Incorporating estimates of data covariance in elastic FWI to combat random and correlated noise

Abstract

As field data applications of FWI increase, dealing with both random and coherent noise in seismic data, and the artifacts they create in FWI models, becomes increasingly important; noise suppression or estimation is also increasingly important when we transition to elastic multiparameter inversion from acoustic approximations. In this study, we analyzed the influence of random and correlated noise on the estimation of model parameter V p, Vs, and density, and, to mitigate their influence, we adopted a two-stage inversion approach, whose second stage involves a modified FWI misfit. The data covariance matrix is calculated from data residuals obtained from an initial run of FWI, and this is incorporated into the misfit function for a second run. With the elastic FWI conducted in the frequency domain, and the data covariance matrix consequently calculated frequency by frequency, the approach, though not computationally inexpensive, places reasonable demands on memory and storage. Random and correlated noise were examined and estimated, and inversion results were compared with those of otherwise identical conventional FWI runs. The bootstrap approach to inclusion of data covariance estimates in FWI appears to be stable, and to have a strong positive impact especially for correlated data noise.