Direct Diffraction Separation by Deep Learning on Pre-Migrated Seismic Data

Abstract

Summary Diffraction imaging is a niche imaging technique which aims to directly image discontinuities in the subsurface by separating diffractions from the rest of the wavefield and processing them independently. However, to separate diffractions is a complicated procedure due to their weak amplitudes and the overlap of energies between diffractions and the much stronger reflections. While analytical methods exist to separate diffractions, they require parameterisation, are comparatively computationally expensive, and leave a volume which contains both diffractions and noise. Here, we aim to use a Generative Adversarial Network (GAN) to automatically separate diffractions from reflections on pre-migrated seismic data without the need for parameterisation. We have applied the GAN to two real datasets, one for validation, which comes from the same dataset used in training, and one which is used solely for prediction. This shows good results for both the validation and prediction data when compared to plane-wave destruction, an analytical separation technique, and is applied in a fraction of the time. The prediction dataset is then added to the overall training data, the network retrained and applied to the same validation data. This further improves the separation on the validation data and suggests that additional data may enhance the separation.