SfM-MVS技術を用いた空撮写真による3次元モデルからの地震建物被害の抽出

抄録

 Large-scale natural disasters have caused a great deal of human and property damage in the past. In the 1995 Kobe earthquake, the greatest cause of death was building collapses that occurred in a wide area. Remote sensing technologies are quick and effective solutions for gathering information on significant and widespread building damage. Most of the techniques use a vertical shot, taken from a satellite or an airplane, and the building damage is extracted using image-processing technologies. However, in many cases, it is difficult to accurately analyze the damage to the middle and lower floors of the buildings with the existing structural analysis techniques that use only two-dimensional information such as photographs (taken vertically).

 The SfM-MVS (Structure from Motion and Multi-view Stereo) technique can reproduce three-dimensional structures from the photographs taken at various angles from a variety of platforms. In this study, we applied this technology to aerial photographs of disaster sites due to the 2014 Hiroshima debris flow to validate the altitude accuracy of the developed three-dimensional models. In addition, we examined the relationship between the altitude accuracy and the spatial resolution of the images. Furthermore, we attempted to extract damaged buildings from the three-dimensional model of the disaster site due to the 1995 Kobe earthquake.

 Comparing the three-dimensional models developed by the SfM-MVS for the affected areas of the Hiroshima debris flow and the DSM generated by LiDAR observation, the RMSE (Root Mean Square Error) of the difference value is 1.75 m in the whole observation area, 0.17 m in the mudslide area, and 1.38 m in the buildings. We found that the difference is large in the area with vegetation and around the outer perimeters of the buildings. By reducing the resolution of the aerial photographs, the altitude accuracy of the reconstructed three-dimensional models was evaluated to clarify the relationship between the spatial resolution of the image and the altitude accuracy. It is found that the altitude accuracy of a model in the whole observation area decreases as the spatial resolution of the input photographs is degraded. In the mudslide area, considerable variations were found in the difference values because the abrupt altitude changes caused by the outflow rocks and the rubble could not be coped with effectively.

 This paper also performed the building damage extraction due to the 1995 Kobe earthquake based on the quantitative examination of the Hiroshima dataset. We used six aerial images taken at a vertical angle before the earthquake on May 8, 1994 and 30 photographs taken by a different institution after the earthquake on January 18 and 20, 1995. In addition, to understand the building exterior wall surface, still images were clipped from a video of the affected area, taken at an obliquely downward angle, from a helicopter. For extracting the building damage, for which visual interpretation is somewhat difficult due to altitude changes, the altitude difference of the building was calculated from the models constructed from the images taken with a vertical angle before and after the earthquake. The extraction ratio was about 80% and a high-accuracy extraction of damage was possible in both severe damage and no damage. In addition, even when the texture variation on the roof was small, the altitude change could be replicated well, as damage. Extraction errors were found to be more in the buildings with fewer reconstructed point groups and greater deflections. Finally, we reconstructed post-event three-dimensional model by adding oblique photographs taken from the helicopter in order to verify the altitude changes and damages of building exterior wall.