2022 Volume 3 Issue J2 Pages 255-267
For the purpose of immediate damage detection over a wide area after the disaster, we created 26,938 training data, by visual interpretation using the WorldView-3 satellite images acquired immediately after the Kumamoto earthquake. The building damage is classified into 3 grades, and the presence of blue sheet covered building is interpreted. Next, 37,121 patch images were cropped from the satellite image into 128 pixel squares, the patch acquired from 80% of the area on the north side of the satellite image is used as the training data, and the patch acquired from 20% of the south area is used as the test data. Next, we developed a program that automatically extracts the building shape and automatically classify the damage and the presence of blue sheets using U-Net, which is a semantic segmentation method that uses deep learning. As a result of evaluating the model using the training and test data, the IoU of the building shape is about 64%, the average F-mesure of the three damage categories is about 74%, and the F-measure for the blue sheet covering is about 89%. It was confirmed the high damage extraction performance that can be used to grasp the degree and location of damage.
