Wavelet Filter Based Low-frequency Data Reconstruction for Time Domain Full Waveform Inversion
Wavelet Filter Based Low-frequency Data Reconstruction for Time Domain Full Waveform Inversion
复制标题
DOI：
10.3997/2214-4609.201600643
复制
发表时间：
2016-05
期刊：
影响因子：
--
通讯作者：
P. Zhang;L. Han;Fengjiao Zhang;Y. Zhou
中科院分区：
文献类型：
--
作者：
P. Zhang;L. Han;Fengjiao Zhang;Y. Zhou
常规的全波形反演（FWI）通常采用局部优化算法来更新速度模型。因此，我们提供的初始模型应该足够好，以避免局部极小值。丰富的低频信息可以补偿初始模型的不准确性，并有助于避免跳周。本文提出了一种基于小波滤波的低频数据重构方法。采用快速迭代收缩保持算法（FISTA）提取地下脉冲响应，并将宽带脉冲响应与带限低频震源子波进行卷积，得到低频数据。卷积过程等效于使用低频源子波进行滤波。精度分析表明，重建数据能够满足FWI的要求。提出了一种新的多尺度时域全波形反演策略，即利用一系列低频重构数据作为观测数据。一个明显的优点是，对于重建的数据，小波是准确已知的，这降低了FWI的不确定性。该方法避免了震源子波的不确定性对反演结果的影响，避免了数据和子波的同时预处理。数值算例表明，该策略能有效地避免跳圈现象，并能收敛到一个坏的初始模型。
The conventional full waveform inversion (FWI) often uses local optimization algorithm to update velocity model. So the initial model we provide should be good enough to avoid local minimum. Abundant low-frequency information can compensate for the inaccuracy of the initial model and help to avoid cycle-skipping. In this paper, we proposed a wavelet filter based low-frequency data reconstruction method. We extracted the subsurface impulse responses using Fast Iterative Shrinkage-Thresholding Algorithm (FISTA), and convolved the broad-band impulse responses with band-limited low-frequency source wavelets to obtain low-frequency data. The convolution process is equivalent to filtering using low-frequency source wavelets. The accuracy analysis demonstrated that the reconstructed data can meet the requirement of FWI. We proposed a new strategy for multiscale time domain full waveform inversion (TDFWI), which using a series of low-frequency reconstructed data as observed data. One distinct advantage is that the wavelet is accurately known for the reconstructed data, which reduces the uncertainty of FWI. This strategy avoids the effect of source wavelets uncertainty on inversion results, and avoids the simultaneously pre-process of data and wavelets. Numerical example shows that our strategy can avoid cycle-skipping effectively and can converge on a bad initial model.