rooftop-type. 
Herein, we introduce two different fully automatic techniques to 
generate 3D city models from DSMs. Both methods need a DSM 
and a segmentation result as inputs. The input DSM is gener- 
ated from stereo satellite images by using the semi-global match- 
ing (SGM) approach based on combination of census and mutual 
information. The segmentation result is obtained automatically 
from the DSM by deriving a terrain model (DTM) and fusing 
the derived high object mask with vegetation and water masks 
derived from the pansharpened multispectral ortho images fitting 
on the DSM. 
The first 3D modeling approach consists of three main modules 
as: detection of building shapes, rooftop classification, and 3D 
reconstruction based on the detected shapes and rooftop classes. 
This approach extracts seed-point locations from the segmenta- 
tion result which indicate approximate building regions. A novel 
active rectangular shape growing approach is used on these seed- 
point locations to approximate the building shapes on the DSM. 
Afterwards, using derivative filters on DSM, we extract roof ridge- 
lines automatically. Obtained ridge-line information helps to clas- 
sify building rooftops as flat or gable shapes. Finally, regarding 
the rooftop class we assign rooftop models to our 3D model. 
The second 3D reconstruction approach is based on hierarchi- 
cal approximation and merging the segmented regions providing 
prismatic models of the buildings according to LOD1 of CityGML 
standard. Building footprints are approximated into regular poly- 
gons by reducing the boundary pixels into the most significant 
nodes which preserve the shape and size of the original segment. 
The building shapes are first classified into rectilinear and non- 
rectilinear by measuring the orientation of the edges. For a rect- 
angular building containing one main orientation of the edges a 
method based on Minimum Bounding Rectangle (MBR) in em- 
ployed. In contrast, a Combined Minimum Bounding Rectangle 
(CMBR) approach is proposed for regularization of non-rectilinear 
polygons, i.e. buildings with not perpendicular edge directions. 
Both MBR- and CMBR-based approaches are iteratively employed 
on building segments to reduce the original building footprints to 
a minimum number of nodes with maximum similarity to orig- 
inal shapes. Finally, prismatic models (LOD1) are created by 
computing an average height of the internal building pixels and 
generating the roof and wall polygons. 
2 OBTAINING BUILDING SEGMENTS 
For the extraction of building segments first a spectral classi- 
fication using a rule based fuzzy method and second a deriva- 
tion of a digital terrain model (DTM) from the digital surface 
model (DSM) delivered from the semi global matching algorithm 
(SGM) is derived. 
The rule based fuzzy classification uses the two fuzzy functions 
  
0 ifm <ic 
Fuzzy up defined as fT x — 5: ife<x<d and Fuzzy 
c . 
1 ifm >d 
down f= = 1 — fT together with the logical functions Fuzzy 
a,b a,b 
and (ja, b,...) = min(a,b,...) and Fuzzy orff(a,5,...) « 
max(a,b,...). The rules used for the classification are based 
on the channels of WorldView-2 (coastal, blue, green, yellow, 
red, red-edge, near infrared, second near infrared) denoted as 
(C, B, G, Y, R, E, N, S) and the three ratio functions ndvi = 
Sex. ent = SS and bgi — £x Using these and a fuzzy 
range d = 2 % the classes vegetation (v), soil (s), water (w) 
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 
      
    
   
   
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
     
  
   
  
   
    
   
    
  
     
     
   
     
   
   
    
   
    
   
    
  
Figure 1: Top-Of-Atmosphere (TOA) reflections image showing 
a section 1 km x 1 km from a World View-2 triple-stereo-scene 
acquired 2011-04-13 over Tunis, channels 2, 3 and 5 (blue, green, 
red) 
and shadow (oc) are defined on the previously calculated top-of- 
atmosphere (TOA) reflectances image in [%] (Fig. 1) as: 
v= f* ndvi (1) 
0,0.2 
sn REG fUeey poyenmys Rn) 
— d.d — d.d —d,d —d,d 
(2) 
w =f (is 2, $a -Y, TY Re J* RN, 
pU ETT %) 
“dd 
(3) 
al f' ndvi-cni, f^ B-5%) (4) 
0.078,0.39 “d,o 
Obtaining from these fuzzy results ranging from 0 to 1 the binary 
masks shown in Fig. 2 the limits w > 0.59, v > 0.89, s > 0.59 
and o > 0.25 are applied. 
In the second step a digital terrain model (DTM) is derived from 
the generated digital surface model (DSM). For this the DSM is 
scaled down by an factor of eight filling already small unmatched 
(no-data) regions. Afterwards a morphological grayscale open- 
ing using a structuring element of radius 20 (i.e. 20 x 8 X 0.5 m) 
with a 10 % or 90 % percentile respectively is used filling also 
remaining no-data areas. Applying the final rescaling to original 
size gives the filled DTM as shown in Fig. 3. Subtracting the 
DTM from the DSM provides us with the so called normalized 
digital elevation model (nDEM) which gives the heights of ob- 
jects above ground. A profile showing the connection between 
DSM, DTM and nDEM is shown in Fig. 3, right. The red marked 
objects are the building segments used in the following steps. 
Applying a threshold filter at the appropriate height (here: 3 m 
and a morphological opening followed by a closing (radius 0 
  
  
Figu 
age, 
dry 
shad 
  
In ti 
dim 
tion 
thei 
3.1 
The 
rece 
moc 
buil 
roo