The National Earthquake Information Center’s Next Steps in Leveraging Machine Learning for Global Earthquake Detection

Abstract:

The U.S. Geological Survey (USGS) National Earthquake Information Center (NEIC) continues to explore how deep-learning tools can be leveraged to improve the NEIC’s current operational systems. As a first step, the NEIC has developed models to classify characteristics of waveforms surrounding standard short-term-average/long-term-average (STA/LTA) picks. These models improve picking accuracy and classify phase type and source-station distance. The development of these tools included specific considerations on how they could be readily incorporated into NEIC’s current real-time seismic event detection framework. The additional information provided by these models is used to refine the NEIC associator’s hypocenter location estimates and to reduce erroneous associations. In this framework we continue to rely on STA/LTA picks as the detection source and therefore do not increase NEIC’s overall detection capabilities. Therefore, NEIC is exploring the use of machine learning models as detectors to directly detect earthquakes from continuous waveforms. We investigate two strategies for generating models useful for detection. First, we explore developing models to detect first arriving phases using three component data from single stations. Second, we explore developing models to detect surface waves of moderate magnitude events lacking near-source observations, for example, earthquakes in the mid-Atlantic. We discuss the considerations needed to run these tools in our real-time detection framework, how these tools may be incorporated into our global earthquake catalog workflow, and the potential pitfalls of relying on these detectors.