International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
interpolation, smoothing and as a last step burning proper 
elevations in. Except for the sea surface, only a small part of the 
edited as in Figure 5. 
  
  
  
  
  
pem CLE Gus 
Figure 5 Failed areas and blunders in the extracted DEM 
2.34 DEM Geo-coding 
After completing the editing, extracted DEM is transferred to a 
geo-coded UTM file. File is geo-coded using the information 
from the model segment. The research area takes place in zone 
35 raw S of UTM projection systems. European Datum 1950 
(ED50) parameters are used for geo-referencing (Figure 6). 
Figure 6. Geo-coded DEM generated from RADARSAT FI-F5 
stereo pairs 
The model is derived from ground control points collected 
previously, the ephemeris and the orbital data from the satellite, 
and from the knowledge of the Earth size and the shape. The 
accuracy of the model depends mostly on the accuracy of 
GCPs. The acquisition of GCPs from SAR imagery is difficult 
as the radar sensor response to ground relief and cover is very 
different from the familiar optical sensors. To collect GCPs 
homogeneously covering the stereo overlapping area and on the 
altitudes of extreme elevations is very important. When 
unexpectedly large Y parallax occurs, the only solution is to 
acquire additional GCPs in the affected area. 
Collecting GCPs are important also to remove the bias, or the 
offset of a DEM. In order to match the actual terrain 
coordinates, bias is removed. To measure absolute accuracy the 
amount of vertical and horizontal shifts must be removed. Zero 
offset is ideal which is aimed in this research by collecting 
GCPs to correct the images. Besides acquiring GCPs in the 
field with static GPS, a 1/ 25 0000 scaled reference digital map 
is used. Reference map is generated by scanning 8 analog 
topographic sheets obtained from the General Command of 
Mapping. Topo sheets are first rectified, and a mosaic map is 
generated. GCPs are collected from this mosaic map as well as 
from other orthorectified images related to the research area. 
2.4 Accuracy Assessment of Stereo RADARSAT DEM 
The absolute accuracy is a measure of the error between the 
extracted DEM and the geographic coordinates of the actual 
terrain. To assess the accuracy, a reference digital terrain model 
(DTM) is generated from | 1/25000, scaled contour maps. 
Planimetric accuracy of 1/25000 scaled maps are about 5 m, 
and elevation accuracy is expected to be equal to one thirds of 
elevation differences between contour lines. Since the contour 
interval is 10m in 1/25000 scaled maps, the accuracy of the 
reference DEM is assumed to be 3m, which is acceptable using 
for the accuracy assessment of the elevations calculated from 
stereo radar images. In order to calculate the difference of the 
elevation models, the DTM is subtracted from the DEM using 
subtraction model of PCI software. 
Absolute accuracy is expressed as the vertical RMSE, or root 
mean square error measured at geographic coordinates. Using 
that disparity map (Figure 7), standard deviations are calculated 
first for the entire training area. 
  
Figure 7. Elevation disparity map calculated as topographic 
DEM (DTM) from radargrammetric DEM 
It is observed that errors are in ascending character according to 
the type of the terrain and the slope (Figure 8), (Table 3). In 
order to increase the confidence level of the accuracy 
assessment, the elevation accuracy of final DEM is evaluated as 
a function of topographic surface. A digital slope model (DSM) 
was computed from the topographic DEM (Figure 9). Accuracy 
assessment is repeated for flat, moderate and mountainous areas 
and histograms are evaluated separately. It is inferred from 
mean, standard deviations and min/max errors that elevation 
accuracy of stereo pairs decrease consistently as the slope 
increases. It is determined that elevation accuracy and slope are 
almost linearly correlated. 
  
  
  
  
Greg Lk Vak 
Figure 8. Error histograms for flat, slope and rough topography 
respectively. 
Internat 
  
Test 
Arcas 
Flat 
Flat 
Flat 
Flat 
Slope 
  
  
Slope 
Slope 
Slope 
Rough 
c 
  
  
Rough 
c 
  
Rough 
Rough 
  
  
  
Table 
Elevatic 
rough a 
better (s 
disparit 
due to t 
two of 
solution 
and mo 
deviatio 
On the 
are calc 
calculat 
ÁS a res 
strongly 
from S/ 
of SAR 
extractic 
slope. R 
due to t 
sensor ; 
towards 
the sens 
extreme 
back s 
derivati 
looks si 
dispariti 
too larg 
In addit 
SAR im 
may cat 
large tei