基于生成对抗网络的地磁数据重建  

Geomagnetic data reconstruction based on generative adversarial network

作　　者：于晓彤 李夕海 曾小牛 刘继昊 谭笑枫 YU XiaoTong;LI XiHai;ZENG XiaoNiu;LIU JiHao;TAN XiaoFeng(Rocket Force University of Engineering,Xi'an 710025,China)

机构地区：[1]火箭军工程大学,西安710025

出　　处：《地球物理学进展》2022年第3期989-997,共9页Progress in Geophysics

基　　金：国家自然科学基金(41804136)资助。

摘　　要：实测地磁数据常因测区不规则导致数据存在空缺.针对地磁数据大面积缺失情况下,现有地磁数据重建方法精度不高的问题,本文提出了一种基于生成对抗网络的地磁数据重建方法.首先,利用生成对抗网络的博弈特性,在深度学习训练中不断优化地磁数据的生成网络和鉴别网络,得到最优的地磁数据生成模型,最后利用该模型可以实现对中心区域存在空缺的地磁数据进行重建.理论模型和实测地磁数据的对比实验结果表明,本文方法的重建精度高于经典的Kriging法、最小曲率法和基于凸集投影的方法;尤其在数据的结构性特征缺失较多的情况下,本文方法的地磁数据缺失部分重建结果的空间分布结构更加合理.Actually measured geomagnetic data is often vacant due to irregularities in the measurement area.Aiming at the problem of low accuracy of existing geomagnetic data reconstruction methods when geomagnetic data is missing in a large area,this paper proposes a geomagnetic data reconstruction method based on generative countermeasure network.First,using the game characteristics of the generative confrontation network,based on the constructed geomagnetic data set,continuously optimize the geomagnetic data generation network and the identification network in the deep learning training,and obtain the optimal geomagnetic data generation model,and finally use this model to achieve the vacant geomagnetic data in the central area is reconstructed.The experimental results of the comparison between the theoretical model and the measured geomagnetic data show that the reconstruction accuracy of this method is better than the classical Kriging method,the minimum curvature method and the method based on convex set projection;especially when the structural features of the data are missing,the spatial distribution structure of the reconstruction result of the missing part of the geomagnetic data of the method is more reasonable.

关 键 词：地磁数据 重建 深度学习 生成对抗网络 

分 类 号：P318[天文地球—固体地球物理学]