2.4 Fusion of the results of both processes and final data 
classification. 
Once obtained, the first segments in the scene (segmentation 
process) and on the other side, the first approximation of the 
DTM represented by the DSM positions labeled as ground 
points (triangulation process), it can be carried out a merging 
process of these results oriented to the initial DTM 
densification. 
This process will consist of two distinct sub-processes. First, 
using triangulation points as seed points, being classified as 
ground segments all those segments that include any point of 
each triangulation. Next, a triangulation densification of the 
initial triangulation must be made. In this sub-process, we 
incorporate all those points of these segments that they are not 
previously incorporated. 
With this process, we can obtain the final triangulation that 
represents the final DTM. Using this surface, the original data 
can be classified into two classes: ground data and non-ground 
data, analyzing the distance between the points and the 
generated surface. Figure 3 shows the densification process and 
the data classification. 
  
Figure 3. Upper: Terrain segments generation process; Lower: 
left) Definitive triangulation densification; right) 
Final classification of ground data. 
3. RESULTS 
To analyzing the results from the proposed methodology, the 
approach has been applied in an area of 550m x 550m size with 
a steep relief with important variations in height and complex 
buildings interspersed with cultivated area, as well as a road 
with several bridges over a river flowing along a ravine quite 
pronounced. As can be seen in table 1, which summarizes the 
data characteristics in this area, there are more than 700.000 
points, with a 2.3 points/m? resolution and fitted with a 0.65m 
spacing regular grid. This is a highly complex urban area for the 
ground classification process, which is an ideal data set to test 
the efficiency of the proposed methodology. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
  
  
  
  
  
  
Number of points | 703.016 
Density 2.44 points/m^ | Spacing 0.65 m 
Minimum 485.52 m Maximum 604.48 m 
Mean 523.24 m Median 518.44 m 
1*-Quartile | 505.47 m 3". Quartile 539.97 m 
Std.Dev. 22233m CV 0.042 
  
  
  
  
  
Table 1. Experimental test dataset characteristics. 
The obtained results are compared to a reference classification 
and several automatic classifications performed with 
commercial software. The reference classification has been 
made manually using TerraSolid (TerraScan module) — 
classification CL-R-. Besides, some automatic classifications 
for comparison have been made with the same software 
(TerraScan) using different parameters configurations, 
considering the recommended values from the user's manual 
(classification TS-1, TS-2 and TS-3) (see Table 2). 
  
  
  
  
  
Clasification Angle (°) Distance (m) 
TS-01 6 1.4 
TS-02 6 0.5 
TS-03 10 2.5 
  
  
  
  
Table 2. Several configurations of the automatic classification 
process of TerraScan software. 
The obtained results of these classifications, and the results 
obtained from the proposed methodology application 
(classification CL-0) is shown in the figure 4. 
  
  
  
0% 20% 40% 60% 80% 160% 
#Overall accuracy ®Omissionerror Commission error 
  
  
  
Figure 4. Obtained results in the data classification for the 
different approaches. 
The analysis has been undertaken at punctual level, always 
taking as reference the classification made manually by an 
operator (TS-R), and comparing the obtained results in each 
case. From this analysis the coincidences are established (skill 
points), the “real ground points” classified as “non ground 
points” (omission error) and “real non-ground points” 
misclassified as “ground points” (commission error). 
It can be seen clearly that the results obtained with the proposed 
methodology, although it presents more commission errors, the 
sum of both errors (omission and commission) is lower than 
those obtained using the commercial software. Additionally this 
method (CL-O) obtains more than 78% of success. 
    
   
    
    
      
     
   
   
   
   
  
    
     
    
    
    
    
   
   
    
    
    
  
   
   
   
  
  
  
  
   
    
    
  
   
  
    
Th 
cla 
Th 
pre 
aut 
lar 
dat 
Te 
are 
the 
sez 
prc 
gr 
Th 
reg 
res 
ch: 
Dx 
A1 
ail 
El 
da 
Se 
Fi 
sc 
Jo