Feature selection for multi-parameter machine learning-driven surrogate modeling of real-time seismic intensity prediction

A real-time estimation of seismic intensity provides a critical basis for earthquake early warning (EEW) and emergency response. Most current methods estimate seismic intensity by relying only on one or a few parameters extracted from the P-wave without systematic feature selection, potentially neglecting essential characteristics from the ground motion (GM) records, such as amplitude and frequency-domain features. To address this issue, a multi-parameter feature selection method for machine learning-driven surrogate modeling is proposed to predict seismic intensity based on GM records obtained from the Kyoshine Network (K-Net) from 2010 to 2018. A sorting importance technique is proposed to assess 24 candidate parameters to eliminate redundancy and identify an optimal parameter combination, enhancing both model accuracy and computational efficiency. To validate the accuracy and generalization ability of the proposed method, earthquakes occurring between 2019 and 2021 are assumed to be new events and used as a verification case dataset. The results demonstrate that the proposed method effectively eliminates redundant input parameters without sacrificing accuracy, achieving an accuracy rate above 90%. Additionally, the efficiency of the optimal models in calculating input parameters, model training, and prediction is improved by factors of two, ten, and two, respectively, compared to models trained using all alternative input parameters. Finally, an actual seismic event is selected and three earthquake early warning decision-making threshold is set to further test the models. These results highlight the promising potential of the proposed method for predicting seismic intensity during emergency response in real time.