Scholarlink

<p>Lithology prediction is crucial for oil and gas reservoir exploration, forming the foundation for reservoir characterization, reserve assessment, and geological modelling. Recent research has shown promising results by integrating time-frequency (T-F) analysis with deep learning (DL) approaches. However, existing methods typically rely on features extracted from a single T-F spectrum resolution as inputs to DL networks, limiting their ability to capture complex lithological patterns across different scales. To address this limitation, we propose a novel multiscale feature-fusion convolutional neural network (MFFCNN) that effectively extracts comprehensive lithological information from T-F spectra. Our method applies the Gabor transform with adaptive window width to convert one-dimensional seismic data into multiple T-F representations with varying T-F resolution. Convolutional neural networks (CNNs) then extract key features from these complementary T-F spectra and fuse them to capture the nonlinear relationship between the seismic characteristics and lithology classifications. This process enhances lithology-related information derived from seismic data. The results show that the proposed method outperforms traditional algorithms, improving the F1-score of lithology prediction by 5 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation> and enhancing the reliability of subsequent reservoir characterization processes.</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

Lithology prediction via multiscale feature-fusion convolutional neural networks

  • 期刊论文

  • 发表日期：

    2025年11月4日

查看全文
  • Wenliang Nie,
  • Heyu Zhang,
  • Jiayi Gu,
  • Wei Huang,
  • Bo Li,
  • Jianfeng Liu,
  • Xiangfei Nie

摘要

Lithology prediction is crucial for oil and gas reservoir exploration, forming the foundation for reservoir characterization, reserve assessment, and geological modelling. Recent research has shown promising results by integrating time-frequency (T-F) analysis with deep learning (DL) approaches. However, existing methods typically rely on features extracted from a single T-F spectrum resolution as inputs to DL networks, limiting their ability to capture complex lithological patterns across different scales. To address this limitation, we propose a novel multiscale feature-fusion convolutional neural network (MFFCNN) that effectively extracts comprehensive lithological information from T-F spectra. Our method applies the Gabor transform with adaptive window width to convert one-dimensional seismic data into multiple T-F representations with varying T-F resolution. Convolutional neural networks (CNNs) then extract key features from these complementary T-F spectra and fuse them to capture the nonlinear relationship between the seismic characteristics and lithology classifications. This process enhances lithology-related information derived from seismic data. The results show that the proposed method outperforms traditional algorithms, improving the F1-score of lithology prediction by 5 \(\%\) and enhancing the reliability of subsequent reservoir characterization processes.