Machine-learning Based Earthquake Detection,Location,Phase Picking And Polarity Determination

Earthquake detection,phase picking,and earthquake location are classical and important processes in seismology.Realizing these processes quickly and accurately is of great significance for tomography based on travel time,earthquake early warning,and related works.The source focal mechanism of an earthquake provides information on fault rupture and regional stress state,which helps understand geological structures and seismic activity.Reliable focal mechanisms of small earthquakes can provide us with a more comprehensive understanding of the process of earthquake occurrence.Seismology is a data-rich and data-driven science.With the increasing global seismic data,it is more and more difficult for analysts to comprehensively analyze and process seismic data manually.Intelligent seismic data analysis and processing can effectively solve this problem.In recent years,deep neural networks have made rapid progress in feature learning and achieved remarkable performance in many seismic data processing and analysis tasks,bringing new opportunities and challenges to the field of seismic data analysis.Currently,there are earthquake detection methods based on single-station and multi-station approaches.These methods directly use waveforms or their timefrequency characteristics as network inputs but neglect the most basic visual shape features of waveforms.Researchers are still trying to use every detail available in seismic data as input for earthquake event detection and other applications,which may affect the accuracy of detection and the generalizability of methods in different regions.In this study,we adopt the idea of image segmentation from computer science,using the overall shape of waveforms instead of detailed waveform vibrations to detect earthquake events.We use two independent neural networks to implement earthquake detection,which is consistent with the process of target recognition and improves the ability of small earthquake detection algorithmically.In the Japanese testing dataset,the precision and recall rates of this method for event data are 98.73%and 96.54%,respectively,and the precision and recall for noise data are 97.41%and 99.05%,respectively.The results demonstrate that the proposed method performs well in different signal-to-noise ratios(SNR)and different filtering frequency bands.In addition to most events recorded in the earthquake catalog,this method also detects more events from continuous data that are not recorded in the earthquake catalog.After detecting earthquake events,it is necessary to locate the events.Most of the new earthquake events detected by machine learning are small or low SNR events,and many traditional earthquake location methods are no longer suitable for these events.In this study,we propose a single-station location algorithm based on neural networks,which determines the epicentral distance and back-azimuth information to achieve earthquake location without using phase information.For the Sichuan testing dataset,the mean absolute error(MAE)of epicentral distance is 3 km;the MAE of backazimuth is 22 degrees.The results of the single-station location algorithm in the Yunnan region show that it can reflect the overall distribution of earthquakes.In addition,this method fully utilizes the station network when events are received by multiple stations.Based on the single-station location algorithm,a small earthquake location system based on a sparse station network is constructed.This system uses sparse network constraints to locate small earthquakes and improve the accuracy of earthquake location as much as possible.This location system is applied to the Sichuan station network.Compared with the Sichuan earthquake catalog,we can obtain almost small earthquake events(0-1M)from the earthquake catalog.In addition,the results show that this method can also locate many small events that were previously missed.Additionally,considering the fusion of seismic station information into machine learning for earthquake location,most current machine learning methods are difficult to use station information as constraints,even if the station location is used,the physical process is not clear.In this study,an earthquake location method that can utilize station information is proposed based on the 3D U-Net network model.To improve the generalizability of the network model,data augmentation is applied,including data shifting,data selection,data rotation,and data fusion.The prediction results are evaluated using the model with the testing dataset from Oklahoma and the average location error is 5 km.This study also tests the location results of time-shifted data,selected data,rotated data,and multi-event data,with average location errors of 4.6 km,8.8 km,5.6 km,6.7 km,and 5.6 km,respectively.This shows that this method can be applied to earthquake locations with an arbitrary number of stations and any location of stations.In addition,this study applies the network model to Sichuan in China and Southern California through transfer learning,which also proves the generalizability of this method and can be applied in other regions.The focal mechanism of an earthquake reflects the state of underground stress and fault activity and plays a key role in the analysis,evaluation,and prediction of geological disasters.Waveform data are commonly used to invert the focal mechanism solution,and these methods have strong robustness for processing large-scale earthquake data.However,for small earthquakes,the focal mechanism solution is usually determined by the polarity of the P-wave first motion.In this study,a novel multi-task neural network structure is designed,incorporating an attention mechanism to simultaneously pick P-wave arrivals and determine first-motion polarities.The network model considers the relationship between phase picking and polarity determination after earthquake detection.We train and validate the model using data from the Southern California Seismic Network(SCSN).The Attention mechanism based neural network for Picks and Polarity(APP)model is used to obtain the phase arrivals and first-motion polarity information and the focal mechanism solutions in Southern California,the phase arrivals and polarity information in Japan,and the focal mechanism solutions in Oklahoma.In the Southern California and Japan datasets,the picking errors are 0.05 s and 0.04 s,respectively,the polarity determination accuracies are 98%and 99%,respectively,and the polarity determination recalls are 87%and 80%,respectively.The focal mechanism solutions obtained by inverting the polarity information predicted by the APP model are in good agreement with those provided in Oklahoma.These results indicate that the network model is reliable for application in different regions.This study is based on machine learning to study the basic work in earthquake data processing,including earthquake detection,earthquake location,phase picking,and first-motion polarity determination.We propose an earthquake detection network model based on the idea of image segmentation,a single-station location method based on feature fusion neural networks,a sparse station network location system,a multi-station location model based on 3D U-Net,and an attention-based multi-task network for phase picking and polarity determination.These methods improve the detection,location,and seismic source analysis capabilities of small seismic events,providing data foundations for our understanding of geological structures,regional stress states,earthquake early warning,and earthquake prediction,among others.