Seismic Facies Identification Using Generative Adversarial Networks

Seismic facies refer to a seismic reflection unit within a certain area that has different seismic attribute parameters from adjacent elements.It represents the lithological assemblage,stratigraphy and sedimentary characteristics of the sediments that produce its reflections,and seismic data can be divided into different seismic facies by dividing the amplitude,continuity and other characteristics of the seismic profile.Seismic facies is the sum of different features reflected on the seismic data.Seismic data can be divided into different seismic facies by dividing the characteristics of the slices such as amplitude and continuity.Accurate identification of seismic facies is helpful for analyzing the subsurface geological environment and further predicting oil and gas reservoirs.The generative adversarial network can generate data to enhance the seismic facies identification effect of traditional algorithms,and at the same time realize the intelligent identification of seismic facies through its discriminator.Seismic facies identification using artificial algorithms takes a long time and is complicated to implement.The results of traditional machine learning algorithms depend on the selection of seismic attributes,which is highly subjective.Unsupervised deep learning algorithms lack human guidance and have low accuracy;supervised deep learning algorithms can achieve high accuracy,but their accurate identification depends on rich label data.When applying it to real data,data augmentation on labeled samples is required.Data enhancement methods such as rotation and flipping are based on raw data rules and have limited enhancement effects.Aiming at the above problems,this thesis proposes a data enhancement method combined with generative adversarial networks,and realizes the intelligent identification of seismic facies.The main work of this thesis includes the following two aspects:(1)Data enhancement based on geological law constraints: On the basis of the ACGAN network,through the feature extraction network,the features of the seismic data are extracted as the input of the generator,and the initial distribution space of the generated data is constrained to speed up the convergence of the network,making the distribution of the generated data closer to that of the real data.Introduce the multi-scale residual convolution module to extract seismic data information of different scales,and prevent the problem of network information loss and model degradation caused by too deep network depth;combine the upsampling layer with the transposed convolution layer to avoid checkerboard artifacts in generated images.Construct a convolutional neural network,add generated data to the training set,and compare with the results of the traditionally enhanced data set.The effectiveness of this data enhancement method is verified through open-source seismic data from the F3 block in the North Sea of the Netherlands.(2)Seismic facies identification based on geological law constraints: modify the loss function of the proposed algorithm,add wasserstein distance and gradient penalty measures to avoid the disappearance of the gradient of the generator,improve the feature diversity of the generated images,to promote the optimization of the discriminator.Introduce spectral norm normalization to constrain the parameter update of the network,and add unilateral label smoothing to avoid over-fitting of the network.The intelligent identification of seismic facies is realized directly while generating images.The superiority of the model is proved by the data of the F3 work area.Also,the method is applied to the fault data and the mound beach body data from the Moxi area of the Sichuan Basin,and its classification effectiveness for different data is verified.