ibul 2004 
)EM 
vhole area 
jeomatica 
generated 
res. Some 
nificantly 
re are lots 
es exist in 
set of the 
generated 
covered a 
30 meter. 
  
x A 
eim 
ing areas 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004 
  
5. ACCURACY ASSESSMENT 
As reference data, an existing 10 meter resolution DEM of the 
study area, which was produced from 1:25000 scale digital 
vector maps, was used. The heights of some 97 points were 
extracted from the generated DEM and compared with their 
correspondings in the reference DEM. Number of test points (n) 
selected for this assessment was determined by Equation 1. 
p NZ” pq 
Nd" 4 Z? pq 
where N = total number of population elements 
Z = confidence level 
d = needed accuracy 
In this equation n is number of test points and N is total number 
of population elements that in this case is infinite. Z is the value 
corresponding to confidence level and d is the needed accuracy. 
The value of p is proportion of elements that have especial 
characteristic and q is the other elements without this 
characteristic (p = 1- q). In the case of infinite number of 
population elements, number of samples for accuracy 
assessment is determined with Equation 2. 
  
ped. 2) 
di 
For example by assuming p = 0.50 and d = 0.10 and confidence 
level equal to 95%, n would be equal to 97. These points have 
been generated from production of some random numbers with 
uniform distribution as x and y coordinates. Since the chance of 
selection of each number must be equal to the others, the 
uniform distribution was applied. For computing the accuracy 
of extracted DEM from stereo images, first the subset of 
generated DEM without any editing with interpolation was 
evaluated. Some points that posed in failing area were 
eliminated from points list. The preliminary statistics of this test 
are as Table 2. 
  
  
  
  
  
  
Number of check points (CPs) 97 
Mean of errors (absolute values) 3.87 m 
Mean of errors 1.31 m 
Maximum error (absolute value) 21.00m 
Standard deviation of errors 5.32 m 
Frequency of errors dz»1m n=89 91.8% 
dz>2m n=67 69.1% 
dz>3m n=48 49.5% 
dz>4m n=39 40.2% 
dz>5m n=33 34.0% 
dz»6m n=20 20.6% 
dz7m n-17. 17.596 
dz»8m n=14 14.4% 
dz»9m n=11 11.3% 
dz>10m n=08 08.2% 
  
  
  
  
Table 2. The statistics of DEM test without editing 
Also, the subset of extracted DEM with editing of failing area 
with interpolation was assessed with 97 points. Then some 
points that posed in failing area were assessed with height 
produced from interpolation. The statistics of this test are as 
Table 3. 
  
  
  
  
  
  
Number of check points (CPs) 97 
Mean of errors (absolute values) 3.94 m 
Mean of errors 0.85m 
Maximum error (absolute value) 13.00 m 
Standard deviation of errors 4.89 m 
Frequency of errors dz>im n=87 89.7% 
dz>2m n=73 75.3% 
dz>3m n=62 63.9% 
dz>4m n=48 49.5% 
dz>5m n=36 37.1% 
dz>6m n=27 27.8% 
dz^7m n-18 18.6% 
dz»8m n-12 12.4% 
dz>9m n=06 06.2% 
dz>10m n=04 04.1% 
  
  
  
  
Table 3. The statistics of DEM test with editing 
For better investigation of this product, the whole of the region 
without any editing was also evaluated with some random 
points. Some points that posed in failing area were eliminated 
from points list. The results of this test were shown in table 4. 
  
  
  
  
  
Number of check points (CPs) 97 
Mean of errors (absolute values) 2.67 m 
Mean of errors 0.53 m 
Maximum error (absolute value) 16.00 m 
Standard deviation of errors 3.99 m 
  
dz»l1m n=76 78.4% 
dz>2m n=53 54.6% 
dz>3m n=38 39.2% 
dz»4m n=28 28.9% 
d7>5m n=17. 17.5% 
dz>-6m n=12 12.4% 
dz>7m n=08 08.2% 
dz>8m n=07 07.2% 
dz»9m n=04 04.1% 
dz>10m n=04 04.1% 
Frequency of errors 
  
  
  
  
Table 4. The statistics of DEM test without editing 
Finally, the whole extracted DEM with editing of failure area 
with interpolation was assessed with 97 points. Some points that 
posed in failure area were assessed with height produced from 
interpolation. The statistics of this test are as Table 5. 
Comparing table 2 and table 3 shows that the standard deviation 
of the subset of extracted DEM with editing of failing areas is 
better than the standard deviation of the subset of extracted 
DEM without any editing. That is because of production of 
more suitable heights using interpolation of neighboring heights 
in failing areas. In the subset of extracted DEM, some small and