Migration Deconvolution via Deep Learning

Abstract

Migration deconvolution is an image domain approach to least-squares migration, which is considered the state-of-the-art algorithm for obtaining seismic reflectivity models of the earth from seismic acquisition results. Seismic imaging is an active research field with the development over the last few years of several techniques that have mitigated imaging issues. Ongoing research aims to improve image resolution and thus provide a more reliable seismic amplitude for the interpreter. Migration deconvolution can be framed as an inverse problem in the image domain to mitigate image resolution problems and reduce migration artifacts. This paper presents a migration deconvolution method via deep learning based on the Hessian filter least-squares migration (HF-LSM) algorithm. The idea is to use deep learning techniques to model the inverse operator instead of directly estimating the inverse Hessian matrix. A data set is generated from a given velocity model by applying Born modeling to the migrated image, followed by application of the reverse time migration algorithm. The resultant data set is then used to train several neural network models. The networks learn the blurring operator that describes the image degradation due to the effects of acquisition geometry. Three different network topologies were developed to handle this problem: a simple fully convolutional neural network, a U-Net and a generative adversarial network. Our results show that the proposed approach provides images of higher resolution and superior quality than the traditional HF-LSM workflow.