# Bulletin of the Seismological Society of America

## Abstract

Earthquake early warning systems (EEWSs) are becoming a suitable instrument for seismic risk management in real time. In fact, they are implemented or are undergoing testing in many countries around the world because EEWSs represent an effective approach to mitigating seismic risk on a short timescale. EEWSs are based on the use of relationships between some parameters measured on the initial portion of seismic signal after the onsets. Here, we address the first approach to the implementation of EEWS in eastern Sicily, a region that has been hit by several destructive earthquakes. We estimated the peak displacement amplitude of the first portion of *P* and *S* waves *P*_{d}, the ground‐motion period parameter *τ*_{c}, and the peak ground velocity (PGV) from earthquakes with *M*_{L}≥2.8 recorded by the broadband stations operated by the Istituto Nazionale di Geofisica e Vulcanologia. We found that the *P*_{d} is correlated with the size of the earthquake and may be used to compute the magnitude for an EEWS in this area. We also derived the relationships between *τ*_{c} and *M*_{L}, and between *P*_{d} and PGV, which can be used to provide on‐site warning in the area around a given station and to evaluate the potential damaging effects. These relationships may be deemed a useful guide for future implementation of the EEWS in the region.

## Introduction

The structural setting of eastern Sicily is connected to the complex tectonic environment of the central Mediterranean basin (Fig. 1a,b), which is subjected to the north‐northwest convergence between the Eurasian and African plates and to the geodynamic processes due to the opening of the Tyrrhenian basin (Faccenna *et al.*, 2001). This tectonic setting makes eastern Sicily one of the most hazardous zones in Italy, characterized by a significant rate of crustal seismicity. The area, in fact, has undergone some disastrous earthquakes. Among them, we can mention the ones occurring on 4 February 1169 (moment magnitude *M*_{w} 6.4 by the 2011 version of the Catalogo Parametrico dei Terremoti Italiani [Rovida *et al.*, 2011]; hereafter, CPTI11; I_{MCS}=X on the Mercalli–Cancani–Sieberg [MCS] scale), 10 December 1542 (*M*_{w} 6.8 by CPTI11; *I*_{MCS}=IX–X), 9 and 11 January 1693 (*M*_{w} 6.2 and 7.4 by CPTI11, respectively; I_{MCS}=IX and X–XI, respectively), 20 February 1818 (*M*_{w} 6.2 by CPTI11; I_{MCS}=VIII–IX), and on 28 December 1908 (*M*_{w} 7.1 by CPTI11; I_{MCS}=X), which is one of the strongest earthquakes to ever occur in the central Mediterranean area (Boschi *et al.*, 1995). In particular, the 1693 and 1908 earthquakes completely destroyed the cities of Catania and Messina, respectively, and were followed by large tsunamis as well. More recently, eastern Sicily has experienced minor events. On 13 December 1990, a seismic event (I_{MCS}=VIII) produced much damage in a wide area despite its magnitude *M*_{s} being equal to 5.4 (Amato *et al.*, 1995). It caused severe damage to the cities of Augusta and Catania, and 19 casualties as well, renewing attention to the seismic potential of eastern Sicily and the sociopolitical consequences of earthquakes in that region.

In eastern Sicily, an important role is played by the volcanic area of Mt. Etna (Fig. 1a). Indeed, its seismicity creates a rather specific scenario, with relatively small, very shallow earthquakes that can produce important damage on a local scale. The most recent examples of destructive seismic events belong to a seismic swarm (maximum *M*_{L} 4.8) occurring at Mt. Etna in October 2002. The most damaging event of the swarm (*M*_{L} 4.5; I_{MCS}=VIII) occurred on 29 October 2002 (Castello *et al.*, 2006). It struck a densely urbanized area on the southeastern flank of Mt. Etna, producing heavy damages even to reinforced concrete structures (Azzaro *et al.*, 2006).

In the framework of several national projects, different research fields are supported and furthered, including earthquake‐hazard mitigation in eastern Sicily. In particular, two closely related development projects deal with implementing an earthquake early warning system (EEWS) in the considered area as a tool for real‐time seismic risk mitigation and management. Indeed, EEWS is the current focus of considerable research effort, and its potential applicability for the immediate activation of safety measures for critical systems is already undoubted (i.e., Wieland *et al.*, 2000; Sato *et al.*, 2011). As a matter of fact, the development of these systems is more related to an actual possibility to immediately trigger actions for the protection of strategic sites and lifelines rather than as an instrument giving a massive alarm to the communities (evacuating people from buildings requires long warning times that are rarely available in urbanized areas). Moreover, the recent impulse to develop EEWSs worldwide is significantly raising the interest toward the potential use of EEWSs for the structural control, adopting structure‐specific applications (active and semiactive control devices) set up within the lead time so as to optimize the expected structural response.

Over the last 10 years, as a result of the technological evolution in the fields of both computing systems and data transmission, it has been possible to develop more effective techniques to analyze seismic data in real time. Indeed, the real‐time seismology (RTS) integrates a real‐time telemetry system, where the transmission of information takes place with a very low latency, with automatic processing of recorded signals, providing fast and reliable estimates of the main earthquake parameters (location and magnitude) in the first few seconds during its occurrence.

On the basis of the configuration of the seismic network, EEWSs can be distinguished into two main types: (1) regional and (2) on‐site (or site‐specific) warning systems (Kanamori, 2005). The regional EEWSs use a dense network of seismic sensors, partially or entirely covering the area where earthquakes are likely to occur, with real‐time capability to estimate the source parameters (event location and magnitude) of earthquake seismic events using the early portion of recorded seismograms. Therefore, the system uses them to predict a specific peak ground motion at distant target sites through an empirical ground‐motion prediction equation. On‐site EEWSs such as the ones installed in Japan (Urgent Earthquake Detection and Alarm System [UrEDAS], Nakamura, 1988), in California (Earthquake Alarm Systems, Allen and Kanamori, 2003), or in Romania (Wenzel *et al.*, 1999; Böse *et al.*, 2007) are based on a single seismic station (single‐station approach) or on an array of seismic stations installed near the target site that needs to be alerted. After detecting the arrival of the faster but weaker *P* wave, the system computes the peak amplitude and the predominant period in the very first seconds of the *P* waves (Wu and Kanamori, 2005a,b) to estimate the associated peak ground motion of the more destructive *S* and surface waves at the target. This approach is relatively simple but less accurate than the regional approach. It also provides smaller effective lead time (e.g., the time span from the arrival of the damaging waves to the alert notification at a given target site) compared with the regional approach, which also has the advantage that the system is constantly run and tested, and the source parameter estimates gain in accuracy as more data are recorded and analyzed. However, the data processing takes some time and the alarm could be issued after the ground motion reaches the sites of interest (defining the so‐called blind zone, Kanamori, 2005). Meanwhile, regional systems are more effective with applications such as shake maps, which are very useful for emergency management immediately after the event (Wald *et al.*, 1999). At the same time, the site‐specific EEWSs are certainly devoted to reducing the exposure of strategic facilities (lifelines, transportation infrastructures, power plants, etc.) in real time by automated safety actions. The Ignalina Nuclear Power Plant in Lithuania takes advantage of a site‐specific EEWS (Wieland *et al.*, 2000). The seismic network, composed of six stations, is installed 30 km from the reactor and ensures that an alarm sounds 4–8 s before the ground motion affects the power plant. This time is enough to activate the control rods because they need only 2 s to come into use.

The regional and on‐site warnings can be combined to give a hybrid EEWS, using the potential of regional seismic networks to protect different critical systems and/or the population at the same time (Kanamori, 2005; Iervolino *et al.*, 2007). Wu and Kanamori (2005b) experimented with the method for on‐site warning based on the measure of the predominant period (*τ*_{c}) of initial portion of *P* wave using the data recorded by accelerometric stations of the seismic network in Taiwan. Recently, Zollo *et al.* (2010) developed a method for integrating the on‐site early warning (EEW) approach into a regional one. It enables estimating a potential damage zone (PDZ) for the forthcoming earthquake, namely the region in which the most damage is awaited. This method is based on the estimation of the peak ground displacement (*P*_{d}) and the predominant period (*τ*_{c}), in real time, at recording sites located at increasing distances from the earthquake epicenter. An alert level is associated with each recording site on the basis of critical values of *P*_{d} and *τ*_{c}. As shown by several authors, these two EEW parameters are empirically correlated to magnitude (e.g., Wu and Kanamori, 2008; Zollo *et al.*, 2010; Colombelli *et al.*, 2012; Carranza *et al.*, 2013) and to peak ground velocity (PGV) and acceleration (e.g., Böse *et al.*, 2007; Zollo *et al.*, 2010; Carranza *et al.*, 2013).

In this work, we determine the two EEW parameters *P*_{d} and *τ*_{c} together with the PGV using a set of more than 200 seismic events and explore the use of *P*_{d} and *τ*_{c} parameters for EEW purposes in the studied area. The goal is to compute specific empirical relationships of these two parameters with earthquake size and peak ground‐motion parameters for future applications of an EEWS in eastern Sicily.

## Data Set and Record Processing

Seismic hazard in eastern Sicily is linked to earthquakes occurring in different seismotectonic areas and associated with various types of faulting mechanisms. This should ensure that the effects due to rupture directivity and focal mechanism on peak amplitudes are averaged out.

The eastern part of Sicily is mainly characterized by two active volcanic regions, the Aeolian Archipelago lying in the southern Tyrrhenian Sea and Mt. Etna located in central–eastern Sicily. From a seismological point of view, northern Sicily and its Tyrrhenian offshore are characterized by the activity of different tectonic structures associated with both the collision between European and African plates, and with the opening of the Tyrrhenian basin. The seismicity is mainly located in the hinge zone between the southern Tyrrhenian Sea and northern Sicily, and is confined to two principal hypocentral sectors (Gueguen *et al.*, 2002; Giunta *et al.*, 2009). The deep seismicity is essentially connected to the subduction processes of the Ionian lithospheric slab beneath the Calabrian arc and affects northeastern Sicily. In this case, the shallow seismicity represents the expression of the strain crossing the whole orogeny (Neri *et al.*, 1996).

Etnean seismic events share their signatures with earthquakes recorded in the near‐tectonic environments of the Hyblean plateau and Peloritani‐Calabrian arc (Patanè *et al.*, 1997; Patanè and Giampiccolo, 2004). The regional tectonic stresses together with the local stresses connected to the magma migration in the Earth’s crust provide the necessary energy for rock failure.

The local surface tectonic structures on Mt. Etna are connected to intense superficial seismic activity essentially characterized by earthquakes often clustered in swarms and having focal depths of generally less than 3 km (Patanè *et al.*, 2004). Although the shallowest events characterize the most central–eastern portion of the volcano, they affect the entire volcanic area, even though they are fewer in number. Occurring in particular geological conditions, they are considered as forming a family of events whose characteristic hypocentral location has effects on both seismic scaling laws and wave propagation phenomena (e.g., Tusa and Langer, 2016). This complex tectonic situation raises questions on the homogeneity of the parent population and on the treatment of the data as a whole. Therefore, we did not include in the analysis any shallow events (*H*≤5 km) occurring in the Mt. Etna area. In doing so, we reduce the introduction of heterogeneities in the data set. The selected data set consists of 232 crustal seismic events (Fig. 2) recorded between 2006 and 2014 by the stations of the “Rete Sismica Permanente della Sicilia Orientale” operated by Istituto Nazionale di Geofisica e Vulcanologia (INGV) Sezione di Catania – Osservatorio Etneo (Fig. 1c). The seismic network is located in a region between the Hyblean plateau and the volcanic archipelago of Aeolian Islands, and comprises about 80 digital stations equipped with Nanometrics Trillium broadband seismometers having an eigenperiod of 40 s. The signals are digitized at each station with 24‐bit resolution at 100 samples/s.

The seismic events considered in the analyses have been extracted from the “Catalogo dei terremoti della Sicilia Orientale ‐ Calabria Meridionale, INGV, Catania” (Gruppo Analisi Dati Sismici, 2017) within an area defined by the rectangle of latitude 35.90°–38.85° N and longitude 13.45°–16.85° E, with local magnitude (*M*_{L}) greater than 2.8 (maximum *M*_{L} 4.8; about 35% of the earthquakes have a magnitude ≥3.5), and focal depth of up to 35 km (Fig. 1d). We chose this interval of magnitude for the quality of data and homogeneity of the instrumental chain. The seismic events were located using the Hypoellipse code (Lahr, 1989) in a seven‐layer 1D velocity structure (Hirn *et al.*, 1991). A constant *V*_{P}/*V*_{S} ratio of 1.73 for travel‐time calculations was assumed. Figure 2a shows the distribution of records of our data set with respect to local magnitude and hypocentral distance. The starting data set has hypocentral distances from 6 to 377 km. However, 99% of the total data have been acquired at a distance of less than about 200 km; therefore, by considering a range up to this distance, we obtain a spatial sampling that is rather homogeneous and dense. Looking at Figure 2a, we can expect that no bias is introduced, because no evident trend between magnitude and hypocentral distance exists. Finally, Figure 2b shows the signal‐to‐noise ratio (SNR) for the *P* phases estimated over windows of 2 s wide. It shows that 97% of windows have an SNR greater than 5 dB, therefore suggesting a good quality of our data set.

As a first step in the processing, the data were checked visually at all stations to exclude traces with electronic glitches and with phenomena of amplitude saturation. The digitized velocity time histories of the three components of ground motion were first baseline corrected by removing the offset and the linear trends, and thus were instrument corrected. Therefore, we identified the *P* phases and their picks on the unfiltered vertical velocity components. As a final step, we integrated the velocity records to obtain the displacement time series, and to remove the low frequencies introduced by numerical integration, a high‐pass Butterworth filter with corner frequency of 0.075 Hz was applied to the data.

## Empirical Correlation Laws for Eastern Sicily

### Peak Ground Displacement, Magnitude, and Distance

Nakamura (1988), with the EEWS known as UrEDAS, was the pioneer in considering the first few seconds of recorded *P* waves to estimate the magnitude of a seismic event. The method of Nakamura consists of continually computing in real time the predominant period from the first 2–4 s of *P* waves for estimating the magnitude of the event. An alternative technique has been proposed by Wu and Zhao (2006) based on the use of the peak displacement amplitude *P*_{d}, measured on the 3‐s window starting from the *P*‐wave arrival‐time picking. They investigated the relationship between *P*_{d}, the hypocentral distance, and the local magnitude in southern California and Taiwan. They found that *P*_{d} can be used to estimate the magnitude of seismic events and can have a practical application in the EEWS. Independently, Zollo *et al.* (2006, 2007) showed that the peak displacement amplitude of the first few seconds of *P*‐ and *S*‐wave seismic signal scales with the earthquake magnitude for 4≤*M*_{w}≤7; it can be used for real‐time estimation of the earthquake size in EEW applications. Two important differences distinguish the approach by Zollo *et al.* (2006, 2007) from that by Wu and Zhao (2006): (1) the time window is not fixed to 3 s, and (2) the initial *S* waves are taken into account as well. Indeed, when a dense seismic network is placed around the potential source area, the *S*‐phases data, which are available at the stations closest to the epicenter, can be used to improve the magnitude estimation before the strong ground shaking reaches the distant target sites.

Following Lancieri and Zollo (2008), *P*‐ (*P*_{d}*P*) and *S*‐wave (*P*_{d}*S*) peak displacement amplitudes are measured on the modulus of displacement defined as (1)in which Z, EW, and NS are the vertical, east–west, and north–south components of ground motion, respectively. Unlike *P*‐waves picking, the onset of *S* waves has been estimated from travel time of *P* waves by assuming a ratio of *V*_{P}/*V*_{S}=1.73 (*V*_{P} and *V*_{S} are the *P*‐ and *S*‐wave velocities, respectively). Therefore, we measure *P*_{d} on time windows of 2 and 4 s of *P* waves (denoted as 2*P* and 4*P*, respectively) and 2 s for the *S* wave (denoted as 2*S*) starting from the *P*‐ and *S*‐waves picked arrivals.

The logarithmic *P*_{d} is generally assumed to be related to magnitude (*M*) and hypocentral distance (*R*) through the standard attenuation expression: (2)(Wu and Zhao, 2006; Zollo *et al.*, 2006; Lancieri and Zollo, 2008), in which *b* and *c* are the coefficients describing the magnitude dependence and the exponent in the distance‐dependent amplitude decay (i.e., the geometrical attenuation, assumed constant in all the investigated distance range), respectively. The model in equation (2) does not include the term representing the anelastic attenuation, that is, the linear *R* term. This is generally removed from the model because it was found not to be statistically significant (Wu and Zhao, 2006). Before testing this assumption for our data, we need to make a number of points that emerge from the distribution of log_{10}(*P*_{d}) as a function of hypocentral distance. In Figure 3, the log_{10}(*P*_{d}) measured on time windows of 2 s (2*P* and 2*S*) and 4 s (4*P*) are plotted versus the hypocentral distance for three narrow ranges of magnitude, 2.8–3.0, 3.6–3.8, and 4.6–4.8, respectively. In doing so, we reduce the scatter in *P*_{d} amplitudes due to the different source sizes. At distances of less than about 60 km, essentially corresponding to attenuation of direct waves and for which the effect of the anelastic attenuation should be smaller, the log_{10}(*P*_{d}) values decay at a rate higher than those shown by log_{10}(*P*_{d}) values at larger distances, for all the three considered magnitude intervals. In fact, beyond about 60 km, the rate of decay of the log_{10}(*P*_{d}) is less severe due to the arrival of energy refracted and reflected from the deeper parts of the crust. This means that if we consider a wide range of hypocentral distance, the coefficient *c* in equation (2) cannot be supposed as constant. Therefore, to estimate the relationship between peak ground displacement, local magnitude, and distance, only records with a maximum hypocentral distance of 60 km have been used. The number of records considered thus drops to 3928 three‐component seismograms. This choice is also based on the general observation that the crustal seismic events have high‐frequency direct body waves with dominant amplitude at distance from the receiver comparable with earthquake rupture length (Zeng *et al.*, 1993). Figure 4 describes the distribution of the number of three‐component seismograms as a function of magnitude that we used to perform further analysis.

We investigate the attenuation due to geometrical spreading and anelastic attenuation reformulating the equation of log_{10}(*P*_{d}) as follows: (3)in which *dR* represents the anelastic attenuation. A robust linear regression has been applied to the data using an iteratively reweighted least‐squares algorithm with a bisquare‐weighting function. In Table 1, the regression coefficients, together with their standard errors for 2*P*, 4*P*, and 2*S*, are reported for both attenuation models. *P*_{d} is in meters and *R* is in kilometers. The coefficient of elastic attenuation *d* was found not to be statistically significant, with values very close to zero and also positive in the case of 2*P*. Additionally, its introduction in the attenuation model does not improve the fit of the data as evidenced by the root mean square errors and the coefficient of determination values (see Table 1 for comparison). It was therefore removed from the model.

A residual analysis has been performed to verify whether the regression models are able to explain as much variation as possible in the dependent variable, assuming that the random error is uniquely distributed over the data set. In particular, the regression analysis can be considered successful in explaining the variation of the dependent variable if the residuals are unstructured and small. Otherwise, the validity of regression is questionable because the residuals are correlated to two independent variables. Figure 5a–c shows the residuals as a function of the hypocentral distance and magnitude. It is clear from the figure that the residuals do not show any significant trends, both versus magnitude and hypocentral distance. Additionally, more than about 90% of residuals are in the −0.5 to 0.5 range for the three considered time windows.

Equation (2) has been used to correct for how distance affects the observed peak amplitudes, by normalizing them to a reference distance of 30 km, a value chosen since it approximates the mean epicentral distance for the considered data set. As shown in Figure 6, there is an evident positive correlation between the logarithmic peak displacement normalized to 30 km (log_{10}(*P*_{d})^{30 km}) and the local magnitude for 2*P*, 4*P*, and 2*S* time windows in the whole investigated magnitude range.

For each magnitude value, we first calculated the mean and the standard deviation of log_{10}(*P*_{d})^{30 km} for both *P* and *S* waves. Therefore, a linear regression line was computed for the means of log_{10}(*P*_{d})^{30 km} weighted by the inverse of standard deviation (*σ*) as (4)(Zollo *et al.*, 2006). The means of log_{10}(*P*_{d})^{30 km} are shown in Figure 6 (black dots), whereas the estimated coefficients *a*′ and *b*′ are listed in Table 2, together with the calculated weighted standard errors (WSEs) that have been computed as (5)in which *w*_{i}=1/*σ*_{i} for the *i*th value of magnitude. The peak amplitudes are log linearly correlated with magnitude in the considered magnitude range (2.8≤*M*_{L}≤4.8) for both *P* and *S* waves, with correlation coefficients greater than 0.98 even for very short time windows from *P*‐wave arrivals.

*τ*_{c} and Magnitude

Nakamura (1988) was the first to develop a method for rapidly estimating the magnitude of an earthquake for EEW purposes using the frequency content of the first *P* wavetrain. Nakamura’s approach is based on the computation of the predominant period for the first *P* wavetrain taking into account the vertical component of ground motion. It has been widely applied to data from both broadband and strong‐motion stations in several seismic regions, demonstrating that the predominant period scales with seismic event magnitude (Allen and Kanamori, 2003; Olson and Allen, 2005; Lockman and Allen, 2007), and up to a few hundreds of kilometers from the seismic source it is independent from the epicentral distance (Allen and Kanamori, 2003; Allen, Brown, *et al.*, 2009; Allen, Gasparini, *et al.*, 2009).

An alternative method has been developed by Wu and Kanamori (2005a) based on the computation of the characteristic period *τ*_{c} defined as (6)in which *u*(*t*) and *v*(*t*) are the ground‐motion displacement and velocity, respectively. The time window of integration starts at the *P*‐wave onset time and has a duration equal to *τ*_{0}, generally set to 3 s. *τ*_{c} is considered to represent the average period of the *P*‐wave signal, and several studies have shown that it reflects the sizes of earthquakes (Kanamori, 2005; Wu and Kanamori, 2005b). Moreover, *τ*_{c} is less affected by the filter parameters and pre‐event noise than the predominant period for the first *P* wavetrain because it is estimated on the actual *P*‐wave window (Shieh *et al.*, 2008).

For the estimation of *τ*_{c}, we considered the ground‐motion‐filtered (high‐pass filtered at 0.075 Hz) displacement *u*(*t*) and velocity *v*(*t*) from the vertical component of ground motion. *τ*_{0} has been set to 3 s.

In the model considered here, the parameter *τ*_{c} depends only on the source characteristics and not on the distance. To verify this assumption, we plot *τ*_{c} as a function of the hypocentral distance in Figure 7a. Looking at the figure, we observe that *τ*_{c} does not show any significant trend with the distance, at least up to 60 km, as further confirmed by *t*‐test with significance threshold equal to 0.05.

As before, a linear regression was estimated for the means of log_{10}(*τ*_{c}) computed for each value of magnitude (Fig. 7b), weighted by the inverse of standard deviation. It has the following equation: (7)in which *τ*_{c} is measured in seconds, suggesting that the average log_{10}(*τ*_{c}) values increase with increasing magnitude. The uncertainties associated with the two coefficients of the model are the 95% confidence intervals. In Figure 7c, we compare our relationship with those obtained by Zollo *et al.* (2010) considering the data from south of Italy, Taiwan, and Japan, and Carranza *et al.* (2013) with the data from south of the Iberian Peninsula, southeast Iberia, and North Africa. If we put our attention on similar magnitude values and take the uncertainty of the predicted log_{10}(*τ*_{c}) into account, we notice that the regression through our data yielded quite comparable results to the findings of both Zollo *et al.* (2010) and Carranza *et al.* (2013), even though the investigated ranges of magnitude are different. However, the relationship by Carranza *et al.* (2013) suggests a closer dependence of the period parameter *τ*_{c} on the magnitude with respect to both our empirical laws and those from Zollo *et al.* (2010). These differences could be attributed to the characteristics of the used data set. In particular, Zollo *et al.* (2010) selected waveform records of events essentially occurring at a depth of less than 50 km and acquired at less than 60 km hypocentral distance, as in our case. However, their magnitude range spanned 4–8.5. Conversely, Carranza *et al.* (2013) considered seismic events with magnitude ranging from 3.8 to 5.9 but recorded at regional distances (of up to 700 km).

### PGV versus *P*_{d}

Wu and Kanamori (2005a) showed that *P*_{d} is correlated with the PGV at the same site, and when *P*_{d}>0.5 cm, the event is most likely able to produce damages. Therefore, in real time, the measured *P*_{d} and *τ*_{c} can be used to calculate the level of shaking (that is PGV) at the target sites and *M*, respectively, even though *M* is not directly used for on‐site EEW purposes.

In Figure 8a, the PGV values, measured as the maximum amplitude between the two unfiltered horizontal components of ground‐motion velocity, then considering the entire seismograms, are plotted as a function of peak displacement *P*_{d}, measured from the high‐pass‐filtered displacement records over a 3‐s time window after the *P* wavepick. The figure shows that, overall, the PGV values increase logarithmically with *P*_{d} in the investigated range of magnitude, in agreement with the findings of several previous studies (e.g., Wu *et al.*, 2007; Zollo *et al.*, 2010). Again, considering a maximum distance of 60 km, we obtain the following best‐fit regression line: (8)in which the units of PGV are cm/s and *P*_{d} are cm. The standard deviation of log_{10}(PGV) is 0.27, whereas the coefficient of determination is 0.80. In Figure 8b, we compare our PGV–*P*_{d} relationship with the others calibrated for several areas worldwide by Wu *et al.* (2007), Zollo *et al.* (2010), and Carranza *et al.* (2013). The comparison suggests that our data distribution is consistent with the empirical regression lines obtained by these authors, independently of the considered maximum distance and magnitude ranges.

## Discussion

In this study, we estimated empirical scaling relationships between the EEW parameters *P*_{d} and *τ*_{c} and both magnitude and PGV, using the broadband velocity seismograms of earthquakes occurring in eastern Sicily. The data have been acquired by the stations of the currently operating network in the area whose distribution ensures a good distance and azimuthal coverage (Fig. 1c).

The *τ*_{c}–*M*_{L} empirical scaling law estimated with our data distribution proved slow magnitude dependence of log_{10}(*τ*_{c}) because the slope is equal to 0.143 (±0.07). However, if we consider the uncertainty, its value is comparable to one obtained by Zollo *et al.* (2010) (slope equal to 0.21±0.01) using strong‐motion data recorded in the south of Italy, Taiwan, and Japan despite the differences in the magnitude ranges covered by the data (see Fig. 7c). We found that our *τ*_{c} shows a worst scaling relation with magnitude, as suggested by the standard deviation and correlation coefficient values (see Fig. 7b). Most likely, this can be associated with the strong sensibility of *τ*_{c} to the SNR of the used records. This represents the most critical aspect of the estimated empirical law and could make *τ*_{c} not sufficiently reliable and an EEW indicator less robust than *P*_{d} for the size of an earthquake.

The estimated PGV versus *P*_{d} relationship is nearly identical to those obtained from data of various regions around the world (see Fig. 8b). This suggests that, despite the scatter of the data around the mean, the correlation between PGV and *P*_{d} is independent of effects such as source, attenuation, site response, or tectonic regime (Zollo *et al.*, 2010; Carranza *et al.*, 2013). The uncertainty bounds associated with the regression lines take into account the potential differences due to the regional context or earthquake mechanisms.

To understand how the estimated scaling laws work, we selected some events that are not included in the data set used to discover the predictive relationships as test data. In particular, we considered 20 events ranging from magnitude *M*_{L} 2.8 to 4.3 (see Table 3) and compared the observed *P*_{d}, *τ*_{c}, and PGV values with the predicted ones. In Figure 9, we plot the difference between the base‐10 logarithms of observed *P*_{d} (*P*_{dobs}) values, for 2*P*, 4*P*, and 2*S* time windows, with the predicted *P*_{d} (*P*_{dpred}) by equation (2), as a function of the hypocentral distance. We note that the differences do not exhibit a structure and appear to be random, suggesting that there is no correlation with the hypocentral distance. Moreover, more than 65% of the observed *P*_{d} are predicted with a difference less than 50%.

The comparison between the average of the observed log_{10}(*τ*_{c}) values and the predicted ones by equation (7) for the 20 earthquakes belonging to the test data set is shown in Figure 10, where the range of one standard deviation is also reported. Taking into account the uncertainty of magnitude estimation (on average ±0.2 units) as well, we can see that the average of the observed *τ*_{c} values are within the predictive uncertainty bounds (±1 std. dev.). However, the figure also demonstrates the large variability of *τ*_{c} values for a given magnitude, making less reliable and more uncertain the prediction of earthquake size.

We compare the predicted and observed PGV of the 20 test events in Figure 11, where the best‐fit line is also shown. It can be seen that the observed PGVs are quite well explained by the model represented by equation (8), and thus it can be considered reliable for predicting future PGV values. For example, for the two earthquakes of *M*_{L} 3.5 (see Table 3), equation (8) predicts log_{10}(PGV) (with PGV in cm/s) from −2.3 to −0.9 with a mean value of −1.8±0.26. Within the uncertainty, these values are consistent with the observed log_{10}(PGV) from −2.5 to −0.9, with a mean of −1.8±0.31. Similarly, for the events in Table 3 with *M*_{L} 4.1, the log_{10}(PGV)_{pred} estimates from −1.6 to −0.1 (mean equal to −1.0±0.4) accord well with the log_{10}(PGV)_{obs} from −1.8 to −0.1 (mean equal to −0.9±0.4). These results confirm the robustness of *P*_{d} as a predictor of PGV for regional earthquake monitoring purposes and EEW operations in our region.

In Zollo *et al.* (2010), the on‐site approach to EEW is addressed through the so‐called threshold‐based method able to give an independent definition of alert levels at each recording site. This approach has the advantage that the potential damaging effects of the earthquake are evaluated without requiring accurate real‐time location of the event. This is particularly useful when the distribution of the stations is very sparse and does not guarantee an optimal azimuthal coverage to obtain an early and reliable location of the epicenter. In the threshold approach, the parameters *P*_{d} and *τ*_{c}, measured in a 3‐s time window after the first *P*‐arrival time at each station, are compared with *a priori* selected threshold values that define four alert levels (from 0 to 3; for details, see Zollo *et al.*, 2010) inside a decisional table. These levels are connected to both the expected on‐site damages and to the damages at distance. Considering a threshold value of macroseismic intensity I_{MCS} for damage effects equal to VII, we can estimate the PGV expected using the regression relationships by Faenza and Michelini (2010). The MCS intensity scale is used in Italy to describe the effects of the earthquake ground shaking on the built environment and communities. For I_{MCS}=VII, the Faenza and Michelini (2010) empirical law predicts PGV=6 cm/s, which we can convert into *P*_{d} threshold value using equation (8) and taking one standard deviation, obtaining *P*_{d}=0.1 cm. The *τ*_{c} threshold can be estimated through equation (7) for a minimum magnitude value fixed as threshold. On the basis of the seismic history of the region, we can select *M*_{L} 5, estimating a *τ*_{c}=0.3 s obtained, again, considering one standard deviation. In case there are a certain number of near‐source stations where *P*_{d} and *τ*_{c} exceed the threshold values, the real‐time mapping of alert levels can be used to predict the PDZ (Zollo *et al.*, 2010). This is particularly important in guaranteeing an efficient planning of rescue operations during emergency phases immediately after an earthquake.

The great number of data we used in this study, acquired by the current seismic network deployed in eastern Sicily, ensured an appropriate sample size for the robustness and accuracy of the empirical laws we estimated. This is an important aspect because the reliability of the predicted ground shaking depends, first of all, on the accuracy of the attenuation law applied to estimate it. At the same time, we believe that the only limiting factor could derive from the lack of events with large magnitude (higher than 5) for which we do not have any instrumental data. In the meantime, the estimated relationships will be useful during the phases of implementation and testing of the prototype EEWS in eastern Sicily in terms of promptness of detection, ground‐motion parameters, and magnitude estimations. In particular, for the EEW purpose in eastern Sicily, the so‐called INGV Network del Sud radio‐link communication network has already been realized. This network is composed of several digital microwave radio links, mainly deployed in eastern Sicily, connected with the INGV Etnean Observatory and Centro Unificato Acquisizione Dati in Catania, the Nicolosi center at Etna, the Lipari Observatory in the Aeolian Archipelago, the Gibilmanna Observatory, and the INGV Palermo. The seismic signals will be transmitted in the 7 and 13 GHz bands with a bandwidth greater than 100 Mbps in full duplex and with a latency less than or equal to 3 ms. The system will be operational when the granting of frequencies is authorized.

## Concluding Remarks

Today, the development of EEWSs represents one of the most useful strategies to mitigate seismic risk in short timescales, and many countries worldwide are promoting and developing such systems. In the frame of seismic risk management, it is considered a reasonable costly solution for the loss reduction. Additionally, the developments of the RTS are opening new scenarios in the framework of interaction of EEW and earthquake‐engineering applications (i.e., Fujita *et al.*, 2011; Kubo *et al.*, 2011; Maddaloni *et al.*, 2011; Nakamura *et al.*, 2011).

Evaluating the practicability of an EEWS in this area is justified by its high level of seismic hazard. Moreover, three big oil refineries and power plants are installed along the eastern coast of Sicily in the cities of Milazzo, Augusta, and Priolo Gargallo, which have the refining of crude oil and its derivatives as their main activity. For them, suitable safety measures such as, for example, the automatic blocking of pipelines or gas lines to prevent fire hazards or the automatic shutdown of the manufacturing operations to avoid equipment casualty, could be adopted for damage reduction. Therefore, the introduction of an EEWS within the practices of the frame of RTS that have been regularly carried out for years in eastern Sicily is a worthwhile objective, because it can be effectively used to reduce the damage caused by the strongest earthquakes.

Taking into account the distribution of the major earthquakes in eastern Sicily, it could be expected that in many cases the regional approach will not give enough time to process the data and divulgate the alarm. Some problems may arise, above all, for the events occurring offshore. For them, more coastal stations might be needed to better constrain the earthquake location. However, for these events, the on‐site threshold‐based EEW approach can issue an alert rapidly to the inland target sites and estimate a potential damage zone within very few seconds (2–3 s) from the origin of the seismic event, increasing the lead time and reducing the blind zone.

In conclusion, a seismic network that includes the real‐time processing of seismic recordings will hopefully be developed in eastern Sicily because it can also be used as a tool to predict ground‐motion measurements in real time and allow any emergency response to be carried out quickly.

## Data and Resources

Seismograms used in this work can be obtained from the European Integrated Data Archive (EIDA) at www.orfeus-eu.org/data/eida/ (last accessed February 2017). Information regarding the recent seismicity in the investigated area is from Castello *et al.* (2006).

## Acknowledgments

We are grateful to Editor‐in‐Chief Thomas Pratt and two anonymous reviewers for their very helpful and constructive comments. We acknowledge the useful suggestions and remarks of Stefano Gresta on the preliminary research. We also thank Eugenio Privitera for the encouragement and support of this study. This work has been supported by the following projects: VULCAMED “Potenziamento strutturale di centri di ricerca per lo studio di aree VULCAniche ad alto rischio e del loro potenziale geotermico nel contesto della dinamica geologica e ambientale MEDiterranea” and SIGMA “Sistema Integrato di sensori in ambiente cloud per la Gestione Multirischio Avanzata,” inside the 2007–2013 infrastructural Italian Program PON (Programma Operativo Nazionale, Ricerca e Competitività ‐ Asse I: “Sostegno ai mutamenti strutturali”); “Attività di sviluppo sperimentale finalizzata alla riduzione del rischio sismico nella Sicilia Orientale” inside the PO‐FESR 2007‐2013 Sicilia; MEDiterranean SUpersite Volcanoes (MED‐SUV) funded from the European Union Seventh Framework Programme (FP7) under Grant Number 308665. Our work benefited from the language correction support given by Stephen Conway.

- Manuscript received 28 July 2016.