Probabilistic Seismic Interpretation Using Bayesian Neural Networks

Summary Recent progress in deep learning and especially convolutional neural networks has brought new advances in the automatic interpretation of seismic datasets. While these methods can achieve high accuracy in predicting seismic facies based on amplitude data alone, the deterministic predictions obtained from traditional neural networks do not reflect the uncertainty of the ground-truth labeled data as well as the diversity of possible interpretations. We present an application of deep Bayesian neural networks in the context of automatic seismic interpretation. Based on a dutch north-sea dataset we show that Bayesian neural networks can be used to obtain high accuracy interpretations of seismic facies with associated measures of predictive uncertainty. The estimates of mean predictive performance and variability are obtained by sampling a number of equally likely interpretations from the Bayesian neural network. We show that the ensemble of stochastic seismic interpretations can additionally be used for horizon extraction and computation of geo-body volume distributions.