奚先^1,,
黄江清²
1. 中国地质大学(武汉)数学与物理学院, 武汉 430074
2. 中国地质大学(武汉), 武汉 430074

基金项目: 国家自然科学基金项目（41630643，11702259，61601417）联合资助


详细信息

作者简介: 奚先, 男, 1964年生, 博士, 教授, 研究方向为随机介质和波动方程的数学模拟.E-mail:3080117816@qq.com

中图分类号: P631

收稿日期:2018-08-13
修回日期:2019-05-10
上线日期:2020-02-05

Location and imaging of scatterers in seismic migration profiles based on convolution neural network

XI Xian^1,,
HUANG JiangQing²
1. China University of Geosciences(Wuhan), School of Mathematics and Physics, Wuhan 430074, China
2. China University of Geosciences(Wuhan), Wuhan 430074, China


MSC: P631

--> Received Date: 13 August 2018
Revised Date: 10 May 2019
Available Online: 05 February 2020


摘要
摘要:本文提出了一种基于深度学习卷积神经网络（CNN）的全波形反演方法，可对地震散射波场中的散射体进行成像和定位.本文的灵感来自如下猜想：在散射波场剖面上的每个点附近的局部波场与该点到各散射体之间的最小距离有关系，并且这个关系可以被CNN网络所识别.我们将该最小距离定义为散射距离场，并将散射距离场的类别（即大小等级）作为CNN网络的预期输出，而输入就是该点附近的局部波场.最后用上述CNN网络对散射波场进行逐点训练和识别.计算结果证实了我们的灵感猜想，即上述CNN网络能够在复杂散射波场中对散射体进行成像.只通过一个训练模型的学习，CNN网络即可反演多种散射模型的偏移剖面，最后得到"类别函数预测值"和"滤波剖面"两种成像结果，由此可以辨识出在复杂的偏移剖面中各散射体的位置.
关键词: 散射波场反演成像/
卷积神经网络/
散射距离场/
类别函数

Abstract:In this paper, a full waveform inversion method based on deep learning Convolution Neural Network (CNN) is proposed, which can image and locate scatterers in the seismic scattered wave field. The inspiration of this paper comes from the following conjecture:the local wave field near each point in a scattered wave field profile is related to the minimum distance between the point and every scatterer, and this relationship can be identified by CNN network. We define the minimum distance as the scattering distance field, and take the category of the scattering distance field (i.e. the size level) as the expected output of the CNN network, with input as the local wave field near the point. Finally, the scattered wave field is trained and identified point by point using the above CNN network. The calculation results confirm our speculation that the above CNN network can image scatterers in complex scattering wave fields. By learning only one training model, the migration profiles of various scattering models can be inverted by CNN network. Finally, two imaging results of "category function prediction value" and "filter profile" care obtained. Thus, the positions of scatterers in complex migration profiles can be identified.
Key words:Scattered wave field inversion imaging/
Convolution Neural Network (CNN)/
Scattering distance field/
Category function



PDF全文下载地址:

http://www.geophy.cn/data/article/export-pdf?id=dqwlxb_15349