 Full paper
 Open access
 Published:
Comparison of statistical lowfrequency earthquake activity models
Earth, Planets and Space volume 76, Article number: 59 (2024)
Abstract
Slow earthquakes are slow fault slip events. Quantifying and monitoring slow earthquake activity characteristics are important, because they may change before large earthquakes occur. Statistical seismicity models are useful for quantifying seismicity characteristics. However, no standard statistical model exists for slow earthquake activity. This study used a highquality catalog of lowfrequency earthquakes (LFEs), a type of slow earthquake, in the Nankai subduction zone from April 2004 to August 2015 and conducted the first comparison of existing statistical LFE activity models to determine which model better describes LFE activity. Based on this comparison, this study proposes a new hybrid model that incorporates existing model features. The new model considers the LFE activity history in a manner similar to the epidemictype aftershock sequence (ETAS) model and represents the LFE aftershock rate (subsequent LFE occurrence rate) with a small number of model parameters, as in the Omori–Utsu aftershock law for regular earthquakes. The results show that the proposed model outperforms other existing models. However, the new model cannot reproduce a feature of LFE activity: the sudden cessation of intense LFE bursts. This is because the new model superimposes multiple aftershock activities and predicts extremely high seismicity rates during and after the LFE bursts. I suggest that reproducing and successfully predicting the sudden cessation of intense LFE bursts is critical for the further improvement of statistical LFE activity models. In addition, the empirical equations formulated in this study for the LFE aftershock rates may be useful for future statistical and physical modeling of LFE activity.
Graphical Abstract
Introduction
Slow earthquakes are a general term used to describe various spontaneous slow faultslip events, such as lowfrequency earthquakes, tectonic tremors, verylowfrequency earthquakes, and slow slip (Ide et al. 2007; Ide and Beroza 2023). They are characterized by longer duration than regular (fast) earthquakes of comparable seismic moment. Slow earthquakes have been observed or inferred to occur at several subduction and transform plate boundaries (e.g., Dragert et al. 2001; Obara 2002; McGuire et al. 2005; Nadeau and Dolenc 2005; Kuna et al. 2019). In particular, very detailed slow earthquake activity has been revealed in subduction zones such as the Nankai Trough, Cascadia subduction zone, and the Hikurangi and Japan Trenches (Beroza and Ide 2011; Obara and Kato 2016; Wallace 2020; Nishikawa et al. 2019).
In subduction zones, slow earthquakes occur mainly on the immediate deep and/or shallow sides of megathrust earthquake sources. The relationship between slow and megathrust earthquakes has been actively studied (e.g., Obara and Kato 2016; Nishikawa et al. 2023). Simulation studies have indicated that slow earthquake activity characteristics may change prior to megathrust earthquakes (e.g., Matsuzawa et al. 2010; Luo and Liu 2019). Therefore, quantifying and monitoring slow earthquake activity characteristics are important for gaining insight into the imminence of future megathrust earthquakes.
Statistical seismicity models are useful for quantifying seismicity characteristics. For example, the epidemictype aftershock sequence (ETAS) model is widely used to quantify the characteristics of regular earthquake activity (Ogata 1988; Zhuang et al. 2002; Kumazawa and Ogata 2013; Nishikawa and Nishimura 2023). This model was also used as a standard model in an international earthquake forecasting experiment named the Collaboratory for the Study of Earthquake Predictability (Schorlemmer et al. 2018). In contrast to the regular earthquake activity, there is no standard statistical model for slow earthquake activity. Statistical modeling of lowfrequency earthquakes (LFEs), a type of slow earthquake, has only recently begun. LFEs are characterized by low dominant frequencies (1–10 Hz) compared to regular microearthquakes of comparable seismic moment (Shelly et al. 2007; Nishikawa et al. 2023). Lengliné et al. (2017) and Tan and Marsan (2020) proposed statistical LFE activity models similar to the ETAS model. In their models, the LFE occurrence rate was assumed to be the sum of the stationary background rate and the effect of past LFEs or external processes behind past LFEs inducing future LFEs.
Ide and Nomura (2022) proposed a statistical model for tectonic tremors (i.e., LFE swarms) (Shelly et al. 2007). Their model is based on an approach different from that of Lengliné et al. (2017) and Tan and Marsan (2020). Ide and Nomura (2022) described the probability distribution of tremor interevent times using a mixture of lognormal and Brownian passagetime (BPT) distributions and forecasted tremor interevent times. Furthermore, their model depends only on the time elapsed since the last event and does not depend on detailed tremor activity history.
Constructing a statistical model that successfully describes the LFE activity is important for better characterization and forecasting. However, existing statistical LFE activity models (Lengliné et al. 2017; Tan and Marsan 2020; Ide and Nomura 2022) have never been compared, and it is unclear which model best describes LFE activity. This study applies the statistical LFE and ETAS models (Ogata 1988; Lengliné et al. 2017; Ide and Nomura 2022) to LFE activity along the Nankai Trough and compare their performances using Akaike’s information criterion (AIC; Akaike 1974) (Sect. “Results”). Based on the model comparison, I propose a new model that incorporates existing model features and discuss LFE activity occurrence patterns (Sect. “Discussion”).
Data and methods
LFE catalog and study regions
A highquality catalog of LFEs along the Nankai Trough in southwest Japan created by Kato and Nakagawa (2020) (Fig. 1a) was used. In the Nankai Trough, the Philippine Sea Plate subducts under the Amurian Plate. This plate subduction causes diverse slow and fast earthquakes at the plate boundary (Obara and Kato 2016; Takemura et al. 2023). Kato and Nakagawa (2020) detected LFEs along the Nankai Trough from April 2004 to August 2015 using a matchedfilter technique, using LFEs in the Japan Meteorological Agency (JMA) catalog as templates. This catalog contained approximately 510,000 LFEs from M − 1.1 to M 1.4 (Fig. 1b), which is approximately 23 times the number of LFEs in the JMA catalog over the same period.
When analyzing the LFE activity, it is necessary to determine the minimum magnitude of the LFEs to be analyzed. In regular earthquake activity analyses, the earthquake catalog completeness is typically evaluated, and the completeness magnitude is used as the minimum magnitude (e.g., Nishikawa and Nishimura 2023). However, no method currently exists for evaluating the completeness of an LFE catalog. Therefore, as a compromise, the goodnessoffit test (GFT) method (Wiemer and Wyss 2000; Woessner and Wiemer 2005), which has been used to evaluate the completeness magnitude of regular earthquake catalogs, was applied to the LFE catalog, and the obtained magnitude was used as a reference to select the minimum magnitude. All LFEs in the Kato and Nakagawa (2020) catalog were used regardless of their source locations. See Woessner and Wiemer (2005) for details of the GFT method.
In the GFT method, the R value is a measure of the goodnessoffit between the observed number of earthquakes and the number of earthquakes expected based on the Gutenberg–Richter relationship (Gutenberg and Richter 1944). Specifically, the R value is the absolute difference between the observed and expected number of events in magnitude bins and is given by
where \({B}_{i}\) and \({S}_{i}\) are the observed and expected number of events in each magnitude bin (\({M}_{i}\Delta M/2\) to \({M}_{i}+\Delta M/2\)), respectively, \(M\) is the magnitude for which the R value is calculated, and \({M}_{{\text{max}}}\) is the catalog maximum magnitude. \(\Delta M\) is the width of the magnitude bins and was set to \(\Delta M=0.1\) in this analysis. A perfect fit between the observed and expected values yields \(R\) of 100%. See Wiemer and Wyss (2000) for more details on R value calculation.
I searched for the smallest magnitude with an R > 95% (Fig. 1c). The resulting magnitude was M 0.4. Considering this estimate of the GFT method, LFEs with M 0.6 or greater were used, which is M 0.2 greater than the estimate. This is because I prefer the minimum magnitude to be as conservative (or as large) as possible. However, if the minimum magnitude is too large, the number of LFEs will be too small for statistical analysis. In the following analyses, I used 3545 M 0.6 or larger LFEs from Kato and Nakagawa’s (2020) catalog. In the above analysis, I followed Bostock et al. (2015) and assumed that the LFE magnitude–frequency distribution (especially for large events in the catalog) obeys the Gutenberg–Richter law. I further discuss this assumption and the influence of the minimum magnitude selection on my results in Sect. “Selection of the minimum magnitude”.
In the Nankai Trough, LFEs occur in bandlike regions downdip of the megathrust seismogenic zones (Fig. 1a). These bandlike regions were divided into subregions in a manner similar to that of Ide and Nomura (2022). In and around the bandlike regions, rectangles of 0.2° in latitude and longitude were placed at 0.1° intervals to create subregions (Fig. 1a). Each subregion was required to contain at least 100 M \(\ge\) 0.6 LFEs, and LFE activity in each subregion was analyzed. Consequently, 43 subregions were included in the analysis.
Statistical LFE activity models
Lengliné et al. (2017) model
Lengliné et al. (2017) proposed a statistical LFE activity model similar to the ETAS model, see Sect. “ETAS model” for details on the ETAS model. Their model assumed the rate of LFE occurrence \(\lambda\) at time \(t\) to be the sum of a stationary background rate \({\mu }_{L}\) and the effect of past LFEs or external processes behind the LFEs inducing future LFEs:
where \({a}_{i}\) (\(\ge 0\)) are parameters controlling the magnitude of the effect of the ith event inducing following events, \({t}_{i}\) is the occurrence time of the ith event, and \(g\) is the normalized aftershock rate kernel. In this study, an event refers to the combined phenomenon of an LFE and possible external processes behind it. Lengliné et al. (2017) interpreted the external process as a transient slip rate increase in the area surrounding LFEgenerating asperities. In the above formulation, \({a}_{i}\) is equivalent to the number of events induced by the ith event.
The most significant difference between the above model and the ETAS model is that in the former, the shape of the normalized kernel \(g\) is not assumed a priori but is determined based on the observed data. In contrast, the ETAS model assumes a simple powerlaw decay (\(\sim {t}^{p}\)) for the aftershock rate kernel. Lengliné et al. (2017) used piecewise constant discretization for \(g\):
where \({T}_{k}\) is the time interval used for discretization and \(k\in \left[1:N\right]\), with \({T}_{1}=0\) (days) and \({T}_{N}=10\) (days). Furthermore, when \({T}_{N}<t\), \(g\) is assumed to be zero:
The aftershock rate kernel \(g\) is normalized to satisfy the following equation:
Lengliné et al. (2017) used approximately 30 piecewise constants to discretize the kernel \(g\). They applied their model to LFE activity in the Mexican and Cascadia subduction zones and on the San Andreas Fault.
The Lengliné et al. (2017) model has the drawback of numerous parameters (usually hundreds or more), because the number of parameters \({a}_{i}\) increases with the number of events. This is a major disadvantage when compared to other models using AIC, because AIC substantially increases with the number of model parameters, see Sect. “Model performance evaluation” for details on AIC. In my preliminary analyses, the original Lengliné et al. (2017) model had a much larger AIC than the Ide and Nomura (2022) and ETAS models, because it had too many model parameters. This suggests that the original Lengliné et al. (2017) model is generally inferior to other models. Therefore, in this study, the following model with significantly reduced model parameters was used:
The triggering effect magnitude (or aftershock rate amplitude) was assumed to be constant for all events (i.e., \({a}_{i}=a\)). Tan and Marsan (2020) reduced the parameters of Lengliné et al. (2017) model in a similar manner. However, the model proposed by Tan and Marsan (2020) is a spatiotemporal LFE model, whereas the above model is purely temporal. I used the temporal model, because Ide and Nomura (2022), which is used for comparison, is a temporal model. Hereafter, the above model is referred to as the Ltype model.
New parameters \({g^{\prime}}_{k}\) and kernel function \(g^{\prime}\) were defined as follows:
Using the new kernel function \({g}^{\prime}\), Eq. 6 can be rewritten as
The likelihood \(L(\theta X)\) of this model is given by
where \(\theta\) represents a set of the model parameters and \(X= \{{t}_{i}  0 \le {t}_{i}\le T\}\) are the data. I obtained the model parameters (\({\mu }_{L}\) and \({g^{\prime}}_{k}\)) that maximized the above likelihood via the Powell method (Powell 1964). In this study, I used 15 piecewise constants (\({g^{\prime}}_{1}\) to \({g^{\prime}}_{15}\)) to discretize the aftershock rate kernels. I assigned \({g^{\prime}}_{1}\) in the period from 0 to 10^{–4} days. I then divided the period from 10^{–4} to 10 days into 14 divisions evenly spaced in logarithmic space and assigned \({g^{\prime}}_{2}\) to \({g^{\prime}}_{15}\), respectively. In addition, Eq. 7 was used to obtain the parameter \(a\), which indicates the number of events induced by a single event.
Lengliné et al. (2017) apply their model to LFEs belonging to the same family. An LFE family is defined as a group of LFEs with similar waveforms. However, as Lengliné et al. (2017) acknowledge, there is no theoretical restriction that this model should only be applied to events belonging to the same LFE family. Therefore, in this study, the model was applied to all LFEs that occurred within the same subregion (see Sect. “LFE catalog and study regions”).
Ide and Nomura (2022) model
Ide and Nomura (2022) represented the probability distribution of interevent times using a mixture of lognormal and BPT distributions, which correspond to shortterm clustering and longterm recurrence of events, respectively. Ide and Nomura (2022) applied their model to deep tectonic tremors along the Nankai Trough. Upon applying the Ide and Nomura (2022) model to LFE activity along the Nankai Trough, an additional lognormal distribution was added to the model. This is because, as discussed in Sect. “Subregion examples”, shortterm LFE clustering often displays two characteristic timescales (tens of seconds and a few hours). Kato and Nakgawa (2020) refer to this LFE activity feature as intermittency. Tan and Marsan (2020) also reported two characteristic timescales for shortterm LFE clustering along the San Andreas Fault.
The model consists of two lognormal distributions (the first and second terms on the righthand side) and one BPT distribution (the third term on the righthand side):
where \(\Delta t\) is the interevent time, and \(f\left(\Delta t\right)\) is the probability density function of \(\Delta t\), \({\mu }_{1}\), \({\mu }_{2}\), and \({\mu }_{3}\) (\({\mu }_{1}<{\mu }_{2}<{\mu }_{3}\)) are parameters representing characteristic time scales of the lognormal and BPT distributions, \({\sigma }_{1}\) and \({\sigma }_{2}\) represent \({\text{log}}\left(\Delta t\right)\) variances of the first and second lognormal distributions, respectively, \({\sigma }_{3}\) controls the variation in LFE burst longterm recurrence periods, and \({\phi }_{1}\) and \({\phi }_{2}\) are parameters controlling the proportions of the first and second lognormal distribution. We call the above model the INtype model.
An important feature of the INtype model is that the probability density distribution \(f\left(\Delta t\right)\) does not depend on detailed LFE activity history. This differs markedly from the Ltype and ETAS models, which consider the superimposition of multiple LFE aftershock rates (see Eq. 6).
The likelihood \(L(\theta X)\) of the INtype model is given by
where \({t}_{i}\) is the ith event occurrence time. Eight model parameters (\({\mu }_{1}\), \({\mu }_{2}\), \({\mu }_{3}\), \({\sigma }_{1}\),\({\sigma }_{2}\), \({\sigma }_{3}\), \({\phi }_{1}\), and \({\phi }_{2}\)) were obtained that maximize the likelihood via Powell’s method (Powell 1964).
The event occurrence rate expected from the INtype model (Ide and Nomura 2022) is given by
These equations were used to compare the expected LFE aftershock rates from the different models in Sect. “Results”.
ETAS model
The ETAS model (Ogata 1988), which is similar to the Ltype model (Lengliné et al. 2017), assumes that the event occurrence rate is the sum of the stationary background rate and the effect of past events inducing future events. However, the aftershock rate kernel of the ETAS model differs from that of the Ltype model, and the ETAS model assumes powerlaw decay a priori (Omori–Utsu’s aftershock law; Utsu 1957; Utsu et al. 1995). In the ETAS model, the event occurrence rate \(\lambda\) at time \(t\) is written as follows:
where \({\mu }_{E}\) is the stationary background rate, the second term of the righthand side is the aftershock rate summation controlled by four parameters (\(\alpha\), \(c\), \(K\), and \(p\)), \({t}_{i}\) and \({M}_{i}\) are the occurrence time and ith event magnitude, respectively, and \({M}_{c}\) is the catalog minimum magnitude. As with the Ltype model, the likelihood of the ETAS model \(L(\theta X)\) is given by Eq. 10. The ETAS parameters that maximized the likelihood were obtained using Powell’s method (Powell 1964).
Model performance evaluation
The Ltype, INtype, and ETAS models were compared based on AIC (Akaike 1974), which is defined as follows:
where \(k\) denotes the number of adjusted parameters. The Ltype, INtype, and ETAS models had 16, 8, and 5 adjusted parameters, respectively. A model with a smaller AIC was considered to be significantly better than a model with a larger AIC when the AIC difference (\(\Delta\)AIC) between the models was 2 or greater.
Results
Subregion examples
This section presents examples of the analyses of the three selected subregions. The selected subregion locations are shown in Fig. 1a (lightgreen rectangles and crosses).
The first subregion is in the Bungo Channel (33.0°N, 132.1°E; subregion #1). Figure 2a shows the magnitude–time diagram of this subregion. The clear increase in the frequency of LFEs in 2010 corresponds to a period of a longterm slowslip event (SSE) in the Bungo Channel (Takagi et al. 2019; Kato and Nakagawa 2020). Model parameter estimates and AICs for this subregion are shown in Table 1 (see subregion #1). Model parameter estimates and AICs for all 43 subregions are shown in Additional files 2, 3, 4, 5: Tables S1–S4 in Additional files. In this subregion, the AICs of the Ltype, INtype, and ETAS models were 524, 499, and 585, respectively (Table 1). The AIC of the INtype model was significantly lower than that of the other models, indicating that the INtype model was superior to the other two models in this subregion.
Figure 2b shows a histogram of the interevent times for this subregion. There are maxima at approximately 10^{–3}, 10^{–1}, and 10^{1.5} days (approximately 30 days). This feature is difficult to reproduce using a model with only two characteristic timescales such as the original Ide and Nomura (2022) model. In contrast, the INtype model (blue curve in Fig. 2b) had three characteristic time constants (\({\mu }_{1}\), \({\mu }_{2}\), and \({\mu }_{3}\)), and they were estimated to be 9.5 × 10^{–4}, 2.3 × 10^{–1}, and 72 days, respectively. Note that the AIC (= 540) for the original Ide and Nomura (2022) model was significantly greater than the AIC (= 499) of the INtype model in this study, suggesting that the model with the three time constants was significantly better.
Figure 2c shows the temporal variation in the aftershock rate (LFE occurrence rate after a single LFE) for up to 10 days in this subregion, as predicted by the L, INtype, and ETAS models. For the ETAS model, the LFE aftershock rate after a single M 0.6 LFE was presented. However, because parameter \(\alpha\) of the ETAS model was zero in almost all subregions, including this subregion (Table 1 and Additional file 4: Table S3), it is safe to assume that the aftershock rate does not depend on the mainshock magnitude. In Fig. 2c, the aftershock rate of the ETAS model decays with a power law of \(p=1.1\). For regular earthquake activity, p values typically range from 0.9 to 1.5 (Utsu et al. 1995). It is intriguing that the p value estimated from LFE activity in this subregion is comparable to that of regular earthquake activity.
In contrast to the simple powerlaw decay of the ETAS model, the aftershock rates expected from the L and INtype models exhibit more complex temporal variations. In the INtype model, the aftershock rate decay stagnates between 1.0 × 10^{–2} and 2.0 × 10^{–1} days. This was because of the second lognormal distribution of the INtype model (Fig. 2b). This feature was also observed in the aftershock rate of the Ltype model (red line in Fig. 2b). As shown in Fig. 2c, the aftershock occurrence rate of the Ltype model tended to be systematically lower than that of the INtype model. This is because the Ltype model includes the effect of the summation of multiple aftershock rates (Eq. 6). This effect resulted in lower individual aftershock rates compared to the INtype model.
In this subregion, the ETAS model was the most inferior in terms of AIC. Considering that the only difference between the ETAS and Ltype models is the functional form of the aftershock rate kernel (Eqs. 6 and 16), the difference in AIC between the Ltype (AIC = 524) and ETAS models (AIC = 585) can be attributed to the difference in the functional form. This result suggests that the simple powerlaw decay assumed by the ETAS model is insufficient to describe the LFE aftershock rate.
The second example is a subregion near the Shima Peninsula (34.7N°, 136.5E°; subregion #34) (Figs. 1a and 3). In this subregion, the AICs of the Ltype, INtype, and ETAS models were 317, 282, and 358, respectively; the INtype model had the lowest AIC and outperformed the other two models (Table 1). The AIC value of the ETAS model was the highest. Figure 3b shows a histogram of the interevent times in this subregion. The three characteristic time constants of the INtype model (\({\mu }_{1}\), \({\mu }_{2}\), and \({\mu }_{3}\)) (blue curve in Fig. 2b) were estimated to be 1.6 × 10^{–3}, 1.8 × 10^{–1}, and 174 days, respectively.
Figure 3c shows the LFE aftershock rates for the Ltype, INtype, and ETAS model. The p value for the ETAS model is estimated as 3.0. This is substantially greater than the typical p value range for regular earthquake activity (0.9–1.5) (Utsu et al. 1995). Owing to this extremely high p value, the aftershock rate decayed rapidly from 1 × 10^{–1} to 10 days (from 2 to 3 × 10^{–4} events/days). However, up to 1 × 10^{–1} days, the aftershock rate was almost constant owing to the large c value (= 0.45 d). A rapid decay from 1 × 10^{–1} to 10 days was also observed for the L and INtype models (red and blue lines, respectively, in Fig. 3c).
The IN and Ltype model aftershock rates showed weaker decay between approximately 1.0 × 10^{–2} and 2.0 × 10^{–1} days. This is similar to the results for the Bungo Channel subregion (Fig. 2c). The aftershock rates predicted by the IN and Ltype models showed a maximum value (approximately 80 events/days) of approximately 3 × 10^{–4} days, in contrast to the ETAS model, which predicted an almost constant aftershock rate until approximately 1 × 10^{–1} days.
The third example is a subregion in the Aichi Prefecture (35.1°N, 137.4°E; subregion #43) (Figs. 1a and 4). In this subregion, the AICs values of the Ltype, INtype, and ETAS models were 618, 626, and 638, respectively, with the Ltype model having the lowest AIC (Table 1). This result contrasts with the results of the first and second examples in the Bungo Channel and near Shima Peninsula, where the INtype model performed best. However, the fact that the ETAS model had the highest AIC was common to all the three examples.
Figure 4b shows a histogram of the interevent times in this subregion. The three characteristic time constants of the INtype model (\({\mu }_{1}\), \({\mu }_{2}\), and \({\mu }_{3}\)) were estimated to be 7.2 × 10^{–4}, 6.2 × 10^{–1}, and 73 days, respectively. Figure 4c shows the aftershock occurrence rates predicted using the Ltype, INtype, and ETAS models. The ETAS model p value was estimated at 1.0. This value is comparable to the typical p values for regular earthquake activity (Utsu et al. 1995). The aftershock rate decay of the L and INtype models stagnated between approximately 3 × 10^{–2} and 8 × 10^{–1} days. This feature is similar to that of the subregions in the Bungo Channel and near the Shima Peninsula.
LFE aftershock rates
The results for the three subregions in Sect. “Subregion examples” suggest that the LFE aftershock rates (Figs. 2c, 3c, and 4c) cannot be described by a simple powerlaw decay and that there is decay stagnation (1.0 × 10^{–2} to 2.0 × 10^{–1} days). This is a general feature of the LFE activity in the Nankai Trough. Figure 5 shows the Ltype model median aftershock rates for all 43 subregions. This figure also shows that the decay of the aftershock rate stagnates between 1.0 × 10^{–2} and 2.0 × 10^{–1} days. The complex shape of the LFE aftershock rate probably explains why the ETAS model, which assumes a simple powerlaw decay a priori, is significantly inferior to the L and INtype models.
A similar shape for the LFE aftershock rate was suggested in LFE analyses of the San Andreas Fault by Lengliné et al. (2017) and Tan and Marsan (2020). However, for LFEs in the Mexican and Cascadia subduction zones, Lengliné et al. (2017) reported a simple power law decay of p = 1.3. The prevalence of stagnation in the LFE aftershock rate decay (Fig. 5) for LFE activity worldwide remains unclear and should be studied in the future.
The physical mechanism of the observed aftershock rate decay stagnation (Fig. 5) is not clear. However, I suggest two hypotheses. One is that triggering between LFEs might involve a characteristic time scale of 10^{–2}–2 × 10^{–1} days in its physical mechanism. However, such a characteristic time scale has not been reported for triggering between regular earthquakes, whose primal physical mechanism is considered to be stress transfer (e.g., Ellsworth and Bulut 2018). Physical models of triggering between regular earthquakes (e.g., Dieterich 1994) also do not suggest the existence of such a time scale. Therefore, the physical mechanism of eventtoevent triggering may be different for LFEs and regular earthquakes. Another possibility is that the processes behind the LFEs (e.g., slow slip in the region surrounding the LFE asperities) may be intermittent, with characteristic time scales of 10^{–2}–2 × 10^{–1} days (Kato and Nakagawa 2020). Although slow slips are usually modeled as smooth slow faultslip events (e.g., Im et al. 2020), actual slow slips may be a series of rapid acceleration/deceleration events with a characteristic recurrence period of 10^{–2}–2 × 10^{–1} days.
Spatial model parameter distribution
Figures 6, 7, 8 show the spatial parameter distributions for the Ltype, INtype, and ETAS models. It is difficult to compare parameter values of different models directly, because they depend on different model assumptions. However, common features were observed in the spatial parameter distributions of different models.
One common feature is the large change in LFE activity parameters between the east and west of Ise Bay (approximately 136.8°E). In the Ltype model (Fig. 6), both the background seismicity rate \({\mu }_{L}\) and the number of LFE aftershocks \(a\) differed between the east and west of Ise Bay. East of Ise Bay (Aichi Prefecture), the background seismicity rate was higher and the number of aftershocks was lower. In contrast, to the west of Ise Bay (near the Shima Peninsula), the background seismicity rate was low and the number of aftershocks was high. In the INtype model (Fig. 7), there was a marked contrast in \({\mu }_{3}\), \({\sigma }_{3}\), and \({\phi }_{2}\) between the east and west of Ise Bay. East of Ise Bay, the longterm clustering time constant (\({\mu }_{3}\)) was short, and the second shortterm clustering fraction (\({\phi }_{2}\)) was small. In contrast, west of Ise Bay, the longterm clustering time constant was longer, and the secondary shortterm clustering proportion was greater. Similarly, the ETAS model (Fig. 8) also shows contrasts in \({\mu }_{E}\), \(c\), and \(p\) between the east and west of Ise Bay. These changes in the LFE activity characteristics may reflect changes in the interplate slip behavior at a spatial scale of approximately 100 km around Ise Bay.
In addition, several parameters show contrasts between the Bungo Channel and western Shikoku (approximately 132.4°E): \({\mu }_{L}\) in the Ltype model; \({\mu }_{1}\), \({\mu }_{3}\), and \({\phi }_{1}\) in the INtype model; and \({\mu }_{E}\), \(c\), and \(p\) in the ETAS model. \({\mu }_{L}\), \({\mu }_{3}\), and \({\mu }_{E}\) are parameters associated with longterm recurrences of LFE bursts. These differences may be related to the occurrence of longterm SSEs in the Bungo Channel (Takagi et al. 2019; Kato and Nakagawa 2020). \({\mu }_{1}\), \({\phi }_{1}\), \(c\), and \(p\) are parameters related to LFE shortterm clustering, suggesting that there are also differences in shortterm LFE clustering features between the Bungo Channel and western Shikoku.
Model performance
Figure 9 shows the AIC difference (\(\Delta\)AIC) between the IN and Ltype models. The INtype, Ltype, and ETAS models were superior in 25, 13, and 1 of the 43 subregions, respectively. The IN and Ltype models were superior in three of the remaining four subregions, with no significant differences between them. In the remaining subregion, the ETAS and Ltype models were superior, with no significant difference between them. These results indicate that although there were more subregions in which the INtype model performed better, the IN and Ltype model performances were competitive. In particular, the Ltype model was superior to the INtype model in several subregions of Shikoku and east of Ise Bay.
Discussion
Hybrid model
In Sect. "Results", although the INtype model is superior in more subregions, the IN and Ltype model performances are competitive. Based on these results, this section examines whether the incorporation of both IN and Ltype model features results in a superior model. The INtype model has the characteristic of using a small number of parameters to represent the LFE occurrence rate (Eqs. 11, 14, and 15). The Ltype model considers the LFE activity history (Eq. 9). I propose a new model incorporating both features. Specifically, a model was developed in which the Ltype aftershock rate kernel was represented by a small number of parameters, similar to the INtype model.
In Figs. 2c, 3c, and 4c, the IN and Ltype model aftershock rates show a common feature: stagnation in the decay of the LFE aftershock rate at approximately 0.1 days. Therefore, a functional form similar to that of the INtype model may represent the shape of the Ltype model aftershock rate kernel well. Note that in the Ltype model, the shape of the aftershock rate kernel is not assumed a priori, and it is not necessarily a trivial task to find a functional form with a small number of parameters that represents the Ltype model aftershock rates well.
Based on these ideas, the following formulation was developed. First, the following equations for the aftershock rate kernel \({g}_{H}\left(t\right)\) were assumed:
These equations represent the LFE occurrence rate predicted by the INtype model, excluding the longterm LFE recurrence effect (Eqs. 11, 14, and 15). Specifically, only shortterm clustering with two time constants was considered. Using the above equations, the LFE occurrence rate can be written as
The above formulation is characterized by a small number of parameters (seven) to represent the LFE occurrence rate and consider the LFE activity history. I call this model a hybrid model. Seven model parameters were estimated (\({\mu }_{H}\), \({\mu }_{1H}\), \({\mu }_{2H}\), \({\sigma }_{1H}\), \({\sigma }_{2H}\), \({\phi }_{1H}\), and \({\phi }_{2H}\)) using the maximum likelihood method (Additional files 1, 6: Fig. S1 and: Table S5).
Figure 10 shows the aftershock rates expected from the hybrid model for the three subregions selected in Sect. “Subregion examples” (orange lines). The hybrid model aftershock rates also showed a stagnation of aftershock rate decay at approximately 0.1 days. The aftershock rates of the Ltype and hybrid models are similar, as indicated by the proximity between the red and orange lines in Fig. 10. This suggests that the functional form of the hybrid model aftershock rate kernel (Eqs. 18, 19, 20) well represents the shape of the aftershock rate kernel of the Ltype model with fewer parameters.
The AIC was used to compare the model performance. Figure 11a shows the AIC differences (ΔAICs) between the hybrid and Ltype models. In most subregions (41 of the 43 subregions), the ΔAIC was < − 2, indicating that the hybrid model was significantly superior to the Ltype model. In the remaining two subregions, the Ltype model performed significantly better than the hybrid model. This may be because in these subregions (subregions #18 and #21), the LFE interevent time histograms have a more complex shape than the other subregions and cannot be represented by a model with only two time constants for shortterm clustering, such as the INtype and hybrid models (see Additional file 1: Figs. S2 and S3).
Figure 11b shows the AIC differences between the hybrid and INtype models. The hybrid model was superior in 26 of the 43 subregions. The INtype model was superior in 15 of the remaining subregions. There were no significant differences in performance between the remaining two subregions. Compared with the results in Sect. “Results”, the number of subregions in which the INtype model performed the best decreased substantially (from 25 to 15).
This analysis demonstrates that the hybrid model is superior to the Ltype model. Furthermore, the hybrid model outperformed the INtype model in many subregions and could be considered the best model in this study. However, there were still subregions where the INtype model performed better. This indicates that considering the LFE activity history in a manner similar to the Ltype and ETAS models (Eqs. 9, 16, and 21) does not necessarily improve LFE forecasts. This point is discussed in detail in Sect. “Toward further model improvement”.
Toward further model improvement
Among the four models used in this study (Ltype, INtype, ETAS, and hybrid), the hybrid model was superior in many regions (Fig. 11). However, there are regions in which the INtype model outperforms the hybrid model. The hybrid model includes the summation of multiple LFE aftershock rates (Eq. 21). This contrasts with the INtype model, in which aftershock rates are not affected by detailed LFE activity history and are determined solely by the time elapsed since the last event (Eqs. 11, 14, and 15). Given this difference between the hybrid and INtype models, in some subregions where the INtype model outperformed the hybrid model, considering the effect of summing multiple LFE aftershock rates may worsen LFE forecasting.
To further understand the possible absence of the effect of summing multiple aftershock rates, the LFE activity was investigated in detail in the subregions where the INtype model outperformed the hybrid model. As a result, it was discovered that, in those subregions, intense LFE bursts were often followed by sudden cessations in LFE activity. In models that apply the summation of multiple aftershock rates, such sudden cessations are difficult to reproduce, because the subsequent predicted seismicity rate increases significantly after an intense burst.
Figure 12 shows an LFE burst in a subregion near the Shima Peninsula (subregion #34; Fig. 3). As shown in Fig. 12a, at approximately 3 days, eight LFEs occurred in 5 h, and the hybrid model predicted a much higher seismicity rate than the INtype model because of the summation of the LFE aftershock rates (Fig. 12b). Contrary to the hybrid model prediction, this burst stopped at approximately 3.3 days, resulting in a significant decrease in the likelihood of the hybrid model (Fig. 12c). This reduction in likelihood is the main reason the hybrid model is considered inferior to the INtype model in this subregion.
Considering the above, I suggest that it is important to incorporate the sudden LFE burst cessations for better LFE activity statistical model development. Models such as the Ltype, ETAS, and hybrid models, which simply apply the summation of multiple LFE aftershock rates, cannot reproduce these sudden cessations. New models are required that consider LFE activity history differently from these models and successfully predict LFE burst cessations.
The sudden cessations could be reproduced if I consider a different variable (\({a}_{i}\)) for each LFE and change the amplitude of the LFE aftershock rates, as in the original Lengliné et al. (2017) model (Sect. “Lengliné et al. (2017) model”). However, such a model would be inferior in terms of AIC, because the number of model parameters would increase substantially. Therefore, when varying the amplitude of the LFE aftershock rates, it is necessary to model the variation with a small number of parameters.
In addition to the additive nature of the LFE aftershock rates, there was another difference between the hybrid and INtype models. The hybrid model assumes a stationary background seismicity rate (Eq. 21), whereas the INtype model considers longterm LFE burst recurrences using a BPT distribution (Eq. 11). This difference may also cause a difference in the performance of the two models.
Transformed time sequence
The number of events \(\Lambda \left(t\right)\) predicted from a model from \({t}_{0}\) to \(t\) is obtained by integrating the event occurrence rate \(\lambda (t)\):
I use Eq. 22 to transform the occurrence time of the ith event \({t}_{i}\) into a transformed time \({\tau }_{i}\equiv\Lambda \left({t}_{i}\right)\) (Ogata 1988, 1992).
If the model can predict events well, the transformed time sequence follows a standard Poisson process, and the plot of the cumulative event count against the transformed time is expected to be linear with a slope of unity (Ogata 1988, 1992). Proximity to the straight line with a slope of unity is usually considered as an indicator of good fitting. However, in this study, the transformed time sequence did not seem to be suitable for visualizing the superiority of model performance.
The proximity to the line with a slope of unity does not necessarily indicate large likelihood or small AIC. For example, if a model repeatedly underestimates and overestimates the event occurrence rate over a short time period, then if the event occurrence rate is integrated over a long period, the predicted and observed number of events will be close and follow the straight line with a slope of unity. However, the model likelihood becomes small, and the AIC becomes large, because the event occurrence rate is underestimated or overestimated for each event. This phenomenon was often observed in this study: in Additional file 1: Fig. S4a, b, the Ltype, ETAS, and hybrid models were closer to the straight line with a slope of unity than the INtype model, but the INtype model had the smallest AIC. This is because the Ltype, ETAS, and hybrid models repeatedly underestimated the event occurrence rate when LFE clusters occurred and overestimated it when the clusters stopped. In other words, the Ltype, ETAS, and hybrid models failed to predict the occurrence of individual events. These results are consistent with the point made in Sect. “Toward further model improvement” that the Ltype, ETAS, and hybrid models do not predict LFE cluster activity well.
Selection of the minimum magnitude
Whether the LFE magnitude–frequency distribution (especially for large events in a catalog) follows the Gutenberg–Richter law (a power law distribution) or an exponential distribution is still under debate and inconclusive (Bostock et al. 2015; Chestler and Creager 2017). Hence, no method currently exists for evaluating the completeness of an LFE catalog.
In seismicity analysis, the use of incomplete catalogs makes it unclear whether the results obtained are true seismicity features or artifacts due to the incompleteness of the catalogs. In light of this, in this study, I followed Bostock et al. (2015), assumed the Gutenberg–Richter law, and used as large a minimum magnitude as possible (M 0.6). This is as conservative an analysis as possible. My intention is to eliminate artifacts as much as possible and to discuss the true features of LFE activity, even if my analysis is limited to the activity of large LFEs.
However, the conservative approach of this study has the problem of ignoring the activity of many LFEs smaller than the minimum magnitude. Therefore, I conducted a reanalysis by changing the LFE minimum magnitude to M 0.4 and M 0.2. The number of events used in the analysis increased to 18,103 (5 times greater than the number of M 0.6 or larger LFEs) and 67,193 (19 times greater than the number of M 0.6 or larger LFEs), respectively. Because the number of LFEs increased substantially as the minimum magnitude was lowered, the computational cost of this study also increased substantially (on the order of the square of the number of LFEs). Therefore, the analysis could not be performed for the minimum magnitude below M 0.2.
As a result, I found two common features among the results for the minimum magnitudes of M 0.6, M 0.4, and M 0.2. The first is that the performance of the INtype model and the Ltype model are competitive, although the INtype model performs better in more subregions (Figs. 9, 13a, b). The second is that the Hybrid model is generally superior to the Ltype model (Figs. 11a, 13c, d).
In contrast, the comparisons between the hybrid and INtype models (Figs. 11b, 13e, f) showed an increase in subregions where the INtype model outperformed the hybrid model for the minimum magnitude of M 0.2. For the minimum magnitude of M 0.2, the performance of the hybrid and INtype models is comparable: the hybrid model was superior in 23 subregions, and the INtype model was superior in 20 subregions. This may be because the effect of the summation of multiple aftershock rates assumed by the hybrid model does not describe the clustering of LFEs well, as suggested in Sect. “Toward further model improvement”. When the minimum magnitude is small, the likelihood and AIC strongly reflect the ability to predict shortterm LFE clustering, because numerous small LFEs are concentrated over short time periods.
The selection of the optimal minimum magnitude for statistical analysis of LFE activity is an important future issue. It is desirable to use as many LFEs as possible in the analysis while ensuring the completeness of the LFE catalog. In future studies, it is important to develop a method to evaluate the completeness of LFE catalogs and to select an appropriate minimum magnitude.
Shortterm aftershock incompleteness
Overlooking small aftershocks due to intense aftershock activity after large mainshocks is called shortterm aftershock incompleteness (STAI) (e.g., Petrillo and Zhuang 2023). For regular earthquake activity, we can easily visually identify STAI in the magnitude–time diagram. In contrast, for LFE activity, STAI is not visually clear in the magnitude–time diagram, because the distinction between mainshocks and aftershocks is unclear for LFE activity.
However, smaller LFEs may be overlooked during intense LFE bursts. To examine this possibility, for M 0.4 and M 0.8 LFEs, I checked the elapsed time from the immediately preceding event to the occurrence of each LFE. In Additional file 1: Fig. S5, I showed the histograms of the elapsed time. Counts were summed and normalized for all 43 subregions. Note that the minimum magnitude of M 0.4 was used in this analysis.
The results showed that M 0.4 LFEs had a smaller percentage of events occurring in less than 10^{−3} days (approximately 90 s or less) compared to M 0.8 LFEs. This is consistent with the hypothesis that during intense LFE bursts (i.e., periods characterized by very short interevent times), small LFEs may be overlooked.
Much is still not known about the temporal variations of the LFE catalog completeness and LFE magnitude–frequency distribution. Further research on STAI for LFE activity is needed.
Factors not considered by the models used in this study
Some important LFE activity features are not considered in the models used in this study (Ltype, INtype, ETAS, and hybrid models). For example, the LFE activity exhibits a high tidal response (e.g., Royer et al. 2015). It is also known that LFE epicenters migrate diffusively on a timescale of approximately 10 days and a spatial scale of approximately 100 km; they also migrate diffusively or at a constant speed on a timescale of approximately 30 min and a spatial scale of 15 km (Kato and Nakagawa 2020). Statistical LFE activity models could be improved by considering these factors. However, as stated in Sect. “Introduction”, this study aims to compare existing statistical LFE activity models, identify which model better describes LFE activities, and create a new model based on this comparison. It is beyond the scope of this study to incorporate new factors that are not considered in existing models into each model. This is an important issue that should be addressed in future studies.
Conclusions
This study applied the existing statistical LFE activity models (L and INtype models) (Lengliné et al. 2017; Ide and Nomura 2022) to LFEs along the Nankai Trough (Kato and Nakagawa 2020) and conducted the first comparison based on AIC (Sect. “Results”). The aftershock rate decay predicted by the L and INtype models stagnated near an elapsed time of 0.1 days. This is the primary reason why the ETAS model (Ogata 1988), which a priori assumes a simple powerlaw decay, does not perform well for LFE activity.
Although there were more subregions in which the INtype model performed better, the IN and Ltype model performances were competitive. Based on these results, a new hybrid model was created that incorporates the features of both models (Sect. “Discussion”). The hybrid model considers the LFE activity history in a manner similar to the Ltype and ETAS models, and represents the LFE aftershock rate with a small number of parameters, similar to the INtype model. The results showed that the hybrid model outperformed the L and INtype models in many subregions, and the hybrid model was considered to be the best currently available model.
However, I found that the hybrid model could not reproduce the sudden LFE burst cessations, because the model summed multiple LFE aftershock rates. In subregions where such a cessation often occurs, the INtype model, which does not consider the LFE activity history, outperformed the hybrid model. For better LFE activity forecasting, new models that consider LFE activity history differently from the hybrid model and successfully predict the abrupt LFE burst cessations are required.
Moreover, the empirical equations formulated in this study for LFE aftershock rates (Eqs. 18, 19, 20) may be useful for future investigations concerning statistical and physical modeling of LFE activity. Specifically, they could be an important observational fact that future physical LFE activity models should reproduce.
Availability of data and materials
The LFE catalog is from Kato and Nakagawa (2020) and is included in its supplementary material. The seafloor topography data (Smith and Sandwell 1997) are available at https://topex.ucsd.edu/marine_topo/. The figures were prepared used the Generic Mapping Tools software package (Wessel et al. 2019), which are available at https://www.genericmappingtools.org.
Abbreviations
 AIC:

Akaike’s information criterion
 BPT:

Brownian passage time
 ETAS:

Epidemictype aftershocksequence
 GFT:

Goodnessoffit test
 JMA:

Japan Meteorological Agency
 LFE:

Lowfrequency earthquake
 SSE:

Slowslip event
 ΔAIC:

AIC difference
References
Akaike H (1974) A new look at the statistical model identification. IEEE Trans Autom Control 19(6):716–723. https://doi.org/10.1109/TAC.1974.1100705
Beroza GC, Ide S (2011) Slow earthquakes and nonvolcanic tremor. Annu Rev Earth Planet Sci 39:271–296. https://doi.org/10.1146/annurevearth040809152531
Bostock MG, Thomas AM, Savard G, Chuang L, Rubin AM (2015) Magnitudes and momentduration scaling of lowfrequency earthquakes beneath southern Vancouver Island. J Geophys Res 120(9):6329–6350. https://doi.org/10.1002/2015JB012195
Chestler SR, Creager KC (2017) Evidence for a scalelimited lowfrequency earthquake source process. J Geophys Res 122(4):3099–3114. https://doi.org/10.1002/2016JB013717
Dieterich J (1994) A constitutive law for rate of earthquake production and its application to earthquake clustering. J Geophys Res 99(B2):2601–2618. https://doi.org/10.1029/93JB02581
Dragert H, Wang K, James TS (2001) A silent slip event on the deeper Cascadia subduction interface. Science 292(5521):1525–1528. https://doi.org/10.1126/science.106015
Ellsworth WL, Bulut F (2018) Nucleation of the 1999 Izmit earthquake by a triggered cascade of foreshocks. Nat Geosci 11:531–535. https://doi.org/10.1038/s4156101801451
Gutenberg B, Richter CF (1944) Frequency of earthquakes in California. Bull Seismol Soc Am 34(4):185–188. https://doi.org/10.1785/BSSA0340040185
Ide S, Beroza GC (2023) Slow earthquake scaling reconsidered as a boundary between distinct modes of rupture propagation. Proc Natl Acad Sci 120(32):e2222102120. https://doi.org/10.1073/pnas.2222102120
Ide S, Nomura S (2022) Forecasting tectonic tremor activity using a renewal process model. Prog Earth Planet Sci 9:67. https://doi.org/10.1186/s40645022005231
Ide S, Beroza GC, Shelly DR, Uchide T (2007) A scaling law for slow earthquakes. Nature 447(7140):76–79. https://doi.org/10.1038/nature05780
Im K, Saffer D, Marone C, Avouac JP (2020) Slipratedependent friction as a universal mechanism for slow slip events. Nat Geosci 13(10):705–710. https://doi.org/10.1038/s4156102006279
Kato A, Nakagawa S (2020) Detection of deep lowfrequency earthquakes in the Nankai subduction zone over 11 years using a matched filter technique. Earth Planet Space 72:128. https://doi.org/10.1186/s40623020012574
Kumazawa T, Ogata Y (2013) Quantitative description of induced seismic activity before and after the 2011 TohokuOki earthquake by nonstationary ETAS models. J Geophys Res 118(12):6165–6182. https://doi.org/10.1002/2013JB010259
Kuna VM, Nábělek JL, Braunmiller J (2019) Mode of slip and crust–mantle interaction at oceanic transform faults. Nat Geosci 12(2):138–142. https://doi.org/10.1038/s4156101802871
Lengliné O, Frank WB, Marsan D, Ampuero JP (2017) Imbricated slip rate processes during slow slip transients imaged by lowfrequency earthquakes. Earth Planet Sci Lett 476:122–131. https://doi.org/10.1016/j.epsl.2017.07.032
Luo Y, Liu Z (2019) Slowslip recurrent pattern changes: perturbation responding and possible scenarios of precursor toward a megathrust earthquake. Geochem Geophys Geosyst 20(2):852–871. https://doi.org/10.1029/2018GC008021
Matsuzawa T, Hirose H, Shibazaki B, Obara K (2010) Modeling short and longterm slow slip events in the seismic cycles of large subduction earthquakes. J Geophys Res 115:B12301. https://doi.org/10.1029/2010JB007566
McGuire JJ, Boettcher MS, Jordan TH (2005) Foreshock sequences and shortterm earthquake predictability on East Pacific Rise transform faults. Nature 434(7032):457–461. https://doi.org/10.1038/nature03377
Nadeau RM, Dolenc D (2005) Nonvolcanic tremors deep beneath the San Andreas Fault. Science 307(5708):389–389. https://doi.org/10.1126/science.110714
Nishikawa T, Nishimura T (2023) Development of an epidemictype aftershocksequence model explicitly incorporating the seismicitytriggering effects of slow slip events. J Geophys Res 128(5):e2023JB026457. https://doi.org/10.1029/2023JB026457
Nishikawa T, Matsuzawa T, Ohta K, Uchida N, Nishimura T, Ide S (2019) The slow earthquake spectrum in the Japan Trench illuminated by the Snet seafloor observatories. Science 365(6455):808–813. https://doi.org/10.1126/science.aax5618
Nishikawa T, Ide S, Nishimura T (2023) A review on slow earthquakes in the Japan Trench. Prog Earth Planet Sci 10:1. https://doi.org/10.1186/s4064502200528w
Obara K (2002) Nonvolcanic deep tremor associated with subduction in southwest Japan. Science 296(5573):1679–1681. https://doi.org/10.1126/science.107037
Obara K, Kato A (2016) Connecting slow earthquakes to huge earthquakes. Science 353(6296):253–257. https://doi.org/10.1126/science.aaf1512
Ogata Y (1988) Statistical models for earthquake occurrences and residual analysis for point processes. J Am Stat Assoc 83(401):9–27. https://doi.org/10.1080/01621459.1988.10478560
Ogata Y (1992) Detection of precursory relative quiescence before great earthquakes through a statistical model. J Geophys Res 97(B13):19845–19871. https://doi.org/10.1029/92JB00708
Petrillo G, Zhuang J (2023) Verifying the magnitude dependence in earthquake occurrence. Phys Rev Lett 131(15):154101. https://doi.org/10.1103/PhysRevLett.131.154101
Powell MJ (1964) An efficient method for finding the minimum of a function of several variables without calculating derivatives. Comput J 7(2):155–162. https://doi.org/10.1093/comjnl/7.2.155
Royer AA, Thomas AM, Bostock MG (2015) Tidal modulation and triggering of lowfrequency earthquakes in northern Cascadia. J Geophys Res 120(1):384–405. https://doi.org/10.1002/2014JB011430
Schorlemmer D, Werner MJ, Marzocchi W, Jordan TH, Ogata Y, Jackson DD et al (2018) The collaboratory for the study of earthquake predictability: achievements and priorities. Seismol Res Lett 89(4):1305–1313. https://doi.org/10.1785/0220180053
Shelly DR, Beroza GC, Ide S (2007) Nonvolcanic tremor and lowfrequency earthquake swarms. Nature 446(7133):305–307. https://doi.org/10.1038/nature05780
Smith WH, Sandwell DT (1997) Global sea floor topography from satellite altimetry and ship depth soundings. Science 277(5334):1956–1962. https://doi.org/10.1126/science.277.5334.1956
Takagi R, Uchida N, Obara K (2019) Alongstrike variation and migration of longterm slow slip events in the western Nankai subduction zone, Japan. J Geophys Res 124(4):3853–3880. https://doi.org/10.1029/2018JB016738
Takemura S, Hamada Y, Okuda H, Okada Y, Okubo K, Akuhara T et al (2023) A review of shallow slow earthquakes along the Nankai Trough. Earth Planet Space 75:164. https://doi.org/10.1186/s40623023019206
Tan YJ, Marsan D (2020) Connecting a broad spectrum of transient slip on the San Andreas fault. Sci Adv 6(33):abb2489. https://doi.org/10.1126/sciadv.abb2489
Utsu T (1957) Magnitudes of earthquakes and occurrence of their aftershocks. Zisin 10(2):35–45
Utsu T, Ogata Y, Matsu’ura R (1995) The centenary of the Omori formula for a decay law of aftershock activity. J Phys Earth 43(1):1–33. https://doi.org/10.4294/jpe1952.43.1
Wallace LM (2020) Slow slip events in New Zealand. Annu Rev Earth Planet Sci 48:175–203. https://doi.org/10.1146/annurevearth071719055104
Wessel P, Luis JF, Uieda L, Scharroo R, Wobbe F, Smith WHF, Tian D (2019) The generic mapping tools version 6. Geochem Geophys Geosyst 20:5556–5564. https://doi.org/10.1029/2019GC008515
Wiemer S, Wyss M (2000) Minimum magnitude of completeness in earthquake catalogs: examples from Alaska, the western United States, and Japan. Bull Seismol Soc Am 90(4):859–869. https://doi.org/10.1785/0119990114
Woessner J, Wiemer S (2005) Assessing the quality of earthquake catalogues: estimating the magnitude of completeness and its uncertainty. Bull Seismol Soc Am 95(2):684–698. https://doi.org/10.1785/0120040007
Zhuang J, Ogata Y, VereJones D (2002) Stochastic declustering of spacetime earthquake occurrences. J Am Stat Assoc 97(458):369–380. https://doi.org/10.1198/016214502760046925
Acknowledgements
I would like to thank Editage (www.editage.com) for English language editing.
Funding
This study was supported by JSPS KAKENHI Grant Number 22H05307.
Author information
Authors and Affiliations
Contributions
TN conducted all analyses, prepared all figures, and wrote the manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The author declares that he has no competing interests.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Additional file 1.
Supplementary figures.
Additional file 2: Table S1.
Ltype model parameter estimates.
Additional file 3: Table S2.
INtype model parameter estimates.
Additional file 4: Table S3.
ETAS model parameter estimates.
Additional file 5: Table S4.
AICs for all models.
Additional file 6: Table S5.
Hybrid model parameter estimates.
Rights and permissions
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/.
About this article
Cite this article
Nishikawa, T. Comparison of statistical lowfrequency earthquake activity models. Earth Planets Space 76, 59 (2024). https://doi.org/10.1186/s40623024020076
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s40623024020076