Reducing the influence of coherent noise on FWI with misfit modification

Luping Qu, Xin Fu, Kristopher A. Innanen

As field data applications of FWI increase in number and ambition, dealing with the presence of both random and coherent noise in seismic data, and the artifacts they create in FWI models, becomes increasingly important. Though noise suppression methods, including various filters and decompositions, applied before the inversion, can mitigate these to some extent, remnant noise still always exist in the processed data. In this study, we carried out a systematic study of the impact of noise on scalar acoustic FWI models and sought mitigation strategies. We found that while random noise with low SNR (<=20%) does not exert a strong influence, correlated noise of all types tends to produce a strong negative affect. Mitigation effort here is likely to pay dividends in model accuracy and reliability. We examine the effect of including the data covariance matrix into the misfit function. Through iterative processing, random and correlated noise, and their combination, can be estimated during FWI. As limited frequency bands contribute most of the signal, high frequency bands (>=35 Hz) were cut off. Considering accuracy and computational efficiency, the data can be resampled every two or three points in frequency domain. The intermediate results of misfit and gradient were calculated by individual frequency and then transformed back to time domain after interpolation. The misfits are then calculated by individual frequency and transformed back to the time domain after interpolation. Numerical testing is suggestivethat this approach improves on more conventional de-noising methods.