Automatic network-based classification of volcanic tremors, localization of their source and estimation of their seismic wave speed
IPGP - Îlot Cuvier
Séminaires de Sismologie
We develop a network-based method to automatically classify volcanic tremors, localize their sources and estimate the associated wave speed. The proposed approach exploits the coherence of tremor signals across the network, that is estimated from the array covariance matrix [Seydoux et al., 2016a]. The method is applied on four and a half years of continuous seismic data recorded by 19 permanent seismic stations in the vicinity of the Klyuchevskoy volcanic group (KVG) in Kamchatka (Russia), where five volcanoes were erupting during the considered time period. We compute and analyze daily covariance matrices together with their eigenvalues and eigenvectors. As a first step, most coherent signals corresponding to dominating tremor sources are detected based on the width of the covariance matrix eigenvalues distribution. With this approach, the volcanic tremors of the two volcanoes known as most active during the considered period, Klyuchevskoy and Tolbachik, are efficiently detected. As a next step, we consider the array covariance matrix’s first eigenvector computed every day. The main hypothesis of our analysis is that these eigenvectors represent the principal component of the daily seismic wavefield and, for days with tremor activity, characterize the dominant tremor sources. Those first eigenvectors can therefore be used as network-based fingerprints of tremor sources. A clustering process is developed to analyze this collection of first eigenvectors, using correlation coefficient as a measure of their similarity. Then we develop a maximization process based on cross-correlations amplitudes, to localize the sources of volcanic tremor and estimate their associated wave speed. As a result, we characterize seven tremor sources associated with different periods of activity of four volcanoes: Tolbachik, Klyuchevskoy, Shiveluch, and Kizimen. The developed method does not require a priori knowledge, is fully automatic and the database of network-based tremor fingerprints can be continuously enriched with new available data.