data has been used as a substitute and used along with simple
filter processes or even no conversion (Sesnie
et al.
, 2008; Barr
and Clark, 2009; Li
et al.
, 2013b). This could certainly bring
adverse effects into the research results. Therefore, to estab-
lish the magnitude of these effects, it is important to assess
the quality of
SRTM
data in relation to the land surface. Such a
study is also a precondition for any attempt to strip the height
of land -cover from
SRTM
to reach a
DEM
.
Elevation errors in
SRTM
data are usually studied by
employing high-accuracy elevation datasets to represent the
actual elevation and comparing them with
SRTM
data. The
high-accuracy elevation datasets primarily include two types:
one is elevation data obtained from ground measurements
using a global positioning system (
GPS
) receiver and the other
is a digital elevation model (
DEM
) that is known to be more
accurate than
SRTM
data. Ground-measured
GPS
elevation data
are of high accuracy; the error can be controlled to within 0.5
m (Rodríguez
et al.
, 2006) and the
GPS
value derived error can
be regarded as an absolute error (Gesch
et al.
, 2012). However,
a
DEM
data with high accuracy that provides continuous and
complete coverage over a large area is often more suitable as
reference data. Accurate
DEM
data have been used here for
studying the associations of
SRTM
elevation error and land-
scape features, because it overcomes some defects of the
GPS
method in diverse landscapes. For example, rugged mountain-
ous terrain and dense forests are unreachable for
GPS
devices
which results in a less than comprehensive evaluation of error
(Gorokhovich and Voustianiouk, 2006). Usually,
DEM
datasets
provide the most suitable data for a country or region. Ex-
amples of previous application include the National Elevation
Dataset (
NED
) with resolution of 1-second used by Shortridge
and Messina (2011) and the high-accuracy
DEM
data with reso-
lution of 3-seconds generated from 1:250 000 scale topograph-
ic maps used by Miliaresis and Paraschou (2005). Because the
error obtained is not the actual bias of
SRTM
data to the bare
land surface, it may be more appropriate to regard the error
derived from the reference
DEM
as a relative error, however,
the reference
Hc-DEM
(see the Data Section) used in this study
is believed to be precise enough to closely characterize the
SRTM
error (
NASMG
, 2008). Over larger areas (e.g., continen-
tal scale), a method typically used to assess the accuracy of
SRTM
data is sample survey. For example, Guth (2006) used
hundreds of thousands of regular-grid sample points to study
elevation errors in
SRTM
data over the United States.
There are many studies that have identified land-cover is-
sues with
SRTM
data (Shortridge, 2006; Castel and Oettli, 2008;
Miliaresis, 2008; LaLonde
et al.
, 2010), but most of them take
a local area as a case study. Current systematic studies of
SRTM
data quality over large areas have primarily focused on the
Amazon basin, Australia, and the United States, and the
SRTM
data error exhibits different characteristics due to variations in
the topography and land-cover in these areas. The Amazon Ba-
sin is primarily covered by tropical rainforests, so the
SRTM
data
quality is largely affected by vegetation effects. In the regions
of umbrageous forest
SRTM
shows a vertical offset of around 30
meters (Blitzkow
et al.
, 2007). Australia is dominated by desert
and has generally low topographic relief, so the
SRTM
data
clearly exhibit stripes in low relief areas with a wavelength of
about 800 m and amplitude of about 0.2 to 4 m (Gallant and
Read, 2009). The United States consists of mountains in the
west and east with plains in the center with relatively large
variations of elevation error in the east and west and small error
variations at the center, with the mean error being just over 2
m (Shortridge and Messina, 2011). Compared with these areas,
the topography and land covers of China are more complex.
Among the different regions there are many diverse landforms
and significant variations in topographic relief, hydrology, and
vegetation types as well as other land coverages. In addition,
glaciers, deserts, and wetlands are also widely distributed (Li
et
al.
, 2013a). Therefore, the distribution of
SRTM
data errors is rel-
atively complex, and has been briefly explored in the prelimi-
nary study of the quality of global
SRTM
data by Rodríguez
et al.
(2006 ). However, although
SRTM
data have been widely used in
studies of China (Yang
et al.
, 2012; Pieczonka
et al.
, 2013; Sun
et al.
, 2013; Neckel
et al.
, 2014), a detailed evaluation of
SRTM
data across the whole of China is still lacking.
The objective of this study is to present the diversities of
effects associated with different land surface features in China
on the quality of
SRTM
data and to provide information about
the applicability and background of uncertainty in the
SRTM
3-second datasets to its users. In this way, the reliability of their
research results can be improved. Because the topography and
land-cover types in China are complex, an elevation error in
one position may be subjected to the effect of a combination of
different factors, and the dominant factor that effects elevation
error may vary across different areas. For
SRTM
data in China,
this study initially endeavored to employ the whole sample
method (all sample points without classification are included
in the analysis of the associations between
SRTM
error and
factors). This has been used in prior estimations of
SRTM
error
(Gorokhovich and Voustianiouk, 2006; Shortridge and Messina,
2011). However, its use did not lead to suitable results that were
consistent with the previous findings. These findings included
SRTM
elevation error increasing with increasing slope and vege-
tation coverage, and the errors in vegetation-covered areas being
usually positive. As a consequence, the single factor method (in
which sub-samples are separated conditionally to ensure they
are dominated by single factor) was adopted for this study to
explore the characteristics of
SRTM
error in complex landscapes
and eliminate interference between the various factors.
Data and Methods
Data
The
SRTM
data used in this study was the
CGIAR
Version 4
dataset obtained from the
CGIAR-CSI
website (
http://
SRTM
.csi.
CGIAR
.org/
) (Jarvis
et al.
, 2008). The data covers the landscapes
of the whole of China and uses the
WGS
84 horizontal datum
and the
EGM96
vertical datum and the resolution is 3-seconds.
The reference
DEM
data was the hydrologically correct
DEM
(
Hc-DEM
) data generated from 1:50 000 topographic maps from
the National Geomatics Center of China (
NASMG
, 2002). The
resolution of the
DEM
data was 25 m, and used the Gauss Kru-
ger Projection in 6° wide zones based on the Xian80 Geographi-
cal Coordinate System and the 1985 National Height Datum.
The reference data was generated from the map sheets using
ANUDEM
software (Hutchinson, 2004) in which the Hutchinson
algorithm (Hutchinson, 1989) is an internationally popular
DEM
interpolation algorithm that can specifically incorporate the hy-
drological correctness of the product. Of the elevation datasets
that cover the whole of China, the 1:50 000 topographic map
has been established as the most accurate (
NASMG
, 2008) with
vertical errors controlled within 3 m for the flat areas, 5 m for
hills, 8 m for mountain areas, and 14 m for the steep mountain
areas. The
ANUDEM
-generated
DEM
data includes topographic
feature lines (terrain shape, streamlines, and ridge lines)
(Clarke and Burnett, 2003) and derived topographic attributes
(slope and aspect, etc.) having high accuracy (Yang
et al.
,
2007).
To investigate the relationships between the
SRTM
eleva-
tion error, the topographic attributes and the land-cover types,
other ancillary datasets were also introduced into this study.
The topographic attributes include two parameters of slope
and aspect that were derived from both the
Hc-DEM
and
SRTM
data for a comparative analysis in the consideration of the
SRTM
data quality assessment and the internal relationship
136
February 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
