GNGTS 2018 - 37° Convegno Nazionale

310 GNGTS 2018 S essione 2.1 physical mechanisms of a seismic region. A spatio-temporal cluster is defined as a rate increase with respect to the typical background level of the area. Two categories of clusters are commonly defined: (1) mainshock-aftershock sequences, in which the largest magnitude event occurs first, possibly preceded by a few foreshocks, and (2) swarms, in which events occur closely clustered in time and space without a single outstanding shock. Static or dynamic stress changes are usually associated to mainshock-aftershock sequences, whereas transient aseismic slow slip has been recently hypothesized to drive a swarm-like activity (Zaliapin and Ben-Zion, 2016). Several declustering (clustering) algorithms are available in the seismological literature (see van Stiphout et al. , 2012 for a review), the most popular ones being the window methods by Gardner and Knopoff (1974) and by Reasenberg (1985). In this study we compare two recent declustering algorithms that allow to recognize clusters in the time-space-magnitude domain: the stochastic declustering method (Zhuang et al. , 2004) and the nearest-neighbor method (Zaliapin and Ben-Zion, 2016 and references therein). Both methods are data-driven tools for the identification of clusters and provide a topological description of the internal cluster structures. The stochastic declustering approach (SD) is based on the space-time ETAS (epidemic-type aftershock sequence) model, a branching point process defined by a hazard function conditional on the observation history. According to the ETAS model, a time-homogeneous Poisson process generates independent background earthquakes that trigger off other events, with a space-time decay modelled by the Omori-Utsu law. An iterative algorithm simultaneously provides the maximum likelihood estimates of the eight model parameters and a nonparametric kernel estimate of the spatial background rate. Simulation by the thinning procedure for point processes allows to easily determine if an event occurred at time t is independent or triggered by previous earthquakes. Thus the whole process stochastically splits into the background process (i.e. the declustered catalog) and the cluster processes triggered by each background event; this means that the SD method does not produce a unique declustered catalog, but many declustered catalogs by simulation. The nearest-neighbor approach (NN) is based on the nearest-neighbor distance between two earthquakes in the space-time-energy domain (Baiesi and Paczuski, 2004): this distance links together the inter-occurrence time, the fractal dimension of the hypocentres distribution, and the Gutenberg–Richter law. Only two parameters need to be estimated, namely fractal dimension d and b-value, which can be jointly and robustly identified by the Unified Scaling Law for Earthquakes (USLE) method. For any earthquake catalog, the histogram of the distances between every pair of nearest-neighbor events clearly shows a bimodal distribution that can be approximated as a mixture of two Gaussian distributions, one associated with the Poissonian background activity and the other with the clustered populations. Further details on the analysis by the nearest-neighbor method are in Peresan and Gentili (2018). An interesting case study is the seismicity of Friuli-Venezia-Giulia (FVG) area. Available instrumental data, as reported by OGS bulletins, range fromMay 1977 toApril 2018 and denote a moderate seismic activity in the region, the largest one being the M5.6 earthquake occurred on 12 April 1998. However, based on historical knowledge, the area can be hit by strong events, such as the M6.5 Friuli earthquake in 1976. Both declustering algorithms, SD and NN, are applied to the OGS dataset with a completeness threshold magnitude 2.0 in the time period 1994-2018. Clusters identified by NN method are compared to the most probable clusters identified by SD method. The cluster structures produced by NN and SD approaches have comparable trend in terms of spatial extent. We notice that, compared to NN method, SD method tends to identify clusters that cover a slightly longer time interval, linking sequences relatively far in time as long as they cover the same area. On the other side, the NN method tends to include in clusters a few events that are quite distant from the mainshock epicenter, when they are very close in time. The percentage of events associated with the background activity is comparable: about 53%