Interactive Seismic Facies Interpretation Method Based On Deep Learning

Seismic facies analysis is the process of describing and interpreting seismic reflection parameters,such as amplitude,frequency,and continuity,within the stratigraphic framework of a depositional sequence.It is used to determine the types of seismic parameters in the sedimentary system that indicate the combined sedimentary structure and rock type.Thus,seismic facies analysis plays an important role in predicting oil and gas reservoirs.Due to the inherent characteristics of seismic data such as low resolution,high noise and complex geological structure,intuitive seismic interpretation based on interpreters can easily lead to misjudgment of identification results,which affects the accuracy of subsequent seismic data interpretation.In addition,with the increasing amount of seismic data,manual interpretation of seismic facies requires a lot of time and energy,which also affects the efficiency of seismic data interpretation.Therefore,the intelligence and automation of seismic facies identification is of great importance for oil and gas exploration.Traditional automatic seismic facies analysis mainly relies on classical machine learning algorithms such as clustering,support vector machines,and random forests.These algorithms usually require manual selection of seismic attributes(root mean square amplitude,coherent volume and chaotic volume)and then further classification of the extracted seismic attributes.However,seismic attributes cannot fully utilize all the information in seismic data in practical applications,resulting in low accuracy of seismic facies classification by machine learning.In recent years,with the improvement of computer hardware and the development of deep learning,deep learning has been widely used in seismic facies analysis with its powerful feature extraction ability.However,there are still some problems with fully automatic seismic facies recognition methods based on deep learning:(1)It is difficult to collect a large amount of seismic data and corresponding labels of seismic facies,and the predicted results by such methods are not satisfactory when the training data samples are insufficient;(2)When such methods are applied to new seismic data,their versatility and accuracy are often less than expected.Therefore,this thesis uses the method of numerical simulation to quickly generate a large amount of seismic data to supplement the data and provide a data foundation for the deep learning.In addition,intelligent optimization algorithms are introduced to achieve interactive seismic facies recognition,based on limited prior information given manually to obtain as many seismic facies features as possible,thereby quickly and accurately identifying the corresponding seismic facies.The contents of this thesis mainly consist of the following aspects:(1)Research on interactive identification method of salt domes in 2D seismic data.The existing automatic identification method of salt dome based on deep learning may be inaccurate due to noise interference.Therefore,in order to solve the problem,this thesis proposes a new interactive segmentation method based on deep learning to extract salt domes.To incorporate human-guided information into the method,we convert humanclicked interaction points(where positive points represent salt domes and negative points represent background)into Euclidean distance maps,and combine them with seismic data to train our network model.The model consists of a UNet network and a pyramid module,and is trained and tested on the TGS salt seismic dataset.In addition,we use the graph cut algorithm to refine the salt dome probability map predicted by the network model to obtain more refined salt dome results.This method can automatically extract accurate salt domes from the actual seismic section by manually clicking only a few interaction points,thus improving the efficiency of salt dome identification.(2)Research on 3D salt dome seismic data simulation method and 3D salt dome intersection recognition method.Aiming at the problem of salt dome identification of 2D seismic data,we propose a method to randomly generate 3D seismic data containing salt dome to enrich the diversity of training data.To improve the accuracy of salt dome interpretation and shorten the work cycle,we use the Unet model to build a 3D encoderdecoder network architecture with skip connection.The model takes the seismic data volume of a certain size as the input,and outputs the probability volume of salt dome of the same size.We train and test the model with seismic data of 3D simulated salt dome.To obtain more detailed 3D salt dome results,we use 3D graph cut to optimize the probability volume of salt dome output by the model.Experimental results on 3D real seismic data show that the proposed method has better performance,which reflects the effectiveness of the proposed method on real data.(3)Research on fluvial facies interactive identification method in 2D seismic data.This method can more accurately identify and segment the channel geological body with fuzzy boundary,poor continuity and even missing data.First,the seismic interpreter clicks on the interior of the seismic facies and converts the artificial interactive information based on the geodesic distance to obtain the geodesic distance map.The map and the seismic attributes commonly used for fluvial facies identification are used as the input of the network model,which reduces the complexity of fluvial facies identification;then the probability map of fluvial facies is output from the network model,and the preliminary results of fluvial facies identification are obtained;finally,the seismic interpreters add extra interaction points,and use the conditional random field to fuse the geodesic distance maps of the extra points with the probability maps of the network model,and finally obtain accurate fluvial facies interpretation results.Compared with seismic attribute and deep learning method,the fluvial facies extracted by this method is more complete and improves the continuity of the channel boundaries.In the case of complex seismic data,this method can still realize fluvial facies interpretation well,which has important guidance significance in practical application.(4)Research on 3D channel seismic data simulation method and 3D channel interactive recognition method.In order to solve the problem of insufficient seismic data,we propose a method that can automatically generate seismic data with real channel.The simulation method is based on the geological numerical simulation to randomly generate a variety of meandering river;then the folded geological model is built and combined with the meandering river to build the reflection coefficient model and the corresponding channel label;finally,the reflection coefficient model is convolved with the Rick wavelet to synthesize the seismic data containing the channel.In addition,to improve the efficiency of channel interpretation in 3D seismic data,we propose an interactive 3D channel interpretation method using two network models to improve the initial results obtained from the network model,and uses manual interaction to achieve higher accuracy in the thinning process.This method takes seismic data and seismic attributes as the input of the first network model to automatically obtain the initial channel identification results,and then adds interaction points to mark the wrong locations based on the the results,and transforms these interaction points with geodesic distance,and finally inputs them to another network model to obtain accurate results.Although the method is trained on synthetic seismic data,compared with fully automatic deep learning methods,it performs better on real seismic data and has good application prospects.The novelties of this thesis are summarized as follows:(1)In order to solve the problem of insufficient sample size for seismic data and corresponding seismic facies interpretation results,this thesis proposes a 3D salt dome seismic data simulation method and a meandering river evolution simulation method.These methods can quickly and randomly generate corresponding seismic data and labels,and the generated geological structures have diversity and complexity,which can provide a data basis for seismic facies interpretation methods based on deep learning.(2)Due to the problems of poor accuracy and practicality of existing seismic facies interpretation methods based on deep learning,this thesis combines deep learning,seismic attributes,and graph cutting algorithms to achieve interactive recognition of salt facies and river facies in 2D and 3D spaces,respectively.These methods can provide new application schemes for seismic facies interpretation.