SUMMARY: Full-waveform inversion has become an essential technique for mapping geophysical subsurface structures. However, proper uncertainty quantification is often lacking in current applications. In theory, uncertainty quantification is related to the inverse Hessian (or the posterior covariance matrix). Even for common geophysical inverse problems its calculation is beyond the computational and storage capacities of the largest high-performance computing systems. In this study, we amend the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm to perform uncertainty quantification for large-scale applications. For seismic inverse problems, the limited-memory BFGS (L-BFGS) method prevails as the most efficient quasi-Newton method. We aim to augment it further to obtain an approximate inverse Hessian for uncertainty quantification in FWI. To facilitate retrieval of the inverse Hessian, we combine BFGS (essentially a full-history L-BFGS) with randomized singular value decomposition to determine a low-rank approximation of the inverse Hessian. Setting the rank number equal to the number of iterations makes this solution efficient and memory-affordable even for large-scale problems. Furthermore, based on the Gauss–Newton method, we formulate different initial, diagonal Hessian matrices as pre-conditioners for the inverse scheme and compare their performances in elastic FWI applications. We highlight our approach with the elastic Marmousi benchmark model, demonstrating the applicability of pre-conditioned BFGS for large-scale FWI and uncertainty quantification.
Least-squares reverse time migration (LSRTM) is an iterative inversion algorithm for estimating the broadband-wavenumber reflectivity model. Although it produces superior results compared with conventional reverse time migration (RTM), LSRTM is computationally expensive. We have developed a one-step LSRTM method by considering the demigrated and observed data to design a deblurring preconditioner in the data domain using the Wiener filter. For the Wiener filtering, we further use a stabilized division algorithm via the Taylor expansion. The preconditioned observed data are then remigrated to obtain a deblurred image. The total cost of this method is about two RTMs. Through synthetic and real data experiments, we see that one-step LSRTM is able to enhance image resolution and balance source illumination at low computational costs.
4 June 2019 in IEEE Geoscience and Remote Sensing Letters
Least-squares reverse-time migration (LSRTM) distinctly improves seismic imaging quality, but at an expensive computation overhead involving tens of iterations. We herein take a computationally cheaper single-step LSRTM solution, which intrinsically performs deblurring through a data-domain Wiener deconvolution. However, the Wiener filter mainly solves the signal estimation problems for stationary signals. Subsalt imaging often suffers from strong salt-related reflections and artifacts. The former and the latter give rise to strong amplitude variance and changed source wavelets in the demigrated data, increasing its nonstationarity and hindering the data-deblurring operation in the single-step LSRTM. To alleviate the nonstationarity during the data-domain deblurring, we consider a Gabor deconvolution method. Testing on the Sigsbee data sets shows that the Gabor deconvolution method is effective, producing subsalt images of more balanced events and fewer artifacts than the raw RTM image. The Gabor deconvolution-related result also outperforms the standard single-step LSRTM result with more robust behavior and better subsalt imaging quality.
13 November 2019 in IEEE Transactions on Geoscience and Remote Sensing
Least-squares reverse time migration (LSRTM) is a powerful tool in seeking broadband-wavenumber reflectivity images. It produces better images over reverse-time migration (RTM) at the expense of computational cost. The Hessian effect can be measured in the image domain with the point-spread function (PSF). Here, we try to measure the Hessian effect in the data domain with the so-called trace-spread function (TSF). The difference between PSF and TSF is that the former originates from LTL in the image domain while the latter from LLT in the data domain. By comparing the TSFs with their original corresponding traces (or beams), we can design adaptive matching filters for preconditioning to alleviate the Hessian effect. However, the full TSF matrix is expensive. In this article, we propose a multiscale solution, which first has a diagonal approximation to LLT in beams, and then handle the full submatrix composed of the one-beam traces using the Sherman-Morrison formula. The preconditioned beams are superimposed into a ``deblurred'' data for remigration. Through synthetic and real data examples, we see that: 1) single-step data-domain LSRTM can yield deblurred RTM images via adaptive matching filters and 2) the beam-by-beam consideration outperforms the trace-by-trace one.