Volcano-Independent Seismic Recognition: detecting and classifying
events of a given volcano using data from others
Abstract
Modern seismic networks provide a huge amount of data received in
real-time, being impossible the manual identification of relevant events
useful to monitor the activity of the volcano. Thus, many volcano
observatories are interested in tools to perform an online, automatic
analysis of the seismic activity. Machine Learning area provides various
of Volcano-Seismic Recognition (VSR) systems designed to classify
seismic events in real-time. However, only a few approaches can also
detect them in a continuous data streams. Most of those VSR systems are
based on the 2-step supervised paradigm: 1. A training database (X-DB)
of a given volcano ’X’ is prepared with hundreds of events manually
detected and classified according to their physical origin. 2.
Statistical models are built analysing this DB, and are later used to
automatically identify events in new data recorded at the volcano X.
This supervised procedure is the major drawback to achieve a fast
deployment of a VSR system for another volcano Y, as the preparation of
its own Y-DB takes considerable time, and requires qualified operators
and previous recordings, which is difficult for volcanoes without recent
activity or which haven’t been monitored. In order to overcome these
limitations, the EU-funded project ’VULCAN.ears’ focused on real-time,
Volcano-Independent VSR (VI.VSR) approaches. It proposes alternative
solutions based on state-of-the-art technologies as universal DBs and
models, waveform standardisation and parallel architectures. Recent
results obtained by mixing DBs from Popocatépetl, Colima, Deception and
Arenal active volcanoes will be presented. We apply VULCAN.ears
technologies to evaluate VSR systems on joint DBs built with data of
several volcanoes. We also use volcano-independent models to
automatically classify events of another volcano, analysing how the
recognition accuracy varies as the training DB becomes more complex. All
tests are carried out by an easy to use, user-friendly graphical
application (geoStudio). All these achievements produce new insights
useful to redesign the next-generation, portable and robust VSR systems.