Wavefield separation of VSP data based on deep neural network

Abstract

Vertical seismic profile (VSP) data contains upgoing and downgoing waves, making the whole wavefields chaotic, affecting subsequent imaging, interpretation, and inversion. It is difficult to obtain accurate results through the conventional separation methods of upgoing and downgoing wavefields. Aiming at the high-precision separation of VSP data, we propose a wavefield separation method based on the deep neural network(DNN). Firstly, the downgoing wavefield data separated by the traditional method are used as the output, and the original data are used as the input for training to obtain the training model. Then, based on the training model, the upgoing wavefield is removed and the downgoing wavefield is output by the convolutional auto-encoder(CAE). Finally, the upgoing wavefield is separated by the same operation. The application of forward modeling data and the field data in the Gulf of Mexico indicates that the wavefield separation method based on the deep neural network is efficacious for obtaining pure upgoing and downgoing wavefields in VSP surveys. And the results of this method are more accurate than the traditional methods, i.e., f-k filtering, SVD, and median filtering.