Earthquake early warning (EEW) and its systems offer a viable solution for disaster mitigation by providing alerts seconds to minutes before destructive seismic waves arrive, especially in areas experiencing rapid urbanization and population growth. Moreover, on-site EEW systems can significantly enhance disaster mitigation by estimating critical information about subsequent ground motions from the initial portion of P-waves, before extensive damage occurs. To address the limitations of conventional approaches for predicting peak ground acceleration (PGA) and estimating seismic intensities, recent advances in machine learning (ML) have introduced sophisticated models capable of capturing complex relationships among seismic waves. In this study, multiple ML-based regressions are integrated via ensemble learning, ultimately yielding accurate PGA predictions and intensity estimates for on-site EEW systems. The proposed framework uses specific site parameters and P-wave features as inputs and trains the model through supervised learning, including neural networks and decision trees. The results demonstrate that individual models exhibit diverse characteristics that significantly influence the accuracy of PGA predictions and intensity estimations, making model selection a trade-off among multiple factors. In contrast, ensemble learning can preserve the strengths of these algorithms while effectively mitigating their inherent weaknesses, like overfitting. The evaluation metrics obtained from the ensemble model, compared with those of other models, reveal its unique capabilities even under Monte Carlo cross-validation. To thoroughly assess the effectiveness and operational reliability of the proposed framework, additional datasets are employed to comprehensively verify the superiority of the well-trained model. Consequently, the performance across varying intensities further elucidates the feasibility of deploying the ensemble model in a field implementation. This study also highlights the effectiveness and timeliness of on-site EEW systems in Taiwan, demonstrating that the ML-based approach can provide valuable information for decision-making and emergency response.
Huang et al. (Thu,) studied this question.