ls and erroneously
after elimination of
yy a region growing
nents
letermination of the
lied by height A;
number of voxels
number of voxels
(change in filling)
4
point density per voxel
Figure 4. Number of filled voxels (thick line), change of number of filled voxels (dots) and local trend function for change of filled
voxels (thin line); the minimum of this function is used to define a threshold of point density per voxel for noise reduction
For this purpose, the method described in section 4.2 is
iteratively applied starting from 1 point per voxel for
elimination. With increasing number of points per voxel, noise
is eliminated and, in consequence, a lot of erroneously filled
voxels disappear and a bigger change in filling is obtained. Due
to the fact that real twigs and branches have a certain minimum
number of points per voxel, there is no larger change in filling
by increasing this threshold until reaching this number.
Exceeding this number of points per voxel, real surface voxels
are eliminated and the surfaces are increasingly perforated,
which leads again to a larger change of fillings. Therefore, a
suitable threshold can be defined as minimum of the function
“change in filled voxels” (Figure 4). This method delivers
appropriate results in most of the investigated cases. In only one
exception however, this function shows a continuous decrease
of filled voxels without a minimum of change. In this case the
threshold for elimination of noise voxels is set to a feasible
minimum value, e.g. t=1 (all voxels with only one point are
eliminated).
5. RESULTS AND QUALITY ASSESSMENT
The scans of nine deciduous trees resulted in point clouds of 20
to 60 million points. After applying both methods of noise
Suppression a strong data reduction could be achieved. The
presented approach of volume calculation yielded good results
ranging between -5.1% and +14.3% compared to the control
volumes derived from the weights of the felled trees (Table 1)
With a slight trend to overestimation for trees with dense
Structure of twigs. Nevertheless, certain differences could be
detected. One problem of tall trees in this study is the
metrological reduction of scan density with increasing height —
and thus distance. Small twigs at the top of a tree would require
à particularly higher scan density. Another problem is a
Significant amount of noise caused by different determining
factors. On the one hand, a tree is not a totally static object, i.e.
due to the sequential scanning mode inevitable movements of
branches and twigs lead to noise of the acquired surface point.
On the other hand, noise (and so-called “phantom” points) is
introduced by the large numbers of edges where the laser beams
are split, introducing errors in distance measurements and
reducing the accuracy from millimetres to centimetres. This
effect occurs mainly in trees with small ending twigs and a high
degree of arborisation, and is further exacerbated by disturbing
objects like thorns, dried leaves or fruits. Additionally,
registration accuracy decreases with increasing tree height due
to the limited spatial distribution of the targets (Pfeifer et al.,
2004). Another major influence concerns occlusion and
shadowing of interior branches dependent on the degree of
arborisation (even if scanning from 4 or 5 directions), which, —
especially in this approach, — leads to missing filled voxels.
Excluded
View Point Control | Estimated | Diffe-
Points Density me yo 1%]
0
[ppv]
Al 7 1-17 1-11 -5.1
A2 1.44 1.46 +1.4
A3 0.79 0.81 +25
C1 0.41 0.39 - 4.9
R1 0.85 0.84 -12
R2 3.28 3.75 + 14.3
Bl 1-13 1.20 1:6.2
TI 1:30 1.34 -3.6
SI SS SS AI A
T2 0.70 0.68 +11.4
Table 1. Estimated and control volume of scanned trees
While the stems and thicker branches are usually completely
filled, filling often failed at higher parts of the crown,
independent of the orientation of the branches. The degree of
correct fillings also depends on the appropriate suppression of
