Development of High-Performance Seismic Phase Picker Using Deep Learning in Hakone Volcanic Area

doi:10.21203/rs.3.rs-2253946/v1

Download PDF

Full paper

Development of High-Performance Seismic Phase Picker Using Deep Learning in Hakone Volcanic Area

https://doi.org/10.21203/rs.3.rs-2253946/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 16 May, 2023

Read the published version in Earth, Planets and Space →

You are reading this latest preprint version

In volcanic regions, active earthquake swarms often occur associated with volcanic activity, and their rapid detection and measurement are crucial for volcano disaster prevention. Currently, however, these processes are ultimately left to human judgment and require much time and money, making detailed verification in real time impossible. To overcome this issue, we attempted to apply machine learning, which has been studied in many seismological fields in recent years. Several models have already been trained using a large amount of training data (mainly crustal earthquakes). Although there are some cases where these models can be applied without any problems, regional dependence on the learned models has also been reported. Since this study targets earthquakes in a volcanic region, existing learned models may be difficult to apply. Therefore, in this study, we created the above publicly available trained model (model0), a model trained with approximately 220,000 seismic waveform data recorded at Hakone volcano from 1999 to 2020 with initialized weights (model1) using the same architecture, and a model fine-tuned with the aforementioned Hakone data using the weight of model0 as initial values (model2), and evaluated their performance. As a result, the detection rates of model1 and 2 were much higher than model0. However, small amplitudes are often missed when multiple seismic waves are in a time window to determine the phase arrival. Therefore, we created training data with two waveforms in the one-time window, retrained the model using the data, and successfully detected waveforms that would have been missed previously. In addition, it was found that more events were detected by setting the threshold to a low value for detection, increasing the number of detections, and filtering by phase association and hypocenter location.

Deep learning

Machine learning

Phase pick

Volcanic earthquake

In volcanic regions, active earthquake swarms are often associated with volcanic activities, and their rapid detection and measurement are vital for disaster prevention. Understanding the seismic activities and its mechanism is also essential for mitigating future disasters. These tasks begin with detecting and identifying seismic phases, but those tasks mainly rely on human labor and are time-consuming and labor-intensive. The general method for detecting and picking seismic phase is applying the AR-AIC method (e.g., Akazawa et al., 2007) to waveforms triggered by STA/LTA (Allen, 1976) to estimate the arrival time of P and S waves. However, in many cases, the final decision is left to human labor. Various machine learning methods, especially deep neural networks, have been developed to detect and pick seismic phase arrival to solve these problems. Early studies used multilayer perceptron networks to detect seismic waveforms (e.g., Dysart and Pulli, 1990; Musil and Plešinger, 1996; Fedorenko et al., 1999; Ursino et al., 2001; Kong et al., 2016), phase-picking techniques were developed (e.g., Dai and MacBeth, 1995; Tiira, 1999; Zhao and Takano, 1999; Wiszniowski et al., 2014). Subsequently, seismic waveform detection and phase-picking techniques were developed using CNNs with convolutional and pooling layers added to deep learning (e.g., Perol et al., 2018; Ross et al., 2018). Zhu and Beroza (2018) used the training data of millions of seismograms picked manually in northern California and built a model to detect P- and S-waves using the U-net architecture, a fully convolutional network. An EQ transformer was also developed by introducing the attention mechanism used in the Transformer(Vaswani et al., 2017) into a conventional FCN to simultaneously perform the detection of earthquake signals and phase picking (Mousavi et al., 2021). Most of the codes and trained models described above are publicly available on GitHub (https://github.com/) and other websites and can be used by anyone. However, it is also known that these models have regional characteristics where the accuracy varies depending on the region where they are applied (e.g., Münchmeyer et al., 2022).

Furthermore, most of the above-trained models are trained on crustal earthquakes as training data, which may be problematic to use as is for detecting earthquakes occurring at volcanoes. Earthquakes in volcanic regions are characterized by earthquake swarm and volcanic earthquakes with relatively low-frequency predominance, and it is possibly necessary to build models suitable for such an environment. The problem, however, is that few volcanoes have accumulated the vast amount of data required to perform such learning, and Lapins et al. (2021), for example, used transfer learning from the published model of Ross et al. (2018) to improve the performance of earthquake motion detection in the area with a small amount of data.

On the other hand, Hakone, the target of this study, has about 28 years of accumulated observation data and sufficient seismic waveform data to be trained from scratch. Therefore, in this study, we used the trained model and architecture of Zhu and Beroza (2018), a publicly available model that has shown high performance in performance evaluation tests in previous studies (e.g., Münchmeyer et al., 2022). We created models using the PhaseNet trained model (model0) and its architecture. We evaluated their performance under various conditions using approximately 220,000 seismic waveform data from about 30,000 events recorded at Hakone Volcano from 1999 to 2020.

Hakone volcano is located in central Japan's northern boundary zone of the Izu-Mariana volcanic arc. It has a caldera topography surrounded by an outer ring of about 10 km diameter (Fig. 1). The most recent large-scale eruption with lava ejecta was about 3,000 years ago. Although no such large-scale eruption has occurred since then, some studies indicated that phreatomagmatic eruption happened between the 12th and 13th centuries (e.g., Kobayashi et al., 2006). In addition, a very small eruption occurred in 2015 at Owakudani for the first time in the Hakone volcano's recorded history (e.g., Mannen et al., 2018). Hakone volcano is still active today, with earthquakes occurring very shallow in the crust of the caldera and occasional large swarms of earthquakes. In this context, the hot spring research institute of Kanagawa (HSRI) has been observing and studying the causes of these seismic events for many years (Fig. 1). Typical seismic activity since 1995 includes active earthquake swarms in 2001, 2006, 2008, and 2009 (e.g., Yukutake et al., 2011a), and active seismic swarms associated with phreatic eruptions were observed in 2015 (e.g., Mannen et al., 2018 ). The mechanism of the seismic swarms at the Hakone volcano has also been extensively studied. The prevailing theory is that they are caused by the migration process of high-temperature, high-pressure hydrothermal water supplied from deep underground through fractures such as micro-faults (Yukutake et al., 2011a). Seismic activity also increased after the 2011 M9.0 Tohoku-oki earthquake, although with different characteristics from the above earthquake swarm (Yukutake et al., 2011b). This study uses 217553 seismic waveforms recorded by the above observation network as input data.

In this study, we use the PhaseNet architecture built by Zhu and Beroza (2018). The method is based on the structure of U-net (Ronneberger et al. 2015), a type of encoder-decoder FCN developed for biomedical image processing, modified to handle 1D time series data. In the decoder part, convolution and upsampling are repeatedly applied to the input data to improve the resolution of the feature map and obtain the desired region extraction results. PhaseNet takes a 3-component seismic waveform as input and outputs the probability of P- and S-wave onsets and noise per sample. See Zhu and Beroza (2018) for more details on the model. While CNN classifies given unknown data, FCN aims to segment what is where, so it does not matter how many waveforms are in the target unknown data. The length of the input data must be the same as the training data for CNN, but it is variable for FCN.

In this study, we used 217553 waveforms from 30565 events recorded from 1995 to 2021 in the Hakone volcanic area mentioned in the previous section. Experts have manually inspected these data for P- and S-waves, and we'll treat the information as the ground truth. In many cases, the picked onsets are stored with the P wave at the same position, but if they are used as is, the model may learn that position as that of the P wave. To avoid the case, in PhaseNet, the P- and S-waves are always included, and the waveforms are randomly shifted back and forth in position. In this study, as in the original, each seismic waveform was cut out in a 90-second time window, and after shifting them in the above manner, they were cut to 30 seconds and used as input data. Other than that, the original 200 Hz sampling was downsampled to 100 Hz as preprocessing.

In this study, three models were first created, and their performance was evaluated as follows:

1) Model0: a trained model built using training data from millions of crustal earthquakes in Northern California (Zhu and Beroza, 2018, publicly available on GitHub).

2) Model1: the model was trained from scratch using the PhaseNet architecture, employing approximately 220,000 seismic waveform data from about 30,000 events recorded at Hakone volcano from 1999 to 2020.

3) Model2: the fine-tuned model with the Hakone volcano training data used in model 1, using the weights of model 0 as initial values.

In the training of model1 and model2, 220424 seismic waveforms were randomly assigned, 75% as training data and 25% as data for validation. To verify the best model for each model, the performance of each model was verified by varying the learning rate and batch size and repeating the computation up to epoch 100 (Fig. 2). The results showed that the best model for model 1 was the one with a batch size of 64 and a learning rate of 0.01, while the best model for model 2 was the one with a batch size of 16 and a learning rate of 0.1. The F1 values were slightly higher for model 1 but almost the same for model 2: 0.870 and 0.867 for the P wave and 0.782 and 0.776 for the S wave, respectively. From now on, the best models for both models will be referred to as model1 and model2, respectively. Using the same data for validation, the F1 values for model 0 were calculated to be 0.832 for the P wave and 0.707 for the S wave, showing a significant improvement mainly in the S wave picking. A heatmap showing F1 scores for each batch size and the learning rate is shown in Fig. 3.

These models were applied to 5559 waveforms (441 events), waveform data from the 5/18–20/2019 swarm earthquakes, which were not used for training or validation. The time window to pick phase onset is 90 seconds, and the detection threshold is 0.6. model0, model1, and model2 show P-wave values of 0.823, 0.860, and 0.857, and S-wave values of 0.641, 0.755, and 0.749, respectively, and as in the validation data, model1 and As was the case with the validation data, model1 showed slightly better values for model1 and model2, indicating that the model trained using the Hakone training data performed better than model0. There was almost no difference between the three models in travel time compared to the human-operated data (Fig. 4). Figure 5 shows an example of phase picking. The results indicate that each model can detect waveforms with high accuracy when there is only one waveform or when there are multiple waveforms with almost the same amplitude in the time window (Fig. 5a), b)). In addition, each model can recognize seismic waveforms with low signal-to-noise ratio or missing components (Fig. 5d)).

On the other hand, when there are multiple waveforms with significantly different amplitudes in the time window, model 0 detects the waveform with the larger amplitude with high accuracy. Still, models 1 and 2 respond to the smaller amplitude with lower probability values, but the accuracy of detecting the larger amplitude is often lower ( Fig. 5c), e), f)). The reason for the overall lower performance of the test data compared to the validation data may be that this is a period of very active earthquake swarms, with many earthquakes occurring in a short period.

Models 0, 1, and 2 were then applied to continuous waveforms for the same period as the test data. The number of manually detected earthquakes during this period was 441. After detecting P and S waves, we used the Rapid Earthquake Association and Location (REAL) (Zhang et al., 2019) for phase association and preliminary location. The parameters used in REAL are 10km for the search range in horizontal centered at the station recorded the initiating phase and a depth of 20 km. The search grid was set to be approximately 2 km horizontally and in depth. The time windows were tested for 1 hour, 30 minutes, 10 minutes, 1 minute, 30 seconds, 10 seconds, and 5 seconds. The best results were obtained at 30 seconds, the same data length used in training, suggesting that shortening the time window may not improve accuracy (Table 1). In addition, when applying the method to continuous waveform data, as was the case with the test above data, there were many cases where small amplitudes were missed when multiple waveforms of widely different amplitudes were included in the time window. Zhu et al. (2020) showed that by having multiple seismic waveforms in the training data, the detection performance improves when many seismic waveforms are in a short period. Thus, it may be necessary to insert multiple seismic waveforms into the training data to improve the detection performance of active swarm earthquakes, where many earthquakes occur in a short period at a volcano. So, in this study, we verified the performance when data containing two seismic waves were added to the time window as training data. Two waveforms were randomly extracted from the validation data and combined so that the P-wave of the following waveform comes 6 to 25 seconds after the first arrives, and 100,000 training data containing two waveforms in a 90 second time window were created. The model re-trained from scratch with only those data was designated model3, and the model trained with the same data as model3 with model1 as the starting model was denoted model4. As a result, both model3 and model4 significantly improved the problems observed in model1 and 2, but model3 was lower than model1 in the total number of detections. The number of event detections in model 4 also exceeded that of model 1, suggesting the importance of selecting training data according to the seismic activity to be applied (Zu et al., 2020).

Table 1

Time window and number of earthquakes detected
Time window	# of event detection
10s	920
20s	1088
30s	1244
1min	1047
10min	787
30min	670
1hr	610

Table 2

Numbers of seismic waves were detected for each model with a detection threshold of 0.6.
Model	# of event detection
Model0	1028
Model1	1244
Model2	1131
Model3	1120
Model4	1311

In this study, we first used a trained PhaseNet model (model0) built by Zhu and Beroza (2018) and its architecture, using approximately 220,000 seismic waveform data from about 30,000 events recorded at Hakone volcano from 1999 to 2020. We created a model trained from scratch (model1), and the model was fine-tuned with the aforementioned Hakone data using model0 as the initial value (model2) and evaluated their performance. The performance of model1 and model2 was significantly improved compared with that of model0, especially in S-wave detection, while the performance of model1 was slightly better than that of model2 but almost the same. This suggests that the number of Hakone data used in this study is sufficient for the PhaseNet architecture to learn the seismic features of this region. Model 1 could find the maximum F1 value when the batch size and learning rate were varied, but model 2 could not find the maximum value in the same test, and the smaller the batch size and the larger the learning rate, the better the performance. If the learning rate is too small, there is a risk that the model will be stuck in a local solution and not perform well, but if it is too large, there is possible that it will jump over the minimum value. Also, suppose the batch size is too large. In that case, there is a risk that the data features will be averaged out and the individuality of the data will be lost. On the other hand, there is an advantage of faster learning without being trapped by local solutions, but in general, it has been shown that performance decreases as the batch size increases (e.g., Keskar et al., 2017). However, Smith et al. (2018) argue that there is an optimal batch size for a given learning rate and that it is not important to reduce the batch size but to choose a learning rate appropriate for that batch size. However, there is no way to determine the optimal batch size or learning rate without benchmark testing. Although detailed hyperparameter testing is computationally demanding, the optimizer called adam used in PhaseNet is known to require only low-level hyperparameter tuning (Kingma and Ba, 2014). Since the values chosen for model 1 are not very different from those used in the original model, we believe the test results are highly reliable. In model2, the smaller batch size showed better performance, which means that each data feature is sensitively captured and learned. In other words, it can be considered that this is the result of starting from model 0 and learning to fit the Hakone data better. However, model 2 is unstable because the loss function does not converge when the batch size is larger than 64, so model 1 is more suitable for Hakone data. Although model1 had the best performance among the three models, when multiple waveforms with different amplitudes were in the time window, the probability values for small amplitudes were lower than the threshold for phase pick, and the prediction accuracy for large amplitudes was often lower. On the other hand, Model 0 did not respond to any small amplitude in such cases and tended to detect only large amplitudes. The training data contains only one seismic waveform per data set, normalized to the maximum amplitude. Since the maximum amplitude also normalizes the data for phase pick, small amplitudes are considered noise and unlikely to be detected. The probability value increased a little in models 1 and 2 because the Hakone seismic waveform has different characteristics from those of the California crustal earthquakes used in the model 0 study, and may have captured those characteristics, even if the amplitude is small. In this study, to overcome the problem observed in model 1, two waveforms were randomly taken from the existing validation data. By creating such data, waveforms of small amplitude in the data normalized by the maximum amplitude can be labeled and trained as P and S waves. We evaluated the performance of model3, which was trained with all weights initialized, and model4, which was trained with the weights of model1 as initial values, and found that both models improved the ability to detect waveforms with smaller amplitude when there were multiple waveforms with different amplitudes in the same time window. Model4 also outperformed model1 in the number of events detected. Model 3, on the other hand, was not as good as model 1 in terms of the number of events detected. This may be because the number of seismic waves is 100,000, one-half of the number of data used in the training of model 1. Deep learning models for seismic phase pick usually use amplitude-normalized data as training data. The amplitude value varies with the distance from the epicenter and the magnitude of the earthquake in each seismic data set. So, when seismic waveforms acquired over a wide area are used as training data, the amplitude values will be in several orders of magnitude range. The model may not converge if trained without normalization. However, in the case of a group of earthquakes that occurred within a narrow range of magnitude and did not vary much in magnitude range, the inclusion of fluctuations in amplitude values within that range in the training data is thought to have improved accuracy.

Model 4 detects events in the continuous seismic waveforms more accurately than model 1, but some waveforms are still missed. Although these may be improved by increasing the number of data, a detailed look at the results shows that many waveforms tended to show even a slight increase in probability value. Therefore, we predicted the same continuous waveform data by setting the detection threshold of model 4 to 0.1 and 0.3 and applying REAL to the results. We further applied VELEST (Kissling et al., 1995) for the earthquake location and finally relocated events using HypoDD (Waldhauser and Ellsworth, 2000). The events were relocated 2094, 1296, and 1091 earthquakes at threshold values of 0.1, 0.3, and 0.6, respectively (Fig. 7, Table 3). From the above, it is not necessary to set the threshold high at the time of waveform detection, but instead, it is more effective to set it low and filter out noise in the subsequent phase association, hypocenter location, and relocation.

Table 3

Number of event detections for each algorithm relative to the detection threshold
Threshold	REAL	VELEST	hypoDD
0.1	7381	6909	2094
0.3	1937	1865	1296
0.6	1311	1302	1091

Regarding the types of earthquakes newly detected by the PhaseNet model, a few are out of the cluster as epicenter locations, but most are within the cluster. Although the magnitudes in the original catalogs are calculated in different ways and cannot be compared, a comparison of the histograms of magnitude and frequency for the earthquakes detected at different threshold values shows that the smaller the threshold value, the more small earthquakes are picked up (Fig. 8). However, as the threshold value decreases, the number of earthquakes out of the cluster also increases. Hence, it is necessary to examine whether these newly detected small earthquakes are real or not in the future. Note that the Matched filter method (MF) (Gibbons and Ringdal, 2006) was used to detect earthquakes in this earthquake swarm (Yukutake et al., 2022a), and the number of earthquakes detected was 2600, which is more than the number detected by model4 with threshold 0.1. A comparison of events detected by model4 and MF shows that many of the earthquakes detected by MF are shallower than those in the original catalog, while the events detected by model4 show a distribution similar to the depth of events in the catalog (Fig. 9). However, since the hypocenter relocation in this study is not based on waveform correlation, depth may change with the operation.

Furthermore, station correction is performed in MF but not in this relocation. As for the processing time, it is possible to improve the performance of both methods by devising the code. The advantage of PhaseNet over MF is that it does not require template earthquakes, which are necessary for MF. It can be generalized to some extent, even if trained on training data from other regions. For example, when model 1 was applied to seismic data from other volcanoes in Japan, it succeeded in detecting eight times more earthquakes than in the catalog (Yukutake and Kim, 2022b). Since not all volcanoes have been monitored with high accuracy for many years like Hakone volcano, it is worth aiming to improve the accuracy of such machine learning models in the future, since it may be possible to create a model specialized to a particular region by tuning a small amount of training data based on the model developed at Hakone volcano. In addition, many of the seismic waveforms may overlap during the very active swam period. By using the know-how of training data with multiple seismic waveforms used in this study, we can artificially create overlapping data and add them to the training data, thereby learning their characteristics and potentially contributing to the event detection capability.

To automate and improve the accuracy of seismic wave detection at Hakone volcano, a seismic wave detection model was created using the PhaseNet architecture with training data of approximately 30,000 earthquakes (~ 220,000 waveform data) that occurred in the area. The newly constructed model outperformed the performance of model 0 (pretrained open-to-public model) when there was only one waveform per training data. Still, when multiple waveforms of different amplitudes existed in the same time window, there were many cases where waveforms with small amplitudes were missed. Fine-tuning with 100,000 data, including two waveforms, using the weights of model 1 as a starting point, significantly improved the above problems and increased the number of seismic detections. Nevertheless, the seismic waveforms that were still missing showed small amplitudes of probability values while not reaching the threshold, so we lowered the threshold and passed the data to REAL, VELEST, and HypoDD, which showed that the number of earthquake detections was higher than in the case where the same process was performed with a higher threshold. These results suggest that, although further investigation is needed, it may be possible to detect a large number of earthquakes if we do not set a strict threshold at the time of seismic wave detection and instead use phase association, hypocenter location, and other methods to filter out the false seismic wave detections. In the future, we will apply this model to earthquakes in volcanic regions where training data is insufficient and verify its generalization performance. We will also consider creating models specific to individual volcanoes using transfer learning.

CNN: Convolutional neural network, FCN: Fully convolutional network, MF: Matched filter method

Ethics approval and consent to participate

Not applicable

Consent for publication

Not applicable

Availability of data and materials

Seismic waveform data from Hi-net (NIED) is available at https://www.hinet.bosai.go.jp/?LANG=en.

Competing interests

The authors declare no conflicts of interest associated with this manuscript.

Funding

This study was supported by JSPS KAKENHI Grant Number JP15KK0171 (to AK) and JP22K03752 (to YY).

Authors' contributions

AK designed this paper and performed the analysis. YN developed the

code for data preparation and analyses. All authors read and approved the final manuscript.

Acknowledgments

We used waveforms from stations that were part of Hi-net (NIED), the Hot Spring Research Institute of Kanagawa Prefecture. The PhaseNet code and model were downloaded from https://github.com/wayneweiqiang/PhaseNet (last accessed September 15, 2022). The codes for phase association (REAL), hypocenter location (VELEST), and hypocenter relocation (hypoDD) are from the following package: https://github.com/Dal-mzhang/LOC-FLOW/blob/main/README.md (last accessed September 15, 2022). Maps were drawn using Generic Mapping Tools (Wessel and Smith 1998).

Authors' information

N/A

Endnotes

N/A

Akazawa T (2004) A technique for automatic detection of onset time of P-and S-Phases in strong motion records. 13th World Conference on Earthquake Engineering
Allen RM, Kanamori H (2003) The Potential for Earthquake Early Warning in Southern California, Science, 300 (5620), 786-789.
Dai H, MacBeth C (1995) Automatic picking of seismic arrivals in local earthquake data using an artificial neural network. Geophys J Int 120:758-774. https://doi.org/10.1111/j.1365-246X.1995.tb01851.x
Fedorenko YV, Huseby ES (1999) First breaks -automatic phase pickings of P- and S-onsets in seismic records. Geophys Res Lett 26:3249-3252.
Gibbons SJ, Ringdal F (2006) The detection of low magnitude seismic events using array-based waveform correlation. Geophys J Int, 165(1):149–166. https://doi.org/10.1111/j.1365-246X.2006.02865.x
Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, Ping Tak Peter Tang (2017) On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima. arXiv:1609.04836
Kingma DP, Ba J (2014) Adam: A method for stochastic optimization. arXiv:1412.6980
Kissling E, Ellsworth WL, Eberhart-Phillips D, Kradolfer U (1994) Initial reference models in local earthquake tomography, J Geophys Res 99:19635-19646.
Lapins S, Goitom B, Kendall JM, Werner MJ, Cashman KV, Hammond JOS (2021). A little data goes a long way: Automating seismic phase arrival picking at Nabro volcano with transfer learning. J Geophys Res 126: e2021JB021910. https://doi.org/10.1029/2021JB021910
Kobayashi M, Mannen K, Okuno M, Nakamura T, Hakamata K (2006) The Owakidani Tephra group: A newly discovered post-magmatic eruption product of Hakone volcano, Japan (in Japanese with English abstract), Bull Volcanol Soc Jpn 51:245–256.
Kong Q, Richard A, Schreier L, Kwon YW (2016) MyShake: A smartphone seismic network for earthquake early warning and beyond. Science Advances 2(2) DOI: 10.1126/sciadv.150105
Mannen K, Yukutake Y, G Kikugawa, Harada M, Itadera K, Takenaka J (2018) Chronology of the 2015 eruption of Hakone volcano, Japan: geological background, mechanism of volcanic unrest and disaster mitigation measures during the crisis. Earth Planets Space 70:68 https://doi.org/10.1186/s40623-018-0844-2
Mousavi SM, Ellsworth WL, Zhu W, Chuang LY & Beroza GC (2020) Earthquake Transformer - an attentive deep-learning model for simultaneous earthquake detection and phase picking. Nat. Commun. 11, 3952, https://doi.org/10.1038/s41467-020-17591-w
Münchmeyer J, Woollam J, Rietbrock A, Tilmann F, Lange D, Bornstein T, Diehl, Giunchi C, Haslinger F, Jozinovic D, Michelini A, Saul J, Soto H (2022) Which picker fits my data? A quantitative evaluation of deep learning based seismic pickers. J Geophys Res 127:e2021JB023499
Musil M, Plešinger A (1996) Discrimination between local microearthquakes and quarry blasts by multi-layer perceptrons and Kohonen maps. Bull Seismol Soc Am 86 (4): 1077–1090 https://doi.org/10.1785/BSSA0860041077
Perol T, Gharbi M, Denolle M (2018) Convolutional neural network for earthquake detection and location. Sci. Adv. 4(2):e1700578.
Pulli J, Dysart PS (1990) An experiment in the use of trained neural networks for regional seismic event classification. Geophys Res Lett 17:977-980 https://doi.org/10.1029/GL017i007p00977
Ronneberger O, Fischer P, Brox T (2015) U-Net: convolutional networks for biomedical image segmentation. Miccai 2015, 9351:234–241. https://doi.org/10.1007/978-3-319-24574-4_28
Ross ZE, Meier MA, Hauksson E, Heaton TH (2018) Generalized seismic phase detection with deep learning, Bull.Seismol. Soc. Am. doi: 10.1785/0120180080
Smith SL, Kindermans PJ, Ying C, Le QV (2018) Don't Decay the Learning Rate, Increase the Batch Size. arXiv:1711.00489
Tiira, T (1999) Detecting teleseismic events using artificial neural networks. Comput. Geosci. 25(8):929–938.
Ursino A, Langer H, Scarfi L, Di Grazia G, Gresta S (2001) Discrimination of quarry blasts from tectonic microearthquakes in the Hyblean plateau (southeastern Sicily). Annals of Geophysics 44:703-722 https://doi.org/10.4401/ag-3569
Yukutake Y, Ito H, Honda R, Harada M, Tanada T, Yoshida A (2011a) Fluid-induced swarm earthquake sequence revealed by precisely determined hypocenters and focal mechanisms in the 2009 activity at Hakone volcano, Japan. J Geophys Res 116:B04308 doi:10.1029/2010JB008036
Yukutake Y, Honda R, Harada M (2011b) Remotely-triggered seismicity in the Hakone volcano following the 2011 off the Pacific coast of Tohoku Earthquake. Earth Planet Sp 63:41 https://doi.org/10.5047/eps.2011.05.004
Yukutake Y, Yoshida K, Honda R (2022a) Interaction between aseismic slip and fluid invasion in earthquake swarms revealed by dense geodetic and seismic observations. Journal of Geophysical Research: Solid Earth 127:e2021JB022933. https://doi.org/10.1029/2021JB022933
Yukutake Y and Kim A (2022b) Detection and hypocenter determination of volcanic earthquakes using Machine Learning: Application to Kirishima volcano, Japan Geoscience Union Meeting, 22-27, May 2022
Vaswani A, Shazeer NM, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L, Polosukhin I (2017) Attention is All you Need. NIPS'17, Proceedings of the 31st International Conference on Neural Information Processing SystemsDecember 2017 6000–6010.
Waldhauser F and Ellsworth WL (2000) A double-difference earthquake location algorithm: Method and application to the northern Hayward fault, Bull. Seismol. Soc. Am., 90:1353-1368.
Wiszniowski J, Plesiewicz BM, Trojanowski J (2014) Application of real time recurrent neural network for detection of small natural earthquakes in Poland, Acta Geophys. 62(3): 469–485.
Zhang M, Ellsworth WL, and Beroza GC. (2019) Rapid Earthquake Association and Location, Seismol. Res. Lett., 90.6:2276-2284.
Zhao Y and Takano K (1999) An artificial neural network approach for broadband seismic phase picking. Bull. Seismol. Soc. Am. 89(3): 670–680.
Zhu W, Beroza GC, PhaseNet: a deep-neural-network-based seismic arrival-time picking method, Geophysical Journal International, 216(1):261–273. https://doi.org/10.1093/gji/ggy423
Zhu W, Mousavi SM, Beroza GC (2020) Chapter Four - Seismic signal augmentation to improve generalization of deep neural networks In: Moseley B, Krischer L(ed)
Advances in Geophysics, Volume 61, Elsevier, pp151-177. https://doi.org/10.1016/bs.agph.2020.07.003.

Gabst.png

Download PDF

Journal Publication

published 16 May, 2023

Read the published version in Earth, Planets and Space →

Reviewers agreed at journal
05 Dec, 2022
Reviewers invited by journal
05 Dec, 2022
Editor assigned by journal
12 Nov, 2022
Submission checks completed at journal
11 Nov, 2022
Editor invited by journal
11 Nov, 2022
First submitted to journal
10 Nov, 2022

You are reading this latest preprint version

Development of High-Performance Seismic Phase Picker Using Deep Learning in Hakone Volcanic Area

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Seismic Activities And Observations Network In Hakone

Construction Of Seismic Phase Picker

Application To Test Data

Application To Continuous Data

Discussions

Conclusions

Abbreviations

Declarations

References

Supplementary Files

Status:

Journal Publication

Version 1