Application of an unbalanced optimal transport distance and a mixed L1/Wasserstein distance to full waveform inversion

Full waveform inversion (FWI) is an important and popular technique in subsurface Earth property estimation. In this paper, several improvements to the FWI methodology are developed and demonstrated with numerical examples, including a simple two-layer seismic velocity model, a cross borehole Camembert model and a surface seismic Marmousi model. We introduce an unbalanced optimal transport (UOT) distance with Kullback-Leibler divergence to replace the L2 distance in the FWI problem. Also, a mixed L1/Wasserstein distance is constructed that preserves the convex properties with respect to shift, dilation, and amplitude change operation. An entropy regularization approach and convolutional scaling algorithms are used to compute the distance and the gradient efficiently. Two strategies of normalization methods that transform the seismic signals into non-negative functions are discussed. The numerical examples are then presented at the end of the paper.