Scholarlink

<p>Seismic data denoising has become a major challenge in seismic exploration, particularly under poor signal-to-noise ratio (SNR) conditions. Therefore, achieving effective separation of signal and noise, especially when weak signals are buried in strong noise, is of critical importance. Traditional denoising methods often fail to simultaneously focus on time–frequency-domain features, limiting their denoising effectiveness. In the TF-UNet architecture, the real and imaginary components of the seismic noise frequency domain are obtained through the fast Fourier transform (FFT) and combined with the noise-free time domain to form a three-channel input for the U-Net. This design enables the model to effectively learn the joint time–frequency representations of seismic data. Furthermore, the TF-UNet is trained using hybrid datasets that integrate field and synthetic seismic data, allowing the network to learn field noise characteristics. Experimental results demonstrate that, compared with the methods discussed in this study, the proposed TF-UNet achieves superior denoising performance on both field and synthetic seismic datasets, effectively suppressing noise and preserving signal fidelity. Although supervised training requires labeled data, the TF-UNet can be effectively applied in geophysical exploration given high-quality annotations. Once trained, the model can process seismic data in real time on systems equipped with parallel computing and time processing capabilities, thereby enhancing the efficiency of seismic data analysis in exploration applications.</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

TF-UNet: a time–frequency feature network for seismic noise attenuation using hybrid synthetic–field data

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Shengrong Zhang,
  • Gaocai Wang,
  • Liuye Lu,
  • Manyi Wei,
  • Xuesha Qin,
  • Liang Zhang

摘要

Seismic data denoising has become a major challenge in seismic exploration, particularly under poor signal-to-noise ratio (SNR) conditions. Therefore, achieving effective separation of signal and noise, especially when weak signals are buried in strong noise, is of critical importance. Traditional denoising methods often fail to simultaneously focus on time–frequency-domain features, limiting their denoising effectiveness. In the TF-UNet architecture, the real and imaginary components of the seismic noise frequency domain are obtained through the fast Fourier transform (FFT) and combined with the noise-free time domain to form a three-channel input for the U-Net. This design enables the model to effectively learn the joint time–frequency representations of seismic data. Furthermore, the TF-UNet is trained using hybrid datasets that integrate field and synthetic seismic data, allowing the network to learn field noise characteristics. Experimental results demonstrate that, compared with the methods discussed in this study, the proposed TF-UNet achieves superior denoising performance on both field and synthetic seismic datasets, effectively suppressing noise and preserving signal fidelity. Although supervised training requires labeled data, the TF-UNet can be effectively applied in geophysical exploration given high-quality annotations. Once trained, the model can process seismic data in real time on systems equipped with parallel computing and time processing capabilities, thereby enhancing the efficiency of seismic data analysis in exploration applications.