1 Introduction

With the increasing depth and intensity of mining in China, mining-induced earthquakes are occurring with unprecedented frequency and intensity, causing widespread concern among government departments and the public (Cao et al. 2023). Since the sources of mining-induced earthquakes are very shallow (usually less than 1 km), the vibration caused by even a small one must be considered as potentially hazardous for miners' safety and pose a risk for population and infrastructures ground (e.g., Rudziński et al. 2016; Wang et al. 2019). Additionally, such shallow vibrations are of international concern due to the implementation of the Comprehensive Nuclear Test Ban Treaty (Wang et al. 2018).

Source type identification is important for better understanding the physical processes and is a crucial and fundamental issue for hazard assessment and emergency rescue in mining environment (Ma et al. 2018; Wang et al. 2019). Recently, with the utilization of microseismic monitoring and the development of seismic source inversion techniques (Cesca and Grigoli 2015; Kühn et al. 2020), the source types of mining-induced earthquakes are classified and defined at the mine scale, including cavity collapses, rock falls, pillar bursts, explosions, tensional ruptures, and fault-slips (Hasegawa et al. 1989; Cesca et al. 2013).

But at regional scale, mining-induced earthquakes monitored by regional seismic networks are routinely classified directly as collapse earthquakes, without further source categorization and reported to the public when M (magnitude) exceeding 2.3 (http://www.ceic.ac.cn/). In addition, research on the ability of regional station networks to identify source types in mining areas is currently unsystematic and limited. Meanwhile, the internal correlation between source types and rupture processes at mine scale are ambiguous and not well investigated.

The moment tensor (MT) theory is provided as a new method for determining the source mechanism of natural and anthropogenic seismicity. It is widely used for monitoring industry-induced seismicity (e.g., Guilhem et al. 2014; Gu et al. 2018), chemical and nuclear explosions (e.g., Yang and Bonner 2009; Wang et al. 2018; Chiang et al. 2018), volcano-tectonic movements (Hrubcová et al. 2021), as well as rockslide parameter estimation (Dammeier et al. 2015). Full Waveform inversion to determine MT solutions is now a standard method to discriminate different types of seismic sources (Chiang et al. 2016). Further, the MT solutions also provide a quantitative constraint on the source deviation degree from DC (equivalent of shear faulting), moment magnitude and source depth. These parameters are directly associated with the physical source properties as the average slip and the area of the ruptured plane (Liu et al. 2023), which are important to derive adequate seismic hazard relationships (Song et al. 2023a).

However, mining-induced earthquakes are more complex, as they involve multiple geophysical processes that are related to specific tectonic structures and mining conditions. These processes interact and overlay, resulting in a composite fracture mechanism (Šílený and Milev 2008; Rudziński et al. 2016). Moreover, errors in the inversion due to a poor knowledge of the velocity model (Cesca et al. 2006; Cesca and Heimann 2018), waveform quality (Sokos and Zahradnik 2013), limited station azimuth coverage (Ford et al. 2010; Domingues et al. 2013), can also cause the actual source types deviate from the real ones. Both scenarios mentioned above pose a challenge for accurate MT inversion of mining-induced earthquakes and their interpretation.

To illustrate the ability of the regional seismic network, and to discriminate the source types of mining-induced earthquakes, as well as to further seek geomechanical explanations of focal mechanisms, we take two reported mine collapse events in Qufu City and Zoucheng City of Shandong Province in North China as a case study (see Table 1). The Qufu event occurred on July 13, 2020, with a magnitude (ML) of 3.2. Accurate regional seismic location shows that the epicenter is within the boundary of the Xingcun Coal Mine (see Fig. 1). Quakes were felt by residents within a few kilometers of the epicenter was also reported. The Zoucheng event occurred on June 09, 2020, in the Dongtan coal mine (see Fig. 1), with ground subsidence of approximately 20 mm above the epicenter the following day. Due to the deployment of microseismic monitoring systems in the coal mine, this event was recorded by both local microseismic network and regional seismic network. The energy and ML measurements by local and regional stations are \(7.6\times {10}^{5}\) J and 2.9, respectively.

Table 1 Basic information of the two selected events reported by Shandong Earthquake Agency
Fig. 1
figure 1

Location of the study area and distribution of regional seismic stations. The inverted triangles on the map represent the distribution of stations in the study area, with the stations used for inversion shown in orange, and the remaining stations in sky blue. The green and yellow stars represent the locations of the two selected mining-induced events

In this paper, we employed the techniques of Full-waveform inversion and MT decomposition to discriminate the source types for selected events. We also investigated the correlation between source types and geo-mechanical processes. Furthermore, we discussed the results and mechanical modelling of two distinct mining-induced earthquakes. Finally, we addressed the scientific challenges associated with MT inversion in underground mines.

Our research shows that the method utilizing regional Full-waveform inversion can obtain robust full MT solutions of shallow mining-induced earthquakes (small to moderate). We also confirm the capability of identifying source types of mining-induced earthquakes by using regional station data, at least for collapse and DC sources. However, because Green’s functions are not sensitive to some components of the MT, accurately determining the depth of shallow source using MT inversion alone remains challenging.

2 Methods

2.1 Moment tensor inversion

Assuming a point source, displacements in the N direction is expressed as a linear convolution of the seismic MT (M) and the spatial derivatives of the Green’s functions (Aki and Richards 2002):

$${\mu }_{n}\left(t,\overrightarrow{x}\right)={M}_{ij}*{G}_{ni,j}\left(\overrightarrow{x},t\right)$$
(1)

In which \(\mu\) stands for displacement, i and j are the directions of the forces and derivatives (force couples), respectively. Subsequently, MT may be expressed through six fundamental solutions, extensively elucidated within the Mttime code (Chiang et al. 2019).

$$M_{ij} = \mathop \sum \limits_{m = 1}^{6} a_{m} M^{m}$$
(2)
$$\begin{aligned} M^{1} & = \left( {\begin{array}{ccc} 0 & 1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 0 \\ \end{array} } \right)\quad M^{2} = \left( {\begin{array}{ccc} 1 & 0 & 0 \\ 0 & { - 1} & 0 \\ 0 & 0 & 0 \\ \end{array} } \right)\quad M^{3} = \left( {\begin{array}{ccc} 0 & 0 & 0 \\ 0 & 0 & 1 \\ 0 & 1 & 0 \\ \end{array} } \right) \\ M^{4} & = \left( {\begin{array}{ccc} 0 & 0 & 1 \\ 1 & 0 & 0 \\ 1 & 0 & 0 \\ \end{array} } \right)\quad M^{5} = \left( {\begin{array}{ccc} { - 1} & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 1 \\ \end{array} } \right)\quad M^{6} = \left( {\begin{array}{ccc} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \\ \end{array} } \right) \\ \end{aligned}$$
(3)

The final MT can be written as:

$$M_{ij} = \left( {\begin{array}{ccc} {a_{2} - a_{5} - a_{6} } & {a_{1} } & {a_{4} } \\ {a_{1} } & { - a_{2} + a_{6} } & {a_{3} } \\ {a_{4} } & {a_{3} } & {a_{5} + a_{6} } \\ \end{array} } \right)$$
(4)

There are two typical calculation methods for MT inversion: full MT inversion and deviatoric MT inversion. Deviatoric MT solutions can be obtained by setting \({a}_{6}\) to zero, since M6 has a nonzero trace, which corresponds to the change of source volume (Křížová and Málek 2021). The deviatoric MT inversion has been often adopted for tectonic earthquakes and whenever the earthquake is expected to be modeled by a DC source (Liu et al. 2023). But in most cases of induced earthquake, MT solutions often include large relevant non-DC terms. Consequently, a full MT inversion is more appropriate for such instances (e.g., Dammeier et al. 2015; Song et al. 2023a, b).

Full Waveform inversion stands as a customary approach for MT inversion, aiming to achieve the maximum variance reduction (VR) percentage between observed waveforms and synthetic seismograms. Synthetics seismograms are linear combinations of elementary seismograms that can be derived by convolving elementary MT from Eq. (3) with Green’s functions. According to Wang and Herrmann (1980), the wave field, resulting from an arbitrary point dislocation source and a point explosion buried in a plane layered elastic medium, could be delineated by ten fundamental fault Green's functions. This set encompasses eight dislocation sources and two explosion sources, with the eight dislocation sources specifically comprising the vertical strike-slip Green's functions (ZSS, RSS, and TSS), the vertical and radial 45° dip-slip Green's functions (ZDD and RDD), and the vertical dip-slip Green's functions (ZDS, RDS, and TDS), while the two explosion sources contain ZEX and REX Green's functions.

2.2 Source interpretation

Seismic MT (M) can be decomposed into three basic source types (Jost and Herrmann 1989; Norman et al. 1999), namely isotropic (ISO) component, double-couple (DC) component, and the compensated linear vector dipole (CLVD) component. The ISO component corresponds to volume changes in the medium. The DC component corresponds to shear displacements and the CLVD component corresponds to the source where one strain axis is shortening while the other two are lengthening (Vavryčuk 2015).

To discriminative representation of the source types in terms of the ISO, DC, and CLVD components, a lune diagram (Tape and Tape 2012a; 2012b; see Fig. 2) is introduced here to ease the visualization of the source types. The fundamental lune longitude γ (deviatoric component) and latitude δ (isotropic component) were calculated:

$$\mathrm{tan}\,\upgamma =\frac{-{\uplambda }_{1}+2{\uplambda }_{2}+{\uplambda }_{3}}{\sqrt{3\left({\uplambda }_{1}-{\uplambda }_{3}\right)}}$$
(5)
$$\mathrm{cos}\,\upbeta =\frac{{\uplambda }_{1}+{\uplambda }_{2}+{\uplambda }_{3}}{\sqrt{3\left({{\uplambda }_{1}}^{2}+{{\uplambda }_{2}}^{2}+{{\uplambda }_{3}}^{2}\right)}}$$
(6)
$$\updelta =\frac{\uppi }{2}-\upbeta$$
(7)

where β is the colatitude and λ1–31 ≥ λ2 ≥ λ3) are the Eigen values of the MT.

Fig. 2
figure 2

A schematic diagram of failure models, their moment tensors, beachballs, and the corresponding positions in the Lune diagram. An example of diagonalised MT matrix is provided to demonstrate the relationships between its Eigen values

In this convention, when γ = 0, the δ =  ± π/2 correspond to ± ISO sources, which represent pure changes in volumes, i.e., spherical explosion and implosion, respectively. When δ = 0, γ =  ± π/6 describes ± CLVD source, reflecting the mechanical properties of compression/extension. A pure DC sources are described by γ = δ = 0.

In the mining environment, an explosion (mining blast) source reflects volume changes and should be projected onto the point of + ISO in the lune diagram (Ma et al. 2018). A collapse source should be projected onto the point of -Crack for a Poisson's ratio of 0.25, which can represented by a pure closing crack component, often driven by gravity (Bowers and Hudson 1999; Tape and Tape 2012a). A shear source (fault-slip) can be represented by a pure DC component, which is equivalent to tectonic earthquakes.

3 Technical details

3.1 Crustal model

We used a one-dimensional layered crustal model, provided by the Shandong Earthquake Agency (referred to as the SD model hereafter), to calculate Green's functions. In order to find a solution with the high VR, we added velocity perturbation within ± 5% (set the step size to 2%) for both P waves and S waves to correct the deviation of crustal anisotropy on the calculation of Green's functions (see Fig. 3). We used seven models to represent the velocities of P waves and S waves, varying from 95 to 105% in 2% increments, based on the standard deviation (SD) model for different stations and sources. The choice of model is determined by the optimal MT solution.

Fig. 3
figure 3

A 1D-layered crustal model is depicted with superimposed velocity perturbations. Solid lines illustrate the original steady-state (SD) model, while dotted lines show the model with additional velocity perturbations

3.2 Calculation procedures

High-quality broadband waveform data from 13 regional stations provided by Shandong Earthquake Agency, characterized by a favourable signal-to-noise ratio (SNR) and no disturbances, were collected and employed. The data were then inverted for deviatoric and full MT solutions with a step of 0.1 km in depth for the search.

We used a Full-waveform MT inversion method in low frequency based on the Mttime software (Chiang and USDOE National Nuclear Security Administration 2020). The waveforms were firstly instrument corrected to ground displacement, tapered and rotated to the great-circle path, and decimated to 5 samples per second. Then, we calculated Green's functions and generated synthetic seismograms by using the frequency-wave number integration method (Wang and Herrmann 1980; Herrmann 2013) based on the velocity structure model of each individual station and source (see Sect. 3.1). To explore the optimal filtering parameters of the full-waveform inversion, we test the different low-frequency filter bands up to 2 Hz to find a good signal with low noise. Finally, after many tests with different frequency ranges and crustal models, we obtained optimal MT solutions, source depths, fault plane parameters (strike/dip/rake), and seismic moment (Mo) estimated, which were calculated according to the maximum VR between the observed waveforms and synthetic seismograms. It is worth noting that one of the advantages of the inversion method is that it allows for minor time offsets between the data and the synthetic seismograms. Thus, it is possible to compensate for uncertainties in origin time, location, and velocity structure (Pasyanos and Chiang 2020).

4 Regional moment tensor inversion

4.1 MT solutions of the Qufu event

High SNR data from 7 regional broadband stations are selected and used for both deviatoric and full MT inversion, based on the SD model. In contrast to full MT inversion, the VR of deviatoric MT inversion is lower, accompanied by poorer fitness of the observed waveforms and synthetic seismograms, especially in the Z (vertical) and R (Radial) directions (see Figs. 4 and 5). This is mainly caused by the lack of ISO components (when set a6 to zero), which are related to two explosion sources Green's functions of ZEX and REX.

Fig. 4
figure 4

Deviatoric MT inversion based on the SD model for the Qufu event. The best solution and waveform fits (the black observed and red synthetic waveforms) at selected stations. Red triangles around the beach ball indicate station azimuth, and the numbers below the Z and R component seismograms are, from left to right: distance, azimuth, and station-specific VR

Fig. 5
figure 5

Full MT inversion based on the SD model of the Qufu event. Symbols and colors are the same as in Fig. 4

Furthermore, we conducted full MT inversion based on the perturbation velocity structure model, utilizing the same records and filter parameters (Table 2). The results show an increase in total VR from 73 to 75%, while the optimal source depth increased from 0.3 to 0.9 km. Additionally, the MW decreased from 3.5 to 3.2, indicating that the difference between the MW and the ML measured by the regional station network has also been reduced (see Fig. 6). Therefore, we believe that the accuracy of the results can be effectively improved by incorporating velocity structure model perturbations in the inversion process. From the MT decomposition results, we can also clearly see that the ISO component increased from 49% to 56%, accompanied by a decrease in the same scale of the DC component from 13% to only 6%. Noteworthy, the phenomenon of weak amplitudes on T (tangential) components at most stations also reflect a small proportion of DC component (see Fig. 6). The very large relevant non-DC terms indicate that the source of this event significantly deviated from the tectonic earthquakes. The results with maximum VR of this event are summarized in Table 3.

Table 2 Crustal models and the frequency ranges applied
Fig. 6
figure 6

Comparison of full MT moment solutions based on the SD model and perturbation models is presented. Observed waveforms are represented by black lines, while synthetic waveforms based on the SD model are shown in red and those based on perturbation models are in blue. The VR of each component is written to the right of each synthetic seismograms. The last column depicts the average VR values for each station

Table 3 Full MT inversion results

4.2 MT solutions of the Zoucheng event

We used the data from 14 components of 8 regional broadband stations to conduct both deviatoric and full MT inversion based on the perturbation velocity structure model. We have also conduct many tests to find suitable filtering bands and velocity structure models to improve the SNR and VR (see Table 2). Figure 7 shows the deviatoric and full MT inversion based on perturbation model for the Zoucheng event. The station azimuths span 178°, primarily due to the poor data quality in the northwest direction of the source, and the maximum epicenter distance is 122 km. Since the ISO component of the full MT solutions is very small (only 8%), the beach ball geometries constrained by both calculation methods are almost the same, indicating a similar focal mechanism. The values of VR are almost equal, while the source depth varies considerably (0.2 km and 0.5 km in deviatoric and full MTs, respectively). The results with maximum VR of this event are summarized in Table 3. It is worth noting that the DC component is dominant in both scenarios. According to the classification method by Vavryčuk (2015), the source type of this event could be identified as a pure shearing one (with the DC component percentage exceeding 60%).

Fig. 7
figure 7

Deviatoric and Full MT inversion were conducted using a perturbation model. The best solutions and waveform fits (the black observed and blue synthetic seismograms) at selected components of stations

To illustrate the accuracy of our calculations, we collected microseismic sensors data in a local mining environment. The strike length of the working face in the mine is 1456.3 m, with a direction of N30°E, a dip length of 261.0 m, and an average buried depth of − 670 m. The microseismic network of this mine consists of 27 sensors. Precise localization based on the microseismic network shows that this event occurred on the roof of panel 6306, No. 6 Mining Area of Dongtan mine, at a depth of − 593 m when the working face had advanced 755 m (see Fig. 8).

Fig. 8
figure 8

The tectonic setting and the division of working faces in the No.6 mining area in Dongtan mine

Although this event was recorded by most of the microseismic sensors, only 9 sensors provided clear P-wave first-motion directions (see Table 4). Therefore, we could not accurately constrain the source mechanism only by P-wave first-motions (Bentz et al. 2018). Nevertheless, the full MT solutions inversion by regional network plot together with signs for first motions shows a high level of confidence, i.e., eight out of nine signs are projected in the ‘right’ place, with positive and negative signs appropriately situated in regions corresponding to compressive and tensile stresses, respectively (see Fig. 9).

Table 4 Basic information of local microseismic stations in Dongtan mine (Zoucheng event)
Fig. 9
figure 9

Beach ball derived from Full MT solutions, plotted using a regional network, accompanied by symbols that represent the initial motion signs from local micro-seismic sensors

5 Further analysis and discussion

Based on the fundamental lune visualization method proposed by Tape and Tape (2012a), we propose an innovative contour lune diagram to visualize source types and source depths. The full MT solutions for different source depths of the Qufu and the Zoucheng events are plotted on the fundamental lune together with VR contours calculated by the Kriging interpolation method (see Fig. 10).

Fig. 10
figure 10

A source-type plot after Tape and Tape (2012a).The VR contours as a function of depth are plotted on the fundamental lune in a and b. The Source types inferred from full MT solutions of two selected events are plotted on the fundamental lune (c), along with the 6 August 2007 Utah collapse (blue square) and the 3 September 2017 DPRK collapse (red square)

5.1 Source type of the Qufu event and its geomechanical interpretations

Xingcun Mine, where the Qufu event occurred, is mining deep coal seams to a depth of more than 1 km. The source, at a depth of approximately 0.9 km constrained by full MT inversion, is more likely to be located on the roof.

The lune diagrams, as illustrated in Fig. 10c, indicate that the projection of the source type for the Qufu event is positioned in the vicinity of negative Crack point, signifying the characteristic of collapse source (see Sect. 2.2). In addition, the projected source type for this event is in the same region as the 6 August 2007 Utah collapse in the USA and the 3 September 2017 DPRK collapse that followed the nuclear explosion. Previous research results show that both collapses were caused by vertical fracturing of the roof under gravitational forces (Pechmann et al. 2008; Chiang et al. 2018).

Theoretically, collapse events in mining areas could be conceptualized either as a falling-block model (Taylor 1994) or as a tabular excavation collapse. Those models are mathematically equivalent to a closing crack, as demonstrated by previous studies (e.g., Pechmann et al. 1995; Bowers and Walter 2002). In this context, the scenario is characterized by a diagonal MT with the ratio \(\mu :\mu :\mu + 2\lambda\) (\(\mu\) and \(\lambda\) are the Lamé constants), while the other six components remain zero.

To quantify the similarity between the source type of the Qufu event and the collapse source, we performed an alternative MT decomposition, following an approach of Ford et al (2008). The Full MT solution is decomposed into the simple closing tensile crack model (Mcrack) plus secondary components contained in a remainder moment tensor (Mrem):

$$M={M}^{Crack}+{M}^{Rem}$$
(8)

In which MCrack can be represented as the sum of a negative isotropic and a CLVD with a negative vertical major dipole:

$${M}^{Crack}={M}^{ISO}+{M}^{Clvd}$$
(9)

Assuming that all volume changes are due to collapse, MCrack moment ratio is [1:1:3] when the Poisson's ratio is 0.25 (Jost and Herrmann 1989).

Thus, the full MT solution solutions for the Qufu event can be decomposed as follows:

$$\left[\begin{array}{ccc}-3.26& -0.26& 0.74\\ -0.26& -5.18& 2.70\\ 0.74& 2.70& -8.49\end{array}\right]=\left[\begin{array}{ccc}-3.38& 0& 0\\ 0& -3.38& 0\\ 0& 0& -10.15\end{array}\right]+\left[\begin{array}{ccc} 0.12& -0.26& 0.74\\ -0.26& -1.80& 2.70\\ 0.74& 2.70& 1.66\end{array}\right]$$

In which each moment is in units of 1020 dyne-cm. This result indicates that the MCrack represents 75% of the total moment, which is consistent with the weak amplitudes on T components at stations (see Figs. 5 and 6), since pure closing crack source does not generate any Love waves (Wang et al. 2019).

Although the proportion of the DC component is only 6%, the T components of some stations, such as JINX, TIA, XIT and LFS, show high fitting coefficients between observed waveforms and synthetic seismograms (see Fig. 6). Accordingly, we cannot attribute the generation of DC components solely to noise in waveforms or anisotropic errors in Green's functions. One possible explanation is that other dynamic mechanisms, such as fault sliding, were superimposed on the collapse process. This fault-slip may have occurred before and induced the collapse, or it may have occurred simultaneously with the collapse (Rudziński et al. 2016).

5.2 Source type of the Zoucheng event and its geomechanical interpretations

The lune diagrams (see Fig. 10) show a clear difference in source types between the Qufu and the Zoucheng event. The projections of the source type for the Zoucheng event are located in the region around the point of DC (see Fig. 10b), reflecting the shear source property.

Generally, the dominant DC component indicates fault activity for natural earthquakes (Gu et al. 2018; Vavryčuk et al. 2021). In mining environment, the dominant DC component is usually related to fault activity or the fracture of the roof caused by stress changes due to mining disturbances (Ma et al. 2019; Mendecki et al. 2020). The previous geological survey in the mineral deposit and three-dimensional (3-D) geophysical data analysis indicate that there are no exploitable faults within 500 m of the source (see Fig. 8). Consequently, the occurrence of the event was most likely independent of tectonic activity (Cai et al. 2019); (see Fig. 11).

Fig. 11
figure 11

Explanatory diagram of the geo-mechanical processes of the Zoucheng (Dongtan) event

The column of the D7 borehole (see Figs. 8 and 11) shows that the immediate roof consists of siltstone intercalated with fine sandstone, with fine sandstone with a thickness of 2.45 m. In particular, about 26 m above the coal seam, there is a thick and hard layer of Jurassic rock, flesh-colored, consisting mainly of sandstone, siltstone, fine sandstone, and mudstone. Analyzed in conjunction with micro-seismic sensor data, the geologic setting, and mining progress, we infer that as the panel continues to advance, the overlying strata continue to buckle and sink due to being unsupported. Shear failure along the weak stress surface occurs in the thick and hard stratum in the case of the roof overhanging area exceeding a certain threshold value (Wang et al. 2016; Jiao et al. 2021). The results from one of the fault plane show that the dip angle of source fracture surfaces is 86° (see Table 5), which can also support our above inferences.

Table 5 Fault planes of Zoucheng (Dongtan) event constrained by full MT solutions

5.3 Resolution of inversion depth

As shown in Fig. 10a and b, we see that shallow depths within 1.0 km are preferred, reflecting the fact that the depth of shallow source can be constrained at kilometer scales by MT solutions. The substantial reduction in VR from 1 to 2 km depth is likely due to the presence of a velocity discontinuity in the structure used to compute the Green's functions (see Fig. 3). However, the depths of both selected events within 1.0 km showed similar fitness with all VR values more than 60%, reflecting the insensitivity of the MT solutions to depths within this range. According to synthetic seismograms study conducted by Chiang et al (2016), the Green's functions are not sensitive to some components of the MT, especially those corresponding to the source of ISO components, so accurately determining the depth of shallow source by MT solutions is still challenging.

5.4 Scientific problems of the MT inversion

The source-types analysis of microseismic monitoring in mines has been extensively studied (e.g., Ma et al. 2018; Song et al. 2023b). However, research on the ability of regional seismic networks to identify source types in mining areas is currently unsystematic and limited. Mining-induced earthquakes, distinct from natural seismic events, are characterized by their occurrence in the shallow near-surface region, resulting in the generation of surface waves that can be detected over relatively short distances by regional stations (Tibi 2021). Consequently, this feature enhances the utility of data from regional seismic networks, enabling the imposition of more stringent constraints on focal mechanisms by regional MT inversion method (Dreger et al. 2008; Whidden and Pankow 2016).

Generally, high-energy mining-induced earthquakes of high energy could be recorded simultaneously by regional seismic networks and local microseismic networks (e.g., Lizurek et al. 2015; Rudziński et al. 2016). Regarding inversion capabilities, the regional MT solutions are less satisfactory compared to those derived from microseismic data, especially when the magnitude is relatively low. However, for high-energy events, the application of low-frequency filtering during the inversion process greatly reduces the dependence on the medium velocity model, thereby enhancing the reliability of the results (e.g., Ford et al. 2008; Wang et al. 2019).

Alternatively, according to Rudziński et al. (2016), certain mining-induced earthquakes are remarkable for their complex source configurations, indicating that multiple stages of the rupture process may be involved. In such case, the microseismic networks could be very useful to obtain information about the initial stage in the development of the seismic process. Regional seismic data, concurrently, are crucial for gaining insights into the largest event within a sequence, enabling a more accurate assessment of seismic energy, seismic moment, and faulting mechanism. Furthermore, it should be noted that the accuracy of inversion results from microseismic networks depends on the reasonable orientation of stations and the scientific setting of instrumental response parameters (e.g., Domingues et al. 2013; Lizurek et al. 2015). Overall, by combining the MT solutions from microseismic networks with those from regional seismic networks, it is possible to gain insights into the entire rupture process of a source.

In this study, we also aimed to investigate whether the rupture behavior of the Zoucheng (Dongtan) event exhibits complexity. However, when attempting to solve for the MT solution at the mine scale, we faced challenges due to partial signal saturation of the S-waves, which imposes constraints on our ability to employ Full-waveform inversion. Then, we attempted to utilize the First Amplitude inversion method proposed by Kwiatek et al. (2016) to carry out MT inversion. However, despite our best efforts to gather high-quality data, we were only able to collect nine waveforms with distinct initial motion signatures. The insufficient coverage of station azimuths ultimately limited our capacity to robustly constrain the source mechanism during the initial rupture phase. Further, we have resorted to collecting P-wave initial motion signs to confirm the regional MT solution. We have found a good consistency between local P-wave initial motion of and the regional results, indicating a high probability of the same source type. However, the findings here alone does not yet have complete persuasive power. Fortunately, the solution from the regional seismic network corresponds to a high VR coefficient, reflecting the robustness of the results, from which we can derive the key source parameters of the maximum energy event throughout the sequence.

6 Conclusion

To illustrate the ability of the regional seismic network, to discriminate the source types of mining-induced earthquakes, and to further seek geo-mechanical explanations of source types, in this study, we employed the method of MT inversion combined with source interpretation to reveal source types of two strong, reported “collapse earthquakes” by the regional seismic network. The main conclusions are drawn as follows:

  1. (1)

    Regional full-waveform inversion is effective to obtain robust full MT solutions for shallow mining-induced earthquakes (small to moderate in size). It can be further used to identify source types, at least for collapse and double-couple (DC) sources.

  2. (2)

    The source of the Qufu event is more consistent with the theoretical models of collapse seismic sources and could be regarded as a collapse event. It is confirmed to be the result of vertical rupture of the roof under the effect of gravity. In contrast, the source of the Zoucheng event, which contained a large proportion of DC components, exhibits the characteristics of natural earthquake. The fracture-slip of a thick-hard roof due to an overlying load is a plausible geo-mechanical interpretation.

  3. (3)

    Although MT solutions can constrain the depth of shallow sources at kilometer scales, accurately determining the depth of a shallow source using MT solutions alone is still challenging.”