# Application of Neural Network Automatic Event Detection for Reservoir-Triggered Seismicity Monitoring Networks

## TL;DR

This study shows that combining different seismic detection methods improves the accuracy of monitoring earthquakes triggered by reservoirs.

## Contribution

The novel contribution is proposing a hybrid approach combining manual and automated methods to enhance seismic event detection in reservoir-triggered seismicity.

## Key findings

- Manual and automated seismic detection methods produce disjointed results.
- Combining detection methods increases event detection by up to 30%.
- Phase association is crucial for eliminating false detections.

## Abstract

What are the main findings?
The study found that manual and various automated seismic detection methods yield disjointed sets of events.Thresholds for detecting weak events recorded by a small number of stations result in a significant number of false detections or incorrect picks; however, detection sensitivity was found to be stable as the RTS evolves.

The study found that manual and various automated seismic detection methods yield disjointed sets of events.

Thresholds for detecting weak events recorded by a small number of stations result in a significant number of false detections or incorrect picks; however, detection sensitivity was found to be stable as the RTS evolves.

What are the implication of the main findings?
The key implication of these findings is that no single seismic signal detection method is adequate.Therefore, we propose combining various detection methods to enhance the overall effectiveness and accuracy of seismic event analysis related to reservoir-triggered seismicity.

The key implication of these findings is that no single seismic signal detection method is adequate.

Therefore, we propose combining various detection methods to enhance the overall effectiveness and accuracy of seismic event analysis related to reservoir-triggered seismicity.

This study examines reservoir-triggered seismicity (RTS) in Poland and Vietnam. The current state of individual RTS seismic networks necessitates detecting earthquakes from only a few stations. The number of P waves is often inadequate for phase association and event location, which underscores the importance of identifying S waves. Given that individual RTS cases may consist of only hundreds of events, it is crucial for algorithms to be trained on small datasets or to detect effectively using external, global training data. To evaluate this, we compared the efficiency of a deep learning global detection model, transfer learning to the RTS database, a specialized neural network designed for RTS, and manual detection of seismic signals. Transfer learning efficiency was database dependent. Additional interpretation and parametrization of detection results are assumed. Therefore, the emphasis is on phase detection, rather than phase picking accuracy, and detection sensitivity is more important than its specificity. Phase association plays a vital role in detecting seismic signals, facilitating the elimination of most false picks. As a result, the comparisons of detections were based on parameters related to the location of seismic events. The findings indicate that neither the automatic signal detection methods nor the manual methods alone are sufficient. However, their combination significantly enhances detectability. The final catalogs cover up to 30% more events compared to the previous manual. It fulfills the main aim of applying a neural network detector, which is to increase the number of seismic events in the catalog. It may also be further utilized in the research of the triggering process, such as identifying fluid paths and determining fault geometry.