Deep learning diffraction separation for seismic and GPR data

Within the last decade, the diffracted wavefield has gained increasing importance for the processing of both seismic and ground-penetrating radar (GPR) measurements. In both communities, the separation of the diffracted wavefield remains a notorious challenge that has been approached with different deterministic methods, ranging from poststack wavefront attributes to plane-wave destruction and coherent wavefield separation. While each of these deterministic methods has characteristic advantages and drawbacks, all of them require the adaptation of processing parameters for each application, particularly when crossing scales from seismic to GPR measurements. In this study, we propose to train a convolutional autoencoder to separate the reflected and diffracted wavefields in a generalized fashion. For this purpose, we have generated highly variable synthetic seismic data that contain reflections, diffractions and noise using an algorithm that allows to compute each component individually, resulting in an automatized generation of data and labels. In order to account for the complexity of field data, we complemented the synthetic data with a large set of reference seismic and GPR field data results from coherent wavefield separation, a deterministic method, in which the reflected wavefield is modeled and adaptively subtracted from the input data. With this dataset we trained a supervised convolutional autoencoder and applied the trained neural network to seismic and GPR field measurements that were not part of the training data. The results show that the trained autoencoder is able to generalize and successfully separate the reflected and diffracted wavefields even for complex field data, resulting in an on-the-fly diffraction separation that requires no choice of parameters and is likewise applicable to both seismic and GPR data.