Dipartimento di Fisica e Astronomia - Università di Bologna

Artifacts often affect seismic catalogs. Among them, the presence of man-made contaminations such as quarry blasts and explosions is a well-known problem. Using a contaminated dataset reduces the statistical significance of results and can lead to erroneous conclusions, thus the removal of such nonnatural events should be the first step for a data analyst. Blasts misclassified as natural earthquakes, indeed, may artificially alter the seismicity rates and then the b-value of the Gutenberg and Richter relationship, an essential ingredient of several forecasting models. At present, datasets collect useful information beyond the parameters to locate the earthquakes in space and time, allowing the users to discriminate between natural and nonnatural events. However, selecting them from webservices queries is neither easy nor clear, and part of such supplementary but fundamental information can be lost during downloading. As a consequence, most of statistical seismologists ignore the presence in seismic catalog of explosions and quarry blasts and assume that they were not located by seismic networks or in case they were eliminated. We here show the example of the Italian Seismological Instrumental and Parametric Database. What happens when artificial seismicity is mixed with natural one?

The destructive 2009 L’Aquila and 2012 Emilia Romagna earthquakes led the Italian Dipartimento della Protezione Civile (DPC) to fund nine groups studying seismic precursors. Three of the groups produced testable predictions by the DPC deadline of 31 May 2013, using: (1) Radon in a well in Friuli, (2) temperature, flow, CO2 flux, and other variables measured in wells in Emilia Romagna, and (3) an artificial neural network (ANN) algorithm applied to seismicity. We evaluated the geochemical precursors by comparing their success to that of an equal number of predictions at the same locations and with the same individual and total durations as the actual predictions, but at random times. This approach avoids modeling seismicity and thereby precludes concluding that predictions are “good” simply because the model for seismicity is bad. Neither precursor predicts significantly better than chance. ANN was a poor predictor of events large enough to affect public safety.