← Back to Sessions

Using Lossy Compression to Speed Up Seismic Event and Ambient Noise Analysis

Description: 

Many of our seismic event detection and ambient noise analysis algorithms are limited by file input/output (I/O). This is becoming a growing issue as we collect more long-term recordings of seismic data from many sensors, such as distributed acoustic sensing and lightweight nodes. Currently, our community data archives struggle to accept such data in their full form, but simple downsampling will lead to spatial or temporal aliasing. We consider lossy compression techniques that maintain an approximation of the broadband information in the data while achieving higher compression rates. A more compressed dataset isn’t just cheaper to store on a data archive, but it also makes remote data quality checks possible during long-term studies and is faster/cheaper to read from the file system of any computer.

We have investigated several common lossy compression strategies for array data: wavelets, Zfp, and low-rank matrix factorizations (e.g. SVD and QR). One appealing quality of these compression techniques is the ability to carry out error propagation considering the representation of the data. In this way, we have calculated bounds on the errors that propagate through some workflows to help inform seismologists and engineers’ compression choices as they collect data and transmit it back to their offices. We have released open-source software that allows users to easily test and compare the errors in template-matching event detection and pick times across compression levels and strategies. In addition to reducing the time to read files, we have created algorithms for template matching and related ambient noise cross-correlations to be carried out in some lossy-compressed domains much more rapidly without decompressing the data. We achieve a particularly large speed-up with cross-correlations of low-rank factorizations. We present results of ambient seismic noise analysis and template matching with continuous distributed acoustic sensing data. We compare the pros and cons in terms of errors, calculation speed and memory footprint.

Session: Building and Decoding High-resolution Earthquake Catalogs With Statistical and Machine-learning Tools [Poster]

Type: Poster

Date: 4/15/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: Eileen

Student Presenter: No

Invited Presentation: 

Poster Number: 37

Authors