DiffractioNet: Deep-learning seismic and GPR diffraction separation

The separation of the diffracted wavefield is a notorious challenge in seismic and ground-penetrating-radar (GPR) subsurface imaging. Over the past decade, numerous studies have attempted to address this challenge with various deterministic schemes. Although each of these schemes has specific advantages and disadvantages, they all require an adaptation of the corresponding processing parameters for each application, especially when crossing scales between seismic and electromagnetic measurements. In recent years, convolutional neural networks (CNNs) have emerged as powerful tools for data analysis. However, their performance strongly depends on training data and labels, the generation of which can be a complex and time-consuming process. In this study, we introduce DiffractioNet, a deep-learning framework for diffraction separation. Our approach is based on the automated generation of synthetic seismic data patches and corresponding labels for the reflected and diffracted wavefields. We augment this synthetic data set with reference field data results from coherent wavefield subtraction, a well-established deterministic method for diffraction separation. With the combined data set, we train a CNN to decompose any input seismic or GPR data into reflected and diffracted wavefields. The trained DiffractioNet provides a solution for efficient on-the-fly diffraction separation that does not require parameter adaptation. We demonstrate this by applying DiffractioNet to unseen seismic and GPR field data.