DASFormer: self-supervised pretraining for earthquake monitoring

Area of Science:

• Geophysics
• Seismology
• Earthquake Science

Background:

• Earthquake monitoring is crucial for understanding earthquake physics and mitigating hazards.
• Distributed Acoustic Sensing (DAS) offers a scalable, cost-effective seismic network solution.
• Supervised learning methods for DAS data are limited by scarce manually labeled datasets.

Purpose of the Study:

• To develop a novel self-supervised pretraining technique for DAS data.
• To address the challenge of limited annotated data in earthquake monitoring.
• To present DASFormer as an effective tool for seismic phase detection and anomaly detection.

Main Methods:

• DASFormer utilizes a coarse-to-fine framework to model spatial-temporal signal correlation in DAS data.
• The method employs self-supervised pretraining, treating earthquake monitoring as an anomaly detection task.
• DASFormer is directly applied as a seismic phase detector without requiring labeled data.

Main Results:

• DASFormer demonstrates effectiveness across multiple evaluation metrics for unsupervised seismic detection.
• The proposed method outperforms state-of-the-art time-series forecasting, anomaly detection, and foundation models.
• Experimental results validate DASFormer's capability in seismic event detection using DAS data.

Conclusions:

• DASFormer provides a powerful self-supervised approach for earthquake monitoring with DAS.
• The technique overcomes limitations of supervised methods due to data scarcity.
• DASFormer shows potential for fine-tuning to various downstream seismic analysis tasks.