De-aliased seismic data interpolation using conditional Wasserstein generative adversarial networks

When sampling at offset is too coarse during seismic acquisition, spatial aliasing will appear, affecting the accuracy of subsequent processing. The receiver spacing can be reduced by interpolating one or more traces between every two traces to remove the spatial aliasing. And the seismic data with spatial aliasing can be seen as regular missing data. Deep learning is an efficient method for seismic data interpolation. We propose to interpolate the regular missing seismic data to remove the spatial aliasing by using conditional generative adversarial networks (cGAN). Wasserstein distance, which can avoid gradient vanishing and mode collapse, is used in training cGAN (cWGAN) to improve the quality of the interpolated data. One velocity model is designed to simulate the training dataset. Test results on different seismic datasets show that the cWGAN with Wasserstein distance is an accurate way for de-aliased seismic data interpolation. Unlike the traditional interpolation methods, cWGAN can avoid the assumptions of low-rank, sparsity, or linearity of seismic data. Besides, once the neural network is trained, we do not need to test different parameters for the best interpolation result, which will improve efficiency.