On the Reduction of Noise-sensitivity in Seismic Tomography

The Iteratively Reweighted Least Squares (IRLS) procedure is combined with the Conjugate Gradient method in defining a robust tomography method (W-CGRAD). In order to make the Simulated Annealing (SA)-based tomography algorithm more robust, a generalized weighted norm is introduced, as generalized energy function. Robustified versions of the SIRT method are also introduced. The generalization is made both in the data space (weighted average using Cauchy weights) and in the model space. A robust SIRT algorithm can also be defined by introducing alpha-trimmed mean instead of the arithmetic mean in the well-known updating formula of the traditional SIRT method. Using synthetic traveltime data the generalized (robust) tomography algorithms are tested in a medium sized tomography example. It is proved that the weighted tomography methods are more robust and less sensitive for noise (especially outliers) then their original versions.