Abstract
Ineffective recognition of weak seismic signals limits fine structural interpretation and reservoir prediction in seismic exploration. Although seismic signal reconstruction has made significant progress, the widespread issue of low Signal-to-Noise Ratio (SNR) remains a major challenge in recovering weak seismic signals from noisy field data. To address this, we propose a prior-based Latent denoising Diffusion model for weak Seismic signal Restoration (LDSR). First, we design a Multi-Scale Information Distillation Attention (MSIDA) network to extract discriminative features from noisy inputs and aggregate the multi-scale spatial information, ensuring more stable model convergence. Second, to effectively establish continuous relationships among features in the latent space, we develop a forward Markovian restoration process involving multiple diffusion iterations. In each iteration, we integrate the MSIDA network and latent neural representation into a unified end-to-end framework. Third, we leverage Non-Subsampled Shearlet Transform (NSST) sub-bands to represent high-frequency prior, which is expected to enhance the details of reconstructed seismic signals. When evaluated on synthetic and logging forward data, our method outperforms state-of-the-art methods. Ablation studies validate the effectiveness of each component in LDSR. When processed on real field data, LDSR successfully restores more effective weak signals near the target layers.
Funding Information
  • National Natural Science Foundation of China (62406051)
  • Taishan Scholar (tsqn202507224)
  • Scientific Research Fund of Liaoning Provincial Education Department (LJ212510167007)
  • National Key Research and Development Program of China (2019YFB2102400)

This publication has 70 references indexed in Scilit: