Waveform Tomography: theory and practice

R. Gerhard Pratt, Queen’s University, Kingston, Ontario, Canada.

Abstract is generally impossible to proceed with wave-

form tomography without first carrying out
Waveform tomography uses the original
traveltime tomography in order to establish
waveform recordings and wave equation mod-
a starting model.
elling to extract high resolution tomographic
images from seismic data. Theoretical and In this presentation I will review the funda-
synthetic studies support the conclusion that mental issues of scale that control the resolu-
waveform tomography can potentially image tion of traveltime methods, I will briefly state
features at sub-wavelength scales. the physical and numerical principles used
in the development of waveform tomography
Introduction and I will provide some synthetic studies that
illustrate both the advantages and potential
In this talk I will review the theory and prac- pitfalls in waveform tomography. I will con-
tice of Waveform Tomography (also referred clude with a brief review of a number of cross-
to as Full Waveform Inversion or Diffraction hole tomography case studies in which wave-
Tomography). Many current tomographic form tomography has made a demonstrable
methods use the arrival time information, ex- contribution to the state of geological knowl-
tracted from the original waveform record- edge of the target.
ings either by manual or automatic, cross-
correlation based methods; I refer to these Scales, wavelengths and Fresnel zones
methods generically as traveltime tomography
methods. Waveform tomography differs from The characteristic length scale for a seismic
more common methods of tomographic inver- wave is its wavelength, λ related to the fre-
sion in two major aspects: quency, f by the fundamental relationship
c
λ= , (1)
• The input data consist of the seismic f
waveforms themselves (as opposed to
traveltimes, amplitudes or some other in which c is the propagation velocity of the
secondary attribute of the recorded data) wave. The general rule is that the high fre-
quencies are lost most rapidly to friction dur-
• The underlying numerical method is ing propagation. Thus, we never seem to have
based on the full wave equation (as op- small enough wavelengths: the further the
posed to a ray approximation or a Born waves travel, the less short wavelength energy
approximation) there is. Wavelength of course is a critical
factor in understanding the spatial resolution
The features above make waveform tomogra- of a given method; we generally accept from
phy less approximate and consequently bet- the outset that the tomographic images we
ter resolved than traveltime tomography. The obtain will be limited by the wavelength in
same features also make algorithms for wave- some way.
form tomography more difficult to develop, A second, less widely appreciated control
less robust and more expensive to use than on the resolution of tomographic methods is
traveltime tomography. As we shall see, it propagation distance. The further a seis-
mic wave travels after encountering a veloc-
ity anomaly, the less evident the effect of the
anomaly on the wavefronts. Due to diffrac-
tion effects, the wavefronts tend to “heal”,
gradually erasing the effect of the original dis-
turbance. The effect of wavefront healing is
that a second scale parameter, the width of
the first Fresnel zone also plays a role in lim-
iting the resolution. This scale parameter is
given by √ kernels in either equation (3) or (4).
λL, (2)
where L is the propagation distance the wave
travels from source to receiver. The Fresnel
zone can be thought of as defining an effec-
tive “ray-width”, implying that structure far
away from the ray path can still influence the
wavefronts (e.g., Nolet 1987). The influence
of this effect on the resolution of traveltime
tomography has been evaluated theoretically
by Williamson (1991) and Schuster (1996),
and further tested numerically by Williamson
and Worthington (1993).
Woodward (1992). The similarity of these
two linear integral equations suggests a uni-
fied discrete representation
d = A m, (5)
where d are the measured data (either trav-
eltime anomalies or waveform anomalies), m
are the model parameters, and the matrix A
is a discrete representation of the integration
Inversion of the linear system in equation
(5) is carried out by iterative methods, of
which the conjugate gradient method is one
of the most effective. In the conjugate gradi-
ent method repeated iterations of the form
m̂k = m̂k−1 + αk−1 AT δd
are carried out, where αk is the step length at
the k th iteration. The quantity AT δd has a
specific meaning in each of the two different
cases:

Mathematical physics of tomography 1. In traveltime tomography this calculates

the backprojection of the traveltime resid-
An interesting comparison of traveltime and uals.
waveform tomography can be made by con-
sidering the following two expressions: for 2. In waveform tomography this calculates
traveltime, ray-based methods we may ex- the backpropagation of the waveform
press the relationship between an anomaly in residuals.
the slowness field, δs(x) and the traveltime
anomaly δT (r, s) between a source at s and a Synthetic example
receiver at r as
Z I illustrate a test of the waveform tomogra-
δT (r, s) = d3 x s(x) Lo (x; r, s), (3) phy approach in Figure 1: A velocity model
containing a random spatial distribution of
where Lo is a thin, pencil-like path through velocity anomalies was computed numerically
the media representing the ray path. For (courtesy of Alan Levander and Colin Zelt);
waveform tomography the relationship be- synthetic waveform data were generated for
tween the anomaly in the squared slowness this model using a finite difference approach.
field and the “scattered wave”, Usc (r, s; ω) at The survey simulated a crosshole geometry,
a frequency, ω is given by with 101 sources and 101 receivers. Typi-
Z cal crosshole scale parameters were used: the
Usc (r, s; ω) = d3 x s2 (x) Lo (x; r, s), (4) dominant frequency of the experiment was
approximately 300 Hz, leading to a dominant
where Lo is a Fresnel-like volume connecting wavelength of approximately 10 m. Wave-
source to receiver, referred to as a wavepath lengths and the first Fresnel zone width are
shown in Figure 1. Traveltimes were picked seismic tomography: in Nolet, G. Ed.,
manually from the waveforms arising in this Seismic Tomography: D. Reidel publishing
model, and traveltime tomography using the company.
methods of Zelt and Barton (1998) was used
to create a starting velocity model (Figure 1, Pratt, R. G., and Shipp, R. M., 1999, Seis-
middle). mic waveform inversion in the frequency
domain, Part 2: Fault delineation in sed-
The resolution of the traveltime tomogram iments using crosshole data: Geophysics,
was somewhat better than the predicted Fres- 64, 902–914.
nel zone width, but much worse than the size
of the dominant wavelength. Waveform to- Pratt, R. G., Hou, F., Bauer, K., and We-
mography methods were then used to extract ber, M. H., Waveform tomography images
a velocity model (starting from the traveltime of velocity and inelastic attenuation from
result) that captured most of the small scale the Mallik 2002 crosshole seismic surveys:,
variation of the true model, at better than Submitted to the Mallik 2002 compendium
wavelength scales. volume, Geological Survey of Canada, 2004.

Crosshole examples Pratt, R. G., 1999, Seismic waveform in-

version in the frequency domain, Part 1:
Waveform tomography has been successfully Theory and vertification in a physical scale
applied to crosshole seismic waveforms in model: Geophysics, 64, 888–901.
a wide variety of geological environments.
Examples have been published from scale Schuster, G. T., 1996, Resolution limits for
model experiments (Pratt, 1999), sedimen- crosswell migration and traveltime tomog-
tary environments (Song et al., 1994; Pratt raphy: Geophys. J. Int., 127, 427–440.
and Shipp, 1999), and most recently from a Song, Z. M., Williamson, P. R., and Pratt,
sub-permafrost investigation of arctic gas hy- R. G., 1994, Full wave 2.5D frequency do-
drates (Pratt et al., 2004). The images in main inversion of real cross-hole data: Pre-
Figure 2 are an example from a site investi- sented at the 56th EAEG meeting in Vi-
gation study in crystalline rocks (Albert et enna, Austria.
al., 1999).
Williamson, P. R., and Worthington, M.,
Acknowledgments 1993, Resolution limits in ray tomography
due to wave behavior: Numerical experi-
Some of the synthetic examples were gener- ments: Geophysics, 58, 727–735.
ated while I was on sabbatical leave at Rice
University. The collaboration of Alan Levan- Williamson, P. R., 1991, A guide to the limits
der, Colin Zelt and Fuchan Gao in producing of resolution imposed by scattering in ray
these results is gratefully acknowledged. tomography: Geophysics, 56, 202–208.

References Woodward, M. J., 1992, Wave-equation to-

mography: Geophysics, 57, 15–26.
Albert, W., Bühnemann, J., Holliger, K., Zelt, C. A., and Barton, P. J., 1998, 3D seis-
Maurer, H. R., Pratt, R., and Stekl, I., mic refraction tomography: A comparison
Grimsel Test Site: Further development of two methods applied to dataq from the
of seismic tomography:, NAGRA Technical Faeroe Basin: Journal of Geophysical Re-
Report 97-05, Switzerland, 1999. search, 103, 7187–7210.
Nolet, 1987, Seismic wave propagation and
 Results with a random velocity model. Left: the “true” velocity model, middle: the traveltime
Fig. 1:
tomography results and right: the waveform tomography results.

(a) (b)

Data example from crystalline rock. Crosshole seismic surveys were carried out between three
Fig. 2:
horizontal boreholes in a site investigation survey for the storage of radioactive waste. a) is the
traveltime tomography image, and b) is the waveform tomography image.