首页|有监督深度学习的地震资料提高分辨率处理方法

有监督深度学习的地震资料提高分辨率处理方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
地震资料分辨率直接影响后续处理和解释成果精度因此备受关注.深度学习方法具备自动提取深层特征和出色的非线性逼近能力,在反问题求解中广泛应用.在地震勘探领域,深度卷积网络中的卷积算子与地震数据的褶积模型相吻合,因而有望通过智能化手段显著提升地震资料的分辨率.目前,针对卷积神经网络提高地震资料分辨率方面的研究发展迅速,但问题的核心在于设计适合、有效的网络结构和损失函数.为此,提出一种基于强监督学习的地震资料高分辨率处理方法.该方法充分利用地下结构的空间连续性,借鉴图像超分辨率重建的思想,设计了一种生成对抗网络结构,用以提高地震资料的纵向分辨率;同时,采用L1损失和多尺度结构相似性(MS-SSIM)损失相结合的损失函数提高感知质量,以提高网络的高分辨率处理效果.合成数据和实际地震数据的应用结果显示,相较于常规损失函数,文中采用的损失函数可以显著提升智能算法的处理效果,明显改善地震数据同相轴的连续性,且高频细节信息更丰富,验证了该方法的可行性和有效性.
High-resolution seismic data processing method based on supervised deep learning
The resolution of seismic data directly influences the subsequent processing and interpretation preci-sion,thus attracting considerable attention.Deep learning is widely used in solving reverse problems due to its capacity for automatic extraction of deep features and excellent nonlinear approximation.In the field of seismic exploration,the convolution operators in deep convolutional networks are consistent with the convolutional model of seismic data,which has the potential to significantly improve the resolution of seismic data through in-telligent means.Currently,enhancing the resolution of seismic data through convolutional neural networks has become a research hotspot.The key to addressing this issue lies in designing suitable and effective network structures and loss functions for resolution enhancement.Therefore,a high-resolution seismic data processing method based on strong supervised deep learning is proposed.Drawing inspiration from image super-resolution reconstruction,this method makes full use of the spatial continuity of the underground structure,and a genera-tive adversarial network structure is designed to enhance the longitudinal resolution of seismic data.Additionally,a loss function combining L1 loss and multi-scale structural similarity loss is employed to improve the perceived quality of deep learning networks.The experimental results of seismic data and actual seismic data show that compared to the conventional loss function,the loss function presented in this study can significantly enhance the high-resolution processing performance of intelligent algorithms.It notably improves the continuity of the seismic events and enriches the high-frequency detail information of seismic data,and the feasibility and effec-tiveness of the proposed method are verified.

supervised deep learningmulti-scale structural similarity loss(MS-SSIM)L1 loss functiongenera-tive adversarial networkimage super-resolution reconstruction

李斐、牛文利、刘达伟、王永刚、黄研

展开 >

长庆油田勘探开发研究院,陕西西安 710018

西安交通大学电子与信息学部,陕西西安 710064

有监督深度学习 多尺度结构相似性损失 L1损失 生成对抗网络 图像超分辨率重建

国家自然科学基金面上项目中国石油集团重大专项

423741352023ZZ25

2024

10.13810/j.cnki.issn.1000-7210.2024.04.006

石油地球物理勘探
东方地球物理勘探有限责任公司

石油地球物理勘探

CSTPCD北大核心
影响因子：1.766
ISSN：1000-7210
年,卷(期)：2024.59(4)