a
U
C
€?
e
as
>st
ral
VS,
en
on
ng
rea
ea.
9 0.
ith
ion
om
me
the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
3. IMAGE MATCHING AND REFERENCE
The different digital elevation models based on space images
have been matched with the Hannover program DPCOR using a
least squares matching in the image space. This program is
independent upon the image geometry — it is not using the
epipolar information and requires some seed points in the
matched area. As seed points the control points can be used, so
it is not an additional effort. The ground points have to be
calculated by programs using the correct geometry of the image
orientation. The y-parallaxes of the intersection are indicating
the quality of the matching in the level of the visible surface.
Depending upon the image geometry, the image orientations
have been made with the Hannover programs BLUH,
CORIKON and BLASPO with standard deviations in X and Y
between 0.8 and 0.9 pixels. The standard deviation for Z (SZ) is
depending upon the height to base relation, but can be
expressed also independent upon this with the root mean square
accuracy of the x-parallax (Spx) respecting the relation: SZ —
h/b * Spx where Spx has the dimension of pixel size on the
ground. The orientation accuracy in Z was in the range of 0.4
pixels < Spx < 1.3 pixels.
The achieved DSMs have been analysed with the Hannover
program DEMANAL separating the area by a classification
layer into forest and open areas. A more detailed separation was
not necessary. The dependency of the accuracy upon the terrain
inclination has been computed in DEMANAL beside a more
detailed analysis.
The reference DEM is based on the topographic map 1 : 25 000
in the Gauss-Krueger coordinate system (transverse Mercator),
using the Hayford ellipsoid, while the GPS-ground positions
like also the SRTM-data are available in UTM with the WGS84
ellipsoid in the ITRF reference frame. The mathematical
transformation is not a problem, but the datum of the Turkish
national coordinate system in the area of Zonguldak was not
known, leading to shifts in X and Y. By this reason the shifts of
the height models have been determined by least squares
adjustment with the Hannover program DEMSHIFT. This may
also solve some orientation problems for example of the
SRTM-data.
4. SRTM-DEMs
With Interferometric Synthetic Aperture Radar (InSAR) by the
Shuttle Radar Topographic Mission (SRTM) in February 2000
the main part of the earth land area between the latitude of 56°
south and 60.25° north has been covered by DSMs. The Space
Shuttle carried the US C-band and the German/Italian X-band
system. By the scan-SAR mode the C-band has had a swath
width of 225km while the X-band was limited to a swath width
of only 45km. Caused by this, the X-band system has not
covered the whole area while the C-band has covered 94.6% of
the land mass twice and approximate 50% three times. This
multiple coverage has improved the accuracy of the C-band
DSMs which are by theory not so accurate like the X-band-
DSMs based on the shorter wavelength. The used X-band with
the wavelength of 3cm cannot penetrate the vegetation, but also
the used C-band with a wavelength of 5.6cm shows nearly the
top of the vegetation. The published C-band elevation models
are unedited and do include artefacts and gaps in steep areas.
The C-band SAR has an incidence angle (nadir angle at the
ground) between 31? and 61? while the X-band is limited to 50?
to 54°. Corresponding to this the radar layover where the
637
returned signal cannot be separated depending upon the location
is in the range of a terrain slope across the view direction of the
same value.
In the area of Zonguldak, Turkey the X-band DSM and also the
C-band DSM from the German aerospace organisation DLR
have been analysed in relation to a DEM generated from the
topographic map 1 : 25 000. The program DEMANAL is able
to separate the analysis depending upon a classification layer.
In this case only the forest and not forest areas (open areas)
have been separated. The Zonguldak area is mountainous with
an average slope of 23%. 6% of the slopes are exceeding 31°
where the problem with the layover of the C-band SAR starts
and 0.6% are exceeding 50°, where the problems with the X-
band are starting. Together with the X-band data a so called
height error map with estimated height accuracies is distributed
showing very well the locations with problems. In the test area
they are especially at the coast line where the terrain is partially
extremely steep up to nearly vertical and in some steep valleys,
visible in figure 2 as elevated parts. The DSM analysis
confirmed the problems in these locations.
Ho Xx
Figure 2. height error map of SRTM X-band in test area
X-band DSM DZ > RMSZ | Bias RMSZ - bias
50m [m] [m] F(slope)
All points 0.55% 11.9 -5.7 19+ 96% uno,
Open areas 0.67% 10.7 -3.5 7.2+87*tan a
Forest 0.39% 13.8 -8.1 7.8* 9.6 * tana
check points 0 5.4 -2.0 1.3 * 40.6 * tana
Table 1: RMS discrepancy of the SRTM X-band DEM against
the reference DEM from the map 1 : 25 000 and check points
The analysis of the SRTM X-band DSM shows a clear
dependency of the root mean square Z-differences upon the
terrain inclination. In the flat areas the accuracy is in the range
of 7m up to 8m. Quite better results can be seen at 43 check
points which have been determined by GPS. The overall
accuracy of 5.4m is still better with 2.9m for the 20 check
points located in an area with an inclination below 7%. The
larger discrepancies in the forest areas are shown in all data sets
described in this presentation. In the case of Radar images the
main influence is coming from the difference between the
visible surface and the reference DEM, indicated by the clearly
higher bias. A filtering of the DSM to a DEM has not improved
this relation because no height values are available from the
bare ground in the forest area (see Passini et al 2002).
Nevertheless the influence of the forest is limited caused by the
not very high trees.
The comparison of the C-band data with the reference DEM is
leading to very similar results like the X-band data. Only the
comparison with check points shows larger discrepancies. In the
flat area with inclination below 7%, the root mean square
discrepancies are 4.9m instead of 2.9m for the X-band data.
