Moving Toward Direct DNN-Based Enhancement of 3D Pre-Stack Seismic Data

Abstract

Summary Pre-stack data enhancement with multidimensional stacking is indispensable part of modern data processing that very compute-intensive since multiple wavefront attributes need to be estimated on dense spatial/temporal grid. At the core of this demand are conventional local or global optimization techniques. We propose two alternative approaches based of artificial intelligence that can greatly reduce computational effort of estimation stage. First approach performs traditional computations on sparser grid and inpaints to dense grid using deep neural network (DNN) with partial convolution layers. Second approach is direct DNN-based attributes estimation from the pre-stack seismic data itself. Both methods incorporate multiparameter attributes by encoding them into RGB-images. On synthetic and real 3D data examples, we demonstrate, that application of these methods for seismic data enhancement using nonlinear beamforming can greatly speed up the computational time while maintaining similar quality of output data.