In a recent work (, ), I introduced the process of velocity continuation to describe a continuous transformation of seismic time-migrated images with a change of the migration velocity. Velocity continuation generalizes the ideas of residual migration (, ) and cascaded migrations (). In the zero-offset (post-stack) case, the velocity continuation process is governed by a partial differential equation in midpoint, time, and velocity coordinates, first discovered by (). () and () describe this process in a broader context of ``image waves''. Generalizations are possible for the non-zero offset (prestack) case (, ).
A numerical implementation of velocity continuation process provides an efficient method of scanning the velocity dimension in the search of an optimally focused image. The first implementations (, ) used an analogy with Claerbout's 15-degree depth extrapolation equation to construct a finite-difference scheme with an implicit unconditionally stable advancement in velocity. () presented an efficient three-dimensional generalization, applying the helix transform ().
A low-order finite-difference method is probably the most efficient numerical approach to this method, requiring the least work per velocity step. However, its accuracy is not optimal because of the well-known numerical dispersion effect. Figure shows impulse responses of post-stack velocity continuation for three impulses, computed by the second-order finite-difference method (). As expected from the residual migration theory (), continuation to a higher velocity (left plot) corresponds to migration with a residual velocity, and its impulse responses have an elliptical shape. Continuation to a smaller velocity (right plot in Figure ) corresponds to demigration (modeling), and its impulse responses have a hyperbolic shape. The dispersion artifacts are clearly visible in the figure.
In this paper, I explore the possibility of implementing a numerical velocity continuation by spectral methods. I adopted two different methods, comparable in efficiency with finite differences. The first method is a direct application of the Fast Fourier Transform (FFT) technique. The second method transforms the time grid to Chebyshev collocation points, which leads to an application of the Chebyshev- method (, , ), combined with an unconditionally stable implicit advancement in velocity. Both methods employ a transformation of the grid from time t to the squared time , which removes the dependence on t from the coefficients of the velocity continuation equation. Additionally, the Fourier transform in the space (midpoint) variable x takes care of the spatial dependencies. This transform is a major source of efficiency, because different wavenumber slices can be processed independently on a parallel computer before transforming them back to the physical space.