Training deep networks with only synthetic data

Abstract

Accurate removal of surface-related multiples remains a challenge in shallow-water cases. One reason is that the success of surface-related multiple estimation (SRME)-related algorithms is sensitive to the quality of the near-offset reconstruction. When it comes to a larger missing gap and a shallower water bottom, the state-of-the-art near-offset gap construction method - the parabolic Radon transform - fails to provide reliable recovery of the shallow reflections due to the limited information from the data and highly curved events at near offsets with strong lateral amplitude variations. Therefore, we have developed a novel workflow that first deploys a deep-learning-based reconstruction of the shallow reflections and then uses the reconstructed data as the input for the subsequent surface multiple removal. In particular, we use a convolutional neural network architecture - U-net that was developed from convolutional autoencoders with extra direct skip connections between different levels of encoders and the corresponding decoders. Instead of using field data directly in network training, the training set is carefully synthesized based on the prior water-layer information of the field data; thus, a fully sampled field data set, which is difficult to obtain, is not needed for training in our workflow. An inversion-based approach - closed-loop SRME - is used for the surface multiple removal, in which the primaries are directly estimated via full-waveform inversion in a data-driven manner. Finally, the effectiveness of our workflow is determined based on 2D North Sea field data in a shallow-water scenario (92.5 m water depth) with a relatively large minimum offset (150 m).