The interval velocity estimation problem can be formulated as a least-square optimization scheme. With this formulation, we search for the velocity parameter that best fits equation (1) considering several points of two adjacent reflections. To avoid division by zero, and to increase the contribution from large horizontal slownesses (where the velocity information is better resolved), I define the objective function as follows:

(2) |

(3) |

Here, equation (3) is implemented, with the help of an
event-picking algorithm (Zhang, 1991) to define the several
reflection curves. The picking algorithm defines the relative
time-shifts between samples of adjacent traces.
This information is used to trace events in the
*x*-*t* domain starting at all samples of the first trace.
Because the events should coincide with the true reflections
wherever they are strong enough,
the algorithm computes the stacking-power (or semblance)
for each event to select the best candidates
for reflections. Figure shows synthetic data generated
from a velocity model (with moderate velocity contrasts to avoid
strong multiples contamination)
overlaid with the six selected picked events. The selection was based
entirely on the stacking power, but some interactive procedure will
be required when strong multiples or other coherent noises are present.

datapk
Synthetic data overlaid with automatically-picked
reflection events.
Figure 3 |

This algorithm also uses the picked reflection-curves to
compute also the horizontal slownesses at the sampled offsets.
For any two pairs of adjacent reflections,
the terms inside the sum of equation (3) are defined by
using the sampled points of the bottom reflection as reference.
For each of these points the algorithm finds the
matching point in the upper curve, that is, the point
with the same horizontal slowness as that of the reference
point in the bottom curve. To find the matching point, the
algorithm uses a spline interpolation of the
inverse function *x*_{i}=*x*_{i}(*p*_{i}) (defined for each sampled point *i*
in the upper curve).
Figure *a* shows the functions *p*(*x*) corresponding
to all six selected events shown in Figure . Clearly
the horizontal slowness evaluated with this method is very sensitive
to even small errors in the *t*_{i}(*x*_{i}) picked curve associated with each
reflection. This sensitivity requires the *p*-curves
to be filtered before the velocity estimation step. First, the
values of *p* that substantially differ from the local mean are
removed. Then the gaps are filled by interpolation, and
convolutional smoothing is applied. Figure *b* shows
the resulting filtered version of Figure *a*.

Figure 4

It is worth mentioning that the horizontal slowness curves *p*(*x*) need to
have a monotonic behavior, so that the inverse functions *x*(*p*)
can be uniquely defined. In principle, this condition should always
be satisfied when the plane-layered assumption is valid, but the
intervention of noise may result in a nonmonotonic *p*(*x*).

After the points sharing the same horizontal slownesses are matched, application of equation (3) is straightforward. Both the resulting estimated model and the true model are represented in Figure . Although the relative errors in the estimated velocities are not large, we should expect to achieve much worse results when dealing with real data because of the high sensitivity of the picked-based horizontal slowness evaluation used in this algorithm.

Figure 5

12/18/1997