AN APPROXIMATE MESSAGE PASSING APPROACH FOR TENSOR-BASED SEISMIC DATA INTERPOLATION WITH RANDOMLY MISSING TRACES

In this paper, we consider the reconstruction of a high-dimensional seismic volume with randomly missing traces. Seismic data in the frequency-space domain are represented via a high-order tensor. Applying the parallel matrix factorization model to the underlying seismic tensor, we propose an iterative approximate message passing (AMP) approach to seismic data interpolation based on loopy belief propagation. In particular, we extend the bilinear generalized AMP (BiG-AMP) approach to incorporate parallel low-rank matrix factorizations by using a "turbo" framework, enabling iterative message passing between the subgraphs of the all mode unfoldings of the seismic tensor. The computational complexity of our algorithmic framework is low and scales linearly with the data size. Simulation results with synthetic seismic data suggest that the proposed algorithm yields better reconstruction performances relative to existing methods.