Seismic Data Reconstruction Using Generative Adversarial Neural Network

Seismic data reconstruction offers the opportunity to interpolate the missing traces and reconstruct regular sampled seismic data,increasing the accuracy of subsequent seismic data processing and interpretation.Conventional seismic data reconstruction algorithms are often based on the assumption of sparsity data and linear seismic events,which is frequently violated in practical applications.Besides,during reconstruction,different parameters need be tested to achieve the optimum output,which introduces additional workloads and decreases efficiency.With the development of artificial intelligence,deep neural networks have been applied widely.A generative adversarial network is a particular artificial neural network and has been used in image processing.This thesis proposes the use of the generative adversarial network to reconstruct seismic data,focusing on network construction and training,parameter testing and merits for real seismic data applications.This thesis designs a cGAN to reconstruct the seismic data.The generator is a U-Net with skip connections and the discriminator is a Patch GAN,which means that the input is divided into small patches with a certain size,and this is different from conventional convolutional neural network,which always uses the whole figure as the input.Using different patch sizes,the performance of the trained generator may be different.To improve the accuracy of reconstruction,a new method is proposed in this thesis,which adds L1 loss to the cGAN loss function.For different situations of data missing,either regularly or irregularly with a large or small gaps,it is important to use different training datasets.In this thesis,geological models are designed to generate training datasets for regular and irregular missing seismic data.Furthermore,two solutions are proposed to solve the gradient vanishing problem in the generator during training,including a conditional Wasserstein generative adversarial network(cWGAN)and a cGAN with Gaussian noise in the discriminator(cGAN-GN).The cWGAN uses the Wasserstein distance,instead of the divergence used in the literature previously,to calculate the difference between the generated data and the real data.And the cGAN-GN has a Gaussian-noise layer in the discriminator,which can fool the discriminator.The testing results show that both solutions can improve the reconstruction performance and the cGAN-GN performs better.In order to determine the best cGAN-GN for data reconstruction,it is necessary to test different patch sizes.For irregular missing seismic data,different training datasets with different missing rates are used to train the network in order to obtain the best training dataset.To reconstruct the missing seismic traces with big gaps,it is recommended to use the seismic data where there is no missing traces to train the network,which yields the best reconstruction results,compared with all other methods.Finally,using the cGAN to reconstruct the missing seismic data can avoid the assumption of data sparsity and linear seismic event,and during reconstruction the cGAN does not need to test different parameters to optimize the reconstructed data,which reduces workloads and improves efficiency.Furthermore,the cost of the cGAN is mainly from the training,and after training,the processing time is negligible.