Analysis of the Epicenter Location Accuracy for the Local Seismic Network Operated in the Mining Area Towards the Automation of Location Procedures

Legnica-Głogów Copper District (LGCD) is one of the most active seismic regions in Europe. Several thousand seismic events induced by underground copper mines are recorded there each year, with the strongest reaching magnitudes above 4. Seismicity in LGCD is monitored by the LUMINEOS surface seismic network and the mine's underground networks. While the horizontal location uncertainty of the LUMINEOS network is about 300–600 m, the declared epicenter uncertainty of dense mine networks is less than 50 m. It gives us a unique opportunity to test the location accuracy of seismic events recorded by the LUMINEOS network determined with various algorithms and automatic procedure. In our work, we compared the location accuracies of traveltime-based algorithms LocSAT and NonLinLoc as well as waveform-based algorithm BackTrackBB. The iterative, Geiger-type LocSAT algorithm is used in routine daily processing for the LUMINEOS. Its downside is the need to use the S wave onset times, which in the case of the LUMINEOS network are very uncertain. As an alternative, we tested the probabilistic NonLinLoc algorithm, and the waveform based BackTrackBB algorithm. The BackTrackBB algorithm is known to work well with local, high noise seismic networks. We aimed to find out if it could be used routinely with the LUMINEOS data. In addition, we conducted a comparative analysis of the location uncertainty of these algorithms to determine the effectiveness of this parameter in assessing accuracy.


Introduction
Seismic activities observed in mining areas are considered very shallow, with weak to moderate (up to M4.5) earthquakes. Due to shallow sources, often less than 1 km, the surface in the vicinity of epicenter can be affected by relatively strong ground vibrations. Moreover, the surface infrastructures around mines are crucial for exploitations works in mines, and in the worst scenario seismicity may be responsible for termination of industrial activity in a part or the whole mine (Garcia-Aristizabal et al., 2020;Grigoli et al., 2017). Thus, it is very important to improve the knowledge corresponding to mining seismicity, starting with seismic source locations, which should be obtained with resolution even better than in case of tectonic seismicity.
Seismic event localization is a typical geophysical inverse problem. The parameters of the seismic source are determined on the basis of the observed data, which are usually the P and S waves onset times (i.e. travel-time-or picking-based algorithms). Sometimes additionally the waves back-azimuths are used. Alternatively, source locations can be determined on the basis of the entire waveforms (i.e. waveform-based algorithms) (Grigoli et al., 2014;Li et al., 2020;Poiata et al., 2016).
The localization methods have evolved over time along with the growth of computing power. Standard techniques are based on the Geiger-type algorithm, which iteratively minimizes the objective function in search of its global minimum (e.g. Geiger, 1912). The objective function is traditionally defined as the sum of the squares (i.e. the L2 norm but sometimes other norms are used) of the time residuals. Time residuals are the differences between the measured arrival times (picked on seismograms) and the theoretical times calculated on the basis of the velocity model used in the specific area. Another way to calculate the location is the time-consuming grid search method, which consists of calculation of the misfit function for all grid points. In such a case, the final location is at the point where the misfit takes the minimum value.
A much more effective method is the probabilistic approach, which narrows the search grid to the area with the highest probability of a hypocenter (Dębski, 2004(Dębski, , 2010Tarantola, 1987).
Seismic waves onset times are usually manually picked by seismologists. However, picking can be subjective and error prone in the case of weak or moderate signal-to-noise ratio. Moreover, the systematically growing amount of seismic data volume increases amount of work. The solution to this problem may be automation of localization procedures. There are two main approaches to this problem. The first one combines automated picking of seismic waves onset times and location with the use of standard algorithms. The latest automatic pickers based on neural networks work correctly even with heavily noisy data. They are constantly being developed and becoming more and more popular in global and local seismology (e.g. Münchmeyer et al., 2022). The second approach is relatively new, fully automated waveform-based location algorithms, which consist of the stacking of characteristic functions computed from seismograms (i.e. functions that enhance the amplitude of seismic waves onset). In this method, the algorithm locates seismic source based on the signal coherence (Li et al., 2020). In recent years, many algorithms of this type have been developed and successfully applied to noisy data, e.g. Source Scanning Algorithm (Kao & Shan, 2004), LOKI (Grigoli et al., 2014), or BackTrackBB-BTBB (Poiata et al., 2016).
The difference between the true (i.e. real location of the rupture starting point) and theoretical location obtained with a selected location algorithm defines the location accuracy. Location accuracy is influenced by many factors. The main ones are geometry of a seismic network and number of sensors with high enough signal-to-noise ratio, quality of velocity model and algorithm which is used for the location procedure (e.g. D' Alessandro et al., 2011;Havskov et al., 2012). In this work, we focused our attention on the location algorithm.
For this purpose, different approaches were tested: the travel-time-based algorithms LocSAT (Bratt & Bache, 1988) and NonLinLoc (Lomax et al., 2000) and the automated waveform-based algorithm BTBB. The purpose of this paper is to verify whether novel quasi-automated waveform-based location algorithm (here BTBB) can be routinely used with the typical local surface mining network and to investigate the effect of uncertain S wave onset times picking, which can be an important issue in mining seismicity (Kokowski & Rudziński, 2022, 2023. In the previous works, the location quality of waveform-based algorithms has been tested and compared with locations calculated by other classical techniques (e.g., Grigoli et al., 2014;Namjesnik et al., 2021;Pesicek et al., 2014;Poiata et al., 2016). Generally, the authors found the location results to be similar. However, they consider waveform-based methods as preferable, because of their higher automation.
What distinguishes our work is the ability to compare the estimated locations obtained using different algorithms with locations estimated with a very local underground mining network, which for our purposes can be considered as almost ''real location'' of seismic sources. In fact, better locations could be obtained only with well know blasting, which are not available in this area (Caputa & Rudziński, 2019). Furthermore, we conducted a comparative analysis of the formal location uncertainty of all these algorithms to evaluate the effectiveness of this parameter in assessing accuracy.
To provide a comprehensive and clear understanding of the terms related to the evaluation of location quality, we aim to establish precise definitions for a few key terms that will be referenced throughout this article: Location uncertainty is calculated on the basis of the measured arrival times and the geometry of the seismic network. It indicates the consistency of measured arrival times with the velocity model used. Typically, it is presented in the form of an error ellipse or ellipsoid.
Accuracy is a difference between actual and measured location: can only be calculated when the true location of the event is known, e.g. for planned mine explosions.
Discrepancy describes the epicentral distance between the location derived from the LUMINEOS surface network and the significantly more precise location obtained from the underground networks. We interpret discrepancy as an indicator of accuracy.

Seismic Monitoring in Legnica-Głogo´w Copper District
The Legnica-Głogów Copper District (LGCD) (Figs. 1 and 2) is one of the most seismically active regions in Europe (Rudziński et al., 2021). Anthropogenic seismicity induced by the extraction of copper ore by KGHM mining corporation appeared several years after mining began in 1968. The strongest earthquake in LGCD had magnitude of M4.  (Lizurek et al., 2015;Rudziński et al., 2016), November 29, 2016(Lasocki et al., 2017, and January 29, 2019 (Ilieva et al., 2020). Seismicity in the LGCD is monitored by the underground seismic networks belonging to the KGHM mines and the LUMINEOS surface network (Fig. 2) belonging to the Institute of Geophysics of the Polish Academy of Sciences (IG PAS).

The LUMINEOS Seismic Network
The LUMINEOS (The Legnica-Głogów Underground Mining INduced Earthquake Observing System) is a surface seismic network belonging to the Department of Seismology of IG PAS (Rudziński et al., 2021) which started observations in 2013.
Because of network improvements, the numbers of sensors and their configuration changed multiple times since 2013 and ultimately, the network got its final shape in 2017 (Fig. 2). Now the LUMINEOS network consists of 27 sensors, of which 10 are GeoSIG AC-73 accelerometers with sampling  Vol. 180, (2023) Analysis of the Epicenter Location Accuracy for the Local Seismic Network 2563 frequency of 250 sps and the rest 17 are GeoSIG VE-53/BB seismometers with sampling frequency of 100 sps. All of them are equipped with GMSPlus recorders developed by GeoSIG (Rudziński et al., 2021). Sensors in the network are usually 2-3 km apart and the whole network rectangular size is about 10 9 25 km (the longest diagonal distance NWLU-TRZS is around 27 km).
In the daily practice, the earthquakes are detected by manual review of seismograms by the IG PAS staff and localized using the manually picked arrival times and the LocSAT algorithm (Bratt & Bache, 1988) incorporated into the SWIP software (https:// docs.cyfronet.pl/display/SWIP/LocSAT, last access Dec. 2022). According to our previous analysis, the horizontal uncertainty of the locations in the center of the network is approximately 400 m (Kokowski & Rudziński, 2023). Relatively high value of uncertainty is probably caused by two factors: an inaccurate and simple 1D velocity model and high noise level connected with operation in the urban and industrial area (Fig. 3). Our analysis of the quality of the LUMINEOS seismic network (Kokowski & Rudziński, 2023) indicated an additional problem with the accuracy of S-wave arrival times, which negatively affects the quality of earthquakes' locations. The

The KGHM's Underground Seismic Networks
The KGHM's copper mines have three seismic networks with the stations located at the deposit level and in mining shafts. Each of them has several dozen single-component short-period Willmore III seismometers, which are located at the mine level and in mining shafts (Caputa et al., 2015;Koziarz & Szłapka, 2010). Such an arrangement of the stations allows to obtain epicenter location accuracy smaller than 100 m (Rudziński & Dębski, 2012) and even approximately 25 m, assuming the P wave picking errors of 5 ms (Koziarz & Szłapka, 2010). Typically, for these networks the source depth is not calculated, and it is assumed just above the deposit level i.e. in average around 650 m below the ground level (Koziarz & Szłapka, 2010;Talaga et al., 2017). The velocity model used for events location assumes three possible wave propagation velocities: 5 km/s for the direct wave and 5.9 km/s and 5.6 km/s for two refracted waves (Caputa et al., 2015;Król, 1998).
Due to the shorter distance from the seismic sources and much simpler seismic wave propagation, the mine catalogs have smaller magnitude threshold and contain more events in comparison with the LUMINEOS seismic catalog. However, they are not complete, because they do not contain seismic events that occurred during the planned blasting works i. e. these events are not located and not used in our analysis (Caputa & Rudziński, 2019). In contrast to the LUMINEOS, the mine catalog is not open and is just used in our work with kind agreement of the KGHM Polska Miedź SA.

Location Techniques Used
In our study, we used data recorded between January 1, 2020 and October 31, 2021. During this time, LUMINEOS locations were routinely determined using the manual picking and LocSAT algorithm. Additionally, for this period, we performed localizations using the NonLinLoc program, just to test the possible effect of S phases picking on locations results. Finally, we calculated locations using the automatic waveform based BTBB algorithm, to verify usefulness of the automated algorithm for shallow induced seismicity data recorded in a very noisy industrial area. For all calculations a simple 1D velocity model obtained with P-wave borehole profiling and the S-wave velocity calculated assuming typical Vp/Vs ratio equal to 1.73 was used (Dec et al., 2011) (Fig. 3a).

LocSAT
LocSAT (Bratt & Bache, 1988) is a Geiger-based (Geiger, 1912) least-squares-inverse location algorithm, which iteratively minimizes the objective function using the observed P-and S-wave arrival times and waves backazimuth estimated for theoretical onsets. In practice, it is the Jordan and Sverdrup (1981) algorithm extended with azimuth data measured at the station clockwise from the north. The disadvantage of the LocSAT algorithm is the necessity to use at least one S-wave phase and the high dependence of the quality of the results on the accuracy of S-waves picking. This issue is especially visible for the data recorded on local distances, as in our use case. For flat seismic networks, the accuracy of the depth determination is much lower than for the epicentral parameters (Havskov et al., 2012), Hence, in order to improve the final solution, the depth is often fixed. In our case, the depths were fixed at the exploitation level, which is between 500-900 m depending on the part of the mining area.

NonLinLoc
NonLinLoc software (Lomax et al., 2000; http:// alomax.free.fr/nlloc/, last access Dec. 2022) is a set of non-linear, global-search probabilistic location algorithms, which allows the use of 1-3D velocity models and generates reliable location uncertainty. The typical global-search method, i.e. exhaustive grid-search, is very time consuming. By contrast, NonLinLoc uses efficient sampling methods to speed up the computation, e.g. the Oct-tree Importance Sampling Algorithm. In OctTree algorithm 3D space is divided into cells and for each cell's center probability density function (PDF) value is calculated. The probability that the earthquake is located in a given cell is calculated as the multiplication result of a cell volume and its corresponding PDF value. Cell with the maximum PDF value is divided into 8 new cells and then the procedure is repeated until the given accuracy criterion is achieved (Lomax & Curtis, 2001). As a result, NonLinLoc returns a PDF from which two earthquake locations are calculated-statistical (average parameter values in PDF) and maximum (for the maximum PDF value). One of the advantages of the program is the ability to correctly estimate location errors for complex error function distributions. Another advantage compared to the LocSAT program is that it allows the use of P-waves arrival times only, which allowed us to study the effect of S-waves on the location results. In our scenario, we used the OctTree Importance Sampling Algorithm, the Equal Differential Time norm and the size of the grid equal to 10 m. As final result, we chose maximum likelihood locations We also fixed the depth with the same values as in the case of LocSAT locations.

BackTrackBB
BackTrackBB (BTBB) is a partial waveform stacking algorithm (PWS) which is the most commonly used type of waveform-based algorithms especially for local seismic networks (Li et al., 2020). PWS, unlike traveltime-based algorithms, does not require manual or automated picking of seismic waves onset times. Instead of this, PWS locates seismic sources based on the coherence of seismic signals. PWS back project seismic waveforms or their characteristic functions into the 4D grid (space and time) and locates earthquake in place where the coherence function reaches its maximum value.
BTBB uses the vertical components of seismograms, which are converted into characteristic functions (CFs) based on kurtosis. CFs are calculated for a set of band-pass filtered time-series obtained with the recursive multiband filtering algorithm Poiata et al., 2016). The final CF is calculated by applying RMS or maximum operator to the set of all CFs. Finally, CF is smoothed with the Gaussian filter. The cross-correlation stacking is then applied for all pairs of characteristic functions to obtain estimate of station-pairs timedelays (TDE). Then, the imaging functions are created by merging the TDE functions of each station pair and mapping them spatially based on the theoretical phase time difference of arrivals. This produces a time sequence of 3D spatial images that depict the probability of each pixel being associated with a source. Seismic event is located where the sum of stacked amplitudes reaches its maximum value 2566 J. Kokowski and L. Rudziń Pure Appl. Geophys. (Poiata et al., 2016). BTBB is able to work on continuous data, using a time window of a specific length. The window is moved, and the calculations are performed for each window separately. For each time window a seismic event can be detected and located.
It is worth mentioning that BTBB software has already been used for the LUMINEOS data. Sobiesiak et al. (2019) used it to analyze waveform data for the short period of time close to one of the strong collapses which occurred in LGCD in the last years. Authors observed an increase in detections of potential seismic events in comparison to STA/LTA triggered catalogs. Sobiesiak et al. (2022) used it also to check the possibility of increase in short-time hazard assessment on a daily basis for Rudna Mine. The work suggests that improved location results can provide support for considering the early warning concept in mining environment.
In our work, we used BTBB only for location procedure. For each event from the catalog, we prepared 5.5-s-long waveforms. The signals started half a second before the origin time obtained by LocSAT. The strong motion data, which are sampled with 250 sps were down-sampled to the sampling frequency of 100 Hz, to be the same as in the velocity seismograms. We configured the software, so that the characteristic functions were calculated in 50 logarithmically spaced filters in the range of 1-49 Hz and time decay constant was equal to 0.4. The example of BTBB result for one of the strong events in Rudna mine is included in Fig. 4. Here the depth was not fixed like in the previous algorithms, however, it is visible that the depth estimation in comparison with epicentral coordinates is solved with high uncertainty (Fig. 4).

Methods
We compared the epicenter locations of seismic events in the time between January 1, 2020, and October 31, 2021 (see events location in Fig. 2). For this period, the mine catalogs contain 7607 events with magnitude completeness 1.15, while the LUMINEOS catalog has 2202 events. However, it should be noted, that the mine catalogs are not truly complete. Especially the catalogs do not contain the events, that occurred during or just after planned blasting works. For this reason, 343 of the 2202 events from the LUMINEOS catalog have no locations in the mine catalogs.
For the entire LUMINEOS catalog, we conducted a detailed completeness magnitude analysis using the maximum curvature method and the code provided by Mignan and Woessner (2012). According to the results the LUMINEOS catalog between January 1, 2020, and October 31, 2021, has magnitude completeness Mc = 1.82. However, the magnitude varies with time and space. The completeness magnitude for the catalog for the period of the day (6AM-6PM) is 1.98 and for the period of the night (6PM-6AM) is 1.7. For the area limited to the center of the network (arbitrarily selected boundaries: Latitude: 51°29 0 -51°34 0 , Longitude: 16°3 0 -16°12 0 , which contain 1031 events) (Fig. 2)  Ultimately, we used 1,859 events to compare the locations accuracy obtained with different approaches. From this group, we distinguished 160 of the strongest events with magnitudes M [ 2.5.
For each of the 1859 earthquakes, we calculated the distances between LocSAT, NonLinLoc (two locations), and BTBB locations from the catalog obtained with in-mine system. Then, we calculated the mean and median distance values for all 1859 events, as well as for the 160 strongest ones. The comparisons are presented in the table (Table 1) and in the histograms (Fig. 5).
Additionally, we performed an analysis of horizontal location uncertainty. To make sure that the uncertainty was calculated with the same method, we calculated it ourselves from variance-covariance matrices of the hypocentral coordinates Eq. (1) (e.g., Havskov et al., 2012). The matrix for a seismic event is obtained by multiplying the variance of the differences between the actual arrival times and the predicted arrival times (i.e. time residuals), with the matrix of partial derivatives of the travel time function with respect to the coordinates of the seismic stations and the origin time of the event Eq. (2). For a matter of simplicity, we decided to calculate partial derivatives for travel time function with constant Vol. 180, (2023) Analysis of the Epicenter Location Accuracy for the Local Seismic Network velocities: 5000 m/s for the P-wave and 2890 m/s for the S-wave. Time variance was calculated according to the formula Eq. (3).
where r 2 is the variance of arrival times multiplied by the identity matrix. G is the known matrix of partial derivatives, which relates changes in travel time to changes in location. G T is G transposed.
where n is a number of observations; ndf is a number of observations minus the number of degrees of freedom, which is equal to 4; tth is a theoretical travel time and t is an observed time at the i-th station. The epicentral location uncertainties were calculated for one standard deviation as the geometric mean of the two axes of the error ellipse. This value is equal to the radius of the circle with the same area as the error ellipse) Eq. (4) (Kijko, 1977).
In order to calculate the uncertainty for the Non-LinLoc, we used all the events in the analyzed catalog for the time period between January 1, 2020, and October 31, 2021. Unfortunately, in the case of LocSAT, we only had access to time residuals for events from January 2020 and January 2021 so in this case the uncertainty was calculated for less events. Time residuals for the BTBB results were calculated as time difference between theoretical times and time corresponding to the maximum values of the characteristic functions in the time windows (Fig. 4b). Therefore, the uncertainty of BTBB can only be considered an approximation and is probably overestimated.
Additionally, we calculated the absolute values of the differences between discrepancy and uncertainty.
Their mean values and medians are also presented in the table (Table 1).

Results
Generally, for the whole dataset, the locations determined with the use of S wave onset have much higher average and median location discrepancies. In the case of locations obtained without the use of S waves onset times (i.e. the NonLinLoc for P phases and BTBB technique), we can observe much smaller discrepancy-especially this applies to mean discrepancy values, which are 2-3 times smaller than for original LocSAT locations. The discrepancy median values are also lower by about 100-150 m. From the Table 1 Mean and median location discrepancy defined as a distance between in-mine and LUMINEOS locations, mean and median values of uncertainty of the locations and mean and median values of absolute values of differences between uncertainty and discrepancy The best results are in green and the worst in red *LocSAT uncertainty was calculated based on January 2020 and January 2021 events only **The uncertainty of the BTBB was estimated as described in the text Vol. 180, (2023) Analysis of the Epicenter Location Accuracy for the Local Seismic Network histograms (Fig. 5) we can conclude that each method returns a similar number of well-located events (discrepancy smaller than 500 m). However, what differs these methods is the large number of very badly located tremors in the case of methods using S waves. This is especially visible for events located with LocSAT. There are almost 300 events located more than 4 km from the ''true'' locations. For events located in the center of the network (orange dashed frame in Fig. 2) and for stronger events, we can observe much smaller discrepancies. However, in the case of methods using S phases, they are still larger. For the best methods (NonLinLoc with P phases and BTBB), the medians and means are always greater than 300 m.
We can also notice that the BTBB works perfectly with stronger events, which is directly related to the number of stations where the events were well registered. The BTBB results for stronger events are actually as good as the ones obtained with Non-LinLoc with manually picked P-wave onsets.
In general, mean and median uncertainties for a 68% confidence interval reflect the mean and median discrepancies of the locations relatively well. However, in the case of locations obtained using S-phases, the uncertainties are clearly lower than discrepancies-especially in the case of the mean values. Interestingly, the opposite is true for the LocSAT method for events in the center of the network. The uncertainty estimated for BTBB reflects true errors relatively well. However, it overestimates the true error values for weaker events. This is probably associated with the way it was estimated. Analyzing the differences between uncertainty and discrepancies allows us to evaluate the value of uncertainty as an indicator of location quality. We can easily notice that the smallest differences occur for events localized by P phases only. For methods using S phases, uncertainty scores are not a reliable indicator of location quality. This is especially visible for locations obtained using the LocSAT algorithm. We can also notice that the method of uncertainty evaluation for the BTBB method is much more reliable for stronger events.
We have also shown the average epicenter location discrepancy on the maps (Figs. 6,7). The mean value was calculated in a 500 m grid as a weighted average, where the weights are distances raised to the -4 power. The map was additionally smoothed with a Gaussian filter. We presented the discrepancy only for seismically active areas, which were defined as areas where at a 2 km distance there are at least 5 events.
Based on the maps, we can conclude that the locations obtained with S-waves have very large discrepancy at the edges of the network (more than 2 km). At the center of the network, the results are comparable for all methods. Interestingly, each of the methods gives the best results in different areas: NonLinLoc (P) achieved the highest accuracy in the northern-central part, LocSAT in the very center of the network and BackTrackBB in the center on the axis of the seismic network.

Discussion and Conclusions
In the work, we used the unique possibility of comparing the locations obtained from the typical surface network seismic catalog, with locations from the underground mine networks, which we consider much more accurate. On this basis, we compared the locations obtained with different algorithms: the iterative LocSAT algorithm, which requires S wave onset times, the probabilistic NonLinLoc algorithm, in two variants-using the P and S wave onset times and with the P wave onset times only, as well as the automated waveform-based BackTrackBB algorithm, which does not use manually picked onset times. Our tests confirmed previous observation (Kokowski & Rudziński, 2023) that in our case the use of S waves negatively affects locations in the LUMINEOS network, especially at the edges of the observing system. These results suggested that it is better to withdraw S-waves from our location procedure. Actually, it is supported by the results of NonLinLoc without S-waves onset times, which give much better locations. Here, the average discrepancy is more than two times smaller in comparison with LocSAT locations.
We also checked to what extent the uncertainty calculated on the basis of residual times and network geometry can allow us to assess the ''true'' location error (i.e. defined by us as a discrepancy). We concluded that for our catalog uncertainty accurately reflects ''true'' errors only for locations obtained using P-phases. The large differences between uncertainty and discrepancy for S-phase methods are probably due to picking errors. The uncertainty estimated by us using the BTBB method gives correct results only in the case of strong events.
Finally, the BTBB automatic locations are very promising, especially in the central area of the network and for strong events, when their discrepancies are comparable to the results of the picking based NonLinLoc algorithm working with the P-phases. Median discrepancies for strong events located by BTBB are even lower than for all other methods. However, at the edges of the network, the algorithm loses its accuracy in the case of weaker events (i.e. the discrepancies are higher), but it is still better than the LocSAT.
These results gave us an important suggestion which improved daily analysis done in the LUMI-NEOS network: we decided to use the P phases only. However, even the best results (i.e. NonLinLoc locations for P wave onset times only) have relatively large epicenter discrepancy. This may be due to the low quality of the velocity model used, which limits the location accuracy for all algorithms. We believe that the current 1D model can be not very accurate and does not take into account the elevation of the stations (i.e. influence of subsurface low-velocityzone is not included). Furthermore, the shape of the network can be changed to surround the seismic sources from the SE and NW with seismic sensors. Our specific conclusion is that in areas with small to moderate induced earthquakes, before any automation is done, the influences of S-wave picking should be carefully checked for all local seismic networks.
Due to the fact that we detect a lot of events at the edge of the network, where the accuracy of the BTBB decreases, we have not yet decided to implement fully automatic BTBB algorithm routinely. However, we are convinced, that if the network completely surrounds the seismically active area and seismic activity is well recorded at a larger number of stations, then automatic BTBB procedure gives as good locations as methods based on travel-times.

Data availability
The data from the LUMINEOS seismic network are available on the EPISODES Platform (IS EPOS, 2017). The data from seismic catalogs from networks belonging to KGHM mines are not publicly available.

Declarations
Conflict of interest The authors have no relevant financial or non-financial interests to disclose.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.