inbul 2004
'ighboured
rors. With
of larger
t.
nge of the
'esponding
beside the
g, but it is
? matching
round are
horter and
gth of the
|! sequence
individual
f the local
is used, in
It will be
even point
All these
nts which
> analysed
clined or
led by the
ce of cells
n be used,
zing to the
f directly
ited if the
re using a
lefined by
in a ditch
d by laser
w the true
etermined
Or.
d by the
d iteration
move also
is may be
ot optimal
ymparison
different
separation
Hannover
NAL can
's. It also
acy as a
to lefine
rer ancies
cte:! from
he HRS-
t location
be found.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004
The points are usually located closely to road junctions, so the
distance to the exact image position had to be estimated ( see
figure 3)
position close to street junction
The image orientation has been determined with the Hannover
program BLASPO for the bundle orientation of satellite line
scanner images. It is using just the given view direction and the
general orbit information (inclination and ellipse specification)
in addition to control points. 4 unknowns have to be determined
together with some additional parameters. At least one
additional parameter has to be used for respecting the satellite
speed. So by theory, the orientation could be determined just
with 3 control points, but finally 46 have been used for the
southern and 14 for the northern model. No control points are
available in the Austrian part, so on the right hand side (figure
4) no points are located. For the northern model control points
have been distributed only for the south-west part (figure 4,
right hand side). This is not an optimal distribution, but the test
area Vilsbiburg is well surrounded by control points, so no
problems have to be expected.
matching in the image space with region growing. The least
squares method is the most accurate possibility of image
matching with advantages especially in inclined areas. A
matching in the image space allows a use for any image
geometry; no special mathematical model is required. At least
one start point with the corresponding positions in both images
must be given. From this seed point neighboured points in any
direction are determined by matching and these are again seed
points for the next.
An image matching is only possible with some image contrast,
so a matching on a water surface is not possible, but also in the
forest some problems may occur because of limited grey value
variations.
Figure 3. typical location of control points with estimated
Figure 5: grey value histogram of both HRS-scenes
for both: standard deviation = 24 grey values
Figure 6: typical grey value
histogram of an open forest,
standard deviation = 6.1 grey
values
Figure 4. distribution of control points:
left: southern scene right: northern scene
The bundle orientation was leading for the southern model to
following root mean square discrepancies at the control points:
SX=6.0m, SY=5.8m, SZ=3.9m and in the northern model to:
SX=7.7m, SY=5.0m, SZ=3.5m. So the results for both models
are similar. Respecting the problems of the point identification
this is a sufficient result in relation to the pixel size of 5m x
10m. The better results for the height are demonstrating the
higher accuracy potential of the SPOT HRS system. The
vertical accuracy corresponds to a standard deviation of the x-
parallax of 0.6 pixels (in relation to the Sm pixel size in orbit
direction).
The orientation accuracy has been confirmed by the image
matching. The rmse between vertical differences of the control
points and the matched DEM is just 3.06m; that means it is
better than the Z-discrepancies of the orientation. This can be
explained by problems of the manual pointing of the control
points which is not so accurate like automatic matching.
4. IMAGE MATCHING
The image matching for the generation of the DEM has been
made with program DPCOR. It is using a least squares
441
The grey value variation for both scenes (figure 5) is not
optimal, but sufficient with the exception of some parts in the
forest (figure 6). Depending upon the area, 85% to 90% of the
possible points in the southern model have been matched with a
sufficient correlation coefficient exceeding the used limit of 0.6.
As it can be seen in figure 7, the success of matching is not
equal distributed. A matching is very difficult if it is too steep.
In addition a matching is also not possible in the snow covered
parts — only the border of the snow can be used. But in general
the coverage is satisfying and sufficient for a DEM generation.
The quality map of the image matching (figure 7) shows very
well the areas where the matching is more difficult. Especially
in the forest areas (dark parts) the contrast is limited causing
also a low correlation coefficient which sometimes is also
below the used tolerance limit of 0.6. The frequency
distribution of the correlation coefficient is varying dependent
upon the landscape. On the left hand side of figure 9, the
frequency distribution of the typical sub-area shown in figure 8
is shown, where 92% of the possible points have exceeded the
acceptance limit of r=0.6, while on the right hand side the
extreme situation of the steep mountains, partially with forest
and also several small lakes is shown, where only 85% of the
points have been above the limit.
