Registration-Based Guided Least-Squares Waveform Inversion

Full waveform inversion with frequency sweeping cannot start from zero frequency because of the lack of low-frequency data, requiring a good starting model. We study a different iterative scheme where the notion of proximity of two traces is not the usual least-squares distance, but instead involves registration as in image processing. In order to create transported data, we propose a nonconvex optimization problem and solve it in a multiscale fashion from low to high frequencies. This process requires defining low-frequency augmented signals in order to seed the frequency sweep at zero frequency. Successful registrations of noisy data, and application of the new method to model velocity estimation are demonstrated. In a crosshole seismic inversion example (transmission setting), we show that the new method decreases the model velocity error while conventional least-squares inversion converges to a spurious model.