DOÇENTLİK SONRASI

YAYIN

Height biases of SRTM DEM related to EGM96:

from a global perspective to regional practice
A. Üstün1, R. A. Abbak∗ 2 and E. Zeray Öztürk2
It is investigated to what extent EGM96 affects the accuracy of digital elevation model (DEM)
produced from the shuttle radar topography mission (SRTM). Global and regional analysis of
EGM96 compared with EGM2008 indicate that locally there are large differences distorting to
the accuracy level of SRTM DEM. In the absolute sense, the overall geoid differences
throughout 1 × 1 arc-degree tiles reach −5 m in the northeast and 2–3 m in the southern parts of
Turkey. A numerical investigation over the test profiles of 200–700 km length running at various
directions proves that a possible vertical datum change from EGM96 to EGM2008 yields
systematically more accurate height information with an improvement of up to 2.5 m. A GPS-
levelling traverse of about 900 km length points out some key patterns of this recovery.
Consequently, a correction for the present version of SRTM DEM should be considered in
critical implementations of Earth sciences like geoid or water flow modelling, especially for areas
where EGM96 shows weak performance.
Keywords: Accuracy assessment, Geoid height, Digital elevation model, Earth gravitational model, Shuttle radar topography mission

Introduction usually 3 × 3 arc-second resolution (1 × 1 arc-second

grid is publicly available in the USA). (ii) A limited spatial
A digital elevation model (DEM) is a representation of distribution of control data tends to over estimate the
the Earth’s physical surface in a computer environment. accuracy of the DEM with regard to the topographical
In 2000, a DEM with 3×3 arc-second resolution was pro- variability. (iii) Usually vertical datums of topographical
duced from the shuttle radar topography mission (SRTM) heights acquired from different sources do not coincide.
by the national aeronautics and space administration Additionally, as a function of the position, geoid model
(NASA). It was found to be one of the most precise global that serves as vertical datum can introduce systematic dis-
DEMs worldwide (Hirt et al., 2010; Jing et al., 2014). tortions of different wavelengths. Hence the height infor-
Detailed DEMs are exploited for a wide range of geos- mation from the control points and the DEM may not be
cientific applications, such as gravity forward modelling directly comparable due to systematic biases. Therefore, a
in physical geodesy, geomorphological simulation in corrector surface or an advanced process is required to
geology, water flow modelling in hydrology, flood risk overcome these drawbacks.
analysis in geographic information systems, topographic In this study, we are focusing on the vertical datum
phase correction in interferometric synthetic aperture problems in the SRTM DEM dataset and resolving it
radar (InSAR) processing etc. by means of better geoid models. According to the
Before employing a DEM, users should pay attention assessment report of the SRTM products, the DEM per-
to its accuracy in project area. Thus the evaluation of formance against the ground-truth data is satisfactory at
DEM errors and the estimation of their impact are critical the level of the mission objectives (Rodriguez et al.,
for the success of a project which is also relevant for geoid 2005). The accuracy validation studies for the C-band
determination (Merry, 2003; Kiamehr and Sjöberg, SRTM data have reported that the absolute and relative
2005), refinement of gravity data bases (Rexer and Hirt, height errors are below 9 and 6 m, respectively, for the
2014), terrain morphology studies (Mukherjee et al., 90% confidence level (e.g. Rabus et al., 2003; Farr
2013) and glacier volume change (Berthier et al., 2006). et al., 2007; Zielinski and Chmiel, 2007; Bildirici
However, some aspects play a substantial role in the et al., 2010). The vertical datum of the SRTM products
final judgment: (i) The spatial structure of the SRTM is the Earth Gravitational Model 1996 (EGM96)
DEM differs from geodetic data, i.e. ground control (Lemoine et al., 1998). On the other hand, the recently
points (GCPs). A specific SRTM height value has to be published Earth Gravitational Model 2008 (EGM08)
predicted from the near neighboring grid points set of (Pavlis et al., 2012) is more accurate than EGM96
(for details see the papers in Newton’s Bulletin Nr. 4
1
Department of Geodesy and Geoinformation Engineering, Kocaeli Univer- entitled ‘External Quality Evaluation Reports of
sity, Kocaeli, Turkey EGM2008’). Thus the present study researches the influ-
2
Department of Geomatics Engineering, Selcuk University, Konya, Turkey
ence of EGM96 on the accuracy of SRTM DEM by
∗
Corresponding author. email: aabbak@selcuk.edu.tr

© 2016 Survey Review Ltd

Received 5 November 2015; accepted 16 July 2016
26 DOI 10.1080/00396265.2016.1218159 Survey Review 2018 VOL 50 NO 358
 Üstün et al. Height biases of SRTM DEM related to EGM96

replacing the default vertical datum of the SRTM with

EGM08 over the territory of Turkey.
The paper starts with the methodology for the essen-
tials of SRTM DEM production and its vertical datum
transformation. Then the computational scheme of absol-
ute and relative validations is shortly summarised. A brief
review of the study area and input data is given as a case
study. A numerical validation of the SRTM DEM from
the point of view of vertical datum is performed in the
study area through kinematic differential global position-
ing system (DGPS) tracks. The impact of datum change
on the accuracy of the SRTM DEM is examined via
GPS-levelling data extending along a levelling line from
the Mediterranean Sea to the Black Sea. Finally some
remarks conclude the present paper.

Methodology
SRTM and vertical datum transformation
In 2000, the NASA, the National Imagery and Mapping
Agency (NIMA), the German Space Agency (DLR) and
the Italian Space Agency jointly carried out the SRTM
which yielded the most complete high-resolution digital
topographic database of the earth to date. SRTM radar
mapping technique is based on two-dimensional measure-
ments attached to a baseline on which inboard and out-
board SAR antennas are mounted. The antennas are
tied together with a mast of 60 m on the Endeavour space-
craft. Interferometric phase measurements of a target on 1 Geometrical aspects of across-track interferometry for
the earth surfaces are two-dimensional arrays of complex- SRTM. Elevations above sea level can be derived by
valued pixels in terms of azimuth (along the orbit) and using geoidal heights estimated from an earth gravity
range (sensor look direction). Phase difference for common model
pixels from stereo image pairs can be converted into a
topographic height value by an expression (cf., Farr et al.,
data coverage and quality are available. For example, it
2007),
can be seen that the overall accuracy of EGM96 geoid
# ! " $
lf remains below 1 m, globally, as listed in the page entitled
h = hS − r1 cos sin− 1 + a (1) ‘Evaluation of models’ on ICGEM website (see e.g.
2p B
ICGEM, 2015). However, it has been reported that differ-
where r1 is range between the target and space vehicle; α is ences of geoidal heights can locally exceed this threshold
the baseline roll angle, λ is wavelength of radar signal, B is by a substantial margin up to 4–5 m over a rough topogra-
the known baseline length between two antennas. As phical region (e.g. Banerjee et al., 1999; Tocho et al., 2008;
depicted in Fig. 1, h and hS that represent ellipsoidal heights Ustun and Abbak, 2010). Taking account of maximum
of topography and spacecraft positions, respectively, are spatial resolution of reference geoid model, i.e. 55 km
referred to WGS84 ellipsoid. for EGM96, topographical heights may be systematically
The C-SAR processing steps were performed at Jet Pro- influenced by the biases in a wide area.
pulsion Laboratory in the USA to transform the raw data By the inclusion of new satellite gravity measurements,
of nearly 80 TB into final topographical heights in the a high degree EGM (like EGM08) reduced the geoid
DTED standards. Finally, the vertical datum of the model errors in specific areas to decimetre or sub metre
DEM data in a regular grid was transferred to EGM96 level. Moreover, an increase in spatial resolution of
geoid, complete to degree and order 360. In order to geoid model leads to much finer and realistic represen-
obtain heights above geoid model which is represented tation for topographical and hydrological information.
by a set of spherical harmonic coefficients, a simplified From this perspective, an improvement in SRTM heights
conversion can be written by can be done by transforming the vertical datum from
EGM96 to EGM08 on the basis of equation (2). Accord-
H = h− N (2)
ing to Fig. 1, the SRTM datum change can be carried out
using geoidal heights,
Before the era of dedicated satellite gravity missions
(e.g. CHAMP, GRACE and GOCE) the earth gravity HEGM08 = HEGM96 + NEGM96 − NEGM08
models (EGM) suffered from a lack of uneven distri- = HEGM96 + dN (3)
bution of satellite and terrestrial gravity observations at
long- and medium-length scales. These models whose per- where dN is a local correction term or geoid height differ-
formances have been commonly tested in terms of geoid ence between EGM96 and EGM08.
heights comparing with GPS-levelling data have been A quantitative analysis on the correction terms over
more successful in some parts of the globe where better land masses mapped by SRTM spacecraft gives an idea

Survey Review 2018 VOL 50 NO 358 27

Üstün et al. Height biases of SRTM DEM related to EGM96

2 Global map of height biases in terms of geoid difference (dN) for SRTM DEM coverage area

about the necessity of this conversion. Figure 2 shows the Since 7-parameter model gave us the most reasonable
absolute discrepancies of geoid heights on a grid with 25 results, it was adopted in this study. The surface model
arc-minute between 60° north and 56° south latitudes is described as follows:
for only land masses. The differences that vary in the ⎡ ⎤
range of −9.5 and 12.0 m are highly correlated with topo- cos wi cos li
graphy and inaccessible regions where terrestrial gravity ⎢ cos wi sin li ⎥
⎢ ⎥
data were not available for particularly EGM96 model. ⎢ sin wi ⎥
⎢ ⎥
The magnitudes of dN in these areas are almost the a= ⎢ cos wi sin wi cos li /Wi ⎥ (5)
⎢ ⎥
same or slightly larger than validation levels of SRTM ⎢ cos wi sin wi sin li /Wi ⎥
⎢ ⎥
heights presented in the literature. Turkish territory has ⎣ sin2 wi /Wi ⎦
taken a portion of the model deviations from each 1
other. In the next section, we will focus on a case study
of EGM96 induced errors in SRTM DEM over the Turk- where wi and li are the ellipsoidal coordinates of ith check
ish peninsula. point and
,+++++++++++++++
Validation of SRTM heights Wi = 1 − e2 sin2 wi (6)
A validation process for SRTM heights was required
before performing a regional assessment of the impact where e denotes the first eccentricity of the reference
of EGM96. The theoretical definition of the absolute ellipsoid.
and relative methods which are suitable for the compari- The a vector contains the parameters of three trans-
son of both EGMs and DEMs are outlined briefly. lations, three rotations and one scale factor, respectively
(Kotsakis and Sideris, 1999). After the calculation of the
Absolute validation design matrix by filling a vector for each point, the para-
Since a DEM is subject to errors distributed systemati- metric model is solved by least-squares adjustment which
cally (e.g. datum shifts, long-wavelength information minimises the sum of squares of the residuals ɛ. Finally,
etc.), two kinds of height data are not directly compar- the estimated root mean squared error (RMSE) is con-
able. In order to remove these errors, a special corrector sidered as the accuracy of the model in question. How-
surface should be created before the detailed discussion. ever, the final RMSE is not exact error of the model
Various corrector surface models can be found in the lit- since it still contains random errors from levelling and
erature for this purpose (e.g. Kotsakis and Sideris, 1999; DGPS measurements as well as DEM interpolation.
Abbak, 2014). The corrector surface can be designed with less parameter
The success of the surface model which is a function of (for example 1 or 4-parameter) to check the presence of
position in geographic coordinates can be evaluated based long-wavelength errors in DEM and/or observations.
upon the following equation:
H SRTM − H GT = aT x + 1 (4)
where H GT and H SRTM are the heights from levelling or Relative validation
DGPS observations (i.e. Ground Truth) and SRTM In order to better understand the real potential of SRTM
model, respectively; a is a vector of known coefficients, heights, a relative comparison can be performed rigor-
x is a vector of unknown parameters, ɛ is random noise ously as well. In this sense the discrepancies between
term at ground control point. With equation (4) one two height data D Hij , which are derived from the model
assumes that the parametric model (aT x) absorbs inher- and DGPS are simply computed for a certain group of
ently all systematic errors in the model dataset. baselines. These differences are traditionally represented

28 Survey Review 2018 VOL 50 NO 358

 Üstün et al. Height biases of SRTM DEM related to EGM96

in the relative form in parts per million (ppm): which translates into 30 and 90 m spatial resolution at
- - the Equator are distributed in 14 400 tiles by 1 × 1 degree
- H SRTM − H SRTM − (H DGPS − H DGPS ) -
- i
D Hij = -
j i j -
(7) (1 arc-second data product is freely presented to public
-
- Sij - users for the USA only). DEM data used in this study
are stored in hgt format (binary) by 1° × 1° tiles at 3×3
where Sij is the length of the geodesic between ith and jth arc-second resolution. Zero for water surfaces (i.e. sea
points on the ellipsoid surface (unit: km), H DGPS and ocean) and −32 768 (minimum integer value of 2-
describes DGPS measurements. Then the mean absolute byte short signed variable) for voids are assigned in the
deviation (MAD) at specific length is the relative evalu- data files. The height values quantised to 1 m are in the
ation of SRTM DEM. format of IEEE byte order. The detailed documentation
and technical specification of SRTM DEM can be
found at the website of the DEM (see SRTM, 2014).
A case study in Turkey The absolute and relative performance requirements on
In this section the input data treated in the investigations vertical height errors are 16 and 10 m, respectively, at the
over the study area is discussed and furthermore their glo- 90% confidence level. Using the ground-truth data col-
bal/regional accuracies are outlined. lected on a continental basis, 90% of absolute and relative
errors have been observed below 9 m. According to Rodri-
guez et al. (2005), the long-wavelength errors reaching up
Study area to 6 m are included within the given estimates.
Our study area which is approximately 780 000 km2
including inner waters covers the peninsula of Turkey
and extends towards the Eastern part of Turkey. It is geo- GPS-levelling and DGPS track data
graphically delimited by 36° to 42° northern latitudes and A total of 47 precise levelling benchmarks which were
by 26° to 45° eastern longitudes. The topography starting occupied with GPS observations were obtained from a
from the coastal plains reaches its peak at Agri (ancient project (cf. Ustun and Demirel, 2006). The ground-truth
Ararat) mountain with 5137 m. Along the mountains control data are located in areas of different topography
that go across the country from the east to west, there along a 900 km profile from the southern part of the
are over 60 peaks with elevations of 3000 m or more. country to the north (Fig. 3). It offers the best way to
The whole area has an average height about 1000 m. determine absolute performance of SRTM DEM that
The region was selected for this research because it is has two different versions based upon EGM96 and
one of the quite complicated areas over the globe from EGM08 geoid models. The levelling heights, which are
the view of the rugged topography as a link of Alpin- in 0–1600 m, referred to the mean sea level with respect
Himalayan collision zone. The geographical boundaries to the tide gauge benchmark in Antalya. On the other
and the topographical view of the study area are shown hand, both GPS and geometric levelling measurements
in Fig. 3. were conducted under the first-order standards of geode-
tic surveys in Turkey. Therefore, the error uncertainties at
the GPS-levelling stations in terms of height component
Data description are expected not to exceed 6 cm from the starting to the
SRTM DEM end points of the levelling line.
A global DEM which is spanning from 56° S to 60° N is The kinematic DGPS tracks, the blue lines in Fig. 3,
publicly available on the NASA servers (SRTM, 2014). were collected in different directions to validate SRTM
The data files of 1 and 3 arc-second the resolution of DEM accuracies and to check the effect of datum

3 Topography of the study area and geographical distribution of ground control benchmarks

Survey Review 2018 VOL 50 NO 358 29

Üstün et al. Height biases of SRTM DEM related to EGM96

4 Contour map of geoid height discrepancies (EGM96 minus EGM08) that mean height biases of SRTM DEM related to EGM96
over Turkey

conversion from EGM96 to EGM08. The total length complete to degree and order 2159 and contains
DGPS tracks including 120 000 ground control points is additional coefficients extending to degree 2190 and
about 5900 km. Two DGPS data sets were gathered in order 2159, which means 5 × 5 arc-minute (≈ 9 × 9 km
the time span of 2006–2008 (Bildirici et al., 2010) and at the Equator) resolution. Ustun and Abbak (2010) con-
2015 for the tracks in the northern territory of Turkey. cluded that EGM08 is the best combined EGM in the ter-
The DGPS heights refer to the mean sea level represented ritory of Turkey, by means of GPS-levelling data and
by EGM96 geoid model. The position info in terms of spectral tools (e.g. the degree variance, error degree var-
latitude, longitude and altitude along a track is parsed iance etc.). Figure 4 depicts the variation of geoid heights
from the NMEA message. differences between EGM96 and EGM08 over Turkey.
The locations where absolute geoid differences are larger
EGM96 versus EGM08 than 1 m could be identified as areas of which EGM96
EGM96 is the result of a collaboration between NIMA, shows weak performance, or in other words an improve-
the NASA Goddard Space Flight Center and Ohio ment of local gravity knowledge in EGM08 (2014).
State University. The joint project obtained various sur-
face gravity surveys from many different regions of the Numerical investigations
globe including data released from the NIMA archives.
In this section we treated, at first, direct evaluation of the
These collection efforts have improved the data holdings
EGMs considering their use as vertical datum of SRTM
over many of the world’s land areas. As a result of all
DEM data set. With this approach, the effect of the
efforts, EGM96 was produced complete to degree and
datum conversion on the SRTM heights can be estimated
order 360, which is equal to 30×30 arc-minute resolution
quantitatively tile-by-tile. In the second step, the DGPS
(Lemoine et al., 1998).
tracks will be utilised for both detecting SRTM accuracy
The availability of accurate and complete gravity
in absolute and relative sense and estimating the magni-
anomaly data throughout the globe is highly important
tude of dN values along different routes. The last step
to build up a successful gravity model. In the time period
identifies to what extent EGM08 affects the accuracy of
of the EGM96 model creation, although a major step in
the SRTM DEM by means of the GPS-levelling data in
collecting and measuring terrestrial gravity data was
the study area.
taken concerning coverage and accuracy, there were
The zero and first degree terms were treated in the same
many areas where gravity anomaly data were sparse,
way for the calculation of absolute geoid heights in both
poor in accuracy or completely non-existent. It seems
EGM96 and EGM08. This enables to remove their effects
that the territory of Turkey had been partly or insuffi-
on geoid height differences, e.g. Fig. 4. When comparing
ciently covered with accurate surface gravity data in the
absolute SRTM heights with GPS-levelling data, the
model development (Ustun and Abbak, 2010). For that
mean difference between national datum (TUD) and
reason, the spatial resolution of EGM96 geoid model in
EGMs is essential. Using nearly 200 control points collo-
the study area is dominated by the long-wavelength signal
cated by GPS-levelling which are evenly distributed over
of the gravity field due to the lack of local gravity support.
Turkey, the mean difference of the TUD from EGM96
Having started the era of satellite gravity mission in the
and EGM08 geoids are 50 and 86 cm, respectively (Tepe-
new millennium and a comprehensive compilation of glo-
koylu and Ustun, 2008; Kilicoglu et al., 2009).
bal 5-minute grid of area-mean free anomalies, the EGM
Development Team of the National Geospatial-Intelli-
gence Agency (NGA) has released the Earth Gravita- Tile-by-tile comparison of EGMs
tional Model 2008 (Pavlis et al., 2012). Compared to The gridded geoid height differences were arranged like
EGM96, EGM08 offers a significant improvement thanks SRTM tiles (1°×1°). The overall median, mean, standard
to the release of new data from more sources. EGM08 is deviation and RMSE statistics of total 114 tiles were

30 Survey Review 2018 VOL 50 NO 358

 Üstün et al. Height biases of SRTM DEM related to EGM96

5 Geographical distribution of median (bottom) and RMSE discrepancies (top) between EGM96 and EGM08 for the SRTM tiles

calculated independently to define variability of model 2009). The sub-metre position uncertainties of the
discrepancies. The median values that were consistent DGPS data enable a conclusive numerical assessment of
with the mean values of 114 tiles are in the range of SRTM data.
− 4.2 and 2.5 m, while the RMSE average of all is above Table 1 gives the statistics of differences between DGPS
1 m (maximum RMSE is 4.1 m). Figure 5 shows the size and corresponding SRTM heights interpolated from hgt
of vertical datum difference from EGM96 to EGM08 (native SRTM format) grid files. With the changing path
for each tile. The maximum separation of − 4.2 m is cap- lengths from 281 to 690 km (total 5866 km), 13 routes
tured at the SRTM tile N40E037 in the northern part of reaching up to 2200 m height represent quite well the
the country. The numerical comparison states that the variability of topography in order to validate SRTM in
absolute correction term of dN shifts the mean position Turkey. The absolute height differences that exceed 35 m
of vertical datum more than 1 m at 33 tiles. The geo- are excluded from the comparison. The threshold of
graphical distribution of median and RMSE statistics 35 m is evaluated as a tolerance by factor 6 of the pub-
show similar pattern throughout Turkey. As mentioned lished accuracy for SRTM heights. The median position
in the previous section, it is believed that the results reflect of SRTM topography along 13 routes is 2 m (mean is
the significant improvement in the gravity field knowledge 2.8) higher than that of DGPS (standard deviation of
with the inclusion of local gravity into the EGM08 model differences is 5.3 m). When considering the frequency dis-
(see Ustun and Abbak, 2010). tribution, the absolute errors for the statistical confidence
level at 90% vary between 6.8 and 13.0 m, while this com-
parison at 118995 points indicates 9.1 m of absolute error.
Assessment of SRTM DEM using DGPS data An investigation regarding long-wavelength error
The DGPS tracks (see Fig. 3) that are known with the geo- implies a term of about 2 m on the differences after
graphic coordinates have been taken into account to gauge using 7-parameter corrector surface described in equation
the accuracy of SRTM DEM data in the study area. Since (4). Though small in quantity, it is believed that a certain
the heights measured by DGPS are based on the same portion of this effect is caused by the reference geoid
horizontal datum of SRTM DEM, an exact comparison model, i.e. EGM96. A relative comparison based on the
can be made at the point locations of tracks with SRTM same sets has been performed for the baseline lengths
heights interpolated by means of grid values. For the from 0.1 to 2 km. The relative error at 1 km that rep-
interpolation, a special software package was developed resents all tracks is about 5.2 m, which is almost same
based on Thin Plate Spline method using heights from with the RMSE derived from the residuals of 7-parameter
nearest 16 points (for further details, see Bildirici et al., comparison (Table 2).

Survey Review 2018 VOL 50 NO 358 31

Üstün et al. Height biases of SRTM DEM related to EGM96

Table 1 SRTM absolute error statistics with respect to the kinematic DGPS data in the various routes

Track Length # Min Max Median Mean StD RMSE 90% AE

(km) of pts (m) (m) (m) (m) (m) (m) (m)

Antalya-Bozüyük 521 10145 −15.1 33.9 3.3 3.9 4.3 5.8 8.4
Ardahan-Erzincan 690 7280 −21.4 34.8 3.5 4.7 5.8 7.5 11.8
Bartın-Kadınhanı 674 16320 −26.3 34.8 1.6 2.8 5.8 6.4 9.4
Beysehir-Ören 526 12326 −17.7 34.7 1.6 2.4 5.2 5.7 8.5
Erdemli-Konya 378 9417 −34.6 28.4 −0.6 0.2 4.9 4.9 6.8
Erzincan-Sivas 228 5432 −13.7 32.8 4.8 6.3 5.2 8.1 13.0
Kocaeli-Samsun 654 12510 −34.6 34.5 0.9 0.8 5.7 5.8 8.3
Konya-Amasra 597 10292 −11.0 33.8 2.0 2.6 5.1 5.7 8.5
Konya-Erdemli 315 7269 −20.9 34.9 1.5 2.5 4.1 4.8 7.4
Konya-Manavgat 267 6295 −24.0 34.3 2.3 3.6 6.2 7.2 11.1
Konya-Yozgat 575 14527 −32.4 34.0 3.2 3.4 4.3 5.4 8.1
Manavgat-Konya 218 4928 −20.5 31.6 2.2 2.8 5.9 6.6 10.2
Trabzon-Ardahan 223 2254 −15.1 33.5 4.7 4.8 6.0 7.7 11.7
Total 5866 118995 −34.6 34.9 2.2 2.8 5.4 6.1 9.1

Table 2 Statistics of SRTM residuals after 7-parameter which is significant at the interval from − 3.9 to 2.2 m.
corrector surface for absolute differences and Under this scenario the numerical results demonstrate
relative comparison for the baselines whose that a datum conversion can improve the SRTM heights
lengths are in the interval of 0.1–2 km (unit in m)
up to 2.6 and 1.2 m in terms of absolute errors and stan-
MAD @ dard deviation, respectively. The gain is equivalent to 25–
Track Min Max Mean RMSE 1km 30% of the error statistics presented in the previous table
for SRTM heights. As an example, the dN variation
Antalya-Bozüyük −20.0 29.7 0.0 4.0 5.5 throughout the profile of Kocaeli-Samsun fluctuates at
Ardahan- −26.2 30.6 −0.0 5.6 5.8
Erzincan
the same levels with the SRTM errors on some parts of
Bartın-Kadınhanı −33.2 30.6 −0.0 5.2 5.2 the plot where the topography is relatively flat (see Fig. 7).
Beysehir-Ören −24.4 33.4 −0.0 4.4 4.7
Erdemli-Konya −35.2 28.5 0.0 4.7 4.2 Datum transformation of SRTM DEM and
Erzincan-Sivas −23.3 23.6 0.4 4.2 4.7
Kocaeli-Samsun −34.1 32.0 −0.0 5.4 6.9
measuring its effect
Konya-Amasra −18.9 27.7 0.0 4.3 5.2 For the transformation of the vertical datum, the correc-
Konya-Erdemli −25.5 30.0 0.0 3.6 4.1 tion term dN is added to the original height of SRTM
Konya-Manavgat −29.3 30.8 0.5 5.9 6.2 DEM by means of equation (3). For this process, dN
Konya-Yozgat −40.0 26.3 −0.0 3.4 3.0 values having 5-arc minute grid resolution were used to
Manavgat-Konya −21.3 29.3 0.0 5.1 6.4
interpolate correction term at computation point by
Trabzon- −20.8 29.3 0.5 5.7 6.5
Ardahan bilinear interpolation. Both SRTM heights (already
Total −37.0 32.0 −0.0 5.2 5.2 exist HEGM96 and newly obtained HEGM08 ) were com-
pared with GPS-levelling data in the study area. Table 4
reveals the effect of correction terms on their SRTM
heights for 47 levelling benchmarks. The RMSE of dN
At this point, to evaluate the necessity for datum estimated from geoid differences along the levelling line
change in SRTM heights, the situation of geoid differ- of approximately 900 km length is 1.5 m. If one disregards
ences can be examined. As seen in Table 3 and Fig. 6, the SRTM height errors, this value might be considered
the absolute differences of EGM96 minus EGM08 the predictable effect of a datum change of SRTM heights
geoid heights based on the DGPS tracks take place, on the line. From the RMSE decline, 3.8–3.0 m, it can be

Table 3 Statistics of d N geoid height differences (height biases in SRTM) along the kinematic DGPS data in the various routes
(unit in m)

Track # Min Max Median Mean StD RMSE 90% AE

Antalya-Bozüyük 10145 −2.7 1.9 0.1 −0.3 1.5 1.6 2.4

Ardahan-Erzincan 7280 −3.8 0.7 −0.4 −0.8 1.2 1.4 2.9
Bartın-Kadınhanı 16320 −2.4 1.2 −0.1 −0.3 1.0 1.1 1.9
Beysehir-Ören 12326 −3.1 0.8 −0.2 −0.7 1.3 1.4 2.8
Erdemli-Konya 9417 −0.3 2.2 0.8 0.8 0.8 1.1 1.8
Erzincan-Sivas 5432 −3.8 0.3 −3.4 −2.9 1.0 3.1 3.6
Kocaeli-Samsun 12510 −3.9 0.3 −0.5 −1.2 1.3 1.8 3.0
Konya-Amasra 10292 −3.1 0.9 −0.2 −0.5 1.1 1.2 2.4
Konya-Erdemli 7269 −1.4 1.2 −0.4 −0.2 0.6 0.7 1.1
Konya-Manavgat 6295 −1.2 0.2 −0.5 −0.4 0.3 0.5 0.8
Konya-Yozgat 14527 −1.2 0.4 −0.3 −0.3 0.4 0.5 0.8
Manavgat-Konya 4928 −1.0 −0.1 −0.5 −0.5 0.2 0.5 0.7
Trabzon-Ardahan 2254 −3.0 −0.8 −1.1 −1.3 0.5 1.4 2.0
Total 118995 −3.9 2.2 −0.3 −0.6 1.2 1.4 2.6

32 Survey Review 2018 VOL 50 NO 358

 Üstün et al. Height biases of SRTM DEM related to EGM96

6 Distribution of the height differences. The left panel designates the height errors of SRTM heights with respect to the kin-
ematic DGPS data in all tracks. The right panel describes the corresponding distribution of NEGM96 minus NEGM08 differences
under the same scenario

7 Variations of SRTM height errors and geoid height difference as a function of distance from Kocaeli to Samsun

Table 4 Validation statistics of two versions of SRTM heights at 47 GPS-levelling benchmarks (unit in m)

Differences Min Max Med Mean StD RMSE

NEGM96 − NEGM08 −3.9 0.6 −0.5 −0.9 1.2 1.5

HEGM96 − Hlev −2.2 8.2 2.8 3.1 2.2 3.8
HEGM08 − Hlev −2.5 6.1 2.2 2.3 1.9 3.0

understood that the datum conversion on the data has Conclusions and remarks
improved the heights by % 2.3 m of RMSE according to
the well-known error propagation rule. The graphical The EGM96 geoid provides the vertical datum for the
view of the improvement can be seen in Fig. 8 displaying global interferometric DEM produced from SAR data
that HEGM08 along the profile has a closer position to the collected by the aid of SRTM. In the published global
levelling heights rather than HEGM96 . As we expected assessment report of SRTM DEM products, it is stated
from the previous comparisons, i.e. Figs. 4 and 5, the sep- that the vertical accuracy meets and exceeds the perform-
aration between the geoid models has shown its effect on ance requirements of the mission (16 m for 90% confi-
the SRTM heights towards the northern Turkey. The mag- dence level) by a factor of nearly two in comparison to
nitude of correction dN term goes up to − 4 m towards the ground-truth data like kinematic GPS trajectories on
end. road networks.

Survey Review 2018 VOL 50 NO 358 33

Üstün et al. Height biases of SRTM DEM related to EGM96

8 Differences of HEGM96 and HEGM08 over 47 benchmarks of the first-order levelling line

Although the error sources presented for the SRTM Considering the statistics, the distribution of control
mission are described very well in the literature, SRTM data is very important in determining the real impact of
height biases due to EGM96 geoid height errors have EGM08 on the SRTM DEM. Although EGM08 is
not been presented before. In this study, at first EGM96 quite better than EGM96, the accuracy improvement of
is checked against to EGM08 to be able to see their differ- SRTM DEM may remain at the level of a few centimetres
ences and possible impacts on SRTM heights in global due to inherent SRTM errors. Therefore, a long traverse of
scale and a certain geographic area. A global perspective GPS-levelling heights that runs from the Mediterranean
on this question implies regional separations exceeding Sea to the Black Sea in north-eastern direction has been
the reported SRTM performances. Accordingly, the terri- used for the comparison. Because the distances between
tory of Turkey has large differences in the range of − 6 to adjacent benchmarks are about 20 km, only absolute
4 m. The overall median and mean statistics of the SRTM height errors were studied at 47 ground-truth control
tiles of 1° × 1° calculated from geoid differences show sub- points. The datum conversion from EGM96 to EGM08
stantial discrepancies that come from the lack of local by means of additive correction terms along the levelling
gravity support in EGM96 model development. The line has reduced the RMSE error from 3.8 to 3.0 m. Most
total surface area of the absolute differences more than of this improvement has been observed on the northern
1 m is about 30% of the Turkish territory. These values portion of the levelling line within the last 300 km of the
should be considered as correction terms which mean a total 900 km length. In this region, the absolute height
significant bias in the original DEM files. In order to biases related to EGM96 rise up to 4 m along the levelling
measure the amount of height biases within the errors line. 7-parameter model analysis reveals the long-wave-
of SRTM heights, a ground-truth control data in 13 tracks length errors above 1 m.
were collected by DGPS techniques in Turkey. Nearly Consequently, from all investigations in this study it is
120 000 points imply an absolute height error of 9.1 m understood that EGM96 can distort the accuracy of
in SRTM data, whereas the absolute geoid height differ- SRTM DEM up to several metres, locally where terres-
ences on the same trajectories of a total of 5866 km trial gravity data are missing or its quality is insufficient.
expose a separation of 2.6 m with 90% confidence level. This deterioration can be repaired by better EGM like
If alternative routes passing through the area that displays EGM08.
large disagreements in the eastern Black Sea region are
chosen, it is obvious that the latter statistics increase.
Beside the absolute SRTM errors, the DGPS tracks
allow us to see the associated long wavelength and relative
Acknowledgements
errors which are equivalent to 1.5 and 5.2 m, respectively. NASA and NGA are cordially acknowledged for provid-
Since both SRTM and DGPS heights use the same geoid ing EGM96 and SRTM DEM, EGM08, respectively. The
model, namely EGM96, there is no geoid effect on the authors gratefully thank the support of The Scientific and
validation estimates related with SRTM heights. Technological Research Council of Turkey under grant

34 Survey Review 2018 VOL 50 NO 358

 Üstün et al. Height biases of SRTM DEM related to EGM96

number 106Y130 in gathering the DGPS data. Dr Orhan Kiamehr R. and Sjöberg L.E., 2005. Effect of the SRTM DEM on the
Kurt helped with technical study in collecting some part determination of high-resolution geoid model: a case study in
Iran. Journal of Geodesy, 79, 540–551.
of data. The authors appreciate the constructive remarks Kilicoglu A., et al., 2009. Evaluation of the earth gravitational model
suggested by two anonymous reviewers in the first version 2008 in Turkey. Newton’s Bulletin, 1 (4), 164–171.
of this manuscript. Kotsakis C. and Sideris M.G., 1999. On the adjustment of combined GPS/
levelling/geoid networks. Journal of Geodesy, 73 (8), 412–421.
Lemoine F.G., et al., 1998. Development of the Joint NASA, GSFC
Disclosure statement and NIMA Geopotential Model EGM96. Technical report,
NASA Goddard Space Flight Center, Greenbelt, Maryland,
No potential conflict of interest was reported by the 20771, USA.
authors. Merry C.L., 2003. DEM-induced errors in developing a quasi-geoid
model for Africa. Journal of Geodesy, 77, 537–542.
Mukherjee S., et al., 2013. Evaluation of vertical accuracy of open
References source Digital Elevation Model (DEM). International Journal of
Applied Earth Observation and Geoinformation, 21 (2013), 205–
Abbak R.A., 2014. Effect of ASTER DEM on the prediction of mean 217.
gravity anomalies: a case study over the Auvergne test region. Pavlis N.K., et al., 2012. The development and evaluation of the Earth
Acta Geodaetica et Geophysica, 49 (4), 491–502. Gravitational Model 2008 (EGM2008). Journal of Geophysical
Banerjee P., et al., 1999. Geoid undulation modelling and interpretation Research: Solid Earth, 117 (B4), 1–38.
at Ladak, NW Himalaya using GPS and levelling data. Journal of Rabus B., et al., 2003. The shuttle radar topography mission – a new class
Geodesy, 73 (2), 79–86. of digital elevation models acquired by spaceborne radar. ISPRS
Berthier E., et al., 2006. Biases of srtm in high-mountain areas: impli- Journal of Photogrammetry & Remote Sensing, 57, 241–262.
cations for the monitoring of glacier volume changes. Geophysical Rexer M. and Hirt C., 2014. Comparison of free high resolution digital
Research Letters, 33 (8). L08502. elevation data sets (ASTER GDEM2, SRTM v2.1/v4.1) and vali-
Bildirici I.O., et al., 2010. Compilation of digital elevation model for dation against accurate heights from the Australian National
Turkey in 3-arc-second resolution by using SRTM data supported Gravity Database. Australian Journal of Earth Sciences, 61 (2),
with local elevation data. In: G. Gartner and F. Ortag, eds. 213–226.
Cartography in Central and Eastern Europe, Lecture notes in geoin- Rodriguez E., et al., 2005. An assessment of the srtm topographic pro-
formation and cartography, ICA symposium, Vienna, 16 –17 ducts. Technical Report JPL D-31639, Jet Propulsion Laboratory,
February 2009 . Springer, 63–76. Pasadena, California.
Bildirici O.I., et al., 2009. Assessment of shuttle radar topography mis- SRTM, 2014. Shuttle Radar Topography Mission, Available from:
sion elevation data based on topographic maps in Turkey. CaGIS, <http//www2.jpl.nasa.gov/srtm> [Accessed November 2014].
36 (1), 95–104. Tepekoylu S. and Ustun A., 2008. Evaluation of global geopotential models
EGM08, 2014. Earth Gravitational Model 2008 (EGM2008). Available by GPS-leveling data in Turkey. Harita Dergisi, 139 (1), 49–65.
from: <http://earth-info.nga.mil/GandG/wgs84/gravitymod/egm2008> Tocho C., Font G. and Sideris M.G., 2008. A new high-precision gravi-
[Accessed November 2014]. metric geoid model for Argentina. In: P. Tregoning and C. Rizos,
Farr T.G., et al., 2007. The shuttle radar topography mission. Reviews of eds. Monitoring and understanding a dynamic planet with geodetic
Geophysics, 45 (RG2004), 1–33. doi:10.1029/2005RG000183. and oceanographic tools. Cairns, Australia: Springer, 416–423.
Hirt C., Filmer M.S. and Featherstone W.E., 2010. Comparison and vali- Ustun A. and Abbak R.A., 2010. On global and regional spectral evalu-
dation of the recent freely available ASTER-GDEM ver1, SRTM ation of global geopotential models. Journal of Geophysics and
ver4.1 and GEODATA DEM-9S ver3 digital elevation models over Engineering, 7 (4), 369–379.
Australia. Australian Journal of Earth Sciences, 57 (3), 337–347. Ustun A. and Demirel H., 2006. Long-range geoid testing by GPS-
ICGEM, 2015. Evaluation of the models. International Center of Global leveling data in Turkey. Journal of Surveying Engineering, 132 (1),
Earth Models, Available from: http://icgem.gfz-potsdam.de/ 15–23.
ICGEM/evaluation/evaluation.html. Zielinski R. and Chmiel J., 2007. Vertical accuracy assessment of srtm c-
Jing C., et al., 2014. Comparison and validation of SRTM and ASTER band dem data for different terrain characteristics. In: Z. Bochenek,
GDEM for a subtropical landscape in Southeastern China. ed. New developments and challenges in remote sensing. Rotterdam:
International Journal of Digital Earth, 7 (12), 540–551. Millpress, 685–693.

Survey Review 2018 VOL 50 NO 358 35