← Back to Sessions

Toward Implementing Earthquake Early Warning in Resource-Limited Regions: Comparing Magnitudes Predicted by Traditional Regressions and by Convolutional Neural Networks

Description: 

Strong earthquakes present major hazards to millions of people and the built environment worldwide. Earthquake early warning (EEW) systems have the potential to provide strong shaking warnings and thus allow proactive measures to be taken to lower the associated risks. However, current EEW models suffer from tradeoffs between timeliness and prediction accuracy and can generate false alerts. Also, the model development procedures are not transparent enough for duplication by low-tech communities. Thus, additional work is needed to make robust EEW systems that are readily reproducible. As part of a larger project to develop EEW systems from scratch for limited-resource communities exposed to high seismic hazards, this research focuses on magnitude prediction. We evaluate the accuracies of earthquake magnitude predictions from the traditional linear regression approach and from a machine learning model using three-component recordings of global earthquakes in the Stanford Earthquake Dataset. The magnitudes of the earthquakes used ranged from 3.5 to 7.9. The linear regression model we evaluated estimates earthquake magnitudes from peak P-phase displacements. We used Convolutional Neural Networks (CNN), which can capture trends in the data that the traditional linear regression method cannot tap into, for feature selection. The CNN takes 10 s time series, consisting of 3 s of pre-P noise, 4 s of signal starting from the first P-wave arrival, and padded with zeros for 3s to construct windows of uniform length. These features are combined with the peak vertical acceleration and are fed to a series of dense layers for magnitude predictions. Preliminary results show that the current CNN model does not outperform the linear regression approach. However, the CNN model training and optimization is ongoing. At a later phase, we plan to use the extracted features in a Transformer to improve the model’s performance. The results of the final CNN model performance will be presented.

Keywords: earthquake early warning, feature selection, machine learning, neural networks

Session: Earthquake Early Warning Optimization and Efficacy [Poster]

Type: Poster

Date: 4/20/2023

Presentation Time: 08:00 AM (local time)

Presenting Author: Chrispin Gabriel

Student Presenter: Yes

Invited Presentation: 

Authors