地球物理学进展 ›› 2020, Vol. 35 ›› Issue (4): 1497-1506.doi: 10.6038/pg2020DD0525

• 应用地球物理学Ⅰ(油气及金属矿产地球物理勘探) • 上一篇    下一篇

基于深度卷积神经网络的地震数据重建

杨冠雨1,2(), 王璐1,2, 孟凡顺1,2   

  1. 1.中国海洋大学海洋地球科学学院,青岛 266100
    2.海底科学与探测技术教育部重点实验室,青岛 266100
  • 收稿日期:2019-12-17 修回日期:2020-04-27 出版日期:2020-08-20 发布日期:2020-09-02
  • 作者简介:杨冠雨,男,1990年生,中国海洋大学海洋地球科学学院在读博士研究生,主要从事地震数据处理和反演计算的研究.E-mail:yangguanyuok@126.com
  • 基金资助:
    国家科技重大专项(2016ZX05027-002)

Seismic data reconstruction via deep convolution neural network

YANG Guan-yu1,2(), WANG Lu1,2, MENG Fan-shun1,2   

  1. 1. College of Marine Geosciences, Ocean University of China, Qingdao 266100, China
    2. Key Lab of Submarine Geosciences and Prospecting Techniques, MOE, Qingdao 266100, China
  • Received:2019-12-17 Revised:2020-04-27 Online:2020-08-20 Published:2020-09-02

摘要:

地震数据重建是一个不适定的反问题, 通常采用正则化方法求解. 正则化方法需要人工建模, 建模的准确性会影响重建结果, 此类方法还存在计算代价高的问题. 为克服正则化方法存在的问题, 本文使用深度卷积神经网络实现了端到端的地震数据重建.此方法是基于数据驱动的, 直接从数据中学习输入与输出的映射关系, 无需人工建模, 经过训练的网络可直接用于非完整数据的重建工作. 数值实验分别使用模拟数据和实际数据并与传统正则化方法对比验证深度卷积神经网络方法的有效性. 实验结果表明, 深度卷积神经网络方法的计算代价主要在于网络的训练阶段, 数据重建阶段仅需花费极短的时间, 与传统正则化方法相比, 对于缺道50%的地震数据, 深度卷积神经网络方法的重建结果质量更高, 速度更快.

关键词: 地震数据重建, 反问题, 正则化方法, 深度卷积神经网络

Abstract:

Due to the limitation of the exploration environment and the cost of exploration, incomplete and irregular phenomena often occur in the raw data obtained from field seismic surveys, incomplete or irregular seismic data will affect the subsequent processing of the data. This paper focuses on seismic data reconstruction method. Seismic data reconstruction is an ill-posed inverse problem. Regularization is a standard method to overcome the ill-posed of the inversion problem. The regularization method is to impose one or more constraints on the target during the inversion process so that the inversion results meet certain characteristics, such as sparsity. From this, the inversion problem can be transformed into a constrained optimization problem, and then this optimization problem can be solved by a suitable optimization algorithm. The regularization method requires manual modeling, and the accuracy of the modeling will affect the inversion results, and this kind of method also faces the challenge of high computational cost. To overcome the problems of traditional regularization methods, in this paper, a Deep Convolutional Neural Network (DCNN) method is used to implement end-to-end seismic data reconstruction. It is a data-driven method, it learns the mapping between input and output directly from the data set. The trained network can be directly used for the reconstruction of incomplete data without manual modeling. Numerical experiments use simulated data and real data to verify the effectiveness of the deep convolutional neural network method. The results show that the computational cost of the proposed method is mainly concentrated on the training of the network, and a trained network only takes a few seconds to reconstruct the seismic data. Compared with the traditional regularization method, the deep convolution neural network method has higher quality and faster speed for 50% of missing seismic data

Key words: Seismic data reconstruction, Inverse problem, Regularization method, Deep convolution neural network

中图分类号: