基于主动学习的地震波阻抗标注实验分析  

Experimental analysis of seismic impedance labeling based on active learning

作　　者：彭真 许辉群[1] PENG Zhen;XU HuiQun(School of Geophysics and Petroleum Resources,Yangtze University,Wuhan 430100,China)

机构地区：[1]长江大学地球物理与石油资源学院,武汉430100

出　　处：《地球物理学进展》2023年第6期2565-2575,共11页Progress in Geophysics

基　　金：中国石油天然气股份有限公司勘探开发研究院地球物理重点实验室开放基金资助课题“考虑标签与网络适应性的高分辨率地震波阻抗反演技术”(2022-KFKT-25)资助。

摘　　要：现阶段的深度学习地震波阻抗反演大都是通过监督学习的方式来完成的,然而监督学习训练需要大量的标签数据,这在实际问题上是不现实的,并且会花费较大的人力和时间成本.为了在低成本下获取大量的地震波阻抗标签,采用主动学习中的机器学习算法对地震数据进行标注可获取数量大、成本相对较低的地震波阻抗标签.该方法的核心在于如何构建主动查询策略,不同的查询策略标注的结果会有所不同,并且在时间成本上也会有较大的差异.鉴于此,为了进一步探讨在不同查询策略下的地震波阻抗标注精度的关系,围绕不同查询策略下地震数据的标注为总目标,笔者设计了三组实验,采用4种不同的查询策略并利用机器学习算法对Marmousi-2数据集开展地震波阻抗标注的实验分析.经计算效率、召回率、皮尔逊系数、均方误差、决定系数评价指标的量化数据表明:基于最小置信度的查询方法标注效率和精度高,并将该方法推广到实际数据,且经实际工区中的验证井证实了方法的有效性.该研究可为高效获取大数据量的地震波阻抗标签提供新的思路,也为半监督学习的反演方法的应用奠定数据基础.Deep learning seismic impedance inversion is mostly completed by supervised learning nowadays.However,supervised learning training requires a large number of label data,which is unrealistic in practical problems and will cost a large amount of labor and time.In order to obtain a large number of seismic impedance labels at a low cost,machine learning algorithm in active learning is used to label seismic data to obtain a large number of low-cost seismic impedance labels.The core of this method lies in how to construct active query strategies,different query strategies will have different labeled results,and there will be great differences in time cost.In view of this,in order to further explore the relationship of seismic impedance labeling accuracy under different query strategies.Focusing on the overall goal of seismic data labeling under different query strategies,the author designs three sets of experiments by using four different query strategies and machine learning algorithm to carry out experimental analysis of seismic impedance labeling in Marmousi-2 data.By calculating efficiency,recall rate,Pearson coefficient,mean square error and determination coefficient evaluation index quantitative data show that:the query method based on Least confidence has high labeling efficiency and accuracy,and the method is extended to the real data,and the effectiveness of the method is verified by the verification wells in the real work area.This study provides a new idea for obtaining a large number of seismic impedance labels,and also lays a data foundation for the application of semi-supervised learning inversion method.

关 键 词：主动学习 地震数据的标注 查询策略 最小置信度 地震波阻抗 

分 类 号：P631[天文地球—地质矿产勘探]