Clustering earthquake signals and background noises in continuous seismic data with unsupervised deep learning

Area of Science:

• Geophysics
• Machine Learning
• Seismology

Background:

• Vast amounts of seismic data challenge traditional supervised analysis methods.
• Existing seismic analysis can be biased by conventional seismological models.
• There is a need for automated, unbiased seismic data analysis techniques.

Purpose of the Study:

• To develop and present a novel unsupervised machine learning framework for seismic signal detection and clustering.
• To address the limitations of human-expert-intensive and potentially biased seismic data analysis.
• To demonstrate the framework's capability in identifying seismic patterns in continuous records.

Main Methods:

• Utilized a deep scattering network combined with a Gaussian mixture model.
• Applied an unsupervised machine learning approach to cluster seismic signal segments.
• Developed a framework for detecting novel seismic structures without prior human labeling.

Main Results:

• Successfully detected and clustered seismic signals in continuous seismic records.
• Demonstrated blind detection and recovery of repeating precursory seismicity before the 2017 Greenland landslide.
• Identified previously unrecognized seismic patterns preceding a major geological event.

Conclusions:

• The unsupervised machine learning framework offers a powerful tool for analyzing large seismic datasets.
• The approach can reveal subtle seismic patterns, such as precursory seismicity, missed by traditional methods.
• This framework has the potential to improve seismic activity forecasting in seismogenic zones.