[S-CG63] Reducing risks from earthquakes, tsunamis & volcanoes: new applications of realtime geophysical data

[SCG63-P01] A Time and Time-Frequency Domain R2U-net-based Model for Phase Picking and On-site Earthquake Early Warning

*Ting Yu Hsu¹, Wen-I Liao² (1.National Taiwan University of Science and Technology, 2.National Taipei University of Technology)

Keywords:On-site Earthquake Early Warning System, Short-Time Fourier Transform, R2U-net, Multi-task Model

A dual-domain attentive phase picker and earthquake reporter (DAPPER) is developed for both phase picking and on-site earthquake early warning tasks. The model architecture of DAPPER is based on Recurrent Residual U-Net (R2U-net), incorporating dual input branches for both time-domain and time-frequency domain features and attention mechanisms for multi-task output. This study combines Z-score standardization with 2.5gal scale normalization for the time-domain input. Short-Time Fourier Transform (STFT) with logarithmic normalization is used for the time-frequency domain input. The study explores various window functions combined with different zero-padding positions to understand their impact on the spectrogram. The final parameters selected for the time-frequency domain input are the Boxcar window function of 150 points window length and stride of 1 point with 72 zeros padded at the beginning of the signal and 77 zeros padded at the end of the signal. The DAPPER is compared with related phase picking and earthquake warning models. Finally, a simulation of the alert issuance for the Hualien earthquake on April 3, 2024, was conducted to understand the model's alert performance during the earthquake event.