Research On High Resolution Reconstruction Method Of Seismic Data Based On Generative Adversarial Network

In seismic exploration,the inversion method is used to obtain the information of rock stratum.However,the seismic information collected has some problems,such as low resolution and small quantity.In order to solve this problem,this dissertation studies the application of Generative Adversarial Network to increase the resolution of seismic data,introduces in detail the use of Feature Pyramid Network and Residual Unit Generative Adversarial Network to extract the characteristics of seismic data,and trains the network to generate high-resolution seismic data in turn.The traditional Generative Adversarial Network uses KL divergence and JS divergence to judge the similarity between data,but KL divergence and JS divergence are unstable,which will cause the gradient to disappear and make the Generator unable to provide effective gradient value.This dissertation studies the distance between real seismic data and generated data based on Wasserstein distance,and proves its effectiveness.The main works of this dissertation are as follows:Firstly,this dissertation mainly adopts the research method of Feature Pyramid Network,and uses multi-convolutional template and multi-path convolutional layer to take multi-level seismic data feature information.Then,the global residual and local residual are combined to iteratively and alternately train Generative Adversarial Network to improve the resolution of seismic data.Secondly,refering to the structure of WGAN(Wasserstein Generative Adversarial Network),Lipschitz constraint is performed by adding a penalty function to the loss function to complete the high-resolution reconstruction of seismic data based on multiscale convolutional kernel and multi-path residual Generative Adversarial Network.In terms of data analysis,20%,40% and 60% randomly sampled seismic data are input into WGAN for high-resolution reconstruction of seismic data.The Peak Signal-to-Noise Ratio,dominant frequency component and bandwidth analysis of seismic data,structural similarity,high-frequency proportion of two-dimensional Fourier Transform,Wassersterin distance value,cosine similarity and traditional data high-resolution reconstruction methods are used to comprehensively evaluate the effect of high-resolution reconstruction.It can be seen from the verification of the above indicators,the application of the modified Generative Adversarial Network used in this dissertation does play highresolution reconstruction effect on the seismic data set.