The Intense 2020–2021 Earthquake Swarm in Corinth Gulf: Cluster Analysis and Seismotectonic Implications from High Resolution Microseismicity

Abstract

The intense 2020–2021 seismic crisis in Corinth gulf, Greece, comprising several hundreds of small earthquakes (maximum magnitude M_w = 5.4 on 17 February 2021) is investigated. The spatial and temporal evolution of the seismicity implied the activation of multiple secondary fault segments. To decipher the geometry of the activated structures, we engaged relocation techniques and obtained the precise locations for 3398 earthquakes and 26 moment tensor solutions. The highly accurate seismicity locations and focal mechanisms illustrate the fine scale faulting geometry of a ~ 10–km–long activated area, almost east west striking and north dipping, and extensional kinematics. We grouped events into clusters using nearest–neighbor distances between them and a temporal stochastic point process, the Markovian Arrival Process (MAP). We identified clusters that evidence seismicity migration and organization in both space and time, deciphering the interaction of even tiny fault segments in a fault network. The exhaustive analysis of the swarm spatiotemporal evolution revealed several either distinct or contiguous activated minor fault segments that evolved in multiple structures, participating in the local fracture mesh. Faulting geometry and kinematics of these structures agree with the ~ N–S extension of the rift and north dipping fault planes.

Appendices

Appendix 1: Earthquake Relocation

We used the recordings of the seismological stations of Hellenic Unified Seismic Network (HUSN) at distances up to 100 km from the study area for compiling the earthquake catalog of the 2020–2021 earthquake swarm in the western Corinth gulf (see Fig. 10).

Fig. 10

Spatial distribution of relocated seismicity in the study area. The inset map shows the seismological stations used for the earthquake relocation. The spatial distribution of seismicity covered an area of approximately 500 km² with a distinct bulk of seismicity revealed in a single cluster in an almost east west elongated zone of ~ 10 to 15 km length

For the application of the Wadati method (Wadati, 1933), spatial and temporal fluctuations of the V_p/V_s ratio were observed with values ranging from 1.6 to 1.9 with an average error of 0.007. For this reason, we performed two different assessments to investigate whether these variations are indicative and driven by the seismic activity. The dataset was divided a) in sliding windows of 100 events and b) in settled time windows determined by the occurrence of the larger magnitude events (20 December–11 January, 12 January–16 February, 17 February–10 March). From this analysis, we did not observe any systematic significant increase or decrease before or immediately after the stronger events occurrence. We observed instead an increase in V_p/V_s ratio towards the east, following the migration of earthquake epicenters. Ultimately, a velocity ratio of 1.80 was deemed appropriate using earthquakes having at least 10 S phases (1842 events) recorded at stations up to 100 km, as shown in Fig. 11 where the origin times of all events were reduced to zero.

Fig. 11

The number of P– and S–wave phases used from the seismological stations in ascending mean epicentral distance (km)

We selected a representative sample of 638 earthquakes, spread out in space and time and followed a two–step approach to define the upper and bottom layers. At first, we used only the closest stations (up to 40 km) to better define the top layers of the model and then we applied the algorithm again, accounting for all available stations up to 100 km to consider rays traveling through the deeper layers. The 1D velocity model proposed by Rigo et al. (1996) (Fig. 12, Table 1) for the study area was used as a reference model. We performed multiple runs of the VELEST algorithm until the changes in the resulting crustal model became negligible and the rms turned stable. The final crustal model (Fig. 12, Table 1), defined after merging the layers with equal velocity, consists of six (6) layers and a half space below 19 km (Table 1) (see Fig. 13).

Fig. 12

Wadati diagram derived from the travel time data. Linear regression provides a V_p/V_s ratio of 1.80

Fig. 13

The 1D crustal velocity model (red line) constructed in this study by applying the VELEST algorithm along with the one proposed by Rigo et al. (1996) as a reference model

The Fig. 14 represents the histograms of the differences in hypocentral locations between catalogs. The differences between our final catalog, namely the double difference and cross correlation catalog (cc in Fig. 14), and the Hypoinverse catalog are notably smaller than the difference between the final catalog and the routine analysis catalog (initial in Fig. 14). This implies that the adapted velocity model and station residuals significantly contribute to the location accuracy and improve the locations of the routine analysis catalog. Furthermore, the procedure that we followed and applied (double difference and cross correlation) returns higher relative location uncertainties with the routine analysis catalog, which used imperfect velocity model without stations delays.

Fig. 14

Location differences among earthquake catalogs. Initial: the routine catalog, cc: the catalog after double difference and cross correlation running, Hypoinverse: the catalog resulted after Hypoinverse running with the velocity model and station residuals defined in this study

Appendix 2: Fault Plane Solutions

We selected the seismic stations used for the calculation of the fault plane solutions based on their azimuthal coverage and epicentral distance (Fig. 15). An additional criterion was the instrument response produced by the poles and zeros for each station. We initially considered more stations and consequently rejected, because of fluctuations in the amplitude response of the instruments, which caused us to question the reliability of the pole and zero files.

Fig. 15

Regional seismic stations used for the moment tensor inversion. Inverse triangles denote the seismic stations while the different colors stand for the different networks that compose the Hellenic Unified seismic network (HUSN). The networks are, HT—Aristotle University of Thessaloniki Seismological Network (https://doi.org/10.7914/SN/HT), HL—National Observatory of Athens Seismic Network (https://doi.org/10.7914/SN/HL), HP—University of Patras, Seismological Laboratory (PSLNET) (https://doi.org/10.7914/SN/HP), HA—Hellenic Seismological Network, University of Athens, Seismological Laboratory (https://doi.org/10.7914/SN/HA). All the fault plane solutions are located inside the red rectangle

We evaluated the solutions quality by using mainly the variance reduction (VR), which examines waveform match and the condition number (CN), which is a measure of the reliability of the inversion. Additional criteria include the Focal Mechanism Variability Index (FMVAR), which compares the optimal solution using the Kagan Angle and the best fit solution, and the Space–Time Variability Index (STVAR), which is complementary to FMVAR and is a measure of the size of space–time area corresponding to a given correlation threshold (Sokos & Zahradnik, 2013).

Appendix 3: Local Magnitude Correction

Local earthquake magnitudes (M_L) reported in the initial earthquake catalog are estimated during the routine analysis performed by the analysts of the Geophysics Department of Aristotle University of Thessaloniki (GD–AUTh; http://geophysics.geo.auth.gr/ss/) using the method of Hutton and Boore (1987) for each individual station. For some of these stations the M_L values to be appear under– or over– estimated when compared to the final M_L assigned to the certain earthquake. Magnitude homogenization in earthquake catalogs is an indispensable component for any further investigation (e.g. Mobarki & Talbi, 2022). The Fig. 10a shows the median difference between the final M_L values and the M_L values of 23 stations (M_L–M_LSTA) located in distances up to 90 km from the center of 2020–2021 earthquake swarm.

Figure 16a evidences that in two stations, namely the EFP (Efpalio) and KALE (Kallithea), the median difference is equal to 0. The differences in the other stations are either positive or negative, with values up to $\pm$ 0.3 magnitude units with three of them exceeding this range. In particular, for KLV (Kalavryta) and LAKA (Lakka) stations the M_L values are underestimated with mean deviations equal to − 0.43 and − 0.37, respectively, and for SERG (Sergoula) station M_L is found overestimated by + 0.48. These observations imply that a persistent overestimation or underestimation of the final M_L calculations exists, depending on the stations used in the M_L calculation during the routine analysis.

Fig. 16

a Median residual value between final M_L and the corresponding M_L of the 23 stations located at distances up to 90 km from the study area. b Frequency histogram of the M_L values from each seismological station considered for the final M_L calculation

Aiming to improve the reliability of the earthquake magnitudes in our catalog, the elimination of these persistent misestimating is necessary. In this respect, we applied an M_L correction approach. For this purpose, we examined the relation between the M_L of selected stations and a reference station, correcting the M_L of the respective stations according to the obtained relations and recalculating the final M_L as the mean value of the corrected M_L values of the selected stations.

We selected the EFP station as the reference station, because it is among the two (EFP and KALE) stations whose median difference from the final M_L is equal to zero, and the most frequently used in the final M_L calculations (3394 times instead of the 3176 times of the KALE station; Fig. 16b). As a next step, we selected the stations used in the final magnitude calculations more than 2000 times, to investigate the relation among them and the reference station. These stations are the ANX (Ano Hora), DRO (Drosia), EVR (Evrytania), KALE (Kallithea), LAKA (Lakka), SERG (Sergoula) and VVK (Vomvokou) ones, participating in 3300, 2518, 2163, 3176, 3076, 2200 and 3020 M_L calculations, respectively. This selection includes stations that (either slightly or significantly) overestimate (ANX, DRO, EVR, SERG) and underestimate (LAKA, VVK) the M_L magnitude, ensuring that the magnitude correction will be objective.

The plots of the available M_L pairs of the seven (7) selected stations and the EFP reference station (Figs. 17a, 18a, 19a, 20a, 21a, 22a and 23a) reveal both linear (3 cases; ANX versus EFP, KALE versus EFP, VVK versus EFP) and nonlinear (4 cases; DRO versus EFP, EVR versus EFP, LAKA versus EFP, SERG versus EFP) trends. Taking into account this fact, the determination of the M_L relations (M_LSTA versus M_LEFP) is implemented by fitting the data with the General Orthogonal Regression (GOR) method (Fuller, 1987), which is a widely used method for magnitude conversions (e.g. Karakostas et al., 2020; Leptokaropoulos et al., 2013), in the cases of the linear trend. In the cases of non–linearity the data are fitted by a 2nd degree polynomial in a least square sense.

Fig. 17

a Local magnitude (M_L) values of the ANX versus EFP stations (blue circles), along with the linear fit among them (red solid line). b Histograms of the initial (blue) and corrected (orange) M_L values of ANX station as obtained after the correction using the respective linear relation

Fig. 18

a Local magnitude (M_L) values of the DRO versus EFP stations (blue circles), along with the 2nd degree fit among them (magenta solid line). b Histograms of the initial (blue) and corrected (orange) M_L values of DRO station as obtained after the correction using the respective 2nd degree relation

Fig. 19

a Local magnitude (M_L) values of the EVR versus EFP stations (blue circles), along with the 2nd degree fit among them (magenta solid line). b Histograms of the initial (blue) and corrected (orange) M_L values of EVR station as obtained after the correction using the respective 2nd degree relation

Fig. 20

a Local magnitude (ML) values of the KALE versus EFP stations (blue circles), along with the linear fit among them (red solid line). b Histograms of the initial (blue) and corrected (orange) ML values of KALE station as obtained after the correction using the respective linear relation

Fig. 21

a Local magnitude (M_L) values of the LAKA versus EFP stations (blue circles), along with the 2nd degree fit among them (magenta solid line). b Histograms of the initial (blue) and corrected (orange) M_L values of LAKA station as obtained after the correction using the respective 2nd degree relation

Fig. 22

a Local magnitude (M_L) values of the SERG versus EFP stations (blue circles), along with the 2nd degree fit among them (magenta solid line). b Histograms of the initial (blue) and corrected (orange) M_L values of SERG station as obtained after the correction using the respective 2nd degree relation

Fig. 23

a Local magnitude (M_L) values of the VVK versus EFP stations (blue circles), along with the linear fit among them (red solid line). b Histograms of the initial (blue) and corrected (orange) M_L values of VVK station as obtained after the correction using the respective linear relation

Figures 17, 18, 19, 20, 21, 22 and 23 show the results of the fitting approach. Both linear and 2nd degree polynomial fits are performing well in respect with the data, indicating rather strong correlation between the M_L of each station under correction and the reference station.

We corrected the M_L of each station by applying the obtained relations on the initial values of M_L derived from the routine analysis (Figs. 17b, 18b, 19b, 20b, 21b, 22b and 23b). We may observe that for the stations with overestimated initial M_L values the corrected ones are now reduced (e.g. SERG station; Fig. 22b), while for the stations with underestimated values the corrected ones are increased (e.g. LAKA station; Fig. 21b), as expected.

The final step of the procedure is the calculation of the final M_L for each earthquake included in the earthquake catalog. We performed this calculation by estimating new mean M_L values for all the earthquakes using the available corrected M_L values of the seven (7) selected stations, along with those of the reference station. Figure 24a shows a summary of the corrected final M_L values in comparison with the initial ones. We observed significant changes of the M_L values from the 1.1 up to the 1.8 magnitude bin, after the correction procedure. Specifically, there are five (5) magnitude bins (1.1, 1.3, 1.5, 1.6 and 1.8) of the initial magnitude values with reduced number of events. This led to a more regular binning distribution in the corrected M_L values. The corrected M_L values follow an increasing linear trend in respect to the initial M_L values for almost the entire magnitude range (from 0.5 up to 4.0 magnitude bins, Fig. 24b). Slight deviations from this linear trend are observed for the earthquakes with M_L < 0.5 and for those with M_L > 4.0, in which the corrected M_L values are higher and almost equal, respectively.

Fig. 24

a Histograms of the initial M_L (blue) versus the corrected M_L (orange) values of the earthquake catalog. b Initial M_L versus corrected M_L (blue circles) values for all the earthquakes of the catalog, along with their linear relation (red solid line)

Focusing on more detail in the difference among the corrected and the initial M_L values (Fig. 25a) it is derived that in 95% the cases the corrected values are modified within the range of − 0.3 up to + 0.2 magnitude units. In the majority of these earthquakes, the corrected values are increasing up to 0.1 magnitude units with a median value equal to 0.05. The magnitude of 95% of earthquakes was corrected by using at least 3 M_L values (stations) with a median value equal to seven (Fig. 25b), leading to implementing the correction procedure in a robust way for improving the reliability of magnitude calculation.

Fig. 25

a Histogram of the difference between the Corrected and the Initial M_L values of the earthquake catalog. b Histogram of the number of stations used in the final corrected M_L calculations of the earthquake catalog

We identified the completeness magnitude, M_c, through the Goodness of Fit (GFT; Wiemer & Wyss, 2000) method at the 95% of residuals (Fig. 26). M_c is found to be equal to the first magnitude bin below the 95% residual bound (blue dashed line), which is equal to 1.2 (M_c = 1.2) having residual percentage equal to 4.29 (Res = 4.29%). The complete earthquake data set consists of 2523 earthquakes with M ≥ M_c.

Fig. 26

Percentage of frequency–magnitude distribution residuals between the final relocated catalog and the ideal synthetic power law as a function of M_c obtained by the Goodness of Fit Method. The red triangle (Residual = 4.29%) depicts the magnitude bin (M = 1.2) having first residual below the 95% bound

Appendix 4: Sensitivity Analysis of MAP-DBSCAN Input Parameters

The DBSCAN algorithm needs two input parameters, the minimum number of neighbors, $minPts$, and the distance threshold, $\epsilon$. The former determines the density level of the cluster i.e., larger values correspond to denser clusters as more neighbors are required for a cluster to be defined. In our case, $minPts=4$, is chosen that seems a reasonable value to avoid trivial cases of clusters with 2 or 3 events. The choice of the distance threshold, $\epsilon$, is based on the k-nearest neighbor plot, which is a tool proposed by Ester et al. (1996) for the determination of the parameter. For each event $i$ we computed its k-nearest neighbor, with $k=minPts$, and all distances are plotted in ascending order. The optimal $\epsilon$ value is based on changes in the slope of the curve that indicate a change in the correlation among the events. In our case, we computed the k-distances for events within the potential clusters that are defined by the MAP model as the DBSCAN algorithm is implemented separately to each one of them. Figure 27 shows the corresponding distances of the three largest potential clusters ($N\ge 100$) for $k=4$. Gradient changes in the slope range between 0.3 and 0.9 km, so we decided to use as input three values, $\epsilon =0.3, 0.6, 0.9$.

Fig. 27

The k-nearest neighbor plot of the potential clusters with N ≥ 100 events. Black horizontal dashed lines indicate the range of ϵ values given as input to the DBSCAN algorithm and each color corresponds to a potential cluster

A temporal merging factor, $T$, is also added to the procedure, in the sense that potential clusters within temporal distance $T$ days are merged into one. We tested three different values, the trivial case with $T=0$ and with $T=2, 4$ days, respectively. Table 4 provides the complete set of the nine tested parameters.

Table 4 The nine tested parameter sets of the MAP–DBSCAN method

We used the cumulative number of the background seismicity for the nine realizations of the clustering algorithm, MAP-DBSCAN, as a tool to test the impact of the parameters on the detection of the clusters. Figure 28 presents the cumulative number of events that have not been assigned to a cluster for each set of parameters along with the initial datasets. We observe pronounced peaks in the cumulative curves for thresholds $\epsilon \le 0.6$ independently of the parameter set. This is an indicator of triggered seismicity wrongly assigned as background so these distance thresholds should be rejected. For the largest threshold ϵ = 0.9 km, a rather stable curve is observed with minor differences among the parameters.

Fig. 28

Cumulative number of the initial data (red line) and cumulative number of background seismicity for parameter sets a PS1, b PS2 and c PS3 for the three different distance thresholds ($\upepsilon = 0.3, 0.6, 0.9$ km)

The differences between the three parameter sets are further explored with the comparison of the space–time evolution between the declustered and the initial datasets. Figure 29 shows similar results among the three datasets, however, there is a concentration of seismicity in Fig. 29b (blue ellipse) that is removed in Fig. 29c, d. We detected the main seismic excitations in Fig. 29c, d while preserving the main patterns of background seismicity. Therefore, parameter PS2 with $T=2$ days seems an appropriate choice although without significant differences with the parameter PS3 ($T=4$ days).

Fig. 29

Space–time evolution of the a initial seismicity and background seismicity for parameter sets b PS1, c PS2 and d PS3