Unsupervised pattern recognition in continuous seismic wavefield records using Self-Organizing Maps

Abstract

SUMMARY Modern acquisition of seismic data on receiver networks worldwide produces an increasing amount of continuous wavefield recordings. In addition to manual data inspection, seismogram interpretation requires therefore new processing utilities for event detection, signal classification and data visualization. The use of machine learning techniques automatises decision processes and reveals the statistical properties of data. This approach is becoming more and more important and valuable for large and complex seismic records. Unsupervised learning allows the recognition of wavefield patterns, such as short-term transients and long-term variations, with a minimum of domain knowledge. This study applies an unsupervised pattern recognition approach for the discovery, imaging and interpretation of temporal patterns in seismic array recordings. For this purpose, the data is parameterized by feature vectors, which combine different real-valued wavefield attributes for short time windows. Standard seismic analysis tools are used as feature generation methods, such as frequency–wavenumber, polarization and spectral analysis. We use Self-Organizing Maps (SOMs) for a data-driven feature selection, visualization and clustering procedure. The application to continuous recordings of seismic signals from an active volcano (Mount Merapi, Java, Indonesia) shows that volcanotectonic and rockfall events can be detected and distinguished by clustering the feature vectors. Similar results are obtained in terms of correctly classifying events compared to a previously implemented supervised classification system. Furthermore, patterns in the background wavefield, that is the 24-hr cycle due to human activity, are intuitively visualized by means of the SOM representation. Finally, we apply our technique to an ambient seismic vibration record, which has been acquired for local site characterization. Disturbing wavefield patterns are identified which affect the quality of Love wave dispersion curve estimates. Particularly at night, when the overall energy of the wavefield is reduced due to the 24-hr cycle, the common assumption of stationary planar surface waves can be violated.