EGU22-5860, updated on 10 Jan 2023
https://doi.org/10.5194/egusphere-egu22-5860
EGU General Assembly 2022
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Seismicity analysis in southern Ghana- I: Detecting local earthquakes by Deep Learning

Hamzeh Mohammadigheymasi1, Nasrin Tavakolizadeh2, S. Mostafa Mousavi3, Graça Silveira4,5, and Rui Fernandes1
Hamzeh Mohammadigheymasi et al.
  • 1Instituto Dom Luiz (IDL) - University of Beira Interior, Covilhã, Portugal
  • 2Institute of Geophysics, University of Tehran
  • 3Department of Geophysics, Stanford University
  • 4Instituto Dom Luiz, Faculdade de Ciências, Universidade de Lisboa, Portugal
  • 5Instituto Superior de Engenharia de Lisboa, Lisbon, Portugal

A large volume of digital seismic data was recorded by six broadband seismic sensors equipped with GPS-clock timing in the Ghana Digital Seismic Network (GHDSN) between October 2012 and April 2014. For this period, no public seismicity catalog was reported by the global data centers, International Seismological Center (ISC), and United States Geological Survey (USGS) for southern Ghana. In this study, this database is processed to detect local earthquakes. To facilitate the challenging and time-consuming process of detecting the earthquakes and picking the arrival times of P and S phases, we utilize EQTransformer, a Deep Learning (DL) model deploying Hierarchical Attention Mechanism (HAM) for simultaneous earthquake detection and phase picking. This model utilizes global and local levels of attention mechanism for identifying earthquake and seismic phases deriving benefits from deep neural networks, including convolutional and recurrent neurons. The thresholding values of 0.2, 0.07, and 0.07 are set for earthquake detection, P-picking, and S-picking, respectively. As a result, a list of events for each station of the network with the associated time of detection, as well as P and S phase arrivals are obtained. Taking these arrival times into account, we have devised a so-called ”conservative strategy” to optimally extract all possible earthquakes in the data set, amenable to locate. Initially, a list of preliminary events recorded by at least two stations is created by comparing the earthquake occurrence and arrival times of the P and S phases for all stations regarding a 100 sec time threshold. The list in this step includes 317 events recorded by at least two stations. Eventually, an analyst controls the obtained waveforms in other stations assesses whether EQTranasformer misses the preliminary list of events in those stations. Consequently, a number of 533 picked phases (282 P and 251 S) recorded by a minimum of 3 stations are finalized. Incorporating these phases and removing the instrument response from the waveforms, the hypocentral parameters for 73 earthquakes with 2.5 ≤ M L ≤ 4.0 are estimated. The main concentration of events is on the intersection of the Akwapim fault zone and the coastal boundary fault, with some scattered seismicity along the Akwapim fault zone. The corresponding set of seismic phases is utilized to estimate an updated 1D crustal velocity model for the study area. This research contributes to the FCT-funded projects SHAZAM (Ref. PTDC/CTA-GEO/31475/2017), RESTLESS (Ref. PTDC/CTA-GEF/6674/2020), SIGHT (Ref. PTDC/CTA-GEF/30264/2017), and IDL (Ref. UIDB/50019/2020).

How to cite: Mohammadigheymasi, H., Tavakolizadeh, N., Mousavi, S. M., Silveira, G., and Fernandes, R.: Seismicity analysis in southern Ghana- I: Detecting local earthquakes by Deep Learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5860, https://doi.org/10.5194/egusphere-egu22-5860, 2022.