Full-waveform inversion (FWI) is a technique that seeks to find a high-resolution high-fidelity
model of the Earth's subsurface that is capable of matching individual seismic
waveforms, within an original raw field dataset, trace by trace. The method begins from a
best-guess starting model, which is then iteratively improved using a sequence of linearized
local inversions to solve a fully non-linear problem. In principle, FWI can be used to recover
any physical property that has an influence upon the seismic wavefield, but in practice the
technique has been used predominantly to recover P-wave velocity, and this is the route that
is followed here. Full-waveform tomographic techniques seek to determine a highly resolved
quantitative model of the sub-surface that will ultimately be able to explain the entire seismic
wavefield including those phases that conventional processing and migration seek to remove
such as refracted arrivals.
Although the underlying theory of FWI is well established, its practical application to 3D land
data, and especially to seismic data that have been acquired using vibrators, in a form that is
effective and robust, is still a subject of intense research. In this study, 2D and 3D FWI
techniques have been applied to a vibrator dataset from onshore Oman. Both the raw dataset
and the subsurface model cause difficulties for FWI. In particular, the data are noisy, have
weak early arrivals, are strongly elastic, and especially are lacking in low-frequency content.
The Earth model appears to contain shallow low-velocity layers, and these compromise the
use of first-arrival travel-time tomography for the generation of a starting velocity model.
The 2D results show good recovery of the shallow part of the velocity models. The results
show a low-velocity layer that extends across the velocity model, but lacking in a high-resolution
image due to the absence of the third dimension. The seismograms of the final
inversion models give a good comparison with the field data and produce a reasonably high
correlation coefficient compared to the starting model.
An inversion scheme has been developed in this study in which only data from the shorter
offsets are initially inverted since these represent the subset of the data that is not cycle
skipped. The offset range is then gradually extended as the model improves. The final 3D
model contains a strongly developed low-velocity layer in the shallow section. The results
from this inversion appear to match p-wave logs from a shallow drill hole, better flatten the
gathers, and better stack and migrate the reflection data. The inversion scheme is generic, and
should have applications to other similar difficult datasets.