Random forest-based multi-hazard loss estimation using hypothetical data at seismic and tsunami monitoring networks

This article presents a novel approach to estimate multi-hazard loss in a post-event situation, resulting from cascading earthquake and tsunami events with machine learning for the first time. The proposed methodology combines the power of random forest (RF) with data that are simulated at seismic and tsunami monitoring locations. The RF model is well-suited for predicting highly nonlinear multi-hazard loss because of its nonparametric regression and ensemble learning capabilities. The study targets the cities of Iwanuma and Onagawa in Tohoku, Japan, where seismic and tsunami monitoring networks have been deployed. To encompass a diverse range of future multi-hazard loss estimation, an RF model is constructed based on 4000 simulated earthquake events with peak ground velocity and tsunami wave amplitude captured at ground-motion monitoring sites and offshore wave monitoring sensors, respectively. The incorporation of 10 ground-motion monitoring sites and five offshore wave monitoring sensors significantly enhances the model's forecasting power, leading to a notable 60% decrease in mean squared error and 20% increase in the R 2 value compared to scenarios where no monitoring sensors are utilized. By harnessing the capabilities of RF and leveraging detailed sensing data, RF achieves R 2 values over 90%, which can contribute to enhanced disaster risk management.