Application Of Convolution Neural Network In Seismic Facies Recognition

Seismic facies analysis is an important means to study stratigraphic sedimentary facies and sedimentary environment,which can provide useful reference for oil and gas reservoir exploration and development.In the past,seismic facies analysis methods often relied on the experience and knowledge of researchers,with the problem of strong subjectivity,and showed the disadvantage of low efficiency in the interpretation of a large number of seismic data.Convolutional neural network can learn and extract the features of seismic data set,so as to recognize seismic facies.This paper explores seismic facies recognition from two methods:convolution classification network and semantic segmentation model.The specific work and results are as follows:（1）The convolution classification network is used to identify the seismic facies of Stanford vi-e reservoir data.Firstly,ten coordinates are randomly selected on Stanford VI-E reservoir data volume,each coordinate point intercepts a cube with a size of 32×32×32 along the depth direction.Then,the orthogonal planes of three dimensions in each cube are extracted,and the training data set for seismic facies classification is generated.For the current mainstream convolution classification networks,such as VGG and Resnet,there are usually four to five down sampling operations,which is not suitable for the problem of small-size data sets.Based on VGG16 and Resnet50 models,this paper modifies the structure to reduce the times of downsampling of the model,so that the model is suitable for small-size data sets.Finally,two new convolution classification network models s VGG and s Resnet are obtained,and the Le Net model is built for experimental comparison.The three networks are applied to Stanford data set.The results show that the convolution classification network can effectively identify seismic facies,and the s Resnet network relying on the residual structure is better than the other two networks in the task of seismic facies classification.（2）The semantic segmentation algorithm is used to identify the seismic facies of the F3seismic data set in the North Sea of the Netherlands.The semantic segmentation algorithm judges the seismic facies category of each pixel in the seismic section,so as to realize the end-to-end seismic facies analysis.The existing semantic segmentation model has the problems of insufficient feature extraction ability,too simple upsampling stage,and difficult to deal with global and local information,resulting in the loss of seismic profile information.This paper proposes a semantic segmentation model CN-Unet,and builds four mainstream semantic segmentation models FCN-8s,Unet,PSPnet and Deep Lab V3+for comparison.In the process of model training,the effects of different loss functions on model training process and training results are discussed.Using the mixed loss function of 0.3 timesCE_Loss and 0.7 times Dice_{Lo ss} can better pay attention to the local and global characteristics of seismic facies.The experimental results show that the CN-Unet model achieves 92.06%and 96.01%in m IOU and MPA,which is better than other mainstream semantic segmentation models.