The development of advanced denoising techniques for seismic data is crucial for accurate geological interpretation. Traditional methods, such as wavelet thresholding and multiscale singular spectrum analysis (MSSA), primarily rely on mathematical and physical priors, which can limit their effectiveness. In contrast, deep learning-based approaches leverage diverse datasets to directly mitigate noise patterns without making physical assumptions. This study introduces a diffusion probabilistic model (DDPM) specifically for prestack seismic data denoising.

The DDPM operates by simulating a process that progressively adds noise to the data and then recovers the original signal through a reverse denoising process. While Ho et al. (2020) demonstrated that noise prediction is optimal for natural images, seismic images exhibit more consistent patterns and simpler distributions. Consequently, we have adjusted the training strategy and modified the loss function to prioritize signal fitting for seismic image denoising. We propose an innovative signal-fitting training strategy to enhance the effectiveness of DDPM in seismic denoising tasks. Moreover, to address the limitations of supervised methods when dealing with unseen data, we incorporate a regularization term into the loss function, thereby improving the model's generalization capabilities.

Our proposed method achieves superior noise suppression while preserving critical signal components with higher precision, resulting in less deviation from the ground truth compared to traditional DDPM. The signal-to-noise ratio (SNR) improvement is approximately 1.5 dB higher than that of DDPM. Furthermore, cross-field data tests reveal that incorporating a total variation (TV) constraint enhances SNR by about 1 dB for mildly noisy data, with diminishing returns as noise levels increase. Field data also confirms the effectiveness of our proposed method.

In conclusion, this study presents an innovative training strategy that significantly improves DDPM's applicability to seismic data denoising. Our approach outperforms traditional methods in noise suppression and signal recovery, demonstrating effective adaptability to seismic data while necessitating careful tuning to prevent overfitting. The findings confirm the potential of this method to enhance data quality for better geological interpretation.