Introduction

In ray-based reflection tomography, picking reflectors is an integral and painful part of the process Clapp (2001b); Kosloff et al. (1996); Stork (1992); van Trier (1990). The common methodology is to pick a series of reflectors from a migrated image. A set of rays are then calculated that reflect at the picked interfaces. A major problem is the human intensive nature of reflector picking, especially for 3-D data. Automatic pickers can help, but significant human quality control (QC) is still necessary. A high level of QCing is required because inaccurate reflector picks lead to inaccurate reflector dip estimates. These poor estimates cause information to be back projected to the wrong portion of the model space, seriously hampering the inversion.

Woodward et al. (1998) and Clapp (2001a) introduced methods to limit the amount of picking required by selecting back projection points based on criteria such as semblance and dip coherence. These methods are successful in reducing the human cost of tomography but have two significant weaknesses. First, moveout is often characterized by a single value. This value is obtained by scanning over a range of moveouts and then selecting the maximum. In areas with significant multiple or converted wave energy, they will often have trouble distinguishing primary reflections (signal) from multiple and converted wave reflections (noise). The second problem is that these automatic point selection methods are generally going to have a larger level of erroneous moveout descriptions that generally increase with depth. These erroneous moveouts will generally cause large residuals which can dominate the inversion procedure.

In this paper, I show three simple methods to combat both problems. Unreasonable moveouts can be avoided by scanning over a large range of moveouts but only selecting points whose maximum is in a narrower range of moveouts. Second, I show that we can account for a higher level of variance by adding a diagonal weight to our model styling goal. Finally, I show that the effect of the remaining non-primary events and other erroneous moveouts can be further diminished by using re-weighted least-squares Claerbout (1998) to simulate a L₁ inversion problem where noisy data points have less of an effect.