[S-SS05] Advancements in Regional Seismic Networks: Operations, Applications, and Development

[SSS05-P01] Improving machine learning-based event detection at sparse regional seismic networks: a foundation for regional seismotectonic analyses.

*Admore Mpuang¹, Takahiko Uchide¹ (1.Geological Survey of Japan, AIST)

Keywords:Machine-learning , Seismic phase picking, Regional seismic networks

Recently, many deep learning methods have been developed and applied for automatic seismic phase picking tasks, enabling construction of comprehensive seismic catalogs needed for understanding and assessing seismic hazards. However, the existing deep-learning models, trained on dense seismic arrays for local event detection, do not perform well when applied to sparse stations at regional distances. This has undermined efforts to better understand regional seismicity and seismotectonics in regions without dense station coverage. We compared the performances of three of the most used machine-learning based phase detection models: PhaseNet (Zhu and Beroza, 2018), EQTransformer (Mousavi et al., 2020) and Generalized Phase Detection (Ross et al., 2018). We applied the models to seismic data from a regional seismic network in Botswana, southern Africa, with epicentral distances up to ~8 degrees. The instrument sampling rates are 20 Hz (typical for regional seismic networks), in contrast to the 100 Hz sampling rate at which the original phase picking models were developed. We used one year data with 4 months before and 8 months after the April 2017 magnitude 6.5 earthquake in the region covered by the seismic network, providing a good opportunity to evaluate the performance of the original phase pickers on the aftershock sequence and improve the available seismic catalog. The individual models performed poorly when applied to the data from Botswana, especially in picking S wave arrivals at longer distances, which resulted in poor event detection. When associating combined phase picks made by all 3 models, we observed minor improvements in event phase arrival time and amplitude residuals estimated by GaMMA phase associator (Zhu et al., 2022), however this approach is computationally inefficient and does not significantly improve event detection. We then retrained the three phase picking models on 20Hz down-sampled versions of the original waveforms used in the development of the models, with no performance degradation in comparison to the original models. This allows us to perform transfer learning of the best performing model in our dataset, using manually labelled phase picks of regional events in the Botswana Seismic network. To minimize regional dependency of the phase picking model, we are collecting additional manually labeled data from various tectonic regions, where available, at epicentral distances of ~3 to 8 degrees and training the phase picking models. The resulting models, with expected improved automatic phase picking at regional distances, will enable the development of comprehensive seismic catalogs in regions such as the East African Rift System, the Himalayan region and parts of the South Pacific Islands, enabling improved seismotectonic studies.
Acknowledgements: This study is supported by MEXT Project for Seismology toward Research Innovation with Data of Earthquake (STAR-E) Grant Number JPJ010217. We used waveform data from the Botswana Seismological Network through the facilities of the IRIS Data Management Center. Event phase data were obtained from the International Seismological Center and contributing agencies: PRE, BUL, LSZ, NAM, and BGSI.