Adaptive multiple subtraction by statistical curvelet matching

In this paper, we present a novel algorithm that performs adaptive subtraction of a multiple model in the curvelet domain. The algorithm is based on the observation that the predicted model is an imperfect estimate of the actual multiples, containing two types of error: 1. a systematic error that manifests as an approximately constant phase and amplitude error within each subband and 2. a localized error that potentially varies fromcoefficient to coefficient. Adaptation of the model is automatically controlled by parameters provided by astatistical modelling step. Results show that the algorithmworks well with different types of multiples and levels of noise.