基于全卷积神经网络的纵横波分解技术研究及其在弹性波成像中的应用

纵波(P)和横波(S)波场分解对弹性介质中的多分量地震波成像至关重要, 但是常规P-S波波场分解方法精度相对较低, 且存在成像假象的问题。为此, 构建了一种基于全卷积神经网络(FCN)的网络结构, 用于二维各向同性弹性介质地震波场的P-S波波场分解。该网络由全卷积神经网络构建, 使用合成波场快照进行训练, 训练完成的网络类似空间滤波器, 可实现高精度的P-S波波场分解。不同于基于傅里叶变换的P-S波波场分解方法, 该方法可以在波场任意空间位置处开展P-S波波场分解, 因此适用于面向目标的地震成像。合成数据的计算示例表明, 基于全卷积神经网络的纵横波波场分解方法可有效分解P波和S波波场, 且精度高于其他空间域分解方法。弹性波逆时偏移成像结果表明, 使用基于全卷积神经网络(FCN)的P-S波波场分解方法所获得的基于P波和S波的地震波成像结果, 可有效减少速度界面处的成像假象, 提高复杂地质条件下的多波成像精度。

P- and S-wave decomposition is essential for imaging multi-component seismic data in elastic media, but conventional decomposition methods suffer from low accuracy and imaging artifacts.A data-driven workflow is proposed to obtain a set of neural networks that are highly accurate and artifact-free for decomposing the P- and S-waves in two-dimensional (2D) isotropic elastic media.The neural networks are fully-convolutional neural networks (FCN) working as a spatial filter to decompose P- and S-waves with a high accuracy.Different from the P-S decomposition algorithms using the Fourier transform, the spatial filters are more flexible in decomposing P- and S-waves at any time step and at any spatial position, which makes this method suitable for target-oriented imaging.Snapshots of synthetic data show that the network-tuned spatial filters can decompose P- and S-waves with improved accuracy compared with other space-domain P-S decomposition methods.Elastic-wave reverse-time migration using P- and S-waves decomposed by the proposed algorithm shows reduced artifacts where there is a high velocity contrast.