Automatic Seismic Event Detection, Characterization and Classiication: a Probabilistic Approach

An application of a wavelet transform and a Bayesian statistical method to seismic event detection and classiication has been studied using data from the New England region. A wavelet expansion forms a new basis set for picking out, from a data stream, important features of a seismic event: time, energy and predominant period of the rst, peak and last waveforms. From these informative features of the seismic event and with some simple tests we are able to discriminate and discard most erroneous triggers. Classiication of the remaining events into one of the following classes: teleseisms, regional earthquakes, near earthquakes, and quarry blasts, is accomplished with conditional class densities derived from training data by nding the maximum a posteriori probability. Using a series of hypothesis tests in this Bayesian framework, we have developed a detection and characterization scheme that operates on a channel of continuously recorded seismic data. We have developed a PC based system, running in parallel with a data acquisition system, that carries out event detections and identiications at remote sites in New England.