Off the grid tensor completion for seismic data interpolation

In this work, we extend our previous low-rank tensor techniques for seismic data interpolation to the “off-the-grid” case, wherein the data is not acquired on a regularly-spaced grid. This irregular sampling scheme causes the underlying tensor to become high-rank, thus preventing our previous Hierarchical Tucker tensor algorithms from successfully recovering the volume. Instead of applying these techniques naively, we introduce a fictitious, regular grid on which the volume exhibits low-rank behaviour and a resampling operator that interpolates the data from the regular, low-rank domain to the irregular sampling domain. By incorporating these changes in to our previous framework, we can interpolate irregularly sampled data volumes with costs determined primarily by the choice of resampling transform. In our case, we use a non-uniform Fast Fourier Transform, but other choices are possible.