1 Introduction

The Armutlu Peninsula is located in the Marmara region, between the northern and middle branches of the North Anatolian Fault Zone (NAFZ), which is one of the most active fault zones in Turkey. The region contains many quarries and active mines, where explosives such as dynamite are prevalently used in order to obtain material easily and economically. The explosives cause artificial vibrations and are clearly recorded at seismic stations of the Armutlu network, which is established to record and analyze earthquakes at even micro-scale. It is very important to detect artificial signals in seismic records and to distinguish them from earthquake events. Many different methods have been proposed to distinguish quarry blasts from tectonic events in seismic records related to parameters as depth of an event, P wave first motion, P and S wave amplitude peak ratio, complexity versus spectral ratio, coda wave decay rate, power spectrum density, Pg/Lg magnitude ratio, and Pn/Sn amplitude ratio (Wüster 1993; Gitterman et al. 1998; Wiemer and Baer 2000; Koch and Fäh 2002; Horasan et al. 2009; Ursino et al. 2001; Yılmaz et al. 2013; Carr and Garbin 1998; Arrowsmith et al. 2006).

Su et al. (1991) performed the discrimination of quarry/mining blasts and earthquakes by using coda wave power spectrum in the area encompassing the south-central Mojave Desert and eastern Transverse Ranges. In 30-s lapse time, especially at low frequencies (1.5–3 Hz), they found higher Qc−1 values for blasts as compared to the earthquakes. It was expressed that this property is not very significant at lapse times of more than 20 s and high frequencies (6–12 Hz). Borleanu et al. (2016) stated in a study they performed for Romania that clearer results could be obtained with the Vp/Vs ratio method with different statistical approaches. Kim et al. (1997) applied the Pg/Lg magnitude ratio technique in southern Russia. While the failure of discrimination was 7% at a 10–20-Hz frequency interval, it was emphasized that Pg/Lg ratios show poor results at a 2–10-Hz interval. Kekovalı et al. (2012) performed a study in western Turkey mining areas, and determined that 344 out of 520 seismic records originated from artificial explosions, by using S/P wave amplitude peak ratio, complexity, and spectral ratio methods. Ataeva et al. (2017) analyzed and compared the P and S wave displacement spectra from local earthquakes and blasts of similar magnitudes. They obtained ratios of P to S wave corner frequencies of 1.23 and 1.86 for earthquakes and explosions, respectively. Shashidhar et al. (2014) examined the seismic events in Koyna-Warna Region. They have checked the temporal and spatial distribution of events and also evaluated waveform data. The surface waves have low energy around 9 Hz. Uniform P wave polarities are the characteristic properties that have been used for a certain discrimination of blasts from earthquakes. Kuyuk et al. (2011) distinguished microearthquakes from quarry blasts in Istanbul, Turkey. Amplitude peak ratio and complexity methods were both applied with unsupervised learning algorithms for event classification. Yıldırım et al. (2011) discriminated the seismic events using software algorithms that feedforward neural networks, adaptive neural fuzzy inference systems, and probabilistic neural networks.

In our study area, where seismic activity and mining areas are prevalent, there has been no study on the separation of natural earthquake activity from artificial sources. We aim to discriminate earthquakes and quarry blasts in the Armutlu Peninsula focusing on the parameters P and S wave amplitude peak ratio, complexity versus spectral ratio, coda wave decay rate, and power spectrum density methods. A combination and comparison of different methods are applied to increase the reliability of the results. Besides that, statistical approaches with linear and quadratic discriminant functions are performed for the event classification. We believe that this study may be helpful for declustering the earthquake catalogs and it should also contribute to seismological studies by providing information on the source type-frequency content relationship of the seismic events.

2 Geological and tectonic structure of the Armutlu Peninsula

Upper Cretaceous-Pliocene sediments cover the northern part of the Armutlu Peninsula. Generally, the western and southwestern parts of the peninsula are characterized by Upper Jurassic-Lower Cretaceous metamorphic rocks and Tertiary granite. Upper Triassic-Middle Jurassic limestones can be widely observed on the southeastern part of the peninsula (Akartuna 1968; Göncüoğlu et al. 1992).

Two main strike-slip fault systems are located in Turkey called NAFZ, which is dextral, and East Anatolian Fault Zone (EAFZ) that is sinistral. The Armutlu Peninsula is positioned at the western part of NAFZ located between the northern and middle strand of this main fault (see in Fig. 1). This main fault zone and its branches are seismically active as revealed by the approximately 24 ± 1 mm/year movement obtained from GPS data (Reilinger et al. 1997, 2006; McClusky et al. 2003).

Fig. 1
figure 1

Location of “seismic” events (red circles) in the first 7 months of 2014 in and around the Armutlu Peninsula. The seismic station KRSK is the yellow triangle. Green squares show the geothermal sources (Eisenlohr 1997), and blue pentagons depict quarries. The black lines indicate the North Anatolian Fault Zone as well as other faults (Kuşçu et al. 2009; Çaka 2012). The yellow star shows the epicenter of the August 19, 1999, Kocaeli Earthquake (M = 7.4) and the dashed yellow line marks the western rupture of this earthquake (Barka et al. 2002)

Full size image

The region is tectonically complex and produces a large number of microseismic events. The 1999 Kocaeli earthquake occurred along the northern branch of NAFZ and stopped north of the peninsula (see in Fig. 1). Within the study area, no big earthquake has been occurred in the last century along the middle strand. The Armutlu Peninsula is located close to important cities such as Istanbul, Bursa, and Kocaeli and its active tectonics have been investigated by many researchers and institutions (Eisenlohr 1997; Armijo et al. 1999; Robertson and Ustaömer 2004; Tunç et al. 2011; Kinscher et al. 2013).

In the Armutlu Peninsula, the seismic activity is a mix of tectonic events, quarry blasts (mostly in limestone areas), and events related to geothermal sources. For example, in 2014, fi a seismic swarm occurred at the geothermal reservoir of Termal (Yavuz et al. 2015).

3 Data set

Station KRSK is one of the seismic stations of the ARmutlu NETwork (ARNET) that is operated by Kocaeli University Earth and Space Sciences Research Center (ESSRC) and GFZ German Research Centre for Geosciences. In this study, KRSK recordings were used to discriminate earthquakes and quarry blasts at different explosion sites as this station has recorded events more clearly than others. It has a 360″ broadband sensor (GÜRALP) and at sampling rate set to 100 samples per second. Only the vertical component of the seismograms was used for calculation. One hundred seven events were cataloged for the first 7 months of the year 2014; however, some of them were compulsorily eliminated due to some technical problems (power loss, non-recording signals, etc.) of station KRSK. Eighty-seven seismic events with high signal/noise ratio records were chosen (Fig. 1).

Within each quarry, explosions took place less than a few hundred meters apart from each other and their energy release is not as high as that of earthquakes. However, regarding the sensitivity of depth solutions, it is useful to increase the depth limit (Aki and Richards 1980; Sertçelik and Başer 2010). Maximum 10 km of focal depth, 0 to 23 km of epicentral distance, and up to Md = 3.0 duration magnitudes were selected for discriminant analyses.

SEISAN software (Havskov and Ottemoller 1999) was used to determine the location of the events. To improve the location quality, minimum five stations’ records for each event which are azimuthally distributed were taken into account.

With respect to the waveforms of events—due to the different source mechanisms—P wave peak amplitudes are much higher than other phases in signals from quarry blast explosions (point source); otherwise, earthquakes (plane source) have higher amplitudes of S and surface waves. When the distance between epicenter and station increases, the seismograms often begin to alter from this general view. Thus, a more detailed investigation of seismograms was needed. In this context, five different methods and two statistical approaches were used to discriminate the quarry blasts from tectonic events.

4 Methodologies and results

In this study, five different techniques that are both in time and frequency domain were applied to the station KRSK recordings of the 87 events with epicenters in the vicinity of the explosion sites.

The signals were first evaluated manually taking into account the following criteria: P wave first motion, no S wave recording, high amplitude of P wave, Rg phase appearance, shallow depth, daytime of occurrence, and distance to quarries. This manual identification is named as “gold standard” and was used to calculate the accuracy of the other techniques (Table 2 in the Appendix).

4.1 Daytime of event

The quantity of explosions increases during day time which corresponds to regular blasting hours of the quarries. To identify preferred blasting times, UTC has been converted to local summer and winter time. In case an “event” occurred before 11:00 and after 19:00 local time, the respective event source was marked as “EQ” (Table 2 in the Appendix). Each day, the average “seismic” activity increases with a sharp onset at 11:00 local time. Increasing day time activity is a clear statistical hint towards the identification of quarry blasts, but it does not help to discriminate distinct events. The graph suggests that the above-average seismic activity during day time is caused by quarry blasts, but it is not possible to reach a definite decision for each single event (Fig. 2, Table 2 in the Appendix).

Fig. 2
figure 2

Time of day frequency distributions of 87 seismic events. The blue and red bins depict GMT and local time, respectively (corrected for summer and winter time periods)

Full size image

4.2 Amplitude ratio As/Ap versus log(As)

This method compares the amplitude ratio of maximum S to P wave to the logarithm of peak amplitude of S wave in time domain (Wüster 1993; Baumgardt and Young 1990). Although this method is quick and easy, it is not completely reliable.

Discriminant analysis is a classification method that assumes different classes to generate data based on different Gaussian distributions. To create a classifier, the fitting function estimates the parameters of a Gaussian distribution for each class. To predict the classes of new data, the created classifier finds the class with the smallest misclassification. Here in this study, linear and quadratic discriminant analyses were applied (Krzanowski 1988; Seber 1984; MATLAB Release 2011). Linear discriminant function (LDF) analysis (Eq. 1) estimates one covariance matrix for all classes, whereas the quadratic discriminant function (QDF) analysis (Eq. 2) estimates one covariance matrix for each class (Kuyuk et al. 2014).

$$ {F}_{\mathrm{LDF}}=K+L(1)\ast x+L(2)\ast y $$
(1)
$$ {F}_{\mathrm{QDF}}=K1+\left[x\kern0.5em y\right]\ast L1+\sum \left\{\left(\left[x\kern0.5em y\right]\ast Q1\right)\ast \left[\begin{array}{cc}x& y\end{array}\right]\right\} $$
(2)

where K is a constant term of the boundary equation, Ls are linear coefficients of the boundary equation, and Q is quadratic coefficient matrix of the boundary equation. Their indices are the coefficients of the linear or quadratic boundaries between the classes.

The events were determined in the time domain using amplitude ratio method. Figure 3a, b, shows the As/Ap versus log(As) distribution. Among the scattering variables, the discrimination line and curve were drawn by LDF and QDF algorithms. The discrimination procedure was used to define a linear and quadratic combination of quantitative variables that characterized the differences between the groups.

Fig. 3
figure 3

Amplitude peak ratio of S to P wave (As/Ap) versus logarithm of amplitude peak of S wave (log(As)) of the vertical component of the velocity seismogram for station KRSK. The classification was performed with a linear discriminant function analysis and b quadratic discriminant function analysis. Blue triangles and red circles depict earthquakes (EQ) and quarry blasts (QB), respectively. The pink line and curve indicate the discrimination of earthquake and quarry blast using LDF and QDF algorithms

Full size image

Using the LDF algorithm and comparing with the gold standard, 21 earthquakes and 56 quarry blasts were clearly discriminated, totally ten events were misclassified. Performing the QDF algorithm, 21 earthquakes and 55 quarry blasts were determined, whereas 11 events could not be classified (Table 1, Table 2 in the Appendix).

Table 1 Classification results of As/Ap–log(As) and complexity-spectral ratio using LDF and QDF algorithms on Figs. 3a, b, 4a, b, and Fig. 5a, b; coda wave decay rate and power spectrum density method analysis (Figs. 7, 8) of all events based on the “gold standard”
Full size table

Calculated parameters for As/Ap–log(As) methods with linear and quadratic forms are Eqs. 3 and 4

$$ {F}_{\mathrm{LDF}}=8.9203+\left(-1.3924\right)\ast \log {\mathrm{A}}_{\mathrm{s}}+\left(-2.6261\right)\ast {\mathrm{A}}_{\mathrm{s}}/{\mathrm{A}}_{\mathrm{p}} $$
(3)
$$ {F}_{\mathrm{QDF}}=\left(-10.0562\right)+\left[\log {\mathrm{A}}_{\mathrm{s}}\kern0.5em {\mathrm{A}}_{\mathrm{s}}/{\mathrm{A}}_{\mathrm{p}}\right]\ast \left[\begin{array}{c}6.3013\\ {}7.4494\end{array}\right]+\sum \left\{\left(\left[\begin{array}{cc}\log {\mathrm{A}}_{\mathrm{s}}& {\mathrm{A}}_{\mathrm{s}}/{\mathrm{A}}_{\mathrm{p}}\end{array}\right]\ast \left[\begin{array}{cc}-0.8487& -0.5677\\ {}-0.5677& -3.6037\end{array}\right]\right)\ast \left[\begin{array}{cc}\log {\mathrm{A}}_{\mathrm{s}}& {\mathrm{A}}_{\mathrm{s}}/{\mathrm{A}}_{\mathrm{p}}\end{array}\right]\right\} $$
(4)

4.3 Complexity versus spectral ratio

In this method, first, the seismogram is divided into two time windows (t0–t1, t1–t2) and the spectral ratio of integrated powers s2(t) of each part is then calculated. It is called complexity (C) and depends on the epicentral distance. Second, ratio of integrated spectral amplitudes a(f) in the different frequency bands (h1–h2, l1–l2, high and low frequency band, respectively) is calculated from the waveform of an event. It is called spectral ratio (Sr) and depends on the frequency content of an event. C and Sr can be written in Eq. 5 and Eq. 6, respectively (Arai and Yosida 2004; Gitterman and Shapira 1993).

$$ \mathrm{C}={\int}_{t_1}^{t_2}{\mathrm{s}}^2(t) dt/{\int}_{t_0}^{t_1}{\mathrm{s}}^2(t) dt $$
(5)
$$ \mathrm{Sr}={\int}_{h_1}^{h_2}\mathrm{a}(f) df/{\int}_{l_1}^{l_2}\mathrm{a}(f) df $$
(6)

Because epicentral distances differ, distance grouping of the events is important in order not to lose the phases within the selected windows. In conjunction with this, the frequency contents differ for each event, too. That is why, in this study, the seismograms were divided into two different time windows and frequency bands for station KRSK. Thus, 59 and 28 records at 0–13-km and 13–23-km epicentral distances were analyzed, respectively. At short distances, the time windows were selected as t0, P wave onset time; t1, 1.5 s; and t2, 3 s; otherwise, in long distances, t0 is not changed, but t1, 2.5 s, and t2, 5 s, were chosen. The frequency bands were determined as h1, 5 Hz; h2, 10 Hz; l1, 1 Hz; and l2, 5 Hz, for all seismograms. Two statistical approaches (LDF and QDF analyses as mentioned in the “Amplitude ratio As/Ap versus log(As)” section) were used for classification of natural and artificial events.

The C versus Sr distributions are shown for 0–13 km of epicentral distances in Fig. 4a, b, and for 13–23 km in Fig. 5a, b. The discrimination line and curve were derived by LDF and QDF algorithms due to sprinkled variables.

Fig. 4
figure 4

Plot of complexity (C) versus spectral ratio (Sr) of the vertical component of the velocity seismograms of KRSK for events with 0–13 km of epicentral distance. The classification was performed with a linear discriminant function and b quadratic discriminant function analysis. Blue triangles and red circles depict earthquakes (EQ) and quarry blasts (QB), respectively. The pink line and curve indicate the discrimination of earthquake and quarry blast using LDF and QDF algorithms

Full size image
Fig. 5
figure 5

Plot of complexity (C) versus spectral ratio (Sr) of the vertical component of the velocity seismogram of KRSK for 13–23 km epicentral distances. For details see Fig. 4

Full size image

When compared to the gold standard, ten earthquakes and 43 quarry blasts were confirmed and six events were misclassified by using both LDF and QDF algorithms for 0–13-km epicenter distance events. On the other hand for the epicentral distance range of 13–23 km, 11 earthquakes and 13 quarry blasts were correctly identified with the LDF algorithm, while four events could not be discriminated. Using QDF algorithm for the same distance range, 11 earthquakes and 14 quarry blasts were correctly determined, while three events were not (Table 1, Table 2 in the Appendix).

Calculated parameters of the C versus Sr analyses for the events between 0–13-km and 13–23-km epicenter distances with linear and quadratic forms are shown below with Eqs. 7–10:

$$ 0-13\ \mathrm{km};{F}_{\mathrm{LDF}}=3.5546+\left(-0.1407\right)\ast \mathrm{Sr}+\left(-3.1691\right)\ast \mathrm{C} $$
(7)
$$ 0-13\ \mathrm{km};{F}_{\mathrm{QDF}}=3.6024+\left[\mathrm{Sr}\kern0.5em \mathrm{C}\right]\ast \left[\begin{array}{c}-2.5060\\ {}3.5140\end{array}\right]+\sum \left\{\left(\left[\begin{array}{cc}\mathrm{Sr}& \mathrm{C}\end{array}\right]\ast \left[\begin{array}{cc}0.5040& -0.1131\\ {}-0.1131& -4.7056\end{array}\right]\right)\ast \left[\begin{array}{cc}\mathrm{Sr}& \mathrm{C}\end{array}\right]\right\} $$
(8)
$$ 13-23\ \mathrm{km};{F}_{\mathrm{LDF}}=2.8097+\left(-0.7458\right)\ast \mathrm{Sr}+\left(-1.7076\right)\ast \mathrm{C} $$
(9)
$$ 13-23\ \mathrm{km};{F}_{\mathrm{QDF}}=\left(-0.1773\right)+\left[\mathrm{Sr}\kern0.5em \mathrm{C}\right]\ast \left[\begin{array}{c}1.6954\\ {}6.3432\end{array}\right]+\sum \left\{\left(\left[\begin{array}{cc}\mathrm{Sr}& \mathrm{C}\end{array}\right]\ast \left[\begin{array}{cc}-0.2608& -1.8429\\ {}-1.8429& -3.9308\end{array}\right]\right)\ast \left[\begin{array}{cc}\mathrm{Sr}& \mathrm{C}\end{array}\right]\right\} $$
(10)

4.4 Coda wave decay rate versus frequency

The solution for a damped harmonic oscillator incorporates damping described through the dimensionless quality factor (Q). It is used to describe the decay of an oscillator or the physical properties of a system that causes a disturbance to attenuate (Stein and Wysession 2003). In other words, it can be define as the ratio of energy loss per cycle to total energy. Coda waves, also known as the wave train of S waves, from small local earthquakes, are interpreted as backscattered waves from numerous heterogeneities distributed uniformly in the earth’s crust (Aki and Chouet 1975). The attenuation characteristic obtained from coda waves defines a coda wave quality factor (Qc). The amplitude of coda waves entirely depends on distance, time, and frequency. Equation 11 indicates the frequency dependency of Qc.

$$ {\mathrm{Q}}_{\mathrm{c}}(f)={\mathrm{Q}}_{\mathrm{o}}{\left(\frac{f}{f_{\mathrm{o}}}\right)}^n $$
(11)

In this equation, fo is the reference frequency; Qo is the Qc at 1 Hz and n indicates the frequency dependence. The different frequency contents of earthquakes and quarry blasts are linked to different source types and thus can be used for event discrimination (Su et al. 1991; Hartse et al. 1995). Quarry blasts have simpler spectra than earthquakes. The coda decay rate, Qc, is significantly lower for quarry blasts than for earthquakes (Kiszely 2005).

The determined coda wave envelope is important for calculating Qc. At lower frequency bands and shorter coda wave envelopes, the difference of coda wave attenuation between earthquakes and quarry blasts can be easily recognized. Otherwise, at increasing frequency bands and lengthened time windows, the difference cannot be observed, because the waves rapidly decay at shallow depths. Generally, attenuation decreases with depth. Thus, it can be easily understood that attenuation is very strong for quarry blasts which are located at shallow depths. This technique aims to detect the decay rate from the slopes.

As a fourth discrimination method, the signals were evaluated by coda wave decay rate. The absorption of energy should be different for earthquakes and quarry blasts, although a medium has a filter effect that leads to attenuation. It is known that earthquakes have wide frequency bands; while quarry blasts have limited frequencies, because explosions have generally frequency-controlled sources (Kiszely 2005). In this method, 10 to 35 s with 5-s interval coda wave envelopes were chosen and 1.5, 3, 6, 9, and 12-Hz center frequency bands were performed for each time window. All seismograms were digitally bandpass-filtered considering the constant bandwidth filtering, to get a comparable amount of energy into each band. The events have such a close epicenter distances to station KRSK that the coda window was selected to start at 1.5-s travel time from the origin. Figure 6 shows an example of the determined coda window for calculating attenuation rate at station KRSK. The CODAQ subroutine integrated in SEISAN software was used to calculate the quality factor (Havskov and Ottemoller 1999).

Fig. 6
figure 6

Example of the 20-s coda window on the vertical component at station KRSK. Event occurred on 22/07/2014 04:18:02.9 GMT (Md = 2.5)

Full size image

Hartse et al. (1995) analyzed the source factor difference in Nevada using coda waves. This method distinguishes between different events while concentrating on smaller magnitudes. Identifying event types with an attenuation rate on various frequencies is effective for this method. Kizsely (2005) used a Qc in Vértes Mountains, Hungary. At 6-Hz and 10-Hz corner frequencies, the difference between natural and artificial events is most prominent. Aki (1995) has successfully discriminated quarry blasts and earthquakes at local distance by using coda Q method with testing lower and higher frequency bands.

In this study, up to M = 3.0 duration magnitudes were used for discrimination and Qc can be easily seen at all corner frequencies. If Qc−1 values of increasing coda wave window length decrease more rapidly (i.e., higher slope), the event should be a quarry blast; otherwise, for earthquakes, the slope is much lower. The classification of the events has been done by visual inspection of the slopes. Figure 7 demonstrates the results of the attenuation rate at different frequencies of earthquakes and quarry blasts, separately. Focusing on the decay rate, events could be discriminated in an easier and more decisive way than with previous methods; yet, the quarry blasts were still indistinguishable after a certain coda wave window length. For these cases, the irregular scattered values prevented the correct identification of earthquakes and quarry blasts. With respect to coda wave decay rate method, 80 events were identified and seven events were not classified in compliance with the gold standard (Table 1, Table 2 in the Appendix). The success rate of this method was calculated as 91.9%.

Fig. 7
figure 7

Coda wave decay rate (Qc−1) versus coda wave window length obtained from all seismic events. The boxes show the attenuation rate for different frequency ranges. EQ, earthquakes; QB, quarry blasts; f, center frequency

Full size image

4.5 Power spectrum density versus frequency

Power spectrum density (PSD) is known as the strength of the energy versus frequency, which may show strong or weak frequency variations in the whole signal or in a selected part of it. Comparing the energies of all events, the power spectra of quarry blasts decrease more sharply with increasing frequency than those of earthquakes showing a more stable pattern (Su et al. 1991). Moreover, the Po(ω) value at 1 Hz (beginning) and the same value at 10 Hz (end) are important for classification. Artificial events have lower end spectra with regard to their beginning. Allmann et al. (2008) compared the P wave power spectra of different events for Southern California. They obtained a constant stress-drop model for earthquakes; otherwise, the model curve has a steeper trend for quarry blasts. Ataeva et al. (2017) analyzed the P and S wave displacement spectra ratio of natural and artificial quakes. It can be pointed out that the difference in the ratios can be effective for discrimination. Aki (1995) noted that the power spectra of quarry blasts decrease sharper with increasing frequencies than those of earthquakes. Hence, the comparison of PSD shows remarkable difference between earthquakes and explosions.

In this study, PSD was calculated by GEOPSY software (Wathelet 2005) within 1 to 10-Hz frequency band between the beginning and the final point of the signal for station KRSK. The discrimination was done by visual comparison of the power spectral values at 1 Hz (beginning) and 10 Hz (end). According to the distribution of power spectra versus frequency graphics, earthquakes and quarry blasts were separated using the frequency contents as shown in Fig. 8. Considering the PSD analysis for 87 seismic events, 35 earthquakes and 52 quarry blasts were clearly discriminated from each other with the accuracy of 93.1% (Table 1, Table 2 in the Appendix).

Fig. 8
figure 8

Power spectrum density versus frequency obtained at station KRSK within 1 to 10-Hz frequency range. The dashed red lines indicate the upper and lower PSD changes of earthquakes and quarry blasts. EQ, earthquakes; QB, quarry blasts

Full size image

4.6 Comparison of different techniques

Table 1 provides a comparison of the performance of different discrimination techniques with respect to the “gold standard,” the PSD technique standing with the highest success rate (93.1% in overall). To reach a final decision on the classification of each event, the results of all methods were evaluated and a decision was made when at least three of five methods yielded the same event type. However, the day time of the event helped to differentiate 21 earthquakes; therefore, for the remaining 66 events, we have results of four methods only. In fact, the five misclassified cases are those for which we cannot make a decision as two of the methods discriminate them as earthquake while the remaining two as blasts. The linear and quadratic forms of the two methods could also be used to make final decision; but, they are not helpful in the present cases, as both forms give the same results. From the 87 events, 27 earthquakes and 55 quarry blasts could be discriminated, while five events were misclassified (Table 2 in the Appendix). This final decision was compared with the “gold standard” and the accuracy is calculated as 94.2%. Epicenters of the earthquakes and quarry blasts according to the final decision are plotted in Fig. 9.

Fig. 9
figure 9

Location of the identified earthquakes (green circles), quarry blasts (red stars), and misclassified events (pink asterisks) through the final decision (with accuracy 97.7%). The seismic station KRSK is the yellow triangle and blue pentagons depict quarries

Full size image

5 Conclusion

In this study, 87 events were classified with manual investigation and five different methods including two kinds of statistical approaches. All events were first classified manually yielding the so-called “gold standard”; then other methods were compared to it. A sharp increase of events at 11:00 local time indicates quarry blast activities, but does not allow the discrimination of natural and artificial events. The amplitude peak ratio method could not provide a clear event classification. Similarly, the C versus Sr method, where classification lines and curves were determined by LDF and QDF for two different distance intervals (0–13 km, 13–23 km), could not clearly discriminate the events. In both methods, the use of the linear and quadratic forms did not provide any significant change in the resulting discrimination. The average slopes of the Qc with increasing window lengths could differentiate the events except 12-Hz central frequency and narrow coda wave window lengths. With respect to the PSD technique, earthquakes have preserved their energies at variable frequencies, whereas quarry blasts have lost more energy with increasing frequencies. Among with 87 seismic events at the first 7 months of 2014, 35 earthquakes and 52 quarry blasts were classified with the accuracy of 93.1% with respect to the “gold standard.” Thus, the frequency domain PSD method could provide more reliable results than the others. For the final decision, all methods were compared to each other and the solutions were combined (Table 2 in the Appendix). The events were classified with five methods including two statistical approaches. Hence, 27 earthquakes and 55 quarry blasts were identified using all different methods and five events were misclassified obtaining a success rate as 94.2%.

The discriminant functions that are evaluated for this study are special for this region and used data sets. The crustal structure, the characteristics of the quarry blasts, and different located events may generally change the parameters of these functions. Therefore, the LDF and QDF can only be useful for future studies in and around the Armutlu Peninsula. Nevertheless, a multi-method application is recommended for areas with similar problems to distinguish better between tectonic and anthropogenic seismic activities.

The discrimination of natural-artificial events is important for tectonically active, geologically complex regions with frequent mine activities. The event classification has to be effective and is crucial for seismological and seismicity studies. Recently, the infrastructure on the Armutlu Peninsula is developing rapidly, with planned constructions of new factories, highways, tunnels, bridges, etc., and it is one of the seismically active regions that should be carefully monitored. Therefore, detailed analyses and interpretations can provide and contribute the basic information about the seismic activity of the peninsula.