parallel lines selected parallels 
Figure 14: Selection of parallel lines enclosing areas with contant 
gray values 
The main structure of this class of objects can be extracted 
using one resolution. This is choosen in such a manner that areas 
have homogenous gray values (Förstner, 1995). To distinguish 
(long) buildings from (short) roads a higher resolution or a DEM 
is needed. 
5.2 Linear objects 
Typical representatives are roads, rivers, and railroads (Huertas 
et al., 1987), (Heipke et al., 1995), (Jedynak and Rozé, 1995), 
(Heipke et al., 1994), (Ilg, 1990), (Li et al., 1992), (Lipari et al., 
1989), (McKeown Jr. and Denlinger, 1988), (Venkateswar and 
Chellappa, 1992), (Vosselman and de Knecht, 1995), (Zerubia 
and Merlet, 1993). Roads are similar to the objects above, but the 
borders are curves and the size (length) is not limited. In addition, 
lines (in contrast to edges) are needed for the interpretation. As 
we have seen in section 4 roads are extracted using different levels 
of resolution. The initial resolution is choosen in such a way that 
roads have a width of a few pixels. In the highest resolution road 
marks have a similar width. Because there are different types of 
roads with typical widths the initial resolution has be to selected 
appropriately. 
The type of model for roads is different to that for buildings. 
This is because roads are unbound in principle and do not have 
a fixed shape. Therefore, the interpretation process is mainly 
bottom up and the model is used for tasks like grouping and se- 
lection of areas which are road candidates and not for matching 
of primitves with a model of a road. An example of this group- 
ing process can be seen in figure 15. The first picture shows 
the initial primitives which are grouped (parallel, colinear), se- 
lected (homogenity) and combined with the results of the initial 
resolution. 
Linear objects which are more complicated are brooks or rail- 
roads: Brooks and rivers often have badly defined banks. Rail- 
roads are defined by a combination of lines and a specific texture. 
For these objects strategies used for roads and arbitrary areas have 
to be combined. 
5.3 Arbitrary areas 
Areas like meadows, forests, or fields define another class. They 
have an arbitrary border and are defined by their specific gray 
value, color, and texture. In this case it is very difficult to extract 
any type of primitives. Direct texture analysis is used instead 
(see section 6.3). A first example can be seen in figure 16. The 
left picture shows a zoomed part with two fields with different 
texture. These textures can be distinguished very easily due to 
the horizontal structure of the left field. 
In figure 17 two examples with forest can be seen. The left 
picture shows the selection of areas with the texture of firs. Look- 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
  
different textures classification 
Figure 16: Separation of areas with different textures 
ing at the road at the right side the invariance with respect to 
illumination can be seen. In the right picture deciduous trees are 
selected. 
  
forest deciduous trees 
Figure 17: Selection of regions with with different type of textures 
A model for this class of objects is mainly implicit (e.g., fre- 
quency or color distribution). Only some features, like minimal 
size, width, or typical shapes (fields), can be used. 
5.4 Special Objects 
Objects like trees or persons have to be treated in a very specific 
manner. A tree, for example, has a complex structure which 
involves texture as well as shape. In the case of deciduous trees it 
is very difficult to extract their boundary, especially if they stand 
close together. As we saw in figure 17 it is possible to detect 
the texture of deciduous trees. Figure 18 shows two extreme 
examples: The single trees in the left image can be extracted 
using the gray values and shape. In addition, 3D information, like 
shadows or height, is useful. But there is no procedure known to 
separate the trees in the right image. 
A more simple class of trees are firs. In figure 19 one tree 
and parts of the neighboring trees can be seen. Looking at the 
3D plot of the gray values the complex structure of the branches 
can be seen. One approach to segment these trees is base on the 
following assumptions (Haenel and Eckstein, 1986): 
e The top of the tree is brighter than the outer parts. 
e All trees a separated by dark areas (shadows). 
e The visible (bright) part of the tree has some texture. 
Smoothing the image with a Gauss filter the influence of the 
texture is eliminated and trees can be interpreted as distinct bright 
blobs. The blobs are segmented using the watershed algorithm 
(inverse gray values). The boundaries are now along the darkest 
part of the shadows. To extract the visible part of the trees a local 
threshold operation is used. 
170 
  
     
  
bot 
  
6 
A variety « 
constructic 
on the obje 
6.1 Pixel 
This is the 
to multich: 
of aerial in 
As we will 
with the he 
is the lack: 
resolution 
the classifi 
6.2 Prim 
The most | 
edges. Dif 
filters like 
developed 
1986), (La 
an object