Interpretability and spatial efficacy of a deep-learning-based on-site early warning framework using explainable artificial intelligence and geographically weighted random forests

Earthquakes pose significant risks globally, necessitating effective seismic risk mitigation strategies like earthquake early warning (EEW) systems. However, developing and optimizing such systems requires thoroughly understanding their internal procedures and coverage limitations. This study examines a deep-learning-based on-site EEW framework known as ROSERS (Real-time On-Site Estimation of Response Spectra) proposed by the authors, which constructs response spectra from early recorded ground motion waveforms at a target site. This study has three primary goals: (1) evaluating the effectiveness and applicability of ROSERS to subduction seismic sources; (2) providing a detailed interpretation of the trained deep neural network (DNN) and surrogate latent variables (LVs) implemented in ROSERS; and (3) analyzing the spatial efficacy of the framework to assess the coverage area of on-site EEW stations. ROSERS is retrained and tested on a dataset of around 11,000 unprocessed Japanese subduction ground motions. Goodness-of-fit testing shows that the ROSERS framework achieves good performance on this database, especially given the peculiarities of the subduction seismic environment. The trained DNN and LVs are then interpreted using game theory-based Shapley additive explanations to establish cause-effect relationships. Finally, the study explores the coverage area of ROSERS by training a novel spatial regression model that estimates the LVs using geographically weighted random forest and determining the radius of similarity. The results indicate that on-site predictions can be considered reliable within a 2–9 km radius, varying based on the magnitude and distance from the earthquake source. This information can assist end-users in strategically placing sensors, minimizing blind spots, and reducing errors from regional extrapolation.