Abstract
In order to understand the damage situation immediately after the occurrence of a disaster and to support disaster response, we developed a method to classify the degree of building damage in three stages with machine-learning using road-running survey images obtained immediately after the Kumamoto earthquake. Machine-learning involves a learning phase and a discrimination phase. As training data, we used images from a camera installed in the travel direction of an automobile, in which the degree of damage was visually categorized. In the learning phase, class separation is carried out by support vector machine (SVM) on a basis of a feature calculated from training patch images for each extracted damage category. In the discrimination phase, input images are provided with raster scan so that the class separation is carried out in units of the patch image. In this manner, learning, discrimination, and parameter tuning are repeated. By doing so, we developed a damage-discrimination model for each patch image and validated the discrimination accuracy using a cross-validation method. Furthermore, we developed a method using an optical flow for preventing double counting of damaged areas in cases where an identical building is captured in multiple photos.
