Efficient 3D Semantic Segmentation of Seismic Images using Orthogonal Planes 2D Convolutional Neural Networks

摘要

Technological advances in oil and gas reservoir characterization such as 3D seismics and seismic attributes enriched the subsurface’s description made by specialists. Nevertheless, the analysis of this now huge volume of data became a complex task. This work explores the use of 2D orthogonal planes convolutional neural networks for 3D seismic cube facies classification, one of the steps of reservoir characterization. Through a sampling method that captures spatial information of seismic data, the proposed model were applied in both synthetic data of the Stanford VI-E reservoir and in a benchmark based on the F3 block, which is part of a real reservoir. Compared to other models in the same benchmark, the classifiers produced here had superior results, with over 88% in pixel accuracy and 90% class accuracy on some instances. The sampling method is also flexible to use in practical cases.