结合台阵策略的震相拾取深度学习方法  

An array-assisted deep learning approach to seismic phase-picking

作　　者：刘芳[1] 蒋一然 宁杰远[2] 张建中[1] 赵艳红 Fang Liu;Yiran Jiang;Jieyuan Ning;Jianzhong Zhang;Yanhong Zhao(Inner Mongolia Autonomous Region Earthquake Administration,Hohhot 010010,China;School of Earth and Space Sciences,Peking University,Beijing 100871,China)

机构地区：[1]内蒙古自治区地震局,呼和浩特010010 [2]北京大学地球与空间科学学院,北京100871

出　　处：《科学通报》2020年第11期1016-1026,共11页Chinese Science Bulletin

基　　金：中国地震局地震科技星火计划(XH18012);内蒙古自治区科技重大专项和中国地震科学实验场专项(2019CSES35)资助。

摘　　要：深度学习的突破性发展及其在地震学领域的初步应用,为有效处理和利用地震资料提供了可能.震相拾取是地震资料处理中的基础性工作,目前已经提出了很多基于深度学习的震相拾取方法,取得了很好的效果.但是,为了满足地震学研究中处理连续地震记录的需要,尚需对方法进行进一步的改进.本研究针对性地设计了结合台阵策略,单独识别P波和S波的长时窗震相拾取深度学习模型PP(phase picker)及其训练方式,提出了具有实用性的震相拾取方法APP(array-assisted phase picker).利用阿里余震AI捕捉大赛和Hi-net数据进行测试的结果表明,模型能够有效地在连续波形上拾取体波震相并具有很好的泛化能力.通过比较该模型与其他模型(较短时窗的模型和同时识别P波、S波的模型)的拾取效果,验证了模型设计的合理性.具体的测试样例显示,该方法能够正确地处理地震密集的波形数据并能避免典型噪声的影响.将该方法运用到内蒙古地区台网的观测数据中,检测到了人工目录中98.1%的地震,地震拾取总数为人工目录数的30倍,进一步表明本研究方法具有很好的实用性.Rapid increase of seismic data poses challenges to seismic data processing,in which phase-picking is a crucial procedure and accordingly attracts attention from seismologists.Among the breakthroughs in deep learning methods for seismic data processing,seismologists have especially proposed many seismic phase-picking methods based on deep learning,which have shown much better accuracy than traditional automatic phase-picking methods.However,further improvements are needed to meet the requirements for processing continuous seismic records.Here we propose a new approach combining seismic array scheme and U-Net,an end-to-end neural network developed in image segmentation.Previous studies have shown that U-Net can classify waveforms and determine the onset of seismic phases accurately with the help of contracting and expanding paths.Based on U-Net,we first design a deep learning phase picker(PP)model with a long(40 s)time window.To avoid difficulties in time window selection and interference between P-and S-phase identifications,the model involves two networks with the same structure,which pick the P and S phases separately.To improve model training,we employ data transformation to expand the sample set,add typical noises to the sample set for better resemblance of real data,and slightly increase the weight of the seismic signal containing seismic records for a better balance between precision and recall.We also adopt a more effective array-assisted phase picker(APP)model which correlates the picking results from different stations to minimize misidentification.Using data from Hi-net and Alibaba Aftershock AI Capture Contest,we design training and test sets of different sizes to examine the performance and generalization ability of the model.Results show that the APP can effectively pick the body-wave phases on continuous records with a good generalization ability.The relationship between the model performance and training set is such that the model performance is directly related to the data quality rather than the s

关 键 词：深度学习 震相拾取 台阵策略 U型网络 

分 类 号：P315[天文地球—地震学]