ESTIMATION METHOD OF ROOF-INJURED BUILDINGS FROM AERO PHOTO IMAGES USING DEEP LEARNING IN EARTHQUAKE DISASTER

Abstract

Issuing disaster certificate which is used to decide contents of victim’s support need accuracy and rapidity. However issuing disaster certificate have taken a long time in earthquake disaster so far. So, government need more efficient issuing disaster certificate mechanism. This study developed estimation system of roof-injured buildings to grasp a big picture view of earthquake damage from aero photo images using deep learning. This system is expected to be able to estimate more rapidly by algorithm to make estimated image data automatically using location information of building polygons on GIS. As a result, this system can predict if house is covered with blue sheet or not with about 93% accuracy. And it can predict if house is damaged or not with about 81% accuracy. Although deep learning needs many and high quality images, there are few images having high quality at the present. So we should consider developing image recognition algorithm and how to collect image data to operate the system in actual earthquake disaster.