Objective The non-uniform interpolation of seismic data is identified as a prolonged challenge in energy exploration. Since geophones cannot be precisely placed at positions corresponding to theoretical grid points, current uniform interpolation techniques frequently suffer deviations and detail distortion.
Methods This study proposed a novel non-uniform interpolation method based on a diffusion probabilistic model, which is an emerging generative model in deep learning that involves the diffusion and generation processes. In the diffusion process, noise is added to the complete seismic data iteratively to train the denoising capability of the neural network. In the generation process, the neural network is employed for iterative denoising of data containing noise to obtain the reconstructed data. In this study, interpolation operators were employed to calculate the deviations between iterative and sampled data. These deviations were then used as the additional inputs of the neural network to improve the non-uniform interpolation capability of the diffusion probabilistic model. In the numerical experiments, the non-uniform sampling was tested using 2D synthetic and actual datasets, and the uniform interpolation model was compared with the model in this proposed study.
Results and Conclusions The results indicate that the proposed method significantly enhanced the processing capability of the diffusion probabilistic model for non-uniform sampling. The tests of synthetic and actual data revealed an increase of approximately 7 dB in the signal-to-noise ratio. Therefore, the proposed method can effectively improve the precision of deep learning for non-uniform interpolation, providing a new approach for non-uniform interpolation algorithms of seismic data.
DownLoad: