Uncertainty estimation in AVO inversion using Bayesian dropout based deep learning

Abstract

Amplitude versus offset (AVO) inversion is the process of transforming seismic reflection into elastic properties such as P- and S- impedance to estimate the interval properties and thickness of underlying geology using well log and post- or prestack seismic data. Recent applications of AVO inversion based on deep learning have shown excellent results and practical applicability. However, traditional deep learning methods yield only prediction results without any associated predictive uncertainty. Two types of predictive uncertainty should be considered: aleatoric uncertainty, which occurs when noisy data are included; and epistemic uncertainty, which is caused by a lack of data. To estimate the impedances and their uncertainties, Bayesian approximation using Monte Carlo dropout is applied, which simply approximates a Bayesian neural network. From the proposed method, we can not only predict impedances but also estimate their predictive uncertainties in the seismic survey area and determine whether prediction results are reliable.