Previous Abstract Return to Session A3 Next Abstract

Session A3: Atmospheric Effects, GNSS Remote Sensing and Scientific Applications

Denoising GNSS Velocities for Earthquake Ground Motions with Deep Learning
Tim Dittmann, EarthScope Consortium; Jade Morton, University of Colorado Boulder; David Mencin, EarthScope Consortium
Location: Seaview Ballroom
Date/Time: Wednesday, Jan. 24, 2:35 p.m.

Peer Reviewed

Distributed measurements of earthquake ground shaking are critical to understanding seismic processes and hazards. High rate, continuous GNSS reference stations are alternative sources of these ground motion observations that complement the dynamic range of traditional inertial-based seismic sensors. GNSS time differenced carrier phase velocity processing (TDCP) has shown to have increased sensitivity to coseismic dynamics when compared to precise point positioning (PPP) while also offering distinct computational advantages. Furthermore, data-driven machine learning (ML) models have proven beneficial for detection and phase picking of seismic waveforms embedded in high-dimensionality noise environments.
In this study, we address this challenging detection problem by modifying existing domain-similar deep learning strategies to take advantage of an established labeled TDCP 5Hz earthquake catalog. We present the results of training a U-net convolutional network architecture used for image processing on denoising temporal windows from a catalog of synthetic 5Hz TDCP waveforms. Unlike traditional fixed bandwidth frequency filtering, the neural network learns a sparse representation mask of time-frequency domain features to separate complex, time-variant noise signatures from a range of signal inputs. These signal inputs from the existing catalog consist of over 2000 observed strong motion velocity waveforms, each augmented with 7 unique noise time series. We use time-frequency domain features from 30 second windows to train a mask on noise-free target waveforms. We report on results of testing denoising unseen synthetic noisy GNSS TDCP time series. We then validate this strategy against both real TDCP waveforms and co-located lower-noise inertial instrumentation.
We find the trained neural network mask filters background noise across a range of time- and spatially-dependent GNSS noise environments. This deep-learned mask permits a spectrum of strong-motion seismic signal signatures to pass through for rapid event characterization and destructive ground motion now-casting. This denoising has the potential to expand the operational role of GNSS in seismic hazard monitoring by increasing signal-to-noise (SNR) of computationally-lightweight TDCP waveforms.



Previous Abstract Return to Session A3 Next Abstract