International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BI. Istanbul 2004
ar
Xu
Figure 7 hing a sub-scene — size of
correlation coefficient displayed as grey value
grey value 255 (white) 7 correlation coefficient 1.0
grey value 51 (dark grey) 7 correlation coefficient 0.6
can also be used as quality indicator. The matching has been
done in 15 main sub-areas separately. The root mean square y-
parallax error is ranging from Spy-4.67m up to 7.11m with a
mean value of Spy=6.0m corresponding to 0.6 pixel (pixel size
in y-direction = 10m). 0.9% of the intersected points have not
been accepted; they exceeded the tolerance limit for the y-
parallax of 30m. In total, approximately 27 million points have
been determined in the southern model where the whole area of
12000 x 12000 pixels has been used and the northern model
where only 12 000 x 8000 pixels where included. The spacing
of 3 pixels in both directions corresponds to 15m in the orbit
and 30m across the orbit direction.
The matching has been done independently for both models. A
comparison of the overlapping part is showing approximately
the same results for the different terrain classes and also only a
very small dependency upon the terrain inclination. The root
mean square difference of 4.Im corresponds to 2.9m for the
individual model and this to a standard deviation of the x-
parallax of 0.5 pixels — similar to the discrepancies at the
control points.
RMS | bias RMS - RMS F(slope)
[m] [m] bias
all points 4.1 0.7 4.0 3.3 + 0.85 x tan a
open area 4.0 0.8 4.0 32+137xtane
forest 4.5 1.6 4.2 3.9 0.22xtana
Table 1. root mean square discrepancies northern against the
independent southern model
5.1 Test areas with laser scanner data as reference
As mentioned in the introduction and shown in figure 2, only
for some sub-areas reference data are available. The test areas
with laser scanner DEMs do have a size of each 5km x 5km.
The reference DEMs with a spacing of 5m do have an accuracy
better than 0.5m, but available only in a resolution of full
meters. These test areas do have a similar topography which is
between flat up to rolling with height variations in the range of
200m. In the average 2096 is covered by forest, which is located
primarily in steep parts.
RMSE | bias | RMSE -
[m] [m] bias
RMSE F(slope)
DSM: 10.2 -5.5 8.5 8.7 + 10.6 x tan à
all points
DSM: 6.7 -3.0 5.9 6.4+4.9xtan à
open areas
DSM: forest 17.0 | -14.3 9.2 16.4 -3.4xtan «
1.0 0.6 ol; 6 5
DEM: 3.7 -2.0 S1 5.0+54xtan a
all points
DEM: 4.4 -1.3 4.1 42+ 1.6x tan à
open areas
Figure 9: frequency distribution of correlation coefficient,
left: area shown in figure 8 right: difficult sub-area
DEM: forest 12.3 -8.5 8.6 10.0 + 6.9 x tan a
horizontal scale = correlation coefficient vertical = frequency
5. ANALYSIS OF THE ACHIEVED DSM / DEM
Based on the orientation determined with program BLASPO,
ground coordinates of the matched points have been computed
by intersection. The matching in the image space has the
freedom in all directions, so the y-parallax of the intersection
Table 2: root mean square difference DEM / DSM of test areas
against reference DEM determined by laser scanner
The automatic matching of all optical images leads to a DSM
with the points on top of the visible objects like trees and
buildings. Also Interferometric Synthetic Aperture Radar
(InSAR) using a shorter wavelength like the X-band, but also
the C-band corresponds to this. Only long wavelength InSAR is
partially penetrating the vegetation. Caused by this, the large
RMSE-values of the first three lines in table 2 can be explained.
Inte
The
groi
neg
heig
loca
buil
rem
the
45%
disc
han
Om
vege
elev
grou
iden
a cle
The:
incli
poin
terra
