Efficient 3D Deep Learning Diffraction Separation for Seismic and GPR Data

The diffracted wavefield plays an important role in the processing and interpretation of both seismic and ground-penetrating radar (GPR) data, as it encodes information about small-scale subsurface heterogeneities such as faults, small objects or water intrusions in glaciers. Separating the faint diffracted wavefield in data that is often dominated by higher amplitude reflected arrivals is therefore a key challenge in the processing of both seismic and GPR data. When using deterministic methods for diffraction separation, parameters have to be adapted for every dataset, and the application to a 3D data cube can become computationally very expensive. We have recently proposed to train a convolutional neural network to decompose a given wavefield into its reflected and diffracted components and demonstrated that the trained neural network is largely able to generalize to unseen data. In this work, we propose a highly efficient approach for diffraction separation on large 3D datasets by applying the trained neural network to all inlines of 3D seismic and 3D GPR data. The results suggest that this approach is able to provide a consistent and highly-resolved 3D diffracted wavefield.