Multi-Vintage Deep Learning 4D Seismic Noise Suppression: Application to Snorre Repeated Seismic Data

In this paper, the deep skipped variational autoencoder model for removal of ambient and repeatability noise across calendar time is applied on repeated 4D seismic streamer data from Snorre field. The described deep neural network’s architecture effectively captures the relevant spatio-temporal patterns to separate the 4D signal from noise without appealing to any a priori knowledge of underlying 4D signals. Noise characterization cross-plot (NCCP) characteristics in overburden and reservoir help understand noise trends and tune the model parameters for the learning process. Filtering 4D signal with this denoising approach can thereby result in improvement of 4D attributes.