A MULTISCALE APPROACH FOR THE AUTOMATIC EXTRACTION OF TREE TOPS 
FROM REMOTE SENSING DATA 
Bernd-M. Straub 
Institute of Photogrammetry and GeoInformation, University of Hannover, Germany 
(SOLVing3D, Schönebecker Allee, Garbsen, Germany) 
Commission III, WG 4 
KEY WORDS: Forestry, Urban, Vegetation, Automation, Scale, Model 
ABSTRACT: 
In this paper we describe our work on the automatic extraction of trees from aerial images and digital surface models for the 
refinement of 3D city models with information about trees. Three different aspects of this issue are discussed in the paper. First we 
give a motivation for the application of the automatic approach for tree extraction. An example is given by a visual comparison of a 
3D city model with trees and the same one without trees. One can see — even in the screenshots - that the impression of quality 
becomes much better, if tree models are placed in the scene. In the second section a short overview on recent work regarding the 
extraction of trees is given. Furthermore, the common elements of the published algorithms are pointed out in this section. The 
strategy of our multiscale approach for the automatic extraction of tree tops from remote sensing data is introduced in the third 
section. The paper closes with a summary, some examplary results and an outlook on further work. 
1. INTRODUCTION 
Obviously, trees are important 3D objects in real as well as in 
virtual cities, not only for the orientation and recognition of 
the virtual city, but also for visibility computations which are 
performed for modern landmark based rout descriptions 
(refer to (Brenner & Elias 2003)). But, even if many people 
seems to be interested in vegetation for 3D city models (cf. 
(Fuchs et al. 1998)) relatively few work was published on the 
automation of the extraction and integration of trees. 
From the viewpoint of data capturing there are some 
obstacles for the extraction and integration of trees: The first 
obstacle comes from the conflict, that the best time for image 
acquisition for the extraction of vegetation is during the 
vegetation period. Which is again not optimal (or necessary) 
for the measurement of buildings and roads. The only 
argument against is, that one could extract buildings and 
roads from images which were recorded during the 
vegetation period, even if there are some problems due to 
exclusions and things like this. But the extraction of 
vegetation parameters is often not possible from images, 
which are captured outside the vegetation period. 
The second drawback seems to be a little bit outdated, but it 
should be mentioned here: For the extraction of vegetation it 
is very helpful to have the optical information in the infrared 
band. Formerly — in the period of analogue aerial cameras - 
one had to decide between colour infrared (CIR) film and 
normal colour film. And often it was decided to capture real 
colour images, because the most users are more familiar with 
this type of imagery for interpretation and due to the 
advantages for visualisation. Nowadays, up-to-date digital 
cameras allow us to capture both types of images in one 
flight. The CIR images can be used for the extraction of 
vegetation parameters and the normal colour ones for the 
production of orthoimages for visualization. And as a result, 
this drawback is no more really relevant. 
The third practical obstacle for the development of automatic 
approaches is perhaps the most important one. A tree can be 
measured interactively by a human interpreter with only one 
or two mouse clicks. The first one defines the tree top and the 
second one defines the radius of the tree. That means, even in 
the case that an automat would be able to detect and measure 
really every tree in the scene, only these two clicks per object 
can be saved. And in fact, that is not really much of one 
compares it with the effort for a single building or a road. 
In the next short section we will give a motivation that it can 
make sense to spent some of effort for saving these two 
clicks. The third section of this paper gives an overview on 
recent work in the domain of “tree counting” and leads over 
to some relevant differences between the extraction of 
individual trees in forest and in urban areas. A short 
description of the multi-scale approach for the automatic 
extraction of tree tops is given afterwards. In the last section 
the results of a performance evaluation are mentioned, an 
overview is given about the scene, which is presented at the 
beginning of the paper. Finally some further work is 
proposed. 
2. MOTIVATION 
Obviously, it would be very helpful if an automatic algorithm 
would be able to detect reliable more than 99.5% percent of 
the visible trees in a given scene. A service provider having 
this algorithm would be able to provide his client with an 
additional nice-to-have doodad without additional costs on 
both sides. Automatically extracted frees as a marketing 
instrument, why not? 
    
  
Internal 
Figure 
In this 
automa 
to hav 
trees. 
The fig 
of the 
Figure 
buildin 
of the | 
inJECT 
an ortl 
impres: 
manua. 
in Fig