Seismic Facies Recognition Based On Deep Convolutional Neural Network

In order to solve the complexity problem of oil and gas exploration,seismic facies recognition technology is used to assist in the analysis of geological stratigraphic structure and improve the accuracy of oil and gas exploration.Traditional seismic facies recognition relies on the theoretical knowledge and experience of relevant researchers,and there are problems such as heavy workload,long time,strong subjectivity,and low efficiency.Deep learning technology can learn features hidden in the data,This paper studies how to use deep convolutional neural networks to learn the nonlinear mapping relationship between seismic data and seismic facies,so as to provide technical support for automatic seismic facies recognition.Aiming at the problems of lack of label data and low accuracy in existing seismic facies intelligent recognition technology,the classification model and semantic segmentation model based on convolutional neural network are studied how to solve the seismic facies recognition problem.The specific work is as follows:(1)Seismic facies recognition model fused with multi-scale features of seismic image dataAiming at the problem of few labels of existing seismic data,the interpreted seismic section is divided into multiple patch images(patch),the global seismic facies recognition of seismic section is realized through the classification of these patch images.Aiming at the problem that the convolutional neural network classification model(VGG16)has too many parameters and cannot fuse multi-scale features to extract more comprehensive seismic facies identification features,a seismic facies identification model(HA-VGG16)that fuses seismic facies multi-scale features is proposed.Firstly,three convolution and pooling operations are used to extract local features;Then,the features of large targets and small targets are extracted by hybrid dilated convolution(HDC),and the multi-scale features are fused by the atrous spatial pyramid pooling(ASPP);Finally,the fusional features are classified by softmax.On the F3 data,compared with the VGG16,Mobile Net V1 and Mobile Net V2 models,the results show that convergence time of the HA-VGG16 model is about 1/3 times of the original,and the accuracy rate is also significantly improved.(2)Semantic segmentation model of convolutional neural network for seismic facies recognitionThe classification model based on convolutional neural network can only perform coarse-grained recognition of seismic facies,but can't perform fine-grained recognition of each pixel in the seismic section.Based on the end-to-end deep learning model of semantic segmentation,the pixel-level classification of the input image can be realized through the "encoder-decoder" network structure.The encoder uses down-sampling to extract the feature information of the seismic section,and the decoder uses up-sampling to reconstruct the feature information to restore the original section resolution.The existing semantic segmentation network structure can only extract a single target feature,the PPM-Unet model is proposed,which uses Unet to extract low-level image features,and then a pyramid pooling module(PPM)is added to obtain spatial distribution features of different scales.On the F3 data set,compared with Unet,PSPNet and Deep Lab V3+ models,the results show that the mean pixel accuracy(MPA)reaches 0.94,and the mean intersection over union(MIOU)reaches 0.89.