The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
287
ADS40 Forest cove
Threshold
Quantize Window green=3x3. blue=axS, red=7x7
Figure 3. Threshold vs. CBO of "Forest cove"
qw=3x3 th=0.25 CBO=92.1
qw=7x7 th=0.80 CBO=30.9
qw=7x7 th=0.85 CBO=26.1
Figure 4. CBO "forest cove"
The evaluation of Figure 3 and 4 reveal the non-linear
behaviour of the parameter space for th. The regression line for
all quantization window sizes shows that higher threshold
values yield lower errors, if the window size is 7x7. With
smaller window sizes and increasing threshold, the error area is
decreasing slower.
The ground truth in Figure 5 contains a typical "straight" forest
boundary and additionally one region as a representative gap in
the forest cover. The smallest CBO-value of 9.8 (threshold=0.95)
from Figure 7 confirms the correct selection by visual
comparison. Missing or additional regions within the forest
increase the CBO more than the comparatively small
differences along the horizontal forest edge. This desired effect
corresponds with the fact, that topological errors should be
weighted more than boundary differences. Forest edges are
often also cluttered with shadow artifacts and therefore the
errors caused by under- and over-segmentations lead to more
robust CBO-values.
qw=7x7 th=0.25 CB067.9
Figure 5. Ground truth "Straight forest edge"
qw=5x5 th=0.85 CBO=64.2