DEVELOPMENT OF THE DEEP LEARNING BASED DAMAGE DETECTION MODEL FOR BUILDINGS UTILIZING AERIAL PHOTOGRAPHS OF MULTIPLE EARTHQUAKES

Abstract

In order to support disaster response activities, we developed an automatic damage classification model using aerial photographs obtained from several earthquakes in Japan. First, we visually classified all buildings into one of four damage levels, then constructed a training and test data set covering four damage levels. By using this training data set, we were able to develop a CNN-based damage detection model with higher performance than previous models. As a result, an average recall value of 70% was obtained, and we confirmed that it is sufficiently accurate to assess the state of disaster damage to wooden buildings in Japan.