Adaptive data-selection elastic full-waveform inversion

The multiscale inversion strategy is widely used to mitigate the cycle-skipping problem in full waveform inversion. There are many different approaches to implement the multiscale inversion and the widely used low-to-high frequency continuation is not applicable when the observed data lack low frequencies. As an alternative to multiple frequencies, offsets continuation is also able to suppress the cycle-skipping problem when the initial model is not perfect. We improve the multiscale strategy of offset-selection by introducing a local similarity criterion. Thus, we formulate an adaptive data-driven selection process that is better than conventional offset continuation approaches. The global-crosscorrelation objective function used here aims to maximize the similarity of two data sets instead of subtracting one from another and it is more consistent with the selection strategy. Besides, the crosscorrelation-based objective function is more sensitive to the phase information of the data and thus is more applicable to field data. We use a modified elastic Marmousi example to verify the effectiveness of the proposed method.