Methods
The bias between the
EGM
96 Height Datum used by
SRTM
data
and the 1985 National Height Datum used by the
Hc-DEM
has
a mean value of 0.357 m and increases gradually from east to
west and south to north (Guo
et al.
, 2004).This bias was first
corrected in this study. The
SRTM
data error includes both
positioning error and the elevation error. Van Niel
et al.
(2008)
demonstrated that the positioning error has a large impact on
the assessment of the differences between
SRTM
data and other
DEM
data. Some studies have quantified the mis-registration be-
tween
SRTM
data and high-resolution
DEM
s (Hofton
et al.
, 2006;
Rodríguez
et al.
, 2006; Smith and Sandwell, 2003) and others
have assessed the impact that the mis-registration has on appli-
cations (Dai and Khorram, 1998; Townshend
et al.
, 1992; Ver-
byla and Boles, 2000; Wang and Ellis, 2005). Eliminating the
positioning error is a prerequisite for exploring the characteris-
tics of elevation error on
SRTM
data applications. It is sufficient
that the relative positioning precision is guaranteed in most
applications of the
SRTM
data with other datasets. In order to
avoid change of
SRTM
elevation values due to the resampling
process during the spatial registration, basing the alignment
on
SRTM
data, the spatial registration was conducted on
Hc-DEM
data and had two steps: preliminary registration and accurate
registration. As the preprocessing step, the original
Hc-DEM
was
resampled to 1-second resolution before the conduct of spatial
registration. The method for preliminary registration was
described by Van Niel
et al.
(2008). To improve the local preci-
sion of this step, the
SRTM
data and resampled
Hc-DEM
data
were segmented into 1.5° × 1° tiles first. Accurate registration
was then conducted to reduce any local positioning errors that
occur during preliminary registration, and the method used in
this step was described by Shortridge and Messina (2011). The
SRTM
elevation error is defined as the bias between
Hc-DEM
and
SRTM
elevation values for the same position so the registration
accuracy is critical. The
SRTM
data had a mean error of −0.81
m with a standard deviation of 27.36 after preliminary registra-
tion and a mean error of −0.35 m with a standard deviation of
14.96 m after accurate registration. Thus, the elevation error
was significantly reduced by the accurate registration leading
to more reliable estimates of the elevation error.
The topographic attributes of slope and aspect used in this
study were derived from the
SRTM
data and the registered
Hc-
DEM
, which were both based on the
WGS84
horizontal datum
and a geographic coordinate system with the unit of cell size
being degrees. The conventional algorithms for slope and as-
pect require that the unit of elevation should be in accordance
with cell size (usually meters).The algorithm was refined so as
to estimate slope and aspect directly in the geographic coordi-
nate system. Zhou and Liu (2004) have evaluated the errors of
conventional grid based algorithms of slope and aspect, among
which the method of third-order finite difference weighted by
reciprocal of squared distance was evaluated as accurate and
so was adopted in this study. Compared with algorithms in
projected coordinate systems, the difference in geographic co-
ordinate systems is that the unit of cell size is degrees, which
need to be converted to the units of length. Supposing that the
Earth is a perfect sphere (the difference of equatorial radius
and polar radius can be ignored in the calculation of slope and
aspect in this study), the length of cell size is a constant in
north-south direction (Y
d
) but varies with latitude in the east-
west direction (X
d
), which can be expressed as:
X
d
= Y
d
cos
α
(1)
where
α
represents the latitude of the center cell in the unit of
radians. The radius of the earth was set to 6,371,393 m in this
study, thus Y
d
is 92.662,437 m. The slope and aspect values
in geographic coordinate system can be derived by combining
the parameters of Y
d
, the latitude value of each cell in
SRTM
data, and Equation 1 with the algorithms of slope and aspect.
Although the
Hc-DEM
dataset is the best elevation dataset
for China, some areas, primarily in the Taklimakan Desert and
southeastern Tibet, are still problematic. In both datasets, the el-
evations of the inland waters were obtained using a special data
processing algorithm and were not based on the radar signals.
Therefore, the water bodies in both datasets and the problematic
areas in the
Hc-DEM
datasets were excluded from the analysis of
the
SRTM
errors. One thing to note here is that there exist over-
laps between water bodies and wetlands which spread over the
middle and lower reaches of the Yangtze River and the Qinghai-
Tibet Plateau, so the remaining wetlands involved in this study
are mainly concentrated in northwest and northeast China.
The sample survey method was adopted in this study.
To ensure the number of sample points meets the statistical
requirements at the continental level, and for the different
zoning and classification systems, as well as to ensure that
the data redundancy is as low as possible, a total number
of 518,587 independent sample points were obtained. The
samples are spread across the whole of China and distributed
on a regular grid. The interval value of the sample points
was set by reference to the study of
SRTM
error in the US
by Shortridge and Messina (2011), in which the value used
was 0.055833°. Considering the more complex landscape in
China, the interval value was reduced to 0.0425° (51-pixel
distance, or about 4.7 km in the north-south direction and 2.8
to 4.5 km in the east-west direction) in this study. The varia-
tion pattern of the
SRTM
elevation error and its associations
with the influencing factors were then explored using basic
statistical analysis of sub-samples, which, conversely, had
been separated according to the attributes of various factors.
The sub-samples based on slope and aspect were separated
using the
SRTM
data and considering the internal relationship
between the elevation error and the
SRTM
slope and aspect.
The design specification of the
SRTM
mission is for an
absolute elevation error of less than 16 m for 90 percent of the
entire region (Rabus
et al.
, 2003), and is one of the most criti-
cal indicators describing the quality of
SRTM
data. In previous
local evaluations of
SRTM
data quality, most regions have met
the specification (Rabus
et al.
, 2003; Van Zyl, 2001), although
in global evaluations, some regions have had absolute errors
exceeding 16 m. Therefore, the original 90 percent error
specification for
SRTM
is used for this study as well.
Results, Analysis, and Discussion
Basic Topographic Attributes
Within China, the distribution of the
SRTM
sample points with
elevations greater than 0 is shown in Figure 3a. The mean el-
evation of the entire sample was 1,787.56 m, and the number
of sample points tended to decrease with increasing elevation.
Specifically, the proportion of sample points in lower altitude
areas was the largest. Peak percentages of points occurred
around 0 to 50 m, 1000 to 1050 m, and 4900 to 4950 m, and
the minimum percentages of points fell between 700 to 750 m
and 2600 to 2650 m, above 6,000 m, the percentage of sample
points tended to be 0. The same pattern was also present in
the
Hc-DEM
elevation distribution curves.
The three very large terrain zones separated by the eleva-
tion values of two minimum percentages of points are generally
similar to the widely accepted three main terrain terraces in the
geomorphology of China (Zhao
et al.
, 1995) (Figure 4). The low-
elevation sample points are primarily distributed in the eastern
part of China, the medium-elevation sample points distributed
in the middle part of China and northwest China, and the high-
elevation sample points distributed in the Qinghai-Tibet Plateau.
The proportion of points with an
SRTM
elevation less than 0
m was 0.11 percent, the majority of which were located in the
Turpan Basin and scattered in the southeastern coastal areas.
138
February 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
