International Archives of the Photogrammetry, Remote Sensing and Spatial Information 
   
  
  
  
7. ACCURACY ANALYSIS 
Two DSMs have been generated using the two orientation 
methods described in Section 4. These DSMs has been 
compared to the reference DEMs provided by the HRS-SAP. 
The main characteristics (location, spacing, source, size and 
height accuracy) of the reference data are shown in Table 2. The 
coordinate system used in the comparison is the Gauss-Krueger 
system, Zone 4, with Bessel ellipsoid and Potsdam datum. 
Two accuracy tests have been performed in 2.5 D and 3D 
respectively. In the first test the differences between the heights 
of the reference DEMs and the corresponding height 
interpolated from our DSMs have been computed (2.5D). The 
limit of this approach is that it is highly influenced by the errors 
of the surface-modelling algorithm. Figure 12 illustrates the 
concept with a step profile: even if the measurements (green 
points) are accurate, the modeled profile can not follow the true 
one. Consequently if the terrain height is compared, in the 
correspondence of the step, a large difference (Ah) may be 
measured. For that reason the computation of the 3D orthogonal 
distance between the surfaces (distance d in Figure 12) is 
theoretically more correct. 
Therefore the second accuracy test is based on the evaluation of 
the normal distance (3D) between our measurements and the 
reference DEMs. This test is fundamental in this case study 
where steep mountains (Alps) are present. 
The two tests have been made separately for each DEMs 
obtained by the procedures described in Section 4. The results 
are reported in the next paragraphs. 
d 
Ah | 
G—- TS. 
Figure 12. Modelling problems. The true profile is the full black 
line, the modelled profile is the dashed line. 
7.1 Accuracy tests on the terrain height (2.5D) 
The results obtained by this test are reported in Table 5. It can 
be observed that the accuracy of the generated DSM is more or 
less on 1.0 fi 2.0 pixels level, depending on the terrain type. As 
expected, better accuracy are achieved in smooth and flat areas, 
while in the mountainous area (DEMs 5-1 and 5-2) the RMSE 
are larger. In all datasets some blunders which failed to be 
detected are still present. In the reference datasets called 5-1 
and 5-2 some blunders are even above 100 meters, with bias up 
to 1.0 pixels. Apart from the results-of reference DEM 6, all the 
biases are negative, indicating that our generated DSMs are 
higher than the reference ones. The results obtained by 
orientation Procedure 2 appear slightly better that the 
corresponding results achieved with Procedure 1. The 
differences between the two orientation approaches are in the 
order of quarter a pixel. 
For further analysis, the frequency distribution of the height 
differences is shown in the second and third columns of 
Table 6. In the frequency distribution of the height differences 
two peaks occur, one located around value 0.0 and the other one 
around a negative value (ca. 8m). The relative frequency values 
are correlated to the percentage of presence of trees. In fact trees 
causes negative height difference (the green areas), while the 
open areas have small height difference values. It can be 
concluded that the bias located at fi8m is mainly caused by the 
trees. This is a main problem for extracting DEM by using the 
optical imaging systems, as the light cannot penetrate the 
vegetation. For this reason the areas covered by trees have been 
manually removed from the images and the accuracy tests have 
been repeated. The percentage of removed points is 25, 26, 17, 
28, 75 and 71 for DEM 1, 2, 3, 4, 5-1, 5-2 and 6 respectively. 
The results obtained by the new accuracy tests are shown in the 
last two columns of Table 5. As expected, the negative bias was 
reduced. This is also graphically confirmed by the new 
frequency distribution reported in the forth columns of Table 5. 
The analysis of the frequency distributions shows that in steep 
mountain areas (DEM 5-1 and DEM 5-2) there are positive 
height-difference values. They are probably caused by the 
presence of blunders or by the local smoothness constraints 
used in our matching algorithm. These constraints smooth out 
some steep and small features of the mountain areas under the 
condition that there are not enough extracted and matched linear 
features. 
7.2 Accuracy tests on the orthogonal distance between two 
3D surfaces 
This accuracy test has been carried out with the commercial 
software Geomatic Studio v.4.1 by Raindrop. This software 
calculates the normal distance between each triangle of a 
surface (in our case the reference DEMs) and the closest point 
Sciences, Vol XXXV, Part B1. Istanbul 2004 
    
  
    
   
   
   
   
    
   
    
   
   
   
   
     
    
   
   
   
   
   
   
   
   
   
   
   
   
   
   
    
    
  
   
   
    
   
   
    
    
  
  
Internati 
belongin 
measure 
the refer 
seen à 
deviatioi 
  
DEM 
  
  
  
  
  
  
  
  
  
Table 6. 
  
DEM