Scholarlink

<p>Full waveform inversion (FWI) is a prevalent method for estimating subsurface model parameters, typically employing a frequency-multiscale serial inversion strategy to achieve the required resolution. However, this approach is computationally costly and often yields imprecise results due to the frequency-dependent resolution of velocity model. To enhance both the efficiency and accuracy of FWI, this study introduces a modified spatial multiscale serial high-resolution inversion strategy underpinned by deep learning. Initially utilizing a coarse grid for low-frequency inversion to capture the general subsurface structure, this strategy employs super-resolution generative adversarial networks (SRGAN) to map coarse grid data onto a fine grid as the inversion frequency increases, facilitating lossless data enhancement. This transition provides superior model details for high-frequency inversion on the fine grid, achieving a scalable, frequency-sequential serial inversion from lower to higher scales, while effectively reducing data space consumption at lower frequencies. Furthermore, the incorporation of a residual network (ResNet) enhances the recovery of high-frequency details and physical property boundaries. Experimental results using the Overthrust and Marmousi-II benchmark standard models demonstrate that the revised spatial multiscale FWI method not only boosts inversion efficiency but also significantly improves inversion stability and data resolution.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Spatial-Temporal Multiscale Full Waveform Inversion of Seismic Waves Based on Superresolution Generative Adversarial and Residual Networks

  • 期刊论文

  • 发表日期：

    2025年8月8日

查看全文
  • Hao Ding,
  • Wangsuo Cai,
  • Wenyue Wu,
  • Chaojin Wang,
  • Shijie Fan,
  • Dongchang Zhao

摘要

Full waveform inversion (FWI) is a prevalent method for estimating subsurface model parameters, typically employing a frequency-multiscale serial inversion strategy to achieve the required resolution. However, this approach is computationally costly and often yields imprecise results due to the frequency-dependent resolution of velocity model. To enhance both the efficiency and accuracy of FWI, this study introduces a modified spatial multiscale serial high-resolution inversion strategy underpinned by deep learning. Initially utilizing a coarse grid for low-frequency inversion to capture the general subsurface structure, this strategy employs super-resolution generative adversarial networks (SRGAN) to map coarse grid data onto a fine grid as the inversion frequency increases, facilitating lossless data enhancement. This transition provides superior model details for high-frequency inversion on the fine grid, achieving a scalable, frequency-sequential serial inversion from lower to higher scales, while effectively reducing data space consumption at lower frequencies. Furthermore, the incorporation of a residual network (ResNet) enhances the recovery of high-frequency details and physical property boundaries. Experimental results using the Overthrust and Marmousi-II benchmark standard models demonstrate that the revised spatial multiscale FWI method not only boosts inversion efficiency but also significantly improves inversion stability and data resolution.