Deep learning diffraction separation for seismic and GPR data

Diffraction separation remains a challenge in both seismic and ground-penetrating radar (GPR) imaging. Different deterministic methods have been proposed for this purpose. Although these methods are of different nature, they have in common that processing parameters have to be adapted for each application and, in particular, when changing scales between seismics and GPR. In the recent years, convolutional neural networks have proven to be a powerful tool for data analysis. However, their performance strongly depends on the training data and labels, whose generation is often a highly time-consuming task. In this study, we propose to generate synthetic seismic data that contain reflections, diffractions and noise and the corresponding labels consisting of only reflections and diffractions, respectively, in an entirely automatized fashion. We complement this data by reference results from field data applications of coherent wavefield subtraction, a deterministic method that extracts the reflected and diffracted wavefields from the input data. With the combined dataset we train a convolutional autoencoder and apply the trained network to seismic and GPR field data. The results show that the network is able to generalize and successfully decompose reflected and diffracted wavefields, resulting in an on-the-fly diffraction separation that requires no adaptation of processing parameters.