Detection of Collapsed Buidings Due to Eathquake in Urban Areas

As the initial investigation on disaster occurrence, it is very important for reducing economical losses to obtain timely observation of damages especially in metropolis. In such situation, information on changed area obtained from aerial photo analysis is promising. In this study, we propose a change detection approach objected for metropolis right after the disaster with automatic image processing of aerial photos. We introduce two different types of approaches. The first method is for the case of no available orientation information. In this case change detection is performed by registration of images, which are taken before and after the disaster. We define this approach as 2D image matching method. The second approach aims at acquiring not only 2D changes’ distribution but also quantitative 3D shifts by matching between digital terrain data and images before and after the disaster. We call this approach 3D image matching method. In the first step of 2D image matching method, initial registration is executed for minimizing the matching process between images before and after disaster. This process is performed automatically by detecting appropriate conjugate points which candidates are derived from images independently with the improved relaxation method and a mathematical model, which is constrained by a photogrammetric principle (relative orientation). In the next step, image rectification is executed. Owing to these processes yparallax in images is eliminated and matching process is restricted in x-axis. In the case of relative orientation failure, formation of imaginary stereo model by perspective projection is executed as a substitutive means. For detecting changed area, the adaptive nonlinear mapping is applied. This method is based on model of self-organization in neural network. The one of the image pair is mapped to the other by iterative local deformation. Changed areas are detected as inconsistent matching in the mapping process. On the other hand, the concept of 3D image matching method is to compare not only changed images but also the additional terrain information created before disaster. In this case, it is assumed that stereoscopic aerial photos and exterior orientation are available. Former digital terrain data such as DSM or digital maps is transformed based on exterior orientation parameters and 3D matching is carried out. Changed areas are detected as inconsistent texture or height anomaly. Evaluation tests were performed with actual aerial photos that were taken right after of the earthquake and several years later respectively. To evaluate the ability of change detection, the results are compared with those by human interpretation. For quantitative estimation, ROC (Receiver Operating Characteristic) chart was applied, which plots sequential probability of detection against probability of false alarm. As a result, 80 % of right change detection was achieved when false alarms were about 30 % in 2D image matching method and 18 % in 3D image matching method respectively. * Corresponding author.