← Back to Sessions

Deepshake: A Machine-Learning Approach to Rapid Estimation of Shaking Intensity

Session: Earthquake Early Warning Live in California! Current Status and Challenges I

Type: Oral

Date: 4/23/2021

Presentation Time: 10:45 AM Pacific

Description: 

The rollout of the ShakeAlert warning system in California has created an impetus for the development of rapid, robust and resilient earthquake early warning algorithms. Early warning systems, provided with a set of real-time ground motion measurements from a network of seismic monitoring stations, aim to predict the intensity of earthquake ground motion at various locations before strong shaking arrives. Traditional methods for earthquake early warning generally consist of two steps -- determining earthquake location and magnitude and calculating ground motion based on ground motion prediction equations (GMPE). These steps are subject to errors due to complex fault rupture patterns and various path and site effects. This makes obtaining accurate ground motion a challenging problem.

We propose a deep spatiotemporal recurrent neural network, DeepShake, to classify future shaking intensity directly from current ground motion observations. DeepShake is a network-based forecasting model, able to predict future shaking intensity at all stations given measured ground shaking from the previous 15 seconds. The model is not given any a priori knowledge of station locations; instead, it learns wave propagation amplitudes and delays solely from training data. We trained DeepShake on 28,543 earthquakes from the 2019 Ridgecrest sequence. Three-component acceleration data is downsampled and smoothed into a single-channel 1Hz feed, diminishing telemetry bandwidth, before being fed into DeepShake. Tasked with alerting for MMI III+ waveforms on 3,568 validation earthquakes at least 5 seconds in advance, DeepShake achieves an equal error rate of 7.9%. For the Mw 7.1 earthquake that hit Ridgecrest on July 5th, 2019, DeepShake was able to provide targeted alerts to stations inside the network between 7 and 13 seconds prior to arrival of MMI III+ waveforms. DeepShake demonstrates that deep spatiotemporal neural networks can effectively provide one-step earthquake early warning with reasonable accuracy and latency.

Presenting Author: Daniel Wu

Student Presenter: Yes

Authors