A deep learning-based method for enhancing isotropic reverse time migration in complex media

In the field of geophysical exploration, reverse time migration (RTM) stands out as an effective seismic imaging technique, offering significant advantages in imaging complex geological structures. However, the seismic data collected in most cases of exploration contain complex geological anisotropy. Employing isotropic RTM methods for processing anisotropic seismic data may result in various issues, including artifacts and inaccuracies in structural imaging. We develop a convolutional neural network (CNN) model that improves isotropic RTM results by learning the results of anisotropic RTM, and the proposed U-net network with ResNet and SmoothL1 loss function can combine the advantages of the two migration methods. The input of the neural network is acoustic isotropic RTM images, and the label is the results of anisotropic RTM based on the tilted transversely isotropic (TTI) acoustic first-order velocity-stress equations. Validation and testing of complex models such as Marmousi model and SEG overthrust model have shown that the trained network effectively improves the imaging quality of isotropic RTM especially for dip structures and suppresses artifacts such as those caused by incomplete convergence of diffraction waves. The application of our CNN model to process isotropic RTM images produces enhanced results, with lower computational burden and implementation difficulty compared to anisotropic RTM methods.