Hierarchical Exploration of Continuous Seismograms With Unsupervised Learning.

Area of Science:

• Geophysics
• Signal Processing
• Machine Learning

Background:

• Continuous seismograms contain diverse signals crucial for understanding geological processes.
• Identifying and classifying these signals is essential but challenging due to signal variety.

Purpose of the Study:

• To develop an unsupervised strategy for identifying and classifying signal classes within continuous single-station seismograms.
• To leverage waveform feature extraction and hierarchical clustering for seismic data exploration.

Main Methods:

• Utilized a deep scattering network combined with independent component analysis for waveform feature extraction.
• Employed agglomerative clustering to group waveforms hierarchically, visualized via a dendrogram.
• Applied the method to a two-day seismogram from the North Anatolian Fault, Turkey.

Main Results:

• Successfully distinguished clusters of anthropogenic/ambient seismic noise and earthquake activity at a low hierarchy level.
• Identified a seismic burst of approximately 200 anthropogenic events with similar waveforms and high-frequency signals at a high hierarchy level.
• Demonstrated the utility of cluster hierarchy for analyzing signal families and subclusters, especially for under-represented signals like earthquakes.

Conclusions:

• The proposed unsupervised method effectively identifies and categorizes diverse seismic signals from continuous data.
• The hierarchical clustering approach provides valuable insights for exploring seismic data and discovering new signal types.
• This data-driven strategy offers a powerful tool for seismic data analysis and interpretation.