An encoder–decoder deep neural network for binary segmentation of seismic facies

Abstract

To explore hydrocarbons, it is necessary to interpret seismic data to identify facies and geological features. Traditionally, this work is performed by visually choosing points representing the limits of seismic features and using a tool to infer the other limit points. This process requires much manual work and may leave some features aside, making the work less accurate. Recently, deep learning has shown promising results in image segmentation. Deep learning methods can help reduce the analysis time and dependence on geometry in facies segmentation to study geological areas. This work investigates the application of neural networks to identify lines that separate seismic facies geometry. In particular, we present a Deep Neural Network for Facies Segmentation (DNFS), which builds upon StNet and U-Net with different hyper-parameters and loss functions to obtain state-of-the-art results concerning facies segmentation. We combine two independent loss functions, namely cross-entropy and Jaccard loss, for training DNFS; together, these functions lead to better results for seismic facies segmentation. The input data comprises small segments of seismic images in the Project Netherlands Offshore F3 Block dataset. Our results demonstrate that it is possible to train an encoder–decoder architecture using a dataset for binary segmentation and a composite loss function and to offer accurate predictions with a neural network designed to be trained within only a few minutes.