Automatic Classification and A-Posteriori Analysis of Seismic Event Identification at Soufrière Hills Volcano, Montserrat

Document Type


Publication Date



automatic classification, Montserrat, neural network, volcano seismicity

Digital Object Identifier (DOI)


Seismic energy radiation at Soufrière Hills volcano, Montserrat, is made up by various types of transient signals, which are distinguished by the Montserrat Volcano Observatory (MVO) in different classes with respect to their characteristics and/or origin. There are five fundamental classes, i.e., Volcano-Tectonic Events, Regional Events, Long-Period Events, Hybrid Events, and Rockfalls. Identification and classification of these transients, which have been hitherto carried out manually by various staff members, yield important information for the assessment of the state of the volcano system. In the frame of the MULTIMO project, we proposed the application of Artificial Neural Networks (ANN) for the classification of these kind of data in order to handle large data sets, and to achieve reproducible results, emulating the expert's analysis. Using the manual routine classification as a-priori information, we obtained a fair performance of such an automatic processing, with 70% of the automatic classifications being consistent with the original ones. From an analysis of the misclassified events, however, we found that for most of them the original a-priori classification was incorrect.

In this study, we first revised manually the original a-priori classification. Based on a data set of 6000 events, we carried out a reanalysis of the seismic traces recorded at different seismic stations. Then, using this new information, we trained and tested the ANN, obtaining a successful classification in ca. 80% of records. Particularly, the automatic classification was excellent in the identification of Rockfalls and Volcano-Tectonic Events. Among the misfits, we observe the erroneous attribution of Long-Period and Hybrid Events to Rockfalls. This may be partly explained by the fact that signals addressed to as Rockfalls contain frequently contributions of various sources. Overall, the failure in the classification between some types of transients highlights the problem of the concurrent activation and/or unclear separation of distinctive sources from which the signals stem. We conclude that the automatic classification with ANN is a powerful tool for handling large data masses as well as for the a-posterior analysis of the consistency of the classification problem.

Was this content written or created while at USF?


Citation / Publisher Attribution

Journal of Volcanology and Geothermal Research, v. 153, issues 1-2, p. 1-10