Locating induced earthquakes with a network of seismic stations in Oklahoma via a deep learning method

Xiong Zhang, Jie Zhang, Congcong Yuan, Sen Liu, Zhibo Chen, Weiping LiScientific 2020 · DOI

Abstract

The accurate and automated determination of small earthquake ($M_L < 3.0$) locations is still a challenging endeavor due to low signal-to-noise ratio in data. However, such information is critical for monitoring seismic activity and assessing potential hazards. In particular, earthquakes caused by industrial injection have become a public concern, and regulators need a solid capability for estimating small earthquakes that may trigger the action requirements for operators to follow in real time. In this study, we develop a fully convolutional network and locate earthquakes induced during oil and gas operations in Oklahoma with data from 30 network stations. The network is trained by 1,013 cataloged events ($M_L \geq 3.0$) as base data along with augmented data accounting for smaller events ($3.0 > M_L \geq 0.5$), and the output is a 3D volume of the event location probability in the Earth. The prediction results suggest that the mean epicenter errors of the testing events ($M_L \geq 1.5$) vary from 3.7 to 6.4 km, meeting the need of the traffic light system in Oklahoma, but smaller events ($M_L = 1.0, 0.5$) show errors larger than 11 km. Synthetic tests suggest that the accuracy of ground truth from catalog affects the prediction results. Correct ground truth leads to a mean epicenter error of 2.0 km in predictions, but adding a mean location error of 6.3 km to ground truth causes a mean epicenter error of 4.9 km. The automated system is able to distinguish certain interfered events or events out of the monitoring zone based on the output probability estimate. It requires approximately one hundredth of a second to locate an event without the need for any velocity model or human interference.

Introduction

In regions like Oklahoma, industrial activities—specifically the underground injection of saltwater from oil and gas operations—have led to a significant increase in induced seismicity. To manage these risks, regulators employ a "Traffic Light System" (TLS), requiring operators to modify or cease activities if earthquakes above specific magnitude thresholds (e.g., $M_L \geq 2.0$) occur.

Locating small earthquakes ($M_L < 3.0$) remains difficult because their seismic signatures are often obscured by background noise. Traditional methods rely on manual "phase picking" of wave arrival times or complex geological velocity models, both of which are time-consuming and prone to error for low-magnitude events. This study introduces an automated, real-time system capable of locating these events with high speed and precision.

Methodology

The researchers developed a Fully Convolutional Network (FCN) to treat earthquake localization as an image-to-volume task. This approach maps raw seismic waveforms directly to a 3D spatial location, bypassing intermediate steps like manual phase picking.

Overview of the FCN architecture, station network in Oklahoma, and training performance.

Model Architecture and Input

The FCN processes three-component seismic data (North-South, East-West, and Vertical) from 30 broadband stations. These components are treated as input channels, analogous to RGB channels in computer vision. The model effectively learns the physics of wave propagation implicitly from the training data.

Training and Data Augmentation

The model was trained using 1,013 cataloged earthquakes ($M_L \geq 3.0$) recorded between 2013 and 2016 by the Nanometrics Research Network. To improve performance on smaller events, the researchers used data augmentation: scaling down high-quality $M_L \geq 3.0$ signals to simulate magnitudes between 0.5 and 3.0 while adding realistic background noise.

Results and Discussion

The FCN produces a 3D probability volume where the peak value identifies the most likely hypocenter. This approach is particularly advantageous for noisy, low-magnitude signals where wave arrivals are nearly invisible to the human eye.

3D visualization of the predicted location probability volume.

Performance Metrics

For earthquakes with $M_L \geq 1.5$, the system achieved a mean epicenter error between 3.7 and 6.4 km, satisfying the operational requirements for Oklahoma's TLS. The system is exceptionally fast, locating an event in approximately 0.01 seconds on a standard GPU. However, accuracy degrades for extremely small events ($M_L \approx 0.5$) where noise dominates the signal.

Robustness and Limitations

The model demonstrates robustness to station failures, as it learns spatial patterns across the entire network. A key finding was that the model's accuracy is currently limited by the quality of the human-generated training catalogs; synthetic tests suggest that with "perfect" labels, error could drop to 2.0 km.

Current limitations include the model's dependency on a specific station geometry and the need for a historical catalog for training. Future work may involve using synthetic seismograms to deploy the system in regions lacking historical data.

Conclusion

This deep learning approach provides a scalable, automated solution for real-time seismic monitoring. By eliminating the need for manual intervention and velocity models, the FCN enables rapid response to induced seismicity, supporting safer industrial operations and more effective regulatory oversight.

Summary updated

Abstract

The accurate and automated determination of small earthquake (M<sub>L</sub> < 3.0) locations is still a challenging endeavor due to low signal-to-noise ratio in data. However, such information is critical for monitoring seismic activity and assessing potential hazards. In particular, earthquakes caused by industrial injection have become a public concern, and regulators need a solid capability for estimating small earthquakes that may trigger the action requirements for operators to follow in real time. In this study, we develop a fully convolutional network and locate earthquakes induced during oil and gas operations in Oklahoma with data from 30 network stations. The network is trained by 1,013 cataloged events (M<sub>L</sub> ≥ 3.0) as base data along with augmented data accounting for smaller events (3.0 > M<sub>L</sub> ≥ 0.5), and the output is a 3D volume of the event location probability in the Earth. The prediction results suggest that the mean epicenter errors of the testing events (M<sub>L</sub> ≥ 1.5) vary from 3.7 to 6.4 km, meeting the need of the traffic light system in Oklahoma, but smaller events (M<sub>L</sub> = 1.0, 0.5) show errors larger than 11 km. Synthetic tests suggest that the accuracy of ground truth from catalog affects the prediction results. Correct ground truth leads to a mean epicenter error of 2.0 km in predictions, but adding a mean location error of 6.3 km to ground truth causes a mean epicenter error of 4.9 km. The automated system is able to distinguish certain interfered events or events out of the monitoring zone based on the output probability estimate. It requires approximately one hundredth of a second to locate an event without the need for any velocity model or human interference.
BibTeX
@inproceedings{zhang2020locating,
  author = {Xiong Zhang and Jie Zhang and Congcong Yuan and Sen Liu and Zhibo Chen and Weiping Li},
  title = {{Locating induced earthquakes with a network of seismic stations in Oklahoma via a deep learning method}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-58908-5},
}
Figure