Peak ground acceleration prediction and seismic intensity estimation based on ensemble framework for earthquake early warning in Taiwan

Among various natural disasters, earthquakes can produce tremendous life loss and extensive damage to public and private property, particularly in densely populated areas. To facilitate immediate protective actions, earthquake early warning (EEW) was developed to provide a lead time of a few seconds to minutes before impending ground motions at specific sites. Conspicuously, enhancing the prediction of peak ground acceleration (PGA) and the estimations of seismic intensities is crucial for EEW systems to be protective. Numerous methods have been employed to predict PGA and estimate intensities, and we aimed to achieve these objectives through the rapid advancements in machine learning (ML) techniques. Instead of using time histories directly, the site parameters and the P-wave features were treated as inputs and trained using supervised learning approaches, including neural networks and decision trees, to generate accurate predictions, thereby creating a more efficient ML model. Apparently, the performance of these models depends on the training inputs, the ML models/algorithms, and the validation processes; thus, they were compared and discussed in this study. Consequently, the importance of the site parameters was demonstrated, including the high-frequency attenuation rate, the shear wave velocity, and the horizontal depth associated with significant shear wave velocities. In addition to the conventional ML models, an ensemble framework combining neural networks and decision trees was proposed to capitalize on the advantageous and disadvantageous characteristics of the individual models. The effectiveness of the proposed ensemble framework in predicting PGA and estimating seismic intensities was evaluated using performance metrics. The enhancement after integrating the site parameters and the proposed ensemble framework was thoroughly demonstrated via the accuracy, recall, precision, F1 score, and false alarm rate (FAR). This comparison also highlights the reliability of applying the proposed ML model in EEW systems, and the recent advances of ML make them especially suitable for emergency response and decision-making.