International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004
418
41.7} 3 approximate coastline
sub-area for direct comparison
acl HRS1
of 2-ray and
3-ray results
g 415 F- HRS2 4
3
a
o
c
9 414 4
413
41.2 4
2 HMA
411F 4
41 À 1 À n A À + 1
16 18 2 22 24 2.6 28 3 32 34
latitude
Figure 4: Location of HRS/HMA scenes for the Catalonia
test site
Matching was done in several steps in order to generate 3-ray
tie points with sub-pixel accuracy. The first step is the
generation of seed points with DLR matching software as
mentioned in chapter 5. A resolution pyramid of 7 levels (factor
2 reduction in resolution from level to level) was used for a
triple of HRSI, HRS2 and a HMA version which was
resampled to HRS1/2 resolution (reduction by a factor of 2 in
across-track direction). This first matching — via maximum of
normalized correlation coefficient and subsequent refinement to
sub-pixel accuracy with local least squares matching (LLSQM)
— generated about 190000 3-ray tie points between HRS1/2 and
reformatted HMA imagery. On the highest resolution level
(original image resolution of HRS1/2) a LLSQM window size
of 29 rows (along-track, 5 m ground pixel size) and 17 columns
accepted points the shifts of the HRS2 coordinates
have a mean of 0 pixel and standard deviations of
0.17 (along-track, 5m resolution) and 0.09 pixel
(across-track, 10m resolution) respectively
The whole matching process ended with a number of 8555000
3-ray points.
7. COMPARISON OF DEM FROM TWO AND THREE
RAY STEREO DATA
Table 3 shows the comparison of the mass points from the Otto-
Chau region growing dense matching algorithm with the
reference DEM after forward intersection and rejection of
points with weak intersection geometry (the threshold for
rejection is determined by the intersection geometry of high
quality homologous points). Image matching with three images
is supposed to provide better control mechanism for the
homologous points by the improved check via a third stereo
image discussed above. The three ray matching process was
based on HMA 5 x 5 m resolution imagery, which results in
about twice as much points for region growing. Slightly better
results of 0.6m less in standard deviation are obtained from
three ray forward intersection. Also minimum and maximum
values are reduced significantly. In all two cases more than
99.96 % of the matching points are fitting the reference DEM
better than 50m after adjusting the mean height.
Table 3: Comparison of mass points derived from two
(HRSI, HRS2) and three ray (HRSI, HRS2, HMA)
intersection with the reference DEM.
(across-track, 10m ground pixel size) has been used to
compensate for the different resolutions in row and column
directions (5 m and 10 m, respectively). Next, to get 3-ray tie
points in original resolution (5m x 5m), the HMA column
coordinates of the tie points from the first step are changed to
original resolution and then put into a new LLSQM matching
with images of differing resolutions (which is easily possible
via LLSQM because of the inbuilt estimation of an affine
transformation — together with some interpolation scheme:
Intersection Intersection
of two rays of three rays
Amount of Points 1418965 2662143
Points « 20m 1407636 2652764
(99.20 94) (99.65 94)
Points < 50m 1418414 2662076
(99.96 %) (~100 %)
Standard deviation [m] 5.76 5.16
Mean height difference [m] 9.54 9.89
Min. height difference [m] -151.5 -61.7
Max. height difference [m] 148.2 1953
bilinear interpolation of grey values is used by the DLR
matching software to get HMA-resolution chips out of HRS
imagery). After careful sub-selection about 20000 tie points
were transferred to the next step of LLSQM with Otto-Chau
region growing.
To exploit the checking possibility available for three stereo
pairs three matching steps with the software for region growing
(see chapter 4) have been performed using a grid spacing of 3x3
for the growing:
I. Matching between HMA and HRSI (original
resolutions) using seed points from DLR matching
software (result: 9720000 points)
Matching between HMA and HRS2 for all HMA
points resulting from step (1.) and combination to 3-
ray points (result: 9404000 points)
3. Check by matching between HRS1/HRS2 for all pairs
found in steps (1-2), (4% could not be matched)
4. Sub-selection by a threshold of 0.5 pixel between
HRS2 coordinates from steps (2.) and (3.) (5% did not
match the threshold condition); for the 91% of
D
472
After regularization of the SPOT DEM into a equidistant grid
by a moving plane algorithm (Linder 1999) (15m x 15m pixel
size), a comparison with respect to different classes (forest,
settlement and open areas) is performed. Table 4 shows that the
mean height differences and the standard deviation are best for
open areas, because the reference DEM is a digital terrain
model compared to the surface model of the SPOT DEM.
Slightly better results of about 0.5 m reduced standard deviation
are obtained for the case of three ray forward intersection.
Table 4: Comparison of a regularized SPOT-DEM derived
from two (HRSI, HRS2) and three (HRSI, HRS2, HMA)
ray intersection with the reference DEM for three different
classes (forest, open areas and settlements)
Intersection of two rays
Forest Open Cities
Areas
Mean height difference| 10.82 9.84 10.69
[m]
Standard deviation [m] 7.24 4.73 5.02
Intern
Min.
Max
Stan
Min.
As sl
chanr
to grc
shift
slight
DEM
It cou
data,
leads
about
about
data)
two £
DEM
comp
refere
classe
deper
about
achie
stand:
expec
whicl
proce
of the
chanr
The a
point:
(Rein
imprc
imprc
Orthc
1 to
in lin
the c
const
grour
S. À
Gach
Revie
High
2003,
A. B
Geon
July :
Sym
