Developing Convolutional Neural Networks as Efficient Tools for Earthquake Detection, Localization and Source Characterization - Work in Progress and Key Challenges

Session: Leveraging Advanced Detection, Association and Source Characterization in Network Seismology
Type: Oral
Date: 4/30/2020
Time: 03:45 PM
Room: 120 + 130
Description: 

Convolutional neural networks (CNN) as part of rapidly evolving machine learning techniques bear new opportunities to efficiently process large amounts of seismic data. Various research groups around the globe have shown that neural networks perform well on routine seismological tasks like event detection and phase picking using single stations. In our studies, we follow a network based approach: We use 1-D and 2-D convolutional filters trained on a set of three-component stations in order to exploit arrival time differences and full waveform information of all stations. In our first study, we trained a CNN to predict hypocenter locations of swarm events from West Bohemia, Czech Republic (Kriegerowski et al., 2018). The CNN was trained using more than 2000 events, recorded by nine local broadband stations. The CNN successfully located 908 validation events with standard deviations between 60 m and 136 m in the three dimensions compared to the reference catalog. We show that the first convolutional layer of our locator-CNN becomes sensitive to features that characterize the waveforms and can hence be exploited as an event detector without training on noise samples. Based on these findings we discuss our work in progress: We develop a pre-trained flexible phase picking tool, that will be easily applicable to new datasets. It gains precision after additional training on the targeted dataset. Furthermore, we use similar network geometries in order to train a CNN for source characterization (moment tensor inversion) of small induced earthquakes. As a test case, we plan to apply the new approach to data collected at a geothermal field in the Hengill area, Iceland. We seek to demonstrate that this approach allows rapid inversion and it is viable for micro-seismic monitoring.

Presenting Author: Gesa M. Petersen

Authors