NH4.3 | Machine learning and statistical models applied to earthquake occurrence
EDI
Machine learning and statistical models applied to earthquake occurrence
Co-organized by SM8
Convener: Stefania Gentili | Co-conveners: Rita Di Giovambattista, Álvaro González, Filippos Vallianatos
Orals
| Wed, 26 Apr, 10:45–12:30 (CEST), 14:00–15:40 (CEST)
 
Room 2.17
Posters on site
| Attendance Wed, 26 Apr, 16:15–18:00 (CEST)
 
Hall X4
Posters virtual
| Attendance Wed, 26 Apr, 16:15–18:00 (CEST)
 
vHall NH
Orals |
Wed, 10:45
Wed, 16:15
Wed, 16:15
New physical and statistical models based on observed seismicity patterns shed light on the preparation process of large earthquakes and on the temporal and spatial evolution of seismicity clusters.

As a result of technological improvements in seismic monitoring, seismic data is nowadays gathered with ever-increasing quality and quantity. As a result, models can benefit from large and accurate seismic catalogues. Indeed, accuracy of hypocenter locations and coherence in magnitude determination are fundamental for reliable analyses. And physics-based earthquake simulators can produce large synthetic catalogues that can be used to improve the models.

Multidisciplinary data recorded by both ground and satellite instruments, such as geodetic deformation, geological and geochemical data, fluid content analyses and laboratory experiments, can better constrain the models, in addition to available seismological results such as source parameters and tomographic information.

Statistical approaches and machine learning techniques of big data analysis are required to benefit from this wealth of information, and unveiling complex and nonlinear relationships in the data. This allows a deeper understanding of earthquake occurrence and its statistical forecasting.

In this session, we invite researchers to present their latest results and findings in physical and statistical models and machine learning approaches for space, time, and magnitude evolution of earthquake sequences. Emphasis will be given to the following topics:

• Physical and statistical models of earthquake occurrence.
• Analysis of earthquake clustering.
• Spatial, temporal and magnitude properties of earthquake statistics.
• Quantitative testing of earthquake occurrence models.
• Reliability of earthquake catalogues.
• Time-dependent hazard assessment.
• Methods and software for earthquake forecasting.
• Data analyses and requirements for model testing.
• Machine learning applied to seismic data.
• Methods for quantifying uncertainty in pattern recognition and machine learning.

Orals: Wed, 26 Apr | Room 2.17

Chairpersons: Stefania Gentili, Álvaro González, Rita Di Giovambattista
10:45–10:50
Spatiotemporal evolution of seismicity
10:50–11:10
|
EGU23-4169
|
solicited
|
Virtual presentation
Eleftheria Papadimitriou

The complexity of seismogenesis requires the development of stochastic models, the application of which aims to improve our understanding on the seismic process and the associated underlying process. Semi Markov models are introduced for estimating the expected number of earthquake occurrences with the classification of the model states based on earthquake magnitudes. The instantaneous earthquake occurrence rate between the model states as well as the total earthquake occurrence rate can be calculated. Seismotectonic characteristic of the study area, incorporated in the model as important component of the model, increase the consistency between the model and the process of earthquake generation and support a classification that integrates magnitudes and style of faulting, thus being more effective for the seismic hazard assessment. For revealing the stress field underlying the earthquake generation, which is not accessible to direct observation, the hidden Markov models (HMMs) are engaged. The HMMs consider that the states correspond to levels of the stress field and its application aim to reveal these states. Different number of states may be examined, dependent upon the organization of observations, and the HMMs are capable to reveal the number of stress levels as well as the way in which these levels are associated with the occurrence of certain earthquakes. Even better results are obtained via the application of hidden semi–Markov models (HSMMs) considering that the stress field constitutes the hidden process and which, compared with HMMs, allow any arbitrary distribution for the sojourn times. The investigation of the interactions between adjacent areas is accomplished by means of the linked stress release model (LSRM), based upon the consideration that spatio–temporal stress changes and interactions between adjacent fault segments constitute the most important component in seismic hazard assessment, as they can alter the occurrence probability of strong earthquakes onto these segments. The LSRM comprises the gradual increase of the strain energy due to continuous tectonic loading and its sudden release during the earthquake occurrence. The modeling is based on the theory of stochastic point process, and it is determined by the conditional intensity function. In an attempt to identify the most appropriate parameterization that better fits the data and describes the earthquake generation process, a constrained “m–memory” point process is introduced, the Constrained–Memory Stress Release Model (CM–SRM) implying that only the m most recent arrival times are taken into account in the conditional intensity function, for some suitable mÎN, instead of the entire history of the process. The performance of the above mentioned models application are evaluated and compared in terms of information criteria and residual analysis.

How to cite: Papadimitriou, E.: Stochastic models for earthquake occurrence, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-4169, https://doi.org/10.5194/egusphere-egu23-4169, 2023.

11:10–11:20
|
EGU23-15321
|
ECS
|
On-site presentation
Aron Mirwald, Laura Gulia, and Stefan Wiemer

The Parkfield section of the San Andreas fault has a history of frequently occurring moderate (M~6) earthquakes with recurrence times ranging from 12 to 38 years. Since 1985, it has been extensively monitored as part of the experiment to predict the next moderate earthquake. Using the rich data resulting from the high-resolution monitoring, studies have revealed several interesting and consistent patterns of the frequency-magnitude distribution (FMD) of earthquakes, measured by the b-value of the Gutenberg-Richter law. The fault consists of patches of low b-values (b < 0.6) that correlate well with locked patches and with the areas that slipped in the 2004 M6 earthquake. High b-values (b > 1.3) were found to correlate with creeping section of the faults, and both observations support the hypothesis of an inverse relation between differential stress and b-values. Further, the b-value was found to increase during the aftershock periods of the 2004 earthquake, but so far, no gradual loading throughout the seismic cycle has been documented at Parkfield.

Here we revisit the b-values along the Parkfield section 19 years after the last M6 event, with the objectives to monitor and better understand the evolution of b-values in space and time as the segment approaches the next rupture. Our aim is first to benchmark and enhance approaches to map and monitor transients, to optimize uncertainty quantification, robustness, and resolving power of our statistical methods. This is best targeted by creating synthetics catalogues with known properties and then benchmarking different methods for spatial mapping and time-series analysis of b-values. We specifically investigate the recently introduced b-positive estimator and convert observed b-values and activity rates to earthquake probabilities. In a second step, we analyse the observed patterns in a context of gradual fault loading and repeated moderate events, to derive insights into the underlying physical processes. Finally, our aim to set up a ‘b-value’ observatory that will continuously monitor the space-time evolution of b-values and earthquake probabilities.

How to cite: Mirwald, A., Gulia, L., and Wiemer, S.: b-value variation through the seismic cycle: Revisiting Parkfield, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15321, https://doi.org/10.5194/egusphere-egu23-15321, 2023.

11:20–11:30
|
EGU23-17092
|
ECS
|
Virtual presentation
Gina-Maria Geffers, Ian G. Main, and Mark Naylor

The Gutenberg-Richter b-value represents the relative proportion of small to large earthquakes in a scale-free population and is an important parameter used in earthquake hazard assessment. Discussion of the amount of data required to obtain a robust b-value has been extensive and is ongoing. To complement these analyses, we show the effect of the b-value with changes to the dynamic range – the difference between minimum magnitude (or magnitude of completeness) and maximum magnitude, which is inherently linked to the sample size, but not proportionately correlated. Additionally, we show that biases in high b-values are due to the bias in the mean magnitude of a catalogue, which asymptotically converges from below.

We derive and analytic expression for the bias that arises in the maximum likelihood estimate of b as a function of dynamic range r. Our theory predicts the observed evolution of the modal value of the mean magnitude in multiple random samples of synthetic catalogues at different r, including the bias to high b at low r and the observed trend to an asymptotic limit with no bias. In the case of a single sample in real catalogues, the situation is substantially more complicated due to the heterogeneity, magnitude uncertainty and lack of knowledge of the true b-value. We summarise that these results explain why the likelihood of large events and the associated hazard is often underestimated in small catalogues with low r, for example in some studies of volcanic and induced seismicity.

How to cite: Geffers, G.-M., Main, I. G., and Naylor, M.: Frequency-size parameters as a function of dynamic range – the Gutenberg-Richter b-value for earthquakes, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-17092, https://doi.org/10.5194/egusphere-egu23-17092, 2023.

11:30–11:40
|
EGU23-14001
|
ECS
|
On-site presentation
Marta Han, Leila Mizrahi, and Stefan Wiemer

The go-to models for developing time-dependent earthquake forecasts are Epidemic-Type Aftershock Sequence (ETAS) models. They model earthquake occurrence as a spatio-temporal self-exciting point process, using basic empirical laws such as the Omori-Utsu law for the temporal evolution of aftershock rate, the Gutenberg-Richter law to describe the size distribution of earthquakes, the exponential productivity law and so on. The main focus and core strength of ETAS lie in modelling aftershock occurrence. An important aspect which holds great potential for improvement is the modelling of background seismicity.

In this study, we focus on the data sets and expert solicitations acquired for building the European Seismic Hazard Model (ESHM) 2020. Since these data sets cover a wide range of space and time, the properties of the earthquake catalogs (completeness magnitude, magnitude resolution, time and space resolution, b-value) vary by region and time period. We address these issues using the model accounting for the time-varying completeness magnitude (Mizrahi et al., 2021) and other adjustments, then develop an ETAS model that allows the background seismicity rate to vary with space and be covariate-dependent. The expectation-maximisation-based algorithm allows for these rates to be given as an input, in our case based on fault locations, estimated long-term seismicity rates and area sources, or estimated during inversion for expert-defined zonations. We test the models retrospectively for self-consistency and pseudo-prospectively to identify the ones that lead to the best operational forecasting model for Europe.

How to cite: Han, M., Mizrahi, L., and Wiemer, S.: Towards an Operational Earthquake Forecasting Model for Europe, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14001, https://doi.org/10.5194/egusphere-egu23-14001, 2023.

11:40–11:50
|
EGU23-1035
|
ECS
|
On-site presentation
Giuseppe Petrillo, Jiancang Zhuang, and Eugenio Lippiello
The seismic gap hypothesis states that fault regions, where no large earthquake has recently occurred, are more prone than others to host the next large earthquake. This could allow an estmate of the occurrence probability of the next big shock on the basis of the time delay from the last earthquakes. Recent results, both numerical and instrumental, have shown that aftershock occurrence can provide important insights about the validty and range of applicability of the GAP hypothes. Here we discuss how to include the information of these new results in Self Exciting Point Process SEPP models, oiginally developed to describe only aftershock spatio-temporal patterns. In particular, using as testing laboratory a numerical model which reproduces all relevant statistical features of earthquake occurrence, we show as the introduction of stress release in SEPP models can improve long term earthquake prediction.

How to cite: Petrillo, G., Zhuang, J., and Lippiello, E.: Is the stress relaxation relevant for long term forecasting?, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-1035, https://doi.org/10.5194/egusphere-egu23-1035, 2023.

11:50–12:00
|
EGU23-5934
|
ECS
|
On-site presentation
Antoine Septier, Alexandra Renouard, Jacques Déverchère, and Julie Perrot

Due to the complexity and high dimensionality of seismic catalogues, the dimensional reduction of raw seismic data and the feature selections needed to decluster these catalogues into crisis and non-crisis events remain a challenge. To address this problem, we propose a two-level analysis.

 

First, an unsupervised approach based on an artificial neural network called self-organising map (SOM) is applied. The SOM is a machine learning model that performs a non-linear mapping of large input spaces into a two-dimensional grid, which preserves the topological and metric relationships of the data. It therefore facilitates visualisation and interpretation of the results obtained. Then, agglomerative clustering is used to classify the different clusters obtained by the SOM method as containing background events, aftershocks and/or swarms. To estimate the classification uncertainty and confidence level of our declustering results, we developed a probabilistic function based on the feature representation learned by the SOM (spatiotemporal distances between events, magnitude variations and event density).

 

We tested the two-level analysis on synthetic data and applied it to real data: three seismic catalogues (Corithn Rift, Taiwan and Central Italy) that differ in area size, tectonic regime, magnitude of completeness, duration and detection methods. We show that our unsupervised machine learning approach can accurately distinguish between crisis and non-crisis events without the need for preliminary assumptions and that it is applicable to catalogues of various sizes in time and space without threshold selection.

How to cite: Septier, A., Renouard, A., Déverchère, J., and Perrot, J.: Can Seismicity Declustering be solved by Unsupervised Artificial Intelligence ?, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5934, https://doi.org/10.5194/egusphere-egu23-5934, 2023.

12:00–12:10
|
EGU23-11684
|
ECS
|
On-site presentation
Ying Zhang and Qinghua Huang

Artificial Neural Networks (ANNs) are well known for their ability to find hidden patterns in data. This technique has also been widely used for predicting the time, location, and magnitude of future earthquakes. Using various data and neural networks, previous works claimed that their models are effective for predicting earthquakes. However, these scores provided by the evaluation metrics with poor reference models, which are non-professional in statistical seismology, are not robust. In this work, we first take the Nighttime Light Map (NLM) as the input of Long-short Term Memory (LSTM) networks to predict the earthquakes with M>=5.0 for the whole Chinese Mainland, and NLM records the lumens of nighttime artificial light, and it is retrieved from the nighttime satellite imagery. The NLM is not physically related to earthquakes; however, the scores provided by Receiver Operating Characteristics curve, Precision-Recall plot, and Molchan diagram with spatial invariant Poisson model indicated that NLM is effective for predicting earthquakes. These results reaffirmed that researchers should be cautious when using these evaluation metrics with poor reference models to evaluate earthquake prediction models. Moreover, the original loss functions of ANNs, such as Cross Entropy (CE), Balanced Cross Entropy (BCE), Focal Loss (FL), and Focal Loss alpha (FL-alpha), contain no knowledge about seismology. To differentiate the hard and easy examples of earthquake prediction models during the training steps of ANNs, the punishment of CE, BCE, FL, and FL-alpha for positive examples will be further weighted by P0 and the punishment for negative examples will be weighted by P1, where P1/P0 is the prior probability provided by the reference model that at least one or no earthquakes will occur for the given example and P1+P0=1. The reference models are supposed to be as close to the real spatial-temporal distribution of earthquakes as possible, and the spatial variable Poisson (SVP) model is the simplest version which is also friendly to data mining experts. In this work, we choose the SVP as the reference model to revise these previous loss functions and take the estimated cumulative earthquake energy in the time-space unit (1 degree*1 degree*10 days) as the input of the LSTM to predict the earthquakes with M>=5.0 in the whole Chinese Mainland, and we use the Molchan diagram (SVP) and Area Skill Score (ASS) to evaluate the performance of these models. Results show that the majority of models (134 out of 144) trained by original loss functions are ineffective for predicting earthquakes; however, the scores of models trained using the revised loss functions have been obviously improved, and 83 out of 144 models are proved to be better than SVP in predicting earthquakes. Our results indicate that designing a more complex structure for ANN and neuron is not the only way to improve the performance of ANNs for predicting earthquakes, and how combining the professional knowledge of data mining experts and seismologists deserves more attention for the future development of ANN-based earthquake prediction models.

How to cite: Zhang, Y. and Huang, Q.: Improved ANN-based earthquake prediction system with reference model engaged in the evaluation metrics, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11684, https://doi.org/10.5194/egusphere-egu23-11684, 2023.

Detection of earthquakes and tremors
12:10–12:20
|
EGU23-13113
|
On-site presentation
Hiromichi Nagao, Ryosuke Kaneko, Shin-ichi Ito, Hiroshi Tsuruoka, and Kazushige Obara

The establishment of the High Sensitivity Seismograph Network (Hi-net) in Japan has led to the discovery of deep low-frequency tremors. Since such tremors are considered to be associated with large earthquakes adjacent to tremors on the same subducting plate interface, it is important in seismology to investigate these tremors before establishing modern seismograph networks that record seismic data digitally. We propose a deep-learning method to detect evidence of tremors from seismogram images recorded on paper more than 50 years ago. In this study, we trained a convolutional neural network (CNN) based on the Residual Network (ResNet) with seismogram images converted from real seismic data recorded by Hi-net. The CNN trained by fine-tuning achieved an accuracy of 98.64% for determining whether an input image contains tremors. The Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps for visualizing model predictions indicated that the CNN successfully detects tremors without being affected by teleseisms. The trained CNN was applied to the past seismograms recorded from 1966 to 1977 at the Kumano observatory, in southwest Japan, operated by Earthquake Research Institute, The University of Tokyo. The CNN showed potential for detecting tremors from past seismogram images for broader applications, such as publishing a new tremor catalog, although further training using data including more variables such as the thickness of the pen would be required to develop a universally applicable model.

How to cite: Nagao, H., Kaneko, R., Ito, S., Tsuruoka, H., and Obara, K.: Detection of Deep Low-Frequency Tremors from Continuous Paper Records at a Station in Southwest Japan About 50 Years Ago Based on Convolutional Neural Network, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-13113, https://doi.org/10.5194/egusphere-egu23-13113, 2023.

12:20–12:30
|
EGU23-15180
|
On-site presentation
Hamzeh Mohammadigheymasi, Nasrin Tavakolizadeh, Peidong Shi, Zhuowei Xiao, S. Mostafa Mousavi, and Rui Fernandes

Modern seismology can benefit from a rapid and reliable earthquake catalog preparation process. In recent years Deep Learning (DL)-based methods attracted seismologists’ attention to keep up with the constantly increasing seismic data for maximally processing and locating the recorded events. This study focuses on deploying an optimized workflow that integrates DL and waveform migration algorithms to achieve a comprehensive automated phase detection and earthquake location workflow. The goal is to deeply scan the seismic datasets for Pand S- phases, associate and locate the detected events, and improve the performance of DL algorithms in processing out-of-distribution and low signal-to-noise ratio data. The workflow consists of six steps, including the preparation of one-minute data segments by employing the framework of ObsPy, deep investigation of the recorded data for P- and S- phases by a low threshold EQTransformer (EQT), and a pair-input Siamese EQTransformer (S-EQT), phase association by Rapid Earthquake Association and Location (REAL) method, applying MIgration Location (MIL) to accurately locate the outputs of REAL, and calculating the local magnitude of the located earthquakes. Eighteen months of the Ghana Digital Seismic Network (GHDSN) dataset (2012-2014), is processed by this integrated and automatic workflow, and a catalog of 461 earthquakes is acquired. Although S-EQT and EQT, with the respective number of 758 and 423 earthquakes, show a figurative superiority in the number of detected events, they are scattered with inaccurate hypo-central depth. Conversely, the compiled catalog show high accordance with the previously interpreted seismogenic sources by Mohammadigheymasi et al. (2023), and a new seismogenic source is also delineated. This workflow significantly enhanced the seismic catalog compilation process and lowered the computational costs while increasing the accuracy of phase detection, association, and location processes. This work was supported by the European Union and the Instituto Dom Luiz(IDL) Project under Grant UIDB/50019/2020, and it uses computational resources provided by C4G (Collaboratory for Geosciences) (Ref. PINFRA/22151/2016). P. S. is supported by the DEEP project (http://deepgeothermal.org) funded through the ERANET CofundGEOTHERMICA (Project No. 200320-4001) from the European Commission. The DEEP project benefits from an exploration subsidy of the Swiss federal office of energy for the EGSgeothermal project in Haute-Sorne, canton of Jura (contract number MF-021-GEO-ERK), which is gratefully acknowledged.

How to cite: Mohammadigheymasi, H., Tavakolizadeh, N., Shi, P., Xiao, Z., Mousavi, S. M., and Fernandes, R.: An automated earthquake detection algorithm by combining pair-input deep learning and migration location methods, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15180, https://doi.org/10.5194/egusphere-egu23-15180, 2023.

Lunch break
Chairpersons: Stefania Gentili, Álvaro González, Filippos Vallianatos
14:00–14:10
|
EGU23-16438
|
Virtual presentation
Luis Carvalho, Hamzeh Mohammadigheymasi, Paul Crocker, Nasrin Tavakolizadeh, Yahya Moradichaleshtori, and Rui Fernandes

A temporary seismic network consisting of 32 broadband seismic sensors was installed in Cameroon between March 2005 and December 2006 to study the seismic structure of the crust and upper mantle beneath the Cameroon Volcanic Line (CVL). This study aims to re-evaluate the seismicity in this period by processing this database and calculating an updated crustal velocity model for the region incorporating the acquired earthquake bulletin. 

The earthquake detection and location procedure applies hybrid deep learning (DL) and phase validation methods. We use an integrated workflow composed of Earthquake Transformer (EqT) and Siamese Earthquake Transformer (S-EqT) for initial earthquake detection and phase picking. Then, PickNet is used as a phase refinement step, and REAL for earthquake association and rough location. A set of thresholding parameters for earthquake detection and P- and S-picking equal to 0.2 and 0.07 are adjusted, respectively. By combining a set of 33282 P and 29251 S-picked phases associated with 743 earthquakes with 1.3 ≤ ML ≤ 4.6, we implement a joint inversion for estimating an updated 1D crustal velocity model. The obtained mode comprises thicknesses of 8, 12, 14, 20, and 30km, from the surface to a depth of 45km, with Vp = 6.1, 6.4, 6.6, 7.6, 8.25, and 8.5km/s, respectively. The newly detected events are primarily concentrated in three main clusters, 1) the east flank of Mount Cameroon, 2) an area between Mount Cameroon and Bioko Island, and 3) southern Bioko Island. The compiled catalog for this time interval is 1.7 times larger than the already reported catalog for this data set. Finally, we present a 3D time-lapsed animation of the detected earthquake sequences.

Acknowledgements: The authors would like to thank the support of the Instituto de Telecomunicações. This work is funded by FCT/MCTES through national funds and, when applicable co-funded EU funds under the project UIDB/50008/2020.

How to cite: Carvalho, L., Mohammadigheymasi, H., Crocker, P., Tavakolizadeh, N., Moradichaleshtori, Y., and Fernandes, R.: Earthquake Detection and Location in the Cameroon Temporary Network Data Using Deep Learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-16438, https://doi.org/10.5194/egusphere-egu23-16438, 2023.

14:10–14:20
|
EGU23-10395
|
ECS
|
Virtual presentation
|
Alireza Niksejel and Miao Zhang

Offshore earthquakes recorded by Ocean Bottom Seismometers (OBS) are crucial to studying tectonic activities in the subduction zones and mid-ocean ridges. In recent years, the ever-advancing Machine Learning (ML)-based phase pickers have shown promise in land earthquake monitoring, but there are few available ML models to handle OBS data due mainly to the lack of labelled training sets and low signal-to-noise ratios. Though land ML-based phase pickers may roughly work for OBS data, they introduce a large number of false negatives and false positives, leading to numerous events being missing and fake.

In this study, we create a tectonically inclusive OBS training data set and develop a generalized deep-learning OBS phase picker - OBSPicker using the EQTransformer (EQT; Mousavi et al., 2019) and the transfer learning approach. To create an inclusive OBS training data set, we collect earthquake waveforms from routine catalogues recorded at 11 OBS networks worldwide with different tectonic settings and geographic locations. Earthquakes are recorded in local and regional distances with diverse magnitudes (ML 0.0-5.8), source depths (0-250 km), and epicentral distances (0-3 deg). To label their phase picks, we adopt a sequence of processing steps including 1) initial phase arrival detection and picking by EQT, 2) identifying and discarding samples with multiple (unwanted) events using STA/LTA method, and 3) refining phase picks using the Generalized Phase Detection method (GPD, Ross et al., 2018), resulting in ~38,000 well-labelled earthquake samples. In addition, we also collect ~150,000 OBS noise samples from the same OBS networks for training augmentation instead of using the commonly adopted Gaussian noises. Those OBS noise samples are used to simulate low-magnitude earthquakes under different marine environments. Initial results show that our transfer-learned OBS phase picker outperforms the EQTransformer base model in both accuracy and precision, especially in presence of higher levels of noise.

How to cite: Niksejel, A. and Zhang, M.: OBSPicker: A generalized transfer-learned OBS phase picker, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10395, https://doi.org/10.5194/egusphere-egu23-10395, 2023.

Seismic event discrimination and characterisation
14:20–14:30
|
EGU23-6343
|
ECS
|
On-site presentation
Céline Hourcade, Mickaël Bonnin, and Éric Beucler

Over the past 15 years, the deployment of dense permanent seismic networks leads to a dramatic increase in the amount of data to process. The seismic coverage and the station quality pave the way toward a comprehensive catalogue of natural seismicity. This means to i) detect the lowest magnitudes as possible and ii) to discriminate natural from anthropogenic events. To achieve this discrimination, we present a new convolutional neural network (CNN) trained from 60 s long three component spectrograms between 1 and 50 Hz. This CNN is trained using a reliable database of labelled events located in Metropolitan France between January 2020 and June 2021. The application of our trained model on the detected events in Metropolitan France between June and November 2021 gives a high discrimination accuracy of 98.18%. To demonstrate the versatility of our approach, this trained model is applied to different catalogues: from a post-seismic campaign in NW France (48 events) and from University of Utah Seismograph Stations, Utah, USA, (396 events between January and March 2016). We reach an accuracy of 100.00% and 96.72%, respectively, for the discrimination between natural and anthropogenic events. Since each discrimination comes with a level of confidence, our approach can be seen as a decision making tool for the analysts. It also allows to build reliable seismic event catalogues and to reduce the number of mislabelled events in the databases.

How to cite: Hourcade, C., Bonnin, M., and Beucler, É.: New CNN based tool to discriminate anthropogenic from natural low magnitude seismic events, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6343, https://doi.org/10.5194/egusphere-egu23-6343, 2023.

14:30–14:40
|
EGU23-6202
|
ECS
|
On-site presentation
Gunnar Eggertsson, Björn Lund, Peter Schmidt, and Michael Roth

Distinguishing small earthquakes from man-made blasts at construction sites, in quarries and in mines is a non-trivial task during automatic event analysis and thus typically requires manual revision. We have developed station-specific classification models capable of both accurately assigning source type to seismic events in Sweden and filtering out spurious events from an automatic event catalogue. Our method divides all three components of the seismic records for each event into four non-overlapping time windows, corresponding to P-phase, P-coda, S-phase and S-coda, and computes the Root-Mean-Square (RMS) amplitude in each window. This process is repeated for a total of twenty narrow frequency bands. The resulting array of amplitudes is passed as inputs to fully connected Artificial Neural Network classifiers which attempt to filter out spurious events before distinguishing between natural earthquakes, industrial blasts and mining-induced events. The distinction includes e.g. distinguishing mining blasts from mining induced events, shallow earthquakes from blasts and differentiating between different types of mining induced events. The classifiers are trained on labelled seismic records dating from 2010 to 2021. They are already in use at the Swedish National Seismic Network where they serve as an aid to the routine manual analysis and as a tool for directly assigning preliminary source type to events in an automatic event catalogue. Initial results are promising and suggest that the method can accurately distinguish between different types of seismic events registered in Sweden and filter out the majority of spurious events.

How to cite: Eggertsson, G., Lund, B., Schmidt, P., and Roth, M.: Supervised Learning for Automatic Source Type Discrimination of Seismic Events in Sweden, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6202, https://doi.org/10.5194/egusphere-egu23-6202, 2023.

14:40–14:50
|
EGU23-16410
|
ECS
|
On-site presentation
Chantal van Dinther, Marielle Malfante, Pierre Gaillard, and Yoann Cano

Recent employment of large seismic arrays and distributed fibre optic sensing cables leads to an overwhelming amount of seismic data. As a consequence, the need for reliable automatic processing and analysis techniques increases. Therefore, the number of machine learning applications for detection and classification of seismic signal augments too.

A challenge however, is that seismic datasets are highly class imbalanced, i.e. certain seismic classes are dominant while others are underrepresented. Unfortunately, a skewed dataset may lead to biases in the model and thus to higher uncertainties in the model predictions. In the machine learning literature, several strategies are described to mitigate this problem. In presented work we explore and compare those approaches.

For our application, we use catalogues and seismic continuous recordings of the RD network in France [RESIF, 2018]. Using a simple 3-layered convolutional neural network (CNN) we aim to differentiate between six seismic classes, which are based on hand-picked catalogues. The training set we obtained is highly skewed with earthquakes as the majority class, containing 77% of the samples.  The remaining classes (quarry blasts, marine explosions, suspected induced events, noise and earthquakes with unquantifiable magnitude) represent 2.1 - 7.5% of the dataset.

We compare four strategies to deal with an imbalanced datasets for a multi-class classification problem. The first strategy is to resample the dataset (i.e. reduction of the majority class). Another approach is the adaptation of the loss function by weighting the classes when penalizing the loss (i.e. increasing the weight of the minority classes). Those class weights can be adjusted either w.r.t. the reciprocal of class frequency [inspired by King and Zeng, 2001] or w.r.t. the effective number of samples [Cui et al., 2019]. Lastly, we have explored the use of a focal loss function [Lin et al., 2020].

Using balanced accuracy as a metric while minimizing the loss, we found that in our case adjusting the class weights in the loss function according to the reciprocal of the class frequency provides the best results.

 

References:

- RESIF, 2018: https://doi.org/10.15778/RESIF.RD

- King, G., & Zeng, L. (2001). Logistic regression in rare events data. Political analysis9(2), 137-163.

- Lin et al. (2020), Focal Loss for Dense Object Detection, IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 42, NO. 2, FEBRUARY 2020

- Cui et al. (2019), Class-Balanced Loss Based on Effective Number of Samples, https://doi.org/10.48550/arXiv.1901.05555

How to cite: van Dinther, C., Malfante, M., Gaillard, P., and Cano, Y.: Increasing the reliability of seismic classification: A comparison of strategies to deal with class size imbalanced datasets., EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-16410, https://doi.org/10.5194/egusphere-egu23-16410, 2023.

14:50–15:00
|
EGU23-5868
|
ECS
|
Virtual presentation
|
Neri Berman, Oleg Zlydenko, Oren Gilon, and Yohai Bar-Sinai

Standard approaches to earthquake forecasting - both statistics-based models, e.g. the epidemic type aftershock (ETAS), and physics-based models, e.g. models based on the Coulomb failure stress (CFS) criteria, estimate the probability of an earthquake occurring at a certain time and location. In both modeling approaches the time and location of an earthquake are commonly assumed to be distributed independently of their magnitude. That is, the magnitude of a given earthquake is taken to be the marginal magnitude distribution, the Gutenberg-Richter (GR) distribution, typically constant in time,or fitted to recent seismic history. Such model construction implies an assumption that the underlying process determining where and when an earthquake occurs is decoupled from the process that determines its magnitude.

In this work we address the question of magnitude independence directly. We build a machine learning model that predicts earthquake magnitudes based on their location, region history, and other geophysical properties. We use neural networks to encode these properties and output a  conditional magnitude probability distribution, maximizing on the log-likelihood of the model’s prediction. We discuss the model architecture, performance, and evaluate this model against the GR distribution.

How to cite: Berman, N., Zlydenko, O., Gilon, O., and Bar-Sinai, Y.: Earthquake Magnitude Prediction Using a Machine Learning Model, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5868, https://doi.org/10.5194/egusphere-egu23-5868, 2023.

Denoising of seismic records
15:00–15:10
|
EGU23-6862
|
ECS
|
On-site presentation
|
Janis Heuel, Meggy Roßbach, and Wolfgang Friederich

Seismogram records always contain seismic noise from different sources. Previous studies have shown that denoising autoencoders can be used to suppress different types of disturbing noise at seismological stations, even when earthquake signal and noise share common frequency bands. A denoising autoencoder is a convolutional neural network that learns from a large training data set how to separate earthquake signal and noise. To train the denoising autoencoder, we used earthquake signals with high signal-to-noise ratio from the Stanford Earthquake Dataset and noise from single seismological stations. We used 160 seismological stations in Germany and surrounding countries and trained a denoising autoencoder model for each station. Afterwards, one year of continuous recorded data have been denoised.

EQTransformer, a deep-learning model for earthquake detection and phase picking, was then applied to the raw and denoised data of each station. Working with denoised data leads to a massive increase of earthquake detections. First results show that in dense seismic networks more than 100% additional earthquakes can be detected compared to events detected in the raw data set. Moreover, the localization accuracy is increased as more stations can be used.

However, like common filter techniques, denoising autoencoders decrease the waveform amplitude. Since earthquake magnitudes are often determined from these amplitudes, we expect a lower amplitude and thus a lower magnitude when using denoised data instead of raw data. So far, we did not find an empirical relation between the raw and denoised magnitude.

How to cite: Heuel, J., Roßbach, M., and Friederich, W.: How does a denoising autoencoder improve earthquake detection and the estimation of magnitude in seismic networks?, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6862, https://doi.org/10.5194/egusphere-egu23-6862, 2023.

15:10–15:20
|
EGU23-1309
|
ECS
|
On-site presentation
Deniz Ertuncay, Simone Francesco Fornasari, and Giovanni Costa

Seismic stations record superpositions of the seismic signals generated by all kinds of seismic sources. In earthquake seismology, seismic noise sources can be natural events such as wind or anthropogenic events such as cars. In this study, we developed a machine learning (ML) based algorithm to remove the noise from earthquake data. This is important since the information related with the features of the seismic event may be overlapped by the noise. The presence of noise in the recordings can affect the performance of the seismic network, lowering its sensibility and increasing the magnitude of completeness of the seismic catalogue. To train ML model, 10000 thousand earthquake records with relatively low signal to noise ratio (SNR) are selected and contaminated by 25 noise records that are magnified up to 50% of peak amplitude of the earthquake signal and frequency content of those signals are stored as three component traces. The architecture used consists of an Attention U-Net, i.e. an encoder-decoder model using an attention gate within the skip connections: the encoder maps samples from input space (the waveform STFTs) to a latent space while the decoder maps the latent space to the output space (the signal-noise mask). Skip connections are introduced to recover, from previous layers, fine details lost in the encoding-decoding process. Attention gates identify salient regions and prune inputs to preserve only the ones relevant to the specific task. The use of attention gates in skip connections allows to pass "fine-detailed" information to high levels of the decoder that the model itself considers useful to the waveform segmentation. Trained model is tested with a new set of data to understand its capabilities. It is found that trained model can significantly improve the SNR of noise signals with respect to standard filtering methods. Hence, it can be considered as a strong candidate for seismic data filtering method. 

How to cite: Ertuncay, D., Fornasari, S. F., and Costa, G.: Seismic Signal Denoising with Attention U-Net, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-1309, https://doi.org/10.5194/egusphere-egu23-1309, 2023.

15:20–15:30
|
EGU23-11985
|
ECS
|
On-site presentation
Nikolaj Dahmen, John Clinton, Men-Andrin Meier, Simon Stähler, Savas Ceylan, Constantinos Charalambous, Doyeon Kim, Alexander Stott, and Domenico Giardini

Marsquake recordings by NASA’s InSight seismometer often have low signal-to-noise ratios (SNR) owing to low marsquake amplitudes - only a handful of events are over M3.5 and epicentral distances are large, due to the single station being located in a seismically quiet region, and highly fluctuating atmospheric, spacecraft and instrumental noise signals.

We have previously shown [1] how deep convolutional neural networks (CNN) can be used for 1) event detection - thereby producing an event catalogue consistent with the manually curated catalogue by the Marsquake Service (MQS) [2], and further extending it from 1297 to 2079 seismic events - as well as for 2) separating event and noise signals in time-frequency domain. Due to the low number of events readily-available for network training, we trained the CNN on synthetic event data combined with recorded InSight noise.

Here, we construct a semi-synthetic data set (with real marsquake & noise data) to assess the denoising performance of the CNN w.r.t. to various evaluation metrics such as SNR, signal-distortion-ratio, cross-correlation, and peak amplitude of the recovered event waveforms, and compare modifications of the CNN architecture and the training data set.

For a large number of identified events [1,2] no distance estimates are available (or only with high uncertainty), and for all but a small subset the back azimuth is unclear, as the relatively high background noise often obscures this information in the waveforms. We explore how the denoised waveforms can support the phase picking and polarisation analysis of marsquakes, and with that their localisation, as well as their general characterisation.

 

References:

[1] Dahmen et al. (2022), doi: 10.1029/2022JE007503

[2] Ceylan et al. (2022), doi: 10.1016/j.pepi.2022.106943

How to cite: Dahmen, N., Clinton, J., Meier, M.-A., Stähler, S., Ceylan, S., Charalambous, C., Kim, D., Stott, A., and Giardini, D.: Denoising InSight’s marsquake recordings with deep learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11985, https://doi.org/10.5194/egusphere-egu23-11985, 2023.

Spatiotemporal evolution of seismicity
15:30–15:40
|
EGU23-9874
|
Virtual presentation
Ilya Zaliapin and Yehuda Ben-Zion

Progressive localization of deformation may signify a regional preparation process leading to large earthquakes [Ben-Zion & Zaliapin, GJI, 2020; Kato & Ben-Zion, Nat Rev Earth Environ. 2021]. The localization framework describes the evolution from distributed failures in a rock volume to localized system-size events. Ben-Zion & Zaliapin (2020) documented robust cycles of localization and de-localization of background earthquakes with M > 2 in Southern California that precede the M7 earthquakes within 2-4 years. This analysis has been done on regional scale, without posterior selection of the examined areas (e.g., around epicenters of large events). Similar results are observed before M7.8 earthquakes in Alaska using background seismicity with M > 4, and in laboratory acoustic emission experiments.

In this work we examine spatial characteristics of the localization process, identifying sub-regions that are responsible for the observed localization and delocalization. The analysis focuses on relative (with respect to other areas) changes in the background intensity. On sub decadal temporal scale, the observed relative seismic activity tends to concentrate on and switch between several subsets of the regional fault network. Within 2-10 years prior to a large event, there is relative activation in a large volume that not necessarily include the impending epicenter. This is followed by a prominent deactivation 2-3 years prior to a large event, reminiscent of the “Mogi donut”, potentially reflecting a transition to aseismic or small events. Some regions may experience multiple activation episodes before a large earthquake. The results emphasize the importance of examining small-magnitude events and joint analyses of seismic and geodetic data.

References:

  • Ben-Zion, Y. and I. Zaliapin (2020) Localization and coalescence of seismicity before large earthquakes. Geophysical Journal International, 223(1), 561-583. doi:10.1093/gji/ggaa315
  • Kato, A. and Y. Ben-Zion (2021) The generation of large earthquakes. Nat Rev Earth Environ 2, 26–39 https://doi.org/10.1038/s43017-020-00108-w

 

How to cite: Zaliapin, I. and Ben-Zion, Y.: Spatio-temporal localization of seismicity in relation to large earthquakes, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9874, https://doi.org/10.5194/egusphere-egu23-9874, 2023.

Posters on site: Wed, 26 Apr, 16:15–18:00 | Hall X4

Chairpersons: Stefania Gentili, Álvaro González, Filippos Vallianatos
X4.72
|
EGU23-5021
Gutenberg-Richter, Omori and Cumulative Benioff strain patterns in view of Tsallis entropy and Beck-Cohen Superstatistics.
(withdrawn)
Filippos Vallianatos
X4.73
|
EGU23-5729
|
ECS
Piero Brondi, Stefania Gentili, and Rita Di Giovambattista

In most of the recent intense earthquakes in Italy, a strong subsequent event (SSE) of comparable or higher magnitude was observed. Its effects, in combination with the strong mainshock, may lead to the collapse of already weakened buildings and to a further increase in damage or even in the number of fatalities, with serious consequences for society. Therefore, the forecasting of an SSE is of strategic importance to reduce the seismic risk during the occurrence of a seismic sequence. To this end, we have recently developed the machine learning-based multi-parameter algorithm NESTORE (Next STrOng Related Earthquake). The first MATLAB version (NESTOREv1.0) was applied to Italian seismicity to forecast clusters where the difference between the magnitude of the mainshock Mm and that of the strongest aftershock is less than or equal to 1. These clusters are called type A by the NESTOREv1.0 software, while the other cases are called type B. NESTOREv1.0 is based on nine seismicity features that measure the number of events with M > Mm-2, their spatial distribution, magnitude, and energy trend over time in increasing time intervals following the occurrence of the mainshock. The software identifies seismic clusters above a threshold for mainshock magnitude Mth, finds appropriate thresholds for features to distinguish A and B cases in a training database, and uses them to provide an estimate of the probability that a cluster is of type A in a test set. For the application of NESTOREv1.0 to Italy, we considered both a national and a regional approach. In the first case, we analysed the seismicity recorded by the INGV network from 1980 to 2021, while in the second case we used the seismic catalogue of the dense OGS network in northeastern Italy for the period 1977-2021. In the nationwide application of NESTOREv1.0, we observed an area between Tuscany and Emilia-Romagna with anomalously high seismic activity concentrated in bursts of short duration. Since this area is almost exclusively populated by type B and therefore not suitable for a specific training procedure, we excluded it from the following analyses. In the remaining national area, we trained NESTOREv1.0 with clusters in the time period 1980-2009 (24 clusters) and tested it in the period 2010-2021 (14 clusters). For the regional case, we considered a rectangular area in northeastern Italy, where we could lower Mth due to the higher local density of seismic stations of the OGS seismic network compared to the mean density of the national network. In this area, 13 clusters from 1977 to 2009 were used as training set, and the performance of NESTOREv1.0 was evaluated using 18 clusters from 2010 to 2021. For both approaches, we obtained good results in terms of the rate of correct forecasting of cluster typology. In the 12 hours following the mainshock, the rate is 86% for the nationwide analysis and 89% for the regional analysis, respectively, which supports the application of NESTOREv1.0 on the Italian territory.

Funded by a grant from the Italian Ministry of Foreign Affairs and International Cooperation

How to cite: Brondi, P., Gentili, S., and Di Giovambattista, R.: Forecasting strong aftershocks in the Italian territory: a National and Regional application for NESTOREv1.0, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5729, https://doi.org/10.5194/egusphere-egu23-5729, 2023.

X4.74
|
EGU23-5738
Stefania Gentili, Eleni-Apostolia Anyfadi, Piero Brondi, and Filippos Vallianatos

It is widely known that large earthquakes are followed by aftershocks that can affect numerous facilities in a city and worsen the damage already suffered by vulnerable structures. In this study, we apply NESTORE machine learning algorithm to Greek seismicity to forecast the occurrence of a strong earthquake after a mainshock. The method is based on extracting features used for machine learning and analyzing them at increasing time intervals from the mainshock, to show the evolution of knowledge over time. The features describe the characteristics of seismicity during a cluster. NESTORE classifies clusters into two classes, type A or type B, depending on the magnitude of the strongest aftershock. To define a cluster, a window-based technique was applied, using Uhrhammer's (1986) law. We used the AUTH earthquake catalogue between 1995 and 2022 over a large area of Greece to analyze a sufficiently large number of clusters. The good overall performance of NESTORE in Greece evidenced the algorithm's ability to automatically adapt to the area under study. The best performance was obtained for a time interval of 6 hours after the main earthquake, which makes the method particularly attractive for application in the field of early warning, as it allows estimating the probability of a future hazardous earthquake occurring after a strong initial event.

 

Funded by a grant from the Italian Ministry of Foreign Affairs and International Cooperation and Co-funded by the Erasmus+ programme of the European Union (EU).

How to cite: Gentili, S., Anyfadi, E.-A., Brondi, P., and Vallianatos, F.: Forecasting Strong Subsequent Earthquakes in Greece Using NESTORE Machine Learning Algorithm, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5738, https://doi.org/10.5194/egusphere-egu23-5738, 2023.

X4.75
|
EGU23-11052
|
ECS
Jongwon Han, Keun Joo Seo, Ah-Hyun Byun, Seongryong Kim, Dong-Hoon Sheen, and Donghun Lee

Earthquake monitoring has been stepped up due to high-density permanent networks in the southern Korean Peninsula, though its relatively low seismicity. With the dramatic increase of data volume, deep learning techniques can be effective ways to process them. In this study, we present a preliminary, but comprehensive earthquake catalog in the southern Korean Peninsula for research purposes by applying a series of deep learning-incorporated methods including for earthquake and phase detection, event discrimination, and focal mechanism determination. We first improved the EQTransformer by re-training it with hybrid local and STEAD datasets to perform earthquake and phase detection for 10-year-long data from 2012 to 2021. Then, the subsequent phase association was carried out using the algorithm based on a Bayesian gaussian mixture model. In the result, 66,855 events were identified and located from 691,077 phase detections. Among them, 27,429 natural earthquakes were separated with a novel CNN model trained using event waveforms and origin time constraints. The natural seismicity suggested various earthquake clusters that constrained by tectonic structures, such as the Okcheon belt and Gyeongsang basin, and showed significantly low rate of occurrence in the Gyeonggi massif. In addition, we developed a CNN model for the determination of focal mechanisms that identify the polarity of initial P-waves in input waveforms, and the application of it resulted in 2,345 reliable solutions. Strike-slip motions were dominant in the inland, while reverse faulting of coastal earthquakes, showing an average P-axis direction of N74E in both areas. Despite the massive volume of data, it took less than a week to perform all of the processes with more cataloged earthquakes than those in the previous one (9,218). The extended earthquake catalog accompanied by focal mechanisms underpins data-driven studies such as tomography, stress field estimation, earthquake hazards assessment, and burial fault mapping.

How to cite: Han, J., Seo, K. J., Byun, A.-H., Kim, S., Sheen, D.-H., and Lee, D.: Preliminary earthquake catalog (2012–2021) of the southern Korean Peninsula by deep learning-based techniques, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11052, https://doi.org/10.5194/egusphere-egu23-11052, 2023.

X4.76
|
EGU23-6844
Nooshin Najafipour and Christian Sippl

As the number of seismic stations and experiments greatly increases due to ever greater availability of instrumentation, automated data processing becomes more and more necessary and important. Machine Learning (ML)methods are becoming widespread in seismology, with programsthat identify signals and patterns or extract features that can eventually improve our understanding of ongoing physical processes.

We here focus on critically evaluating the performance of a popular machine-learning-based seismic event detection and arrival time picking program, EQ-Transformer, forseismic data from the IPOC deployment in Northern Chile, using handpicked benchmark datasets.By using the open-source framework SeisBench, we can test the effect of using different pre-trained models as well as modify critical parameters such as probability thresholds.

By performing this evaluation, we want to decide whether it is necessary to retrain EQTransformer with local data, or if its performance with one of the supplied pre-trained sets is sufficiently good for our purposes.

We prepared alarge handpicked benchmark dataset for Northern Chile, which we use to find the optimal configuration of EQTransformer. For this benchmark dataset, we select a total of 35 days distributed throughout the 15 years covered by the IPOC deployment. Our goal was to pick all of the many small events in the dataset, even when they are only visible at one or two stations, with high accuracy. We found around three hundred events per day, which highlights the very high seismic activity of the region. We then ran EQTransformer for the same days, using a wide range of parameter choices and pre-trained models.

We need to find if our data is similar to seisbench benchmark dataset or if we should use our data to calibrate the EQTransformer for picking in subduction zones.

We use our handpicked benchmark dataset to evaluate the detection rate (missed events, false detections) as well as the picking accuracy (residuals to handpicks) achieved with EQTransformerin the various tested configurations. We present results of choosing different event detection thresholds, showingtrue positive rate vs. false positive rate plots in order to find optimal thresholds, and evaluate the pick accuracy of obtained arrival time picks by comparing to the handpicked benchmark. This comparison of picking times (P & S) is visualized with residual histograms. Lastly,we also show examples for a visual comparison of picks fromEQTransformer with manual picks.

The present study is the first step towards the design of an automated workflow that comprises event detection and phase picking, phase association and event location and will be used to evaluate subduction zone microseismicity in different locations.

How to cite: Najafipour, N. and Sippl, C.: Optimizing the performance of EQTransformer by parameter tuning and comparison to handpicked benchmark datasets in a subduction setting, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6844, https://doi.org/10.5194/egusphere-egu23-6844, 2023.

X4.77
|
EGU23-6806
Jorge Antonio Puente Huerta and Christian Sippl

Seismic phase association is a fundamental task for earthquake detection and location workflows, as it gathers individual seismic phases detected on multiple seismic stations and associates them to events.

Current phase picking algorithms are capable of generating large phase datasets, and together with new improved phase association algorithms, they can create larger and more complete earthquake catalogs when applied to dense seismic networks (permanent or temporary).

As part of project MILESTONE, which aims at the automatic creation of large microseismicity catalogs in subduction settings, the present study evaluates the performance of three different phase association algorithms, both by comparing their outputs with a handpicked benchmark dataset and by the retrieval of synthetic events.

For this purpose, we used seismic data from the IPOC (Integrated Plate boundary Observatory Chile) permanent deployment of broadband stations in Northern Chile.

We manually picked P and S phases of raw waveforms on randomly chosen days, with event rates in excess of 100-150 per day. All events that were visually recognizable were picked and located, leading to a dataset to be used as “ground truth”.

We do the phase picking with EQTransformer (Mousavi et al. 2020) and evaluate the performance of three seismic phase associators: 1) PhaseLink (Ross et al. 2019), a deep learning based approach trained on millions of synthetic sequences of P and S arrival times, 2) REAL (Zhang et al. 2020), that combines the advantages of pick-based and waveform-based methods, primarily through counting the number of P and S picks and secondarily from travel-time residuals, and 3) GaMMA (Zhu et al. 2021), an associator that treats the association problem as an unsupervised clustering problem in a probabilistic framework.

In the synthetic test we use NonLinLoc raytracer and add random noise, as well as false picks to simulate an automatic picker output.

In both experiments we evaluate the number of correctly associated and lost events, and the number of constituent picks per event.

How to cite: Puente Huerta, J. A. and Sippl, C.: Comparison of seismic phase association algorithms and their performance., EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6806, https://doi.org/10.5194/egusphere-egu23-6806, 2023.

X4.78
|
EGU23-7028
|
ECS
Pierre-Yves Raumer, Sara Bazin, Dorian Cazau, Vaibhav Vijay Ingale, Aude Lavayssière, and Jean-Yves Royer

Hydrophones arrays have proven to be an efficient and affordable method to monitor underwater soundscape, in particular magmatic and tectonic events. Indeed, thanks to the sound fixing and ranging (SOFAR) channel in the ocean, acoustic waves undergo a very low attenuation over distance and thus propagate further than they would do across the solid Earth. The MAHY array, composed of four autonomous hydrophones, has been deployed off Mayotte Island since October 2020. It contributes to monitor the recent volcanic activity around the island, and enabled to detect short and energetic acoustic events sometimes reffered to as impulsive events. As for their cause, it has been proposed that these signals are generated by water-lava interactions on the seafloor. So far, these events have been searched by visually inspecting the data, which is a cumbersome and somewhat observer-dependent task. To face these issues, we have developped an automatic picking algorithm tailored for these impulsive events. After some initial signal processing, a supervised neural network model was trained to detect such signals, which can be later checked by a human operator. Taking advantage of the genericity of this detection framework, we applied it to other hydroacoustic data sets (OHASISBIO and IMS-CTBT) to explore the feasibility of detecting T-wave generated by submarine earthquakes. The next step will be to improve the model with unsupervised or semi-supervised feature learning, in order to improve our metrics and, in the end, facilitate the study of specific acoustic signals.

How to cite: Raumer, P.-Y., Bazin, S., Cazau, D., Ingale, V. V., Lavayssière, A., and Royer, J.-Y.: Application of machine learning to hydro-acoustic seismic and magmaticevents detections, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-7028, https://doi.org/10.5194/egusphere-egu23-7028, 2023.

X4.79
|
EGU23-11042
|
ECS
Javad Kasravi and Jonas Folesky

One of the vital open questions in seismology is rapid, high quality phase identification and picking. Measurements of earthquake arrival time or phase picking are often done by expert judgment with many years of experience. Due to advances in technology and seismometer deployment, the amount of recorded data has increased dramatically in the previous decade, leading up to a point, where it has become almost impossible for humans to deal with this amount of data flow. Therefore, automatic picking algorithms are being used.  In recent years multiple machine learning algorithms have been introduced that bear the potential to combine both, high picking accuracy and the capability of processing large amounts of data. 
In this contribution, we demonstrate the performance of the EQTransformer autopicker, when applied to continuous seismic data from the Northern Chilean subduction zone. To test this deep neural network, we chose a random day and carefully hand picked the continuous data on 18 IPOC stations, selecting only combinations of picks which should lead to locatable events (e.g. with at least five picks). This results in the identification of  3040 P and 2310 S picks. We compare the results of two different training versions of EQTransformer with hand-picked data and with the IPOC seismicity catalog. As it turns out, the comparison is not straightforward, because the evaluation of the picks is highly complicated, given that the true number and type of phase arrivals is and remains unknown. However, the autopicker is able to detect most of the hand-picked phases and arrival times. It outperforms the IPOC catalog by a factor of about 10-15 and thusly it appears to be a valid alternative for advanced seismic catalog construction.

How to cite: Kasravi, J. and Folesky, J.: Benchmarking Study of EQTransformer Autopicker for Seismic Phase Identification in Northern Chile, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11042, https://doi.org/10.5194/egusphere-egu23-11042, 2023.

X4.80
|
EGU23-13927
Chun-Ming Huang, Li-Heng Chang, Hao Kuo-Chen, and YungYu Zhuang

Deep learning has greatly improved the efficiency of earthquake detection and phase picking tasks, as demonstrated by neural network models such as PhaseNet and EQTransformer. However, the code released by these authors is not production-ready software that can be easily integrated into our lab's workflow. To solve this problem, we developed "SeisBlue," a platform that brings all the necessary steps together in one place. It includes these major components: database client, data inspector, data converter, model trainer, model evaluator, and pick associator, and is designed to be modular and interchangeable to allow for easy experimentation with different combinations.

SeisBlue has been used in several major earthquake events in Taiwan, including the 918 Taitung earthquake (magnitude 6.9 Mw). In this event, we were able to capture over 1,200 events near real-time in just two days - a task that would have taken over a month to complete manually. The quickly-released earthquake catalog provided insight into the complex behavior of the blind fault.

How to cite: Huang, C.-M., Chang, L.-H., Kuo-Chen, H., and Zhuang, Y.: SeisBlue: a deep-learning data processing platform for seismology, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-13927, https://doi.org/10.5194/egusphere-egu23-13927, 2023.

Posters virtual: Wed, 26 Apr, 16:15–18:00 | vHall NH

Chairpersons: Stefania Gentili, Álvaro González, Rita Di Giovambattista
vNH.14
|
EGU23-11763
Parthena Paradisopoulou, Eleni Karagianni, Ioanna Karagianni, Areti Panou, Odysseus Galanis, Dominikos Vamvakaris, Vasileios Karakostas, Despina Kementzetzidou, and Eleftheria Papadimitriou

Intense and continuous seismicity in the last two years (October 2020 – September 2022) that took place in a rather small area near the city of Thiva is investigated here. The activity started with an M4.6 earthquake (2 of December 2020) followed by its own aftershock sequence. The activity migrated slightly to the west, with a persistent swarm in July 2021-September 2022 (largest magnitudes M4.3, on 11 July 2021, M4.0 on 2 September 2021 and M4.3, on 10 April 2022). Aiming to constrain the geometry and kinematics of the activated fault segments along with the spatiotemporal evolution of the seismic activity, processing of the recording of the regional seismological network was accomplished that includes the determination of the focal coordinates using the HYPODD software. The above information will provide a better understanding of the seismic sequence and seismic hazard in the region so that there is better prevention and preparation against a future strong earthquake. Aiming to study in detail the properties of this seismicity manifestation, the recordings of the Hellenic Unified Seismological Network (HUSN) are used to accurately determine the seismic parameters of earthquakes with magnitudes M≥1.5. Phases (P, S phases) are gathered from the Geophysical Department of the Aristotle University of Thessaloniki and the Geodynamics Institute of National Observatory of Athens from October 2020 to September 20212. Then, the bulletins were merged, and an initial earthquake catalogue was compiled containing ~6000 events. Earthquake relocation was initially performed using HYPOINVERSE software and all the available manually picked P and S phases. An appropriate local velocity model and the VP/VS ratio were necessary to defined. The Wadati method was applied to the dataset and the resulting VP/VS ratio equals to 1.76. The one-dimensional velocity model used for this study is calculating by the VELEST software. Time corrections relative to the crustal model were calculated considering the mean residual for each station. For the relocation of the events the calculated time delays were taken into account. To improve the obtained locations, we relocate the earthquakes using the double difference inversion algorithm, hypoDD with differential times derived from phase-picked data.

Τo define the stress regime in the area, the moment tensors of earthquakes with ML ≥ 3.5 were estimated using the ISOLA and FPFIT software. The fault plane solutions from the largest earthquakes of the seismic sequence have been used for Coulomb stress changes calculation. The stress field is calculated according to the focal mechanism of the next large event, whose triggering is inspected, so it can be checked if foreshocks contributed to the occurrence of the largest earthquakes of the sequence and the possible sites for future strong earthquakes can be assessed as well.

How to cite: Paradisopoulou, P., Karagianni, E., Karagianni, I., Panou, A., Galanis, O., Vamvakaris, D., Karakostas, V., Kementzetzidou, D., and Papadimitriou, E.: Cascading occurrence of moderate magnitude seismicity in the active fault system of Thiva (central Greece), EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11763, https://doi.org/10.5194/egusphere-egu23-11763, 2023.