A new look at autoregressive low frequency reconstruction of seismic data

This abstract proposes a novel, easy to deploy and computationally fast solution to reconstruct the low frequency content of seismic data for FWI application. This solution is needed when the SNR is very poor or when denoise fails to enhance sufficiently the low frequency content of the data. The method follows a signal processing approach and does not need to train an offline model on auxiliary data, as is the case for solutions based on machine learning. The reconstruction is done from the higher frequencies using a recursion filter which is estimated from the data itself. The novelty of the proposed method is that it transforms the data locally from the time-space domain to the time-slowness domain where the reconstruction is performed. This transformation enforces the time domain sparsity needed to justify the use of recursion modeling in the frequency domain and exploits the spatial coherency in the data. The method also implicitly uses the signal cone limits of the seismic wavefield to provide a physically constrained solution. The proposed method was tested on many datasets to condition it for FWI and proved to be successful to help the inversion to mitigate cycle skipping and to improve the final model.