International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
resolution and uses the green channel by default for the image
matching. No weighting of the three channel was possible. The
approach takes advantage of a feature-based matching technique
and a robust DTM filtering based on finite elements. It was
originally designed for DTM generation, especially in open
areas. Because of the robust statistics the approach is capable of
eliminating single small objects like trees if the majority of the
matched 3D points is representing the terrain. Also, the original
approach is restricted to a stereo pair and a 2.5 DTM. The latter
is due to the fact that the DTM is a simple raster allowing just
one height value for a planimetric position (Krzystek and Wild
1992). Since it was planned to reconstruct the forest canopy as
good as possible a small grid spacing was envisaged. In order to
get the sufficient redundancy the point density was maximised
by deriving interest points in each epipolar line. The resulting
density of matched 3D points was 18 pts/m? based on a mean
point distance of llcm across track and 66cm along track.
After the robust filtering of the 3D points the cleaned point
cloud showed a mean density of 11 pts/m? indicating an error
percentage of 30%. The grid spacing was set to Im although 0.5
m might have been better. However, the current version of
ISAE did not allow for a smaller grid width. Since the DSM's
still showed slight shifts both in planimetry and height if
compared to the laserscanning DSM's they were locally co-
registered by comparing profiles at single tree crowns. The
remaining discrepancies between the DSM's were in the order
of 10cm.The absolute height accuracy of a single matched 3D
point can be estimated to 11cm by taking into account the side
lap, the image scale and a matching precision of 1/3 pixel.
3.1.2 Laser scanner data: The processing of the
laserscanner data for each strip was performed in house by
TopoSys using a standard procedure called TopPit. The DSM
was filtered from the last pulse data by emphasising the highest
points of the laserscanner data and eliminating outliers. The
DSM contains in this study only height information about
vegetation. The DTM was derived from the last pulse data by
filtering the vegetation points. The original grid spacing of 0.5m
was subsequently resampled to a 1m grid. The georeferencing
of the data was achieved by a standard transformation using the
parameter set DHDN Süddeutschland. Slight deviations from
the local datum were compensated by an adjustment with
polygons of buildings derived from the digital cadastral map
and with control points measured by GPS. Checks at some
control points indicated a positional accuracy of less than 0.5m.
3.2 Data analysis
Profiles were measured manually in the digital stereo
workstation SSK from Z/l-Imaging across each stand to
compare the two types of DSM's. Only true 3D points which
could be clearly identified in both stereo images were measured.
The measured points can be grouped into the following classes:
tree tops, tree crowns, valleys between trees and ground surface.
The manually measured profiles were compared against profiles
which were calculated from the lasercanning DSM and the
photogrammetric DSM by height interpolation at the same
planimetric positions.
A comparison of the tree height captured by the field technique
was only carried out if the tree could be clearly identified in the
stereo workstation. Tree heights derived from laserscanning
DSM, photogrammetric DSM and tree heights measured
manually in the stereo workstation were compared to field
measurements by interpolation of the stem positions. The data
analysis and visualisation was carried out in ARC GIS 8.3.
86
3.3 Automated individual tree delineation
The individual tree recognition was achieved by a special
approach developed at the University of Freiburg. The
procedure is based on a rasterised DSM and DTM. The
algorithms are implemented using the HALCON image
processing software which is a developing system mainly used
in the area of machine vision applications. In the first step a
digital crown model (DCM) is calculated by subtracting the
DTM from the DSM. The DCM is subsequently smoothed by a
gaussian filter. The following main step is a pouring algorithm
for the estimation of tree tops and crown areas. After the
calculation of local maxima an expansion is performed until the
valley bottoms are reached. The areas referring to one maxima
are considered as single trees. In order to enhance the results
additional functions have been introduced to this main concept.
For more details see (Diedershagen et al. 2003).
4. RESULTS
4.1 Accuracy of tree height measurements
From the results shown in table 3 the following conclusions can
be drawn. The laser scanner underestimated the tree heights
about 2m in comparison to the ground measurements. This
effect is quite similar for the deciduous and the coniferous
ground ground ground stereo stereo
[m] laser stereo image laser image
correl. correl.
Plot 22 (n=26)
SD 1.29 1.36 5.37 1.19 5.33
Mean 1.40 0.42 3.78 0.99 3.37
R^ 0.94 0.96 0.37 0.94 0.37
Plot 50 (n=39)
SD 0.02 0.72 1.77 1.00 1.81
Mean 2.08 0.92 2.09 1.16 2.08
R? 0.84 0.89 0.36 0.81 0.30
Plot 60 (n=18)
SD 1.13 1.08 1.08 0.27 0.50
Mean 237 1.58 2.10 0.79 0.51
R? 0.74 0.76 0.73 0.97 0.90
Table 3: Differences of tree heights determined by different
methods in meter. Laser: DSM Laser Scanner.
Image Correl.: DSM derived by image correlation.
Stereo: Stereo measurements. Ground: Tree
measurements on ground.
stand. It is less evident in the old growth stand, since the ground
measurements were performed one growing season earlier. The
standard deviation is about Im in the deciduous and the
coniferous stand and 1.3m in the old growth stand. Also the
stereo measurements underestimated the tree heights. The
amount of this underestimation is about Im if compared to the
laser data. One reason for this is that the DMC flight was
performed on year later. However, this contributes only 0.2 to
0.25cm to the total difference. Consequently, the stereo
measurements are closer to the tree heights measured with field
technique. The standard derivations are almost equal to the
corresponding value of the laserscanner data. The image
correlation underestimated the tree height, too. In relation to the
laser data the underestimation is similar in the deciduous stand,
almost Im larger in the coniferous and even 2.4m larger in the
old growth stand. The standard deviation is almost the same for
Internatioi
EEE es
Figure 1:
the deciduou
and reaches
The relations
in terms of tl
other studies
the height int
height instea:
closest to the
to the stem
show a coni
implies a sig
coefficients 1
3).
4.2 Measu
The mean dif
measurement
the deciduou
shows that th
which seem:
place one ye:
L—— ÁÀSÁÀ————-——
Figure 3: Df
