基于改进残差网络的大地电磁反演研究  

Magnetotelluric inversion based on an improved residual network

作　　者：李思平 刘彩云[2] 熊杰[1] 田慧潇 王方 LI Si-Ping;LIU Cai-Yun;XIONG Jie;TIAN Hui-Xiao;WANG Fang(School of Electronics Information,Yangtze University,Jingzhou434023,China;School of Information and Mathematics,Yangtze University,Jingzhou434023,China)

机构地区：[1]长江大学电子信息学院,湖北荆州434023 [2]长江大学信息与数学学院,湖北荆州434023

出　　处：《物探与化探》2023年第6期1508-1518,共11页Geophysical and Geochemical Exploration

基　　金：国家自然科学基金项目(62273060);长江大学大学生创新创业项目(Yz2022055)。

摘　　要：针对传统反演方法存在依赖初始模型、反演时间较长等问题,提出一种基于改进残差网络的大地电磁反演方法。该方法首先构造不同形状和不同电阻率的地电模型,在TM模式下正演得到视电阻率数据,组成数据集;然后在经典的残差网络ResNet基础上进行改进得到一种新的反演网络iResNet(improved residual network),并使用上述数据集训练该网络;最后将视电阻率数据输入到训练好的网络中,直接得到反演结果。实验结果表明,该方法能快速、准确地反演出地电模型的位置、形态和电阻率值,具有较好的泛化能力和抗噪能力,并能有效解决大地电磁实测数据问题。Traditional inversion techniques rely on initial models and exhibit prolonged inversion times.This study proposed a magneto-telluric inversion method based on an improved residual network.Specifically,geoelectric models of varying shapes and resistivity val-ues were established,and apparent resistivity data were obtained using the TM mode,forming a dataset.Then,a novel inversion net-work-iResNet(an improved residual network)-was established by improving classic residual network ResNet,and the new network was trained using the afore-mentioned data set.Finally,the apparent resistivity data were input to trained network,directly producing in-version results.The experimental results demonstrate that the method proposed in this study can accurately determine the positions,shapes,and resistivity values of the geoelectric models through swift inversion,suggesting high generalization and anti-noise capabili-ties.Therefore,this method can effectively deermine measured magnetotelluric data.

关 键 词：残差网络 大地电磁 反演 

分 类 号：P631.1[天文地球—地质矿产勘探]