SCEC Award Number 20208 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Geometric-Mean Back-Projection for Dense-Array Imaging of Seismic Sources
Name Organization
Lei Yang Stanford University Xin Liu Stanford University Gregory Beroza Stanford University
Other Participants
SCEC Priorities 3b, 3a, 3e SCEC Groups Seismology, FARM, EFP
Report Due Date 03/15/2021 Date Report Submitted 05/11/2021
Project Abstract
W developed a deep-learning-based seismic denoising algorithm – UrbanDenoiser – to suppress the strong cultural noise in seismic recordings inherent to urban environments. The algorithm is trained using a waveform data set that combines noise sources from the urban Long Beach dense array and high signal-to-noise ratio earthquake signals extracted from the rural San Jacinto dense array, and is based on the framework of DeepDenoiser. We apply UrbanDenoiser to denoise the Long Beach dense array data and seismograms recorded by isolated stations from regional seismic network and find that seismic noise levels are strongly suppressed relative to seismic signals, so that the seismic signals can be recovered even from noisy seismic data with signal-to-noise ratio (SNR) around one. The seismic detection/location results based on denoised data preserve real earthquake events and exclude large amplitude non-earthquake sources. To explore whether or not widespread seismicity in the upper mantle beneath Los Angeles is real, we perform back-projection imaging/location on the denoised continuous data. We do not find widespread earthquakes below 20 km, but observe seismicity distributed beneath the surface fault trace of Newport-Inglewood fault at 0 – 5 km, which becomes more diffuse at 5 – 15 km, before concentrating again near the fault trace at 15 – 20 km depth. This suggests a fault model featuring shallow creep, intermediate locking, and localized stress concentration at the base of the seismogenic zone.
Intellectual Merit This project revisited a significant finding and demonstrated that the previous results were not reliable. This project also illustrates the potential of deep-learning-based denoising for urban earthquake monitoring.
Broader Impacts The project funded Lei Yang, who was a postdoctoral associate and is now an assistant professor at IGG Chinese Academy of Sciences. During the project she transformed her research research from what was previously purely model-driven, to what is now data-driven.
Exemplary Figure Fig. 2 Application of UrbanDenoiser to the 40-minute seismograms (3:20 – 4:00 UTC, Mar. 29, 2014, vertical component) from the five stations of SCSN (Station CI.BRE, CI.FUL, CI.OLI, CI.RHC2 and CI.WLT): (a1-e1) during the aftershocks of the La Habra earthquake. Raw seismograms; (a2-e2) on left and denoised seismograms on right; (a4-e4). Small panels show raw vs. denoised earthquake waveforms.