandesitic- Trachybasaltic
Ettb tuffs -lava flows
QUATERNARY

Recent alluvium
Etb Trachyandesitic-
Qt2 Younger gravel fans CF Ets
Ers
Trachybasaltic lava flows
Tuffs-tuffites
Ev

Qt1 Older gravel fans Etab Trachyandesitic-basaltic

dc-a Dacite and andesite
PF Et
Red tuffs
EOCENE

d Diorite
Ngar
Altered rocks
Eab Andesite-basaltic rocks
Ngpy Pyroclastics El Limestones
Hornblende phenoandesite Limestones and marls
NEOGENE

a Elm
(lava flow) Fault
Ers Red sandstones
Ngvc Volcanomict conglomerates
Conglomerates Road
(pyroclastics fall and flow) Ec
Sample sites
Ngp Pyroclastics surge Elf Flysch Residential
area
gd Granodiorite

Figure 1. Location map of the study area. Geological map of northwest Shahr-e-Babak, Iran, showing the
Masahim caldera at center (modified from 1:100,000 geological map of Shahr-e-Babak, Geological Survey
of Iran, 1973).

along the inner slopes of the caldera rim. The lower-most
exposures in the proximal facies are pyroclastic surge
deposits (tuff-breccias) that mantle the older intrusive
country rocks (Ahmadipour, 1994).
Pyroclastic flows include block and ash with
stratigraphic layering that reflects the history of the
volcano, as well as abundant angular blocks embedded
in volcanic ash and lapilli (Figure 1, Ahmadipour, 1994).
Pyroclastic falls are poorly sorted layers of welded, lithic
tuff rich in lithic fragments (andesitic-dacitic), which in
particular show enrichment in lithic fragments towards the
top of each layer. Lava flows overlie a thick accumulation
of pyroclastic flows and falls of the volcano and mainly
consist of andesite and andesite-basalt rocks (Ahmadipour,
1994). Medial and distal facies mainly contain pyroclastic
falls and flows.

629
 TAYEBI et al. / Turkish J Earth Sci

3. Materials and methods
VNIR and SWIR datasets of one ASTER (Level 1B,
0410120707401112190002) scene, acquired on 5 October
2004, were used to map alteration units of the Masahim
composite volcano. The ASTER instrument measures
reflected radiation in 3 bands between 520 and 860 nm
(VNIR) and 6 bands from 1600 to 2430 nm (SWIR), with
15- and 30-m resolution, respectively (Fujisada, 1995).
The following steps (Figure 2) show brief descriptions of a
procedure along which an integration of ASTER data and
SOM neural network model was implemented to deal with
the objectives of the research:
(1) Pre-processing of ASTER data, including geometric
correction using 1:25,000 topographic maps and
atmospheric correction using the dark object subtraction
method. Dark object subtraction is a simple empirical
procedure (Gupta, 2003) that adjusts the brightness
values (digital number) of a given scene based on image
information. The digital number to subtract from each band
can be the band minimum, an average, or a specific value
(Research Systems Inc, 2003). The scattering is removed by
subtracting band minimum from every pixel in the band.
Then 3 sets of spectral ratio image data including 36 ratio
images from VNIR+SWIR bands, a subset of 15 images
from SWIR bands alone, and yet smaller subset of 5 ratio
images from SWIR bands but based on characteristic ratio
codes of key representative minerals were produced and
used as inputs to the SOM neural network (Table 1).
(2) Field observation, sample collection, spectral
reflectance measurement of fresh and weathered surfaces of
the rock samples using a Fieldspec Pro spectroradiometer
and X-ray diffraction (XRD) analysis, and identification
of key representative minerals, which are composed of the
mineral assemblages.
(3) Calculation of spectral ratio codes for the field
samples and the USGS spectral library, determination of 3
ratio images for each representative mineral, and creation
of color ratio composite images.
(4) Training area selection using overlay (AND and
XOR logic) method for representative minerals.

Preproces s ing of AS TER datas ets Training area

(s patial reges tration-geometric correction and s election
dark object s ubtraction)

1 2 3 Identifying repres entative

alteration minerals
us ing field s amples
5 ratio 15 ratio 36 ratio imag e s
imag e s (S WIR) imag e s (S WIR) (VNIR+S WIR)

Calculating s pectral ratio

codes for field s amples
and US GS library

Comparing s pectral ratio

S OM codes of field s amples and
ne ural ne twork US GS library

Determining 3 ratio images

Training area for each mineral and
creating color ratio compos ite
image

Feature maps
GIS overlay method Producing
(AND and XOR) 3 ratio value layers

Accuracy as s es s ment

Figure 2. The flowchart of process steps. After preprocessing, training area is selected using field samples and
representative alteration minerals. Spectral ratio codes of USGS library and field samples are calculated and
compared for finding 3 ratio images for each mineral and color ratio image is created. GIS overlay method
is used to select training area. SOM is then separately applied to 5, 15, and 36 spectral ratio image data sets.
Feature maps are generated for each input. Finally, the results are compared using confusion matrices and of R2
coefficients for accuracy assessment.

630
 TAYEBI et al. / Turkish J Earth Sci
Table 1. Ratio image data sets which were input to the SOM neural network.
Input to the SOM Ratio images
5 ratio images (SWIR-Based on ratio
6/5, 7/5, 7/6, 8/6, 9/7
codes of key representative minerals)
15 ratio images (SWIR) 5/4, 6/4, 7/4, 8/4, 9/4, 6/5, 7/5, 8/5, 9/5, 7/6, 8/6, 9/6, 8/7, 9/7, 9/8
2/1, 3/1, 4/1, 5/1, 6/1, 7/1, 8/1, 9/1, 3/2, 4/2, 5/2, 6/2, 7/2, 8/2, 9/2, 4/3, 5/3, 6/3,
36 ratio images (VNIR+SWIR)
7/3, 8/3, 9/3, 5/4, 6/4, 7/4, 8/4, 9/4, 6/5, 7/5, 8/5, 9/5, 7/6, 8/6, 9/6, 8/7, 9/7, 9/8
(5) Application of SOM to spectral ratio image data
sets (5, 15, and 36 ratio image data sets) for subpixel
classification and generation of feature maps.
(6) Accuracy assessment and comparison of the
output results via a confusion matrices approach and R2
coefficient. The flowchart of processes is shown in Figure 2.
3.1. Training area selection
For subpixel classification of ratio image data sets with
a neural network, the analyst must select representative
pixels as a training area for each of the key minerals. If
the multispectral image contains sufficiently distinct visual
cues, it may be possible to find suitable training areas by
visual examination. Frequently, however, one must resort
to additional sources of information, such as field data
or existing maps, to find representative areas for each
class. The process of finding training areas can, therefore,
be rather labor intensive (Schowengerdt, 2007). Since
coded spectral ratio imaging (Vincent, 1997) information
enhances chemical compositional differences by using the
most unique spectral characteristics (absorption features)
of a given mineral or rock and a neural network has no
way of finding the most characteristic spectral parameters
like the ratio codes, ratio codes information can be used to
select training areas for a given mineral or rock, especially
when we do not have access to the area and only have
access to archived library spectra of the USGS and field
samples (collected from the study area) as references for
material identification.
3.1.1. Spectral features of caldera (central facies)
The spectral reflectance of field samples was measured
using a Fieldspec Pro spectroradiometer (ASD) with
spectral range of 350–2500 nm, and a quartz-tungsten-
halogen (QTH) lamp was used as light source (Figures
3 and 4). USGS spectral library (Clark et al., 1993) and
XRD analysis were used for comparative analysis and
characterization of field samples. Sample site locations are
shown in Figure 1.
The result of characterizing field samples by
interpretation of reflectance spectra showed alteration to
sericite+illite+goethite+hematite+jarosite (phyllic zone),
kaolinite+ montmorillonite+goethite+hematite (argillic
zone), and chlorite (Figures 3 and 4). Rock samples 219,
228, 263, 245, and 100 display multiple absorption features
near 1.4 µm, 1.9 µm, 2.20 µm, and 2.35 µm. However,
rock sample 100 does not exhibit strong absorption
features in 1.4 µm or 2.20 µm. For example, the spectrum
of sample 263 shows alteration to sericite+illite, sample
219 to sericite+illite+hematite, and samples 228 and 245
to sericite+goethite (Figure 3a). The resampled spectra
of all rocks exhibit Al-OH absorption bands around
2.20 µm, which falls within the spectral limits of ASTER
band 6 (Figure 3b). Rock samples 218, 222, 266, and 264
exhibit multiple absorption features near 0.48 µm, 0.92
µm that can be attributed to the goethite, and near 1.39
µm, 1.41 µm, 1.9 µm, 2.17 µm, and 2.20 µm that can be
attributed to kaolinite (Hunt and Salisbury, 1970). Rock
sample 239 exhibits multiple absorption features near 1.39
µm, 1.41 µm, 2.17 µm, and 2.20 µm that can be attributed
to kaolinite (Hunt and Salisbury, 1970). The spectra of
samples 218 and 264 indicate mixtures of kaolinite+mo
ntmorillonite+goethite, rock sample 239 is a mixture of
montmorillonite+kaolinite, and sample 222 is a mixture
of montmorillonite+kaolinite+goethite (Figure 3c). The
resampled spectra of rocks (Figure 3d) exhibit an Al-OH
absorption band around 2.20 µm (coincided on band 6
of ASTER), and a high reflectance in band 4 of ASTER,
which is typical in argillic (kaolinite) and phyllic (sericite)
alterations. However, the difference between kaolinite and
sericite is in the reflectance of band 5, which is lower in
kaolinite relative to sericite (Figure 3d).
The rock samples 229, 232, 233, 236, 237, 238, 259,
and 262 typically exhibit multiple absorption features
near 1.4 µm 1.9 µm, and 2.20 µm, and weaker features
near 2.35 µm, 1.83, 2.1 µm, 0.88 µm, and 2.44 µm that can
be attributed to the sericite+illite. The spectra of samples
236, 238, 259, and 262 show mixtures of sericite+illite+
hematite (Figures 3e and 4a); samples 237 and 232 show
goethite, in addition to the preceding minerals (Figure
3e); samples 233 and 229 show mixtures of sericite+illite
(Figures 3e and 4a), and sample 257 reveals a mixture of
jarosite+sericite (Figure 4a). The resampled spectra of all
rocks exhibit Al-OH absorption bands around 2.20 µm,
which falls within the spectral region of ASTER band 6
(Figures 3f and 4b). Rock samples 250 and 254 exhibit
multiple absorption features near 0.48 µm, and 0.92 µm
that can be attributed to goethite, and 1.39 µm, 1.41 µm,
631
 TAYEBI et al. / Turkish J Earth Sci

a c e
100 232

264
245
Value (offset for claritiy)

233
266

236
239
263
237
228 222

238
219 218

2000 2500 500 1000 1500 2000 2500 500 1000 1500 2000 2500
500 1000 1500

b 100 d f
232

245 264
Value (offset for claritiy)

266 233

263 239
236
228
222
219 237

218
238

2500 1000 1500 2000 2500 1000 1500 2000 2500

1000 1500 2000

X axis: Wavelength (nm)

Figure 3. Reflectance spectra of volcano samples measured in the laboratory. (a) Reflectance spectra of lava breccias exposed at
the caldera, and (b) spectra after being resampled to ASTER VNIR+SWIR bands. (c) Reflectance spectra of altered lava breccias
exposed at the caldera and (d) spectra after being resampled to ASTER VNIR+SWIR bands. (e) Reflectance spectra of precaldera
stage intrusive rocks exposed at the caldera and (f) spectra after being resampled to ASTER VNIR+SWIR bands.

1.9 µm, 2.17 µm, and 2.20 µm that can be attributed to
kaolinite (Figures 4c and 4d).
Rock samples 252, 224, and 258 (plugs and dikes,
Figures 4e and 4f) do not exhibit strong absorption
features in the VNIR-SWIR region (Figure 4e). However,
some of them exhibit weaker absorption features near
0.48 µm (goethite), 0.90 µm, 2.2 µm, 2.3 (chlorite), 1.4
µm, and 1.9 µm. Absorption minima at 1.4 µm and 2.20
µm, with shallower depth at 1.9 µm, indicate the presence
of hydroxyl groups, rather than just water of hydration,
and can be attributed to weak alteration (Figure 4e). The
resampled spectra of these rocks also do not exhibit strong
absorption features (Figure 4f). The phyllic and argillic
alteration units (sericite and kaolinite, respectively) were
typically mixed at the center of the volcano. Therefore, we
tried to map sericite (phyllic zone) and kaolinite (argillic
zone) in the study area.
3.1.2. Coded spectral ratio imaging
Coded spectral ratio imaging (Vincent, 1997) was
developed to enhance chemical compositional differences
by using the most unique spectral characteristics
(absorption features) of a given mineral or rock. Each
digit in a ratio code of a particular dark-object-subtracted
spectral ratio represents the decile of dark-to-bright in that
spectral ratio image in which a particular mineral or rock
would be displayed. For instance, for a given spectral ratio
image, a ratio code of 0 signifies that a particular mineral
or rock is in the darkest 10% (decile) of all the materials in
the spectral library for that spectral ratio image. Likewise,
a ratio code of 9 signifies that a particular rock or mineral

632
 TAYEBI et al. / Turkish J Earth Sci

a c e 258-E

254 224-E
Value (offset for claritiy)

258-B
257

250
259 224-B

251
262 264-B
229

500 1500 2000 2500 500 1000 1500 2000 2500 500 1000 1500 2000 2500
1000

b d f 258-E
Value (offset for claritiy)

254 224-E
258-B
257

250

259
224-B

264-B 251
262
229
1000 1500 2000 2500 1000 1500 2000 2500 1000 1500 2000 2500

X a xis : Wa ve le ng th (nm)
Figure 4. Reflectance spectra of volcano samples measured in the laboratory. (a) Reflectance spectra of precaldera stage intrusive
rocks exposed at the caldera and (b) spectra after being resampled to ASTER VNIR+SWIR bands. (c) Reflectance spectra of
precaldera stage intrusive rocks exposed at the caldera and (d) spectra after being resampled to ASTER VNIR+SWIR bands.
(e) Reflectance spectra of postcaldera stage intrusive rocks exposed at the caldera and (f) spectra after being resampled to ASTER
VNIR+SWIR bands.

is brighter in that spectral ratio image than at least 90% of
the materials included in the spectral library. To produce a
color image that shows the mineral you wish to map as red,
you must find 3 spectral ratios (from the ratio code tables
for that particular sensor) that have ratio codes of 9, 0, and
0 and display them as red, green, and blue, respectively.
3.1.3. Selecting training area using GIS
The spectral ratio codes of 13 spectra of sericite, 8 spectra
of kaolinite from the USGS spectral library (Table 2), and
field samples were compared to select the best ratio images
to provide contrast between the minerals of interest
(Figure 5). Three spectral ratios with very high ratio values
(ideally, a ratio code of 9) and spectral ratios with very low
values (ideally, with a ratio code of 0) were selected. The
best ASTER spectral ratios for mapping sericite were 7/6
(with ratio code 9), 6/5 (with ratio code 0), and 8/6 (with
ratio code 9); and for kaolinite were 7/5(with ratio code 9),
7/6 (with ratio code 9) and 6/5 (with ratio code 1). Figure 6
shows the resulting color ratio image for enhancing sericite
as yellow (Figure 6a, where the ratios with ratio codes of
9 were displayed as red and green, and the one with the
lowest ratio code was displayed as blue), and kaolinite as
yellow (Figure 6b, similarly displayed).
The brighter (ratio code of 9) or darker pixels (ratio
code of 0) in the spectral ratio images indicate a better
match to the reference material in those pixels. However,
higher and lower values could be artifacts of the shadow
and topographic effects that remained near cliffs after the
resampling of 30 m pixels to 15 m pixels, and should not
be used. The actual values of each ratio image were divided
into 7 intervals or slices using density slicing, and the
output pixels exported into a GIS environment as a shape
file layer, yielding 3 ratio value layers for each mineral.

633
 TAYEBI et al. / Turkish J Earth Sci

Table 2. ASTER VNIR+SWIR ratio codes for muscovite (sericite) and kaolinite (USGS spectral library).

Ratio image/Mineral 2/1 3/1 4/1 5/1 6/1 7/1 8/1 9/1 3/2 4/2 5/2 6/2 7/2 8/2 9/2 4/3 5/3 6/3
muscovi1 8 7 7 5 2 5 4 4 7 7 4 2 4 3 3 7 3 2
muscovi2 6 6 5 4 2 4 4 4 6 5 3 2 4 3 3 5 3 2
muscovi3 3 4 7 7 6 6 6 6 5 7 7 6 6 7 6 8 7 6
muscovi4 7 7 6 4 2 5 4 4 6 6 3 2 4 3 3 6 3 2
muscovi5 4 5 5 5 2 4 4 4 6 6 5 2 4 4 3 6 4 2
muscovi6 6 6 7 7 5 6 6 5 7 7 6 5 6 6 5 7 6 4
muscovi7 3 3 5 3 2 3 3 3 2 5 3 2 3 3 3 6 3 2
muscovi8 4 5 6 6 3 5 5 4 6 6 6 3 5 5 4 6 6 3
muscovi9. 6 6 5 3 2 4 3 3 6 5 3 2 3 3 3 5 3 2

Table 2. (Continued). ASTER VNIR+SWIR Ratio codes for muscovite (sericite) and kaolinite (USGS spectral library).

Ratio image/Mineral 7/3 8/3 9/3 5/4 6/4 7/4 8/4 9/4 6/5 7/5 8/5 9/5 7/6 8/6 9/6 8/7 9/7 9/8
muscovi1 4 3 3 2 1 3 3 2 0 4 2 2 9 9 8 2 1 2
muscovi2 4 3 3 3 2 4 4 4 0 4 3 3 9 9 7 3 3 2
muscovi3. 7 7 6 5 3 3 4 4 0 1 2 2 7 7 5 6 4 2
muscovi4 4 3 3 2 1 3 3 2 0 8 3 3 9 9 8 2 2 2
muscovi5 4 4 3 3 2 4 4 3 0 2 3 3 9 9 7 4 3 2
muscovi6 6 5 4 4 2 3 4 3 0 1 2 2 8 8 6 4 3 2
muscovi7 4 3 3 2 1 3 3 2 0 3 3 3 9 9 8 3 2 2
muscovi8 5 5 4 4 2 4 4 4 0 2 3 3 8 8 6 4 3 2
muscovi9. 3 3 3 3 2 3 3 3 0 3 3 3 9 9 7 3 2 2
Ratio image/Mineral 2/1 3/1 4/1 5/1 6/1 7/1 8/1 9/1 3/2 4/2 5/2 6/2 7/2 8/2 9/2 4/3 5/3 6/3
muscovia 6 6 6 4 2 4 4 4 6 6 4 2 4 3 3 6 3 2
muscovib 5 6 7 6 3 5 5 4 7 7 6 3 6 5 4 7 6 3
muscovic 6 7 7 6 4 6 5 5 7 7 6 3 6 5 4 7 6 3
muscovid 6 6 6 6 5 6 6 6 7 6 6 5 6 6 5 6 6 4
kaolini1 4 4 4 1 1 1 1 1 5 4 1 1 1 2 2 4 1 1
kaolini2 6 6 5 1 1 2 2 2 6 4 1 1 2 2 2 4 1 2
kaolini3 6 6 5 1 1 2 2 2 5 4 1 1 2 2 2 4 2 2
kaolini4 6 6 4 1 1 2 2 1 6 3 1 1 2 2 2 2 1 1
kaolini5 6 6 5 1 1 2 2 2 7 5 1 1 2 2 2 4 1 1
kaolini6 6 6 5 1 1 2 2 2 6 4 1 1 2 2 2 3 1 1
kaolini7 6 6 5 1 1 2 2 2 6 4 1 1 2 2 2 4 1 1
kaolini8 6 6 4 1 1 2 2 2 6 3 1 1 2 2 2 2 1 1
Ratio image/Mineral 7/3 8/3 9/3 5/4 6/4 7/4 8/4 9/4 6/5 7/5 8/5 9/5 7/6 8/6 9/6 8/7 9/7 9/8
muscovia 3 3 3 3 2 3 4 3 0 3 3 3 9 9 7 3 3 2
muscovib 5 4 4 2 1 3 3 2 0 2 2 2 9 9 7 3 1 2
muscovic 5 4 4 3 2 3 3 2 0 1 2 2 9 9 6 3 2 2
muscovid 6 5 5 5 3 5 5 4 0 2 3 4 8 8 7 5 4 2
kaolini1. 2 1 2 1 1 1 1 1 2 9 5 4 9 7 4 2 1 2
kaolini2 2 2 2 1 1 2 2 2 2 8 6 5 9 8 5 2 2 2
kaolini3 2 2 2 1 1 2 3 2 1 8 5 4 9 8 5 2 2 2
kaolini4 2 2 2 1 1 2 2 2 0 8 4 3 9 8 5 2 1 1
kaolini5 2 2 2 1 1 2 2 2 1 9 5 4 9 8 5 2 1 2
kaolini6 2 2 2 1 1 2 2 2 1 9 5 4 9 8 5 2 1 2
kaolini7 2 2 2 1 1 2 2 2 1 9 5 4 9 8 5 2 1 2
kaolini8 2 2 2 1 1 2 2 2 1 8 5 4 9 8 5 2 1 2

634
 TAYEBI et al. / Turkish J Earth Sci

X a xis : Ra tio ima g e numbe rs

1: 2 /1 5 : 6 /1 9 : 3 /2 1 3: 7 /2 1 7 : 5 /3 2 1 : 9 /3 2 5: 8 /4 29 : 8 /5 33 : 9 /6
2: 3 /1 6 : 7 /1 1 0: 4 /2 1 4 : 8 /2 1 8 : 6 /3 2 2: 5 /4 2 6 : 9 /4 3 0 : 9/5 3 4 : 8/7
3: 4 /1 7 : 8 /1 11 : 5 /2 1 5 : 9/2 1 9 : 7 /3 23 : 6 /4 2 7 : 6/5 3 1 : 7 /6 3 5 : 9 /7
4: 5 /1 8 : 9 /1 1 2: 6 /2 1 6 : 4 /3 2 0 : 8 /3 2 4: 7 /4 2 8 : 7 /5 3 2 : 8/6 3 6 : 9/8

10 10
9 259 expo.000.sco.rt 9 b muscovi 7
a
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Sample 259 (6 /5 7 /6 8 /6 ) Muscovite 7 (U SGS Library) (6 /5 7 /6 8 /6 )

10 10
9 c 218 expo.000.sco.rts 9 d kaolini
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Sample 218 (7 /5 7 /6 6 /5 ) Kaolinite (U SGS Library ) (7 /5 7 /6 6 /5 )

Figure 5. Spectral ratio code signature pattern of sample 259 (a), muscovite (sericite) (b), sample 218 (c), and kaolinite (d). The
best ratio images of sericite and kaolinite were selected (red circles) by comparing spectral ratio code signature pattern of field
spectra and USGS spectral library.

AND and XOR logic were used according to Eqs. (1), (2),
and (3) to create a training area layer. The intersection
(AND logic) method was performed on the ratio layers to
select and generate new layer (T sericite) for sericite and (T
kaolinite
) for kaolinite. This is an integration of 3 ratio layers
that preserves pixels that fall within the spatial extent
common to 3 ratio layers. Therefore, pixels that pass this
AND-logic method are pixels covered by the sericite that
have ratio codes 9, 0, and 9 and covered by kaolinite that
have ratio codes 9, 9, and 1. When a color ratio image is
made with display colors of red, green, and blue assigned
to the first, second, and third ratios, respectively, it yields
a color of yellow for kaolinite (red+green) and fuchsia/
purple for sericite (red+blue). The AND logic equations
for the sericite and kaolinite layers are given by Eqs. (1)
and (2):
T sericite = Rv7/6 ∩Rv6/5 ∩Rv8/6 or T sericite = Rv7/6 * Rv6/5* Rv8/6 the n-dimensional feature vector (Mather et al., 1998) or
(1)
T kaolinite = Rv7/5 ∩Rv7/6 ∩Rv6/5 or T kaolinite = Rv7/5 * Rv7/6* Rv6/5, image data sets) into a typically 2-dimensional array of
(2)
where T sericite is the training area layer for sericite, T kaolinite
is the training area layer for kaolinite, Rv i/j are the 3 ratio
value layers for each mineral, and i,j are the 2 spectral band
numbers used in the Ri/j spectral ratio. Then another GIS
overlay (XOR logic) method was used to create a training
area layer (TA) according to Eq. (3). This is an integration
of T sericite and T kaolinite that preserves pixels that fall within
the spatial extent not common to these 2 layers.
TA= (T sericite + T kaolinite) – (T sericite * T kaolinite), (3)
where TA is training area. The training area was then used
for training the SOM with all of the inputs.
3.2. Mapping alteration units
3.2.1. Self-organizing map (SOM) model
A SOM (Kohonen, 1995) is a 2-layer network that maps
reflectance bands (input layers, which is 5, 15, and 36 ratio
nodes referred as the competitive layer (output layer) or

635
 TAYEBI et al. / Turkish J Earth Sci

a b

)"")") ")")")")
")") )"
")
") ")")
)"") )"
") )"")
") ")")")
") ")
")
)"")")
")
)"
)" S a mple 259
") ")
")")
S a mple 218

)"")")")")")")")")")")
)"")
R a tio c o lo r c o mpo s ite R a tio c o lo r c o mpo s ite
(R GB : 7/6 8/6 6/5) (R GB : 7/5 7/6 6/5)
)"

S a mple s ite s

P lug s a nd dike s
c Alte re d ro c ks d e
Alte ra tio n units Alte re d ro c ks

La v a b re c c ia s

Figure 6. Color ratio images displayed as RGB with sampling sites as red dots in Figure 6a only. (a) Muscovite (sericite) is
displayed as yellow (RGB: 7/6 8/6 6/5), (b) Kaolinite is displayed as yellow (RGB: 7/5 7/6 6/5). Color field photos of the caldera
showing alteration units, pre- and postcaldera stage intrusive rocks (c, d, and e).

codebook vector map (Lee and Lathrop, 2006). Input and

output neurons are connected with each other by synaptic
weights between 0 and 1 (Mather et al., 1998), and, like (4)
a biological neural system, lateral interaction between
neurons is a common feature (Bian and Zhang, 2000). At where X ti is input to neuron i at iteration t, W ijt is the
first, synaptic weights between the output and input neurons synaptic weight from input neuron i to input neuron j
randomly assigned (0–1 range): a vector of reflectance for at iteration t, and j* is the index of the winner node. In
a single pixel as input vector is chosen at random from the cooperation stage the neighborhood of the winner
the set of training data. The SOM unsupervised training node is considered. The radius of this neighborhood
procedure can be divided into several steps: the competitive starts large. Typically it is set to be the radius of the
stage, the cooperative stage, and the adaptation stage. In network. According to a special neighborhood function
the competition stage the network is trained by the use some scalar values are attributed to all the lattice nodes.
of coarse tuning, which is an unsupervised classification This function is selected in such a way that the scalar
stage and leads to a fundamental regional organization (a values of the nodes are reversely proportional to their
topology) of neuron weights representing the underlying distances from the winner node. In the adaptation
clusters and subclusters in the input data. In fact for each stage, the weights of the winner node and its neighbors
input vector all the nodes are examined to indicate which within the radius γ are altered according to their scalar
nodes’ weights are most like the input vector. A node with values and the learning rate (a value between 0 and 1)
the smallest Euclidean distance to this input vector is (Ji, 2000; Tso and Mather, 2001). This adjustment of
called the “winner” node. The winner is determined by the weights proceeds for N iterations from global order to
use of Eq. (4): local adjustments.

636
 TAYEBI et al. / Turkish J Earth Sci
Then the next step is to tune the weight vectors in a
supervised mode. A majority voting procedure (Tso and
Mather, 2001) is used for labeling the training data to a
neuron that is assigned most frequently. On the other
hand, to refine the decision boundaries between training
classes, learning vector quantization (LVQ) (Kangas et al.,
1990; Ji, 2000; Tso and Mather, 2001) is used. In vector
quantization, the input space is partitioned into a number
of distinct regions and for each region a reconstruction
vector is defined. The collection of reconstruction vectors
is called the code book. A new input is then represented
by the reconstruction vector representing the region it
falls into. Accordingly, the weights of the neurons with the
most similar weights to the input vector of reflectance are
adjusted as notified in Eqs. (5)–(7) below:
Wct+1 = Wct + d t (Xi – Wct) If the class of the input vector
and the class of the winner neuron’s weight vector match area was divided into a 60 × 60-m grid pattern and the
(5)
Wct+1 = Wct – d t (Xi – Wct) If the class of the input vector
and the class of the winner neuron’s weight vector do not
match (6)
Wit+1 = Wit other neurons (7)
where Wc is weight vector of the winner, and d t is a gain
term that will decrease with time and is within the range of
0 and 1, Xi is input vector to the network, and C is the class.
In this manner, weights move closer to an input feature
if they match the label of the corresponding winner (Eq.
(4)), while moving farther from it if it does not (Eq. (5)) (Ji,
2000; Tso and Mather, 2001). In this study, we implemented
LVQ1, in which the known patterns (from training sites)
are fed again to the SOM, and finally classifying an image
to generate feature maps.
4. Accuracy assessment
The test data were selected independently from the
training data, based on the field observations. In each case,
an overall accuracy, producer and user accuracies, kappa
coefficient, and errors of commission and omission were
measured and determined. The overall accuracy (OA) is
calculated by using Eq. (8):
(8)
where A is the number of pixels assigned to the correct
class and B is the number of pixels that actually belong
to that class. The kappa coefficient (K) (Cohen, 1960) is
a statistical measure of the agreement (Chen et al., 2007),
and is calculated by the use of Eq. (9):
(9)
where N is the total number of pixels in all the testing
samples, ∑k Xkk is the sum of the confusion matrix diagonal
(Xkk), Xk∑ is the sum of the testing samples in class k, and
X∑k is the sum of the predicted pixels in class k (Cohen,
1960). The kappa coefficient is always in the range of –1
to 1, where the latter indicates complete agreement, and
is often multiplied by 100 to give a percentage measure
of classification accuracy. Meanwhile, the fraction cover
of surface materials was also assessed by comparing the
fraction cover of surface materials estimated in the field
to the corresponding pixel fraction values in the image
data using the coefficient of determination (R2). The study
proportion of minerals was calculated for each site using
the Pima view software. Mineral fraction maps generated
from the SOM model were resampled to the 60 × 60-m
pixels, and the R2 was calculated.
5. Results and discussion
Thirty-six nonreciprocal spectral ratios of 9 ASTER
reflective bands were created for key minerals including
sericite and kaolinite. Spectral ratio code signature
patterns of field samples were then compared to the USGS
spectral library to find 3 spectral ratio images for each key
mineral. Based on these 3 ratio images, color composite
ratio images were created and training areas were selected
by a GIS overlay approach and were used to train the SOM
model.
The best spectral ratios for mapping sericite were 7/6,
6/5, and 8/6, and for kaolinite were 7/5, 7/6, and 6/5. Figure
6 shows the results of displaying 2 spectral ratios with high
ratio code in red and green and 1 spectral ratio with very
low ratio codes as blue for the sericite and kaolinite. In
each image, the mineral of interest should be yellow, if it
is present in the imaged scene (Figures 6a and 6b). The
SOM was then applied to the 5, 15, and 36 ratio image data
sets (Table 1), separately, for subpixel classification of the
Masahim volcano.
Sampling intervals, the times between samples, were
set to 3 and 7, respectively. The training area was used
to define the categories that should be classified. A 15 ×
15 array of nodes was selected as output layer nodes. The
minimum and maximum learning rates were adjusted
between 0.5 and 1.0. The value for setting the fine tuning
rate including the minimum and maximum gain terms
were in the range of 0.0001–0.0005. Fine tuning was done
by using the LVQ algorithm to improve classification
accuracy by defining the class boundaries in the output
637
 TAYEBI et al. / Turkish J Earth Sci

layer more precisely. Pixels within the training areas were
again presented to the SOM and the output neuron with
the minimum Euclidean distance between a training pixel
and its weight vector was selected as the best matching
unit. The feature maps of alteration units of the volcano
derived from 5, 15, and 36 ratio image data sets are shown
in Figures 7–9. Maps include the fraction and spatial
distribution of sericite (phyllic zone) and kaolinite (argillic
zone) in the Masahim volcano. The zoomed feature maps
of the caldera are likewise shown.
The result of the SOM applied to the 5 ratio images
showed that the fractions of sericite at the caldera mainly
consisted of 62%–75% (yellow pixels) and 87%–100% (red
pixels) (Figure 7b), and the fraction of kaolinite mainly
consisted of 75%–87% (green pixels) and 87%–100% (red
pixels) (Figure 7d). The result of SOM applied to the 15
ratio image showed that the fraction image of sericite at
the caldera mainly consisted of 75%–87% (green pixels)
and 87%–100% (red pixels) (Figure 8b). However, the
fraction image of kaolinite at the caldera mainly consisted
of 50%–62% (magenta pixels) and 62%–75% (red pixels)
(Figure 8d). The result of SOM applied to the 36 ratio
image showed that the fraction image of sericite at the
caldera mainly consists of 62%–75% (yellow pixels) and
87%–100% (red pixels, Figure 9b). Meanwhile, the fraction
image of kaolinite at the caldera mainly consists of 37%–
50% (cyan pixels, Figure 9d).
To evaluate the results of SOM classifications maps,
the accuracy of the alteration units was assessed using
ground reference information by determining the overall,
user’s, and producer’s accuracies and kappa coefficients
(Tables 3–5). The overall accuracy of the SOM applied
to the 36, 15, and 5 ratio image data sets was 87%, 86%,
and 85%, respectively. The confusion matrix of the SOM
applied to the 36 ratio image data set (Table 3) shows good
accuracy for sericite and kaolinite. Errors of commission
and omission (errors of C and O) for sericite in the SOM
applied to the 36 ratio image data set (Table 3) were 0.07
and 0.15, and the user accuracy and producer accuracy
were 93% and 85%, respectively. Sericite has 15% error of
O and 7% error of C with kaolinite, which again means that
15% of sericite pixels are omitted from the sericite class

b
0.37 to 0.50
0.50 to 0.62
0.62 to 0.75
0.75 to 0.87
a b
0.87 to 1.00

d
0.37 to 0.50
0.50 to 0.62
0.62 to 0.75
0.75 to 0.87
c d
0.87 to 1.00

Figure 7. Feature and zoomed feature maps of Masahim volcano derived from 5 ratio images. Maps include the fractions
of sericite (a and b) and kaolinite (c and d). The yellow solid line indicates the geologic unit boundaries (Figure 1).

638
 TAYEBI et al. / Turkish J Earth Sci

b
0.37 to 0.50
0.50 to 0.62
0.62 to 0.75
0.75 to 0.87
a 0.87 to 1.00 b

d
0.37 to 0.50
0.50 to 0.62
0.62 to 0.75
0.75 to 0.87
c d
0.87 to 1.00

Figure 8. Feature and zoomed feature maps of Masahim volcano derived from 15 ratio images. Maps include the fraction
of sericite (a and b) and kaolinite (c and d). The yellow solid line indicates the geologic unit boundaries (Figure 1).

and 7% of the pixels that should be included as sericite
are assigned to kaolinite. User accuracy and producer
accuracy for kaolinite were 79% and 90% respectively,
with 15% error of O and 7% error of C with sericite. The
results also indicated a kappa coefficient of 0.73 for the
SOM (Table 3).
Errors of C and O for sericite in the SOM applied to
the 15 ratio were 0.12 and 0.10, respectively (Table 4).
Meanwhile, user accuracy and producer accuracy were
88% and 90%, respectively. Sericite has 10% error of O and
12% error of C with kaolinite, which means that 10% of
sericite pixels are omitted from the sericite class and 12%
of the pixels that should be included as sericite are assigned
to kaolinite. Errors of C and O for kaolinite were 0.17 and
0.20, and the user accuracy and producer accuracy were
80% and 83%, respectively. Moreover, kaolinite showed
20% error of O and 17% error of C with sericite. The
results also indicated that the operating SOM has a kappa
coefficient of 0.70 (Table 4).
The confusion matrix of the SOM applied to the 5 ratio
image data set (Table 5) showed good accuracies for sericite
and kaolinite. For example, errors of C and O for sericite
were 0.05 and 0.20, and the user and producer accuracies
of that were 95% and 80%, respectively. Sericite has 20%
error of O and 5% error of C with kaolinite, which means
that 20% of sericite pixels were omitted from the sericite
class and 5% of the pixels that should have been included
as sericite were assigned to kaolinite. Errors of C and O
for kaolinite were 0.26 and 0.07, and its user and producer
accuracies were 74% and 93%, respectively. Kaolinite
has 7% error of O and 26% error of C with sericite. The
results also indicated that the operating SOM has a kappa
coefficient of 0.69 (Table 5).
The most accurate classification is achieved by applying
the SOM to the 15 and 36 ratio data sets (Figures 8 and 9).
Output results of the SOM applied to the 15 and 36 ratio
image data sets generally enhanced similar areas spatially.
However, the fractions of pixels in SOM applied to 15 ratio
image data sets are higher than the output result of 36 ratio
image data sets (Figures 8 and 9). The accuracy of fractions
estimated by the SOM applied to 15 and 36 ratio image
data sets were also assessed using R2 coefficient. A total of

639
 TAYEBI et al. / Turkish J Earth Sci

b
0.37 to 0.50
0.50 to 0.62
0.62 to 0.75
0.75 to 0.87
a b
0.87 to 1.00

d
0.37 to 0.50
0.50 to 0.62
0.62 to 0.75

c 0.75 to 0.87 d
0.87 to 1.00

Figure 9. Feature and zoomed feature maps of Masahim volcano derived from 36 ratio images. Maps include the fraction of
sericite (a and b) and kaolinite (c and d). The yellow solid line indicates the geologic unit boundaries (Figure 1).

Table 3. Confusion matrix of SOM for the classification image of 36 ratio images input.

Ground truth

Phyllic Argillic Error User’s

Classified Total
(Sericite) (Kaolinite) commission accuracy
Phyllic (Sericite) 414 30 444 0.07 0.93
Argillic (Kaolinite) 71 270 341 0.20 0.79
Total 485 300 785
Error omission 0.15 0.10 0.13
Producer’s accuracy 0.85 0.90
Overall accuracy 0.87
Kappa coefficient 0.73

40 sites of 60 m × 60m were sampled and used for this
assessment. The scatter plots of the accuracy assessment
are shown in Figure 10. The R2 coefficient for sericite was
nearly 0.35, and for kaolinite was nearly 0.52 in the SOM
applied to the 15 ratio images. Moreover, sericite showed
0.64 and kaolinite 0.68 of R2 coefficients in the SOM
applied to the 36 ratio images.
Triplets of spectral ratio images displayed in colors of
red, green, and blue (as a ratio color composite) were used
to find training areas. When these 3 colors are combined,

640
 TAYEBI et al. / Turkish J Earth Sci

Table 4. Confusion matrix of SOM for the classification image of 15 ratio images input.

Ground truth

Phyllic Argillic Error User’s

Classified Total
(Sericite) (Kaolinite) commission accuracy
Phyllic (Sericite) 437 61 498 0.12 0.88
Argillic (Kaolinite) 48 239 287 0.17 0.83
Total 485 300 785
Error omission 0.10 0.20 0.14
Producer’s accuracy 0.90 0.80
Overall accuracy 0.86
Kappa coefficient 0.70

Table. 5. Confusion matrix of SOM for the classification image of 5 ratio images input.

Ground truth

Phyllic Argillic Error User’s

Classified Total
(Sericite) (Kaolinite) commission accuracy
Phyllic (Sericite) 386 20 406 0.05 0.95
Argillic (Kaolinite) 99 280 379 0.26 0.74
Total 485 300 785
Error O 0.20 0.07 0.15
Producer’s accuracy 0.80 0.93
Overall accuracy 0.85
Kappa coefficient 0.69

1.2 1.2 y = 0.9304x + 0.0977

R 2 = 0.5224
1
a y = 0.775x + 0.0053 b
1
Estimated Fraction

R 2 = 0.3477
0.8 0.8
0.6 0.6
0.4 0.4
2
0.2
R2 =0.3477 0.2
R =0.5224

0 0
0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Actual Fraction (sericite) Actual Fraction (kaolinite)
1.2 y = 0.8391x - 0.0159 1.2
y = 0.9462x + 0.0671
1
c R 2 = 0.6413 d
1 R 2 = 0.6841
Estimated Fraction

0.8 0.8
0.6 0.6
0.4 0.4

0.2 2 2
R =0.6413 0.2 R =0.6841
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Actual Fraction (sericite) Actual Fraction (kaolinite)
Figure 10. Scatter plots of accuracy assessment results for SOM applied to the 15 (a and b) and 36 (c and d) ratio image data sets.

641
 TAYEBI et al. / Turkish J Earth Sci
they produce different colors in the visible spectrum. The
resulting color ratio image generally enhances specific
minerals of interest in different colors. This mineral may
appear in shades of blue, green, and red or various color
combinations of the 3, depending on which ratios are
employed to yield the color ratio image. More pointedly,
we can use a ratio code method to determine which
spectral ratios are very high and very low in values for a
given mineral, compared to other minerals and rocks for
which we have an archived spectral library of mineral and
rock reflectances, and we can create a spectral ratio image
in which we assign the red display color to the highest
spectral ratio for a mineral that we wish to highlight,
and assign the blue and green display colors to 2 spectral
ratios that are very low for that mineral. The resulting
target mineral is displayed as red, and almost all and
sometimes all other surface areas are displayed as other
colors, except red. Red in the scene, for that ratio-code-
selected color ratio composite, is the mineral we seek to
find (the target mineral). Then it is possible to input dark-
object–corrected spectral ratio imaging data sets to the
SOM model that are rid of the topographic effects to map
and estimate abundance fractions of minerals of interest.
While this method can identify minerals of interest, more
study areas need to be tested in the future.
6. Conclusion
The spatial distribution of altered units at the Masahim
volcano was determined using ratio image/ratio codes
and a SOM neural network model. The result of the SOM
References
Ahmadipour H (1994). Dynamism of Volcanic Activities in
Mosahim volcano (Shahr-e-Babak). MSc, Department of
Geology, Shahid Bahonar Univerity, Kerman, Iran.
Bagan H, Wang QX, Watanabe M, Yang YH, Jianwen M (2005).
Land cover classification from MODIS EVI times-series
data using SOM neural network. Int J Remote Sens 26:
4999–5012.
Bedini E (2009). Mapping lithology of the Sarfartoq carbonatite
complex, southern West Greenland, using HyMap imaging
spectrometer data. Remote Sens Environ 113: 1208–1219.
Bedini E (2012). Mapping alteration minerals at Malmbjerg
molybdenum deposit, central East Greenland, by Kohonen
self-organizing maps and matched filter analysis of HyMap
data. Int J Remote Sens 33: 939–961.
Bian ZQ, Zhang XG (2000). Pattern Recognition, 2nd ed, Beijing,
China: Tsinghua University Press.
Chen X, Warner TA, Campagna DJ (2007). Integrating visible,
near-infrared and short-wave infrared hyperspectral and
multispectral thermal imagery for geological mapping at
Cuprite, Nevada. Remote Sens Environ 110: 344–356.
642
applied to 36, 15, and 5 ratio image data sets showed good
identification of phyllic and argillic units. We achieved
overall accuracies of 87% (kappa 0.73) for SOM applied to
the 36, 86% (kappa 0.70) for 15, and 85% (kappa 0.69) for 5
ratio image data sets. The results also showed that the SOM
applied to 15 and 36 data sets generally enhanced similar
areas spatially. However, the output results showed some
differences in the fractions of minerals within each pixel.
The R2 coefficient for sericite was 0.35 and for kaolinite was
0.52 in the SOM applied to the 15 ratio images. Sericite
showed 0.64 and kaolinite 0.68 of R2 coefficients in the
SOM applied to the 36 ratio images.
It is concluded that coded spectral ratio imaging is
efficient for selecting training areas for a given mineral or
rock, especially when we do not have access to the area and
we only have access to the archived spectral library spectra.
Meanwhile, the SOM neural network model applied to
36 ratio image data sets can also be effective for mapping
alteration minerals, yielding estimates of pixel fractions
covered by each target mineral of interest. However, more
study areas need to be tested in the future, and the search
for a better method than the ratio code method for finding
the most characteristic spectral parameters of target
minerals should be sought.
Acknowledgments
The authors thankfully acknowledge the constructive
suggestions received from the TJES’s editor and 2
anonymous reviewers.
Clark RN, Swayze GA, Gallagher A, King TVV, Calvin WM (1993).
The U.S. geological survey digital spectral library: Version 1:
0.2 to 3.0 µm. USGS Open File Report 93–592, U.S. Geological
Survey, p. 1340.
Cohen J (1960). A coefficient of agreement for nominal scales. Educ
Psychol Meas 20: 37– 44.
Crosta AP, Souza Filho CR, Azevedo F, Brodie C (2003). Targeting
key alteration minerals in epithermal deposits in Patagonia,
Argentina, using ASTER imagery and principal component
analysis. Int J Remote Sens 24: 4233–4240.
Crowley JK, Zimbelman DR (1997). Mapping hydrothermally altered
rocks on Mount Rainier, Washington, with Airborne Visible/
Infrared Imaging Spectrometer (AVIRIS) data. Geology 25:
559–562.
Crowley JK, Hubbard BE, Mars JC (2003). Analysis of potential
debris flow source areas on Mount Shasta, California, by
using airborne and satellite remote sensing data. Remote Sens
Environ 87: 345–358.
 TAYEBI et al. / Turkish J Earth Sci
Cudahy T, Hewson R (2002). ASTER geological case histories:
porphyry-skarnepithermal, iron oxide Cu-Au and Broken hill
Pb-Zn-Ag. Annual General Meeting of the Geological Remote
Sensing Group ‘ASTER Unveiled’, Burlington House, Piccadilly,
London, UK.
Dimitrijevic M (1973) Geology of Kerman region. (Iran Geological
Survey Report No. Yu/52.) Institute for Geological and Mining
Exploration and Institution of Nuclear and Other Mineral Raw
Materials, Beograd Yugoslavia, p334.
DuBose P, Klimasauskas C (1989). Introduction to Neural Networks
with Examples and Applications.Pittsburgh, PA, USA:
NeuralWare Inc.
Ehsani AH, Quiel F (2008). Geomorphometric feature analysis using
morphometric parameterization and artificial neural networks.
Geomorphology 99: 1–12.
Emaruchi B, Jin Y, Sauchyn D, Fuller G (1997). Estimation of runoff
index using Landsat (TM) data and an artificial neural network
system. Can Water Resour J 22: 155–166.
Fujisada H (1995). Design and performance of ASTER instrument.
In: Fujisads H, Sweeting MN, editors. Proceedings of SPIE
(International Society for Optical Engineering), 25–28
September 1995, Paris, France, 2583, pp.16–25.
Geaga JV, Le TC, Sailer CT (1992). Neural net spectral pattern
recognition. In: Proc. SPIE, 1766. SPIE, Bellingham, WA: pp.
609–618.
Gomez C, Delacourt C, Allemand P, Ledru P, Wackerle R (2005).
Using ASTER remote sensing data set for geological mapping,
in Namibia. Phys Chem Earth 30: 97–108.
Grebby S, Cunningham D, Naden J, Tansey K (2010). Lithological
mapping of the Troodos ophiolite,  Cyprus,  using airborne
LiDAR topographic data. Remote Sens Environ 114: 713–724.
Gupta RP (2003). Remote Sensing Geology. Berlin, Germany:
Springer-Verlag.
Hassanzadeh J (1993). Metallogenic and tectonomagmatic events
in the SE sector of the Cenozoic active continental margin
of central Iran (Shahr e Babak area, Keman Province). PhD,
University of California, Los Angeles, Los Angeles.
Hellman MJ, Ramsey MS (2004). Analysis of hot springs and
associated deposits in Yellowstone National Park using ASTER
and AVIRIS remote sensing. J Volcanol Geotherm Res 135:
195–219.
Hosseinjani M, Tangestani MH (2011). Mapping alteration minerals
using sub-pixel unmixing of ASTER data in the Sarduiyeh area,
SE Kerman, Iran. Int J Digital Earth 4: 487–504.
Hu X, Weng Q (2009). Estimating impervious surfaces from medium
spatial resolution imagery using the self-organizing map and
multi-layer perceptron neural networks. Remote Sens Environ
113: 2089–2102.
Hubbard BE, Sheridan MF, Carrasco-Núñez G, Díaz-Castellón R,
Rodríguez SR (2007). Comparative lahar hazard mapping at
Volcan Citlaltépetl, Mexico using SRTM, ASTER and DTED-1
digital topographic data. J Volcanol Geotherm Res 160: 99–124.
Hunt GR, Salisbury JW (1970) Visible and near-infrared spectra of
minerals and rocks, I. Silicate minerals. Mod Geol 1: 283–300.
Ji CY (2000). Land-use classification of remotely sensed data
using Kohonen self organizing feature map neural network.
Photogram Eng Remote Sensing 66: 1451–1460.
Jianwen M, Bagan H (2005). Land-use classification using ASTER
data and self organized neutral networks. Int J Appl Earth
Obs 7: 183–188.
Joyce K, Samsonov S, Manville V, Jorgens R, Graettinger A, Cronin
S (2009). Remote sensing data types and techniques for lahar
path detection: a case study at Mt Ruapehu, New Zealand.
Remote Sens Environ 113: 1778–1786.
Kangas JA, Kohonen TK, Laaksonem JT (1990). Variants of Self-
Organizing Maps. IEEE T Neural Networ 1: 93–99.
Kerle N, Froger JL, Oppenheimer C, De Vries BV (2003). Remote
sensing of the 1998 mudflow at Casita volcano, Nicaragua. Int
J Remote Sens 24: 4791−4816.
Kohonen T (1995). Self-Organizing Maps. Berlin, Germany:
Springer.
Kruse FA (1988). Use of Airborne Imaging Spectrometer data to
map minerals associated with hydrothermally altered rocks in
the northern Grapevine Mountains, Nevada and California.
Remote Sens Environ 24: 31–51.
Lee S, Lathrop RG (2006). Subpixel analysis of Landsat ETM+ using
Self-Organizing Map (SOM) neural networks for urban land
cover characterization. IEEE T Geosci Remote 44: 1642–1654.
Li Z, Eastman JR (2006). The nature and classification of unlabelled
neurons in the use of Kohonen’s self-organizing map for
supervised classification. Trans GIS 10: 599–613.
Mars JC, Rowan LC (2006). Regional mapping of phyllic- and
argillic-altered rocks in the Zagros magmatic arc, Iran, using
Advanced Spaceborne Thermal Emission and Reflection
Radiometer (ASTER) data and logical operator algorithms.
Geosphere 2: 161–186.
Mars JC, Rowan LC (2010). Spectral assessment of new ASTER
SWIR surface reflectance data products for spectroscopic
mapping of rocks and minerals. Remote Sens Environ 114:
2011–2025.
Mather PM, Tso B, Koch M (1998). An evaluation of Landsat
TM spectral data and SAR-derived textural information for
lithological discrimination in the Red Sea Hills, Sudan. Int. J.
Remote Sens 19: 587–604.
Perry JL, Vincent RK (2009). ASTER Brightness and Ratio Codes
for Minerals: Application to Lithologic Mapping In West-
Central Powder River Basin, Wyoming. Rev Econ Geol 16:
143–168.
Rajendran S, Thirunavukkarasu A, Balamurugan G, Shankar K
(2011). Discrimination of iron ore deposits of granulite
terrain of Southern Peninsular India using ASTER data. J
Asian Earth Sci 41: 99–106.
Research Systems Inc (2003). ENVI Tutorial, ENVI Software
Package Version 4.0.
643
 TAYEBI et al. / Turkish J Earth Sci
Rowan L, Wetlaufer PH, Goetz AFH, Billingsley FC, Stewart
JH (1974). Discrimination of rock types and detection of
hydrothermally altered areas in south-central Nevada by use
of computer-enhanced ERTS images. US Geological Survey
Professional, Paper, 883.
Rowan LC, Mars JC (2003). Lithologic mapping in the Mountain
Pass, California, area using Advanced Spaceborne Thermal
Emission and Reflectance Radiometer (ASTER) data. Remote
Sens Environ 84: 350–366.
Rowan LC, Hook SJ, Abrams MJ, Mars JC (2003). Mapping
hydrothermally altered rocks at Cuprite, Nevada using the
Advanced Spaceborne Thermal Emissivity and Reflection
Radiometer ASTER. A new satellite-imaging system. Econ
Geol 98: 1019–1027.
Schowengerdt RA (2007). Remote Sensing: Models and Methods for
Image Processing. New York, NY, USA, Academic Press.
Souza Filho CR, Tápia CH, Crósta AP, Xavier RP (2003). Infrared
spectroscopy and ASTER imagery analysis of hydrothermal
alteration zones at the Quellaveco Porphyry-Cooper Deposit,
Southern Peru. Proceedings of the American Society of
Photogrammetry and Remote Sensing (ASPRS), Annual
Conference, Technology: Converging at the Top of the World,
pp.1−12.
Tangestani MH, Mazhari M, Agar B, Moore F (2008). Evaluating
Advanced Spaceborne Thermal Emission and Reflection
Radiometer (ASTER) data for alteration zone enhancement in
a semiarid area, northern Shahr-e-Babak, SE Iran. Int J Remote
Sens 29: 2833–2850.
644
Terunuma T, Nishida K, Amada T, Mizuyama T, Sato I, Urai M
(2005). Detection of traces of pyroclastic flows and lahars
with satellite synthetic aperture radars. Int J Remote Sens 26:
1927−1942.
Tso B, Mather PM (2001). Classification Methods for Remotely
Sensed Data. New York, NY, USA: Taylor and Francis.
Villmann T, Merenyi E, Hammer B (2003). Neural maps in remote
sensing image analysis. Neural Networks 16: 389–403.
Vincent RK (1972). An ERTS multispectral scanner experiment
for mapping iron compounds. International Symposium on
Remote Sensing of Environment, 8th, Ann Arbor, University
of Michigan, Proceedings, 2, pp. 1239–1247.
Vincent RK (1997). Fundamentals of Geological and Environmental
Remote Sensing, , Upper Saddle River, NJ, USA, Prentice Hall.
Yamaguchi Y, Naito C (2003). Spectral indices for lithologic
discrimination and mapping by using the ASTER SWIR bands.
Int J Remote Sens 24: 4311–4323.
Zhang X, Pamer M, Duke N (2007). Lithologic and mineral
information extraction for gold exploration using ASTER
data in the south Chocolate Mountains (California). Isprs J
Photogramm 62: 271–282.
Zhuang X, Engel BA, Baumgardner MF, Swain PH (1991). Improving
classification of crop residues using digital land ownership data
and LANDSAT TM Imagery. Photogram Eng Remote Sensing
57: 1487–1492.