Cosparse seismic data interpolation

Over the years we have investigated seismic data interpolation and redatuming algorithms rely on on the assumption that seismic records and images permit sparse approximations under certain representations, such as Curvelet coefficients. Recent findings have suggested that for redundant representations (of which Curvelet is an example), the analysis operator that maps the physical signal to coefficients may also play a crucial role in recovering data from incomplete observations. This insight elevates the significance of a question that often goes unaddressed: is it better for the transform-domain sparsity to be achieved through explicit construction of sparse representations (e.g., by thresholding of small transform-domain coefficients), or by demanding that the algorithm return physical signals which produces sparse coefficients when hit with the forward transform? Recent results show that the two approaches give rise to different solutions when the transform is redundant, and that the latter approach imposes a whole new class of constraints. In particular, the number of zero-valued coefficients given by the analysis operator acting on the signal, referred to as its "cosparsity", have an analogous role to the sparsity of the signal in terms of the coefficients. From this framework, a new reconstruction algorithm is proposed which may allow better reconstruction from subsampled signaled than what the sparsity assumption alone would predict. In this work we apply the new framework and algorithm to the case of seismic data interpolation under the curvelet domain, and show that it admits better reconstruction than some existing L1 sparsity-based methods derived from compressive sensing for a range of subsampling factors. We will also investigate different analysis operators and their impact on both sparsity and cosparsity-based algorithms.