Automated Detection of Building Damage Due to Recent Earthquakes using Aerial Television Images
3. Application to the Recent Earthquakes
The Kocaeli, Turkey and the Chi-Chi, Taiwan earthquakes occurred on August 17, 1999 and September 21, 1999, respectively (Earthquake Disaster Mitigation Research Center, 2000a,b). The Japanese Geotechnical Society captured images of stricken areas, a few weeks after the Turkey earthquake using a digital video camera from a helicopter. Two and four days after the Taiwan earthquake, NHK captured images of severely damaged areas using a HDTV camera from a helicopter. In this study, we used images taken at Golcuk in Turkey and Chungliao in Taiwan, which had suffered severe damage after the earthquakes. We applied the Kobe method to images taken in Turkey and Taiwan, as shown in Figure 1. The damage areas extracted from the Turkey image was much smaller than the actual distribution of damaged buildings, although more accurate information on building debris was extracted in the Taiwan image. Because some factors attaching the images may be different, such as the influence of sunshine, built environment and so on, whereas damage distribution in the case of the Taiwan image may fortunately be similar to the built environment of the Kobe image used by Hasegawa et al. (2000b). We selected training data and threshold values for each image in order to obtain an extracted damage distribution close to the actual damage situation. Figure 2 and Table 1 represent the selected training data and threshold values for each image of the Turkey and the Taiwan earthquakes. In this study, training data were also selected for moderately damaged buildings, such as D-3 in Turkey, D-3 and D-4 in Taiwan. Damaged pixels were then extracted
Table 1: Threshold values of image characteristics for three earthquakes
| Characteristics
|
Range of threshold values
|
| Kobe
|
Turkey
|
Taiwan
|
| Hue
(degree)
|
92 - 148
|
0 - 360
|
14 - 148
|
| Saturation
|
0 - 90
|
16 - 95
|
25 - 109
|
| Intensity
|
0 - 175
|
63 - 176
|
52 - 170
|
| Edge
intensity
|
32 - 90
|
15 - 103
|
14 - 70
|
| Variance of
edge intensity
|
0 - 30
|
5 - 77
|
1-25
|
Figure 3: Pixels satisfying the threshold values for the selected small areas in Figure 2. The
identified pixels (white area) from each image corresponds mostly to "collapsed" buildings while the shaded gray-color pixels indicate mostly "undamaged" buildings
Table 2: Number of pixels and ratio of extraction in the training data of the two earthquakes
Figure 4: Cumulative relative frequencies versus local density of the selected pixels in a
31 x 31 pixel window for the texture analysis
from the training data in combination with these threshold values for each case, as shown in Figure 3. Table 2 presents the ratios between total and extracted pixels for all the training data. In this table, the differences in the ratios between damaged and undamaged training data are observed, although the ratios for damaged training data are not particularly high.
