LSTM Regression Models for Real-Time Earthquake Source Localization From Single Station

Abstract

Real-time earthquake source localization plays a vital role in Earthquake Early Warning (EEW) systems, yet remains a significant challenge, especially in regions with sparse seismic station coverage. This study explores the potential of Long Short-Term Memory (LSTM) networks for estimating source-to-station distance and hypocentral depth from single-station three-component waveform data. We utilize the Stanford Earthquake Dataset (STEAD) and apply a rigorous preprocessing pipeline to extract a clean subset of over 100,000 labeled seismic event waveforms, ensuring completeness and consistency. We develop and tune three LSTM-based regression models using the Hyperband optimization algorithm. The best-performing model achieves a mean absolute error (MAE) of 26.1 km for distance and 10.1 km for depth. While these results are less accurate than those of baseline models based on Temporal Convolutional Networks and deep CNNs, our approach emphasizes architectural simplicity and operational efficiency. All LSTM models maintain a low number of parameters (as few as 6,098) and exhibit fast inference speeds under 60 ms on a standard GTX 1080 GPU with an Intel i7-7700K CPU. These findings suggest that LSTM-based architectures provide a promising lightweight alternative for rapid deployment in EEW systems, especially in low-resource or single-station scenarios. Future work will explore hybrid neural architectures and attention mechanisms to improve localization performance while maintaining real-time feasibility.