Seismic Data Interpolation Via Generative Adversarial Networks

Owning to the limitations of natural environment and economic conditions,there often contain missing data in the collected field seismic data.The occurrence of missing data will seriously affect the processing of subsequent data and final data interpretation.Seismic data interpolation is the process of reconstructing missing data in seismic data,and it is an important part of seismic data processing.Traditional seismic data interpolation algorithms are generally based on certain assumptions,such as assuming that the data is linear event or sparsity.Deep learning is widely used in various fields in recent years.It depends on the data itself,uses back propagation algorithms to learn and extract data features through the constructed deep network models and finally applies to data prediction.Generative adversarial networks is one of the model in deep learning.This paper completes the seismic data interpolation via generative adversarial networks which has been improved in loss function and model structure.Firstly,using the sum of Wasserstein distance and interpolation error as the loss function,the seismic data interpolation algorithm via Wasserstein generative adversarial networks is introduced.The performance of the algorithm on seismic data interpolation under different sampling forms is verified by a large number of simulated seismic data and field seismic data.For the random sampling seismic data,the algorithm can fit the data distribution roughly,but the reconstruction error is larger in the continuous missing region of the data,and some alias frequencies are not completely removed.When the data is sampled in a regular way,the algorithm shows superior performance in both data fitting and alias frequencies removing.However,the reconstruction error is larger at the boundary of data,meaning that there is a small amount of false data in boundaries.Finally,the model is improved in structure,and the seismic data interpolation algorithm via conditional Wasserstein generative adversarial networks is introduced.The improved algorithm effectively improves the performance of the model,reduces the error of the interpolation result and suppresses the generation of false data in boundaries in some degree.Both algorithms will have good application value when dealing with field seismic data interpolation problems.