Stacked vector multi-source lithologic classification utilizing Machine Learning Algorithms: Data potentiality and dimensionality monitoring

Abstract

Machine Learning Algorithms (MLAs) have recently introduced considerable lithologic mapping. Thus, this study scrutinizes the efficacy of Artificial Neural Network (ANN), Maximum Likelihood Classifier (MLC) and Support Vector Machine (SVM) over hybrid datasets including optical (Sentinel 2, ASTER, Landsat OLI and Earth-observing 1 Advanced Land Imager (ALI)), radar (Sentinel 1 and ALOS PALSAR), DEMs and their derivatives (Slope, and Aspect). The study aims to (1) monitor the effect of data dimensionality in enhancing categorization accuracy. (2) disclose the most efficient MLA and most powerful dataset in labeling rock units accurately. (3) highlight the impact of embedding topographical and radar data in lithologic classification. (4) outline the best relation between the number of training pixels and number of utilized bands, in delivering reliable allocation. To achieve these aims, we selected training and testing pixels meticulously, in concordance with a recently published geological map of the study area. We adopted a stacked vector approach for handling the implemented multi-sensor data. Results show that diversifying information sources raised the classification accuracy by approximately 10% for each classifier. SVM and MLC are much better than ANN. Slope is better than aspect and both are less qualified when compared to DEM. Sentinel 1 (C-band) and ALOS PALSAR (L-band) effects are not so different whatever the implemented polarization. Landsat OLI is less qualified in lithologic classification when compared to Sentinel 2, ASTER and ALI. The utilized training pixels should be at least 30N for (N) channels submitted to the classifiers.

Full text

Remote Sensing Applications: Society and Environment 24 (2021) 100643
Available online 2 October 2021
2352-9385/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Stacked vector multi-source lithologic classication utilizing Machine
Learning Algorithms: Data potentiality and dimensionality monitoring
Ali Shebl
a
,
b
,
*
, ´
Arp´
ad Cs´
amer
a
a
Department of Mineralogy and Geology, University of Debrecen, Hungary
b
Department of Geology, Tanta University, Egypt
ARTICLE INFO
Keywords:
Lithologic classication
Support vector machine
Articial neural network
Maximum likelihood classier
Sentinel 2
Sentinel 1
ASTER
ABSTRACT
Machine Learning Algorithms (MLAs) have recently introduced considerable lithologic mapping. Thus, this study
scrutinizes the efcacy of Articial Neural Network (ANN), Maximum Likelihood Classier (MLC) and Support
Vector Machine (SVM) over hybrid datasets including optical (Sentinel 2, ASTER, Landsat OLI and Earth-observing
1 Advanced Land Imager (ALI)), radar (Sentinel 1 and ALOS PALSAR), DEMs and their derivatives (Slope, and
Aspect). The study aims to (1) monitor the effect of data dimensionality in enhancing categorization accuracy.
(2) disclose the most efcient MLA and most powerful dataset in labeling rock units accurately. (3) highlight the
impact of embedding topographical and radar data in lithologic classication. (4) outline the best relation between
the number of training pixels and number of utilized bands, in delivering reliable allocation. To achieve these aims,
we selected training and testing pixels meticulously, in concordance with a recently published geological map of the
study area. We adopted a stacked vector approach for handling the implemented multi-sensor data. Results show that
diversifying information sources raised the classication accuracy by approximately 10% for each classier. SVM
and MLC are much better than ANN. Slope is better than aspect and both are less qualied when compared to DEM.
Sentinel 1 (C-band) and ALOS PALSAR (L-band) effects are not so different whatever the implemented polarization.
Landsat OLI is less qualied in lithologic classication when compared to Sentinel 2, ASTER and ALI. The utilized
training pixels should be at least 30N for (N) channels submitted to the classiers.
1. Introduction
Accurate geological mapping is the key for economic minerals
localization and geo-hazards mitigations. Enormous improvements in
lithologic mapping have been introduced by the availability of a wide
range of remote sensing datasets (Grebby et al., 2011). The digital
format of the latter exceedingly prompted Machine Learning Algorithms
to be applied in land cover and geologic mapping; using multispectral
and/or hyperspectral data (Foody and Mathur, 2004; Ham et al., 2005;
Lee et al., 2012; Liesenberg and Gloaguen, 2012; Pal and Mather, 2005).
Efcient lithologic classication was a matter of interest for several re-
searchers over the last decade. The classier performance may vary
depending on the implemented datasets, thus a wide range of datasets
are implemented for mapping rock units using several classiers. Most of
the results highlighted SVM (Bachri et al., 2019; Bentahar and Raji,
2021; Kumar et al., 2021; Manap and San, 2018), MLC (Fatima et al.,
2013; Ge et al., 2018; Hadigheh and Ranjbar, 2013; Jellouli et al., 2016;
Mehr et al., 2013), and ANN (Grebby et al., 2011; He et al., 2015). Thus,
we decided to intensively investigate the performance of these classiers
over a wide range of the utilized data sets. On the other hand and for
datasets, previous studies highly recommend using ASTER in lithologic
mapping (Jellouli et al., 2016; Othman and Gloaguen, 2014, 2017; Yu
et al., 2012), others (Bentahar and Raji, 2021; Fatima et al., 2013; He
et al., 2015) utilized Landsat data effectively. With the advent of the
recently launched Sentinel 2 (S2) data, several studies highlighted the
advantages of S2 in categorization rock units due to its higher spectral
resolutions (Bentahar and Raji, 2021; Ge et al., 2018). However (Pour
and Hashim, 2014), stated that ALI is an efcient source for lithological
and mineralogical mapping due to its excellent VNIR spectral separa-
bility. Thus and for the rst time, this study implied the four widely used
sensors in lithologic mapping.
With the proven role of geomorphological and topographical data in
enhancing lithological allocation (Bachri et al., 2019; Grebby et al.,
2011; Othman and Gloaguen, 2017; Yu et al., 2012), the current study
embedded not only DEM but also its derivatives (slope and aspect maps)
to inspect their role in identifying rock units. Recently (Kumar et al.,
* Corresponding author. Department of Mineralogy and Geology, University of Debrecen, Hungary.
E-mail address: ali.shebl@science.tanta.edu.eg (A. Shebl).
Contents lists available at ScienceDirect
Remote Sensing Applications: Society and Environment
journal homepage: www.elsevier.com/locate/rsase
https://doi.org/10.1016/j.rsase.2021.100643
Received 14 June 2021; Received in revised form 22 August 2021; Accepted 29 September 2021

Remote Sensing Applications: Society and Environment 24 (2021) 100643
2
2021), utilized sentinel 1 data in geological mapping, thus Sentinel 1
radar data is also considered in this study to enhance the mapping
process. Authors of (Kumar et al., 2021) highly recommend the inte-
gration of several sensors for superb geological mapping especially with
MLAs. Consequently, and for the rst time, This study integrated
Sentinel 2, ASTER, Landsat OLI, and Earth-observing 1 Advanced Land
Imager, Sentinel 1, ALOS PALSAR-1, ALOS PALSAR DEM, Slope, and
Aspect to map 9 lithological units. Moreover, and due to a large number
of the included bands the study closely monitors the effect of data
dimensionality on the classier performance and overall accuracy.
Furthermore, a detailed examination of the effect of topographic and
radar data when added to each optical sensor is inspected to highlight
the reliable data combinations for delivering superior outputs, through
performing more than 100 classication process. The investigated area
was selected for this work as it is famous for gold mining and have
various rock units, as well as, it has a recently published geological map
which is useful for comparing the results and verication.
2. Materials and methods
2.1. Study area description
Um Salatit –Mueilha is located at the extreme southern part of the
Central Eastern Desert (CED), Egypt as shown in Fig. 1. It extends be-
tween latitudes 24◦40′′ to 25◦20′′ N and longitudes 33◦45′′ to 34◦05′′ E,
and covered by an intricate complex of Precambrian rugged igneous and
metamorphic rocks, in addition to, Phanerozoic sedimentary Nubian
sandstone. Authors of (Zoheir et al., 2019a) stated that the area is
covered mainly by a widely distributed stretch of Neoproterozoic
ophiolitic m´
elange constituted mainly of allocthonous ophiolitic frag-
ments embedded in a sheared matrix, as well as, other different map-
pable units (Fig. 1). The vastly disseminated unit is the ophiolitic
m´
elange. The other mappable units in the area include metavolcanics,
metagabbro-diorites and granitic rocks.
2.2. Data characteristics and preprocessing
The widely used Landsat-8 (L8) was launched on February 11, 2013,
toting two sensors namely OLI (Operational Land Imager) and TIRS
(Thermal Infrared Sensor) to acquire spectral data in the visible and
Near-Infrared (VNIR), Short-Wave Infrared (SWIR) and Thermal
Infrared (TIR) regions. OLI data are recorded in nine spectral bands,
while TIRS data give information only in two bands (Roy et al., 2014), as
shown in Table 1. Advanced Spaceborne Thermal Emission and
Reection Radiometer (ASTER) are commonly used in lithologic
discrimination (Pour and Hashim, 2014), and detect radiance in four-
teen bands covering spectral areas from VNIR, SWIR, and TIR regions as
shown in Table 1. Advanced Land Imager (ALI) sensor of Earth
Observing-1 satellite, recorded data in ten spectral bands (Czapla-Myers
et al., 2016), as shown in Table 1. ALI operates in a pushbroom fashion,
with a spatial resolution of 30 m for the multispectral bands, and 10 m
for the panchromatic band. The standard scene width is 37 km, while the
standard scene length is 42 km. Sentinel 2 (S2) was developed by the
European Space Agency (ESA) and launched (S2A) on June 23, 2015, to
provide spectral data in 13 bands in various spectral regions and vari-
able spatial resolutions (10, 20, or 60m) (Drusch et al., 2012), as shown
in Table 1.
Radar data includes Sentinel-1, the premier satellite of the Coper-
nicus Programme satellite constellation conducted by ESA. The rst
satellite, Sentinel-1A, was launched on April 3, 2014, and Sentinel-1B
was launched on April 25, 2016. They carry a C-band (λ =5.6 cm, f
=5.4 GHz) synthetic-aperture radar (SAR) instrument, which provides a
collection of data in all weather conditions, day or night. Sentinel-1A of
Interferometric Wide (IW) mode (VH, VV polarizations), Ground Range
Detected High Resolution (GRDH) is obtained from ESA to fulll the aim
of this study. Phased Array L-type band Synthetic Aperture Radar
(PALSAR) device was on the board of Advanced Land Observing Satellite
(ALOS), which was launched on January 24, 2005. PALSAR sensor is an
active microwave sensor. Fine Beam Double polarization (HH, HV),
Geocoded product, 1.5 processing level is utilized in this study. L-band
(1.27 GHz) synthetic aperture radar aids at achieving high-resolution
DEM products. Thus, ALOS-PALSAR FBS (Fine Beam Single polariza-
tion mode, HH) RT1 (Radiometric Terrain Corrected with Pixel spacing
is 12.5 m) DEM is used to form topographical combinations with optical
and radar data, as well as, derivation for slope and aspect of the study
area.
All of these data are attained through the U.S. Geological Survey (htt
ps://earthexplorer.usgs.gov/) and the European Space Agency (ESA)
with the following IDs: LC81740432019298LGN00 (Landsat 8),
AST_L1A_00303062007083043 (ASTER), S2A_MSIL1C_2020050
5T081611_N0209_R121_T36RWN_20200505T095132 (Sentinel 2),
EO1A1740422003070110PZ (EO1 ALI), S1A_IW_GRDH_1SDV_2020
0926T154707_20200926T154736_034532_0404DC_2A0B (Sentinel 1),
ALOS-FBD_GEC_1P-ORBIT__00022767_D20100503-T201857038 (ALO
S-PALSAR 1), and AP_10689_FBS_F0480_RT1 (ALOS-PALSAR DEM). The
cloud-free multispectral data of Landsat-8 level 1T, ASTER, and EO-1
ALI are introduced to Fast Line-of-Sight Atmospheric Analysis of Spec-
tral Hypercubes (FLAASH) atmospheric correction, and data resizing to
the borders of the study area. All of these operations were carried out at
the Environment for Visualizing Images (ENVI) software version 5.6.
Sentinel-2 Dataset was reprojected to UTM WGS 84 zone 36N then,
radiometrically-corrected using sen2cor processor in Sentinel Applica-
tion Platform (SNAP). Sentinel-1 data was preprocessed by applying the
precise orbit of acquisition, removing thermal and image border noise,
radiometric calibration, and range doppler and terrain correction using
Sentinel-1 toolbox. ALOS PALSAR1 data are also subjected to repro-
jection, radiometric calibration and speckle reduction using SNAP.
2.3. Training and testing samples
Nine main information classes cover the study area. Thus, repre-
sentative, and well-distributed nine classes (for training and testing
samples), are delineated from Landsat 8 and ASTER data, based on the
geological map and visual interpretation of the images. The full
Fig. 1. Location of the study area (small black rectangle over the Egyptian
Eastern Desert); and geological map of the study area showing its lithologic
units, after (Zoheir et al., 2019a).
A. Shebl and ´
A. Cs´
amer

Remote Sensing Applications: Society and Environment 24 (2021) 100643
3
description of training and ground truth pixels is given in Table 2, and
their spectral signatures from Landsat OLI are shown in Fig. 2. Prototype
pixels are used to train the classiers to produce thematic maps for the
classes, and ground truth samples are employed to assess the accuracy of
these outputs.
2.4. Machine learning classiers
2.4.1. Articial Neural Network (ANN)
A frequently used non-parametric supervised classier. Mimicking
the human brain, ANN is chiey formed from neurons (processing
nodes) embedded into input, middle (maybe more than one), and output
layers (Haykin, 2010). Binding and connecting these nodes during the
process forms a complicated network depending on the number of the
utilized layers and classes. Through this network, a decision surface
could be managed, and various classes could be separated. One of the
main ANN strengths is the ability to create non-linear decision surfaces
for inherently not separable data with a simple linear decision surface
(Richards and Jia, 1999). For this study, we used ENVI 5.6 software to
perform multi-layer feed-forward ANN by the logistic activation func-
tion. The best results are revealed by assigning the training root mean
square (RMS) exit criterion as 0.1, training threshold contribution value
as 0.9, training rate as 0.2, and training momentum as 0.9.
2.4.2. Maximum Likelihood Classier (MLC)
Statistical parametric supervised classier, where the classes are
separated depending on probability distributions. Each class signature is
specied by its mean vector and covariance matrix. Unclassied pixel
belongs to a certain class only when it possesses a higher probability for
that class (Chen et al., 2007). Consequently and based on discriminant
functions, decision surfaces could be handled and separating different
classes, according to the following equations (Richards and Jia, 1999),
x∈
ω
i,if p(
ω
i|x)>p(
ω
j⃒
⃒x)for all j ∕= I,(1)
Where x represents a pixel vector (column of values),
ω
i spectral class.
Bayes’ theorem control the relation between the coveted p(
ω
i|x) and the
obtainable (from training data) p(x|
ω
i) probabilities,
p(
ω
i|x) =p(x|
ω
i) p(
ω
i)/p(x) (2)
x∈
ω
i,if p(x|
ω
i)p(
ω
i)>p(x|
ω
j)p(
ω
j)for all j ∕= I(3)
Where p(
ω
i) is the probability that class
ω
i occurs in the image, and is
called prior probability. Then last equation (3) is transformed to deci-
sion rule (utilized in MLC);
x∈
ω
i,if gi(x)>gj(x)for all j ∕= I(4)
Where gi(x) denotes discriminant functions. For the current study,
lithologic generalization using MLC is carried out using ENVI 5.6
software.
2.4.3. Support vector machine (SVM)
SVM is a non-parametric, supervised classier that proved its
leverage in remote sensing classications since introduced by (Bayliss
et al., 1998). The unique character of SVM is that pixels in the proximity
of the separating hyperplane, accurately detect, where it should pass.
Thus, these pixels are called support vectors, as they delicately locate the
optimal separator hyperplane, via maximizing the distance between
support pixels at class margins, then constructing two marginal hyper-
planes. Generally, for (N) number of bands, (w
i
) set of coefcients,
delineate the linear surface as
w
1
x
1
+w
2
x
2
+… … +w
N
x
N
+w
N+1
=0
or, w
tx
+w
N+1
=0 (5)
Where x is the co-ordinate vector and w is called the weight vector. The
transpose(t) operation for transforming column vector to row vector.
After training and specication for decision surface, the decision rule is
x ∈class 1, if w
tx
+w
N+1
>0 (6)
x ∈class 2, if w
tx
+w
N+1
<0 (7)
Kernel function (such as linear, polynomial, radial basis function and
sigmoid (Wang et al., 2017) is required for SVM. In this study, radial
basis function kernel is applied due to its considered interpolation ca-
pabilities (Zhu et al., 2011), the penalty parameter was set to 100, and
the gamma parameter in the kernel function was the inverse of the band
number (Othman and Gloaguen, 2014).
Table 1
Characteristics of the utilized multispectral datasets.
Landsat 8 ASTER Sentinel 2 EO1 ALI
B.n C.W. (
μ
m) S.R (m) B.n C.W. (
μ
m) S.R (m) B.n C.W. (
μ
m) S.R (m) B.n C.W. (
μ
m) S.R (m)
1 0.442 30 1 0.560 15 1 0.443 60 Pan. 0.585 10
2 0.483 30 2 0.660 15 2 0.490 10 1 0.443 30
3 0.561 30 3N 0.820 15 3 0.560 10 2 0.482 30
4 0.654 30 3B 0.820 15 4 0.665 10 3 0.565 30
5 0.864 30 4 1.650 30 5 0.704 20 4 0.660 30
6 1.609 30 5 2.165 30 6 0.740 20 5 0.790 30
7 2.203 30 6 2.205 30 7 0.782 20 6 0.867 30
8 0.598 15 7 2.260 30 8 0.842 10 7 1.250 30
9 1.373 30 8 2.330 30 8a 0.865 20 8 1.650 30
10 10.90 100 9 2.395 30 9 0.945 60 9 2.215 30
11 12.00 100 10 1.375 60
11 1.610 20
12 2.190 20
Band number (B.n), Central wavelength (C.W), Spatial Resolution (S.R) and Pan for panchromatic.
Table 2
Areas, training and testing pixels, and abbreviations of the Lithologic classes.
Lithologic Unit Area
(Km
2
)
Training
Samples
Testing
Samples
1 Wadi Deposits (Wdp) 1.39 3546 879
2 Nubian Sandstone (Nss) 0.61 1599 338
3 Post-orogenic granite
(Pog)
0.45 1190 292
4 Syn-orogenic granite (Sog) 0.86 2215 599
5 Metagabbro-Diorite (MgD) 0.62 1605 367
6 Metavolcanics (Mvs) 0.79 2080 286
7 Ophiolitic Melange (Ome) 0.94 2404 438
8 Ophiolitic Metagabbro
(Omg)
0.54 1410 188
9 Ophiolitic Serpentinite
(Osp)
0.96 2518 389
Total 7.16 18567 3776
A. Shebl and ´
A. Cs´
amer

Remote Sensing Applications: Society and Environment 24 (2021) 100643
4
2.5. Stacked vector method
The utilized data are hybrid and have various spectral and spatial
characteristics. Consequently, all the datasets are reprojected to the
same coordinate system (WGS, 1984 Zone 36), then, cubic convolution
resampling to a pixel size 20 m, was performed, to ensure the same size
and location for each pixel in the dimensional data. Combinations are
established as extended pixel vectors by aggregating spectral (i.e. optical
sensors) and nonspectral (i.e. DEMs and Radar) datasets, according to
the following form
X=[xt
1,xt
2, ......xt
s]t(8)
Where x
1
… …x
S
are data vectors included in S number of individual
layers (Richards and Jia, 1999). The superscript “t” indicates a vector
transpose operation. X denotes the stacked vector introduced to the
three classiers. In this study, 37 stacked vectors are originated and
tested for the classication process as described in Table 3.
3. Results
We nd several results from fullling 111 classication processes (37
processes for each classier). For all the outputs, we assessed the allo-
cation accuracy via the widely used error matrix and kappa coefcient.
Overall Accuracies (OAs) and kappa coefcient (K) for all the outputs
are shown in Table 3. The latter clearly shows the notability of SVM and
MLC over ANN, and obviously, conrm (just by comparing the numbers
at the top and the bottom of the table) the effect of data dimensionality
(by increasing the number of participated channels) in boosting classi-
cation accuracy. What is better displayed in Fig. 3a, where the OAs are
gradually increased from around 75% to up to 91% (i.e. growing by
16%) by augmentation of the data input to the classier. Fig. 3a, also
demonstrated SVM superiority over MLC in lithologic classication,
approximately at all the processes. Moreover, the increasing rate be-
tween (10–20) bands is higher than that between bands (20–30). An OA
peak value is obtained when the participated bands are 31, beyond this
number, no considered OA increase is evident whatever the participated
number of bands. Up and down curve swinging for the increasing trend
is attributed to the various characteristics of datasets introduced to the
classier. Moreover, data spectral characteristics can variously affect the
generalization OAs, even though at the same dimensionality, as shown
in Fig. 3b.
Among the utilized multispectral datasets (S2, L8, AST, and ALI), the
highest labeling efciency is recorded for ALI, the lowest for L8. ASTER
and S2 give reasonable results. Combining two optical sensors results in
variable accuracies depending on the spectral characteristics of the
combined datasets, the number of bands, and the utilized classier. SVM
highest OA was 87.57% for AST +ALI, MLC gives OA of S2+ALI as
86.65%, and ANN best result is specied to L8+ALI by 79.34%. Jointing
three different optical datasets takes the categorization OAs a little bit
higher (89.67% by SVM and 89.35% by MLC) When the classiers are
fed by ALI +S2+ASTER data. Stacking the four optical datasets raised
the accuracy of the resultant thematic maps to 89.40% and 90.01%
when using MLC and SVM, respectively.
Embedding topographical data (especially DEM), to any collection of
multispectral data is prodigious. DEM introduces a permanent but var-
iable OA raise, depending on the other included data vectors and di-
mensions, as shown in Fig. 3c. For DEM derivatives, the slope is
preferable than aspect in boosting the allocation output, but DEM is
more advisable when compared to both. Immersing radar (S1 and ALOS
PALSAR1) data in multispectral and topographical data can faintly
improve the lithologic identication as shown in Fig. 3d.
However, tangible OA excess is achieved by incorporating all the
datasets for the classication (Fig. 3d). It is worth mentioning to note
that, the highest two OAs are delivered by ALI +S2+AST +DEM +
S1+PALSAR1+slope (43 band, 91.41% by SVM) and ALI +S2+AST +
DEM (31 band, 91.23% by SVM). This in turn, boosts the second com-
bination to be applied in further classications (only 0.18% difference in
OAs despite the presence of 13 excessive bands in the rst combination).
By investigation of the confusion matrix of this combination, reasonable
accuracies are observed. Average of producer and user accuracies for
each class of ALI +S2+AST +DEM combination are shown in Fig. 3e.
Consequently, the number of utilized training pixels and (31) channels
are supposed to be well related in the current study. The lowest number
of training pixels, employed in this study was 1190 for Pog (i.e. 38 N).
Thus, we believed that classication OA reaches its peak for a (N)
number of bands (depending on the data, classiers, and classes) by
implementing number training pixels ranging from 30 to 40N, and no
Fig. 2. Average of surface reectance from training pixels selected from Landsat OLI data (7 bands) for nine information classes.
A. Shebl and ´
A. Cs´
amer

Remote Sensing Applications: Society and Environment 24 (2021) 100643
5
considered increase is obtained beyond that limit unless a new training
data could be included for the classication process. Consequently and
for achieving potent categorization, this study assumed that the number
of training pixels for a single region of interest (ROI) is preferable to be
at least 30 N, rather than the wide range (10N–100N per class) assumed
by the authors of (Davis et al., 1978). The resultant thematic maps of the
previously mentioned stacked data collections are shown in Fig. 4.
4. Discussion
Various rock units could be reasonably discriminated by multispec-
tral data depending on their spectral characteristics and the imaging
sensor properties as displayed in Table 4, which shows continuous in-
terest in lithologic mapping over the last 10 years. Table 4 compares
results of 15 previous studies with the current study that implemented
the frequently recommended MLAs over a wide range of integrated data
sets.
Using MLAs and adding topographical and/or radar data could
enhance the discrimination process, especially if the classes have closer
spectral characteristics and are poorly recognized by optical sensors
solely. This study revealed that discrimination ability is higher for DEM
and slope when compared to radar data. For instance, serpentinites,
post-orogenic granites and metavolcanics are easily identied by their
higher elevations (from DEM) and steeper slopes as shown in Fig. 5a, b.
On contrary, wadi deposits and syn-tectonic granites are distinguished,
by their relative lower elevations. In between, m´
elange assemblages and
the other rock units could be differentiated with the aid of their spectral
signatures (from multispectral data). Thus, topographical data can set
apart most of the lithologic classes. However, radar data can hardly
disclose differences between rock units (due to approximately their
similar interaction with active microwaves) and the obvious discrimi-
nation is evident only for ne wadi deposits and weathered syn-tectonic
granites from other hard rocks, as shown in Fig. 5c.
Classifying and mapping Precambrian rocks (particularly ophiolitic)
is a complicated process not only from remote sensing data (often mixed
spectral signatures are encountered) but also during eldwork. For
example, several researchers introduced geological maps for parts of the
study area, based on eld studies, and many variations are detected
between the produced maps (Aboelkhair et al., 2010; Helba et al., 2001;
Zoheir et al., 2019b; Zoheir and Weihed, 2014). Consequently,
Table 3
Overall accuracies (OAs) and Kappa coefcients (K) for the utilized datasets and classiers.
Data N ANN MLC SVM
OA K OA K OA K
L8 7 71.05 0.664 74.81 0.710 76.43 0.726
AST 9 73.11 0.688 77.07 0.735 80.27 0.771
ALI 9 77.15 0.735 81.36 0.784 83.87 0.812
S2 12 74.29 0.701 78.71 0.753 79.42 0.761
Optical sensor +DEM
L8+DEM 8 18.94 0.0565 81.78 0.7896 80.93 0.7791
AST +DEM 10 70.84 0.6614 82.47 0.797 86.84 0.8475
ALI +DEM 10 42.85 0.332 84.75 0.8238 86.89 0.8481
S2+DEM 13 82.07 0.7927 81.51 0.786 82.73 0.7998
2 Optical sensors together
S2+AST 21 78.681 0.7525 86.626 0.8451 86.89 0.8481
S2+L8 19 76.85 0.7315 78.416 0.7509 80.985 0.7794
S2+ALI 21 77.171 0.7345 86.652 0.8453 85.725 0.8344
AST +L8 16 73.384 0.6915 79.449 0.7633 84.666 0.8221
AST +ALI 18 74.973 0.71 84.401 0.8194 87.579 0.8559
L8+ALI 16 79.343 0.7595 84.322 0.8187 85.301 0.8294
2 Optical sensors +DEM
S2+AST +DEM 22 82.256 0.7937 87.314 0.8531 88.268 0.8641
AST +ALI +DEM 19 73.119 0.6873 87.394 0.8543 89.91 0.8831
3 Optical sensors together
S2+AST +L8 28 77.092 0.7337 86.732 0.8465 87.791 0.858
AST +L8+ALI 25 73.49 0.693 87.764 0.8585 88.135 0.8623
L8+S2+ALI 28 78.125 0.7459 84.613 0.8213 86.99 0.849
ALI +S2+AST 30 77.807 0.7422 89.3538 0.8766 89.671 0.8802
3 Optical sensors +DEM
S2+AST +L8+DEM 29 81.197 0.7818 88.082 0.8621 89.036 0.873
AST +L8+ALI +DEM 26 71.18 0.666 88.638 0.8687 89.48 0.878
L8+S2+ALI +DEM 29 81.56 0.785 86.573 0.8444 86.99 0.849
ALI +S2+AST +DEM 31 83.26 0.805 89.857 0.8826 91.23 0.898
Synergetic use of 4 Optical sensors
S2+AST +L8+ALI (4Opt) 37 79.21 0.758 89.40 0.877 90.01 0.884
4 Optical sensors +topographic data
S2+AST +L8+ALI +DEM 38 82.28 0.7944 89.6 0.88 91.18 0.897
S2+AST +L8+ALI +slope 38 78.76 0.75 89.72 0.880 90.78 0.893
S2+AST +L8+ALI +aspect 38 75.45 0.71 89.85 0.882 89.88 0.882
4 Optical sensors +radar data
S2+AST +L8+ALI +S1VV 38 76.50 0.727 89.83 0.882 90.09 0.885
S2+AST +L8+ALI +S1VH 38 76.24 0.72 89.93 0.883 90.30 0.887
S2+AST +L8+ALI +S1VH +S1VV 39 76.69 0.72 89.93 0.883 90.09 0.885
S2+AST +L8+ALI +Palsar (HH +HV) 39 76.48 0.72 90.06 0.885 89.80 0.881
4Opt +S1+Palsar 41 77.48 0.73 90.22 0.886 90.14 0.885
4 Optical sensors +topographic and radar data
4Opt +DEM+2 S1+2pal 42 82.83 0.80 90.59 0.891 90.51 0.890
4Opt +DEM+2 S1+2pal +slope 43 75.92 0.72 90.28 0.887 91.41 0.90
4Opt +DEM+2 S1+2pal +aspect 43 81.48 0.78 89.67 0.880 90.17 0.886
4Opt +DEM+2 S1+2pal +slope +aspect 44 82.28 0.793 89.98 0.884 90.83 0.893
The four optical sensors together (4 Opt), Sentinel 1 (S1), ALOS Palsar data (pal) and Number of bands (N).
A. Shebl and ´
A. Cs´
amer

Remote Sensing Applications: Society and Environment 24 (2021) 100643
6
homogenous monotone training pixels are carefully selected, aiming to
resolve these discrepancies. The resultant thematic maps are closely
analogous to the reference geological map as shown in Fig. 4. Mis-
classications are sometimes introduced among ophiolitic m´
elange,
ophiolitic metagabbro, and metagabbro-diorite complex. This is attrib-
uted to closeness in their spectral signatures (as shown in Fig. 2),
cognate elevations, and similar appearance in radar data. Thus, these
ground cover classes are not exemplary resolved when compared to the
reference map. ALI +S2+AST +DEM combination (31 bands) gather
the advantages of ALI (six unique VNIR wavelength bands (Hubbard
et al., 2003; Hubbard and Crowley, 2005) and improved Signal-to-Noise
Ratio (Bryant et al., 2003; Lencioni et al., 1999; Lobell and Asner, 2003;
Mendenhall et al., 2000), ASTER (high spectral separability of most
functional groups and their minerals using eminent six SWIR bands
(Pour and Hashim, 2014)), S2 (higher spectral characteristics for both
VNIR and SWIR (Ge et al., 2018) and DEM (topographical characteristics
discernment), thus delivered precise results. Furthermore, this combi-
nation is strongly recommended for lithologic classications due to two
main reasons: 1- adding any supporting bands can hardly enhance the
allocation accuracy, as seen for ALI +S2+AST +DEM +
Fig. 3. (a) The effect of increasing data dimensionality on classication accuracy. Beyond 31 participated band, the Overall Accuracy (OA) excess is barely
perceptible; (b) various OA results for different datasets having the same dimensions, differences could reach to about 9% (between S2+L8 and AST +ALI +DEM);
(c) DEM inuence in boosting the classication OA; (d) Topographical and radar data impact on categorization OA; (e) Average accuracies of the information classes,
classied by ANN, MLC, and SVM, from ALI +S2+AST +DEM.
A. Shebl and ´
A. Cs´
amer

Remote Sensing Applications: Society and Environment 24 (2021) 100643
7
S1+PALSAR1+slope combination (43 bands), only gives 0.18%
improvement by adding 11 supporting layers. 2- Removing any con-
stituent can signicantly reduce the OA as shown in Table 3. Our result
is harmonized with (Ge et al., 2018), who preferred S2+AST +DEM in
lithologic classication, however, we strongly recommend adding ALI to
the previous combination as it improves OAs by approximately 1%,
2.5%, and 3% for ANN, MLC, and SVM respectively. It should be
emphasized that this study applied transfer learning as it is still a
powerful method when multisource remote sensing images are imple-
mented in the allocation process (Dong et al., 2021), which is the case in
the current study. However, other studies efciently combined deep and
transfer learning to get efcient results (Dong et al., 2021; Shi et al.,
2021; Xie et al., 2021; Yuan et al., 2020; Zhou et al., 2019).
MLC reasonable results are interpreted by representing training
signatures for each class by both covariance matrices and mean vectors.
However, sometimes and due to data dimensionality and heterogeneity,
incompatible statistics and quadratic cost increase are encountered
resulting in lower OAs compared to SVM results. For SVM, the class
boundaries are determined by the collection of weights (equations (5)–
(7)) that make the accurate labeling process is feasible, as well as a
penalty for misclassications is introduced to achieve better results
compared to ANN. For data with N dimensions, 30–40N is assigned to be
the appropriate numbers of training pixels required for fullling
accurate classication (by attaining the exact location of the decision
surface). If the used training pixels are insufcient (less than 30N), the
location of the separating surface may be inadequate, resulting in
various misclassications depending on the utilized algorithm. Also,
Fig. 4. Lithologic classication outputs utilizing ALI +S2+AST +DEM data
from (a) ANN; (b) MLC and; (c) SVM. (d) and (e) are the results of general-
ization over ALI +S2+AST +DEM +S1+PALSAR1+slope data utilizing MLC
and SVM, respectively. (f) Geological map for comparison.
Table 4
Comparison of the obtained results over the last decade with the current study.
Study Classiers Data used Results
1 Grebby et al.
(2011)
ANN (Self-
Organizing
Map), MLC
Airborne
multispectral and
LiDAR data
ANN results shows
greater
enhancement
2 Yu et al.
(2012)
MLC, SVM ASTER, DEM SVM outputs is
better than MLC
results
3 Mehr et al.
(2013)
MLC Thematic Mapper
5 (TM5)
MLC enhances the
separability of units
in the image by
7.3%
4 Fatima et al.
(2013)
MLC, PP,
MDM, MD,
SAM
Landsat ETM+
and ASTER
MLC has the highest
correlation with the
geological map
5 Hadigheh and
Ranjbar
(2013)
MLC, SAM,
SID
ASTER, IRS MLC is the better
than SAM and SID
6 A. A. Othman
and Gloaguen
(2014)
SVM ASTER SVM is efcient tool
in mineral and
lithological
mapping
7 He et al.
(2015)
ANN, RF,
SVM, MLC
Landsat-7 and
Landsat-8
ANN, RF, and SVM
outperformed MLC
8 Jellouli et al.
(2016)
MLC, SVM ASTER MLC results is better
than that of SVM
9 Othman and
Gloaguen
(2017)
MLC, SVM,
RF
ASTER, DEM RF and SVM are
better than MLC
10 Ge et al.
(2018)
ANN, k-NN,
MLC, RF,
SVM
Sentinel-2A MLC method
offered the highest
overall accuracy
11 Manap and
San (2018)
MLC, RF,
SVM
ASTER MLC and RF have
approximately the
same accuracy, and
SVM is more
accurate than them.
12 Bachri et al.
(2019)
SVM Landsat 8,
PALSAR DEM
SVM performs
effective
classication with
85% accuracy
13 Karimzadeh
and H.
Tangestani
(2021)
SVM WorldView-3 SVM performs
effective
classication with
88.36% accuracy
14 Bentahar and
Raji (2021)
MLC, SVM Landsat OLI,
ASTER, and
Sentinel 2A
the effectiveness of
Sentinel 2A data
and SVM is better
than MLC
15 Kumar et al.
(2021)
SVM and RF ASTER, PALSAR,
Sentinel 1
SVM is better than
RF
16 Current Study ANN, SVM,
MLC
Sentinel 2, ASTER,
Landsat OLI
Advanced Land
Imager, Sentinel 1,
ALOS PALSAR-1,
ALOS PALSAR
DEM, Slope,
Aspect
SVM is better than
MLC which is in
turn better than
ANN,
DEM enhances the
classication,
S2, ASTER and ALI
are preferred when
compared to
Landsat OLI,
Integration of
multi-sensors
strongly boost the
outputs.
Articial neural network (ANN), Maximum likelihood (MLC), support vector
machine (SVM), k-nearest neighbor (k-NN), maximum likelihood classication
(MLC), random forest classier (RF), and parallelepiped (PP), minimum distance
to mean (MDM), mahalanobis distance (MD), Spectral Angle Mapper (SAM),
Spectral Information Divergence (SID), Indian Remote Sensing Satellite (IRS).
A. Shebl and ´
A. Cs´
amer

Remote Sensing Applications: Society and Environment 24 (2021) 100643
8
excessively utilized training pixels (exceeding 40N) or exaggerated di-
mensions (for a constant number of training pixels) cannot signicantly
improve the classication because the separator hyperplane is already
optimally located, and any support (either by training pixels or chan-
nels) cannot introduce signicant revamping for the hyperplane place-
ment between classes. we believe that the possible way to boost the
allocation process is by increasing both training data and dimensions
together, keeping the same percentage (30–40N).
5. Conclusions
In this study, three classiers (ANN, MLC, and SVM) are fed by 37
data inputs including (Sentinel 2, ASTER, Landsat OLI, Earth-observing
1 Advanced Land Imager, Sentinel 1, ALOS PALSAR 1, ALOS PALSAR
DEM, Slope, Aspect and their combinations) to enhance the lithologic
mapping of the study area. Also, the dimensionality power in improving
the classication output and the effect of these datasets are outlined. The
study concludes the followings results:
•SVM, MLC, and ANN are eligible in lithologic classication. Results
of SVM and MLC outperforms that of ANN.
•Data dimensionality increment can enhance the allocation process
only when there is sufcient training data, compared to the number
of utilized bands (N). The minimum number of the desired training
pixels for executing accurate classication is 30N.
•DEM remarkably improves the allocation compared to slope, aspect,
sentinel 1 and Alos palsar data.
•S2, ASTER and ALI are preferred when compared to Landsat OLI, in
lithologic categorization. Finally,
•For further lithologic classications, ALI+S2+AST+DEM combina-
tion is robustly recommended. Moreover, this study can help re-
searchers in detecting a suitable data input for their classication, as
this study introduces 111 classication results for commonly used
remote sensing datasets and their combinations.
Author statement
Ali Shebl: Conceptualization, Data curation, Formal analysis,
Methodology, Project administration, Resources, Software, Validation,
Visualization, Writing - original draft, Writing - review and editing.
´
Arp´
ad Cs´
amer: Funding acquisition, Investigation, Supervision, Vali-
dation, Resources, Writing - review and editing.
Funding
This research received no external funding. This research is sup-
ported by University of Debrecen. Ali Shebl is funded by Stipendium
Hungaricum scholarship under the joint executive program between
Hungary and Egypt.
Data availability statement
The data that support the ndings of this study are openly available
in gshare at [https://figshare.com/s/fd6a261ce7e636307ddd], refer-
ence number[10.6084/m9.gshare.14701209].
Declaration of competing interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Acknowledgments
The authors are thankful to U.S. Geological Survey, and European
Space Agency (ESA), for providing the data. Thanks to Prof. Mahmod
Ashmawy, Prof. Mohamed Abdelwahed and Prof. Samir Kamh of Tanta
University, for their support.
Fig. 5. Reasonable lithologic discrimination fro m (a) DEM (approximately brown Osp, green to cyan Mvs, white Wdp, and pink Sog) compared to; (b) Slope and; (c)
RGB (HH, HV, HH/HV) ALOS PALSAR 1 combination.
A. Shebl and ´
A. Cs´
amer

Remote Sensing Applications: Society and Environment 24 (2021) 100643
9
References
Aboelkhair, H., Ninomiya, Y., Watanabe, Y., Sato, I., 2010. Processing and interpretation
of ASTER TIR data for mapping of rare-metal-enriched albite granitoids in the
Central Eastern Desert of Egypt. J. Afr. Earth Sci. 58, 141–151. https://doi.org/
10.1016/j.jafrearsci.2010.01.007.
Bachri, I., Hakdaoui, M., Raji, M., Teodoro, A.C., Benbouziane, A., 2019. Machine
learning algorithms for automatic lithological mapping using remote sensing data: a
case study from Souk Arbaa Sahel, Sidi Ifni Inlier, western Anti-Atlas, Morocco.
ISPRS Int. J. Geo-Information 2019 8. https://doi.org/10.3390/IJGI8060248. Page
248 8, 248.
Bayliss, J.D., Gualtieri, J.A., Cromp, R.F., 1998. Analyzing hyperspectral data with
independent component analysis. 26th AIPR Work. Exploit. New Image Sources
Sensors 3240, 133–143. https://doi.org/10.1117/12.300050.
Bentahar, I., Raji, M., 2021. Comparison of Landsat OLI, ASTER, and sentinel 2A data in
lithological mapping : a case study of rich area (central high Atlas, Morocco). Adv.
Space Res. 67, 945–963. https://doi.org/10.1016/J.ASR.2020.10.037.
Bryant, R., Moran, S., McElroy, S.A., Holied, C., Thome, K.J., Miura, T., Biggar, S.F.,
2003. Data continuity of earth observing 1 (EO-1) Advanced land imager (ALI) and
Landsat TM and ETM+. IEEE Trans. Geosci. Rem. Sens. 41, 1204–1214. https://doi.
org/10.1109/TGRS.2003.813213.
Chen, X., Warner, T.A., Campagna, D.J., 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. https://doi.org/
10.1016/j.rse.2007.03.015.
Czapla-Myers, J., Ong, L., Thome, K., McCorkel, J., 2016. Validation of EO-1 hyperion
and Advanced land imager using the radiometric calibration test site at railroad
valley, Nevada. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 9, 816–826. https://
doi.org/10.1109/JSTARS.2015.2463101.
Davis, S.M., Landgrebe, D.A., Phillips, T.L., Swain, P.H., Hoffer, R.M., Lindenlaub, J.C.,
Silva, L.F., 1978. Remote Sensing: the Quantitative Approach. mhi.
Dong, Y., Liang, T., Zhang, Y., Du, B., 2021. Spectral-spatial weighted Kernel Manifold
embedded distribution alignment for remote sensing image classication. IEEE
Trans. Cybern. 51, 3185–3197. https://doi.org/10.1109/TCYB.2020.3004263.
Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B.,
Isola, C., Laberinti, P., Martimort, P., 2012. Sentinel-2: ESA’s optical high-resolution
mission for GMES operational services. Remote Sens. Environ. 120, 25–36.
Fatima, K., Khattak, U.K., Kausar, A.B., 2013. Selection of appropriate classication
technique for lithological mapping of Gali Jagir area, Pakistan. Int. J. Earth Sci. Eng.
7, 964–971.
Foody, G.M., Mathur, A., 2004. A relative evaluation of multiclass image classication by
support vector machines. IEEE Trans. Geosci. Rem. Sens. 42, 1335–1343. https://
doi.org/10.1109/TGRS.2004.827257.
Ge, W., Cheng, Q., Jing, L., Armenakis, C., Ding, H., 2018. Lithological discrimination
using ASTER and Sentinel-2A in the Shibanjing ophiolite complex of Beishan
orogenic in Inner Mongolia, China. Adv. Space Res. 62, 1702–1716. https://doi.org/
10.1016/j.asr.2018.06.036.
Grebby, S., Naden, J., Cunningham, D., Tansey, K., 2011. Integrating airborne
multispectral imagery and airborne LiDAR data for enhanced lithological mapping in
vegetated terrain. Remote Sens. Environ. 115, 214–226. https://doi.org/10.1016/j.
rse.2010.08.019.
Hadigheh, S.M.H., Ranjbar, H., 2013. Lithological mapping in the eastern part of the
central Iranian volcanic belt using combined ASTER and IRS data. J. Indian Soc.
Remote Sens. 2013 414 41, 921–931. https://doi.org/10.1007/S12524-013-0284-1.
Ham, J.S., Chen, Y., Crawford, M.M., Ghosh, J., 2005. Investigation of the random forest
framework for classication of hyperspectral data. In: IEEE Transactions on
Geoscience and Remote Sensing, pp. 492–501. https://doi.org/10.1109/
TGRS.2004.842481.
Haykin, S., 2010. Neural Networks: a Comprehensive Foundation. 1999. Mc Millan, New
Jersey, pp. 1–24.
He, J., Harris, J.R., Sawada, M., Behnia, P., 2015. A Comparison of Classication
Algorithms Using Landsat-7 and Landsat-8 Data for Mapping Lithology in Canada’s
Arctic, pp. 2252–2276. https://doi.org/10.1080/01431161.2015.1035410 36.
Helba, H.A., Khalil, K.I., Abou, N.M.F., 2001. Alteration patterns related to hydrothermal
gold mineralizaition in meta-andesites at Dungash area, Eastern Desert, Egypt.
Resour. Geol. 51, 19–30. https://doi.org/10.1111/j.1751-3928.2001.tb00078.x.
Hubbard, B.E., Crowley, J.K., 2005. Mineral mapping on the Chilean-Bolivian Altiplano
using co-orbital ALI, ASTER and Hyperion imagery: data dimensionality issues and
solutions. Remote Sens. Environ. 99, 173–186. https://doi.org/10.1016/j.
rse.2005.04.027.
Hubbard, B.E., Crowley, J.K., Zimbelman, D.R., 2003. Comparative alteration mineral
mapping using visible to shortwave infrared (0.4-2.4
μ
m) Hyperion, ALI, and ASTER
imagery. IEEE Trans. Geosci. Rem. Sens. 41, 1401–1410. https://doi.org/10.1109/
TGRS.2003.812906.
Jellouli, A., El Harti, A., Adiri, Z., El Ghmari, A., Bachaoui, E.M., Jellouli, A., El Harti, A.,
Adiri, Z., El Ghmari, A., Bachaoui, E.M., 2016. Lithological mapping using ASTER
data in the Moroccan Anti Atlas belt. EGUGA 18, EPSC2016–13872.
Karimzadeh, S., Tangestani, M.H., 2021. Evaluating the VNIR-SWIR datasets of
WorldView-3 for lithological mapping of a metamorphic-igneous terrain using
support vector machine algorithm; a case study of Central Iran. Adv. Space Res. 68,
2421–2440. https://doi.org/10.1016/J.ASR.2021.05.002.
Kumar, C., Chatterjee, S., Oommen, T., Guha, A., Mukherjee, A., 2021. Multi-sensor
datasets-based optimal integration of spectral, textural, and morphological
characteristics of rocks for lithological classication using machine learning models.
https://doi.org/10.1080/10106049.2021.1920632.
Lee, S., Song, K.Y., Oh, H.J., Choi, J., 2012. Detection of landslides using web-based
aerial photographs and landslide susceptibility mapping using geospatial analysis.
Int. J. Rem. Sens. 33, 4937–4966. https://doi.org/10.1080/01431161.2011.649862.
Lencioni, D.E., Digenis, C.J., Bicknell, W.E., Hearn, D.R., Mendenhall, J.A., 1999. Design
and performance of the EO-1 Advanced land imager. Sensors, Syst. Next-Generation
Satell. III 3870, 269–280. https://doi.org/10.1117/12.373195.
Liesenberg, V., Gloaguen, R., 2012. Evaluating SAR polarization modes at L-band for
forest classication purposes in eastern Amazon, Brazil. Int. J. Appl. Earth Obs.
Geoinf. 21, 122–135. https://doi.org/10.1016/j.jag.2012.08.016.
Lobell, D.B., Asner, G.P., 2003. Comparison of earth observing-1 ALI and Landsat ETM+
for crop identication and yield prediction in Mexico. IEEE Trans. Geosci. Rem. Sens.
41, 1277–1282. https://doi.org/10.1109/TGRS.2003.812909.
Manap, H.S., San, B.T., 2018. Lithological mapping using different classication
algorithms in western antalya, Turkey. Int. Multidiscip. Sci. GeoConference Surv.
Geol. Min. Ecol. Manag. SGEM 18, 551–556. https://doi.org/10.5593/SGEM2018/
2.2/S08.069.
Mehr, S.G., Ahadnejad, V., Abbaspour, R.A., Hamzeh, M., 2013. Using the mixture-tuned
matched ltering method for lithological mapping with Landsat TM5 images.
http://doi.org/10.1080/01431161.2013.853144 34, 8803-8816.
Mendenhall, J.A., Lencioni, D.E., Evans, J.B., 2000. Earth Observing-1 Advanced Land
Imager: Radiometric Response Calibration.
Othman, A., Gloaguen, R., 2014. Improving lithological mapping by SVM classication
of spectral and morphological features: the Discovery of a new chromite body in the
mawat ophiolite complex (kurdistan, NE Iraq). Rem. Sens. 6, 6867–6896. https://
doi.org/10.3390/rs6086867.
Othman, A.A., Gloaguen, R., 2017. Integration of spectral, spatial and morphometric
data into lithological mapping: a comparison of different Machine Learning
Algorithms in the Kurdistan Region, NE Iraq. J. Asian Earth Sci. 146, 90–102.
https://doi.org/10.1016/J.JSEAES.2017.05.005.
Pal, M., Mather, P.M., 2005. Support vector machines for classication in remote
sensing. Int. J. Rem. Sens. 26, 1007–1011. https://doi.org/10.1080/
01431160512331314083.
Pour, A.B., Hashim, M., 2014. ASTER, ALI and Hyperion sensors data for lithological
mapping and ore minerals exploration. SpringerPlus 3, 130.
Richards, J.A., Jia, X., 1999. Remote Sensing Digital Image Analysis.
Roy, D.P., Wulder, M.A., Loveland, T.R., Woodcock, C.E., Allen, R.G., Anderson, M.C.,
Helder, D., Irons, J.R., Johnson, D.M., Kennedy, R., 2014. Landsat-8: science and
product vision for terrestrial global change research. Remote Sens. Environ. 145,
154–172.
Shi, Y., Ma, D., Lv, J., Li, J., 2021. ACTL: Asymmetric Convolutional Transfer Learning
for tree species identication based on deep neural network. IEEE Access 9,
13643–13654. https://doi.org/10.1109/ACCESS.2021.3051015.
Wang, F., Zhen, Z., Wang, B., Mi, Z., 2017. Comparative study on KNN and SVM based
weather classication models for Day Ahead short term solar PV power forecasting.
Appl. Sci. 8, 28. https://doi.org/10.3390/app8010028.
Xie, F., Gao, Q., Jin, C., Zhao, F., 2021. Hyperspectral image classication based on
superpixel pooling convolutional neural network with transfer learning. Remote
Sens. 13, 930. https://doi.org/10.3390/rs13050930.
Yu, L., Porwal, A., Holden, E.J., Dentith, M.C., 2012. Towards automatic lithological
classication from remote sensing data using support vector machines. Comput.
Geosci. 45, 229–239. https://doi.org/10.1016/J.CAGEO.2011.11.019.
Yuan, X., Ou, C., Wang, Y., Yang, C., Gui, W., 2020. Deep quality-related feature
extraction for soft sensing modeling: A deep learning approach with hybrid VW-SAE.
Neurocomputing 396, 375–382. https://doi.org/10.1016/J.NEUCOM.2018.11.107.
Zhou, P., Han, J., Cheng, G., Zhang, B., 2019. Learning compact and discriminative
stacked autoencoder for hyperspectral image classication. IEEE Trans. Geosci.
Remote Sens. 57, 4823–4833. https://doi.org/10.1109/TGRS.2019.2893180.
Zhu, X., Zhang, S., Jin, Z., Zhang, Z., Xu, Z., 2011. Missing value estimation for mixed-
attribute data sets. IEEE Trans. Knowl. Data Eng. 23, 110–121. https://doi.org/
10.1109/TKDE.2010.99.
Zoheir, B., El-Wahed, M.A., Pour, A.B., Abdelnasser, A., 2019a. Orogenic gold in
transpression and transtension zones: eld and remote sensing studies of the
barramiya–mueilha sector, Egypt. Rem. Sens. 11, 2122.
Zoheir, B., Steele-MacInnis, M., Garbe-Sch¨
onberg, D., 2019b. Orogenic gold formation in
an evolving, decompressing hydrothermal system: genesis of the Samut gold deposit,
Eastern Desert, Egypt. Ore Geol. Rev. 105, 236–257.
Zoheir, B., Weihed, P., 2014. Greenstone-hosted lode-gold mineralization at Dungash
mine, Eastern Desert, Egypt. J. Afr. Earth Sci. 99, 165–187.
A. Shebl and ´
A. Cs´
amer