Regional-scale FWI of wide-angle OBN data from a crude initial model using graph-space optimal transport

Overview & Synopsis

Regional-scale imaging with full-waveform inversion of sparse longoffset ocean-bottom node data is inherently difficult. In particular, the large number of wavelengths that have to propagate within the volumetric model-space makes the inversion prone to cycle-skipping. While building an accurate initial model that predict the synthetic data with half-wavelet accuracy might be problematic due to the sparsity of the ocean-bottom nodes, solutions can be found from full-waveform inversion schemes based on alternative misfit functions. The recently developed graph-space optimal transport misfit function is designed for an improved convexity with respect to shifted patterns in the seismic signals-and therefore to the kinematic errors of the velocity model. In principles, this shall make it possible to successfully invert the data starting from a crude initial model. Here we show how using full-waveform inversion based on this misfit function, combined with multiscale data-selection strategy, we are able to make inversion converge from a simple 1D starting model. Despite the fact that the kinematic inaccuracy between the observed and the synthetic first-arrivals is reaching five cycles in the initial stage, we are able to bring the data in phase and obtain geologically consistent velocity model of the complex subduction zone. The presented approach can therefore significantly relax the constraint on the kinematic accuracy of the initial FWI model which is typically derived by traveltime tomography.