| 2004
—
rmat of
zle was
inimum
n, these
junding
'nerated
s edges
ike into
lext, the
letected
ws. The
This is
joint on
> | share
ected as
S (di, d»,
e corner
ounding
listances
he edges
shadow
distances
; shadow
e. If d, >
ige 1 and
nnecting
anking is
low edge
es d, and
angle of
ing # 175
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
3.2.2 Buffer Zone Generation:
A three pixel wide buffer zone was generated around the
shadow producing edges of each building (figure 6). It was
divided into two sub-zones (1) inside building, and (ii) outside
building. The inside building part of the buffer zone (zone B)
was used for building analysis while the outside building part of
the buffer zone (zone S) was used for shadow analysis. The
purpose of the buffer zone generation was to deal with the
shadow and building areas around the shadow producing edges
of the buildings. These areas can also be called 'the most
significant parts’ of a building for the damage assessment.
Buffer zone,
outside +, Buffer zone,
building area "*, « inside
e
(S) dd Building area
(B)
Figure 6. Buffer zone generation along the shadow producing
edges
3.2.3 Watershed Segmentation:
The watershed segmentation was performed based on the idea
of flooding from selected sources (Beucher ef al., 1992). These
sources represent the markers. Two sets of markers were
needed, one for shadow areas and the other for the building
regions. These markers were utilized to avoid the over-
segmentation. After the gradient image was found, the shadow
and the building markers were selected within the outside
building buffer zone (S) and the inside building buffer zone (B)
respectively. The locations of the markers were seeded
randomly. Figure 7a shows an example for the marker
orientation on a gradient image.
The watershed segmentation algorithm was run to generate a
binary image. After running the watershed algorithm, the two-
region output image was obtained. Of these regions, one refers
to shadow areas while the other corresponds to the building
areas. In figure 7b, the shadow and the building areas are shown
in blue and yellow colors respectively.
Markers
for the
Markers building
for the - area (S) I The region
shadow The region detected
area (S) | detected as
as . BPH *
“shadow” building
(a) (b)
Figure 7. (a) The starting pixels (markers) for watershed
transform, and (b) the segmented output after the watershed
transform.
3.2.4 Assessing the Conditions of the Buildings:
After detecting the shadow and building areas, for each
building, the agreement was measured within the buffer zone of
the shadow producing edges between the pixels labeled as
building and the pixels labeled as shadow (figure 8). To do that
the pixels inside the shadow buffer (S) and the building buffer
(B) were counted and categorized as shadow pixel or building
pixel. Then, a ratio was computed between those pixels labeled
as building and the total number of pixels falling inside the
building of the buffer zone. Similarly, a ratio was also computed
between those pixels labeled shadow and the total number of
pixels falling inside the shadow region of the buffer zone. This
can be illustrated with an example. The pixel distribution of
building # 175 is shown in table 1. For this building, the shadow
detection algorithm detected two shadow edges, which are 1 and
2. The total pixels inside the buffer zone along the shadow
edges were calculated and labeled as “Total Assessed Pixels”
(Table 1). Totally, 99 pixels were generated for shadow buffer
and 99 pixels were generated for building buffer. After the
watershed transform, 91 shadow pixels (blue pixels) fell into the
shadow buffer and 66 building pixels (yellow pixels) fell into
the building buffer. Then, the building and the shadow
percentages were calculated as 66/99 = 66.67% and 91/99 =
91.92% respectively. A user-defined threshold was used to
make a decision about the building. If the building ratio or the
shadow ratio is below the threshold value, then the building is
labeled as collapsed. If on the hand, both the building and the
shadow ratios are over the threshold value then, the building is
labeled as un-collapsed. The building and the shadow ratios
were used together in deciding the building condition in order to
reduce the misdetection of the buildings.
Significant
regions for
building
analysis
Figure 8. The regions used in the building analysis.
Total Assessed Pixels: 99
Detected Shadow Pixels: 91
Detected Building Pixels: 66
Shadow Ratio: 0.9192
Building Ratio: 0.6667
Table 1. Calculation of (building / shadow) pixel percentages.
3.3 Results
Table 2 provides the accuracy indices computed for the
threshold values between 20% and 80%. These are overall
accuracy, overall kappa, average user’s accuracy, average
producer’s accuracy, combined user’s accuracy and combined
producer’s accuracy. Of the six indices, four gave the highest
percentage in the 50% threshold level. The remaining two
indices did not reach to the maximum value at 50%. But, their
percentages were not quite different from the maximum. For this
reason, 50% level was chosen as the optimum threshold. The
trend of the overall accuracies versus varying threshold values is
also shown graphically in figure 9. It can be clearly seen in the
645
RGR iuri
CS TRRE
