CDDIP: Constrained Diffusion-Driven Deep Image Prior for Seismic Data Reconstruction

Paul Goyes-Peñafiel, Ulugbek S. Kamilov, Henry ArguelloIEEE GRSL 2025 · OpenReview

TL;DR

This paper introduces a novel seismic data reconstruction method that combines the power of diffusion models for generating realistic seismic data with the Deep Image Prior (DIP) technique to enforce data consistency. This approach significantly improves the reconstruction of missing seismic traces, demonstrating robust generalization across diverse geological scenarios and reducing inference time compared to existing diffusion-based methods.

What Problem Does This Paper Address?

Seismic data acquisition often results in missing or corrupted traces, which severely hinders subsequent seismic processing and interpretation. While deep learning has shown promise in reconstructing this missing data, existing methods struggle with generalization to new geological settings and can be computationally expensive during inference. This paper tackles the challenge of creating a robust and efficient seismic data reconstruction method that can handle complex geological variations and reduce computational burden.

Methods

The core of the proposed method, CDDIP (Constrained Diffusion-Driven Deep Image Prior), lies in its hybrid approach. It leverages a pre-trained generative diffusion model (DDPM) to synthesize realistic seismic data. Crucially, instead of relying solely on the diffusion model's sampling process, CDDIP incorporates a Deep Image Prior (DIP) solver. This DIP acts as a data consistency regularizer, ensuring that the generated seismic data aligns with the available partial measurements.

The process involves an iterative refinement where the DIP solver guides the diffusion model's generation at specific timesteps. This allows for a more constrained and accurate reconstruction. A key innovation is the ability to skip diffusion sampling steps by utilizing the DIP solver, which significantly speeds up the inference process. The DIP solver itself is designed to leverage parameters from previous states, acting as a warm-starting model for improved efficiency.

Datasets

The authors trained their diffusion model on a substantial collection of seismic datasets, including:

  • Post-stack datasets: 14,398 patches from 6 databases (TGS salt Identification challenge, SEAM Phase I, F3 Netherlands, 1994 BP, AGL Elastic Marmousi, Kerry 3D).
  • Pre-stack datasets: 3,400 patches from 8 databases (Alaska 2D land, AvoMobil, BP model 94, SEG-C3, synthetic cross-spread, Seam Phase I, Seam Phase II/ Foothills, Stratton 3D).

These datasets encompass both synthetic and field data, representing a wide range of geological complexities. The evaluation also included datasets entirely outside the training domain (Stratton, Penobscot, Blake Ridge, and Alaska) to specifically test generalization capabilities.

Key Insights and Ideas

  • Synergy of Diffusion Models and DIP: Combining the generative power of diffusion models with the constraint-enforcing nature of DIP proves highly effective for seismic reconstruction. Diffusion models excel at capturing the underlying data distribution, while DIP ensures fidelity to the observed data.
  • Skipping Diffusion Steps for Speed: The ability to skip diffusion sampling steps by using the DIP solver as a regularizer is a major breakthrough for inference efficiency. This drastically reduces computation time without sacrificing reconstruction quality.
  • Generalization Beyond Training Domain: The method demonstrates remarkable robustness and generalization capabilities, performing well on datasets that were not part of the training set. This is crucial for real-world applications where geological variability is high.
  • DIP as a Warm-Starter: The DIP solver's ability to leverage previous states as initialization further enhances its efficiency and contributes to the overall speedup.

Key Results

  • Improved Reconstruction Quality: The proposed CDDIP method consistently outperforms state-of-the-art methods like DiffPIR and CCSeis-DDPM in terms of PSNR and SSIM metrics, especially in complex scenarios with removed traces. For instance, it achieved up to 2.2 dB higher PSNR and 0.23 higher SSIM on the Blake Ridge dataset.
  • Significant Inference Speedup: CDDIP reduces the number of sampling timesteps by up to 4x compared to other diffusion-based methods. For example, achieving good results with k=25 timesteps and an average computation time of 23 seconds.
  • Robustness to Noise and Complex Structures: The method effectively reconstructs missing traces and reduces seismic noise, even in datasets with inclined structures, curves, and coherent noise (e.g., ground roll).
  • Effective for Both Post-stack and Pre-stack Data: The approach shows strong performance across both post-stack and pre-stack seismic data, highlighting its versatility.

Limitations and Future Work

  • DIP Solver Epochs: While the method reduces sampling timesteps, the DIP solver still requires a few epochs to converge. The paper notes that strict control over the number of epochs and early stopping is necessary to avoid overfitting noise.
  • Computational Cost of Training: Training the diffusion model itself can be computationally intensive, as indicated by the 6-day training time for 5000 epochs.
  • Future Work: The authors suggest exploring different timestep schedules for the DIP solver and investigating further optimizations for inference speed. The potential for applying this method to other seismic processing tasks beyond reconstruction is also a promising avenue.

Summary updated

Abstract

Seismic data frequently exhibit missing traces, substantially affecting subsequent seismic processing and interpretation. Deep learning-based approaches have demonstrated significant advancements in reconstructing irregularly missing seismic data through supervised and unsupervised methods. Nonetheless, substantial challenges remain, such as generalization capacity and computation time cost during the inference. This work introduces a reconstruction method that uses a pretrained generative diffusion model for image synthesis and incorporates deep image prior (DIP) to enforce data consistency when reconstructing missing traces in seismic data. The proposed method has demonstrated strong robustness and high reconstruction capability of poststack and prestack data with different levels of structural complexity, even in field and synthetic scenarios where test data were outside the training domain. This indicates that our method can handle the high geological variability of different exploration targets. Additionally, compared to other state-of-the-art seismic reconstruction methods using diffusion models, during inference, our approach reduces the number of sampling timesteps by up to $4\times $ . Our implementation is available at https://github.com/PAULGOYES/CDDIP.git.
BibTeX
@article{goyespenafiel2025cddip,
  author = {Paul Goyes-Peñafiel and Ulugbek S. Kamilov and Henry Arguello},
  title = {{CDDIP: Constrained Diffusion-Driven Deep Image Prior for Seismic Data Reconstruction}},
  journal = {IEEE Geoscience and Remote Sensing Letters},
  year = {2025},
  url = {https://openreview.net/forum?id=DtYeyN6sHx},
}
Figure