基于RFD-U-Net的探地雷达缺失数据重构  

GPR missing data reconstruction based on RFD-U-Net

作　　者：胡磊磊 邹仁辉 杨富强 莫亚军 曾宣源 区小毅 HU LeiLei;ZOU RenHui;YANG FuQiang;MO YaJun;ZENG XuanYuan;OU XiaoYi(Geophysical Survey Instituteof Guangxi Zhuang Autonomous Region,Liuzhou 545005,China;Guangxi Zhuang Autonomous Region Geological Environment Monitoring Station,Nanning 530029,China)

机构地区：[1]广西壮族自治区地球物理勘察院,柳州545005 [2]广西壮族自治区地质环境监测站,南宁530029

出　　处：《地球物理学进展》2023年第4期1775-1786,共12页Progress in Geophysics

摘　　要：在探地雷达勘探中,由于受到采样点数与测量速度不匹配、地表不均匀或仪器内外部干扰等因素的影响,容易造成信息不完整,甚至缺失.如何有效重构缺失信号是提高原始GPR数据精度和进行后续高分辨率成像的关键环节.本文提出了一种残差特征提取网络(RFD-U-Net)来针对性处理数据缺失这一问题.其中,RFD采用信息蒸馏网络的方法对缺失数据进行重构,首先使用多个特征提取连接来进行特征学习,然后在残差单元上增加跳跃映射,补充卷积过程中损失的特征信息,最后与轻量网络U-Net相结合,极大程度地缩减了网络冗杂.合成测试数据的对比结果及UQI和SNR两种评价指标的量化分析结果可知,RFD-U-Net获得了最优的重构效果和最佳的计算数值,验证了本文所提方法重构缺失GPR剖面的准确度和优越性.将RFD-U-Net应用于实测数据中,能够较好地重构缺失数据,在细节上突出了有效信号的连续性,并采用自动聚焦技术这一评价指标进行佐证,证明了该方法的实用性.In Ground-Penetrating Radar(GPR)exploration,various factors such as mismatch between sampling points and measurement speed,surface irregularities,and internal/external interferences in the instruments can lead to incomplete or even missing information.Effectively reconstructing missing signals is a crucial step for improving the accuracy of raw GPR data and achieving subsequent high-resolution imaging.In this paper,we propose a Residual Feature Distillation U-Net(RFD-U-Net)to address the issue of missing data.Specifically,RFD utilizes the information distillation network approach to reconstruct the missing data.It first employs multiple feature extraction connections for feature learning and then incorporates skip connections in residual units to complement the lost feature information during the convolution process.Finally,it combines with the lightweight U-Net network,significantly reducing network redundancy.Comparative results of synthetic test data and quantitative analysis using two evaluation metrics,namely UQI and SNR,demonstrate that RFD-U-Net achieves the optimal reconstruction performance and the best computational values,validating the accuracy and superiority of the proposed method in reconstructing missing GPR profiles.When applied to real measured data,RFD-U-Net effectively reconstructs the missing data,highlighting the continuity of valid signals in the details.Furthermore,the application of the automatic focusing technique as an evaluation metric provides evidence for the practicality of this method.

关 键 词：探地雷达 深度学习 自动聚焦技术 数据重构 

分 类 号：P631[天文地球—地质矿产勘探]