Chengping Chai1, Jonas Kintner2,
Kenneth M. Cleveland2, Jingyi Luo3,
Monica Maceira1, and Charles J.
Ammon4
1Oak Ridge National Laboratory, Oak Ridge, Tennessee,
USA.
2Los Alamos National Laboratory, Los Alamos, New
Mexico, USA.
3School of Data Science, University of Virginia,
Charlottesville, Virginia, USA.
4Department of Geosciences, Pennsylvania State
University, University Park, Pennsylvania, USA.
Corresponding author: Chengping Chai
(chaic@ornl.gov)
Key Points:
- We applied five machine learning algorithms to a waveform quality
control problem using a labeled dataset of 400,000 surface-wave
samples
- Neural networks and random forests outperformed other algorithms with
a higher accuracy, a faster execution speed, and a smaller storage
- The trained neural network and random forest performed equally to
human analysts but used only 0.5% of time of human analysts
Abstract
Surface-wave seismograms are widely used by researchers to study Earth’s
interior and earthquakes. Reliable results require effective waveform
quality control to reduce artifacts from signal complexity and noise, a
task typically completed by human analysts. We explore automated
approaches to improve the efficiency of waveform quality control
processing by investigating logistic regression, support vector
machines, k-nearest neighbors, random forests (RF), and artificial
neural networks (ANN) algorithms. Trained using nearly 400,000 waveforms
with human-assigned quality labels, the ANN and RF models outperformed
other algorithms with a test accuracy of 92%. We evaluated the trained
models using seismic events from geographic regions not used for
training. The results show the trained models agree with labels from
human analysts, but required only 0.5% time. Although the quality
assignments assessed general waveform signal-to-noise, the ANN or RF
labels can help facilitate detailed waveform analysis, reducing
surface-wave measurement outliers without human intervention.