Yasuno Takato, Amakata Masazumi, Fujii Junichiro and Shimamoto Yuri, Yachiyo Engineering Co., Ltd., Japan

Whenever a natural disaster occurs, it is important to quickly evaluate the damage status in high-priority locations. Frequently, owing to the restrictions imposed by the availability of disaster management resources, spatial information is predicted where the infrastructure manager makes an initial response. It is critical that an initial response be effective to mitigate social losses. In recent years, Japan has experienced several great earthquakes with magnitudes of around 6, most notably the Great East Japan earthquake of March 2011 (M9), as well as those striking Kumamoto (April 2016 (M7)), Osaka (June 2018 (M6.1), and Hokkaido (September 2018 (M6.7)). These huge earthquakes occur not only in Japan but around the world, with an earthquake and tsunami striking Indonesia as recently as October 2018. The initial response to future earthquakes is an important issue related to knowledge of natural disasters and to predict the degree of damage to infrastructure using multi-mode usable data sources. In Japan, approximately 5 million CCTV cameras are installed. The Ministry of Land, Infrastructure and Transportation uses 23,000 of these cameras to monitor the infrastructure in each region. This paper proposes a feature extraction damage classification model using disaster images with five classes of damage after the occurrence of a huge earthquake. We present a support vector damage classifier for which the inputs are the extracted damage features, such as tsunami, bridge collapses, and road damage leading to a risk of accidents, initial smoke and fire, and non-disaster damage. The total number of images is 1,117, which we collected from relevant websites that allow us to download records of huge earthquake damage that has occurred worldwide. Using ten pre-trained architectures, we have extracted the damage features and constructed a support vector classification model with a radial basis function, for which the hyper parameters optimize the results to minimize the loss function value with an accuracy of 97.50%, based on the DenseNet-201. This would provide us with further opportunities for disaster data mining and localized detection.

Disaster Response, Damage Mining, Feature Extraction, Support Vector classifier, Bayesian Optimization