The Application of Speaker Recognition Techniques in the Detection of Tsunamigenic Earthquakes

Abstract:

Tsunami warning procedures adopted by national tsunami warning centres largely rely on the classical approach of earthquake location, magnitude determination, and the consequent modelling of tsunami waves. Although this approach is based on known physics theories of earthquake and tsunami generation processes, this may be the main shortcoming due to the need to satisfy minimum seismic data requirement to estimate those physical parameters. At least four seismic stations are necessary to locate the earthquake and a minimum of approximately 10 minutes of seismic waveform observation to reliably estimate the magnitude of a large earthquake similar to the 2004 Indian Ocean Tsunami Earthquake of M9.2. Consequently the total time to tsunami warning could be more than half an hour. In attempt to reduce the time of tsunami alert a new approach is proposed based on the classification of tsunamigenic and non tsunamigenic earthquakes using speaker recognition techniques. A Tsunamigenic Dataset (TGDS) was compiled to promote the development of machine learning techniques for application to seismic trace analysis and, in particular, tsunamigenic event detection, and compare them to existing seismological methods. The TGDS contains 227 off shore events (87 tsunamigenic and 140 non-tsunamigenic earthquakes with M≥6) from Jan 2000 to Dec 2011, inclusive. A Support Vector Machine classifier using a radial-basis function kernel was applied to spectral features derived from 400 sec frames of 3-comp. 1-Hz broadband seismometer data. Ten-fold cross-validation was used during training to choose classifier parameters. Voting was applied to the classifier predictions provided from each station to form an overall prediction for an event. The F1 score (harmonic mean of precision and recall) was chosen to rate each classifier as it provides a compromise between type-I and type-II errors, and due to the imbalance between the representative number of events in the tsunamigenic and non-tsunamigenic classes. The described classifier achieved an F1 score of 0.923, with tsunamigenic classification precision and recall/sensitivity of 0.928 and 0.919 respectively. The system requires a minimum of 3 stations with ~400 seconds of data each to make a prediction. The accuracy improves as further stations and data become available.