Airborne
ous Laser
998) 245-
ing Data.
e LiDAR
"alifornia,
a. ISPRS
ity Press,
ser Swath
jy 1997,
Attitude.
tographs.
urements.
ne 1996.
'SCription
Uwe Bacher
AUTOMATIC EXTRACTION OF TREES IN URBAN AREAS FROM AERIAL IMAGERY
Uwe Bacher‘, Helmut Mayer?
! Tiefenbach GmbH, Munich Branch, Telematic Systems
PapinstraBe 53, D-81249 Munich, Germany
uwe.bacher@ gmx.de
*Institute for Photogrammetry and Cartography
University of the Federal Armed Forces Munich, D-85577 Neubiberg, Germany
Helmut.Mayer @ UniBw-Muenchen.de
http://www.BauV.UniBw-Muenchen.de/institute/inst10
KEY WORDS: Automatic tree extraction, tree model, Hough transform, hysteresis thresholding
ABSTRACT
In this paper we propose an approach for the automatic extraction of leafless deciduous trees from high resolution, i.e.,
ca. 4 cm ground resolution, aerial imagery taken in spring. In analogy to approaches for building extraction, we make
use of the dark shadow of the tree as well as of the fact that the vertical trunk is imaged as a straight line pointing to the
nadir point. Hypotheses for the trunk are found via the Hough transform. Branches are tracked by means of hysteresis
thresholding. With our approach, it is possible to determine the trunk base, height, and width of the tree. Results show
the potential of the approach.
1 INTRODUCTION
The modeling of trees is a great challenge because of their complex structure and their interaction with other trees as well
as with other objects, such as buildings and roads. We focus on single trees and groups of trees or trees in a straight line,
but not on trees in forests. One reason for this limitation is that this kind of trees is of specific interest for cadastre and
three dimensional (3D) urban modeling. Furthermore, there is only few work about the automatic extraction of single
trees or groups of trees. Most work deals with forest, e.g., (Pollock, 1996, Gougeon, 1995, Krummheuer, 1998).
The two most state-of-the-art approaches are described more in-detail. In (Larsen, 1998), trees are modeled by a gen-
eralized rotational ellipsoid and a distribution function for the density of the branches and leaves. This model is used
to generate templates of the trees, taking the geometry of the sensor, the time of data capture, the illumination, as well
as the reflectance of the tree and the ground into account. The templates are varied with respect to sizes and shapes of
the trees and are matched with the imagery, in this case a line scanner with 30 cm ground resolution. The disadvantage
of the approach is that it works well only for imagery containing mostly trees, i.e., forest. Moreover, many templates
are necessary if the variability of trees is high. The approach proposed in (Brandtberg, 1996) is based on color-infrared
imagery with a ground pixel size of about 3 cm. The approach uses information about the structure to classify the tree.
Second order derivatives result into lines which are used to describe the branching structure. From the distribution of the
orientations of the lines, a parallel structure can be distinguished from a radial structure. Herewith it is possible to classify
three types of trees: pine, spruce, and birch. This approach works, however, only in case the regions for the tree crowns
are given.
Photogrammetric data acquisition for geographic information systems (GIS) of urban areas is in most cases based on high
resolution aerial imagery. Nowadays, colored images taken primarily in spring are used. From this follows that deciduous
trees, which make up the majority of the trees in urban areas, are leafless. Moreover, conifers are rarely found in public
areas, and are therefore not of first interest for urban planning and tree information systems.
A model for a leafless deciduous tree, which is represented both directly and as its shadow in the imagery, is proposed in
Section 2. This gives way to the strategy described in Section 3. Section 4 demonstrates the feasibility of the proposed
approach. Finally, conclusions are given.
2 MODEL
Our model of a leafless deciduous tree in an aerial image with high resolution, i.e., 2-5 cm ground pixel size, is presented
as semantic network in Figure 1. The basic ideas for the modeling are similar to those used for building extraction,
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 51