Unsupervised Ground Roll Filtering on a Seismic Gather Using Generative Networks

Processing seismic data is a long, complex and iterative process, which is now practiced at Gazprom Neft. One of the stages of this process is the filtering of ground roll wave noise. Implementation of the filtering requires highly qualified geophysicists, who currently solve this problem by manually selecting combinations of various predetermined filters. On the other hand, generative adversarial networks have proved to be very effective for image processing: style transfer, random face generation, etc. The company has been dealing with this task for a long time, so during this time it has accumulated a sample of data where manual separation of noises has been done. This data can be used as an automatic hypothesis testing for new filtering methods, as well as for training neural networks. This project will show how to make an automatic system of metrics for hypothesis testing, as well as consider the applicability of the generative networks approach, and present a baseline solution based on heuristic rules.