1192
The International Archives of the Phutogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
For vehicles (Figure 1 a and b), eCognition is more responsive to
parameter settings, judging from the more pronounced curvature
in the distribution of parameter combinations over S. It is also
readily apparent that eCognition produces results closer to an
optimal segmentation near the origin. For larger training objects,
such as buildings (Figure 1. c and d), the opposite is true, with
ASTRO producing a result closer to the origin (smaller D). The
software is less distinguishable from the other training object sets
(trees, Figure 1 e and f; and combined vehicles, buildings and
trees, Figure 1 g and h).
For both ASTRO and eCognition (relative to combined vehicles,
buildings and trees), the parameter combinations with the lowest
D values differ from the combinations with the lowest D when
averaged over training objects. For ASTRO, the scale=50,
color=0.1, smoothness=0.5 combination minimizes D while the
scale=40, color=0.1, smoothness=0.5 minimizes D when it is
computed by averaging over training objects. For eCognition, the
scale=60, color=0.3, smoothness=0.3 combination minimizes D
while the scale=40, color=0.3, smoothness=0.1 minimizes D when
it is computed by averaging over training objects. The
segmentations that result from parameter combinations that
minimize D are shown in Figures 2 and 3 for a subset of the image
we used. Figure 4 shows the training objects in the same subset.
The results are quite obviously qualitatively different, suggesting
that visual interpretation of the segmentation is relevant to the
ultimate selection of a particular parameter combination or
segmentation software.
Figure 2. The ASTRO result that minimizes D: scale=50,
color=0.1, smoothness=0.5.
Figure 3. The eCognition result that minimizes D: scale=60,
color=0.3, smoothness=0.3.
Figure 4. The training shapes corresponding to the area in
Figures 2 and 3.
DISCUSSION
The problem of finding an optimal configuration of parameter
settings has been addressed by Holt et al. (accepted). The index D
can be used for this purpose, though it not been attempted for this
study. The procedure involves fitting a convex function of scale,