← Back to Sessions

Reducing the Computational Cost of Seismic Interferometry With Compressed Array Data

Session: Recent Development in Ultra-Dense Seismic Arrays with Nodes and Distributed Acoustic Sensing

Type: Oral

Date: 4/20/2021

Presentation Time: 05:30 PM Pacific

Description: 

Ambient seismic noise interferometry lets geophysicists perform seismic imaging in new locations, avoid the cost of seismic source crews and permits. In exchange we have accepted computational analysis requiring increasingly more data movement as the number of sensors grows. This is problematic in the face of: (1) growing seismic array density due to new technologies (e.g. nodes, DAS, MEMS), (2) the growing ratio of data movement operations to arithmetic operations and (3) more frequent long-term monitoring. We are motivated to improve computational methods for ambient noise interferometry, particularly to reduce data movement required during analysis. We reduced costs of ambient noise dispersion images and double beamforming transforms via scalable algorithms with implicit frequency-domain interferometry (Martin, 2018; Martin, 2021), but imaging methods (e.g. tomography) still require explicit ambient noise interferometry.

We present two methods of crosscorrelating dense seismic data stored in lossy compressed forms without requiring raw data reconstruction. One method uses data stored in low-rank form, reducing the cost from being proportional to the product of the number of time samples with the number of sensors squared, to the sum (Martin, 2019). Low-rank compression does not capture some high-frequency features, so we turn to sparse wavelet compression. We propose a new algorithm to perform crosscorrelations via a sparse set of largest wavelet coefficients, essentially an outer product of wavelet coefficients with crosscorrelations of wavelet basis functions (precomputed, and independent of data). The resulting sparse coefficients can be accumulated prior to reconstruction of average time-domain crosscorrelations across many time windows. Redundancies in some wavelet bases greatly reduce the memory footprint. We introduce new software, show tradeoffs in compression ratio with accuracy in individual time windows as well as averages across many time windows.

Presenting Author: Eileen Martin

Student Presenter: No

Authors