Walkaway VSP data conditioning for FWI

Raul Cova, Kristopher A. Innanen, Marianne Rauch

Full waveform inversion (FWI) applications on land seismic data remain very limited. The presence of strong anelastic effects, near-surface heterogeneities, unknown source and receiver signature and poor signal-to-noise ratio, among other reasons, challenge the capabilities of most modelling and inversion algorithms. Here, we perform an elastic FWI using land VSP data acquired in a walkaway configuration. We pre-process the data with the intent of improving the signal-to-noise ratio and removing undesired anelastic effects. Elevation differences among source locations were accounted for by applying elevation static corrections. Signal-to-noise ratio was improved by using a predictive filter in the FX domain. Two datasets with different deconvolution conditions were generated. A deterministic deconvolution using the recorded downgoing wavefield was applied to one of the datasets to remove the source signature. Even though this process partially accounts for changes in the wavelet with depth, a single operator is used for all the events recorded on a given trace. For this reason, we also computed a Gabor deconvolution to account for non-stationarity in the source signature. Then, we performed an elastic FWI using a multi-scale approach, with four frequency bands (4-8 Hz, 4-12 Hz, 4-16 Hz and 4-20 Hz) and three different depth windows (250-1000 m, 750-2250 m and 2000-3500 m). The FWI performed on the data deconvolved with the deterministic operators converged toward a solution that was closer to the sonic logs available in the well. Despite providing a wider frequency spectrum, the FWI using the Gabor deconvolved data did not converge toward an optimal solution. A closer examination of the input data revealed that in addition to removing some of the multiples, the deterministic deconvolution resurfaced some downgoing S-wave events that were not evident before. Providing data with less complexity and enhancing prominent events provided us with a more robust initialization of the inversion problem.