Istanbul 2004 
vad results, in 
ve 10 pixels. 
GCPs it was 
h to orientate 
ts. In order to 
tment at least 
llowed by the 
M GCPs.is a 
ipply satellite 
ifficult. 
ple, to fix the 
attitude to be 
ts only if the 
with ancillary 
Ps. For a grid 
ojection rays, 
yrojected onto 
143 (13 by 11 
mage 3B, and 
curacy of the 
were used to 
r polynomials 
r both images. 
system, vector 
gure 5 shows 
espectively. 
  
d 3B images 
4l. 
icy of the orbit 
eight above the 
ters. It can be 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
     
  
observed that the coincidence is better for the 3N image (errors 
around 35 meters) than for 3B (lack of coincidence around 60 
meters, in opposite direction). À larger error is expected for the 
3B image due to the larger incidence angle. Among these errors 
there is a component resulting from the orbit and attitude data, 
which is only of a few tens of meters. 
It would be possible to improve the orientation method, making 
use of these data, which would require only very few GCPs. À 
possible alternative would be to apply small corrections in 
image space by an affine transformation, as described by Dial 
and Grodecki (2002), for Ikonos. 
2.4 DEM extraction and accuracy analysis 
Once images are oriented, the epipolar images were generated 
and the DEM was extracted for the full image using half 
resolution (30 m spacing). The height range was known to be 
from 0 to 1000 meters and was given to the matching program. 
The matching success was of 98%. The image had a few clouds, 
where the matching failed. However, some lower clouds were 
extracted. Figure 6 shows an image of the DEM were clouds 
can be identified (some not extracted, in black, and the ones 
extracted, with heights near the maximum, in white). Small 
gaps were filled by interpolation. 
  
Figure 6. DEM extracted for the whole area. 
Failures also occurred in water surfaces, which are very 
uniform. That is the case of the sea, in the bottom left of the 
image. 
Figure 7 shows the extracted DEM (a) without any 
interpolation of gaps or smoothing, for a region of 
approximately 11 by 8 km and the corresponding image (b). 
There are two rivers in the area where some heights could not 
be extracted. However, some heights where wrongly extracted. 
Editing should be done in such areas according to the 
knowledge of the terrain, in this case a flat water surface. 
In order to assess the accuracy of the heights obtained, the 
extracted DEM was compared with a control DEM, derived 
from aerial photographs. This DEM covers an area of 16 by 10 
km and was produced by the Portuguese Army mapping 
service. It is known to have an accuracy better than 2 meters. 
Figure 8 shows shaded relief images of both DEMs. 
  
  
(a) (b) 
Figure 7. DEM extracted (a) for a small area, without any 
interpolation, and corresponding image (b) 
kg 
zs EAE 
  
Figure 8. Shaded relief representations of the control DEM (a) 
and DEM derived from ASTER (b) 
The statistics of the differences between the DEMS are listed in 
table 3 (all values in meters). Figure 9 shows the histogram of 
the differences. 
  
  
  
  
  
  
Min -47 
Max 69 
Mean 4.0 
St. Dev. 8.7 
  
  
Table 2. Statistics of the differences between the two DEMs. 
  
10000 
LA LU LE 
5000 4 
  
  
  
Eee 
40 -30 20 10 0 10 20 30 40 50 60 
Height differences 
Figure 9. Histogram of the height differences between the 
extracted DEM and the control DEM 
The standard deviation of 8.7 m is a very good result and is 
according to what is referred by other authors (Toutin and 
Cheng, 2001, Hirano et al., 2003). There were no clouds or very 
uniform areas, such as water surfaces, and the area contains 
essentially hills with low vegetation and agricultural fields, 
which is the appropriate type of terrain for good results. The 
accuracy analysis was made with very accurate data but over a 
small area. More tests should be done, distributed over the 
whole image, in order to get a more significant figure of the 
global accuracy. 
Two cross sections, one in direction east-west and the other 
north-south, were plotted for both DEMs (figure 10). It can be 
seen that in some slopes the ASTER DEM tends to be above the 
actual terrain.