Deep Learning Phase Detection Models for Indonesian On-site EEWS Using Strong Motion Accelerograph Network

Indonesia is located in one of the most seismically active regions in the world, making early earthquake warning systems (EEWS) crucial for mitigating disaster impacts. This paper explores developing and implementing deep learning models for phase detection in Indonesia's on-site EEWS using a network of Strong Motion Accelerographs (SMA). Traditional seismic detection methods often face limitations in speed and accuracy when distinguishing between P-waves and S-waves, which are critical for early warnings. This study aims to assess phase detection models utilizing a deep learning method to predict P-Phase arrival times, harnessing Indonesian SMA data from 45,000 seismic occurrences. We used pre-trained from several published phase detection models and performed the models on data collected from Indonesian seismic networks, addressing challenges such as noisy data and varying geological conditions. The findings of this investigation reveal that the appropriate model for Indonesian EEWS is the GPD model, which exhibits a commendable equilibrium in performance. The results of this study indicate that the GPD model achieves an excellent time error value in determining the onset of the P-wave arrival with a very short data length of 2 seconds, yielding an error of 0.02 seconds, followed by PhaseNet with 0.05 seconds and EQTransformer with 0.09 seconds. Results demonstrate significant possibilities in detection speed and accuracy compared to conventional methods, making this approach a promising solution for enhancing the effectiveness of Indonesia's on-site EEWS. Our findings have important implications for disaster preparedness and response in earthquake-prone regions.