Automated Detection of Building Damage Due to Recent Earthquakes using Aerial Television Images
Hajime Mitomi, Fumio Yamazaki and Masashi Matsuoka
Earthquake Disaster Mitigation Research Center, RIKEN
2465-1 Mikiyama, Miki, Hyogo 673-0433, Japan.
Tel: +81-794-83-6632 Fax: +81-794-83-6695
E-mail: mitomi@miki.riken.go.jp
Keywords
Image Processing, Building Damage, Texture Analysis, the 1999 Kocaeli Earthquake, the 1999 Chi-Chi Earthquake
Abstract
The characteristics of severely damaged buildings were examined by image processing of aerial television images taken after the 1999 Kocaeli, Turkey and the 1999 Chi-Chi, Taiwan earthquakes. In image processing, color indices and edge elements were used for the extractions of information on damaged buildings. After pixels indicating damaged buildings were detected on the basis of these characteristics, the texture analysis of the image was attempted. The result of the texture analysis was compared with the results of visual inspection. Using this approach, collapsed buildings were properly identified. The automated damage detection method proposed here can be used efficiently in emergency management shortly after a large-scale natural disaster.
1. Introduction
After the 1995 Hyogoken-Nanbu (Kobe) earthquake, the delay of initial measures by central and local governments was pointed out. It is important to estimate and grasp damage situations during the early stage of recovery activity without depending on information sent from the interiors of the stricken area. Several methods for gathering information on damage from outside of the stricken area are available, such as aerial television imagery, aerial photography and satellite imagery. Aerial television images and photographs, which have higher spatial resolution than satellite images, were visually inspected (Hasegawa et al., 2000a; Ogawa and Yamazaki, 2000). Although the severely damaged buildings could be visually inspected, a significant amount of time was required for visual interpretation. Therefore, Koga et al. (1998) and Hasegawa et al. (2000b) carried out preliminary studies on automated damage detection using image processing. In particular, Hasegawa et al. (2000b) examined the characteristics of aerial images of severely damaged wooden buildings taken by high-definition television (HDTV) cameras operated by the Japan Broadcasting Corporation (NHK) after the Kobe earthquake, and we developed a method of automated damage detection based only on the post event images. The damage distribution extracted using the proposed method agreed relatively well with the ground truth data and the results of visual inspection of the HDTV images. In this paper, we report the results obtained by applying the method proposed by Hasegawa et al. (2000b) to the 1999 Kocaeli, Turkey and the 1999 Chi-Chi, Taiwan earthquakes.
2. Damage Detection Method
In the damage detection method proposed by Hasegawa et al. (2000b), the characteristics of damage to wooden buildings were defined based on hue, saturation, brightness (intensity) and edge elements. Using the threshold values of these parameters, the typical areas were classified into damaged and undamaged pixels. Texture analysis was then carried out on these pixels and the damaged buildings were identified. The steps of the automated detection method (described as the Kobe method) are as follows:
- Some training data are selected from typical damaged and undamaged wooden buildings in a HDTV image.
- The intensities of edge elements are calculated by a general gradient filter with a 3 x 3 pixel window and are allocated to one byte (256) value.
- The pixels with the edge intensity value between 32 and 90 are selected.
- The variances in the edge intensity are evaluated for the area of 7 x 7 pixels and are allocated to one byte value for the center pixel.
- The pixels with the variance between 0 and 30 are selected.
- The relative frequencies of color indices such as hue, saturation and intensity (HSI) are also calculated using the RGB values and are allocated values from 0-360 degrees for hue and one byte values for saturation and intensity.
- The pixel, which contains the range of 92-148 degrees (this color range is from red to yellow) for hue, 0-90 for saturation, and 0-175 for intensity, are selected.
- The local density of the selected pixels (described as Rpx) is calculated by texture analysis. For the texture analysis, 31 x 31 to 63 x 63 pixel windows are selected to be proportional to the image scale, depending on the location of the area in the image.
- The pixel blocks whose density values are smaller than 14%, and larger than or equal to 14% are assigned as belonging to undamaged and collapsed buildings, respectively.
Figure 1: Results of automated damage detection by the Kobe method
Figure 2: Training data used in this study
