The seismic wavefield arises from interactions of a source wavefield with subsurface heterogeneities as the wavefield propagates through the earth. In a conventional seismic processing workflow, the reflected portion of the wavefield is enhanced at the expense of the rest of the wavefield. While this is useful, considerable information which is contained outside of the reflections is lost. To alleviate this issue, we propose a deep learning technique which aims to separate the wavefield into three of the wavefield components: reflections, diffractions, and noise. This technique involves first performing several data domain transformations on the input and applying these as input classes on a pixel-by-pixel basis to guide the neural network. A simple separation is performed to remove the reflections and noise, allowing the diffractions to be processed independently. This technique, called Multi-domain diffraction identification, gives a high standard classification of the diffractions in a fraction of the time and computational cost of plane-wave destruction. These diffractions have then been removed from the data and compared with separation results from plane-wave destruction, showing similarities in the diffractions and demonstrating the denoising capability of the method.

当波场在地球中传播时，地震波场产生于震源波场与地下非均质性的相互作用。在传统的地震处理工作流程中，波场的反射部分以波场的其余部分为代价而增强。虽然这很有用，但包含在反射之外的大量信息丢失了。为了缓解这个问题，我们提出了一种深度学习技术，旨在将波场分成三个波场分量：反射、衍射和噪声。该技术涉及首先对输入执行多个数据域转换，并将这些转换为逐个像素的输入类以引导神经网络。执行简单的分离以去除反射和噪声，允许独立处理衍射。这种称为多域衍射识别的技术在平面波破坏的时间和计算成本的一小部分内提供了高标准的衍射分类。然后将这些衍射从数据中去除，并与平面波破坏的分离结果进行比较，显示了衍射的相似性并证明了该方法的去噪能力。