Sparsity promoting seismic imaging and full-waveform inversion

This thesis will address the large computational costs of solving least-squares migration and full-waveform inversion problems. Least-squares seismic imaging and full-waveform inversion are seismic inversion techniques that require iterative minimizations of large least-squares misfit functions. Each iteration requires an evaluation of the Jacobian operator and its adjoint, both of which require two wave-equation solves for all sources, creating prohibitive computational costs. In order to reduce costs, we utilize randomized dimensionality reduction techniques, reducing the number of sources used during inversion. The randomized dimensionality reduction techniques create subsampling related artifacts, which we mitigate by using curvelet-domain sparsity-promoting inversion techniques. Our method conducts least-squares imaging at the approximate cost of one reverse-time migration with all sources, and computes the Gauss-Newton full-waveform inversion update at roughly the cost of one gradient update with all sources. Finally, during our research of the full-waveform inversion problem, we discovered that we can utilize our method as an alternative approach to add sparse constraints on the entire velocity model by imposing sparsity constraints on each model update separately, rather than regularizing the total velocity model as typically practiced. We also observed this alternative approach yields a faster decay of the residual and model error as a function of iterations. We provided empirical arguments why and when imposing sparsity on the updates can lead to improved full-waveform inversion results.