Wave field prediction for denoising of seismic measurements using an autoencoder network

The adoption of machine learning in different sectors demonstrates a huge potential of multiple techniques applicable to learn specific features of a dataset. We aim to make use of machine learning methods to predict the development of seismic wave fields between two seismic stations and use this information to remove random noise post-measurement, considering the phase and time information of the signal. Thereby, the initial approach follows the use of an autoencoder network in a self-supervised fashion. Aiming to reconstruct its input, the form of the autoencoder corresponds to the traditional U-Net structure but expands with residual blocks for increased network capacity. To refine results, we modify the interrelated training process of encoding and reconstruction and separate it into sequential phases. To make sure that the dataset includes multiple sources and thus provides various features, we use field data gathered at a seismic exploration site in an area containing several roads, wind turbines, oil pump jacks and railway traffic. Using the well-known autoencoder network structure and applying it in the context of transfer learning enables us to automatically learn a representation of the wave field and, more importantly, predict its spatial development based on different frequency bands. The gained knowledge can be used in future directions to exclude non-relevant parts of the data in the context of denoising and to compare results to currently used methods such as the Wiener optimum filters.