DeepImaging: Deep Neural Networks Driven Full Waveform Migration

Jiachun You

doi:10.1109/TGRS.2025.3648341

DeepImaging: Deep Neural Networks Driven Full Waveform Migration

2026-01-01

0

PRE

AI

黄兴国 (Xingguo Huang) *

J

Jianping Huang

DOI：10.1109/TGRS.2025.3648341

原文链接

原文求助

摘要

En 中文

Multiples, including sea-surface-related and internal multiples, are often regarded as seismic noises. However, they also contain valuable information about subsurface geological structures and can enhance illumination when both primary and multiple waves are fully utilized-a strategy known as full waveform migration (FWM). Conventional FWM typically relies on standard one-way wave propagators as its computational core. With the advancement of deep learning, and leveraging the strong nonlinear fitting capability of convolutional neural networks (CNNs), we propose using a CNN to project the Helmholtz operator into a one-way wave propagator, and develop a DeepPropagator model. Based on this DeepPropagator model, we further develop a novel core engine for FWM, termed DeepModeling and DeepImaging models. In our numerical experiments, we use a lens and salt models that generate strong internal multiples to demonstrate the effectiveness of the proposed method. Compared to traditional one-way phase shift plus interpolation (PSPI) migration without multiple-wave attenuation, the DeepImaging approach achieves higher resolution imaging and more accurately recovers the reflectivity of the model. Moreover, we apply both the proposed DeepImaging method and the conventional one-way PSPI migration to a real seismic dataset. The results show that DeepImaging provides significantly clearer and more detailed images of geological structures compared to the conventional method. In addition, the wavenumber content of the imaging results demonstrates that our proposed DeepImaging method achieves an improvement in imaging resolution. Both numerical experiments and field data applications indicate that employing a CNN model for full-wavefield simulation and imaging is entirely feasible. Our proposed DeepImaging model establishes a novel approach in which deep learning is used to perform seismic modeling and imaging in a practical and effective manner.

Keyword:

Imaging

Numerical models

Mathematical models

Adaptation models

Deep learning

Computational modeling

Lighting

Geology

Data models

Computer architecture

Convolutional neural network (CNN)

deep learning

full waveform migration (FWM)

one-way propagator

DeepImaging: Deep Neural Networks Driven Full Waveform Migration

摘要

期刊

机构

引用论文