J.M. Rekoske | Geology Reports

USGS2024

Automated, near real-time ground-motion processing at the U.S. Geological Survey

We describe automated ground‐motion processing software named gmprocess that has been developed at the U.S. Geological Survey (USGS) in support of near‐real‐time earthquake hazard products. Because of the open‐source development process, this software has benefitted from the involvement and contributions of a broad community and has been used for a wider range of applications than was initially envisioned. Here, we give an overview and introduction to the software, including how it has leveraged other open‐source libraries. We highlight some key features that gmprocess provides, compare response spectra calculated with the automated processing approach of gmprocess to the response spectra provided by the Next Generation Attenuation projects, and summarize projects that have utilized gmprocess. These use‐cases demonstrate that this software development effort has been successfully leveraged in earthquake research activities both within and outside the USGS.

Seismological Research Letters↗

USGS2021

Automated detection of clipping in broadband earthquake records

Because the amount of available ground‐motion data has increased over the last decades, the need for automated processing algorithms has also increased. One difficulty with automated processing is to screen clipped records. Clipping occurs when the ground‐motion amplitude exceeds the dynamic range of the linear response of the instrument. Clipped records in which the amplitude exceeds the dynamic range are relatively easy to identify visually yet challenging for automated algorithms. In this article, we seek to identify a reliable and fully automated clipping detection algorithm tailored to near‐real‐time earthquake response needs. We consider multiple alternative algorithms, including (1) an algorithm based on the percentage difference in adjacent data points, (2) the standard deviation of the data within a moving window, (3) the shape of the histogram of the recorded amplitudes, (4) the second derivative of the data, and (5) the amplitude of the data. To quantitatively compare these algorithms, we construct development and holdout datasets from earthquakes across a range of geographic regions, tectonic environments, and instrument types. We manually classify each record for the presence of clipping and use the classified records. We then develop an artificial neural network model that combines all the individual algorithms. Testing on the holdout dataset, the standard deviation and histogram approaches are the most accurate individual algorithms, with an overall accuracy of about 93%. The combined artificial neural network method yields an overall accuracy of 95%, and the choice of classification threshold can balance precision and recall.

Seismological Research Letters↗