摘要
机器学习算法在地球物理领域的应用越来越广泛、深入。在地震资料解释中,目前主要利用实际或人工合成的断层样本,训练浅层卷积神经网络识别断层。实际断层样本需要人工标记,消耗大量时间成本;人工合成的断层样本虽然容易获得,但训练出的网络在应用于实际地震数据时效果不佳。为此,将深度残差网络与迁移学习结合并应用于断层识别。首先构建性能更优秀的深度残差网络训练人工合成的断层样本,然后使用少量实际断层样本进行迁移学习,增强网络的泛化能力,优化网络的识别结果。迁移学习后的网络能够有效提高实际断层的识别准确率,实际地震数据验证了该方法的可行性和有效性。
The application of machine learning algorithms in the field of geophysics has been expanded and deepened.In fault recognition on seismic data,the main approach is training a shallow convolutional neural network to achieve fault recognition using actual or synthetical fault samples.Actual fault samples require manual marking,which is very time-consuming.Synthetic fault samples are easy to obtain,but the effect of the trained network model is inadequate when applied to actual seismic data.For this reason,this paper combines a deep residual network with transfer learning to fault recognition.First train synthetical fault samples by constructing a deep residual network with better performance,then use a small number of actual fault samples for transfer learning.This way the generalization ability of the network can be enhanced,and the recognition results can be optimized.After transfer learning,the network can more effectively improve the recognition accuracy of actual faults than ever before.Actual seismic data have proved the feasibility and effectiveness of the method.
作者
张政
严哲
顾汉明
ZHANG Zheng;YAN Zhe;GU Hanming(Institute of Geophysics and Geomatics,China University of Geosciences,Wuhan,Hubei 430074,China)
出处
《石油地球物理勘探》
EI
CSCD
北大核心
2020年第5期950-956,929,共8页
Oil Geophysical Prospecting
基金
国家自然科学基金项目“基于地震属性的地震资料自动解释方法研究”(41574115)资助
关键词
地震资料解释
断层识别
深度残差网络
迁移学习
网络结构优化
seismic data interpretation
fault recognition
deep residual network
transfer learning
network structure optimization
作者简介
严哲,湖北省武汉市洪山区鲁磨路388号中国地质大学(武汉)地球物理与空间信息学院,430074。Email:yanzhe@cug.edu.cn;张政,硕士研究生,1994年生,2018年获中国地质大学(武汉)勘查技术与工程专业学士学位,现在中国地质大学(武汉)地球物理与空间信息学院攻读地质工程专业硕士学位,主要从事地震资料解释方面的学习和研究。
