One of the significant challenges in seismic interpretation is to accurately delineate subsurface features and quantify the uncertainty of the interpretation results due to the non-unique nature of seismic processing and imaging. Salt interpretation usually has limited resolution and relies upon an interpreter's experience with a limited set of geological concepts. In seismic interpretation, especially salt interpretation, researchers have focused on improving the accuracy of pixel predictions by developing various neural network architectures, such as Dense U-Net, Attention U-Net, Residual U-Net, etc. Studying uncertainty quantification of point predictions is important in assessing prediction quality. In this paper, we implemented Monte-Carlo dropout analysis in the variational inference setting with a Bayesian Neural network (BNN) to analyze the aleatoric and epistemic uncertainty of the salt classification. Our approach helps to analyze the posterior distribution from the variational inference and quantitively measure the range of predictive probability distribution.

No abstract