← Back to Sessions

Exploring Urban Distributed Acoustic Sensing Datasets With Scattering Networks

Description: 

Distributed Acoustic Sensing (DAS) probes fiber-optic cables with repeated laser pulses to measure the Earth deformation along the cable. The extended spatio-temporal nature of DAS time series provides a vast amount of information about the seismic wavefield and the related processes occurring at the Earth's surface and interior. Urban DAS datasets have been shown to contain various types of natural (e.g., earthquakes) and anthropogenic (e.g., carry blasts, car traffic) signals, but new signals and/or hidden patterns are still likely to be discovered. We mine continuous DAS data in urban areas in an unsupervised manner using a deep scattering network — a convolutional neural network with convolutional filters restricted to wavelets. Deep scattering networks allow us to directly build an accurate representation of each DAS dataset from which we extract the most relevant information with independent component analysis (ICA) and Uniform Manifold Approximation and Projection (UMAP) algorithms. The ICA representation of the data is then clustered with dendrograms together with UMAP to provide a 2D representation of the data. We use external measurements, such as earthquake catalogs and weather stations to understand the output of the clusters and the UMAP representations. This study shows the potential of unsupervised methods to investigate the extent of the nature of signals contained in terabytes of data recorded by DAS.

Session: Advancing Seismology with Distributed Fiber Optic Sensing - III

Type: Oral

Date: 5/3/2024

Presentation Time: 03:00 PM (local time)

Presenting Author: Loic

Student Presenter: No

Invited Presentation: 

Authors