Wavefield decomposition for diffraction separation with convolutional neural networks

The process of wavefield decomposition is a notorious challenge in seismic and electromagnetic imaging, because often only specific components of the full measured wavefield are targeted during the processing, while other components are undesired. This specifically applies to the separation of the reflected and diffracted wavefields. While the diffracted wavefield has often been treated as noise in the past, recent advances have demonstrated its importance for both high-resolution imaging of subsurface heterogeneities such as faults and depth-velocity model building for multi-channel and single-channel seismic and electromagnetic data. The potential of convolutional neural networks (CNNs) for sophisticated data analysis has been demonstrated in numerous studies in the recent past. However, the training of deep neural networks generally requires a large amount of labeled data and the application of trained networks to previously unseen data often poses a challenge. In this study, we generate synthetic seismic data composed of three separate wavefield components - reflections, diffractions and noise - with randomly distributed reflectors and diffractors and use these data to train a deep convolutional neural network, which tries to decompose the wavefield of any input seismic or ground-penetrating-radar data into reflections, diffractions and noise. Applications to both unseen validation data and seismic field data show promising results.