Abstract
Missing seismic data is largely present problem in the world. Lack of seismic data usually occurs due to some form of natural obstacle or legislative prohibitions of seismic exploration. Restoration of seismic data would allow locating of new oil traps and reduce the risk of unsuccessful drillings. The approach is based on deep learning (image inpainting) techniques, which will be applied on inline and crossline sections of a given 3-d seismic cube, in order to restore missing parts of sections. The study was provided for non-commercial purpose for the aims of scientific research. Data used in our experiments comes from open source typical Western Siberia field. Our approach uses Generative Adversarial Networks (GANs) for completing missing parts of images (sections), based on known parts. Method can be used for restoration of arbitrarily-shaped missing parts of seismic cube, but also for extrapolation purposes. Metrics used for model evaluation are correlation coefficient and mean absolute percentage error (MAPE) between original and inpainted parts of data.
This paper applies modern approach from growing image inpainting field to restore missing data, even if it's irregularly-shaped and very large. Using very powerful GANs is what gives this model ability to learn difficult inpainting scenarios, but also implicates challenging and time-consuming training process. Accurate estimation of model performances in different scenarios provides an exact instruction manual for a geologist, which helps him to identify cases where our model should be applied.
An Efficient Seismic Data Acquisition Based on Compressed Sensing Architecture With Generative Adversarial Networks