Full waveform inversion based on dynamic data matching of convolutional wavefields

Abstract

Cycle skipping problem caused by the absent of low frequencies and inaccurate initial model makes full waveform inversion (FWI) deviate from the true model. A novel method is proposed to mitigate cycle skipping phenomenon by dynamic data matching which improves the matching of synthetic and observed events to regulate the updating of initial model in a correct direction. 1-dimentional (1-D) Gaussian convolutional kernels with different lengths are used to extract features of each time sample in each trace which represents the integrated properties of wavefield at different time ranges centered on each time sample. According to the minimum Euclidean distance of the features, the optimally matched pairs of time samples in the observed and synthetic trace can be found. A constraint evaluates the reliability of dynamic matching by attenuating the amplitude of synthetic data according to the values of traveltime differences between each pairs of optimally matched time samples is proposed to improve the accuracy of data matching. In addition, Gaussian kernels have the capability to extract features of time samples contaminated by strong noises accurately to improve the robustness of the propose method further. The selection scheme of optimal parameters is discussed and concluded to ensure the convergence of the proposed method. Numerical tests on Marmousi model verify the feasibility of the propose method. The proposed method provides a new approach to tackle the convergence problem of FWI when using the field seismic data.