Improving the Accuracy of Building Damage Estimation Model Due to Earthquake Using 10 Explanatory Variables

Abstract

Aiming to support disaster recovery, we have developed a new method to extract damaged buildings by using machine learning that combines 10 explanatory variables obtained from analysis of aerial photographs and observation data. We used site amplification factors, seismic intensities of foreshock and mainshock, distance from faults, estimated building structures and ages, coverage by blue tarps, texture analysis, and digital surface model differences before and after the earthquake as explanatory variables, in addition to convolutional neural network prediction results based on post-earthquake aerial photographs. The random forest method resulted in an overall accuracy of about 81% and an average F-measure of three classes was about 70%, indicating that it can classify possible damage to buildings more accurately than our previous studies.