Development of a Method to Estimate Damage Caused by Heavy Rain Disasters Using Machine Learning

Abstract

Despite increasing frequency and severity of disasters, it has been difficult to put into practical use the prediction of damage costs due to heavy rain disasters. In response to this situation, this study aims to quickly predict the amount of direct damage to farmland and land improvement facilities caused by heavy rain disasters, as an easy method in administrative practice. We have developed a 3-layered ensemble model and a system that can estimate the total amount of damage on a national scale. In this examination, we have addressed issues such as the constraints on the number of data points and lack of precision, and the handling of outstanding values in the actual damage amounts. We conducted predictive calculations for 8 disasters from 2022 to 2023. As a result, we have confirmed the accuracy of the model and the reproducibility of the total damage amount, and we evaluate that it has reached a level that can serve as the basis for administrative operation.