IsoMapGen: Framework for early prediction of peak ground acceleration using tripartite feature extraction and gated attention model

Abstract

Time series data associated with seismic activities pose significant challenges in disaster preparedness. These challenges underscore the need for reliable and timely damage assessments, critical for developing effective response strategies. The computation of Peak Ground Acceleration (PGA) is central to these assessments, serving as a crucial element in generating dynamic damage maps essential for managing rescue operations. Traditional approaches usually derive PGA from full-length accelerograms after an event, a process that is often complicated and prone to delays. In this work, Isoseismal Map Generator (IsoMapGen) is an end-to-end deep-learning framework engineered to predict early PGA using the initial few seconds of the primary waveform. This model integrates a novel spatio-temporal learning approach with gated component-wise attention mechanisms to enhance PGA and magnitude predictions for real-time damage mapping. It employs a chained prediction methodology that dynamically updates damage maps in response to incoming seismic data. The waveform, as well as tabular features extracted from the waveform, are passed in the model. The data imbalance in high-magnitude earthquake records of the tabular datasets has been addressed through synthetic data using a Conditional Tabular Generative Adversarial Network (CTGAN). CTGAN’s application in generating synthetic earthquake indicator data is largely unexplored. A detailed comparative analysis of IsoMapGen has been designed against established baseline models, highlighting its strong performance in real-time applications. The models’ efficacy was demonstrated by successfully predicting site-specific PGA from early three seconds of ground motion related to three recent earthquakes of magnitude 7.6, 6.1, and 5.8 $M_{JMA}$, that occurred on January 01, 2024. This represents notable progress in earthquake damage mitigation using early PGA prediction. Furthermore, this work could be utilized for other short-length time series characterization problems.