"Hei- 
order 
'oach. 
n area 
were 
y. The 
from 
south 
louds. 
scene, 
nous 
lation 
These 
and 5 
coor- 
these 
lotter 
:. The 
ns for 
GCP 
plate 
ce the 
nin, 
of the 
- 25 pixels for the maximum difference to the initial 
values (corresponding to a maximum height diffe- 
rence between adjacent points of about 30 m), 
- a maximum of 10 adjustment iterations. 
For STEP different values were investigated. Since the 
computing time increases roughly by STEP?, STEP 
should be chosen as large as possible. On the other hand 
it was found that neighbouring template matrices must 
overlap in order to achieve acceptable results. This 
means that STEP must be smaller than the size of the 
template matrix. Best results were obtained for values 
between 10 and 15. For STEP = 10 the algorithm pro- 
vided more than 230.000 conjugate points. From the 
criteria mentioned above pmin = 0.6 proved to be the 
most important one. More than 70 % of those points, 
which were found to be incorrect, had a correlation 
coefficient below 0.6. The coverage of the images was 
fairly equal. 
For the bundle adjustment the following information 
was introduced: 
interactively measured image coordinates of the 834 
check points. The observations were treated as un- 
correlated with an equal standard deviation of 
Oo= Sum. 
- automatically derived image coordinates of about 
6600 equally distributed conjugate points such that 
correlation coefficient of each point is a local maxi- 
mum with the same o,, 
- avarying number of equally distributed GCP with a 
standard deviation of Oxccr = Oyccer = 3 m and 
OZGCP = Sm. 
- the XYZ object coordinates for the projection cen- 
tres of the 2 * 4 orientation images with a relative 
accuracy of 1 m and with an absolute accuracy of 
1000 m. 
4.1.2. Results. The results can be seen in table 1. 
Besides the introduced number of GCP the root mean 
square errors of all three object coordinates derived 
from the check points are given. The following conclu- 
sions can be drawn: 
- according to the accuracy requirements five to six 
GCP are sufficient in order to obtain accurate re- 
sults, 
- the accuracy in Z lies between 5 and 7 m, 
469 
  
  
  
  
  
  
  
  
"— of RMS errors of object coordinates 
X Y Z 
5 13.1m 21.5 m 7.0 m 
6 13.6 m 17.4 m 5.5m 
10 12.9 m 16.2 m 6.2m 
15 12.8 m 15.5 m 4.3m 
  
  
  
Table 1: Results of point determination 
- the planimetric accuracy is worse by a factor of 2 to 
3. This phenomena has also been observed by other 
authors /Dowman 1992/. A possible explanation for 
this result is the following: most of the check points 
are road crossing centres, and thus lie in relatively 
flat areas. If a point at the border of the road rather 
than in the middle is measured by accident, the 
resulting height is still correct, but the derived XY 
coordinates are not. These identification errors can 
not be detected in the adjustment. 
- the accuracy in X is better than in Y. This is a 
consequence of the use of line imagery: the parallel 
projection in flight direction (similar to the Y axis) 
is less stable than the central perspective in the 
direction perpendicular to the flight path. 
With the determined elements of exterior orientation 
from six GCP a forward intersection was performed for 
each pair of conjugate points obtained from image mat- 
ching. Subsequently a DTM was generated using the 
HIFI programme package /Ebner et el. 1988/. In order 
to determine the DTM quality, and thus to check the 
matching results, heights for the 834 check points were 
interpolated from the derived DTM and compared to 
the known values. An empirical standard deviation of 
10.8 m was obtained. This value represents an inde- 
pendent check of the whole procedure (matching, point 
determination, DTM generation) over the entire image. 
It must be regarded as a very good result considering the 
small base-to-height ratio of 0.4. 
The generation of an orthophoto is a standard task given 
the results computed so far. In this project each ortho- 
photo pixel was projected into one of the images (the 
cloudfree image from November 8) using the pixel-by- 
pixel method /Mayr, Heipke 1988/. Since the necessary 
orientation elements are given for each image line and 
thus for each point in image space after the bundle 
adjustment, but not explicitly for an arbitrary point in