Deep learning for improving the image quality of elastic seismic data

Abstract Elastic reverse time migration is currently an advanced imaging method in the petroleum industry. Generally, we use a finite difference algorithm to discretize forward-propagated wavefield and the back-propagated wavefield , then obtain the final image profile through cross-correlation imaging conditions . To overcome the numerical dispersion of elastic reverse time migration, one all-new strategy proposed in our letter, we design a kind of residual learning network of deep CNN for the seismic date to correct the dispersion of lower order finite difference operator, that is, both for seismic receiver data and wavefields data . In the process of elastic reverse time migration, we modify the numerical dispersion of forward-propagated wavefield, back-propagated wavefield and receiver data respectively, and then calculate the cross-correlation to obtain a more accurate image of underground structure . Residual learning and bath normalization were used to improve the trained seismic network . The test verified that our algorithm is more accurate when compared with the original finite difference method.