부산대학교 도서관

In this study, Convolutional Neural Network (CNN) was applied to aerial photographs after the 2011 tsunami for detection of tsunami debris. Five different sample areas were chosen respectively from Miyagi and Iwate prefectures. The sample images were divided into small square images, and the divided images were classified into 5 classes (debris, vegetation, road, building and others) by visual identification as the training data for the network. The trained networks were able to differenciate the 4 main classes (debris, vegetation, road and building) with higher accuracy, while the class “others” had more misidentifications to reduce the total accuracy. Although there are some problems to be improved in the current method for the classification, it was found that CNN is effective in tsunami debris detection from the ortho images.