'stanbul 2004 
ons for the 
rol points in 
TE) 
d looking 
ird looking 
1A) 
  
-xl 
1,22 
  
96 | 
control, the 
sor is in the 
and standard 
ce the values 
the French 
(Bouillon et 
> in the nadir 
n difference 
ed} 
  
  
es derived 
s of SPOT 
a test area 
ged) 
  
  
es derived 
SPOT HRS 
, Catalonia 
  
  
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004 
An interesting behavior shows up when looking at the 
comparison of the three orthoimages in detail. The automatic 
matching of the orthoimages reveals that the difference vectors 
(each a mean value in a 200 x 200 pixel squared area) show a 
very homogeneous behavior, mean length is about 15 meter 
(Fig. 1) — the shift shows up predominantly in flight direction. 
This means that by using one very good and exact ground 
control point, the absolute accuracy of the orthoimage can be 
improved and the images can be used as matched correctly (see 
e.g. Nonin et al. 2003). Only the nadir looking channel (HMA) 
shows a different behavior, since here the matching differences 
(arrows in Fig. 2) show variable shift, which depends on the 
position in the CCD array. 
Remark: This systematic behavior, which shows mainly a 
constant shift between the two images, is a result of good 
relative orientation for the single images, but an absolute 
pointing change between the forward and backward data 
acquisitions (-90 sec. time difference). By using the values of 
table 1 the corrections of the angular changes are 0.0013? for 
HRSI and 0.0008? for HRS2 for the Catalonia test site. These 
values are in line with the accuracy specifications of the data 
provider. The reason for these residual rotations are probably 
due to the uncertainty of the initial attitude values, but should 
be discussed with the data producers for further investigations. 
The measured residual orientation values have been applied for 
corrections to the attitude values, which leads to nearly accurate 
matched orthoimages. Similar results are found for test site 
Bavaria (Reinartz 2004). 
5. DEM PRODUCTION FROM TWO RAY STEREO 
DATA 
The first matching of the two images is performed purely in 
image space with DLR software. Details on this software are 
described in Lehner et al. 1992. It relies on a 7-step image 
resolution pyramid and applies intensity matching in two forms: 
normalized correlation coefficient for pixel accuracy and 
subsequent local least squares matching (LLSQM) for 
refinement to sub pixel accuracy (for mass points 0.1 to 0.3 
pixel standard deviation, depending on the radiometric quality 
of the imagery). First interest points are generated with a 
Fórstner operator and the homologous points are searched for in 
the other image. Only points with high correlation and quality 
figure are selected as tie points for bundle adjustment (see 
chapter 7) and a less stringent criterion is valid for the usage as 
seed points for the subsequent Otto-Chau region growing 
procedure for dense matching (Heipke et al 1996). This local 
least squares matching starts with template matrixes of 13. x 13 
pixels around the seed points with a constant step in each 
direction (here three pixel). For cross checking a backward 
match is performed for all points found. From the differences of 
the image coordinates a standard deviation of about 0.14 pixel 
is found. Points showing differences larger than 0.5 pixel in the 
backward matching are eliminated. 
Having the mass points from the matching process as well as 
the exterior and interior orientation of the camera system, the 
Object space coordinates can be calculated using forward 
intersection. This is done by least squares adjustment for the 
intersection of the image rays. Intersections with weak 
geometry (threshold determined using intersection constraints 
of high quality homologous points) are rejected. 
The irregular distribution of points in object space after the 
forward intersection has to be regularized into a equidistant grid 
of about 15 x 15 meter pixel size. The interpolation process is 
performed by a moving plane algorithm (Linder 1999). The 
resulting DEM, which are surface models, are compared to the 
reference DEM, which are terrain models. Therefore a distinct 
difference is expected e.g. in forest areas. 
The area covered show besides the city of Barcelona, the very 
steep mountainous area of Montserrat as well as the moderate 
mountains of Tibidabo and others. The comparison of the DEM 
is therefore performed in different areas: cities, open areas and 
forest areas, which are masked using classification results of the 
orthoimages. Fig. 3 shows the derived SPOT-DEM calculated 
by using two ray intersection. 
  
Figure 3: SPOT-DEM of Barcelona and surroundings 
First the “best” homologous points for two-fold imagery as 
projected to object space, are investigated. The result is very 
close to the result achieved in Bavaria (Reinartz 2004) and 
shows again a very good absolute accuracy without using any 
ground control information (table 2). 
Table 2: Comparison of height for high quality homologous 
points in SPOT-DEM derived from two ray intersection and 
the reference DEM of 67 x 67 km? with 1.1 m accuracy 
  
Mean Height Difference [m] | Std. Dev. [m] # Points 
  
  
  
8.8 3.4 101858 
  
  
6. DEM PRODUCTION FROM THREE RAY STEREO 
DATA 
For Catalonia test site the images of four cameras are available, 
the off-nadir looking HRS1/2 and the nadir looking HMA/B 
(two 5 meter resolution bands). This offers the possibility to 
derive DEM from the stereo channels HRS1 and HRS2 (called 
two ray intersection) and additionally to take into account the 
nadir looking bands (called three ray intersection). For the 
investigation only the band HMA was included for DEM 
generation (no interpolation to 2.5m resolution of the HMA / 
HMB Supermode image was performed). The overlap region 
can be seen in figure 4.