Keyword: Aerial HDTB Images, Image Processing, Building Damage, 1995 Kobe Earthquake.

Abstract:
The characteristics of aerial high -definition television ( HDTV) images of damaged building due to he 1995 Hyogoken-Nanbu ( Kobe ) earthquake were examined in order to evaluate the possibility of automatic detection of damaged buildings . In this study , the relationships between the degree of building damage and the color information and edge intensity from aerial images were examined by image processing, and were used for the automated detection of damaged buildings. Our automated interpretation results from aerial HDTV images were in fairly good agreement with those of ground surveys, although the extraction accuracy was not particularly high.

1. Introduction
Several methods for gathering information on damage due to natural disasters are available, such as field surveys, aerial television imagery , aerial photography and satellite imagery. Aerial television images by means of which each and every building in a large area can be easily and quickly monitored, may provide effective information during athe early stage of recovery activity. The authors have already reported preliminary studies on the possibility of visually identifying earthquake damage using aerial high-definition television ( HDTV) images ( Hasegawa et al., 1999, 2000) ) . however, it is difficult to interpret the building damage in a large area in a short time. Therefore, the difficult to interpret the building damage in a large in a short time. Therefore, the automated detection of damaged buildings is requisite. Characteristics of severely damaged wooden buildings were examined using image processing of aerial HDTV images taken after the Kobe earthquake. The detection of damaged areas was performed by comparing satellite images taken pre-and post-earthquake ( Matsoka and Yamazak, 1998; Aoki et al., 1998). However, the detection of damaged building s by comparing pre-and post-earthquake aerial images is impractical, because it is difficult to obtain pre-earthquake aerial images. In this study, we only used aerial television images taken after event. Characteristics of damaged building were examined form aerial images for automated detection of the distributions of these buildings. Damaged buildings were identified by extracting the edge elements present only in the aerial HDTV images of collapsed buildings based on color and edge information. The results of automated detection of damaged buildings were compared with the visually interpreted results.

2. Aerial Television Images And The Studied Area
Arial imaging of areas damaged due to the kobe earthquake was started shortly after the event by the Japan . Broadcasting Corporation ( NHK). In this study, we used some of these images taken 10 days after the event. These images were taken at a 30-45 degree angle from the vertical direction, from a height of about 300m by NHK's HDTV cameras. The HDTV images were converted to RGB image data prior to use. The region of this image and this studied are shown in Figure 1 and 2. the spatial resolution of this image is approximately 9 and 17 cm for near and far distance from the camera, respectively.

The areas covered in this study include several blocks ( about 200m x 230m and 46000m2 ) in Nishinomiya City. In this area, many wooden buildings suffered severe damage due to the earthquake.

Figure 1 Aerial HDTV image



Figfue 2 The area of HDTV recording
3. Data of Visual Extraction of Damaged Buildings
The damaged buildings in the studied area were already extracted visually using aerial HDTV images ( Hesegawa et al., 1999), in the visual extraction, the buildings damage level was classified into three categories; collapsed, damaged and nondamaged buildings. Outlines of nondamaged buildings were distinctly observed in the images and the form of buildings was confirmed. As for the collapsed buildings, the soil under the roofing tile and the roofing tile were mixed in the images. Hence, images characteristics of the area representing collapsed buildings were examined using color information and edge elements, in this study.

4. Quantification of Image Characteristics of Damaged Buildings
The area representing collapsed and nondamaged building were selected by visual extraction from the aerial HDTV images. In addition, a ground area was selected. The characteristics of selected buildings are shown in Table 1, and selected pixels from aerial HDTV images are shown in Figure 3. The number of pixels fro each object extracted from the aerial HDTV image was different, because they were selected based on the area of each building had already been covered with blue vinyl canvas sheets as the images were taken 10 days after the earthquake. This characteristic ( blue vinyl canvas sheet ) of damaged buildings was not used in this study.