B3. Istanbul 2004 
's along roads or 
ccur in compact 
om forest scenes 
over close to the 
process for the 
ironments, which 
It should be able 
as far as possible 
ations single tree, 
nother important 
1 environment in 
e not covered by 
an vary from two 
1. 2002)), and in 
jyften stand close 
rm some kind of 
best (or optimal) 
atures. The scale 
In (Schardt et al. 
ction mechanism 
based on the 
formation. In our 
1 higher semantic 
'e, and not before 
). This allows an 
s semantic level, 
. is necessary to 
ts (as well as in 
COACH 
uild a strategy for 
er data. The first 
one could try to 
to reconstruct the 
is to remove the 
o create a surface 
k-Model. In the 
an be found: In 
: the typical fine 
images for the 
| et al. 2002) the 
stochastic process 
e structure of the 
  
  
   
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
The other strategy, the removal of noise, was proposed by 
(Schardt et al. 2002), (Persson et al. 2002), and (Straub 
2003a). The main problem of the second type of approach is 
the determination of an optimal low pass filter - which is of 
crucial importance for the segmentation - for every single 
tree in the image. It is kind of a chicken-and-egg problem: 
the optimal low pass filter depends mainly on the diameter of 
the individual tree one is looking for, which is not known in 
advance. In the case of trees this size can neither be assumed 
to be known nor is it constant for all trees in one image. The 
size of trees depends on the age, the habitat, the species and 
many more parameters, which cannot be modelled in 
advance. 
Process of object extraction from images and/or surface 
models generally depends on an object model as well as a 
strategy for extraction of image features, their combination, 
and their relation to the model. A generic geometric model of 
a tree is used which basically consists of a function 
describing the tree top. Based on this model features are 
identified, which are used to recognise single tree tops from 
the image data. The basic idea for this strategy consists of 
two steps (cf. Figure 4). At first, the often very complex fine 
structures are removed from the surface model by using 
multiple scale levels in linear scale space. As a result of 
scale-space transformation the tree top can be identified in 
the surface model based on the coarse structure. Here, the 
main problem is, that on the one hand the diameter of a 
single tree continuously varies in reality, but also strongly 
influences the choice of filter parameters. To overcome this 
difficulty, the image data was examined at different scale 
levels. 
The basis idea of our approach is to use a multi-scale 
representation of the surface model (assigned as H in 
sigma 
Figure 4) and of the orthoimage (assigned as / in Figure 
sigma 
4) in order to reduce get rid of the fine structures of the tree 
crown, similar to the proposal described in (Persson et al. 
2002). Whereas sigma is the parameter of the Gaussian, 
Which is used to create the multi-scale representation (refer to 
(Lindeberg 1994b) for details on Linear Scale-Space 
transformation). 
  
À ef Ne À 7 ~ 
| Surface Model ) Segmentation ( Segment ] Evaluation Y Tree ) 
sigma ; ; sigma YN 
NS e EN e e Vet net 
eu proe f | — 
ran TT = 
Figure 4: Strategy for the automatic extraction of trees 
A Watershed transformation is used as segmentation 
Every S is a 
sigma 
algorithm, leading to the segments S 
sigma * 
hypothesis for a tree (see Figure 5, for an example). The 
evaluation of the segments is performed according to fuzzy 
membership values. A tree is an object with a defined size, 
circularity, convexity and vitality (NDVI value). 
   
   
   
D». Pe BEN = 
Figure 5: Segmentation results in three different scale levels, 
left fine scale, right coarse scale 
The evaluation phase is divided in two independent steps: 
First, the hypotheses for trees are selected regarding their 
membership values (refer to Figure 6). Than, in the second 
step, the best hypothesis in scale-space is selected. As at one 
and the same spatial position in the scene, more than one 
valid hypothesis can exist, the best one — considering the 
membership value — is selected (refer to the marked segments 
in Figure 6). 
   
Figure 6: Valid hypotheses for trees in different scale levels, 
depicted are the borderlines in different grey 
values. Best hypotheses are marked with a white 
circle. 
  
A detailed description of the approach is given in (Straub 
2003c) and (Straub 2003a). 
5. SUMMARY 
An approach for the automatic extraction of trees from 
remote sensing data - aerial imagery and surface models — 
was shortly depicted in this paper. A detailed description of 
the most important considerations, leading to the 
development of the approach, is given: for the model of an 
individual tree, which is the base of the approach and for the 
strategy for low-level feature extraction and generation of 
hypotheses. 
Recently, the approach was applied on different data sets. 
Results of a performance evaluation of the approach are 
presented in (Straub 2003d) (and, more detailed in (Straub 
2003a)). The test was carried out with one and the same 
parameter settings for all data sets in order to demonstrate its 
robustness and the stability of the underlying model and 
strategy. The Hanover example (c.f. Figure 2 and Figure 7) 
was produced using image and height data from Toposys 
Falcon system, which were acquired by Toposys GmbH in 
summer 2003 by order of the institute of cartography and 
geoinformatics (University of Hannover). An overview of the 
results is given in Figure 7, the automatically extracted trees 
are printed are depicted as white circles: