基于深度神经网络的地震强反射剥离方法

In reservoir prediction, it is often encountered that the weak reflection signal is submerged in the strong reflection, which is disadvantageous to accurately identify and describe reservoir structure. In this study, we propose a method to remove the strong seismic reflection using the deep neural networks to help detect weak reflection signals of reservoirs. In the framework of the convolution model, the proposed method first decomposes the strong reflection prediction problem into two optimization sub-problems: seismic wavelet prediction and strong reflection prediction, which are solved by AIDNN and U-Net, respectively. The mapping relationship between seismic data and strong reflection can be established directly through training, which avoids the artificial empirical parameter adjustment, and is fast in the calculation and suitable for massive seismic data processing. Tests on synthetic and real data show that the proposed method can predict and remove strong seismic reflection with good amplitude preservation and fidelity. Base on this approach we predict the distribution of sand bodies in reservoirs and achieve good results.