基于粗-细网络模型分步训练的地震数据重建方法

Seismic Data Reconstruction Method Based on Coarse-Refine Network Model with Stepwise Training

由于地形等复杂条件的限制,叠前地震数据在空间上存在不完整或不规则分布的情况,导致数据出现缺失或混淆等现象。近年来,基于卷积神经网络的方法已经广泛应用于缺失地震数据重建工作。然而一步训练过程的网络模型不足以重建具有宽振幅范围的缺失地震数据,低振幅缺失部分的重建结果仍需改进。因此本文提出一种具有分步训练过程的粗-细网络模型。该模型由粗网络和细网络组成,分步恢复宽振幅范围内的缺失地震数据。在细网络中引入离散小波变换代替池化操作,其可逆性在上采样阶段有利于保留细节特征。模型采用混合损失函数重建缺失信号的真实细节。粗网络的初步恢复结果经过掩码操作处理后输入到细网络,细网络进一步精确恢复缺失部分的低振幅信号。实验结果表明,与残差网络(ResNet)、U型网络(U-Net)和多级小波卷积神经网络(MWCNN)的重建方法相比,本文方法在合成数据和真实数据上展现出更卓越的重建性能:在缺失75%的合成数据上,信噪比为18.818 5 dB;在缺失50%的真实数据上,信噪比为12.2551 dB。在消融研究中,本文模型重建的均方误差为1.689 3×10^(-4),信噪比为19.284 6 dB,峰值信噪比为43.743 5 dB,结构相似性为0.984 1,均优于其他三组对照实验。

Due to the limitation of complex conditions such as topography,the pre-stack seismic data are spatially incomplete or irregularly distributed,resulting in phenomena such as missing or confusing data.In recent years,methods based on convolutional neural networks have been widely used in the reconstruction of missing seismic data.However,the network model of one-step training process is not enough to reconstruct the missing seismic data with a wide amplitude range,and the reconstruction results of the low-amplitude missing part still need to be improved.Therefore,a coarse-fine network model with a stepwise training process is proposed in this paper.The model consists of a coarse network and a fine network to recover the missing seismic data with a wide amplitude range in a step-by-step process.Discrete wavelet transform is introduced in the fine network instead of pooling operation,and its reversibility facilitates the preservation of detailed features in the up-sampling stage.Using a hybrid loss function,the model reconstructs the true details of the missing signals.The preliminary recovery results of the coarse network are processed by masking operation and input to the fine network,which further accurately recovers the low amplitude signal of the missing part.The experimental results show that compared with the reconstruction methods of residual network(ResNet),U-shaped network(U-Net)and multilevel wavelet convolutional neural network(MWCNN),the method in this paper demonstrates superior reconstruction performance on both synthetic and real data:the signal-to-noise ratio is 18.8185 dB on synthetic data with 75%missing,and 12.2551 dB on real data with 50%missing.In the ablation study,the mean square error of the model reconstruction in this paper is 1.6893×10^(-4),the signal-to-noise ratio is 19.2846 dB,the peak signal-to-noise ratio is 43.7435 dB,and the structural similarity index is 0.9841,all of which are better than the other three sets of control experiments.