International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
numbers point at a very complex situation or a very poor 
approach, and significantly higher numbers would raise 
suspicion, that the results of the automatic approach are a 
little bit sugar-coated. This holds also for the automatic 
extraction of trees. Let us expect a success rate of 66% for an 
approach which automatically extracts trees and "plants" 
them into the virtual city, without additional costs for the 
service provider. This is exactly, what we see in Figure 2. 
This is an interesting give-away for class A customers, or 
not? 
In other words: As a potential customer of a 3D city model, 
who can select between the one product with 66% of the 
trees and another one without any trees in it. Both for the 
same amount of costs, which one would you choose? 
3. STATE OF THE ART 
The first trial to utilize an aerial image for forest purposes 
was performed in 1897 (Hildebrandt 1987). Since that time 
the scientific forest community is working on methods for 
the extraction of tree parameters from aerial images. Early 
work was carried out on the manual interpretation of images 
for forest inventory (Schneider 1974), (Lillesand & Kiefer 
1994). The pioneers in the field of the automation of the 
interpretation task "extraction of individual trees from 
images" proposed first approaches about one and a half 
decade ago (Haenel & Eckstein 1986), (Gougeon & Moore 
1988), (Pinz 1989). Recent work in the field was published in 
(Pollock 1996), (Brandtberg & Walter 1998), (Larsen 1999), 
(Andersen et al. 2002), (Persson et al. 2002), (Schardt et al. 
2002). 
A in depth state of the art overview regarding the automatic 
extraction of trees is given in (Straub 2003a). There are 
mainly two common elements in the most approaches: The 
first one is the use of a rotationally symmetric geometric 
model of a tree, as it was proposed by R.J. Pollock in 
(Pollock 1994). A three dimensional surface which simplifies 
the shape of the crown to an ellipsoid of revolution (assigned 
as Pollock-Model in the following). The surface of a real tree 
is of course very noisy in comparison to this simplification. 
This “noise” is not caused by the measurement of the surface, 
it is simply a consequence of the simplification for a very 
complex shape like the real crown of a tree. The idea is, that 
the coarse shape of the crown is well modelled with such a 
surface description. This leads over to the next common 
element of the most approaches, the use of some kind of low 
pass filtering in order to get rid of the "noisy" fine structures. 
Most authors propose to apply — with good reasons -a 
Gaussian function as low pass filter in this early processing 
stage, refer to (Dralle & Rudemo 1996), (Brandtberg & 
Walter 1998), (Schardt et al. 2002), (Straub 2003b), and 
(Persson et al. 2002). 
Some work with focus on the automatic extraction of trees in 
urban areas was also published. In (Haala & Brenner 1999) it 
was proposed to use node points of the region skeletons of 
groups of trees as hypothesis for trees. Morphological 
processing of automatically extracted tree groups is also used 
in (Straub & Heipke 2001) for the computation of tree 
hypotheses. Local maxima of the digital surface model are 
used in (Vosselman 2003) for the detection of trees. The 
proposed solutions are constrained to elongated regions with 
trees (Haala & Brenner 1999), (Straub & Heipke 2001), or 
less complex scenes (Vosselman 2003). But, not all the trees 
in urban environments are standing in rows along roads or 
lines of buildings. In many cases they occur in compact 
arrangements, which are not fare away from forest scenes 
(refer to Figure 3). 
  
Figure 3: An “urban forest” inside of Hannover close to the 
University. 
This was the motivation to develop a process for the 
automatic extraction of trees in urban environments, which 
should fulfil the following pre-conditions: It should be able 
to handle trees in different local context, i.e. as far as possible 
it should be the same algorithm for the situations single tree, 
row of trees, compact group of trees. Another important 
aspect for the extraction of trees in urban environment in 
contrast to a forest is, that smaller trees are not covered by 
the bigger ones. As the diameter of a tree can vary from two 
meter up to fifteen meters (cf. (Gong et al. 2002)), and in 
urban environment, small and big trees often stand close 
together. Therefore it is necessary to perform some kind of 
mechanism for the selection of the locally best (or optimal) 
scale for the extraction of the low level features. The scale 
selection is also a problem in forest areas: In (Schardt et al. 
2002) it was proposed to use the scale selection mechanism 
proposed in (Lindeberg 1994a), which based on the 
maximum response after Scale-Space transformation. In our 
approach the scale selection is applied on a higher semantic 
level, i.e. after the segmentation of the image, and not before 
as it was proposed in (Schardt et al. 2002). This allows an 
internal evaluation of the segments on this semantic level, 
which is particularly then important if it is necessary to 
distinguish between trees and other objects (as well as in 
urban environments). 
4. STRATEGY OF OUR APPROACH 
In principal, there are two possibilities to build a strategy for 
the automatic extraction of trees from raster data. The first 
possibility is to model the crown in detail: one could try to 
detect and group the fine structures in order to reconstruct the 
individual crowns. The second possibility is to remove the 
fine structures from the data with the aim to create a surface 
which has the character of the Pollock-Model. In the 
literature examples for both strategies can be found: In 
(Brandtberg 1999) it was proposed to use the typical fine 
structure of deciduous trees in optical images for the 
detection of individual trees. In (Andersen et al. 2002) the 
fine structure of the crown is modelled as a stochastic process 
with the aim to detect the underlying coarse structure of the 
crown. 
  
Internat 
UE ii 
The oth 
(Schard 
2003a). 
the dete 
crucial 
tree in | 
the opti 
the indi 
advance 
to be kn 
size of t 
many 1 
advance 
Process 
models 
strategy 
and thei 
a tree 
describi 
identifie 
the ima 
two step 
structure 
multiple 
scale-sp 
the surf 
main pr 
single ti 
influenc 
difficult 
levels. 
The ba: 
represen 
Figure 4 
4) in or 
crown, 
2002). ' 
which is 
(Lindeb 
transfori 
( Surface 
Ha 
y^ Ima 
© ls 
Figu 
A Wat 
algorith 
hypothe 
evaluati 
member 
circulari