International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
figure that the peak of the overall accuracy (80,6396) is reached
when the threshold value is 50%.
Overall
Threshold| Overall Average Combined
Kappa
(%) Accuracy X Accuracy Accuracy
(96) 100 JUser'sProducer'sUser'sProducer's
20% 72,89 | 4,71 |73,93| 51,66 |73,41| 62,28
30% 76,06 | 22,24 |77,40| 58,52 |76,73| 67,29
40% 77,46 | 35,32 |73,09| 65,33 [75,28] 71,40
50% 80,63 | 51,19 |75,93| 75,30 |78,28| 77,97
60% 74,65 | 45,01 |71,41| 75,83 |73,03| 75,24
70% 64,08 | 32,98 |68,15| 72,01 |66,12| 68,05
80% 42,61 | 11,28 [62,98] 59,08 |5S2,80| 50,85
Table 2. The accuracy indices for the threshold values between
20% and 80%.
All the buildings contained within the study area were analyzed
using the optimum threshold of 50%. An error matrix was
generated by comparing the analyzed results with the reference
data. The error matrix contains the overall accuracy, the user’s,
and the producer’s accuracies for collapsed and un-collapsed
buildings (Table 3). The overall accuracy (80,63%) was
computed by dividing the sum of the diagonal of the error matrix
(highlighted in gray) by the total number of the buildings (284).
The producer's for collapsed, the producer's for un-collapsed,
the user's for collapsed, and the user's for un-collapsed
buildings were also computed as 63,29%, 87,31%, 65,79% and
86,06% respectively. It is evident that, 55 buildings were
incorrectly detected. Of these buildings, 29 were not detected as
collapsed through the analysis. Instead, 26 un-collapsed
buildings were detected as collapsed. The mis-detected buildings
represent the omission and commission errors respectively.
Reference
Collapsed | Un-collapsed | Total
Collapsed 50 26 76
Un-collapsed 29 179 208
Total 79 205 284
Producer's 63,29 87,31
User's 65,79 86,06
Overall 80,63
Table 3. Error matrix and accuracies
80.63%
100,00% >
80,00% A
> f |
9 60.00% | {N -
3 N —— accuracy
9 40,00% | i
« X j T. Hi
20,0096
0,0096
9v o\e o\o ov
Q v d$ $
Threshold
Figure 9. The change of the overall accuracy as the threshold
changes.
4. CONCLUSIONS
In this study, we presented an approach for detecting the
earthquake-damaged buildings through shadow analysis of the
watershed segmented post-event aerial imagery. The approach
was implemented in an urban area of the city of Golcuk. A total
of 284 buildings were analyzed to measure their conditions. The
results are quite encouraging. Of the 79 collapsed buildings, 50
were detected correctly providing a producer's accuracy of
63.2994 and a user's accuracy of 65.7994. On the other hand, of
the 205 un-collapsed buildings, 179 were labeled correctly
providing a producer's accuracy of 87.3176 and a user's
accuracy of 86.0696. The overall accuracy was computed as
80.63%.
We found that determining the optimum threshold for
separating the damaged buildings from non-damaged is
important. In the present case the optimum threshold was
computed as 50%. This threshold value is valid for this study
only, and should not be considered global. The proposed
method has several shortcomings to be improved in the future.
The selection of the the initial markers is one problem. The
buffer zones that are defined by the user can be expanded or
shrinked to find a better aggreement between the shadow pixels
(generated by the algorithm) and the actual shadow pixels.
REFERENCES
Beucher, S. and Meyer, F., 1992. The morphological approach
of segmentation: the watershed transformation. In:
Mathematical Morphology in Image Processing, E. Dougherty,
Ed., chapter 12, pp. 433-481. Marcel Dekker, New York.
Gamba, P. and Casciati, F., 1998. GIS and Image Understanding
for Near-Real-Time Earthquake Damage Assessment.
Photogrammetric Engineering and Remote Sensing, 64(10), pp.
987-994.
Hasegawa, H., Aoki, H., Yamazaki, F. and Sekimoto, I., 1999.
Attempt for Automated Detection of Damaged Buildings Using
Aerial HDTV Images. Proceedings of the 20" Asian Conference
on Remote Sensing, Vol. 1, pp. 97-102.
Huertas, A. and Nevatia, R., 1988. Detecting Buildings in Aerial
Images. Computer Vision, Graphics, and Image Processing,
41(2), pp. 131-152.
Irvin, R.B., and McKeown, D.M., 1989. Methods for Exploiting
the Relationship Between Buildings and Their Shadows in
Aerial Imagery. IEEE Transactions On Systems, Man, and
Cybernetics, 19(6), pp. 1564-1575.
Ishii, M., Goto, T., Sugiyama, T., Saji, H. and Abe, K., 2002.
Detection of Earthquake Damaged Areas from Aerial
Photograph by Using Color and Edge Information. Proceedings
of the Fifth Asian Conference on Computer Vision, pp. 27-32.
Mitomi, H., Yamazaki, F. and Matsuoka, M., 2000. Automated
Detection of Building Damage due to Recent Earthquakes
Using Aerial Television Images. 21” Asian Conference on
Remote Sensing, Vol. 1, pp. 401-406.
646
Intern
San, I
Space
Schoo
Geogr
Unive
Shafar
Water
IEEE
Sonka
Analys
pp. 18
Turket
detecti
Intern
2450.
Turkeı
earthq
and p
Journc
Turker
buildir
digital
Journa
Vincer
an effi
Transa
13(6), |