Noise attenuation across multiple 4D seismic vintages using deep variational autoencoder

A new deep learning approach is proposed to attenuate noise across multiple 4D seismic vintages. In this approach, a deep variational model is used during training to force the large amount of input seismic data to follow a distribution that samples to a latent space. This allows the model to avoid overfitting and exploit the spatio-temporal behaviour to efficiently separate 4D signal from background noise without using any prior information of the underlying 4D signals. The method is tested on frequently repeated towed streamer seismic data acquired over the Sleipner Field to improve the interpretation of the CO2 plume development. The results shows that a substantial noise can be removed while effectively preserving the 4D seismic signal across all vintages, despite the large amount of data and non-repeatability noise level.