Wavelet decomposition and advanced denoising techniquesn for analysis and classification of seismic signals

Abstract

This work describes an automatic classification procedure for seismic signals suitable for the analysis of complex, broad-band waveforms commonly associated with fluid-rock interaction in volcanic and hydrothermal systems. Based on Discrete Wavelet Transform, a set of significant seismic signal features that characterize the type of event is identified (e.g. noise, volcano tectonic, long period). These features are initially assessed for events whose category (class) can be previously determined by an expert analyst. A Bayesian Pattern Recognition supervised technique based on these features is adopted to classify a new ‘unlabelled pattern’, whose class is unknown. In this way values computed for known events are used to classify events of unknown identity ('supervised classification'). A test was performed on seismological data recorded at Campi Flegrei (Italy), which was divided into three classes. Automatic classification accuracy ranges from 82% to 100% over a broad range of datasets.