Scholarlink

<p>As seismic acquisition systems advance, how to effectively suppress complex noise and restore weak signals in seismic data has become a key task. Seismic data denoising through deep learning approaches has grown significantly in recent years, with U-Net showing great potential, yet limitations remain in its ability to adequately address both complex noise and weak signal restoration. This study resolves these challenges through a refined U-Net network based on multi-scale double-layer convolution and boundary enhancement combined with an attention mechanism for seismic data denoising. The proposed model replaces standard U-Net convolutions with multi-scale double-layer convolutions to capture richer hierarchical features; a boundary enhancement module is used to recover weak seismic signals, and an attention-based fusion module is designed to effectively combine multi-scale features with enhanced edge information. When testing synthetic data and field seismic data with complex noise levels, we evaluated our algorithm against competing methods that utilize both convolutional neural networks and transformer architectures. The results indicate that the algorithm significantly improves denoising performance while effectively preserving seismic details.</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

MSBE-UNet: A deep learning denoising method for effective seismic noise suppression

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Hongtao Xi,
  • Jingrui Luo,
  • Jiangchao Liu,
  • Wenze Shi,
  • Guoxin Chen,
  • Naijian Wang,
  • Xingguo Huang

摘要

As seismic acquisition systems advance, how to effectively suppress complex noise and restore weak signals in seismic data has become a key task. Seismic data denoising through deep learning approaches has grown significantly in recent years, with U-Net showing great potential, yet limitations remain in its ability to adequately address both complex noise and weak signal restoration. This study resolves these challenges through a refined U-Net network based on multi-scale double-layer convolution and boundary enhancement combined with an attention mechanism for seismic data denoising. The proposed model replaces standard U-Net convolutions with multi-scale double-layer convolutions to capture richer hierarchical features; a boundary enhancement module is used to recover weak seismic signals, and an attention-based fusion module is designed to effectively combine multi-scale features with enhanced edge information. When testing synthetic data and field seismic data with complex noise levels, we evaluated our algorithm against competing methods that utilize both convolutional neural networks and transformer architectures. The results indicate that the algorithm significantly improves denoising performance while effectively preserving seismic details.