基于机器学习的地表破裂自动识别和填图——以2021年青海玛多M_(W)7.4级地震为例  

Machine learning in identifying and mapping the surface rupture of the 2021 M_(W)7.4 Madoi earthquake, Qinghai

作　　者：曾宪阳 刘静 王伟[1] 姚文倩[2] 吴静 刘小利 韩龙飞[2] 王文鑫 邢宇堃 杜瑞林 杨绪前 ZENG XianYang;LIU Jing;WANG Wei;YAO WenQian;WU Jing;LIU XiaoLi;HAN LongFei;WANG WengXin;XING YuKun;DU RuiLin;YANG XuQian(State Key Laboratory of Earthquake Dynamics,Institute of Geology,China Earthquake Administration,Beijing 100029,China;Institude of Surface-Earth System Science,School of Earth System Science,Tianjin University,Tianjin 300072,China;Hangzhou Five Generation Communication and Big Data Research Institute,Hangzhou 310012,China;Key Laboratory of Earthquake Geodesy,Institute of Seismology,China Earthquake Administration,Wuhan 430071,China)

机构地区：[1]中国地震局地质研究所地震动力学国家重点实验室,北京100029 [2]天津大学表层地球系统科学研究院,地球系统科学学院,天津300072 [3]杭州五代通讯与大数据研究院,杭州310012 [4]中国地震局地震研究所地震大地测量重点实验室,武汉430071

出　　处：《地球物理学报》2023年第3期1098-1112,共15页Chinese Journal of Geophysics

基　　金：国家重点研发计划(2021YFC3000605-04);国家自然科学基金(U1839203,42030305);中国地震局地质研究所基本科研业务项目(IGCEA1812);中国地震局地震科技星火计划项目(XH22003C)联合资助。

摘　　要：大地震的同震地表破裂高分辨率填图对于理解破裂传播机制、量化地震灾害和地震危险性等至关重要;无人机航片和地形数据为地表破裂研究提供大量的高精度数据.同时基于海量数据的人工填图耗时费力,效率较低;机器学习(Machine learning)技术的发展为快速处理这类高分辨率图像数据提供了新思路.本文以2021年玛多M_(W)7.4级地震震后高精度无人机航片为基础数据,展示了机器学习技术快速、高效识别地表破裂的潜力.基于卷积神经网络Canny算法,详细讨论了无人机数字正射影像的处理流程和关键步骤,包括准备训练数据、训练和后处理.对比人工识别和机器识别的结果显示,本文所提出的方法可以有效地绘制地表破裂,为未来研究大地震地表破裂提供新思路.同时,展示了机器学习在地震地质、地表过程和地貌等定量研究中的巨大优势和广阔前景.High-resolution mapping of coseismic surface rupture of large earthquakes is very important for better understanding the behavior and mechanism of earthquake rupture and for quantifying earthquake hazards. High-resolution UAV imagery and topographic data provide a large volume of valuable images of the surface rupture. Manual mapping of fractures on many high-resolution images could be labor-intensive, time-consuming, and thus inefficient. Machine learning provides more possibilities for the rapid processing of such big-data images. In this paper, we demonstrate the potential of Machine learning techniques to rapid, efficient, and complete identification of fractures of the surface rupture zone using high-precision UAV images of the 2021 Madoi M_(W)7.4 earthquake. We applied the canny algorithm(based on Convolutional Neural Networks) to discuss the processing flow and key steps of UAV digital orthophoto in detail, including preparing training data, training, and post-processing. By comparing the interpretations of manual mapping and machine recognition, the proposed method can effectively map surface rupture, providing a tool for studying future large earthquakes. Machine learning has advantages and broad prospects in quantitative studies of earthquake geology, surface processes and geomorphology.

关 键 词：机器学习 遥感影像 卷积神经网络 2021年青海玛多地震 地震地表破裂 精细填图 

分 类 号：P316[天文地球—地震学]