Deep Generative Networks For Seismic Facies Analysis

Seismic facies play a key role in geological stratigraphic analysis and reservoir prediction.In recently,researchers have proposed some seismic facies classification methods based on deep learning,but deep learning often requires a large number of labeled data for training.However,it is difficult and costly to obtain labeled data in actual seismic prospecting.Aiming at the problem of insufficient existing seismic labeled data,in this paper,deep generative models for seismic facies analysis are studied from the perspective of selfsupervised,unsupervised and semi-supervised learning.Our work is as follows:(1)Pre-trained models for seismic features extractor.For seismic facies analysis,the deep neural networks need to extract semantic features at different levels,but they depend on different annotation data.Similar to the mask models in self-supervised learning,reconstruction of masked seismic traces is used as a pretext task,then we get backbone networks from massive unlabeled seismic data to provide some seismic features for downstream tasks.The experimental results show that,the pre-trained model can effectively improve the seismic facies classification performance.(2)Seismic facies clustering model fused with deep generative networks and clustering algorithms.For to overcome low accuracy with traditional of traditional seismic clustering algorithms,clustering of seismic facies clustering with clustering algorithm based on deep neural networks,named deep clustering.Deep clustering is fused with deep neural networks and clustering algorithms.Clustering loss guides deep neural network to extract friendly clustering features,which can obtain better clustering results.In this paper,we proposed three different deep generative networks,which were used as feature extractors,and the network loss and clustering loss are combined as the total loss.The experimental results show that the accuracy of seismic clustering using deep clustering has been significantly improved.(3)Semi-supervised learning for seismic facies segmentation.Semantic segmentation network predicts seismic data at the pixel level.In actual seismic prospecting,there is still a little precious annotated data.Through semi-supervised learning,annotated data and annotated data are both involved in the training process.The generator is based on a multiclass classification semantic segmentation network,which predicts seismic data at the pixel level.The discriminator is based on a binary class classification semantic segmentation network,which classifies labeled seismic facies and the result of the generator.The proposed model has been tested successfully on three real seismic data fields,which indicates that the model is effective with limited labeled data.(4)Unsupervised learning for seismic facies segmentation.Deep learning often requires a large number of labeled seismic data for training.To overcome this,we proposed an invariant information clustering model for unsupervised seismic facies segmentation.The model uses the similarity of segmentation results between seismic patches and their nearest neighbors to maximize the mutual information learning objectives of these seismic facies segmentation results,so as to learn a seismic facies segmentation network.The experimental results show that the method is effective in seismic facies segmentation.