Transfer learning seismic and GPR diffraction separation with a convolutional neural network

The separation of the reflected and diffracted wavefields has been a crucial challenge in both seismic and ground-penetrating radar (GPR) data processing for many years. Different deterministic methods based on wavefront attributes, adaptive subtraction of reflections or plane-wave destruction have been proposed and applied for this purpose. While all these methods are characterized by different advantages and drawbacks they generally have in common that the choice of parameters has to be adapted for each data application - particularly the different scales of seismic and GPR measurements have to be accounted for. With the aim of overcoming this drawback, in this study we train a deep convolutional autoencoder on synthetic seismic data and apply the trained neural network to seismic and GPR field data. The results demonstrate that the trained neural network is capable of successfully separating reflections and diffractions in complex data although no field data was part of the training.