Design of two-dimensional randomized sampling schemes for curvelet-based sparsity-promoting seismic data recovery

The tasks of sampling, compression and reconstruction are very common and often necessary in seismic data processing due to the large size of seismic data. Curvelet-based Recovery by Sparsity-promoting Inversion, motivated by the newly developed theory of compressive sensing, is among the best recovery strategies for seismic data. The incomplete data input to this curvelet-based recovery is determined by randomized sampling of the original complete data. Unlike usual regular undersampling, randomized sampling can convert aliases to easy-to-eliminate noise, thus facilitating the process of reconstruction of the complete data from the incomplete data. Randomized sampling methods such as jittered sampling have been developed in the past that are suitable for curvelet-based recovery, however most have only been applied to sampling in one dimension. Considering that seismic datasets are usually higher dimensional and extremely large, in the present paper, we extend the 1D version of jittered sampling to two dimensions, both with underlying Cartesian and hexagonal grids. We also study separable and non-separable two dimensional jittered sampling, the former referring to the Kronecker product of two one-dimensional jittered samplings. These different categories of jittered sampling are compared against one another in terms of signal-to-noise ratio and visual quality, from which we find that jittered hexagonal sampling is better than jittered Cartesian sampling, while fully non-separable jittered sampling is better than separable sampling. Because in the image processing and computer graphics literature, sampling patterns with blue-noise spectra are found to be ideal to avoid aliasing, we also introduce two other randomized sampling methods, possessing sampling spectra with beneficial blue noise characteristics, Poisson Disk sampling and Farthest Point sampling. We compare these methods, and apply the introduced sampling methodologies to higher dimensional curvelet-based reconstruction. These sampling schemes are shown to lead to better results from CRSI compared to the other more traditional sampling protocols, e.g. regular subsampling.