Earthquake Detection And Seismic Tomography Using Machine Learning And Template Matching:Application To Sichuan And Yunnan Regions

With the wide deployment of dense seismic arrays and the rapid development of microseismic detection techniques,microseismicities are receiving extensive attention in the seismological community.In addition to hand pick-based conventional earthquake detection and location methods,the other two types of methods are being promoted more and more widely:(1)template matching-based microseismic detection and location methods;(2)machine learning-based microseismic detection and location methods.The former one is based on waveform cross-correlation technique,which is very sensitive to weak signals.A pair of earthquakes,occurring closely in space and sharing similar focal mechanisms,have similar propagation paths to the same station.As a result,their waveforms should be similar.Using the waveforms of known earthquakes as templates,small earthquakes can be detected through stacked crosscorrelograms between continuous waveforms and template waveforms at multiple stations.The advantage of this method is that it can detect weak seismic signals with signal-to-noise ratio lower than the noise level.Its disadvantage is that it cannot detect earthquakes that are not similar to templates,and time-consuming is a major concern.The latter one is based on machine learning techniques,which are sensitive to feature classification and extraction.First,a machine-learning phase picker is trained using the convolutional neural network method based on available manual phase picks.Then the well-trained picker is adopted to pick phase arrivals from raw continuous waveforms accurately and efficiently.Finally,absolute and relative location methods are utilized to obtain high-precision earthquake catalogs.The advantage of this method is that it uses a well-trained machine-learning model to pick arrival times from raw continuous data,sequentially builds a high-precision earthquake catalog without additional prior information,and the calculation efficiency is very high.The disadvantage is that the lower limit of magnitude detected by machine learning-based methods is usually not as low as that from template-based methods.The best strategy is to combine the two methods:First,earthquake catalog can be obtained by the machine-learning method;second,using these cataloged events as tempaltes template matching methods detect more smaller events to improve the magnitude of completeness.Although small earthquakes are small in magnitude,their recurrence period is short and their occurrence frequency is high.As a result,they can reveal more detailed distribution characteristics of seismic activities in time and space.Small earthquakes can be used to describe high-precision three-dimensional fault geometry,study the spatiotemporal distribution characteristics and triggering mechanism of aftershock sequences and earthquake swarms,explore foreshock activity and earthquake nucleation mechanism,detect repetitive seismic activity to gain insight into fault creep characteristics,explore the relationship between induced earthquakes and human activities,study the relationship between remote dynamic triggering phenomenon and stress state,and detect tremors to understand the characteristics of deep fault activity,etc.In our study,we use template-based and machine-learning-based methods to detect and locate earthquakes in the Sichuan and Yunnan regions,China,respectively.Based on high precision earthquake location results,we discusse the triggering mechanism,nucleation process,fault geometry,and seismogenic mechanism of earthquakes in the study regions.The research work and corresponding conclusions are as follows:We use the Match and Locate method to study the 2018 ML 4.0 Shimian intraplate earthquake and its foreshocks in the northern segment of the Anninghe fault zone,and discuss its triggering mechanism and nuclear process.We detect a total of 1864 earthquakes 30 days before and 44 days after the mainshock,approximately three times more events than that reported in the routine catalog.The distribution of the earthquake sequence suggests that the seismogenic fault is a blind strike-slip fault in the east of the Anninghe fault.Forty-one foreshocks are detected in the 4 hours before the mainshock,and do not show an accelerating pattern leading up to the mainshock.The b value of the foreshocks is smaller than that of the aftershocks.The inversion results of focal mechanisms reveal that focal mechanisms of the foreshocks and the mainshock are consistent,indicating that the mainshock and the foreshocks occurred on the same fault plane.The high precision HypoDD relocations show that the mainshock and foreshocks are mainly clustered within a compact volume of 300 m×100 m×400 m.Most foreshocks rupture adjacent fault patches with little or partial overlap,which corresponds to cascade stress triggering from foreshocks to foreshocks to the mainshock.We use machine learning-based earthquake detection and location workflow(LOC-FLOW)and the double-difference seismic tomography method(TomoDD)to study the fault geometry and seismogenic mechanism of the middle-northern segment of the Xiaojiang fault zone.The Xiaojiang fault zone is an important tectonic boundary between the Sichuan-Yunnan diamond block and the South China block.Almost the entire fault zone has been ruptured by paleoearthquakes in history.Due to the lack of dense stations,the seismogenic environment in the middle-northern segment of the fault zone is not well understood.Based on the recently deployed Qiaojia dense seismic array in the middle-northern segment of the Xiaojiang fault,we utilize the recently developed machine-learning-based LOC-FLOW method to analyze one-year continuous seismic waveform data.First,P/S arrivals are picked up by the machine-learning phase picker Phasenet,and then a series of absolute and relative earthquake location methods are used to construct a high-precision earthquake catalog.The seismic distribution can reveal the detailed characteristics of the main faults and blind faults,and also identify two seismic gaps that may have the potential of hosting future moderate-large earthquakes.Based on machine-learning travel times and cross-correlation differential travel times,we further use the double-difference seismic tomography method to simultaneously invert earthquake locations and high-resolution velocity structures of the study area.The velocity structures show strong lateral heterogeneities.Most earthquakes are distributed on the boundary of high-velocity and low-velocity regions or within high-velocity regions above low-velocity bodies.This study demonstrates that it is practical to use LOC-FLOW and TomoDD to rapidly construct high-precision earthquake catalog and high-resolution velocity structures from raw continuous waveform data.