Automatic Recognition of Landslides Based on Neural Network Analysis of Seismic Signals: An Application to the Monitoring of Stromboli Volcano (Southern Italy)

Research paper by Antonietta M. Esposito, Luca D’Auria, Flora Giudicepietro, Rosario Peluso, Marcello Martini

Indexed on: 23 Oct '12Published on: 23 Oct '12Published in: Pure and Applied Geophysics


In the last 9 years, the amount and the quality of geophysical and volcanological observations of Stromboli's' activity have undergone a marked increase. This new information highlighted that the landslides on the Sciara del Fuoco flank are tightly linked to the volcanic activity. Actually, at the beginning of the December 28, 2002, effusive eruption, the seismic monitoring network was less dense than now, and therefore it is not known if there was an increase in the landslide rate before the eruption. Despite this, it is known that a big landslide occurred 2 days after the beginning of the eruption which caused a tsunami (December 30, 2002). More recently, the effusive eruption in February 2007 was preceded by an increase in landslides on the Sciara del Fuoco flank, which were recorded by the seismological monitoring system that had been improved after the 2002–2003 crisis. These episodes led us to believe that monitoring the Sciara del Fuoco flank instability is an important topic, and that landslides might be significant short-term precursors of effusive eruptions at the Stromboli volcano. To automatically detect landslide signals, we have developed a specialized neural algorithm. This can distinguish between landslides and the other types of seismic signals usually recorded at the Stromboli volcano (i.e., explosion quakes and volcanic tremor). The discrimination results show an average performance of 98.67 %. According to the experience of the crisis of 2007, to identify changes that can be considered as precursors of effusive eruptions, we set up an automatic decision-making method based on the neural network responses. This method can operate on a continuous data stream. It calculates a landslide percentage index (LPI) that depends on the number of records that are classified by the net as landslides over a given time interval. We tested the method on February 27, 2007, including the beginning of the effusive phase. The index showed an increase as early as at 09:00 UTC on that day and reached its maximum value (100 %) at 12:00, about 40 min before the onset of the eruption. After the beginning of the effusive phase, the index remains high due to the blocks that roll down along the slope from the front of the lava flow. On the basis of these tests, we propose a decision-making method that is able to recognize a trend in the LPI similar to that of 2007 eruption, allowing the identification of precursors of effusive phases at the Stromboli volcano.