P/S-wave separation of multicomponent seismic data at the land surface based on deep learning

Abstract

P/S-wave separation is a key step for data processing in multicomponent seismic exploration. The conventional methods rely on either the prior information of near-surface elastic properties or the carefully selected parameters to estimate the polarization directions of the P- and S-modes when arriving at the geophones. In the case of complex wave propagation or the near-surface elastic parameters being unavailable, the conventional methods fail to achieve a high-quality P/S separation. The P/S-wave separation of multicomponent seismic data recorded at the land surface can be regarded as a nonlinear problem of point-by-point image prediction and addressed by using the powerful ability for feature extraction of a deep neural network. To obtain an accurate, abundant, and representative labeled data set, a variety of elastic models are constructed through geologically reasonable data augmentation, followed by elastic forward modeling and Helmholtz decomposition. A convolutional neural network is trained with these labeled data to predict scalar P- and S-wave recordings from the multicomponent data in a data-driven manner. Eventually, a novel approach of P/S-wave separation is established for multicomponent seismic imaging with the help of deep learning, which can work without prior knowledge of the near-surface elastic parameters. Numerical examples indicate that, by constructing an augmented labeled data set for training, our method significantly improves the generalization of the deep neural network and thus enables the trained neural network to achieve more accurate P/S-wave separation than the conventional model-driven method for the target data.