Lossy Compression and Reconstruction of Distributed Acoustic Sensing Data Using Deep Learning
Description:
Distributed Acoustic Sensing (DAS) is a new seismic observation method. DAS dramatically expands the ability of dense seismic observation and has been used for ocean observation, sub-surface imaging, volcano monitoring, and earthquake characterization. While the high rate of data benefit research analysis, they are problematic for data transmission and storage and limit the real-time or large-scale application of DAS. Data compression algorithms can accelerate the transmission by transforming the raw data into smaller size , however, at the cost of more computing time and/or data distortion. Current state-of-the-art data compression techniques for DAS involve either a low compression rate (40%) for a lossless compression (Dong et al., 2022) or lossy compressions (Issah and Martin, 2023) that retain the low rank representation of the data.
In this study, we compress and reconstruct an ocean bottom fiber-optic sensing dataset using Deep Learning, particularly Implicit Neural Representation and Shallow Recurrent Decoder. The dataset recorded on two seafloor telecommunication cables in Alaska’s Lower Cook Inlet samples ~16k channels at 250 Hz. On the 2.5 Hz decimated dataset, we reach a 8x (12.5%) and 10x (10%) compression ratio while retaining a maximum mean square error of 0.03 and 0.05, respectively. In the reconstruction, ocean surface gravity waves that are dominantly observed in the raw data are effectively reconstructed by both methods and can be further used for physical oceanography. We test the methods on earthquake waveforms and demonstrate feasibility of the compression for seismological use cases. At last, we evaluate these methods by comparing their accuracy, adaptivity, computing expense and generalizability on the near real-time data transmission using our edge computing experiment in Alaska Cook Inlet.
Session: Advancing Seismology with Distributed Fiber Optic Sensing - III
Type: Oral
Date: 5/3/2024
Presentation Time: 02:45 PM (local time)
Presenting Author: Yiyu
Student Presenter: Yes
Invited Presentation:
Authors
Yiyu Ni Presenting Author Corresponding Author niyiyu@uw.edu University of Washington |
Marine Denolle mdenolle@uw.edu University of Washington |
Bradley Lipovsky bpl7@uw.edu University of Washington |
Qibin Shi qibins@uw.edu University of Washington |
Shaowu Pan shawnpan@uw.edu University of Washington |
J. Nathan Kutz nathankutz@googlemail.com University of Washington |
|
|
|
Lossy Compression and Reconstruction of Distributed Acoustic Sensing Data Using Deep Learning
Category
Advancing Seismology with Distributed Fiber Optic Sensing