Stochastic FWI on Wide-angle Land Data with Different Order of Approximation of the 2D Acoustic Wave Equation

We experience the application of a genetic algorithm driven full-waveform inversion (GA FWI) on two expanding spread records acquired in a forward-reverse configuration along a 2D seismic land profile. Maximum source to receiver offset reach up to 40 km for the forward record and up to 30 km for the reverse record. The FWI is performed in time domain and with the acoustic 2D approximation. The data area is characterized by rough topography and complicate near surface and the seismograms show a low S/N ratio. We test whether, given the poor data quality and using a simple data misfit computation based on waveform envelopes and L1norm, the order of approximation of the spatial derivatives can be reduced without losing anything significant with respect to higher order approximations. The GA FWI experiments consider only the direct and diving waves of the shot records and constitute part of a wider project in which GA FWI is applied to both the expanding spread and the standard data to estimate a low-resolution velocity model apt to be used as a starting model for gradient based FWI. It turns out that reducing the order of approximation in the spatial derivatives computation from the 4th to the 2nd order does not appreciably change the matching between observed and predicted data as well as the estimated velocity models, while the computing time is drastically reduced. Also, in such data conditions, the adoption of more sophisticated misfit functions does not seem to produce significant improvements in the results.