ılly the
1996).
rom its
g. The
-called
nk and
crown
'kness,
s thick
e. The
terrain
: level,
direct
n. The
Ived at
nk is a
Uwe Bacher
To make the extraction robust, we model also disturbing objects. In the real world, they correspond to medium-sized
volumes, e.g., cars, or they are vertical linear objects, such as poles and street-lamps. The latter can be modeled at the
material and geometry level as vertical cylinders, which then appear in the shadow projection at the image level as long
distinct dark lines. They are like the trunks directed in sun direction and can therefore be mixed up with them. The
medium-sized volumes often appear as dark blobs disturbing the short dark lines of the branches and the blurred edge of
the tree crown in shadow projection. To keep the model simple and compact, we decided not to integrate occlusions and
shadows cast by buildings, though they are necessary in principle.
3 STRATEGY
The strategy on top of the model is based on "hypothesize and verify". The focus of attention is laid on objects with
high contrast to make the extraction as productive as possible. We start with distinct dark lines in sun direction which
are hypotheses for trunks. After the trunk is verified by integrating more and more evidence, we continue by tracking the
branches, determination of the outline of the tree, and its classification.
1. Extraction and Verification of Hypotheses for Trunks
First, distinct dark lines representing the shadow projections of the trunks are extracted. Knowledge about existing
objects, e.g., roads or buildings, may help to limit the search space geometrically, speeding up the extraction and
making it more robust. Since the shadow of the trunk is oriented in sun direction, only these lines are selected and
combined to form hypotheses for trunks. The direct projection of the tree is used as evidence to verify a hypothesis
by extracting lines in nadir direction. By intersecting the shadow of the trunk with its direct projection, the position
of the base of a trunk can be determined more accurately and robustly. If the base is occluded, it might be recovered
using only this second kind of information.
2. Tracking of Branches and Determination of the Outline of the Tree
Starting from a hypothesis for a shadow projection of a trunk, the dark lines representing the branches in shadow
projection are extracted. It can be seen as a further verification of the hypothesis for the trunk and it is based on the
knowledge that the branches start at the trunk, branch out and become thiner and thiner. The latter can be used to
separate trees with overlapping shadow projections. What is more, it can also be employed to separate trees from
vertical disturbing objects. For this, the structure of the surrounding branches should be investigated.
By using the end points of the branches together with the fact that the twigs form a blurred edge and the assumption
that the shape of the tree is more or less compact and symmetric, the outline of the tree can be reconstructed using
a snake-based approach (Kass et al., 1987). The diameter of the tree crown can be derived directly from the outline.
The height can be computed from the known sun direction and the shadow outline projected onto a DTM. Based on
the base of the extracted trunk and a DTM, the 3D position of the base of the trunk can be determined.
3. Derivation of the Type of the Tree
The type of the tree is derived by classification based on the shape of the outline, the ratio of the width of the tree
crown and the height of the tree, and the structure of the branches. Because of the correlation between height and age
of the tree, one can determine the approximate age of a tree when its type is known (Baumann, 1957, Spurr, 1960,
Hildebrandt, 1996).
The proposed strategy has been outlined for a single gray-scale image. We do not use color information, because the
shadow has no color and the trunk is only weakly colored. In principle, n 7 2 images might be used to find corresponding
structures of the branches. We have not done this due to the following reasons: Matching the shadow projection results in a
refined DTM only. In suburban areas roads and other areas can in most cases be represented by plane surfaces sufficiently
accurately. By matching the direct projection of the branches in two or more images, the 3D branching structure could be
reconstructed. However, there are not only large problems in line extraction, but above all, the complexity of matching
these structures is extremely high. The reason for this is that because of the complex distribution of the branches in space,
their order can change totally in different images. Contrary to this, the information for our application can be derived
which much less effort from the shadow projection of the tree.
4 INVESTIGATIONS
Figure 2 shows a part of an image showing a suburban area of the German town Tamm near Stuttgart. The original color
image was scanned in black and white at a pixel size of 14 um corresponding to a pixel size of 4.2 cm on the ground at an
image scale of 1:3000. The image shows a suburban area comprising buildings, roads, and trees. The image was taken on
March 23. 1995. at about 11 o'clock in the morning. In early spring the deciduous trees are leafless. The shadows of the
trees are mostly cast on roads and clearly show the structure of the branches. Figure 3a) is a part of Figure 2.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 53