Spectro-temporal features applied to the automatic classification of volcanic seismic events

Abstract

Feature extraction and selection are very relevant processes in the design of automatic classifiers. In the context of volcanic seismic signal classification, most of the features presented in the literature have been extracted separately from a single domain, such as time or frequency. However, spectrograms, which combine time and frequency information, are widely used by experts during the classification of manual seismic events. This paper proposes to evaluate the performance of classifiers trained with features extracted from the spectro-temporal domain, individually or combined with other conventional features. The parameters were extracted from the spectrogram, based on a curve which combines the high energy components and the frequency bandwidth information through the duration of the event. The tests were performed at the Llaima volcano and seismic events were classified into four classes: long-period, tremor, volcano-tectonic and tectonic, using a database of signals recorded between the years 2009 and 2017. The main achievements of this study were the reduction of more than 70% of the error and false positive rates and also a reduction of approximately 30% of the number of features, compared with a baseline established in previous studies. Thus, the inclusion of spectro-temporal information was considered relevant to complement the conventional features and to support classification.