数据驱动的块排列正则化多道叠前地震反演

Data-driven multichannel pre-stack seismic inversion via patch-ordering regularization

叠前地震反演充分利用地震记录振幅随偏移距变化的信息,在储层预测中发挥着重要作用.由于地震观测数据的带限性和其中所含有的噪声,地震反演是一个典型的不适定问题.目前,提出的各种正则化方法通过对反演参数施加约束,使反演结果具备期望的特征.然而,传统的反演方法忽略了地下构造的空间连续性,当地震数据信噪比较低时,单道反演方法无法得到稳定、准确的反演结果.本文提出了一种基于自适应块排列正则化的多道叠前地震反演方法,利用叠后地震资料高信噪比的特点,提取了一个块排列矩阵来记录地下构造信息.该方法通过将地震剖面分解为地震记录块,利用地震数据的局部相似性来记录相邻采样点的位置.在此基础上,构造了多道块排列正则项,建立了多道叠前地震反演的目标函数.在多道块排列正则项中,可以自适应调整权重系数,在增强空间连续性的同时保护边缘.利用L-BFGS算法可以有效地求解多道叠前反演目标函数.合成数据和实际数据测试均表明,当观测记录信噪比较低时,该方法的反演结果优于传统的基于单道模型约束的反演结果.

Prestack seismic inversion plays an important role in reservoir prediction,which takes full advantage of the Amplitude Variation with Offset(AVO)information.Due to the band-limited nature and noise presented in observed seismic data,seismic inversion is a typical ill-posed problem.Various regularization schemes are proposed to impose specified characteristics on the inverted parameters.However,traditional inversion approaches ignore the spatial continuity of the subsurface structures.When seismic record is contaminated with serious noise,single-channel inversion scheme cannot provide stable and accurate inversion results.This paper presents a multichannel pre-stack seismic inversion scheme via adaptive patch-ordering regularization.Taking advantage of the high signal-to-noise ratio of post-stack seismic data,a patch-ordering matrix is extracted to record the subsurface structure information.By decomposing seismic profile into patches,the patch-ordering matrix utilizes the local similarity of seismic data to record the locations of neighboring sampling points.Then,we construct the multichannel patch-ordering regularization and formulate the objective function for pre-stack seismic inversion.The patch-ordering regularization can adaptively adjust the weighted coefficients to enhance the spatial continuity and preserve edges.The multichannel pre-stack inversion objective function can be effectively solved by limited-memory BFGS algorithm.Both synthetic and field data tests demonstrate the superior performance of the patch-ordering regularized inversion scheme over conventional single-trace model-based inversion method especially for observed data with low signal to noise ratio.