Probabilistic Seismic Facies Classification

Seismic interpretation is a fundamental process in basin and reservoir scale assessments, however, the process is often time and labor intensive. Manual interpretation of seismic surveys is prone to error, subjective to the interpreter’s expertise and rarely comes with an associated measure of uncertainty.

Dense labeling (i.e. interpreting) of seismic observations based on seismic amplitudes, facies or structural patterns are commonly described as image segmentation tasks in computer vision. Deep convolutional neural networks have led to a number of recent advances in producing high-quality image segmentations given large datasets of densely labeled training examples. While high accuracies can be achieved in the big-data regime, these deep neural networks provide no measure of uncertainty in their evaluation. Moreover, in contrast to large image datasets, manually-created dense labels of seismic datasets typically only consist of a very limited number of training examples. Transfer learning, where a model is pre-trained on a related dataset with much more available data and is then adapted to the existing dataset by slow annealing of the weights of the neural network presents one solution to this problem. A second approach is to heavily regularize deep neural networks by dropping individual connections while training.

We use a previously developed approximation to Bayesian variational inference to train deep convolutional neural networks on densely labeled seismic interpretations. These Bayesian neural networks produce a per-pixel (or per data point) estimate of the model uncertainty given the model as well as the available training data and have been shown to perform favorably for small-data tasks. We show the ability of deep Bayesian neural networks to be trained on a small set of manually-interpreted dense labels of the Dutch F3 3D seismic dataset and highlight the inherent uncertainty in the predictions of the produced models. The associated uncertainty measures enable greater collaboration within subsurface teams and principled decision-making when evaluating and interpreting seismic data.