GNGTS – Atti del 19° Convegno Nazionale / 12.10 A. Ursino (1), H. Langer (1), L.Scarfì (1), G. Di Grazia (1) and S. Gresta (1,2) (1) (2) Sistema Poseidon, C.da Cava Sorciaro, Priolo-Gargallo (SR) Dipartimento di Scienze Geologiche, Università di Catania, Corso Italia, Catania DISCRIMINATION OF QUARRY BLASTS FROM TECTONIC EARTHQUAKES IN THE IBLEAN PLATFORM (SOUTHEASTERN SICILY) Abstract. The seismic network set up in the Iblean Plateau (Southeastern Sicily) in the framework of the POSEIDON project is aimed at the seismic surveillance of the zone, and in particular the identification of faults with enhanced activity. The seismic activity as inferred from the records of local events recorded in the time span September 1999-February 2000 showed an apparent concentration of events in the zone between Augusta and Siracusa. However, the heterogeneity in the distribution of events with daytime suggested that the seismicity maps are severely biased by artificial events, such as quarry explosions. We have distinguished between tectonic earthquakes and quarry blasts by the inspection of waveforms of certain key stations, and by spectral analysis. As a general rule we found that the local tectonic microearthquakes are richer in high frequencies than the quarry blasts. We tested our discrimination based on waveforms and spectra with a data of local events recorded in the time span September 1999-February 2000. All events identified as quarry blasts have occurred during daytimes between 08:00 a.m. and 03:00 p.m. GMT and on weekdays from Monday to Friday. Automatic discrimination was carried out in a straightforward way using artificial neural networks (ANN) in a supervised classification. The application of the ANN to our data set gave a success of about 95%. DISCRIMINAZIONE DI SCOPPI IN CAVA DA TERREMOTI TETTONICI NELLA PIATTAFORMA IBLEA (SICILIA SUD-ORIENTALE) Riassunto. La rete sismometrica installata nella piattaforma iblea (Sicilia sud-orientale) nell'ambito del progetto POSEIDON è finalizzata alla sorveglianza sismica dell'area ed in particolare all'identificazione di faglie attive. La sismicità registrata nel periodo settembre 1999-febbraio 2000 mostra un'apparente concentrazione di eventi nella zona fra Augusta e Siracusa, ma il loro aumento diurno suggerisce la presenza di eventi artificiali, come scoppi in cava. Gli eventi tettonici sono stati separati da quelli artificiali tramite l'analisi spettrale delle forme d'onda registrate da alcune stazioni. Abbiamo trovato che, in generale, i microterremoti tettonici presentano maggior contenuto di alte frequenza e che gli scoppi avvennero tra le 8 e le 15 dei giorni lavorativi. È stata definita una discriminazione automatica degli scoppi tramite l'uso di reti neurali artificiali che ha dato ottimi risultati. INTRODUCTION Southeastern Sicily, in particular the Iblean Plateau, belongs to the zones in Europe with the highest seismic potential. During the past millenium Southeastern Sicily was struck by several strong earthquakes (Fig. 1), which reached magnitudes of about M = 7.0 and caused some 10,000 deaths each. Actually, the Ioanian coast belongs to the most populated areas in Sicily, while between the cities of Augusta and Siracusa there are important industrial petrochemical facilities bearing an elevated secondary seismic risk. In the light of the considerable seismic potential, a permanent seismic network in the Iblean Plateau was set up in the framework of the POSEIDON project. This network actually consists of 9 digital three-components stations (Fig. 1) and is aimed at the seismic surveillance of the zone, and in particular the identification of faults GNGTS – Atti del 19° Convegno Nazionale / 12.10 with enhanced activity. The earthquakes recorded during the operation time of the network are of small magnitudes which did not exceed M = 4.0 yet. Despite the modest seismic energy release, the distribution of hypocenters may give useful seismotectonic hints such as the existence of active faults, orientation of seismic dislocations, the relation of seismicity to tectonic structures visible at the surface, and so on. The geometrical distribution, the technical specifications of the network together with the low noise level recorded at the stations permit a sufficient completeness of the catalog for events whose epicenters fall within the area represented shown in Fig. 1 and whose magnitudes are greater than 1.0. The errors of hypocentral coordinates calculated with the program HYPOELLIPSE (Lahr, 1989) are typically of the order of 2 km (or less) for the epicenter, and 5 km (or less) for the focal depth (Bollettino Sistema Poseidon, 2000). SEISMICITY DISTRIBUTION IN TIME AND SPACE The seismicity map in the Iblean Plateau inferred from events recorded in the time span September 1999-February 2000 (Fig. 2) showed an apparent concentration of activity along the coastal zone between the towns of Augusta and Siracusa. This heterogeneity makes us suspect that the data set is affected by the presence of artificial events, in particular quarry blasts. Quarry blasts are supposed to occur more frequently during working hours of the day. We therefore may investigate possible biases introduced to seismicity maps by filtering the data set with respect to day-time. SR2 SR1 1169 Lentini A SE < 6.5 < 6.0 < 5.5 < 5.0 < 4.5 N M M M M M A NI 1693 7.0 <= M < 7.5 6.0 <= 5.5 <= 5.0 <= 4.5 <= 4.0 <= IO MAGNITUDE 1990 SR9 Augusta SR3 1542 SR4 SR6 SR5 Siracusa BL EA N PL AT EA U 1693 I SR8 SR7 Ragusa SICILY CHAN NEL 0 10 20 Km Fig. 1 - Historical seismicity (open dots) in Southeastern Sicily. Epicenter coordinates were taken from Azzaro and Barbano (2000), coordinates of the event of 13 December 1990 from Amato et al. (1995). Black triangles represent station of the seismic network. GNGTS – Atti del 19° Convegno Nazionale / 12.10 AN NI IO A SE Lentini O RM Augusta IB LE AN PL A TF Siracusa Ragusa SICILY CHAN 10 0 NE L 20 Km Fig. 2 - Overall “seismicity” in the time span September 1999-February 2000. AN NI IO A SE Lentini Siracusa IB LE AN PL AT FO RM Augusta SICILY CHANN EL Ragusa 0 10 20 Km Fig. 3a - Distribution of the overall seismicity in the time span September 1999-February 2000 occurred between 03:00 p.m. and 08:00 a.m. All hours are in GMT. The hammer symbols give the location of quarries as reported by the Distretto Minerario of Catania. GNGTS – Atti del 19° Convegno Nazionale / 12.10 AN NI IO A SE Lentini Siracusa IB LE AN PL AT FO RM Augusta SICILY Ragusa CHANN EL 0 10 20 Km Fig. 3b - Distribution of the overall seismicity in the time span September 1999-February 2000 occurred between 08:00 a.m. and 03:00 p.m. All hours are in GMT. The hammer symbols give the location of quarries as reported by the Distretto Minerario of Catania. During the daytime between 03:00 p.m. and 08:00 a.m. GMT the quarry explosions can be supposed to be substantially absent. The activity during these hours shows a quite uniform distribution over the area covered by the seismic network, whereas the aforementioned concentration of events between Augusta and Siracusa widely disappears (Fig. 3a). Vice-versa, in the hours from 08:00 a.m. to 03:00 p.m GMT the majority of events occurs in the coastal range between Augusta and Siracusa, some more close to the town of Lentini (Fig. 3b). It is no surprise that many of these events were located close to quarries, even though some of these locations are affected by high uncertainties. A good deal of these events has waveforms which we found to be quite typical for quarry blasts, at least at certain stations. On the other hand we can provide a training set of signals, whose origin as tectonic microearthquakes is out of doubt since quarries do not launch explosions during night times. WAVEFORMS AND SPECTRA The superficial position of explosion sources together with the techniques of firing during quarry explosions should produce specific signal characteristics, which should help for discrimination purposes. In ripple firing quarries P-onsets are often emergent, clear S-waves are lacking, whereas Raleigh-waves are observed. In Fig. 4 the waveforms and spectra of a microearthquakes and of a quarry blast both, recorded at station SR6 are compared. We note indeed, for quarry blast, an emergent P-onset, and a clear S (in terms of a “secondary” waves) cannot be identified. A typical feature of explosion seismograms at this station is an envelope of GNGTS – Atti del 19° Convegno Nazionale / 12.10 1.E-03 Ground velocity (m/s) Ground velocity (m/s) fish-like shape. Amplitudes of the vertical component are dominating over the horizontal components ones. The quarry blasts close to the city of Lentini can be distinguished rather clearly using the station SR2, where we note a very sharp S-wave onset on all three components of the earthquake seismogram (Fig. 5). Here peak amplitudes of the horizontal components are about three times higher than one the vertical. For the explosion seismogram amplitudes of the vertical and horizontal components are of the same order. A first visual distinction between earthquakes and explosions thus may carried out more safely using stations SR2, SR6 as key stations. 5.E-04 0.E+00 -5.E-04 -1.E-03 0 2 4 6 8 10 3.E-03 2.E-03 1.E-03 0.E+00 Z -1.E-03 -2.E-03 -3.E-03 12 0 2 4 2.E-03 1.E-03 5.E-04 0.E+00 -5.E-04 -1.E-03 -2.E-03 -2.E-03 0 2 4 6 8 10 1.E-03 N -1.E-03 -2.E-03 -3.E-03 0 Ground velocity (m/s) Ground velocity (m/s) 5.E-04 0.E+00 -5.E-04 -1.E-03 -2.E-03 6 2 4 6 8 10 12 8 10 3.E-03 2.E-03 1.E-03 0.E+00 E -1.E-03 -2.E-03 -3.E-03 12 0 2 4 Time (s) 6 8 10 12 Time (s) 1.E-02 Spectral density ground velocity (m) 1.E-03 Spectral density ground velocity (m) 12 Time (s) 1.E-03 4 10 0.E+00 12 2.E-03 2 8 2.E-03 Time (s) 0 6 Time (s) Ground velocity (m/s) Ground velocity (m/s) Time (s) 1.E-04 1.E-05 1.E-06 1.E-07 1.E-03 1.E-04 1.E-05 1.E-06 0.1 1 10 Frequency (Hz) 100 0.1 1 10 100 Freque ncy (Hz) Fig. 4 - Example of seismograms and spectra at SR6 station: earthquake (left); explosion (right). From top to bottom: vertical (Z), north-south (N) and east-west (E) components. The spectra analysis was carried to establish a simple criterion for the discrimination between quarry blasts and earthquakes, which might be applied also by an automatic discrimination. The spectra for which we show examples have been obtained from the vertical components after applying a Butterworth band-pass filter with corner frequencies of 1.5 and 20 Hz and two filter sections. We considered a GNGTS – Atti del 19° Convegno Nazionale / 12.10 6.E-03 4.E-03 2.E-03 0.E+00 -2.E-03 -4.E-03 -6.E-03 -8.E-03 Ground velocity (m/s) Ground velocity (m/s) time window of length 8 s, starting at the P-wave arrival without taking care about effects of mixing P and S-waves. Despite the crude preprocessing, the differences between the spectra of quarry explosions and earthquakes are clear. Contrary to nuclear explosions the seismograms of quarry blasts are poorer in high frequencies than seismograms of earthquakes with comparable magnitudes (Evernden, 1969; Elvers, 1974). The ground velocity spectra of the earthquakes studied here show dominating frequencies of up to 20 Hz, whereas the peaks in the spectra of the explosion seismograms fall into ranges between 3 an 10 Hz. Dominating frequencies are particularly low for the quarries between the towns of Augusta and Siracusa, where they hardly exceed 5 Hz. The spectra of the blasts close to Lentini show dominating frequencies between 5 and 10 Hz, anyway fairly below the dominating frequencies of earthquake spectra. 0 2 4 6 8 10 2.E-03 1.E-03 Z 0.E+00 -1.E-03 -2.E-03 12 0 2 4 2.E-02 1.E-02 5.E-03 0.E+00 -5.E-03 -1.E-02 -2.E-02 -2.E-02 0 2 4 6 8 10 6 1.E-03 N -1.E-03 -2.E-03 2 4 Ground velocity (m/s) 6 8 10 12 Time (s) 8 10 2.E-03 1.E-03 0.E+00 E -1.E-03 -2.E-03 -3.E-03 0 12 2 4 6 8 10 12 Time (s) Time (s) 1.E-03 1.E-03 Spectral density ground velocity (m) Spectral density ground velocity (m) 12 0.E+00 0 Ground velocity (m/s) 4 10 2.E-03 12 2.E-02 2.E-02 1.E-02 5.E-03 0.E+00 -5.E-03 -1.E-02 -2.E-02 2 8 3.E-03 Time (s) 0 6 Time (s) Ground velocity (m/s) Ground velocity (m/s) Time (s) 1.E-04 1.E-05 1.E-06 1.E-07 1.E-04 F i g u r e 5 c 1.E-05 1.E-06 1.E-07 0.1 1 10 Freque ncy (Hz) 100 0.1 1 10 100 Freque ncy (Hz) Fig. 5 - Example of seismograms and spectra at SR2 station: earthquake (left); explosion (right). From top to bottom: vertical (Z), north-south (N) and east-west (E) components. Having established sound criteria for the discrimination, we may reconsider the distribution of the quarry blasts and earthquakes with respect to space and daytime. GNGTS – Atti del 19° Convegno Nazionale / 12.10 Infact, as suspected earlier, the “true” seismicity is dispersed rather uniformly over the coastal zone and inland area (Fig. 6). The cluster to the north is related to two earthquake swarms occurring in November 1999 and January 2000 (Scarfì et al., 2001). On the other hand quarry blasts are indeed concentrated between Augusta and Siracusa, some more blasts were localized close to the town of Lentini. The daytime distribution reveals, that most of the quarry explosions are launched between 09:00 a.m. and 10:00 a.m. GMT and are restricted to the working days Monday to Friday (Fig. 7). AN NI IO A SE Lentini Siracusa IB LE AN PL AT FO RM Augusta SICILY C HANN EL Ragusa 0 10 20 Km Fig. 6 - Epicenter distribution of earthquakes (open circles) and quarry explosions (full dots) on the base of location with HYPOELLIPSE. The time span considered is September 1999 to February 2000. Note that some of the quarry blasts were localized in the sea which is probably an effect of the small number of good onset readings available. DISCRIMINATION OF QUARRY BLASTS FROM EARTHQUAKES WITH ARTIFICIAL NEURAL NETWORKS AND APPLICATION OF ANN TO IBLEAN PLATEAU DATA In order to verify our ‘visual’ classification we have applied so called Artificial Neural Networks (ANN) to our discrimination problem. This technique is straightforward for this task since, contrary to conventional classification techniques, it does not require any a priori knowledge about the mathematical structure of the discrimination function. The afore mentioned generality of ANNs requires at minimum a topology with three layers of nodes, i.e., layers with input, hidden and output nodes, respectively. The input layer serves for the storage of the data vector U, with components normalized with respect to a maximum range between –1 to 1. This input vector U is passed along the interconnections with the weights wij to the next, the "hidden" layer. After applying an activation function to the values in the nodes of the hidden layer, GNGTS – Atti del 19° Convegno Nazionale / 12.10 the values are passed to the nodes of the output layer along the interconnections with the weights wik. In the output layer the values obtained by the network are compared with the known ones Y. During the “training phase” the weighting coefficients are adjusted in an iterative procedure minimizing the mismatch between target and calculate out Y. Here we have been using the “Back Propagation Algorithm” (see Werbos, 1974, Rummelhart et al., 1986). Finally we obtain a function, which maps the input vector U to the output vector Y: NH yˆ k ( U) c j ( wTj U t j ) c0 j 1 where: ŷ k = k-th element of Y estimated by the network; U = input vector; wj, = vectors of weights between input and hidden layer; cj = weights between hidden and output layer; tj = biases.; () = sigmoid activation function (z)=1/(1+e-z) 20 a) ) 18 NUMBER OF BLASTS 16 14 12 10 8 6 4 2 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 HOUR (GMT) 14 12 b) NUMBER OF BLASTS 10 8 6 4 2 0 Monday Tuesday Wednesday Thursday Friday Saturday Sunday Fig. 7 - Daily (a) and weekly (b) distribution of quarry blasts recorded in the time span September 1999 - February 2000. GNGTS – Atti del 19° Convegno Nazionale / 12.10 A suitable transformation of the data facilitates the task of signal classification with ANN. As Falsaperla et al. (1996) and Langer et al. (1996) point out, the use of autocorrelation functions instead of plane waveforms improves the results, probably because no problems with phase alignment occur. Besides this, autocorrelation functions represent practically the same information as do amplitude spectra, which have proven to provide significant criteria during the ‘visual’ classification. Autocorrelation functions are normalized by definition, therefore, they can be directly used as input vectors in ANN applications. In data preprocessing autocorrelation functions have been preferred to plane waveform and spectra. 1 a) Score Mismatch [dl] 0.5 0 -0.5 -1 1 b) Score Mismatch [dl] 0.5 0 -0.5 -1 Fig. 8 - Classification mismatch of test set data at stations SR2 (a) and SR6 (b). The time span considered is September 1999-February 2000. We applied the ANN to the data recorded at the key stations SR2 and SR6 in the time interval from September 1999 to February 2000. The autocorrelation functions were obtained from the filtered traces. We divided the data set into two almost equal parts for the training and test data set, choosing the test set events randomly 61 events made up the test set of station SR2 and 60 events at station SR6. We used a length of the input vector of 100, i.e., the first 100 points of the GNGTS – Atti del 19° Convegno Nazionale / 12.10 autocorrelation functions. The output vector is made up by two elements. Thus, the desired output vector Y for an earthquake reads as (1,0), conversely (0,1) for a quarry blast. A classification of a test set event is considered as “failed” if the difference between calculated and desired output of either of the two elements of Y is larger than 0.5. After some trial and error experiments we adopted an ANN with 5 hidden nodes (together with 100 nodes in the input layer and 2 nodes in the output layer). The results of the classification are showed in Fig. 8 where we note two misclassified signals (earthquakes) at station SR2, and one misclassified signal (quarry blast) at station SR6. DISCUSSION AND CONCLUSION The presence of quarries in an active seismic zone may cause severe misunderstandings concerning the distribution of microseismicity. In our particular case an uncritical representation of event location makes believe that the coastal zone between Augusta and Siracusa has an enhanced level of activity compared to the inland areas of the Iblean Plateau. However, the epicenters of events occurring during night times are quite uniformly distributed over the area covered by the seismic network. On the other hand, the events occurring at critical hours, i.e., between 09:00 a.m. and 10:00 a.m. GMT are concentrated in a dense cloud of epicenters between Augusta and Siracusa. As it turns out difficult to distinguish between quarry blasts and tectonic earthquakes on the base of hypocenter locations we have been analyzing waveforms and spectra defining some kind of training set by taking only events occurring during night times. These events represent true tectonic earthquakes, since quarries launch explosions only during working hours. The visual distinction between waveforms of earthquakes and of quarry blasts, particularly when recorded on paper drum recorders, is not evident at all stations. Nevertheless, it was possible to identify two key stations (SR2 and SR6) where the differences between digital waveforms of earthquakes and of quarry blasts appear rather clear. The differences between earthquakes and quarry blasts turned out most clearly from their spectra. Even reducing the signal preprocessing to a level which could be reached also by an automatic treatment, we find dominating frequencies of earthquakes being significantly higher than those of quarry blasts. This is observed also at stations where the visual inspection of waveforms led to some uncertainty of the correct event classification. As a ‘visual’ discrimination can always be argued as arbitrary, we have applied artificial neural networks in a supervised classification of signals as tectonic earthquakes or quarry explosions. This means that we defined training sets of events, whose origin (tectonic or artificial) was supposed to be known, and estimated a discrimination function, which can be arbitrarily complex. The performance of this function was checked by applying it to a set of test events, again with known origin, but not used for establishing the discrimination function. Given the high rate of success, on average about 95%, we conclude that our ‘visual’ classification first carried out is mathematically reproducible and thus far from arbitrary. As the preprocessing steps are simple, it is possible use ANN as a reliable tool for an automatic discrimination of quarry blasts from local tectonic microearthquakes. GNGTS – Atti del 19° Convegno Nazionale / 12.10 REFERENCES Amato, A., R. Azzara, A. Basili, C. Chiarabba, M. Cocco, M. Di Bona, and G. Selvaggi (1995). Main shock and aftershocks of the December 13, 1990, Eastern Sicily earthquake, Annali di Geofisica, 38, 255-266. Azzaro, R., and M. S. Barbano (2000). Analysis of the seismicity of Southeastern Sicily: a proposed tectonic interpretation, Annali di Geofisica, 43 (1), 171-188. Bollettino Sistema Poseidon, 2000. Bollettino trimestrale delle attività, 1, 194 pp. Elvers, E. (1974). Seismic event identification by negative evidence. Bull. Seism. Soc. Amer, 64, 16711683. Evernden, J. F. (1969). Identification of earthquakes and explosions by use of teleseismic data. J. Geophys. Res., 74, 3823-3856. Falsaperla, A., S. Graziani, G. Nunnari, and S. Spampinato (1996). Automatic classification of volcanic earthquakes by using multy-layered neural networks. Natural Hazards, 13, 205-228. Lahr, J.C. (1989). Hypoellipse/version 2.0: A computer program for determining local earthquake hypocenter, magnitude and first motion pattern of local earthquakes. U.S. Geol. Surv., Open File Rep. 75-311. Langer, H., G. Nunnari, and L. Occhipinti (1996), Estimation of Seismic Waveform Governing Parameters with Neural Networks. J. Geophys. Res, 101, p. 20109-20118. Rummelhart, D. E., Hinton, G. E., and R. J. Williams (1986), Learning internal representation by error propagation, in Parallet distributed processing, 1, p. 318-62, MIT Press, Cambridge, MA, U.S.A. Scarfì, L., H. Langer, G. Di Grazia, A. Ursino, and S. Gresta (2001). Analysis of Two Swarms of MicroEarthquakes in Southeastern Sicily: Evidence for Active Faults? J. Seismology Submitted. Werbos, P. J. (1974), Beyond regression: New tools for prediction and analysis in behavioral sciences. Master thesis, Harvard University.