地震绕射波弱信号U-net网络提取方法

绕射波携带大量小尺度非均匀地质体信息,对于提高地震勘探分辨率具有重要意义.绕射波能量远小于反射波,在地震记录中常被强反射波掩盖,因此分离并单独成像绕射波,为探测小尺度非均匀地质体的关键问题.传统绕射波分离方法受限于理论模型假设,对陡倾角反射波去除效果不佳,且易对绕射波造成损伤.基于经典编码-解码框架下的U-net网络和注意力机制,本文提出了一种绕射波智能分离方法,通过编码器自动提取地震数据中的绕射波特征,再由解码器恢复绕射波,从而隐性去除反射波.该方法作为端到端的机器学习,训练后的U-net网络可自适应地分离绕射波.本文通过数值模拟数据与实际数据构建训练数据集,利用训练后的U-net网络分离绕射波,并将结果偏移成像.数值模型测试和实际资料应用表明,融合了注意力机制的U-net网络能够有效压制反射波能量,保留绕射波动力学特征,克服了传统绕射波分离方法难以去除陡倾角反射的局限性,其提取的绕射波弱信号特征较为完整,能够进一步提高地震成像分辨率,在小尺度断裂刻画上具有优势.

Seismic diffraction extraction method using U-net for weak signals

Seismic diffractions carry extremely rich information associated with small-scale geologic inhomogeneities, which is of great significance to improve the resolution of seismic exploration. The energy of diffractions is much weaker than that of reflections, and diffractions are normally masked by strong reflections in the seismic data. Therefore, separation and imaging of diffractions are the key to depicting small-scale geological discontinuities. The traditional diffraction separation method has poor performance in the removing high-slope reflections and may harm diffractions interfered with reflections due to the limitation of theoretical model assumptions. Here, we propose a diffraction separation method based on the U-net in the classical encoder-decoder framework and the attention mechanism, which extracts the diffraction features in the seismic data through the encoder and then reconstructs the diffractions from these features through the decoder. The reflections are implicitly removed in this encode-decode process. As an end-to-end machine-learning approach, the trained U-net realizes adaptive diffraction separation. We train U-net based on synthetic and field data and use the trained U-net to separate diffractions subsequent to obtain diffraction image. The numerical test and field data application demonstrate that U-net incorporating attention mechanism can efficiently suppress the reflection energy and preserve the kinematic characteristic of diffraction. It overcomes the limitation of the traditional diffraction separation method difficult to remove the high-slope reflection and the extracted weak signal characteristics of diffraction are relatively complete, which can further improve the seismic imaging resolution and have advantages in small-scale fracture characterization.