- © 2016 by the Seismological Society of America
A low‐magnitude regional ground‐motion prediction equation (GMPE) for peak ground acceleration and 5% damped spectral acceleration at 11 oscillator periods ranging from 0.1 to 3 s was developed for applications of the earthquake early warning system operating in southern Italy. The accelerometric dataset was built out of 2270 waveforms from 319 earthquakes with local magnitude ranging from 1.5 to 4.2, recorded from 60 stations located on soil type B, according to the Italian Building Code and Eurocode 8, with hypocentral distances ranging from 3 to about 100 km. The GMPE coefficients were computed through a nonlinear weighted damped least‐squares algorithm, attributing a higher weight to ground‐motion data recorded at stations for which soil category was established using geotechnical and geophysical measurements. Predictions of the proposed GMPE were tested against independent ground‐motion data recorded in southern Italy and compared with other regional and global GMPEs through the log‐likelihood method. The comparison shows that the proposed GMPE performs better than any other GMPE considered in this study.
Tables of seismic stations and main features of selected GMPEs.
Historically, ground‐motion prediction equations (GMPEs) have been developed based on data from large earthquakes (moment magnitude Mw>∼5), as shown by Douglas (2016) in his most recent comprehensive worldwide summary of attenuation relationships developed in the last 50 years. It has been well established that empirical GMPEs generally do not extrapolate reliably to smaller magnitudes (e.g., Bommer et al., 2007). Therefore, in the last few years, the development of GMPEs for low‐magnitude earthquakes has grown rapidly. Several small‐magnitude GMPEs have been published, for example Bommer et al. (2007), Massa et al. (2007), and Chiou et al. (2010). It is important to point out that regional differences in ground‐motion characteristics become more apparent at small magnitudes (e.g., Chiou et al., 2010).
A regional GMPE for southern Italy (Campania and Basilicata regions) was developed by Emolo et al. (2011) using horizontal peak ground acceleration (PGA) and peak ground velocity (PGV) from events with local magnitudes (ML), ranging between 1.5 and 3.2. Even though it was constructed using data recorded by the local seismic network, named Irpinia Seismic Network (ISNet), this GMPE is not suitable to be used in the earthquake early warning (EEW) system currently implemented in southern Italy for reasons that will be discussed hereinafter.
The southern Italy EEW system aims at protecting critical facilities located in a region where large and destructive earthquakes have occurred in the past, such as, for instance, the 1980 Mw 6.9 Irpinia earthquake. It is based on the ISNet, managed by analysis and monitoring of environmental risk (AMRA, see Data and Resources) and deployed since 2005 along the southern Apennines belt in Campania and Basilicata regions (Weber et al., 2007; Zollo et al., 2009). The system uses a software platform named PRobabilistic and Evolutionary early warning SysTem (PRESTo, Zollo et al., 2014), which continuously processes the real‐time stream of ground‐motion data recorded by the ISNet stations. When a seismic event occurs, PRESTo promptly detects the location of the epicenter jointly with an estimate of the magnitude and ground‐motion intensity at the sites to be protected. These estimates are continuously updated as new ground‐motion data become available. When the ground motion predicted at the target sites exceeds an assigned threshold, an alarm is issued.
PRESTo is designed to handle both low‐magnitude and moderate‐to‐large earthquakes using two different sets of operational parameters depending on the magnitude (e.g., signal filters, regression laws, thresholds). This allows on one hand to test the real‐time response of the system using the continuous, low‐intensity activity background (Mw<3) occurring in the Irpinia region (Iannaccone et al., 2010) and on the other to issue early warnings in case strong earthquakes would strike in the future (e.g., Zollo et al., 2014). The ground motion at the target sites is assessed in real time from the estimated epicenter location and magnitude through GMPEs. PRESTo can use two different GMPEs for low and moderate‐to‐large earthquakes, respectively, which can be specified by the user.
The GMPE by Emolo et al. (2011) is not suitable to be used within PRESTo as low‐magnitude GMPE for two main reasons. First, it was developed only for PGA and PGV, despite the most appropriate intensity measure to characterize ground motion for earthquake engineering applications of EEW systems is an ordinate of the acceleration response spectrum (Convertito et al., 2008). The second motivation is due to its upper magnitude bound, which is limited to ML 3.2. This value is significantly smaller than the lower bound limitations of most GMPE for strong ground motion, which typically have a minimum magnitude larger than 4 (e.g., Bindi et al., 2011, for Italy). Therefore, the use of the Emolo et al. (2011) GMPE in the EEW system would generate a magnitude gap in the computation of GMPE passing from low‐magnitude earthquakes to moderate‐to‐large events, which should be filled by extrapolation.
The aim of this work is to illustrate a GMPE developed to overcome the aforementioned limitations associated with the use of the Emolo et al. (2011) GMPE in the EEW system operating in southern Italy. The study was carried out in the framework of the Tools and Technologies for Risk Management of Transportation Infrastructures (STRIT) project, within a specific work package devoted to the development of tools to be used in the EEW system operating in southern Italy.
A ground‐motion database was purposely built up by gathering accelerograms from low‐magnitude earthquakes mainly recorded in the Campania and Basilicata regions, extending the database used by Emolo et al. (2011). A significant effort was made to define ground conditions at the recording sites. Almost all seismic stations turned out to be located on site class B, according to Italian Building Code (NTC08, 2008) and Eurocode 8 (2003; hereinafter, EC8). Thus, the regional GMPE was developed for PGA and 5% damped spectral acceleration (SA) at 11 oscillator periods from 0.1 to 3 s, considering stiff ground conditions (site class B). Predictions of the proposed GMPE were compared with independent ground‐motion data (i.e., data not used for the development of the GMPE) and other empirical GMPE models, both regional and global, showing that the proposed GMPE seems to perform better than the others considered in the study.
Ground‐Motion Data and Signal Processing
Ground‐motion data were gathered from 1131 earthquakes with ML ranging from 1.5 to 4.4. To avoid the inclusion in the final accelerometric dataset of events generated in different tectonic settings (e.g., crustal and subduction earthquakes), the maximum focal depth was cut at 30 km; this is consistent with the highest class of effective depth (12–20 km) associated with the seismogenic zones that were identified in the area of study (Meletti et al., 2008). Figure 1 illustrates the time and magnitude ranges covered by each database used as source for assembling the accelerometric ground‐motion data (namely SeismNet Manager and ITalian ACcelerometric Archive [ITACA] 2.0, see Data and Resources for more details).
Among different magnitude scales, the local magnitude was used because it is the reference magnitude in ITACA for Mw<5.0 earthquakes (Luzi et al., 2010). In fact, most of the events of the ITACA accelerometric database are associated with ML estimate only. It must be clarified that the ISNet local magnitude scale (Bobbio et al., 2009) has a different formulation from the National Institute of Geophysics and Volcanology (INGV) local magnitude scale, which is that used in ITACA. However, Bobbio et al. (2009) demonstrated that the two magnitude scales do not differ substantially in the magnitude range considered herein, and therefore they can be used in the same dataset without introducing a significant bias.
Following Emolo et al. (2011), uncorrected recordings were preliminarily processed for a linear detrending and to remove the mean acceleration. Subsequently, a band‐pass filter (a four‐pole Butterworth filter in the range 0.075–20 Hz) and a 2% cosine taper window were applied to the signals. For each seismogram the signal‐to‐noise ratio (SNR) was then computed to exclude records having a dominant noise contamination. The SNR was computed according to the procedure proposed by Vassallo and Cantore (2010), which is based on comparing the pre‐event noise amplitude with respect to a portion of the signal centered at the time of occurrence of maximum amplitude. Only records with SNR greater than or equal to 10 in both the horizontal components of ground motion were selected. Furthermore, signals recorded before 2002 coming from ITACA 2.0 were rejected because most of them were identified as late‐triggered.
Definition of Ground Category at the Seismic Stations
It is well known that earthquake ground motion is strongly affected by the geological setting and geotechnical properties of soil deposits at the recording sites. Several building codes worldwide (e.g., EC8, NTC08) take into account the role played by local site conditions in defining the seismic action typically via a simplified approach based on introducing appropriate ground categories. These are typically defined using VS30, the average shear‐wave velocity in the upper 30 m of subsoil. At stable sites, ground categories are divided into five (NTC08; EC8): A for VS30≥800 m/s, B for 360 m/s≤VS30<800 m/s, C for 180 m/s≤VS30<360 m/s, and D if VS30<180 m/s. Category E includes soil profiles consisting of a surface alluvium layer with VS30 corresponding to type C or D subsoil and thickness less than 20 m, underlain by stiffer material with VS>800 m/s.
The total number of investigated stations is 87. ITACA 2.1 specifies the site classification for most of them. However, it has been established based on VS30 only at few stations, whereas for the remaining stations the site class was defined using geological data only. On the contrary, SeismNet Manager does not include any information on the subsoil conditions at the sites where the stations are located. The classification adopted by Emolo et al. (2011), named Quaternary‐Volcanic‐Tertiary‐Mesozoic (QVTM), was obtained using geological information.
To increase the number of stations geotechnically characterized based on VS30, geotechnical, geophysical, and geological data were collected for the area under investigation, including 103 direct measurements of VS. They were retrieved from 70 downhole (DH), 15 crosshole (CH), and 18 noninvasive geophysical tests, such as multichannel analysis of surface waves (MASW), refraction microtremor (ReMi), extended spatial auto correlation (ESAC), and frequency–wavenumber (f‐k) passive methods. Less than 30% of VS profiles were downloaded from ITACA 2.1, whereas the majority of the ground‐motion data were gathered from various sources, including scientific publications (e.g., for the area surrounding Naples see, among others, Nunziata et al., 2004), urban plans at the municipal level, and reports concerning the design of strategic infrastructures (e.g., retrofit of a railway tunnel crossing the Municipality of Somma Vesuviana; see Data and Resources for more details). The collected data were digitized and processed within a purposely developed Geographic Information Systems (GIS) database (Fig. 2). This allowed the analysis of geological information jointly with the data obtained from field testing to come up with an approximate description of the subsoil conditions at most station sites. Interpretation of the acquired data led to the definition of soil category based on VS30 for eight stations belonging to the ISNet (BENI, BSC3, CLT3, MRN3, NAPI, PGN3, SALI, and STN3).
At the end, the number of stations for which soil category was defined based on VS30 measurements is 28. For the remaining 59 stations, the soil category was established using the classification based on geological information from ITACA 2.1. When the latter was not available, the classification provided by AMRA, also based on geological information, was adopted. The stations selected for the study and the associated soil categories are listed in Ⓔ Table S1 (available in the electronic supplement to this article). Most of the selected stations (61) turned out to be located on soil category B, 14 on soil type A, 11 on soil type C (one of which, SLC1, is located inside a building), and 1 (SGR) on soil type E. Most of the accelerograms (93%) were recorded at stations classified as soil type B, 4% as soil type A, 3% as soil type C, and only one recording as soil type E.
The constructed dataset did not allow the performance of reliable regressions taking into account a subdivision of stations based on EC8, due to the limited number of recordings from stations on soil classes A, C, and E. Therefore, it was decided to limit the development of the regional GMPE only for soil type B.
At the end, the original dataset was reduced to 2270×2 waveforms (only the two horizontal components were considered) associated with 319 earthquakes with ML ranging from 1.5 to 4.2 and hypocentral distances Rhypo from 3 to 100 km. Ⓔ Table S1 lists the number of recordings distributed among the 61 seismic stations located on soil class B. The epicenters of the selected earthquakes are shown in Figure 3, along with the two major fault systems controlling the background low‐seismic activity of the area under study: a sequence of northwest–southeast (NW–SE) trending normal faults (along the Apenninic chain) and approximately east–west (E–W)‐oriented strike‐slip faults, transversely cutting the chain (De Matteis et al., 2012). The recordings were selected to obtain a distribution in magnitude–distance space as homogeneously as possible. The distribution of events as a function of ML and Rhypo is shown in Figure 4. The depth of the earthquakes varies between 1 and 28 km.
FUNCTIONAL FORM AND REGRESSION
The GMPE proposed in this study has the following functional form: (1)in which Y (log is in base 10) is the intensity measure to be predicted (in m/s2), M is the local magnitude, and R is the hypocentral distance (in km). The geometric mean of horizontal components of PGA along with 5% damped SA computed over 11 oscillator periods (0.1, 0.2, 0.3, 0.4, 0.6, 0.8, 1.0, 1.5, 2.0, 2.5, and 3.0 s) are the calculated intensity measures. The coefficients a, b, and c were computed through a nonlinear regression analysis using the Levenberg–Marquardt algorithm, also known as the damped weighted least‐squares method, implemented by the MATLAB function nlinfit (see Data and Resources). A weighted version of this algorithm was used to give higher weights to accelerometric data recorded at station sites for which VS profiles were available. Specifically, weights of 0.7 and 0.3 were associated with the latter and remaining recordings, respectively. Finally, σlogY is the standard error associated with the logY variable. Even though two different faulting styles characterize the low‐activity seismic background in the studied area (normal and strike slip), it was decided not to consider a specific term in the GMPE model taking into account the focal mechanism, due to the inherent difficulties in associating a focal mechanism to low‐magnitude earthquakes immediately after the occurrence (it is worth noting that the proposed GMPE has been developed for EEW applications).
The coefficients and standard errors for the GMPE proposed in this study are given in Table 1.
PERFORMANCE ASSESSMENT OF THE PROPOSED GMPE
Predictions of the proposed GMPE were tested against independent ground‐motion data and compared with other GMPEs from the literature by means of the log‐likelihood (LLH) method proposed by Scherbaum et al. (2009), which allows assessing the relative performance of various GMPEs against a ground‐motion dataset. The algorithm is based on the probability that an observed ground motion is actually realized under the hypothesis that a model is true (Beauval et al., 2012). The negative average LLH (Delavaud et al., 2009) measures the distance between a model and the data‐generating distribution as (2)in which N is the number of observations xi, and is the probability density function (assumed to have a normal distribution) predicted by the GMPE. The smaller the value, the closer is the candidate GMPE to the model that has generated the data.
According to Scherbaum et al. (2009), the LLH values should be computed using a dataset different from the one used to develop the GMPE. Therefore, an independent dataset was built by collecting ground‐motion data recorded by the ISNet in 2015. The available data were processed using the same procedure as previously described. The corresponding dataset is composed of 262 good‐to‐high‐quality accelerograms from 41 earthquakes with depth<30 km, recorded at stations located on class B soil. The magnitude–distance distribution is shown in Figure 5, both in terms of ML and Mw (see Data and Resources) to allow a comparison with global GMPEs, which typically are based on Mw.
Predictions of the proposed GMPE were compared with those of other regional GMPEs (Frisenda et al., 2005 [hereafter, FRI05]; Massa et al., 2007 [hereafter, MAS07]; Emolo et al., 2011 [hereafter, EMO11]) and global (Bindi et al., 2014a,b [hereafter, BIN14]; and Cauzzi et al., 2015 [hereafter, CAU15]).
A summary of the main features of the adopted GMPEs is shown in Ⓔ Table S2 (available in the electronic supplement to this article). It is important to note that CAU15 was developed using the rupture distance (RRUP) as distance metric. However, RRUP can be approximated to Rhypo for Mw<5.7 (Cauzzi et al., 2015), that is, in the magnitude range considered in this study. EMO11 proposed a station‐dependent functional form; thus ground motion was predicted herein only at the 21 stations considered in that study. Except EMO11 and FRI05, the remaining GMPEs were extrapolated below their magnitude range of validity.
The performance assessment of the aforementioned GMPEs against independent data is shown in Figure 6, in which the LLH value is plotted as a function of the structural period. It is apparent that the GMPE developed in this study provides the lowest LLH values (<1.0) for the entire period range, which means that it performs better than any other considered GMPE. Low values are also obtained for PGA for the remaining regional GMPEs. Among them, only MAS07 predicts SA for periods different from zero (i.e., PGA). Despite the fact that LLH values for PGA are close to those computed for the GMPE proposed in this study, a significant deviation can be observed starting from T>0.5 s. The two global GMPEs show the highest LLH values (up to 5.6). LLH<1.0 are found only in two different limited period ranges: 0.2<T≤1.0 s for BIN14 and 1.25<T≤2.0 s for CAU15. This supports the idea that GMPE models based on moderate‐to‐large magnitudes are not able to correctly predict the ground motion for low‐magnitude earthquakes.
Comparison among recorded ground‐motion data and the GMPEs considered in this study is also given in Figures 7–9 for two earthquakes that occurred in the study area, namely the 5 May 2015 ML 3.2 (Mw 3.1) and the 18 August 2015 ML 2.1 (Mw 2.3) events. Figures 7 and 8 show a comparison for PGA and SA at three structural periods (T=0.1, 0.3, and 1.5 s) as a function of the hypocentral distance, for the ML 3.2 and the ML 2.1 earthquakes, respectively. Figure 9 shows the predicted acceleration response spectra versus the response spectra of the recordings for two hypocentral distances for each earthquake. They confirm the outcome obtained from the LLH values. Furthermore, the oscillator period of the peak of the response spectrum is well captured by all the GMPEs, except by CAU15, which is shifted to higher structural periods. Figure 10 shows the standard deviation of the selected GMPEs as a function of the oscillator period. Concerning PGA, EMO11 has the largest standard deviation (0.417), whereas MAS07 has the lowest (0.282). The two global GMPEs exhibit a standard deviation close to that associated with the GMPE proposed in this study (0.354). The four GMPEs that predict SA for periods different from zero roughly converge to a peak around 0.1 s and diverge significantly for higher structural periods, except the two regional GMPEs (the one proposed in this study and MAS07), which exhibit a similar trend and SAs.
This article proposed a low‐magnitude regional GMPE to be used in the EEW system operating in southern Italy.
It was developed to overcome two major limitations of the Emolo et al. (2011) GMPE model, which currently represents the reference low‐magnitude GMPE for southern Italy. First, the proposed GMPE is based on a purposely developed ground‐motion dataset larger than that used by Emolo et al. (2011), because it includes also earthquakes recorded from 2010 to 2014 at ISNet stations as well as strong‐motion data from ITACA 2.0. This allowed the extension of the upper validity bound of the GMPE, previously limited to 3.2, to ML 4.2, thereby filling the gap between low‐magnitude and moderate‐to‐large magnitude GMPEs. Furthermore, whereas the GMPE proposed by Emolo et al. (2011) allows the prediction of peak ground‐motion parameters only (namely PGA and PGV), the model proposed in this study allows real‐time computation of SAs (up to an oscillator period of 3 s), which are more suited intensity measures to characterize the ground motion for earthquake engineering applications of EEW (Convertito et al., 2008).
Special attention was given to ground classification of the station sites according to NTC08 and EC8 building codes. Geotechnical, geophysical, and geological data were gathered for the area under investigation, and organized under a GIS platform, to define the soil category based on measured VS profiles for as many stations as possible. For the remaining stations, the classification was based on available geological information. It turned out that very few recordings are associated with soil categories different from B. This prevented the development of a GMPE with a soil‐type dependent term. As such, it was assumed statistically meaningful to generate a GMPE for soil category B only. The coefficients of the GMPE were computed through a nonlinear damped least‐squares weighed algorithm, attributing a higher weight to accelerometric data recorded at stations where VS profiles were available.
Finally, predictions of the proposed regional GMPE were compared against an independent set of ground‐motion data recorded in 2015 within the area of study. A comparison was also carried out with respect to other regional and global GMPEs. The goodness of the predictions against the observed ground‐motion data was assessed through the LLH parameter. The proposed GMPE exhibited the best performance when compared with others GMPE models, thus it is recommended for EEW applications in southern Italy.
DATA AND RESOURCES
The main source for retrieving the ground‐motion data used to carry out this study was SeismNet Manager (http://seismnet.na.infn.it, last accessed May 2016), a web‐based application that allows the handling of ground‐motion data acquired by the Irpinia Seismic Network (ISNet). Mw values were retrieved from the ISNet Bulletin (http://isnet.na.infn.it/cgi-bin/isnet-events/isnet.cgi, last accessed May 2016). The accelerometric data downloaded from SeismNet Manager (up to December 2014) were integrated with strong‐motion records downloaded from the ITalian ACcelerometric Archive, ITACA 2.0 (http://itaca.mi.ingv.it/, last accessed April 2015, Luzi et al., 2008; Pacor et al., 2011). VS measurements at some recording sites have been downloaded from ITACA 2.1 (http://itaca.mi.ingv.it/, last accessed May 2016). Subsoil data have been collected from various sources: data for the Municipality of Napoli have been downloaded from http://www.aracneeditrice.it/pdf/9788854811843.pdf (last accessed May 2016); for the Municipality of Somma Vesuviana from ftp://host52-163-static.138-193-b.business.telecomitalia.it/PA-1222/PA-1222_ELABORATI-1-di-2/53.pdf (last accessed May 2016); for the Municipality of Benevento from http://www.comune.benevento.it/ (last accessed May 2016); for the Municipality of Sant’Angelo dei Lombardi from http://www.comune.santangelodeilombardi.av.it/ (last accessed May 2016); and for the Municipality of Melfi from http://www.comune.melfi.pz.it/ (last accessed May 2016). The soil categories of recording sites based on geological information have been retrieved from ITACA 2.1 and provided by analysis and monitoring of environmental risk (AMRA, personal comm., 2015; http://www.amracenter.com/, last accessed February 2015). Maps have been prepared using ESRI ArcMap 9.1. The MATLAB functions are available from http://www.mathworks.com (last accessed February 2015).
The current research work has been carried out within the Tools and Technologies for Risk Management of Transportation Infrastructures (STRIT) project (Code PON01_02366), in the framework of the National Operational Programme for Research and Competitiveness 2007–2013 (NOP for R&C), cofounded with the European Regional Development Fund and national resources. We would like to express our gratitude to Iunio Iervolino (Università degli Studi di Napoli Federico II), Vincenzo Convertito (National Institute of Geophysics and Volcanology, INGV – Osservatorio Vesuviano), and Raffaella De Matteis (Università del Sannio) for their insightful discussions on some points of this work. We are also grateful to Claudio Martino (analysis and monitoring of environmental risk [AMRA]) for his assistance with the Irpinia Seismic Network (ISNet) sensors. Finally, we would like to acknowledge Sanjay Singh Bora, GFZ Potsdam, and an anonymous reviewer, whose comments significantly contributed to improve the article.