基于Fourier近似Wasserstein距离的隐式全波形反演方法OA

Full-waveform inversion(FWI),as a high-resolution subsurface imaging technique,still faces several challenges in practical applications,including missing low-frequency data,cycle skipping,and noise contamination.To address these issues,this study proposes an implicit FWI method based on the Fourier-approximated Wasserstein distance.At the objective function level,a low-frequency-enhanced objective function is constructed using a Fourier approximation,exploiting the bounded relationship between the Wasserstein distance and Fourier coefficients.By weighting and enhancing the low-wavenu-mber components of seismic data,the proposed objective function improves the recovery of background velocity,thereby ef-fectively mitigating the cycle-skipping problem.At the model reparameterization level,a neural network with sinusoidal acti-vation functions is introduced to provide an implicit representation of the velocity model.Its continuous high-dimensional mapping capability replaces the traditional discrete grid-based representation,increasing the degrees of freedom in model rep-resentation while implicitly introducing smoothness constraints into the inversion process through the parameterization.Three representative models-Marmousi,Marmousi2,and SEG Salt-are employed to evaluate the performance of the proposed method under various challenging conditions,including missing low frequencies,linear initial models,noisy data,and strong scatter-ing media.The results demonstrate that,even under complex scenarios such as poor initial models,low-frequency deficien-cy,noise interference,and strong scattering,the proposed method exhibits strong robustness and favorable convergence be-havior,effectively alleviating cycle skipping and significantly improving inversion accuracy.