Summary
Full Waveform Inversion (FWI) has been regarded as an effective tool to build the velocity model for the following pre-stack depth migration. While traditional methods, which are built on the Born approximation, are initial model dependent. Introducing Transmission matrix (Tmatrix), which includes all orders of scattering effects, can avoid the initial model dependence. From the T-matrix to estimate the velocity perturbation, it requires matrix inversion which is always time consuming. In order to achieve that efficiently, previously we have proposed Inverse Thin-Slab Propagator (ITSP) which is suitable for smooth media, and we study domain decomposition strategy to estimate the velocity perturbation efficiently in this abstract. Numerical examples demonstrate the validity of the proposed method.
Introduction
As the development of seismic exploration and exploitation, it requires more and more accurate seismic processing technologies. Full waveform inversion (FWI) can provide accurate parameter distributions of the sub-surface media, while it is always time consuming and initial model dependent (Virieux and Operto, 2009). In order to improve the inversion efficiency, the GPU, phase encoding and source encoding technologies are adopted (Ben-Hadj-Ali et al., 2009; Luo et al., 2012). In order to weaken the initial model dependence, many authors have done much work. Bunks et al. (Bunks et al., 1995) proposed a multi-scale seismic waveform inversion strategy: the result from long scale seismic data is regarded as the initial model for the short scale seismic data which can weaken the initial model dependence. Shin and Cha (Shin and Cha, 2008; Shin and Cha, 2009) proposed Laplace domain and Laplace-Fourier domain waveform inversion strategy to provide initial model for the following FWI. Wu et al (Luo and Wu, 2015; Wu et al., 2014) proposed envelope inversion strategy, which can use the ultra low frequency components compared with source frequency band, to provide initial model for FWI. While these methods are all built on born approximation and the differences lie in the objective functions.