Seismic pattern recognition techniques to predict large eruptions at the Popocatépetl, Mexico, volcano

Using pattern recognition techniques, we formulate a simple prediction rule for a retrospective prediction of the three last largest eruptions of the Popocatépetl, Mexico, volcano that occurred on 23 April–30 June 1997 (Eruption 1; VEI ~ 2–3); 11 December 2000–23 January 2001 (Eruption 2; VEI ~ 3–4) and 7 June–4 September 2002 (Eruption 3; explosive dome extrusion and destruction phase). Times of Increased Probability (TIP) were estimated from the seismicity recorded by the local seismic network from 1 January 1995 to 31 December 2005. A TIP is issued when a cluster of seismic events occurs under our algorithm considerations in a temporal window several days (or weeks) prior to large volcanic activity providing sufficient time to organize an effective alert strategy. The best predictions of the three analyzed eruptions were obtained when averaging seismicity rate over a 5-day window with a threshold value of 12 events and declaring an alarm for 45 days. A TIP was issued about six weeks before Eruption 1. TIPs were detected about one and four weeks before Eruptions 2 and 3, respectively. According to our objectives, in all cases, the observed TIPs would have allowed the development of an effective civil protection strategy. Although, under our model considerations the three eruptive events were successfully predicted, one false alarm was also issued by our algorithm. An analysis of the epicentral and depth distribution of the local seismicity used by our prediction rule reveals that successful TIPs were issued from microearthquakes that took place below and towards SE of the crater. On the contrary, the seismicity that issued the observed false alarm was concentrated below the summit of the volcano. We conclude that recording of precursory seismicity below and SE of the crater together with detection of TIPs as described here, could become an important tool to predict future large eruptions at Popocatépetl. Although our model worked well for events that occurred in the past, it is necessary to verify the real capability of the model for future eruptive events.