Source-independent time-domain waveform inversion using convolved wavefields: Application to the encoded multi-source waveform inversion

Source-independent time-domain waveform inversion using convolved wavefields: Application to the encoded multi-source waveform inversion

Yunseok Choi and Tariq Alkhalifah, "Source-independent time-domain waveform inversion using convolved wavefields: Application to the encoded multisource waveform inversion", Geophysics 76 (2012): R125-R134. doi: 10.1190/geo2010-0210.1​​.​
Yunseok Choi, Tariq Alkhalifah
inversion, sources, wavelet, modeling, wave equation
2011
​Full waveform inversion requires a good estimation of the source wavelet to improve our chances of a successful inversion. This is especially true for an encoded multisource time-domain implementation, which, conventionally, requires separate-source modeling, as well as the Fourier transform of wavefields. As an alternative, we have developed the source-independent time-domain waveform inversion using convolved wavefields. Specifically, the misfit function consists of the convolution of the observed wavefields with a reference trace from the modeled wavefield, plus the convolution of the modeled wavefields with a reference trace from the observed wavefield. In this case, the source wavelet of the observed and the modeled wavefields are equally convolved with both terms in the misfit function, and thus, the effects of the source wavelets are eliminated. Furthermore, because the modeled wavefields play a role of low-pass filtering, the observed wavefields in the misfit function, the frequency-selection strategy from low to high can be easily adopted just by setting the maximum frequency of the source wavelet of the modeled wavefields; and thus, no filtering is required. The gradient of the misfit function is computed by back-propagating the new residual seismograms and applying the imaging condition, similar to reverse-time migration. In the synthetic data evaluations, our waveform inversion yields inverted models that are close to the true model, but demonstrates, as predicted, some limitations when random noise is added to the synthetic data. We also realized that an average of traces is a better choice for the reference trace than using a single trace.
(print): 0016-8033 (online): 1942-2156