Introduction

Multiple reflections often erect the highest barrier to the successful imaging and interpretation of marine seismic data. Despite their nuisance, however, multiples illuminate the prospect zone, and moreover, illuminate different angular ranges and reflection points than primaries. In theory and in practice, multiples provide subsurface information that primaries do not. To use the information in the multiples, we must map multiples and primaries to a domain where they are comparable, like a prestack image domain, and then combine them.

The semi-independent measurements provided by primaries and multiples overlap each other in one data record. In theory, averaging the multiple and primary images can improve signal fidelity and fill coverage gaps, but this encounters two problems in practice. First, unless we correct multiples for their different raypaths and additional reflections, the signal events are incomparable. Secondly, just as multiples represent noise on the primary image, primaries and higher order multiples represent noise on a first-order multiple image. Corresponding noise, or ``crosstalk'' events on the images are often kinematically consistent, so adding the images may actually degrade signal fidelity. Prestack separation of multiples and primaries is a prerequisite to image averaging.

I presented the LSJIMP method to simultaneously solve the separation and integration problems, as a global inversion Brown (2003b). The model space is a collection of images, with the energy from each mode partitioned into only one image. The forward model contains amplitude corrections which ensure that the signal events in the multiple and primary images are directly comparable, in terms of both kinematics and amplitudes. I presented an efficient kinematic Brown (2003c) and amplitude Brown (2003a) modeling scheme which combines the efficiency necessary to realistically run on real 3-D data with the accuracy to model pegleg multiples in a moderately complex earth. Three model regularization operators exploit multiplicity between and within the images to discriminate between crosstalk and signal, combine the images, fill coverage gaps, and increase signal fidelity.

I previously applied LSJIMP to a difficult 2-D Gulf of Mexico data example Brown (2003b). In this paper, I apply LSJIMP to a 3-D data example, also from the Gulf of Mexico. CGG donated the data, which were acquired in Green Canyon. I demonstrate that LSJIMP can cleanly separate the multiples and primaries in the data, in spite of sparse data geometry that inhibits many other multiple suppression techniques. I show that LSJIMP compares favorably to least-squares Radon demultiple, both in terms of computational performance and result quality. Last, I show that the LSJIMP estimated primaries are more robust with respect to amplitude-versus-offset (AVO) analysis than the raw data.