International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
different to that of the Melbourne images. The San Diego area
has a much more homogeneous texture since it is almost
entirely forested. Conversely, the Melbourne scene is an urban
area with a more heterogeneous texture and perhaps more
ideally suited to image matching.
6.2 Comparison with reference data
Although an analysis of the correlation coefficients can be
instructive, it does not reveal sufficient information by which to
assess the success of the matching algorithm and the quality of
the matched points: it is entirely possible that some of the well
matched points with correlation coefficients greater than 0.8
may be blunders. Further conclusions can be inferred by
triangulating the matched points to give object space
coordinates (using the affine model), and comparing those
coordinates with a reference digital elevation model (DEM).
This process was carried out with the Melbourne data where a
high quality reference DEM was available. This DEM was
created from aerial photography and resampled to a grid
spacing of 25m. Significant manual post-processing was carried
out on this DEM to ensure it gave an accurate representation of
terrain heights. As a result, it would be expected that points
triangulated from the matched Ikonos points would be
uniformly higher, since they represent the surface of the objects
in the imagery, and not the terrain.
Geometrie Matched points with correlation
constraitit coefficient > 0.8
Mean (m) Std. Dev. (m)
Epipolar 3.41 6.80
Affine 3.64 7.60
Table 3. Differences between triangulated points and reference
DEM - Melbourne Test Area
As expected, it can be seen from table 3 that the differences
between the triangulated points and the reference DEM are all
positive, and of a magnitude that could realistically represent
the mean building heights within the test areas. Also
noteworthy is the fact that the differences between the affine
constraint results and the epipolar constraint results are small,
although the surface created from the points matched using the
affine constraint is marginally higher than the surface created
from the epipolar constraint.
Since no reference DEM was available for the San Diego test
area, a similar comparison was not possible.
6.3 Visual analysis
Finally a visual analysis of the results was carried out. A
surface model was created for each test area from the
triangulated points. Figure 3 shows a 5m grid DSM of the
Melbourne test area, while figure 4 shows a similar DSM of the
San Diego test area. In each figure the lighter tones of grey
represent high elevation, whilst the darker tone represent low
elevation.
Figure 3. DSM of Melbourne test area
Figure 3 clearly illustrates the potential of matching stereopairs
of Ikonos images. The resulting DSM shows many features,
including buildings, vegetation, bridges, overpasses and even
roads and railways. Although some blunders are apparent,
particularly in the river in the eastern part of the test area, the
majority of the matching has been successful. Even with the
presence of the surface features, it is still possible to recognise
the pattern of the underlying terrain, with areas of high ground
in the northern part of the test area. Note that this image has
been shown from a near-nadir perspective due to the fact the
terrain varies very little across the image.
Figure 4. DSM of the San Diego test area
Figure 4 shows the San Diego DSM displayed in perspective
view (with the vertical scale exaggerated). It is clear to the
observer that the DSM is a very good representation of the
underlying terrain, with hydrological and topographic features
standing out very clearly. There are very few, if any, obvious
blunders.
7. DISCUSSION
This paper has presented the use of two different geometric
constraints on two stereopairs of high resolution satellite
imagery. The results of the matching procedures have been
analysed by assessing the correlation coefficients, triangulating
International Archi
individual points
Consequently these
constraint (based o
good as the imag:
epipolar model). A
that the matching
coordinates. If ap
matching search sp
and hence processi
can be reduced. (W
parameters are muc
further confirm the
geometric processit
Images not aligned
have also been su
epipolar model. Sir
those from the affi
of the quadratic epi
result since it mean
stercopair) can be n
Success of these
largely contingent
linearities. Fortunat
effects. Future wor
multi-temporal, san
well as multi-tempo
8...
This work was cai
fellow at the D
Melbourne. The au
Dr. Harry Hanley
Yamakawa for the
research.