结构张量引导的多阶可转向马尔可夫随机场先验波阻抗反演方法

A Structure-Tensor-Guided Multi-Order Steerable Markov Random Field Prior Method for Acoustic Impedance Inversion

摘要: 地震反演是油气勘探与开发中的关键技术之一，而波阻抗反演是利用地震数据、低频地质信息(如低频背景模型、趋势模型)反演出地下介质的波阻抗，从而为地层结构分析、岩性识别与界面追踪提供重要依据。然而，波阻抗反演往往是一个病态问题，易受到噪声、带限效应与初始模型不确定性的影响；同时传统的单道反演无法充分利用采样道之间的构造信息，容易产生横向连续性不足等现象，尤其在地层结构复杂的区域更加明显。因此，为了解决上述问题，我们提出了一种结构张量引导的多阶可转向马尔可夫随机场先验波阻抗反演方法。该方法在贝叶斯框架下将先验建模为各向异性马尔可夫随机场，通过提取采样点的局部主方向，并引入考虑不同距离的多阶邻域约束，构造出沿层与垂直于层的差分约束，以增强对层理连续性与构造边界的刻画能力，同时提高反演稳定性。实验结果表明：在相同设置下，本方法相较于其他基线方法在RMSE、NRMSE、相关系数等指标上取得更优结果，并且对噪声具有更强的鲁棒性。

Abstract: Seismic inversion is one of the key technologies in oil and gas exploration and development. Wave impedance inversion is a method that uses seismic data and low-frequency geological information (such as low-frequency background models and trend models) to invert the wave impedance of underground media, thereby providing important basis for stratigraphic structure analysis, lithology identification, and interface tracking. However, wave impedance inversion is often a pathological problem, which is easily affected by noise, band-limited effects, and uncertainty of the initial model; at the same time, traditional single-channel inversion cannot fully utilize the structural information between sampling channels, and is prone to phenomena such as insufficient lateral continuity, especially in areas with complex stratigraphic structures. Therefore, to solve these problems, we propose a structure tensor-guided multi-order steerable Markov random field prior wave impedance inversion method. This method models the prior in the Bayesian framework as an anisotropic Markov random field, by extracting the local principal direction of the sampling points and introducing multi-order neighborhood constraints considering different distances, constructing differential constraints along the layer and perpendicular to the layer, to enhance the ability to depict layer continuity and structural boundaries, and improve the inversion stability. Experimental results show that, under the same settings, this method achieves better results in terms of RMSE, NRMSE, correlation coefficient, etc., compared to other baseline methods, and has stronger robustness to noise.