International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
to carthquake is satisfactory. A high degree of aggreement is
evident between the assessment results and the reference
data. A total of 282 buildings falling within the selected area
of study were assessed. Of these buildings, 205 were
correctly labeled as collapsed and un-collapsed, providing an
overall accuracy of 72.6%. Of the total 79 collapsed
buildings, 63 were correctly detected. The producer’s and the
user’s accuracies were computed, for collapsed buildings, as
79.5% and 50.8% respectively. Compared with the
producer’s accuracy, the user’s accuracy is significantly low.
It appears that sixteen buildings were omitted from the
collapsed category.
For un-collapsed buildings, the user’s accuracy was found to
be remarkably higher than that of collapsed buildings. The
producer’s and user’s accuracies of un-collapsed buildings
were computed as 70.0% and 89.9% respectively. Of the total
203 un-collapsed buildings, 61 were omitted from this
category while 142 were labeled correctly. The assessment
results are also illustrated graphically in figure 6. In the
figure, the green colored buildings represent the un-collapsed
buildings. The red colored buildings represent the collapsed
buildings. The yellow and the blue colored buildings
represent the omisssion error for collapsed and un-collapsed
buildings, respectively.
Reference
Collapsed | Un-collapsed | Total
Collapsed 63 61 124
Un-collapsed 16 142 158
Total 79 203 282
Producer’s (%) | 79.7 70.0
User’s (%) 50.8 89.9
Overall (%) 72.6
Table 1. The error matrix.
The érroneously categorized buildings were further
investigated to find out what might have caused them to
deviate from reference data. It appears that the main reason
for 16 collapsed buildings to be wrongly categorized as un-
collapsed is the spacing between the buildings. The rather
short distance between the buildings caused the line segments
detected through perceptual grouping to match with the
wrong buildings. Therefore, 16 collapsed buildings were
wrongly labeled as un-collapsed. Similarly, 61 un-collapsed
buildings were wrongly categorized as collapsed. There
seems to be two main reasons that cause these false
negatives. The first one was the absence of the edges on the
boundaries of the buildings. It was therefore, impossible to
find a line segment corresponding to the edges of the
buildings. Of the 61 un-collapsed buildings, 38 were found to
be erroneously labeled due to this reason. The second reason
was the vectorization process, through which the line
segments were detected and smoothed. The smoothing
process might have changed the orientations of the line
segments. Therefore, the difference between the orientations
of the line segments and the edges of vector building
polygons might have stayed above the pre-set threshold level.
Therefore, because the rule of orientation does not work in
such a case the un-collapsed buildings are wrongly labeled as
collapsed, therefore.
Figure 6. The results of the assessments.
5. CONCLUSIONS
In this study, an approach was presented for detecting the
collapsed buildings due to earthquake. The proposed
approach is based on the perceptual grouping and utilizes the
relationship between the buildings and the cast shadows. The
results of the analysis reveal that cast shadows can provide
very useful cues for detecting the collapsed buildings. The
results achieved in this study are satisfactory. The overall
accuracy was computed as 72.7%. Of the 79 collapsed
buildings, 63 were detected correctly, providing 79.7%
producer’s accuracy. Compared with the collapsed buildings,
a lower degree of agreement is evident between the
assessment results and the reference data for un-collapsed
buildings. Of the total 203 un-collapsed buildings, 142 were
labeled correctly, providing 70.0% producer’s accuracy.
REFERENCES
Turker, M. and San, B.T., 2003. SPOT HRV data analysis
for detecting earthquake-induced changes in Izmit, Turkey.
International Journal of Remote Sensing, 24(12), pp. 2439-
2450.
Turker, M. and Cetinkaya, B., in press. Automatic detection
of earthquake damaged buildings using DEMs created from
pre- and post-earthquake stereo aerial photographs.
International Journal of Remote Sensing.
Turker, M. and San, B.T., in press. Detection of collapsed
buildings caused by the 1999 Izmit, Turkey earthquake
through digital analysis of post-event aerial photographs.
International Journal of Remote Sensing.
KEY W
ABSTR
The par
enhance
The firs
texture
allowin;
simplifi
higher |
eCognit
classific
The pa]
part of
of the |
extracti
these ac
result
classific
signatui
assump
differen
the oth
The sig
not ha
referenc
monoct
were a
space
classifi
the ob
image «
Segmer
various
1994,
segmen
Approa
Schäpe
eCogni
The pr
relative
The res
try and
sary.
