Scholarlink

<p>Deep learning (DL) has shown great promise for seismic facies classification (SFC). However, the absence of systematic evaluation frameworks currently hinders fair comparisons between models. Existing studies often use different data partitioning strategies, leading to inconsistent benchmarks and making it hard to identify the most effective approaches. This paper systematically evaluates eleven state-of-the-art DL models for SFC, including both traditional and lightweight architectures, 2D and 3D models, and also a foundation model. Using three public datasets, Parihaka, Penobscot, and Netherlands F3, we analyze the effects of data partitioning strategies, training data size, model complexity, and accuracy. Beyond assessing conventional Large Rectangular Prisms (LRP), we introduce two novel data partitioning strategies: Rectangular Prisms with Equally Distant Slices (RPEDS) and Equally Distant Slices (EDS). These strategies offer diverse global information, aiming to reduce overall annotation demands. Notably, EDS achieves accuracy comparable to or superior to LRP, while requiring up to 13<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation> less labeled data. We also demonstrate that lightweight DL models can match or outperform heavyweight architectures in various scenarios, providing significant computational savings with minimal to no accuracy loss. Our systematic evaluation framework, combined with these partitioning strategies, offers new baselines and practical guidelines for designing and scalabling SFC solutions using DL.</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

A systematic evaluation methodology of deep learning on seismic facies classification

  • 期刊论文

  • 发表日期：

    2025年10月13日

查看全文
  • Gabriel Lima,
  • Gabriel Amarante,
  • Willian Barreiros Jr.,
  • Matheus T. P. Souza,
  • Wagner Meira Jr.,
  • Renato Ferreira,
  • George Teodoro

摘要

Deep learning (DL) has shown great promise for seismic facies classification (SFC). However, the absence of systematic evaluation frameworks currently hinders fair comparisons between models. Existing studies often use different data partitioning strategies, leading to inconsistent benchmarks and making it hard to identify the most effective approaches. This paper systematically evaluates eleven state-of-the-art DL models for SFC, including both traditional and lightweight architectures, 2D and 3D models, and also a foundation model. Using three public datasets, Parihaka, Penobscot, and Netherlands F3, we analyze the effects of data partitioning strategies, training data size, model complexity, and accuracy. Beyond assessing conventional Large Rectangular Prisms (LRP), we introduce two novel data partitioning strategies: Rectangular Prisms with Equally Distant Slices (RPEDS) and Equally Distant Slices (EDS). These strategies offer diverse global information, aiming to reduce overall annotation demands. Notably, EDS achieves accuracy comparable to or superior to LRP, while requiring up to 13 \(\times\) less labeled data. We also demonstrate that lightweight DL models can match or outperform heavyweight architectures in various scenarios, providing significant computational savings with minimal to no accuracy loss. Our systematic evaluation framework, combined with these partitioning strategies, offers new baselines and practical guidelines for designing and scalabling SFC solutions using DL.