04
xth
es.
1al
1S
/as
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
hed Rid
matching results in Alpine area.
E dus Er >. s
Figure 6. Shaded 5m DSM generated from IKONOS. The city
of Thun in the upper left seen from North-West.
CF
The procedure mainly contains the following characteristics:
I) It is a combination of feature point, edge and grid point
matching. The grid point matching procedure uses
relaxation-based relational matching, and can bridge-over
areas with no or little texture through local smoothness
constraints. The matched edges are introduced to control
the smoothness constraints in order to preserve the
surface discontinuities.
2) The adaptive determination of the matching parameters
results in a higher success rate and less blunders. These
parameters include the size of the matching window, the
search distance and the threshold value for cross-
correlation and MPGC. For instance, the procedure uses a
smaller matching window, larger search distance and a
smaller threshold value in rough terrain area and vice
versa. The roughness of the terrain can be computed from
the approximate DSM on a higher level of the image
pyramid.
Linear features are important for preserving the surface
discontinuities. À robust edge matching algorithm, using
the multi-image information and adaptive matching
window determination through the analysis of the image
content and local smoothness constraints along the edges,
is combined into our procedure. One example of edge
matching is shown in Fig. 5.
4) Edges (in 3D) are introduced as breaklines when a TIN-
based DSM is constructed. This DSM provides good
approximations for matching in the next pyramid level.
The computation of the approximate DSM in the highest
pyramid level uses a matching algorithm based on the
2
_
r2
"region-growing" strategy (Otto and Chau, 1988), in
which the already measured GCPs and tie points can be
used as "seed points".
5) If more than two images are available, the MPGC
procedure can use them simultaneously and matching
results are more robust. Here, the resulting DSM from an
image pair can be used as approximation for the MPGC
procedure.
6) Through the quality control procedure, e.g. using the local
smoothness and consistency analysis of the intermediate
DSM at each image pyramid, the analysis of the
differences between the intermediate DSMs, and the
analysis of the MPGC results, blunders can be detected
and deleted.
For each matched feature, a reliability indicator is assigned
based on the analysis of the matching results from cross-
correlation and MPGC. This indicator is used for assigning
different weights for each measurement, which are used when
a regular grid is interpolated.
4.2 Test Results
For Thun, we used for initial matching the images (and the
respective triangulation results) of the triplet and stereopair
separately and for the final MPGC all 5 images. The patch size
varied from 7^ to 17^ for initial matching and was 11^ for
MPGC. Some areas like lakes and rivers were manually
defined as "dead areas" via a user-friendly interface. A regular
grid DSM with 5m spacing was interpolated from the raw
measurements. Fig. 6 shows a visualisation of the generated
DSM. In spite of smoothing due to the large area used in each
point measurement, the discontinuities are quite well
preserved.
Tables 9 and 10 show the DSM accuracy results, without any
manual editing. The results are evaluated based on the
differences between the heights interpolated in the reference
laser DSM at the planimetric position of the DSM from
matching and the heights from matching.
The tables show that the DSM accuracy is in the 1-5 m range,
depending on the landcover and terrain type. A very high
accuracy can be achieved in open areas. In these areas, more
than 80% of the differences are less than 2 m. In urban and
vegetation areas, the accuracy is worse, which is due to the fact
that the reference LIDAR measurements and the parallaxes
determined in matching refer to partly different objects.
Matching measures higher than LIDAR at trees (in addition, at
tress LIDAR sometimes measures below the tree tops) and
narrow low-lying objects (like streets). Apart from that, the
time difference between LIDAR and IKONOS data acquisition
was 3-4 years, and the triplet of IKONOS had snow, up to 2-3
m in the mountains. Other factors that influenced matching
were the long shadows (sun elevation was just 19 deg),
occlusions, espec. in the W-E mountains, very low textured
snow areas (which were improved with our preprocessing) and
the patch size used in matching which unavoidably leads to
smoothing of abrupt surface discontinuities. The accuracy
values deteriorate also due to the high bias (see mean values
espec. in Table 9), while height accuracy also gets worse due to
the suboptimal base/height ratio (see sensor elevation and
azimuth in Table 1). Taking all above factors into account, it
becomes clear that IKONOS has a very high geometric
accuracy potential and with sophisticated matching algorithms
a height accuracy of 0.5 m — 1 m can be achieved in open areas
c
Un
