SCEC Award Number 21057 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Detecting Low-frequency Earthquakes with a Deep Convolutional Neural Network
Investigator(s)
Name Organization
Gregory Beroza Stanford University
Other Participants Rosamiel Ries
SCEC Priorities 5a, 5d, 5e SCEC Groups Seismology, CS, FARM
Report Due Date 03/15/2022 Date Report Submitted 12/13/2022
Project Abstract
Tectonic tremor is composed of many repeating low-frequency earthquakes (LFEs, LFE analysis provides valuable insights into slow slip processes. LFEs have seismic amplitudes slightly above background noise, making them difficult to detect. In most locations, they are detected using prior waveform templates. We use a 15-year catalog of more than 1 million LFEs along the San Andreas Fault (Shelly, 2017), to train a convolutional neural network (CNN) for LFE detection. We detect new LFEs with low signal amplitude, even below the noise level, without a prior template, with an accuracy of 95.9%. We cluster the detections by waveform similarity to characterize new templates. Our method is scalable and computationally efficient.
Intellectual Merit The project contributes to understanding tectonic tremor and low frequency earthquakes through automating the process of LFE detection using machine learning.
Broader Impacts LFEs are used to locate tremor precisely and to explore their mechanism. Our approach should be applicable to any environment where increasing the size of LFE catalogs is of interest.
Exemplary Figure Figure 1. Clusters obtained by DBSCAN on the test dataset projected on the principal components of the computed waveform similarity. Each LFE family is plotted in a different color. We note that despite the clustering being performed without prior knowledge of LFE families, the obtained clusters correspond to distinct LFE families. The second and third principal components were selected for visualization purposes. Some clusters visually overlap in this 2-dimensional projection.