Abstract
Predicting the shaking at a given point due to an assumed earthquake is called “ground motion prediction,” and is used for earthquake hazard assessment, emergency earthquake early warning, and damage estimation after a large earthquake. There have been many studies on the prediction of ground motion indices such as seismic intensity, and empirical equations based on multiple regression analysis and machine learning have been used. On the other hand, numerical simulations of seismic wave propagation have been used to predict time series of seismic waveforms and seismic motion indices, but their computational cost has been an issue. In recent, machine learning has been applied to this prediction problem, and many results have been produced. In this study, we propose a prediction method for time series ground-motion index that combines feature extraction using a graph-based nonlinear dimensionality reduction method (UMAP) and regression using a random forest. The prediction target is real-time seismic intensity, which is an index corresponding to the seismic intensity every second. We believe that the black box nature of machine-learning-based prediction can be reduced by visualizing the features of time series with the feature extraction. We report the results of applying this method to actual records to demonstrate its usefulness.
