| 基于深度学习到时拾取自动构建长宁地震前震目录 |
| |
| 引用本文： | 赵明,唐淋,陈石,苏金蓉,张淼.基于深度学习到时拾取自动构建长宁地震前震目录[J].地球物理学报,2021,64(1):54-66. |
| |
| 作者姓名： | 赵明 唐淋 陈石 苏金蓉 张淼 |
| |
| 作者单位： | 中国地震局地球物理研究所,北京 100081;北京白家疃国家地球科学野外观测研究站,北京100095;四川省地震局,成都610041;中国地震局地球物理研究所,北京 100081;北京白家疃国家地球科学野外观测研究站,北京100095;四川省地震局,成都610041;加拿大戴尔豪斯大学地球和环境科学系,新斯科舍省哈利法克斯B3H 4R2 |
| |
| 基金项目： | 国家自然科学基金青年基金（41804047）、中国地震局地球物理研究所基本科研业务专项（DQJB19A0114）资助. |
| |
| 摘 要： | 将深度学习到时拾取、震相关联技术与传统定位方法联系起来,构建一套连续波形自动化处理与地震目录自动构建流程,对于高效充分利用地震资料,提升微震检测能力具有十分重要的意义.我们应用最新发展的迁移学习震相识别技术、震相自动关联技术,对长宁M S6.0地震震中附近21个台站震前半个月(6月1日—6月17日)的连续记录波形进行P、S震相识别、震相自动关联和初步定位,并应用传统绝对定位和相对定位技术得到了长宁地震震前微震活动的绝对和相对定位目录.其中绝对定位目录能在较小的误差范围匹配85%的人工处理目录,其发震时刻平均误差为0.36±0.07 s,震级平均误差为0.15±0.024级,水平定位平均误差为1.45±0.028 km,其识别的1.0级以下微震数目是人工的8倍以上,将长宁地震震前微震目录的检测下限提升至M L-1左右,证明了基于深度学习到时识取和REAL(Rapid Earthquake Association and Location,快速震相关联和定位技术)震相自动关联来构建微震目录具有较好的实用性.我们的自动地震目录揭示了长宁M S6.0主震所发生的区域震前异常频繁的微震活动,以及与区域内盐矿注水井的关联性,更好地描绘了这些微震活动的时空演化特征,其空间活动性分布特征与长宁M S6.0余震序列的分布一致.
|
| 关 键 词： | 深度学习 前震 长宁地震 绝对定位 相对定位 |
| 收稿时间： | 2020-11-09 |
Machine learning based automatic foreshock catalog building for the 2019 MS6.0 Changning,Sichuan earthquake |
| |
| ZHAO Ming,TANG Lin,CHEN Shi,SU JinRong,ZHANG Miao.Machine learning based automatic foreshock catalog building for the 2019 MS6.0 Changning,Sichuan earthquake[J].Chinese Journal of Geophysics,2021,64(1):54-66. |
| |
| Authors: | ZHAO Ming TANG Lin CHEN Shi SU JinRong ZHANG Miao |
| |
| Institution: | 1. Institute of Geophysics, China Earthquake Administration, Beijing 100081, China;2. Beijing Baijiatuan Earth Science National Observation and Research Station, Beijing 100095, China;3. Sichuan Earthquake Administration, Chengdu 610041, China;4. Department of Earth and Environmental Sciences, Dalhousie University, Halifax, NS B3H 4R2, Canada |
| |
| Abstract: | We establish an automatic workflow for cataloging foreshocks of the 2019 MS6.0 Changning, Sichuan earthquake from continuous raw seismic data by sequentially applying the transfer learning phase picking, automatic phase association, and traditional earthquake location methods. In the workflow, we first pick seismic P and S picks using a deep-neural-network-based phase picker (PhaseNet). By applying the Rapid Earthquake Association and Location (REAL) method, seismic phases are associated into particular earthquakes, whose locations are preliminarily determined at the same time. Then we further sequentially improve earthquake absolute and relative locations using the VELEST and HYPODD software. We apply this workflow to study foreshocks of the 2019 MS6.0 Changning, Sichuan earthquake from June 1 to June 17, 2019. We automatically pick seismic picks from 28 GB continuous waveforms recorded at 21 three-component seismic stations and associate 4358 P phases and 5022 S phases into 1067 earthquakes. We apply the VELEST program to refine their locations, resulting in 870 events after applying critical event selection thresholds based on station gap and travel time residual. At last, 416 earthquakes are relatively relocated by the HYPODD software. Compared with 101 cataloged earthquakes from the Sichuan Earthquake Administration, our absolute VELEST catalog and relative HYPODD catalog can match 85% and 74% of the routine catalog, respectively. The average origin time difference of common events between the absolute VELEST catalog and the routine catalog is 0.36±0.07 seconds, the average magnitude difference is ML0.15±0.024, and the horizontal location error is 1.45±0.028 kilometers. We recover all ML>1.5 cataloged earthquakes with consistent location, except for one ML2.0 event, which was located with large error in the routine catalog. The number of events with ML<1.0 in our VELEST catalog is more than 8 times than that in the routine catalog. Our earthquake catalog shows that plenty of small foreshocks occurred close to a nearby water injection well for salt mining and the epicenter of the Changning MS6.0 mainshock. Thus, we suggest the MS6.0 Changning mainshock and its foreshocks are most likely related to salt mining in the region, which is supported by the recent InSAR observations as well. This study demonstrates that our workflow not only enables us to automatically and efficiently detect, locate and associate foreshocks of the MS6.0 Changning earthquake, but also significantly improves the completeness magnitude and extends the detection limit to ML-1. |
| |
| Keywords: | Deep learning Foreshock Changning earthquake VELEST catalog HYPODD catalog |
| 本文献已被 CNKI 维普 万方数据 等数据库收录！ |
| 点击此处可从《地球物理学报》浏览原始摘要信息 |
| 点击此处可从《地球物理学报》下载免费的PDF全文 |
|
