Introduction

While seismic recorders are recording seismic events in a region, they record not only natural seismic activities but also man-made events such as quarry blasts. These events exist in seismicity catalogs together. This situation may cause some errors for scientific studies in these areas, such as in the determination of the b-value in seismic hazard studies. In order to determine the real seismic activity in a study area, seismic catalogs should be cleaned of quarry blasts. The use of location, distance and origin time are simply not enough to achieve this. Therefore, waveforms should be carefully examined (Horasan et al. 2006).

Many different methods exist in the literature on the subject of the discrimination of natural and man-made seismic activity. These include the Pn/Sn and Pn/Lg ratio methods (Baumgardt and Young 1990), the Lg/Pg and Lg/Rg ratio methods (Wüster 1993), the artificial neural networks (ANNs) method (Dowla et al. 1990), the linear discriminant function (LDF) method (Horasan et al. 2006, 2009; Deniz 2010; Öğütçü et al. 2010; Kartal 2010; Kekovalı et al. 2010, 2012; Badawy et al. 2019; Ceydilek and Horasan 2019) and short-time Fourier transform (STFT) methods (Yılmaz et al. 2013), quadratic discriminate function (QDF), diquadratic discriminate function (DQDF) and Mahalabonis discriminate function (MDF) methods (Küyük et al. 2011), QDF methods (Yavuz et al. 2018), spectral seismograms (Korrat et al. 2008), declustering-dequarry discriminate methods (Kalafat 2010), the Fisher–Shannon discrimination method (Telesca et al. 2011), the corner frequency discriminant (CFD) method and P- and S- wave corner frequencies (Ataeva et al. 2017), and histograms of time versus seismic events (Naserieh et al. 2019).

In addition to these methods, natural and artificial seismic events have also been distinguished from each other using several algorithms of artificial neural networks (ANNs). In this study seismic activities occurring during May 2009 and March 2014 in Edirne (Turkey) and its vicinity were examined. The location of the stations and the distribution of seismic events are shown in Fig. 1. Most of the quarry blasts recorded in the study area are related to mineral and construction material extraction. The purpose of this study is the discrimination of quarry blasts from earthquakes by applying the linear discriminate function (LDF) and artificial neural networks (ANNs) methods on digital vertical component velocity seismograms recorded at the ERIK, ENEZ and GELI seismic stations. The obtained values of accuracy percentage were compared to determine the real seismic activity. This will improve the quality of earthquake catalogs, help to better determine their completeness, reduce errors in seismic hazard studies by ensuring that quarry sites are not identified as fault zones and allow the calculation of reliable b-values.

Fig. 1
figure 1

Location of the stations ERIK, ENEZ and GELI and seismic events with magnitudes of Md ≤ 3.5 in the study area between May 2009 and March 2014 (RETMC). Faults were taken from Şaroğlu et al. (1992), Emre et al. (2013) and Yaltirak et al. (2012). NAFZ: North Anatolian Fault Zone

Full size image

Data and methods

In this study, a total of 150 seismic events with Md ≤ 3.5 recorded at three stations, ERIK, ENEZ and GELI, were investigated between May 2009 and March 2014 in the region bounded by 40–41° N and 25.30–26.80° E (Fig. 1). Data were taken from Boğaziçi University, Kandilli Observatory and Earthquake Research Institute, and the Regional Earthquake–Tsunami Monitoring Center (RETMC).

The vertical component velocity seismogram data recorded at station ENEZ have the lowest number of data points.

The seismic activity distribution versus day and time (in GMT) are shown in Fig. 2. The histogram at the right side of Fig. 2 shows seismic events after removing quarry blasts. Time and location domain discrimination may not be sufficient, hence the vertical component velocity seismogram and spectrum were also examined.

Fig. 2
figure 2

Distribution of seismic activity (number of events) occurring between May 2009 and March 2104 versus hours (in GMT) in the study area (40–41° N and 25.80–26.70° E). a The maximum activity is observed at 09:00 and 15.00 in GMT through the day. b Distribution of seismic events after removing quarry blasts

Full size image

The seismogram and spectrum of the natural and artificial seismic events at ERIK are seen in Figs. 3 and 4, respectively. The P-wave amplitude of artificial seismic events is dominant compared to the amplitude of earthquakes in Fig. 3. In this study, we used some parameters such as the amplitude ratio of the maximum S-wave and maximum P-wave of the vertical component of the velocity seismograms, power ratio (complexity) and total signal duration of the waveform, as described in detail by Horasan et al. 2009. The complexity versus ratio of the maximum P-wave amplitude and maximum S-wave amplitude at the vertical component velocity seismograms of the selected stations allowed the determination of the linear discriminant function (LDF) using Statistical Package for the Social Sciences (SPSS) Analysis Program (SPSS 2005) to discriminate natural and man-made seismic events.

Fig. 3
figure 3

The vertical component of velocity seismograms recorded at station ERIK. a Earthquake, b quarry blast

Full size image
Fig. 4
figure 4

Normalized amplitude spectrum of signals recorded at station ERIK. a Earthquake, b quarry blast

Full size image

Artificial neural networks (ANNs) method

Back propagation feed forward neural networks (BPNNs) learning algorithm

In this study we used the back propagation feed forward neural networks (BPNNs) learning algorithm because of its advantages, such as reducing backward-propagating error, namely from output to input (Çetin et al. 2006) and having an easy neural structure at the expense of slowness in the learning process (Çayakan 2012). According to the size of the errors between the expected output and the real value of the output, weights were organized by using BPNNs learning algorithm in order to obtain the most suitable output values (Yıldırım 2013). Generally, the members of the network topology are shown in Fig. 5 (Gülbağ 2006).

Fig. 5
figure 5

(Modified from Gülbağ 2006)

a Members of the network topology, a neural network structure for seismic events. b Ratio versus complexity.

Full size image

According to the type of the problem, after the learning algorithm was determined, a network structure that included an input layer, a hidden layer and an output layer was obtained. Generally, members of the network architecture were determined as inputs, outputs, weights, the sum function and the activation function (Rumelhart et al. 1986). Inputs were information that entered the cell from other cells or out-medias and entered the cell using connections that were on the weights (w) (Fig. 5).

In this study the topology of the artificial neural network is a feed-forward backpropagation artificial neural network (BPNNs) and its learning algorithm is supervised learning. In the supervised learning algorithm input and output values entered together into the computing system. In this learning algorithm we used the amplitude peak ratio of the S to P-wave and complexity values as input (Fig. 5).

Selection of the number of neurons (Nn)

While deciding the topology of the artificial neural network, the selection of the number of neurons (Nn) is an important criterion in the ANNs method (Gülbağ 2006). Kermani et al. (2005) emphasized that Nn has an important role in neural networks. Nn is one of the significant factors for the discrimination of different data groups. If we used fewer neurons than we needed at the hidden layer, it might cause us to obtain very few sensible results. Often, when the number of neurons is low in hidden layer, it fails to validate the connection of input and output factors. Similarly, when the number of neurons in the hidden layer is high, it causes overfitting (Molga 2003). While the structure of the ANNs was being obtained, Nn was decided by trial and error (Yıldırım 2013; Kaftan et al. 2017).

At the stage of making a decision to determine a suitable model, Nn was tried at an interval and incremented. Then, the artificial neural network model that had the highest accuracy percentage was chosen for the determined ANNs model (Gülbağ 2006). In the literature, researchers used different intervals using different increments. Gülbağ (2006) obtained their ANNs model using increments of 10 neurons between 0 and 100. Küyük et al. (2009) determined their model using increments of 1 neuron between 1 and 20 in the model and selected their number of neurons as 5 because its accuracy percentage was the highest. Yıldırım (2013) obtained their network architecture of ANNs by incrementing the number of neurons by 2 between 0 and 22. Kaftan et al. (2017) determined their model using an increment of 1 between 1 and 6 in their artificial neural network study.

In this study, to define the ANN network we incremented the number of neurons by five between 5 and 25. To determine the best neuron for ANN network we calculated determination coefficient (R2) values. The highest determination coefficient value is 0.99 for the number of neurons equal to 10 (Table 1). So we selected the best number of neurons as 10 in this data set (Table 1).

Table 1 The variation of determination coefficient values (R2) obtained using the ANNs method for the pair of ratio versus complexity parameters that belonged to the Edirne study area versus values of the number of neurons per data set
Full size table

An additional training algorithm used was Levenberg–Marquardt, and the activation function used in this study was the Hyperbolic Tangent-Sigmoid activation function. The application of Levenberg–Marquardt to neural network training is described in the literature (Hagan and Menhaj 1994; Kermani et al. 2005). This algorithm has an efficient application in MATLAB software (Charrier et al. 2007; MATLAB 2011). The network trainlm function can train every member of the artificial network (MATLAB 2011). The Levenberg–Marquardt training algorithm has an efficient implementation (Levenberg 1944; Marquardt 1963).

We have to discuss about the training algorithm on a vast scale. Gülbağ and Temurtaş (2007) showed the equations of the standard back propagation and the Levenberg–Marquardt (LM) algorithms. They explained the reasons of using the LM method and the attitudes of that method versus the generalization learning by heart as follows. When the number of the data was inadequate, learning by heart could occure and in that state it might be difficult for the generalization. But according to Gülbağ and Temurtaş (2007) that problem might be solved as: While training with the training set at the same time they were testing it simultaneously by using the test set until an error level determined achieved. When the error of the test set reached to an acceptable level, they recorded the network.

Another activation function selected, denoted by \(\varphi \left( x \right)\) and defined the output of a neuron in terms of the induced local field v. In this study, we used the hyperbolic tangent sigmoid function in this network architecture. In fact, this activation function assumed a continuous range of values from − 1 to + 1. Therefore, the activation function was an odd function of the induced local field as shown in Eq. (1).

$$\varphi \left( x \right) = \left\{ {\begin{array}{*{20}l} 1 \hfill & {{\text{if}}\quad x > 0} \hfill \\ 0 \hfill & {{\text{if}}\quad x = 0} \hfill \\ { - 1} \hfill & {{\text{if}}\quad x > 0} \hfill \\ \end{array} } \right.$$
(1)

which is commonly referred to as the signum function. For the corresponding form of a sigmoid function, we may use the hyperbolic tangent sigmoid function, defined by

$$\varphi \left( x \right) = \tanh \left( x \right)$$
(2)

It is a hyperbolic tangent sigmoid activation function to assume positive and negative values as prescribed by Eq. (2) (Haykin 2009).

Based on the initial investigation, the hyperbolic tangent sigmoid was used for all output layers except the first output layer. The hyperbolic tangent sigmoid activation function can be defined using Eq. (3).

$$\varphi \left( x \right) = \frac{2}{{1 + {\text{e}}^{{\left( { - 2x} \right)}} }} - 1$$
(3)

Here: hyperbolic tangent sigmoid activation function

The dynamic variation interval is [− 1 1] and this function shows variations according to the number of neurons and total input (Gradshteyn and Ryzhik 2007).

Preparation of data set

After determining the number of neurons, we started to prepare the data sets that belonged to inputs and outputs. Next, the normalization process was applied. Then, a significant percentage of the data was selected as the training data and the remainder was taken as the testing data. These processes were materialized randomly. Kermani et al. (2005) selected their data randomly in a similar way. We then obtained the new data set. We trained with the training data using the BPNNs learning algorithm. When R2 was approximately 1, the training was stopped and completed. Next, using the testing process, the ANNs method was applied. This means that “The expected artificial neural network learned the learning algorithm from the training data, so it can test its information using the test data.” The data had to provide that rule. The obtained outputs were then compared with the tested outputs. Consequently, the accuracy percentage was calculated.

Different researchers prepared their data using different percentages for training and test data. Ursino et al. (2001) used 50% of their data as training data and 50% of their data as testing data. Gülbağ (2006) used 84% of their data for training and 16% as testing data. Yıldırım et al. (2011) used 25% of their data set as training data and 75% of the data set as testing data for their study. Kundu et al. (2012) used 51% of their data as training data and 49% as testing data in their study. Yıldırım (2013) selected the data randomly and then separated them into two parts. Their training data made up 70% of all data and the remaining part of the data was testing data. Kaftan et al. (2017) selected the data arbitrarily and then separated them into two parts, with 85% of all data used as training data and 15% of the data used as testing data.

In this study the data were randomly selected from all data sets belonging to ERIK, ENEZ and GELI. We arranged a common data set called E_ALL and used 70% of all data as training data and 30% as testing data (Table 2). The all data set consists of 235 seismic events that were recorded in the ERIK, ENEZ and GELI. We then separated the E_ALL data set into two parts. The number of training data points was 165, and the number of test data points was 70 for the E_ALL data set (Table 4).

Table 2 Features of the data set according to the pair of ratio versus complexity parameters using the ANNs method
Full size table

The ANNs method was applied to a pair of parameters ratio versus C for E_ALL data set and then accuracy percentage was obtained for it (Table 3).

Table 3 Number of events in the training set, test set, misclassified earthquakes and misclassified quarry blasts for the pair of ratio versus complexity parameters according to E_ALL data set using the ANNs method
Full size table

Further, we applied the k-fold cross-validation technique to all of the data (James et al. 2017). Suitable results were obtained. All results were obtained using the ANNs method in MATLAB (MATLAB 2011).

Results and discussion

Before we applied the LDF and ANNs methods to the data, we tried to identify earthquake and quarry blasts according to the amplitude of the signal. We observed that the P-wave amplitude of quarry blasts is dominant compared to the amplitude of earthquakes. The frequency content of the events is shown in Fig. 4. We observed the spectral modulation on the quarry blast spectrum. These identification methods were not sufficient for satisfactory discrimination of earthquakes from quarry blasts. For this reason, different pairs of parameters, such as the ratio of the amplitude of the maximum S-wave to the amplitude of the maximum P-wave, the logarithmic value of the amplitude of the maximum S-wave (log S), the ratio of power at two time windows of the signal (complexity) and the total duration of the signal were used.

The classification of natural and artificial seismic events was realized using the linear discriminate function (LDF) and the artificial neural networks (ANNs) methods. As a result, 81 (54%) of the total studied 150 seismic events were determined to be earthquakes and 69 (46%) of them were determined to be quarry blasts (Fig. 6).

Fig. 6
figure 6

Distribution of earthquakes (filled red circles) and quarries (filled green stars) in the study area. Faults were taken from Şaroğlu et al. (1992), Emre et al. (2013) and Yaltırak et al. (2012). NAFZ: North Anatolian Fault Zone

Full size image

Ratio versus complexity

The amplitude ratio versus complexity values for the LDF and the ANNs methods are plotted in Figs. 7 and 10 for E_ALL data set. The results of the classification method LDF for pairs of criteria 1 (ratio vs C), 2 (ratio vs log S) and 3 (ratio vs duration) are given in Table 5 for the E_ALL data set. In the first criterion in Table 5, 69 earthquakes out of 80 were classified correctly and 11 earthquakes were misclassified as quarry blasts, whereas 155 quarry blasts were classified correctly. Using LDF method we obtained an accuracy percentage of 95% for the E_ALL data set.

Fig. 7
figure 7

Plot shows the distribution of ratio versus complexity for the data set including all stations (E_ALL) in the study area. The accuracy percentage obtained is 95% for LDF

Full size image

After using the LDF method, we discriminated earthquakes and quarry blasts with the ANNs method for the amplitude ratio and complexity parameters. The number of neurons versus the determination coefficient (R2) for ratio versus Complexity values are given in Table 1. In this table determination coefficient values change between 0.97 and 0.99. The highest value of the determination coefficient (R2) is a very important criteria for decision of the Nn for the pair of parameters. We selected the best number of neurons as 10 in this data set (Table 1). This situation indicates that the BPNNs learning algorithm is successful for those parameters on that structure of the network architecture. The comparison of determination coefficient (R2) values that were obtained using the ANNs method for the pair of ratio versus complexity parameters in Edirne and the values of the number of neurons which were increased by 5 between 5 and 25 is given in Table 1. The comparison of determination coefficient versus Nn is not sufficient for the ANN model performance. In order to evaluate the performance of the model we obtained the variance account value as 99% (Table 2).

We created test and training data set in Table 2. In this study, we used 70% of the total data for training set and 30% for testing (Tables 3, 4). The comparison of the two methods is shown in Table 6.

Table 4 Number of training and test data sets used for modeling classification status and model accuracy
Full size table

Ratio versus log S

The amplitude ratio versus log S values for the LDF method are plotted in Fig. 8 for the E_ALL data set. The results of the classification method between natural and artificial seismic events using the LDF method for criteria pair 2 (ratio vs log S) are given in Table 5 for E_ALL data set. In the second criterion in Table 5, 66 earthquakes out of 80 were classified correctly and 14 earthquake were misclassified as a quarry blast, whereas 155 quarry blasts were classified correctly. Using the LDF method we obtained an accuracy percentage of 94% for E_ALL data set.

Fig. 8
figure 8

Plot shows the distribution of ratio versus log S for the data set including all stations (E_ALL) in the study area. The accuracy percentage obtained is 94% for LDF

Full size image
Table 5 The results of the discriminant analysis using the LDF method for Criterion 1: ratio versus C, Criterion 2: ratio versus log S and criterion 3: ratio versus duration parameters for the E_ALL data set
Full size table

Ratio versus duration of the signal

The amplitude ratio versus duration of the signal for the LDF method was plotted in Fig. 9 for E_ALL data set. The results of the classification method for natural and artificial seismic events using the LDF method for criteria pair 3 (ratio vs duration) are given in Table 5 for the E_ALL data set. In the third criterion in Table 5, 75 earthquakes out of 80 were classified correctly and 5 earthquakes were misclassified as a quarry blast, whereas 154 quarry blasts were classified correctly. Using the LDF method we obtained an accuracy percentage of 97% for the E_ALL data set (Fig. 10).

Fig. 9
figure 9

Plot shows the distribution of ratio versus duration for the data set including all stations (E_ALL) in the study area. The accuracy percentage obtained is 97% for LDF

Full size image
Fig. 10
figure 10

Plot shows the distribution of ratio versus complexity for the data set including all stations (E_ALL) in the study area. The accuracy percentage obtained is 99% for ANNs

Full size image

The LDF method is one of the most popular and successful techniques in earth science worldwide for the classification of natural and artificial seismic events. Horasan et al. (2009) obtained accuracy percentage values for pairs of parameters (Ratio vs log S) of 98.6%, 93.8%, 97.7% and 95.8% for Gaziosmanpaşa, Çatalca, Gebze-Hereke, and Ömerli, respectively. Yılmaz et al. (2013) determined accuracy parameters of 96.3%, 89.3%, 100%, 100%, 96.5%, and 100% for stations KTUT, ESPY, BAYT, PZAR, GUMT, and BCA, respectively, in their study. In their Egypt study, Badawy et al. (2019) obtained lower accuracy percentage values (91.7%, 83.7% and 83.2%) than those seen in other countries using the S-wave/P-wave amplitude peak ratio, complexity and spectral ratio. The accuracy percentage of these parameters depends on the quantity of data, geological features, and local site effects.

Number of neurons versus determination coefficient values for ratio and complexity are given in Table 2. In this table the highest determination coefficient value is 0.99. This situation indicates that the BPNNs learning algorithm was successful for these parameters on that structure of the network architecture in the area considered in this study.

According to the results of this study, the number of seismic events recorded at all three stations by using LDF and ANNs methods will be investigated for whether the accuracy percentage is directly proportional. When we compared the accuracy percentage values for LDF and ANNs methods, both of these methods are successful but the ANNs method is more successful than the LDF method (Table 6). Yıldırım et al. (2011) used three methods to distinguish natural and artificial seismic events in Istanbul and its vicinity. They obtained model success rate of 99% for feed-forward backpropagation neural networks (FFBPNN), 97% for probabilistic neural networks (PNN) and 96% for adaptive neural fuzzy inference systems (ANFIS). These ratios are similar to our E_ALL data set results. Hence, we conclude that the quarry blasts were discriminated very effectively in this study, and this will improve seismic hazard studies of the region. While this is true as a generic principle, the main seismic hazard in the study region is due to the W segment of North Anatolian Fault/NE segment of North Aegean Fault.

Table 6 Comparison of the accuracy percentage values for the E_ALL data set according to the LDF and ANNs methods
Full size table