Uncertainty estimation in impedance inversion using Bayesian deep learning

Abstract

Impedance inversion estimates interval property and thickness of underlying geology using seismic survey data. Recently there are lots of studies using deep learning algorithm for estimating elastic properties. However, traditional methods only produce simple prediction without uncertainty information. There are two kinds of uncertainty that may be of interest: aleatoric and epistemic uncertainty. Aleatoric uncertainty refers to the notation of randomness caused by noise in observed seismic data. Epistemic uncertainty refers to model uncertainty caused by lack of knowledge such as model uncertainty. In this paper, we estimate the aleatoric and epistemic uncertainty in the estimation of P-impedance, S-impedance, and density using Bayesian deep learning framework approximated by dropout. From the proposed method, we can estimate the uncertainty about predicting elastic properties and quantify how reliable the results are.