Parameter optimization of automatic phase detection and picking algorithms - Application in Sao Paulo University Seismological Center and Colombian National Seismic Network.

• Ano de defesa: 2021
• Autor(a) principal: Lopez, Camilo Eduardo Muñoz
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Automatic processing system; Detecção; Detection; Grid-search and Bayesian.; Grid-search e algoritmo Bayesiano.; Optimization; Otimização; Picking; Sistema de processamento automático
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/14/14132/tde-20092021-120029/

The accurate and efficient analysis of seismic data requires the implementation of automatic rocessing algorithms. Therefore, the reliability and quality of these automatic results have become critical requirements for seismological networks. Two methodologies, Grid-search and Bayesian optimization, were used to optimize the automatic detection and phase picking parameters in SeisComP. These methodologies were tested using a set of stations selected from two seismological networks the Brazilian Seismic Network (RSBR) and Colombian National Seismological Network (RSNC). A comparison of manual and previous automatic locations, revealed numerous missing events and others with low-quality locations in automatic databases. We selected 508 manual events from 2017/01/01 to 2020/07/31 in Brazil, and 532 manual events from 2019/02/01 to 2019/02/15 near the Bucaramanga Nest in Colombia, as training data sets for the optimization process. A code was implemented to use an iterative grid-search to optimize the picking parameters. In addition, the Optuna Python package was used to implement the Bayesian optimization. Selected events were used as a training set, and an iterative process according to the Bayesian method was used. The results of both methodologies were compared. Both methodologies improved the system performance by increasing the number of picks and detections. Grid-search allowed us to perform a complete analysis of the results examining the entire space of parameters. However, Grid-search lose efficiency while increasing the number of parameters being optimized. On the other hand, the Bayesian algorithm is computationally more efficient by not exploring the entire parameter space. After the optimization process, automatic picks associated with P phases increases by 78% (76 picks) and 56% (903 picks) for RSBR and RSNC, respectively. Although not all new picks belong to new events, the number of locations calculated using new automatic picks doubled the automatic locations determined by the systems before the optimization process for both networks. Seismological centers could implement methodologies such as Grid-search or Bayesian optimization to improve their automatic processing systems. Besides, the standardization of these methodologies would help to make their implementation easier.