| Seismic facies identification is the basis of oil and gas reservoir exploration and reservoir prediction.The traditional seismic facies division method is usually based on the change of seismic attributes to identify seismic facies.For different seismic facies,it is necessary to manually select appropriate seismic attributes that can highlight their characteristics.With the continuous development of seismic exploration technology,the data scale of seismic data is increasing,and manual interpretation of seismic facies requires a lot of time and energy.At the same time,manual interpretation has great subjectivity and uncertainty,which directly affect the accuracy of later seismic data interpretation.For this reason,this thesis has carried out the research on the automatic classification method of seismic facies based on deep learning,in order to greatly improve the work efficiency of interpreters.The main research contents and results of this thesis are as follows:(1)An automatic classification method of seismic facies based on LinkNet is proposed.The existing seismic facies division methods based on deep learning algorithm pay little attention to the effective use of neural network parameters.The number of network parameters and operations is often very large,and the computational efficiency is low.LinkNet superimposes the input of each coding layer from the second coding layer onto the output of the corresponding decoding layer,so that the decoding layer can share the characteristics learned from the coding layer at each layer.Therefore,the decoding layer can be designed more concise and use fewer parameters,so LinkNet can perform large-scale computation in a faster and more efficient way,greatly saving computing resources.In addition,the method uses a combination of multivariate cross-entropy loss function and Tversky loss function as the loss function of the training network,which can improve the characterization of a few types of seismic facies boundaries in the unbalanced data.(2)An automatic seismic facies classification method based on LinkNet and hybrid pooling is proposed.A lightweight hybrid pooling module consisting of pyramid pooling and strip pooling is added between the coding layer and the decoding layer of LinkNet.Lightweight pyramid pooling is used to collect short-range correlation.It uses convolution layers of different scales to extract regional features,and then combines the features;Horizontal and vertical strip pooling is used to collect long-range correlation.Strip pooling uses strip convolution to extract regional features,which can connect discrete areas in the whole scene.At the same time,because of its narrow core shape,it can also capture local details.The hybrid pooling structure enables LinkNet network to fully extract and aggregate multi-scale feature information,and then more accurately divide seismic facies.(3)An automatic seismic facies classification method based on UPer Net and migration learning is proposed.UPer Net is based on the feature pyramid network(FPN),which connects the high-level features of low resolution and general information and the low-level features of high resolution and detailed information from top to bottom,so that the features at all scales contain rich stratigraphic structure information,and then carries out multi-scale feature fusion to complete the automatic classification of seismic facies.UPer Net has better generalization ability than LinkNet.This thesis uses the open source data of F3 work area in the north of the Netherlands to pre-train the UPer Net,and then uses only a small amount of SEG open source data as training data to fine-tune the trained UPer Net on the data of F3 work area through migration learning,and then uses it for the automatic classification of seismic facies of SEG open source data.The experimental results show that this method has higher accuracy of automatic classification of seismic facies than the method without migration learning initialization when the amount of tag data is small.(4)In order to reduce the dependence on manual labeling,this thesis proposes a semi-supervised seismic facies automatic classification method based on Mean Teacher.Mean Teacher includes a student model and several teacher models,and their network structure is completely consistent.Add different noises to the data and input the student model and teacher model respectively;Then calculate the loss function of student model output and label and the consistency loss function of teacher model output;Then the weight of student model is updated by gradient descent method;The weight of the teacher model is updated by the exponential moving average method(EMA)of the student model weight.Part of the weight comes from the weight of the historical teacher model,and the other part comes from the weight of the current student model.In this thesis,the student and teacher model in Mean Teacher is selected as UPer Net,three teacher models are set up,different learning parameters are given,and then the relative majority voting integration strategy is used to make full use of the feature information in the unlabeled data,reducing the demand for the amount of labeled seismic phase data.In this thesis,the effects of supervised learning,single teacher Mean Teacher,multiple(three)teacher Mean Teacher,and the combination of multiple teacher Mean Teacher and migration learning method on automatic seismic facies classification are compared.The results show that the accuracy of automatic seismic facies classification of the above methods is improved in turn with less tag data. |
