Leveraging domain adaptation for efficient seismic denoising

Abstract

The selection of training data for deep learning procedures dictates both the neural network's performance and its applicability to further datasets. Particularly in seismic applications, the selection is non-trivial with the common approaches of manually labelling field data or generating synthetic data both exhibiting severe limitations. The former in its inability to outperform conventional approaches required for the label generation and the later in its inability to properly represent future application data. Domain adaptation, through input features and label transformations, offers the potential to leverage on the benefits of both these approaches while reducing their drawbacks. In this work we illustrate how vital information from field data can be incorporated into a training procedure on synthetic data with the trained network successfully applied on the field data afterwards, despite large differences between the training and inference, i.e., synthetic and field, datasets. Furthermore, we illustrate how an inverse correlation procedure can be incorporated into the training procedure in an attempt to maintain the original wavefield properties.