Non-linear primary-multiple separation with directional curvelet frames

Predictive multiple suppression methods consist of two main steps: a prediction step, during which multiples are predicted from seismic data, and a primary-multiple separation step, during which the predicted multiples are ’matched’ with the true multiples in the data and subsequently removed. This second separation step, which we will call the estimation step, is crucial in practice: an incorrect separation will cause residual multiple energy in the result or may lead to a distortion of the primaries, or both. To reduce these adverse effects, a new transformed-domain method is proposed where primaries and multiples are separated rather than matched. This separation is carried out on the basis of differences in the multiscale and multidirectional characteristics of these two signal components. Our method uses the curvelet transform, which maps multidimensional data volumes into almost orthogonal localized multidimensional prototype waveforms that vary in directional and spatio-temporal content. Primaries-only and multiples-only signal components are recovered from the total data volume by a non-linear optimization scheme that is stable under noisy input data. During the optimization, the two signal components are separated by enhancing sparseness (through weighted l1-norms) in the transformed domain subject to fitting the observed data as the sum of the separated components to within a user-defined tolerance level. Whenever, during the optimization, the estimates for the primaries in the transformed domain correlate with the predictions for the multiples, the recovery of the coefficients for the estimated primaries will be suppressed while for regions where the correlation is small the method seeks the sparsest set of coefficients that represent the estimation for the primaries. Our algorithm does not seek a matched filter and as such it differs fundamentally from traditional adaptive subtraction methods. The method derives its stability from the sparseness obtained by a non-parametric (i.e. not depending on a parametrized physical model) multiscale and multidirectional overcomplete signal representation. This sparsity serves as prior information and allows for a Bayesian interpretation of our method during which the log-likelihood function is minimized while the two signal components are assumed to be given by a superposition of prototype waveforms, drawn independently from a probability function that is weighted by the predicted primaries and multiples. In this paper, the predictions are based on the data-driven surface-related multiple elimination method. Synthetic and field data examples show a clean separation leading to a considerable improvement in multiple suppression compared to the conventional method of adaptive matched filtering. This improved separation translates into an improved stack.