Warning of impending volcanic unrest is of high priority for volcano observatories. It is one of the main surveillance tasks of the Osservatorio Etneo, a branch of the Istituto Nazionale di Geofisica e Vulcanologia (INGV) in Italy, which runs the 24-hours surveillance of Etna. With its frequent Strombolian explosions, lava fountains and lava flows, Etna is the largest and most active volcano in Europe. Hundreds of lava fountains occurred from 2000 to 2021 alone. They produce abundant ash fallout along with lapilli, posing a serious threat to air and road traffic, with heavy social and economic impacts. Early changes detected by a monitoring network of sensors are therefore of paramount importance for highlighting variations that herald impending eruptions.
In the specific case of Etna, the seismic background radiation known as volcanic tremor plays a key role in volcano surveillance. The application of Pattern Recognition techniques, namely the “Self Organizing Maps” (SOM), has allowed us to recognize unrest in its early stage, when the eruptive activity was still mild (Fig. 1).
However, volcanic activity on Etna may hover on low to intermediate levels (e.g., mild Strombolian eruptions) even for months without being followed by a paroxysmal event (see Fig. 2). As a consequence, a warning system focused only on the very early signs of unrest may flag an alert over long time spans. As this jeopardizes the credibility of the warning, there is a need of identifying “points of no return”. These points should mark levels of activity which, once reached, have a high probability of being followed by a paroxysm within short time spans, on the order of hours.
Applying SOM to long volcanic tremor data (seismic records spanning years), we were able to verify that such points of no return can be indeed identified. We designed a specific alert system for paroxysmal events, which focuses on recognizing these points; it gains robustness when the alert flag is reached at a sufficient number of recording stations. The criteria setting of such a multi-station warning system for paroxysmal events regards both the energy levels of the signal encountered in each seismic station as well as the number of stations in which these levels are reached. Suitable settings can be found using so-called 'Receiver Operation Characteristic' (ROC) curves (Fig. 3). These are based on the comparison of the “true positive” rate, where a true positive is a correctly flagged paroxysmal event, versus the “false positive rate”, where a false positive is encountered when the system raises the flag, but no paroxysm occurs. There is an intrinsic tradeoff between the two rates. By choosing a sensitive configuration, the rate of correctly identified events improves, but at the cost of a high false positive rate. We were able to establish settings regarding the number of stations as well as the intensity of the signal such that ~80% of events (90% for the whole year 2021) were correctly recognized with a number of false positives being less than one third of the true positives. Besides, more than 50% of the false positives were found shortly before the event occurred. These come as a precursor and may be welcomed by operators rather than being considered an error of the alert system.
We applied our multi-station system in a blind test, namely using it for data not considered when fixing the configuration. The good performance mentioned above confirmed the results previously obtained. Being completely automatized, the alert system is appropriate for online processing in real time, providing the staff of the Osservatorio Etneo with rapid and reliable information. In particular, such a system may become extremely useful when visual monitoring is hindered for adverse meteorological conditions. Besides, it may allow us to scroll through the large datasets accumulated over the years, checking if any relevant episode of activity had escaped the attention of previous analyses.