Deep Learning of On-Site Photos to Detect Collapsed Buildings after Earthquakes

Abstract

To support an efficient and rapid investigation of seismically induced damage to buildings, this study applies a deep learning algorithm to on-site photos taken after earthquakes. The convolutional neural network (CNN) is employed to detect the collapsed wooden houses after the 2016 Kumamoto earthquake. The accuracy of discrimination was improved when the on-site photos of collapsed buildings after the other Japanese earthquakes were included in the image dataset. The dropout layer was also very effective to prevent overfitting. The overall accuracy was 80.8%, and this study is helpful to support damage investigation performed by municipalities soon after an earthquake.