Extraction of Seismic Properties and Models From, and Full Waveform Inversion of, Dispersed Seismic Waves





Journal Title

Journal ISSN

Volume Title




Surface waves, which propagate along boundaries between two different media, play an important role in resolving geological structures of different scales targeted from global seismology, exploration seismology, geotechnical engineering, to nondestructive testing. Over the past half century or so, different methods have been explored to process and invert surface waves for underground model properties, especially the shear wave velocity. However, there are still many problems waiting to be solved. Conventional dispersion curve inversion (DCI) is limited to 1-D model assumption and has increased uncertainty when the structure is complicated. It also requires picking of dispersion curves from field data, which is often a labor intensive process. Although, methods in the framework of full waveform inversion of surface waves yield models with good resolution, both laterally and vertically when carefully implemented and applied, they are computationally intensive and can easily suffer from the cycle skipping problem. Wave-equation based dispersion curve inversion method combine some of the advantages of those in conventional dispersion curve inversion and full waveform inversion, but also requires picking of dispersion curves from both field and synthetic data. This dissertation focuses to partially solve some of the above issues and leads to more work that can be done in the future. To automate the picking of dispersion curves from surface waves, which is required for many approaches for shallow-subsurface characterization using surface waves, my first project presents a convolutional-neural-network (CNN) based machine learning approach to automatically pick the curves for the fundamental and higher modes along the two azimuths of any 2D seismic profile. Various attributes such as amplitudes, coherency, and local phase velocity as well as frequency and wavenumber of dispersion curves are derived; different sub-sets of these are tested in the CNN training process to assess the best combinations. We use a U-net architecture that is modified to convert the conventional 2D image segmentation problem in the (f,k) domain into direct multi-mode curve fitting and a subsequent picking process. To make the automatic picking algorithm more practical, we (1) introduce a second loss function that combines conventional wavenumber residuals and curve slope residuals; (2) use the transfer learning strategy, in which the network is pre-trained with synthetic data and then with a relatively small portion of the field data, to improve the efficiency of the algorithm; (3) evaluate two categories of uncertainty, the epistemic uncertainty from the method itself and input data, and uncertainty from non-deterministic factors such as random initialization of model weights and random shuffle of samples in training in the CNN, and in GPU parallelism. The epistemic uncertainty is an important indicator of the picking quality and can be used as a weighting of data in subsequent inversion; (4) perform post-processing to determine the effective dispersive frequency range of the picked curves by using different criteria, such as long/short moving average ratios (MAR) of squared picked wavenumbers, posterior uncertainty etc. The effectiveness of the automatic picking process is demonstrated in this study through applications to a field OBN dataset where different modes of Scholte waves were recorded. To reduce cycle skipping in FWI and to increase resolution of the estimated model, my second project develops and illustrates concurrent elastic full-waveform inversion (FWI) of P and S body waves and Rayleigh waves using interleaved envelope- and waveform- based misfit functions, in a gradually-increasing frequency, multi-scale, inversion strategy, to estimated both lateral and horizontal variations of models, which breaks the 1D assumption of conventional DCI. Computing correlation coefficients between the observed and predicted data, and between the inverted and correct models, provides quantitative measures of the composite contributions, of the starting model, the chosen data flow, and the depth extent of the solution space, to the fits of the corresponding solutions. Treating the whole wavefield as a single data set means that it is not necessary to separate, or even to identify, different types of body and surface waves.