PaperlayerView exact revision on arXiv

Source-linked AI summary

Seismic Arrival-time Picking on Distributed Acoustic Sensing Data using Semi-supervised Learning

Weiqiang Zhu, Ettore Biondi, Jiaxuan Li, Jiuxun Yin, Zachary E. Ross, Zhongwen Zhan

arXiv:2302.08747v2physics.geo-ph

TL;DR

DAS phase picking is difficult because effective algorithms and manual labels are lacking, and its array structure complicates transfer from conventional seismic data. The paper transfers PhaseNet's picking capability through pseudo-labeling and develops PhaseNet-DAS, which the authors report detects and picks phases with few false positives and high temporal accuracy.

  • Problem

    Effective DAS algorithms for detecting earthquakes and picking phases are lacking, and scarce manual labels make accurate phase-picking model development difficult.

    Source: Introduction · p. 2

  • Method

    The approach uses pretrained PhaseNet picks, GaMMA association, and a new 2D PhaseNet-DAS model to train on unlabeled DAS recordings.

    Source: Introduction · p. 3, Neural Network Model · p. 4, Conclusions · p. 12

  • Results

    PhaseNet-DAS detects and picks P- and S-phase arrivals with few false positives and high temporal accuracy.

    Source: Applications to Earthquake Monitoring · p. 7

  • Takeaways & Limitations

    The semi-supervised approach bridges seismic and DAS data modalities to transfer phase-picking capability to DAS recordings.

    Source: Discussion · p. 11

    Training on two California arrays sampled at 100 Hz and around 10 m spacing may limit generalization to other arrays and samplings.

    Source: Discussion · p. 12

Abstract

from arXiv · show

Distributed Acoustic Sensing (DAS) is an emerging technology for earthquake monitoring and subsurface imaging. The recorded seismic signals by DAS have several distinct characteristics, such as unknown coupling effects, strong anthropogenic noise, and ultra-dense spatial sampling. These aspects differ from conventional seismic data recorded by seismic networks, making it challenging to utilize DAS at present for seismic monitoring. New data analysis algorithms are needed to extract useful information from DAS data. Previous studies on conventional seismic data demonstrated that deep learning models could achieve performance close to human analysts in picking seismic phases. However, phase picking on DAS data is still a difficult problem due to the lack of manual labels. Further, the differences in mathematical structure between these two data formats, i.e., ultra-dense DAS arrays and sparse seismic networks, make model fine-tuning or transfer learning difficult to implement on DAS data. In this work, we design a new approach using semi-supervised learning to solve the phase-picking task on DAS arrays. We use a pre-trained PhaseNet model as a teacher network to generate noisy labels of P and S arrivals on DAS data and apply the Gaussian mixture model phase association (GaMMA) method to refine these noisy labels to build training datasets. We develop a new deep learning model, PhaseNet-DAS, to process the 2D spatial-temporal data of DAS arrays and train the model on DAS data. The new deep learning model achieves high picking accuracy and good earthquake detection performance. We then apply the model to process continuous data and build earthquake catalogs directly from DAS recording. Our approach using semi-supervised learning provides a way to build effective deep learning models for DAS, which have the potential to improve earthquake monitoring using large-scale fiber networks.

Introduction

DAS offers dense, potentially low-cost seismic sensing, but its noise and spatial structure challenge conventional processing. The paper addresses the shortage of labeled DAS data with semi-supervised phase picking that transfers knowledge from seismic waveforms.

Introduction

  • DAS can turn existing fiber-optic cables into arrays with meter-scale spacing, offering dense sensing and potential low-cost deployment for earthquake monitoring.The introduction contrasts this with seismic-network spacing of tens of kilometers.Source: Introduction · p. 1
  • DAS phase picking and earthquake detection remain challenging because effective algorithms are lacking, while high noise and dense spatial sampling limit traditional methods.STA/LTA handles a single trace and is vulnerable to anthropogenic noise; template matching requires existing templates and substantial computation.Source: Introduction · p. 2
  • Deep learning has been effective for seismic detection and phase picking, but its success depends on plentiful analyst-labeled arrivals that are largely unavailable for DAS.Building a large manually labeled DAS dataset is costly and time-consuming.Source: Introduction · p. 2
  • Existing low-label approaches have not proven effective for DAS phase picking, in part because dense DAS arrays differ structurally from sparse seismic networks and accurate timing is required.These differences make fine-tuning and transfer learning difficult, while self- and unsupervised methods struggle to achieve accurate picks.Source: Introduction · p. 2
  • The proposed approach uses a pretrained PhaseNet teacher to generate DAS pseudo-labels and trains PhaseNet-DAS to use spatial and temporal information across channels.It tests the method on Long Valley and Ridgecrest arrays and evaluates picking and earthquake detection and location.Source: Introduction · p. 3

Method

The method transfers seismic phase-picking knowledge to DAS through pseudo-labeling and develops a spatial-temporal picker for DAS arrays. Training uses event records from two California regions, with PhaseNet picks associated by GaMMA to form the dataset.

Semi-supervised Learning

  • A pretrained PhaseNet generates independent P- and S-arrival pseudo-labels on DAS channels; GaMMA associates picks and filters false or inconsistent picks using a 1-second window.The window trades pseudo-label quantity against quality: smaller windows yield fewer, higher-quality labels, while larger windows can admit noisier arrival times.Source: Semi-supervised Learning · p. 3
  • Training uses the filtered pseudo-labels, with noise superposition and other augmentations to improve generalization to noisy DAS waveforms and weak signals.The approach could be iterated, but this work uses one iteration and leaves optimization of iteration count for future work.Source: Semi-supervised Learning · p. 4

Neural Network Model

  • PhaseNet-DAS extends a 1D PhaseNet-style U-Net to 2D convolutions so it can use spatial and temporal information in DAS recordings.It uses larger 7×7 kernels and 4×4 stride steps to increase receptive field; the architecture predicts arrivals channel by channel.Source: Neural Network Model · p. 4, Figure 2 · p. 5
  • The model retains a simple U-Net design, while the authors leave exploration of other architectures, including transformers, for future research.The semi-supervised approach does not require the student to share the pretrained model's architecture.Source: Neural Network Model · p. 5

Training Data

  • The training set combines Long Valley and Ridgecrest DAS event records, with PhaseNet picks associated by GaMMA and events retained when they have at least 500 P and 500 S picks.The resulting dataset contains four cable subsets with 542, 410, 284, and 419 events, respectively; lacking manual ground truth, the authors use a 90/10 training-validation split.Source: Training Data · p. 5, Figure 3 · p. 6

Results

PhaseNet-DAS produced accurate P- and S-phase picks and supported earthquake detection from Long Valley DAS recordings. Its picks enabled catalog association, while the model also processed a day of dense DAS data in minutes.

Phase Picking Performance

  • Jointly analyzing channels lets PhaseNet-DAS reduce traffic-related false detections and produce picks that vary less across channels than PhaseNet.Traffic signals are often local over a few kilometers, whereas earthquakes tend to be recorded across the array with body and surface waves.Source: Phase Picking Performance · p. 6, Figure 4 · p. 8
  • Differential-time errors averaged 0.001 s for P waves and 0.005 s for S waves, with standard deviations of 0.06 s and 0.25 s, assuming waveform cross-correlation as ground truth.These differential-time and PhaseNet absolute arrival-time errors are not directly comparable, though their similar scales support effective transfer of picking accuracy to DAS.Source: Phase Picking Performance · p. 7, Figure 5 · p. 9

Applications to Earthquake Monitoring

  • PhaseNet-DAS produced ~36 million P-picks and ~53 million S-picks; GaMMA associated 9,588 events, with 65% within 3 s and 75% within 15 s of catalog origin times.It detected small-magnitude events near the array and most events of magnitude ≥M2 within 100 km; locations became less constrained with distance because a single DAS array has limited azimuthal coverage.Source: Applications to Earthquake Monitoring · p. 7, Figure 6 · p. 10, Figure 7 · p. 11
  • For 24 hours of 200-Hz DAS data with 5,000 channels, PhaseNet-DAS prediction took 14.7 minutes using 8 GPUs.The fully convolutional model can process different input lengths and channel counts within server memory limits; segmenting data enables parallelization across more GPUs.Source: Applications to Earthquake Monitoring · p. 7

Discussion

PhaseNet-DAS transfers phase-picking capability from conventional seismic traces to DAS arrays, while the discussion addresses pseudo-label bias and limits on generalization. The authors identify seismic tomography and other DAS signal tasks as potential applications.

Discussion

  • Semi-supervised learning bridges conventional seismic traces and DAS recordings, transferring phase-picking capability from 1D time series to 2D spatio-temporal measurements.The approach combines DAS's dense spatial resolution with deep learning models developed for conventional seismic waveforms.Source: Discussion · p. 11
  • Pseudo-labels may miss very weak first arrivals or confuse phases; phase association filters inconsistent picks, while strong noise augmentation targets weak arrivals in low-SNR data.The association step uses phase-type consistency across nearby channels and arrival-time moveout based on channel locations and wave velocities.Source: Discussion · pp. 11–12
  • Training used two DAS arrays at Long Valley and Ridgecrest, California, with 100 Hz time sampling and approximately 10 m spatial sampling, which may limit generalization to other arrays and samplings.The authors note the approach can be applied at other locations where historical seismic labels are available.Source: Discussion · p. 12
  • PhaseNet-DAS could measure phase arrivals for seismic tomography, and the semi-supervised approach could be applied to other DAS signals such as tremors where large seismic archives are available.The authors frame these as potential applications.Source: Discussion · p. 12

Conclusions

The conclusion presents a semi-supervised pipeline that uses PhaseNet and GaMMA to create training labels for a DAS-specific model. PhaseNet-DAS detects P and S arrivals with fewer false picks and similar temporal accuracy to pretrained PhaseNet, supporting potential applications in earthquake monitoring and other seismic analyses.

Conclusions

  • PhaseNet-DAS detects P and S arrivals with fewer false picks and similar temporal accuracy compared with pretrained PhaseNet.The model detects arrivals across multiple DAS channels and measures P- and S-arrival times in each channel.Source: Conclusions · p. 12
  • Pretrained PhaseNet generates noisy P- and S-phase picks on single-component DAS traces, and GaMMA associates and selects consistent picks as pseudo-labels for training PhaseNet-DAS.The new model is designed to use temporal and spatial information from DAS arrays.Source: Conclusions · p. 12
  • The approach addresses scarce DAS training labels by using historical seismic data and could support earthquake detection, early warning, seismic tomography, and other DAS applications.The authors also identify developing models for other DAS seismic signals as a potential use.Source: Conclusions · p. 12

Related papers

Loading 2302.08747v2…