Evaluation of Building Damage due to Natural Disaster using CNN and GAN

After destructive natural disasters, it is necessary to quickly grasp the damage situation for the initial response. In recent years, studies on the method of the automatic evaluation of building damages due to disasters using the convolutional neural network (CNN), which is a deep learning methodology for image recognition, were conducted. In these studies, it was clarified that a large number of images are necessary to train the CNN with sufficiently high accuracy. However, the number of images of damaged building is limited. Therefore, in the present study, we used the generative adversarial network (GAN) to automatically generate a large number of imitation images of damaged and undamaged buildings and trained the CNN using imitation images to obtain a higher accuracy rate of the CNN. Then, the validity of the CNN for judgment of “damaged” and “undamaged” using imitation images was confirmed. In addition, photographs of actual buildings were input to the trained CNN as test data.