Content uploaded by Zhikun Ren
Author content
All content in this area was uploaded by Zhikun Ren on Oct 27, 2022
Content may be subject to copyright.
DEM generation from Worldview-2 stereo imagery and vertical accuracy
assessment for its application in active tectonics
Siyu Wang
a,b
,ZhikunRen
a,b,
⁎,ChuanyongWu
c,d
, Qiyun Lei
e
,WenyuGong
a
,QiOu
f
, Huiping Zhang
a
,
Guangxue Ren
b
, Chuanyou Li
b
a
State Key Laboratory of Earthquake Dynamics, Institute of Geology, China Earthquake Administration, Beijing 100029, China
b
Key Laboratory of Active Tectonics and Volcano, Institute of Geology, China Earthquake Administration, Beijing 100029, China
c
Guangdong Provincial Key Laboratory of Geodynamics and Geohazards, School of Earth Sciences and Engineering, Sun Yat-sen University, Guangzhou, China
d
Earthquake Administration of Xinjiang Uygur Autonomous Region, Urumchi 830001, China
e
Earthquake Administration of Ningxia Hui Autonomous Region, Yinchuan 750001, China
f
COMET, Department of Earth Sciences, University of Oxford, Oxford OX1 3QR, UK
abstractarticle info
Article history:
Received 16 August 2018
Received in revised form 12 March 2019
Accepted 15 March 2019
Available online 19 March 2019
The DEM-generation technology from high-resolution satellite imagery enables us to generate a wide range of high-
resolution topographic data rapidly, improving the efficiency of data acquisition greatly. This method is more effi-
cient than airborne Light Detection And Ranging (LiDAR) in terrain reconstruction, since satellite imagery covers a
larger area without difficulties in field deployment. Previous researches evaluated the accuracy of DEMs generated
from stereo imagery of different satellite sensors, however there is not enough quantitative analysis concerning the
capability of satellite-based DEM in active tectonic studies. Therefore, in this paper, we presented a study to inves-
tigate the accuracy to measure heights of fault scarps using the DEM obtained from Worldview-2 stereo imagery
and chose the Kumysh fault in the southern margin of Kumysh Basin (Eastern Tian Shan, China) as our test site.
Point cloud data were obtained from stereo satellite imagery, both with and without GCPs respectively. Prior to
the generation of the DEM, we compared the overall elevation differences of the point clouds and the fault scarp
swaths. The overall elevation difference ranges from −1.2 to 0.4 m, with the mean value of −0.57 m, while the el-
evation difference of fault scarp swaths range from −1.1 to 0 m, with the mean value of −0.4 m. Afterwards, we
generated a 0.5 m resolution DEM of 5-km swath along the Kumysh fault, measured the heights of fault scarps
on different levels of alluvial fans, and compared the topographic profiles obtained from DEM and post-
processed differential GPS (ppGPS) survey. Our results show that: (1) the elevation difference between the topo-
graphic profiles ranges from −2.82–4.47 m, the shape of the fault scarp can be accurately reconstructed by
satellite-based DEM with the deviation of 0.29 m after elevation correction; (2) the accuracy of the height measure-
ment of fault scarps can reach 0.25 m. These findings indicate that the DEM generated from Worldview-2 stereo
imagery is capable of measuring relative deformed topographic features, which could be of great interest to profes-
sionals exploring the use and accuracy of satellite stereo imagery for active tectonic applications.
© 2019 The Authors. Published by Elsevier B.V. This is an open access article under the C CB Y-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).
Keywords:
Stereo imagery
DEM
Fault scarp
Active tectonics
Accuracy analysis
1. Introduction
Digital Elevation Model (DEM) enables us to quantify the stages of
surface processes using discrete mathematical model to characterize
surface topography (Burbank and Anderson, 2011). High spatial resolu-
tion terrain data could be used to monitor earth surface deformation, es-
pecially for some inaccessible areas by comparing multi-temporal DEM,
including active tectonics (Binet, 2005;Mitchell and Chadwick, 1999;
Ren et al., 2014;Zhou et al., 2015), landslide (Mora et al., 2003;
Stumpf et al., 2014), polar resea rch (Shean et al., 2016) and glacier mon-
itoring (Fieber et al., 2018). In the study of active tectonics, substantial
researchhas been conducted using high-resolution DEM, which can dis-
play detailed displacement offault and provides a new method to study
the pattern of surface rupture and fault geometry (Barišin et al., 2015;Bi
et al., 2018;Hashimoto et al., 2010;Turker and Cetinkaya, 2005).
Several techniques are suitable for high quality DEM generation,
such as real-time dynamic data from GPS, Interferometric Synthetic Ap-
erture Radar (InSAR) images, Light Detection And Ranging (LiDAR)
technology, Unmanned Aerial Vehicles (UAV) survey, high resolution
satellite stereo imagery and so on. Traditional field topographic survey
is time-consuming, labor-intensive and inefficient; InSAR has shown a
Geomorphology 336 (2019) 107–118
⁎Corresponding author at: State Key Laboratory of Earthquake Dynamics, Institute of
Geology, China Earthquake Administration,Beijing 100029, China.
E-mail addresses: rzk@ies.ac.cn,lzkren@gmail.com (Z. Ren).
https://doi.org/10.1016/j.geomorph.2019.03.016
0169-555X/© 2019 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Contents lists available at ScienceDirect
Geomorphology
journal homepage: www.elsevier.com/locate/geomorph
great potential for capturing surface deformation caused by various geo-
physical phenomenon with high resolution and precision (Hashimoto
et al., 2010;Massonnet et al., 1992;Wright, 2004;Zebker et al., 1994).
But the quality is limited by the geometry and temporal baseline, atmo-
spheric artifacts and existence of ground deformation; LiDAR is an effec-
tive way to study the distribution characteristics of seismic deformation
and the mode of strong earthquake rupture (Kondo et al., 2008;Ren
et al., 2015;Ren et al., 2018;Zielke et al., 2010) since it can scan the
fault zone in a wide range and gets topographic data of decimeter
level. However, high cost limits its wide application; In recent years,
UAV technology combined with Structure from Motion (SfM) has
greatly improved the efficiency of the acquisition of topographic data
(Angster et al., 2016;d'Oleire-Oltmanns et al., 2012;James and
Robson, 2012;Westoby et al., 2012). Nevertheless, it is limited by the
weather condition and can only obtain photos over small areas. With
the development of digital photogrammetry and image matching tech-
nology, high-resolution satellite imagery has become another source
high-resolution DEMs (Aguilar et al., 2013;Shean et al., 2016;Zhou
et al., 2015). In the study of active tectonics, a small number of stereo
imagery is sufficient to cover the entire study area, which greatly re-
duces the cost and is especially helpful for the regions with harsh cli-
mate and accessible difficulties, providing a robust method for
measuring the surface deformation.
High-resolution surface topography and the height of fault scarps
are valuable for studying coseismic deformation and fault geometry. It
enables us to measure three-dimensional surface displacements in the
earthquake and to calculate accurate slip rate. High accuracy height of
scarp can reduce the uncertainty of slip-rate calculation. With the ad-
vent of the very high-resolution (VHR) optical satellites, such as
IKONOS, Pleiades, GeoEye-1, Worldview-2 and so forth, a large number
of researches have documentedthe accuracy of DEM extracted from ste-
reo imagery and its effectiveness in terms of different applications
(Deilami and Hashim, 2011;Fabris and Pesci, 2005;Poli et al., 2015;
Reinartz et al., 2006;Saldaña et al., 2012;San and Suzen, 2005;Shaker
et al., 2010;Zhou et al., 2015). Saldaña et al. (2012) analyzed the plani-
metric and vertical accuracy of DSM generated from GeoEye-1 stereo
pairs and the results were better than 0.5 m. Shaker et al. (2010) tested
Indian Remote Sensing (IRS) - 1D stereo images and IKONOS stereo im-
ages. The accuracy of the extracted features was found to be within a
pixel-level by empirical mathematical models. The vertical accuracy of
DEM extracted from Pleiades Tri-stereo-pair can reach 0.3 m, compared
with LiDAR data (Zhou et al., 2015). However, there are still few studies
on the quantitative analysis of the accuracy of DEM generated from
Worldview-2 stereo imagery in studying active tectonics. In this study,
we chose the Kumysh fault as the research subject with less vegetation,
used Worldview-2 panchromatic band stereo imagery to generate DEM
and Geo 7× post-processed differential GPS (ppGPS) to survey ground
control points (GCPs). Then we analyzed the accuracy of generated
point clouds under the condition of without GCPs and with GCPs. In
order to evaluate the accuracy of DEM in terms of describing the feature
of fault scarps and its validation in measuring the height of scarps, four
topographic profiles across fault scarps in the field were surveyed and
Fig. 1. Active tectonic map of the study area and stereo imagery coverage region. (a) Tectonic setting of the Eastern Tian Shan. The Cenozoic tectonism due to Eurasia-India collision is
responsible for the north-south shortening of theTian Shan. The black dashed rectangle is the location of (b). (b) Themajor faults of the Eastern Tian Shan.Black rectangle shows the cov-
erage of Worldview-2 stereo imagery (c). (c) Distribution of GCPs and topographic profiles surveyed across the fault scarps. YMTF: YaMaTe fault; BGDSF: Bogda southern fault; HSKF:
HuoYanShan fault; BAF: Bolokenu-Aqikekuduke fault; BETF: Baoertu faut; KMSF: Kumysh fault; YQNF: North-edge fault of Yanqi basin; YQSF: South-edge fault of Yanqi basin.
108 S. Wang et al. / Geomorphology 336 (2019) 107–118
compared to profiles measured from DEM. This work will provide infor-
mation about geometric distribution and deformation characteristics of
Kumysh fault. The aim of this study is to quantitatively evaluate
whether this method is suitable for a wide range application of active
tectonic research.
2. Study site and datasets
2.1. Study site
In the past few decades, many studies focused on the deformation
characteristics and slip rates of northern Tian Shan and southwestern
Tian Shan (Burchfiel et al., 1999;Dumitru et al., 2001;Huang et al.,
2015;Thompson et al., 2002;Wang, 2004;Xiao et al., 2013;Yin et al.,
1998). There is still little attention on the internal tectonic deformation
of Tian Shan. The study site Kumysh Basin is located in the eastern seg-
ment of Tian Shan (Fig. 1). The late Quaternary movement characteris-
tics of faults around the Kumysh Basin are rarely reported. It is a typical
basin-mountain landform with multiple levels of alluvial fans develop-
ing at the edgeof the basin. There are different heights of fault scarps de-
veloping on the alluvial fans formed by the thrust fault. Because of the
strong collision between the Indian and Eurasian plate, the Tibetan Pla-
teau had been greatly uplifted and it blocks warm wet air from the
Indian Ocean. Since the Quaternary, theKumysh Basin had been charac-
terized by a typical continental arid climate with annual precipitation
less than ~ 150 mm/yr and sparse vegetation coverage. Thus, it is an
ideal site to test the applicability of VHR stereo imagery in active
tectonics.
2.2. Dataset
Worldview-2 is a high-resolution commercial satellite launched by
DigitalGlobe in the United States on October 8, 2009. It runs in a sun-
synchronous orbit at the altitude of 770 km and the average revisit pe-
riod is 1.1 days. The Worldview-2 sensor provides panchromatic imag-
ery with the resolution of 0.46 m (0.52 m resolution imagery when
deviating from sub-satellite point of 20°) and multispectral band
image with the resolution of 1.8 5 m (2.07 m resolution imagery
when deviating from sub-satellite point of 20°) (Jawak and Luis,
2011). The data used in this paper are Worldview-2 panchromatic ste-
reo imagery with sub-meter resolution. It covers the south edge of
Kumysh Basin, and were acquired on November 25, 2013. The overlap
of images is 90% without clouds. There was no recorded strong earth-
quake that may cause surface deformation in the study area and sur-
rounding area since the images were acquired. More detailed
information is shown in Table 1.
Field surveying and mapping of GCPs and cross-scarp topographic
profiles were completed on November 15, 2017. GPS observations
were recorded using the Trimble Geo 7× handheld Global Navigation
Satellite Systems with an external Trimble Zephyr 2 antennas (Fig. 2).
This measurement system uses the FastStatic GNSS surveying mode,
whose horizontal precision is 3 mm + 1 ppm RMSand vertical precision
is 3.5 mm + 0.5 ppm RMS. The maximum distance between the rover
and master is more than 12 km (http://www.trimble.com/, accessed
on 19 December 2018). In the mountainous area, this measurement sys-
tem does not need real-time communication of two stations and would
not be affected by the terrain, which is portable and much lighter than
the real time kinematic (RTK) differential system. The coordinate sys-
tem employed was UTM Zone 45 N, each point was recorded every 1 s
until its vertical precision was stable within 2 cm. Post-processing cor-
rections were applied in the software GPS Pathfinder. Error reports
showed the error was controlled within 1–5cm,whichcanmeetthere-
quirements of this research to measure topographic profilesacrossfault
scarps and accuracy analysis.
3. Methodology
The basic theory of using satellite stereo imagery to acquire DEM is
to match the feature points according to the image overlaps and gener-
ate the 3D coordinates from these points, then further reconstruct a spa-
tial stereo model and obtain fine 3D information of the ground (Fig. 3).
Worldview-2 provides a rational polynomial coefficient (RPC) file con-
taining a rational function model representing the relationship between
object and image coordinates (Dolloff and Theiss, 2012). The ratio poly-
nomial is defined as below:
L¼LSNumlU;V;WðÞ
DenlU;V;W
ðÞ
þL0;S¼SSNumSU;V;WðÞ
DenSU;V;W
ðÞ
þS0;ð1Þ
NumlðU;V;WÞ
DenlðU;V;WÞ;
NumsðU;V;WÞ
DensðU;V;WÞis the normalized image coordinates, (U,V,W)
is the normalized ground coordinates, (L
s
,S
s
) is the normalized propor-
tional parameter, (L
0
,S
0
) is the normalized translation parameter. Num
l
Fig. 2. Geo 7x and Zephyr 2 antennas for ppGPS surveying. (a) The master station. (b) The rover station.
Table 1
The metadata information of the stereo images.
File name Rows Columns Acquisition
time
Solor elevation
Angle
Solar
azimuth
Satellite elevation
Angle
Satellite
azimuth
Cloud
coverage
Forward Image 13NOV25051910 16,280 35,820 2013-11-25
05:19:01 26.7° 170.8° 71.0° 52.2° 0%
Backward
Image 13NOV25052028 16,284 35,860 2013-11-25
05:20:21 26.7° 171.1° 61.0° 169.9° 0%
109S. Wang et al. / Geomorphology 336 (2019) 107–118
(U,V,W), Den
l
(U,V,W), Num
s
(U,V,W), Den
s
(U,V,W) can be expressed as
the following polynomial respectively:
m¼a1þa2Vþa3Uþa4Wþa5VW þa6VW þa7UW þa8V2þa9U2
þa10W2þa11 UVW þa12 V3þa13VU2þa14VW 2þa15 V2U
þa16U3þa17 UW2þa18V2Wþa19 U2Wþa20 W3
ð2Þ
The total coefficients of 4 polynomial functions are stored in one file
containing 80 rational polynomial coefficients (RPC), through which we
can calculate the ground coordinates.
Erdas 2015 Leica Photogrammetry Suite (LPS) platform was used to
carry out relative orientation, absolute orientation, automatically
matching tie points and aerotriangulation. Its core module contains
all the functions required for orthophoto mapping, supports a
Fig. 4. Extracted DEM fromstereo imagery. (a)The DEM generated without addingGCPs. (b) The zoom-inpicture of DEM withoutGCPs (c) The DEM generated withadding GCPs. (d) The
zoom-in picture of DEM with GCPs, same location as (b). The red stars represent the location of zoom-in picture. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)
Fig. 3. Principle diagram of DEM generation from stereo imagery.
110 S. Wang et al. / Geomorphology 336 (2019) 107–118
wide range of satellite sensor models, and provides an application
platform for geometric correction, orthophoto correction, information
extraction and engineering mapping. We used the eATE module to
extract digital terrestrial model as its more advanced algorithm
for high precision terrain model generation and terrain analysis
than ATE. It can output the extracted elevation information in high-
density point cloud format, which can be then filtered, and interpolated
to DEM (Zhou et al., 2015). The major steps are illustrated as
followed.
3.1. Create a block file
The initial task is to determine the geometric model, with the devel-
opment of satellites, some new commercial high-resolution satellites
starts to provide Rational Polynomial Coefficient (RPC) model only,
which contains the relationship between object coordinates and
image coordinates without sensor parameters (Büyüksalih et al., 2012;
Dolloff and Theiss, 2012). This study selected the Worldview RPC
model and UTM projection.
3.2. Automatic relative orientation
After importing forward and backward images, we need to create
pyramid model and carry out interior orientation based on the RPC
model to establish the conversion relationship between object and
image coordinates.
3.3. Absolute orientation
There are 12 GCPs evenly distributed on the image. They were sur-
veyed by Geo 7x ppGPS in the field, with centimeter level measurement
accuracy. LPS predicts the position of GCPs automatically when we im-
port them. If the matching position is inaccurate, manual fine-tuning
can be performed to improve accuracy.
3.4. Automatically match tie points
We took the corresponding image points to calculate the relation-
ship of two images according to the position of the GCPs, and to com-
plete relative orientation in LPS. Setting larger search window size can
get higher accuracy. We then carried out absolute orientation by calcu-
lating external orientation elements.
3.5. Aerial triangulation
The number of iteration is set to 10. The generated accuracy reports
show that the accuracy is 0.166 pixels with GCPs, and 0.921 pixels with-
out GCPs.
Fig. 5. (a) Elevation difference of the point cloudswith and without GCPs.X-X′and Y-Y′are two swath profiles across the faultscarps. (b) Thedistribution of elevation difference. (c)Gauss
distribution of elevation difference.
Table 2
Accuracy analysis for the DEM without ground control points.
Points number Δx(m) Δy(m) Δz(m)
1 0.01 −0.02 0.17
2−0.56 −0.80 −0.12
3−0.79 0.11 −0.42
4 0.10 0.39 0.01
5−0.70 −0.44 −1.32
6 0.13 −0.27 2.39
7−0.62 −0.03 0.42
8−0.12 −0.44 −0.45
9 0.18 −1.04 4.88
10 0.47 0.68 1.70
11 0.13 −0.10 −0.95
12 0.03 −0.03 −0.89
RMSE 0.44 0.51 1.82
111S. Wang et al. / Geomorphology 336 (2019) 107–118
Fig. 6. Compared point clouds across fault scarp. Swath profileX-X′across the fault data (a) and corresponding distribution of elevation difference (b). Projected elevation difference for
profile X-X′in latitude direction (c) and Gauss distribution of elevation difference (d). Swath profile Y-Y′across the fault in the compared point cloud data (e) and corresponding
distribution of elevation difference (f). Projected elevation difference for profile Y-Y′in latitude direction (g) and Gauss distribution of elevation difference (h). The locations of X-X′
and Y-Y′are showed in the Fig. 5a.
112 S. Wang et al. / Geomorphology 336 (2019) 107–118
3.6. Generate DEM
Using eATE module to generate point cloud file and convert it into
text format, we then filtered it in 0.5 m cell, and used continuous tension
curvature spline interpolate method (Smith and Wessel, 1990) with
tension coefficient of 0.75 to generate DEM.
The generated DEMs are shown in Fig. 4. Apart from the mountain
area that is in the low quality due to the shadow effects from the optical
satellite imagery, the topography of the rest area is well described. Com-
paring the DEMs generated with GCPs and without GCPs, the terrain
relief of the later one is more distinct and elevation value is about 1 m
larger than the DEM with no GCPs added. As can be seen from the
zoom-in pictures (Fig. 4b, d), the DEM is much smoother and has less
null value after adding GCPs.
4. Accuracy analysis
The application of digital photogrammetry, computer vision and
image matching technology achieves large area three-dimensional
measurement, but there are many factors affecting its accuracy
Fig. 7. Fourtopographicprofiles surveyedby ppGPS in the field (see Fig. 1c for locationon the DEM). (a) The locationsof Profile 1, Profile 2. (b) Fieldphoto of fault scarp of Profile1. (c) Field
photo of faultscarp of Profile 2. (d) The location of Profile 3. (e)Field photo of faultscarp of Profile 3. (f)The location of Profile 4.(g) Field photo of faultscarp of Profile 4. Thered arrows in
(a), (d), and (f) represent the fault trace.
113S. Wang et al. / Geomorphology 336 (2019) 107–118
including rational function model, distribution and measurement of
GCPs, image matching methods and so on. The accuracy of DEM de-
termines the results of quantitative active tectonics studies. In the
common accuracy assessments of active tectonics, although some
of them had performed accuracy assessment of topographic data
generated from different methods (James and Robson, 2012;
Johnson et al., 2014;Salisbury et al., 2015;Zhou et al., 2015), there
are rare researches regarding the accuracy analysis of measuring
the height of fault scarp. Previous studies have shown that the
plane accuracy of DEM generated from Worldview-2 stereo imagery
is much higher than vertical accuracy (Toutin et al., 2012), so the
goal here is (1) to assess the elevation difference between the DEM
with GCPs and without GCPs; (2) to evaluate vertical accuracy for
measuring the height of fault scarp by comparing topographic pro-
files derived from ppGPS and DEM.
4.1. Elevation comparison of generated DEM with and without GCPs
Generating high quality DEM without GCPs has been applied in
many studies (Poli et al., 2015;Toutin et al., 2012;Zhou et al., 2015).
Poli et al. (2015) applied well-defined texture points in theoriented ste-
reo pair as GCPs and used these points for image orientation, the RMSEs
they got was around 0.3 m in plane and elevation; Toutin et al. (2012)
processed Worldview-1 and -2 stereo images without GCPs, elevation
Fig. 9. Comparison of topographic profiles 1–4 across the fault scarp, derived from theDEM and that surveyed by ppGPS in the field. (a), (e), (i), (m) Digital elevation model of four fault
scarps. (b) (f) (g) (n) arethe histogram graphs of elevation differences distribution. (c), (g), (k),(o) Searching for the elevation correction valueΔH. Black solidlines in (d), (h), (l),(p) are
the profiles surveyed by ppGPS; Grey solid lines are profiles extracted from the DEM; Grey dashed lines are profiles after elevation correction.
Fig. 8. The two fault outcrops detected in the field. Red arrows indicate the faults; black arrows indicate the fault strike. (a) Deformation evidence shows the displaced gravel layer and
unconformity between strata. (b) The wedge-shaped layers deposited in front of the scarp.
114 S. Wang et al. / Geomorphology 336 (2019) 107–118
errors with 68% confidence level compared to the accurate GCPs was
10–20 cm; Zhou et al. (2015) extracted DEM from Pleiades stereo imag-
ery without GCPs and compared it with LiDAR data, the results show
that the vertical accuracy of DEM can reach about 0.3 m. A small number
of GCPs can improve the position accuracy of rational function model
(Aguilar et al., 2013). In this study, 12 GCPs selected from the intersec-
tion of road and river on the images were uniformly distributed and ac-
curately surveyed in the field. We generated DEM from stereo imagery
without GCPs based on rational function model and used 12 GCPs mea-
sured in the field (Fig. 1) as check points to evaluate the plane and ver-
tical accuracy. The root mean square error (RMSE) in Table 2 clearly
shows that the plane accuracy is better than the vertical accuracy.
Thus, it is necessary to compare the elevation difference of DEMs with
GCPs and without GCPs.
In order to discuss the elevation difference between the DEM de-
rived with and without GCPs, the point cloud data were compared in
Cloud Compare software. Using point cloud data with GCPs as a refer-
ence, we calculated the distance between them (Fig. 5a). Distribution
of elevation difference histogram graph (Fig. 5b) shows 95.7% points'
value is less than 0.8 m. The value of points without GCPs is smaller
than that with GCPs and the overall elevation difference is −1.2–
0.4 m, the average difference is −0.57 m. The differences of the point
cloud in the mountain area are higher and reach 1.0 m, which might
be due to the large elevation variation, shadow effects of the stereo im-
agery etc. The average elevation differences in alluvialand village have a
lower range of 0.5 m. This part is also the main study area that the fault
scarps developed.
The goal of this research is to obtain the most accurate height of fault
scarp. Therefore, we extracted two swath profiles across the fault in the
elevation difference point cloud data (Fig. 6a, e) and projected the coor-
dinates in the latitude direction (Fig. 6c, g). Compared with the overall
elevation difference of −1.2–0.4 m, the difference near the fault scarp
is −1.1–0.01 m. The average elevation difference is −0.4 m, with only
a few abnormal points. The result of vertical difference distribution
shows that the position accuracy of rational function model of
Worldview-2 satellite is high. We could also get high quality DEM
under harsh natural condition without surveying GCPs in the field.
4.2. Comparison of topographic profiles
Measuring height of fault scarp requires fitting the hanging wall and
footwall according to the topographic profile. Hence, accuracy of the
surveyed topographic profile is the basis for obtaining accurate height
of a fault scarp. There are a number of alluvial fans displaced by Kumysh
fault, which are obvious on the Google Earth images (Fig. 7). Two fault
outcrops were detected in the field (Fig. 8). We chose four sites to get
topographic profiles derived from DEM and surveyed by ppGPS across
the fault scarps (Fig. 1,Fig. 7). The results are projected in the latitude
direction (Fig. 9). The overall outline and trend of profiles measured
by two different methods are in good agreement, but there is a certain
elevation deviation. This might due to the coordinates of the reference
station calculated by the satellite are different from GCPs survey or
lose of fine feature due to the resolution limit of satellite imagery. So,
we first made the elevation correction for each profile of the generated
DEM, and then calculated the RMSE of two-method profiles to evaluate
the elevation difference between the DEM and ppGPS. In the field
measurement of topographic profiles using ppGPS, the point was
surveyed about every 0.25 m, while the point interval was about
0.5 m surveyed from DEM because of the spatial resolution. So, we
used the points on the ppGPS profiles as reference, searched the
nearest points on the DEM profiles as corresponding points (Fig. 10),
and calculated elevation difference di between the two points.
Assuming that the translation of elevation correction is ΔH,which
makes the minimum quadratic sum of elevation difference after transla-
tion [(d1−ΔH
2
)+(d2−ΔH
2
)+ …+(dn −ΔH
2
)]. After the trans-
lation, the ppGPS data is used as the reference value, and the root mean
square error (RMSE)of the whole profile iscalculated to evaluate the ac-
curacy of DEM data to portray the profile of fault scarp.
It can be seen from the profile comparison and difference histogram
in Fig. 8 that there is deviation between two groups of profiles. The off-
set ranges from −2.82–4.47 m and the average RMSE of the two groups
is 0.29 after elevation correction. Hence, the average difference is 0.29 m
for the topographic profiles across fault scarps, measured from the ex-
tracted DEM and ppGPS in the field respectively. (See Table 3.)
4.3. Comparison of the height of fault scarps
In the study of active tectonics, measuring the offset along the fault is
essential. When measuring the height of fault scarp, it is necessary to be
perpendicular to thefault scarps and select the geomorphic surface with
less gully and erosion.The results are projected in the vertical direction
of scarps and fitted to the hanging wall, footwall (h1, h2). We then used
Table 3
DEM accuracy analysis for profiles 1–4.
Profiles Profile1 Profile2 Profile3 Profile4
|di_min|(m) 1.85 2.52 0.87 1.06
|di_max|(m) 3.43 4.48 2.38 2.82
Translation ValueΔH(m) 2.76 3.24 −1.42 −1.37
RMSE(m) 0.27 0.33 0.29 0.26
Fig. 10. Schematic diagram of calculating elevation difference of surveyed profiles between DEM and ppGPS.
115S. Wang et al. / Geomorphology 336 (2019) 107–118
the average height of hanging wall and footwall to calculate the height
of fault scarps and also computed the standard deviation between the
elevation and average values as the measurement error (Hetzel et al.,
2004). Furthermore, we compared four groups of scarps' heights to an-
alyze the accuracy for measuring the height of scarp, by comparing the
heights measured from the extracted DEM and ppGPS field survey
(Fig. 11).
The RMSE between the heights of fault scarps measured by DEM and
ppGPS is 0.25 m, and the deviation range is −0.3 to −0.17 m. It is much
smaller than the elevation difference of −2.82 to −4.47 m for the topo-
graphic profiles. The measurement accuracy is significantly increased,
which indicates that the relative error is small even though the overall
deviation is large. So, the DEM generated from stereo imagery has
high relative accuracy and can identify fault scarps higher than
0.25 m. However, in the quantitative study of active tectonics, it is not
enough to only use height difference of hanging wall and footwall as
true heightof fault scarp. The erosion of hanging wall and accumulation
of footwall caused by geomorphological evolution cannot be ignored.
We should use trench near the fault to identify the true height of fault
scarp formed by paleoearthquakes according to the stratigraphic
information.
5. Discussion
We suggest that the vertical uncertainty is mainly due to the follow-
ing three factors:
(1) Error introduced by the complex topography and geomorphol-
ogy. In the mountainous areas with steep slopes, large relief am-
plitude, vegetation (Chaplot et al., 2006;Fisher and Tate, 2006),
due to the influence of solar elevation angle, azimuth, and sensor
angle, all rises shadows that lead to low contrast and poor
matching accuracy. The accuracy of DEM is the highest in the
plain area without vegetation in our study, followed by the resi-
dential areas and the worst in the mountain areas (Fabris and
Pesci, 2005;Fisher and Tate, 2006).
(2) Error caused by rational function model. The rational function
model uses “terrain independent”method relating ground coor-
dinates to corresponding image coordinates by a ground-to-
image polynomial generated from the physical sensor model,
the effect of RPC errors is the same as physical sensor (Aguilar
et al., 2013;Dolloff and Theiss, 2012), affected by satellite orbit
parameters, sensor parameters and flight attitude. The error of
generated DEM in the absence of GCPs mostly comes from ratio-
nal function model error. Topographic profiles measured by DEM
and DGPS have deviation along the whole line, but the
distribution is not uniform. This may be largely affected by ratio-
nal function model.
(3) Data errors introduced by GPS surveying and mapping. The dis-
tribution and number of GCPs have a great impact on the
accuracy of DEM generation (Aguilar et al., 2013;Poli et al.,
2015). Since there are few intersection points of road and drain-
age with obvious features, errors will be caused when the sur-
veyed GCPs are projected onto the image and adjusted manually.
(4) Cloud noise and vegetation of stereo imagery are also non-
negligible factors that affect the quality of DEM. The existence
of cloud and mist makes it difficult to obtain clear feature infor-
mation and image matching. This problem can be mostly over-
come by three methods: 1) Using multitemporal images to
mosaic the cloud-free areas and to generate a new one (Lin
et al., 2013;Tseng et al., 2008); 2) Using the multi-spectral data
to detect and remove clouds (Wang et al., 2005); 3) Treating
cloud as noise to remove (Feng et al., 2004). In addition, point
cloud data obtained from stereo imagery contain information of
both ground points and vegetation. Several filtering algorithms
have been developed for this drawback (Deng and Shi, 2013;
Hu et al., 2014;Sithole and Vosselman, 2004), which have dispa-
rate performance in different landscape and environment. We
should choose appropriate cloud denoising and vegetation elim-
ination method according to our study area.
In addition to the above factors, spatial resolution of satellite images,
band characteristics and the method of interpolation could also affect
DEM-generation accuracy to some extent.
6. Conclusion
This paper mainly introduces the basic principle and processing flow
of DEM generation using Worldview-2 stereo imagery by extracting
0.5 m resolution DEM of the southern margin of Kumysh Basin in the
eastern Tian Shan. Three main findings can be summarized as follows:
(1) Using GCPsas check points to evaluate accuracy of the DEM with-
out GCPs, the vertical accuracy is ~1.82 m. The elevation differ-
ence between the DEM with and without GCPs is −1.2–0.4 m
while the elevation difference is −1.1–0.01 m for the point
cloud swath profiles across the fault scarps;
(2) Under the condition of 12 GCPs, the morphological difference be-
tween the two methods is ~0.29 m, as obtained by comparing the
topographic profiles across fault scarps derived from the gener-
ated DEM and ppGPS field survey;
(3) In the research of active tectonic, by analyzing the height of the
fault scarps using DEM and ppGPS, we got the accuracy of scarp
height measured by DEM is 0.25 m.
The results show that Worldview-2 stereo imagery is feasible to es-
tablish accurate DEM with very few GCPs. In the absence of GCPs, the
Fig. 11. Comparison of the height of fault scarps measured from the DEM and ppGPS survey in the field.
116 S. Wang et al. / Geomorphology 336 (2019) 107–118
vertical accuracy of DEM is higher than 1.82 m. The topographic profiles
can be accurately depicted under a small number of GCPs and the mea-
surement accuracy can reach 0.29 m. With this method, we can obtain
the fine geometric structure of faults quickly and accurately. It also pro-
vides us with reference to choose trench-excavation site and
chronological-sample collection site. Generating DEM from stereo im-
agery costs less than airborne LiDAR and has a wide flight range of cov-
erage compared with UAV, which has a broad application prospect in
the study of active tectonic.
There are still some deficiencies in this paper: 1) The study area is lo-
cated in the piedmont alluvial fans. There is a certain error in GCPs sur-
vey because of the lack of landmark features; 2) Mountain shadow of
the stereo imagery we purchased is severe, which would affect the
DEM accuracy. This error may be carried over to our follow-up analysis.
Fortunately, the faults we concerning with are located on the alluvial
fans uncovered by shadows rather than on the mountain, but it is nec-
essary to use mountain part to carry out analysis of tectonic-
geomorphology, which will affect the results; 3) Due to the limitation
of Worldview-2 stereo imagery resolution, this method will be ineffec-
tive for small fault scarps less than 0.25 m.
Acknowledgements
This research was funded by National Nonprofit Fundamental
Research Grant of China (IGCEA1607，IGCEA1803), and National Natu-
ral Science Foundation of China (41472201, 41761144071). We thank
Ming Ai dueto all the heights of fault scarps aremeasured with the help
of him. We appreciate Jiahong Luo for his help in writing Matlab code.
The authors would also like to thank the editor and reviewers for their
useful comments.
References
Aguilar, M.A., Saldana, M.D.M., Aguilar, F.J., 2013. Generation and quality assessment of
stereo-extracted DSM from GeoEye-1 and WorldView-2 imagery. IEEE Trans. Geosci.
Remote Sens. 52 (2), 1259–1271.
Angster, S., Wesnousky, S., Huang, W.l., Kent, G., Nakata, T., Goto, H., 2016. Application of
UAV photography to refining the slip rate on the Pyramid Lake fault zone, Nevada.
Bull. Seismol. Soc. Am. 106 (2), 785–798.
Barišin,I., Hinojosa-Corona, A., Parsons, B., 2015. Co-seismic vertical displacements from a
single post-seismic lidar DEM: example from the 2010 El Mayor-Cucapah earth-
quake. Geophys. J. Int. 202 (1), 328–346.
Bi, H., Zheng, W., Ge, W., Zhang, P., Zeng, J., Yu, J., 2018. Constraining the distribution of
vertical slip on the South Heli Shan Fault (Northeastern Tibet) from high-resolution
topographic data. J. Geophys. Res. Solid Earth 123 (3), 2484–2501.
Binet, R., 2005. Horizontal coseismic deformation of the 2003 Bam (Iran) earthquake
measured from SPOT-5 THR satellite imagery. Geophys. Res. Lett. 32 (2).
Burbank, D.W., Anderson, R.S., 2011. Tectonic Geomorphology, Second Edition. Environ-
mental & Engineering Geoscience, 19(2), 198–200.
Burchfiel, B.C., Brown, E.T., Qidong, D., Xianyue, F., Jun, L., Molnar, P., Jianbang, S.,
Zhangming, W., Huichuan, Y., 1999. Crustal shortening on the margins of the Tien
Shan, Xinjiang, China. Int. Geol. Rev. 41 (8), 665–700.
Büyüksalih, G., Baz, I., Alkan, M., Jacobsen, K., 2012. Dem Generation with WORLDVIEW-2
Images. Journal of Epidemiology & Community Health, XXXIX-B1(12), 203–207.
Chaplot, V., Darboux, F., Bourennane, H., Leguédois, S., Silvera, N., Phachomphon,K., 2006.
Accuracy of interpolation techniques for the derivation of digital elevation models in
relation to landform types and data density. Geomorphology 77 (1–2), 126–141.
Deilami,K., Hashim, M., 2011.Very high resolution optical satellites for DEM generation: a
review. Eur. J. Sci. Res. 49 (4), 542–554.
Deng, S.S., Shi, W.Z., 2013. Integration of different filter algorithms for improving the
ground surface extraction from airborne LIDAR data. Implement. Sci., XL-2/W1(2),
105–110.
d'Oleire-Oltmanns, S., Marzolff, I., Peter, K., Ries, J., 2012. Unmanned aerial vehicle (UAV)
for monitoring soil erosion in Morocco. Remote Sens. 4 (11), 3390–3416.
Dolloff, J.T., Theiss,H.J., 2012. Temporal correlation of metadata errors for commercial sat-
ellite images: representation and effects on stereo extraction accuracy. Int. Arch.
Photogramm. Remote. Sens. Spat. Inf. Sci. 39, 215–223.
Dumitru, T.A., Zhou, D., Chang, E.Z., Graham, S.A., Hendrix, M.S., Sobel, E.R., Carroll, A.R.,
2001. Uplift, exhumation, and deformation in the Chinese Tian Shan. Memoir of the
Geological Society of America, 194, 71–99.
Fabris, M., Pesci, A., 2005. Automated DEM extraction in digital aerial photogrammetry: pre-
cisions and validation for mass movement monitoring. Ann. Geophys. 48 (6), 57–72.
Feng, Chun, Ma, Jianwen, Dai, Qin, Chen, Xue, 2004. An improved method for cloud re-
moval in ASTERdata change detection,IEEE InternationalGeoscience & Remote Sens-
ing Symposium.
Fieber, K.D., Mills, J.P., Miller, P.E., Clarke, L., Ireland, L., Fox,A.J., 2018. Rigorous 3D change
determination in Antarctic Peninsula glaciers from stereo WorldView-2 and archival
aerial imagery. Remote Sens. Environ. 205, 18–31.
Fisher, P.F., Tate, N.J., 2006. Causes and consequences of error in digital elevation models.
Prog. Phys. Geogr. 30 (4), 467–489.
Hashimoto, M., Enomoto, M., Fukushima, Y., 2010. Coseismic deformation from the 2008
Wenchuan, China, earthquake derived from ALOS/PALSAR images. Tectonophysics
491 (1–4), 59–71.
Hetzel, R., Tao, M., Stokes, S., Niedermann, S., Ivy-Ochs, S., Gao, B., Strecker, M.R., Kubik,
P.W., 2004. Late Pleistocene/Holocene slip rate of the Zhangye thrust (Qilian Shan,
China) and implications for the active growth of the northeastern Tibetan Plateau.
Tectonics, 23(6), n/a-n/a.
Hu, H., Ding, Y., Zhu, Q., Wu, B., Lin, H., Du, Z., Zhang, Y., Zhang, Y., 2014. An adaptive sur-
face filter for airborne laser scanning point clouds by means of regularization and
bending energy. ISPRS J. Photogramm. Remote Sens. 92 (2), 98–111.
Huang, W.L., Yang, X.P., Li, A., Pierce, I.K.D., Thompson, J.A., Angster, S.J., Zhang, L., 2015.
Late Pleistocene shortening rate on the northern margin of the Yanqi Basin, south-
eastern Tian Shan, NW China. Seismol. Geol. 112, 11–24.
James, M.R., Robson, S., 2012. Straightforward reconstruction of 3D surfaces and topogra-
phy with a camera: Accuracy and geoscience application. J. Geophys. Res. Earth, 117
(F3), n/a-n/a.
Jawak, S.D., Luis, A.J., 2011. Applications of WorldView-2 satellite data for extraction of
polar spatial information and DEM of Larsemann Hills. East Antarctica, International
Conference on Earth Sciences & Engineering.
Johnson, K., Nissen, E., Saripalli, S., Arrowsmith, J.R., Mcgarey, P., Scharer, K., Williams, P.,
Blisniuk, K., 2014. Rapid mapping of ultrafine fault zone topography with structure
from motion. Geosphere 10 (5), 969–986.
Kondo, H., Toda, S., Okumura, K., Takada, K., Chiba,T., 2008. A fault scarp in an urban area
identifiedby LiDAR survey: a case studyon the Itoigawa–Shizuoka TectonicLine, cen-
tral Japan. Geomorphology 101 (4), 731–739.
Lin, C.H., Tsai, P.H., Lai, K.H., Chen, J.Y., 2013. Cloud removal from multitemporal satellite
images using information cloning. IEEE Trans. Geosci. Remote Sens. 51 (1), 232–241.
Massonnet, B.D., Rossi, M., Carmona, C., et al., 1992. The displacement field of the Landers
earthquake mapped by radar interferometry, Nature.
Mitchell, H.L., Chadwick, R.G., 1999. Digital photogrammetric concepts applied to surface
deformation studies. Geomatica (4), 53.
Mora, P., Baldi, P., Casula, G., Fabris, M., Ghirotti, M., Mazzini, E., Pesci, A., 2003. Global po-
sitioning systems and digital photogrammetry for the monitoring of mass move-
ments: application to the Ca' di Malta landslide (northern Apennines, Italy). Eng.
Geol. 68 (1), 103–121.
Poli, D., Remondino, F., Angiuli, E., Agugiaro, G., 2015. Radiometric and geometric evalua-
tion of GeoEye-1, WorldView-2 and Pléiades-1Astereo images for 3D information ex-
traction. ISPRS J. Photogramm. Remote Sens. 100, 35–47.
Reinartz, P., Müller, R., Lehner, M., Schroeder, M., 2006. Accuracy analysis for DSM and
orthoimages derived from SPOT HRS stereo data using direct georeferencing. ISPRS
J. Photogramm. Remote Sens. 60 (3), 160–169.
Ren, Z., Zhang, Z., Dai, F., Yin, J., Zhang, H., 2014. Topographic changes due to the 2008
Mw7.9 Wenchuan earthquake as revealed by the differential DEM method. Geomor-
phology 217, 122–130.
Ren, Z., Zhang, Z., Chen, T.,Yan, S., Yin, J., Zhang, P., Zheng, W.,Zhang, H., Li, C., 2015. Clus-
tering of offsets on the Haiyuan fault and their relationship to paleoearthquakes.
Geol. Soc. Am. Bull. B31155, 31151.
Ren, Z., Zhang,Z., Zhang, P., 2018. Different earthquake patterns f or two neighboring fault
segments within the Haiyuan Fault zone. Earth Planet. Phys. 2 (1), 1–7.
Saldaña, M.M., Aguilar, M.A., Aguilar, F.J., Fernández, I., 2012. DSM extraction and evalua-
tion from GeoEye-1 stereo imagery, XXII Congress of the International Society for
Photogrammetry and Remote Sensing. pp. 113–118.
Salisbury, J.B., Haddad, D.E., Rockwell, T., Arrowsmith, J.R., Madugo, C., Zielke, O., Scharer,
K., 2015. Validation of meter-scale surface faulting offset measurements from high-
resolution topographic data. Geosphere 11 (6) GES01197.01191.
San, B.T., Suzen, M.L., 2005. Digital elevation model (DEM) generation and accuracy as-
sessment from ASTER stereo data. Int. J. Remote Sens. 26 (22), 5013–5027.
Shaker, A., Yan, W.Y., Easa, S., 2010. Using stereo satellite imagery for topographic and trans-
portation applications: an accuracy assessment. Geosci. Remote Sens. 47 (3), 321–337.
Shean, D.E., Alexandrov, O., Moratto, Z.M., Smith, B.E., Joughin, I.R., Porter, C., Morin, P.,
2016. An automated, open-source pipeline for mass production of digital elevation
models (DEMs) from very-high-resolution commercial stereo satellite imagery.
ISPRS J. Photogramm. Remote Sens. 116, 101–117.
Sithole, G., Vosselman, G., 2004. Experimental comparison of filter algorithms for bare-
Earth extraction from airborne laser scanning point clouds ☆. ISPRS J. Photogramm.
Remote Sens. 59 (1), 85–101.
Smith, W.H.F., Wessel, P., 1990. Gridding with continuous curvature splines in tension.
Geophysics 55 (3), 293–305.
Stumpf, A., Malet, J.P., Allemand, P., Ulrich, P., 2014. Surface reconstruction and landslide
displacement measurements withPléiades satelliteimages. ISPRS J. Photogramm. Re-
mote Sens. 95, 1–12.
Thompson, S.C., Weldon, R.J., Rubin, C.M., Abdrakhmatov, K., Molnar, P., Berger, G.W.,
2002. Late Quaternary slip rates across the central Tien Shan, Kyrgyzstan, central
Asia. J. Geophys. Res. Solid Earth, 107 (B9), ETG 7–1-ETG 7-32.
Toutin, T., Schmitt, C.V., Wang, H., 2012. Impact of no GCP on elevation extraction from
WorldView stereo data. ISPRS J. Photogramm. Remote Sens. 72, 73–79.
Tseng, Din-Chang, Tseng, Hsiao-Ting, Chien, Chun-Liang, 2008. Automatic cloud removal
from multi-temporal SPOT images. Appl. Math. Comput. 205 (2), 584–600.
Turker, M., Cetinkaya, B., 2005. Automatic detection of earthquake-damaged buildings
using DEMs created from pre- and post-earthquake stereo aerial photographs. Int.
J. Remote Sens. 26 (4), 823–832.
117S. Wang et al. / Geomorphology 336 (2019) 107–118
Wang, C., 2004. Crustal structure of the northern margin of the eastern Tien Shan, China,
and its tectonic implications for the 1906 M?7.7 Manas earthquake. Earth Planet. Sci.
Lett. 223 (1–2), 187–202.
Wang, Z., Jin, J., Liang, J., Yan, K., Peng, Q., 2005. A new cloud removal algorithm for multi-
spectral images. Proceedings of SPIE - the International Society for Optical Engineer-
ing, vol. 6043, 60430W-60430W-60411.
Westoby, M.J., Brasington, J., Glasser, N.F., Hambrey, M.J., Reynolds, J.M., 2012. ‘Structure-
from-motion’photogrammetry: a low-cost, effective toolfor geoscience applications.
Geomorphology 179, 300–314.
Wright, T.J., 2004. Toward mapping surface deformation in three dimensions using InSAR.
Geophys. Res. Lett. 31 (1).
Xiao, W., Windley, B.F., Allen, M.B., Han, C., 2013.Paleozoic multiple accretionary and col-
lisional tectonics of the Chinese Tianshan orogenic collage. Gondwana Res. 23 (4),
1316–1341.
Yin, A., Nie, S., Craig, P., Harrison, T.M., Ryerson, F.J., Xiangli n, Q., Geng, Y., 1998. Late
Cenozoic tectonic evolution of the southern Chinese Tian Sh an. Tectonics 17 (1),
1–27.
Zebker, H.A., Rosen, P.A., Goldstein, R.M., Gabriel, A., Werner, C.L., 1994. On the derivation
of coseismic displacement fields using differential radar interferometry: the Landers
earthquake. J. Geophys. Res. Solid Earth 99 (B10), 19617–19634.
Zhou, Y., Parsons, B., Elliott, J.R., Barisin, I., Walker, R.T., 2015. Assessing the ability of
Pleiades stereo imagery to determine height changes in earthquakes: a case study
for the El Mayor-Cucapah epicen tral area. J. Ge ophys. Res. Sol id Earth 120 (12 ),
8793–8808.
Zielke, O., Arrowsmith, J.R., Grant Ludwig, L., Akciz, S.O., 2010. Slip in the 1857 and earlier
large earthquakes along the Carrizo Plain. San Andreas Fault. Science 327 (5969),
1119–1122.
118 S. Wang et al. / Geomorphology 336 (2019) 107–118
