tically creating 
ling features, 
sition process 
extraction and 
sequence with 
ith the features 
investigations, 
  
attan site 
the Manhattan 
d from digital 
ire extraction" 
llow a rooftop 
er of operating 
| Square, static, 
). All of these 
ctual building 
g were created 
( by digitizing 
o the building 
Xtraction was 
traction and 
traction tools 
te. The major 
lilding volume 
een performed 
he elevation of 
1 was used for 
perators’ input 
tprint was not 
ion process but 
curacy. 
ock on the east 
n. Considering 
in posts, it was 
a into eleven 
    
smaller subsites. Figure 2, showing a part of the area 
nicely illustrates the complexity of this very dense 
urban scene. 
  
  
   
Figure 2. Detail image of the Manhattan site 
Based on size and complexity, building structures 
were extracted differently. Blocks of smaller 
buildings attached to each other were measured as 
generalized features, while more complex structures 
were individually extracted. The vertical extension 
of the features was obtained by digitizing at their 
footprint. 
The DTM extraction started with a larger 10m grid 
interval selected for initial processing and the 
required Im by Im grid spacing was obtained by 
densification. Several automated DTM extraction 
strategies were tested on small test areas. The one 
offering the highest accuracy, measured in Figure of 
Merit (FOM) and by visual inspection, was selected 
even though this selection frequently required 
extensive manual editing in areas of limited 
visibility (occlusions) and elimination of rooftop 
elevations. The average speed of automated DTM 
extraction was about 3600 points per minute in the 
strategy we have employed, which is a very 
impressive number. On the other hand, even the 
highest accuracy strategy applied, for automated 
DTM extraction, would yield lower vertical 
accuracy, compared with analytical plotter 
measurements, because of the image matching 
difficulties at this scale of photography. As an 
example, image matching in parking lot areas would 
give us ground elevations as well as elevations on 
top of cars, vans, and trucks. Another example is a 
highway overpass, where occlusions and different 
imagery of moving vehicles from two consecutive 
images cause difficulties. In those cases, the built-in 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
    
   
  
   
  
   
   
  
    
   
   
   
   
   
   
   
   
DPW 770 interpolation methods that we applied 
produced slightly less accuracy, compared with 
direct elevation readout by an experienced operator 
on an analytical plotter. For the reasons stated 
above, once we determined initial DTM elevations, 
a number of editing tools were employed to create 
refined, ground DTM elevations. Figure 3 shows 
contour lines representing the terrain; the areas 
under the already digitized features were 
interpolated for completeness. 
  
Figure 3. Terrain contour lines 
The merged DTM areas combined with the features 
of the Manhattan site resulted in a total of 4,344,171 
elevation posts at the 1m grid spacing. Obviously, it 
would not be feasible to obtain this type of DTM 
information from analytical plotters because of the 
time necessary to manually digitize such a large 
number of ground elevations (Toth and Schenk, 
1990). 
  
Figure 4. Perspective view of a mixed area 
EE YN ST Te STE SE BRR LS i