Seismic facies identification method and system based on Dice index

Abstract

According to the seismic facies identification method and system based on the Dice index provided by the invention, seismic facies segmentation training of the section image is performed on the full convolutional neural network in advance, and the seismic section obtained through observation processing is input into the full convolutional neural network, so that the section image with low seismic facies segmentation precision is obtained; and carrying out post-processing optimization on the coarse segmentation result, wherein the post-processing optimization comprises carrying out iterative optimization on the segmentation boundary by using a full-connection CRF and carrying out optimization on classification of small connected domains by using small connected domain statistical analysis. The problems that seismic facies classification boundaries are not fine enough, segmentation precision is not high, and interference is caused by profile image noise are well solved. The method provided by the invention powerfully supports the segmentation of the seismic facies profile image, and is of great significance for helping geological prospecting workers to improve the working efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing, and particularly relates to a seismic facies identification method and system based on the Dice index. Background Art

[0002] In the field of oil exploration, by studying the sedimentary facies, the paleogeographic features and the development history of the basin during the geological period can be understood and studied. Furthermore, the conditions and distribution laws favorable for oil and gas formation can be found to guide the exploration and development of oil. The study of sedimentary facies can also help understand the physical properties structure of the reservoir and the distribution of sand bodies in space, thereby improving the economic benefits of the oilfield. When searching for subtle non-structural oil and gas reservoirs, the study of sedimentary facies and sedimentary environment becomes even more important. Based on the accurate division of seismic facies, according to the changes in seismic stratigraphic parameters, the stratigraphic units with similar parameters in the same seismic sequence are connected to make a planar distribution map of seismic facies, and it is analyzed to find out the corresponding relationship between seismic facies and sedimentary facies, so as to "transform the facies" of the seismic facies map to form a sedimentary facies map. The sedimentary facies map or sedimentary environment map is the comprehensive expression or final result of seismic facies analysis. Seismic waveform classification analysis is mainly applicable to areas with parallel structures, small changes in formation thickness, and relatively simple structures. When opening time windows downward or upward along the standard layer for waveform classification analysis, it can be maintained within an isochronous framework without crossing seismic traces, and the effect is better.

[0003] For areas with large changes in formation thickness, such as the wedge-shaped formations at the edge of the basin, due to unequal thickness, uniformly opening time windows often leads to the phenomenon of time crossing caused by crossing the axis, resulting in a decrease in research accuracy. In areas with complex structures, due to many influencing factors, it is necessary to first perform target-preserving amplitude processing to obtain better results. Therefore, the change in seismic reflection characteristics caused by the change in formation structure is a key factor that needs to be considered when performing seismic waveform classification. It is very necessary to introduce formation structure information seismic parameters for seismic facies analysis. Traditional sedimentary facies research is to establish sedimentary facies models based on core observations, logging curve morphology research, etc., and perform single-well facies division; use attributes such as the strength, continuity, frequency, reflection geometry, and internal structure of seismic reflections to compile seismic facies maps and perform the conversion of seismic facies to inter-well sedimentary facies. This method is not only time-consuming and laborious, but also has a large degree of artificiality and certain uncertainty, and its accuracy is difficult to meet the needs of sedimentary facies research for the exploration and development of subtle oil and gas reservoirs. Seismic waveform classification seismic facies analysis uses neural network and pattern recognition technologies to compare the waveform characteristics and seismic attribute characteristics of each actual seismic data trace in a certain section one by one, and carefully depicts its lateral changes to obtain the planar distribution law of seismic anomalies, that is, the seismic waveform classification map. At present, the seismic facies plan obtained by seismic waveform classification seismic facies analysis is intuitive and rich in color, and is increasingly valued in sedimentary facies research, and has achieved good results in the preliminary application in China.

[0004] The following problems exist in the seismic facies image segmentation based on convolutional neural network: (1) The labels of adjacent pixels in the image are independent, without considering the consistency and correlation between labels; (2) The size of the convolution kernel limits the size of the receptive field, and the convolutional network can only extract local features, resulting in the network being unable to segment fine formation boundaries; (3) The convolutional neural network performs convolution on each adjacent pixel block one by one, resulting in high redundancy in calculations. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a seismic facies recognition method and system based on the Dice index that overcomes the above problems or at least partially solves the above problems.

[0006] According to one aspect of the present invention, a seismic facies recognition method based on the Dice index is provided. The recognition method includes:

[0007] Step 1, after processing the original seismic data obtained from field exploration, obtaining a seismic profile image set;

[0008] Step 2, dividing the seismic profile image set, constructing a seismic profile image set sample, and dividing the seismic profile image set sample into a training set and a validation set;

[0009] Step 3, constructing a fully convolutional DeepLabV3+ neural network model;

[0010] Step 4, inputting the training set into the fully convolutional neural network model for training to obtain a trained fully convolutional neural network model;

[0011] Step 5, inputting the seismic profile image to be recognized into the trained fully convolutional neural network model to obtain a rough result image of seismic facies segmentation of the profile image;

[0012] Step 6, according to the rough segmentation result output by the fully convolutional neural network, obtaining a fine segmentation fusion result processed by the fully connected CRF model;

[0013] Step 7, dividing the connected domains that need to be reclassified according to the fine segmentation fusion result.

[0014] Optionally, in Step 1, after processing the original seismic data obtained from field exploration, obtaining a seismic profile image set specifically includes:

[0015] Performing data decoding on the original seismic data obtained from field exploration to obtain seismic trace data;

[0016] Performing editing processing on the seismic trace data;

[0017] After gain recovery, extracting CDP gather.

[0018] Remove interference noise;

[0019] Adopt correction offset processing to obtain a seismic profile image set.

[0020] Optionally, the editing process for the seismic trace data specifically includes:

[0021] Zero-fill the seismic traces with a signal-to-noise ratio lower than the threshold;

[0022] Correct the working traces with reversed polarity.

[0023] Optionally, the adopting correction offset processing to obtain a seismic profile image set specifically includes: After static and dynamic corrections, common depth point stacking, and migration processing, a seismic profile image set is obtained.

[0024] Optionally, the removing interference noise specifically includes: Adopt deconvolution and filtering to remove interference noise.

[0025] Optionally, step 3, constructing a fully convolutional DeepLabV3+ neural network model specifically includes:

[0026] Construct a fully convolutional DeepLabV3+ neural network model, and introduce a spatial pyramid pooling module with dilated convolutions;

[0027] Adopt dilated convolutions with different dilation coefficients;

[0028] Use a pre-trained model to initialize the model weights.

[0029] Optionally, step 4, inputting the training set into the fully convolutional neural network model for training to obtain a trained fully convolutional neural network model:

[0030] Input the training set into the fully convolutional neural network model for training;

[0031] During the learning process, adopt the Dice coefficient loss function L Dice and the multi-class cross-entropy loss function L CE to form a combined loss function l loss to describe the closeness between the classification result and the true formation;

[0032] Based on the multi-class cross-entropy loss function, introduce the Dice coefficient loss function, so that the fully convolutional neural network model can still better evaluate the similarity of the classification result under the condition of extremely unbalanced classification samples;

[0033] And use the validation set to adjust the model parameters during the training process to obtain a trained fully convolutional neural network model.

[0034] Optionally, the hybrid loss function L loss The calculation formula of

[0035]

[0036] where λ CE represents the weight of the multi-class cross-entropy loss function; λ Dice represents the weight of the Dice coefficient loss function; C represents the total number of seismic facies categories; N represents the total number of pixel points contained in the input seismic section; y c,n represents the label data, whether the nth pixel point on the actual seismic section is the cth seismic facies. If so, it is 1; otherwise, it is 0; z c,n represents the output result of the neural network, and its physical meaning is the probability that the corresponding point is predicted as the cth seismic facies; p c,n represents the result after normalizing z c,n by the Softmax function; ε represents the smoothing coefficient.

[0037] Optionally, step 6, obtaining the refined segmentation fusion result processed by the fully connected CRF model according to the coarse segmentation result output by the fully convolutional neural network specifically includes:

[0038] According to the coarse segmentation result output by the fully convolutional neural network;

[0039] Initialize the energy function in the fully connected conditional random fields (CRF) model to obtain the original attribution label probability value of the pixel points;

[0040] Calculate the fully connected CRF model, iteratively correct the maximum probability map in the seismic section segmentation image, and obtain the refined segmentation fusion result processed by the fully connected CRF model.

[0041] Optionally, the steps of calculating the fully connected CRF model specifically include:

[0042] Step 6.1, construct the energy function E(x|I), and the energy function is composed of the unary potential function θ i (x i ) and the binary potential function θ ij (x i , x j );

[0043] In step 6.1, the calculation formula for defining the energy function E(x|I) of the fully connected CRF for the label assignment scheme x is:

[0044]

[0045] θ i (xi ) = -logP(x i )

[0046]

[0047] where I represents the current input pixel vector; x represents the label predicted for the current observed pixel; θ i (x i ) represents the unary potential function, which is defined on the state feature function of the observation position x i ; x i represents the current input pixel; P(x i ) is the probability value of the seismic facies classification of this pixel;

[0048] θ ij (x i , x j ) represents the binary potential function, which is defined on the transition feature function at different observation positions; μ(x i , x j ) represents the label compatibility term, which is 1 when the predicted classification labels of the seismic facies of x i and x j are inconsistent, and 0 otherwise; ω 1 represents the weight parameter of the surface kernel; ω 2 represents the weight parameter of the smoothing kernel; p i represents the position encoding of pixel i; p i represents the position encoding of pixel j; Ii represents the color encoding of pixel i; I j represents the color encoding of pixel j; σ α represents the variance coefficient of the position vector p; σ β represents the variance coefficient of the color vector I; σ γ represents the variance coefficient of the smoothing term;

[0049] Step 6.2: Normalize the energy function value to obtain the probability value of the pixel belonging to the seismic facies classification label, and take the label corresponding to this pixel with the maximum probability until the iteration of the probability values of all pixel points belonging to their respective classification labels reaches more than 90% and then exit the iteration;

[0050] In Step 6.2, the formula for calculating the posterior probability value of a pixel point is:

[0051]

[0052] where Z(I) represents the normalization factor, and x represents the label predicted for the current observed pixel;

[0053] Obtain the seismic facies segmentation map with refined boundaries.

[0054] Optionally, step 7, dividing the connected regions that need to be reclassified according to the fine segmentation fusion result further includes:

[0055] Statistically classify the boundaries of the connected regions that need to be reclassified into eight-domain seismic facies classification categories, and use the voting system to update the pixel position iterated to currently with the seismic facies classification outside the domain. If the statistical results show that the votes of two categories are the same, the processing is postponed.

[0056] Among them, if a single pixel is not updated, randomly select the seismic facies classification with equal votes as the final optimized classification category for covering the pixel point position.

[0057] Optionally, step 7, dividing the connected regions that need to be reclassified according to the fine segmentation fusion result specifically includes:

[0058] The present invention also provides a seismic facies recognition system based on the Dice index, which applies the above-mentioned seismic facies recognition method based on the Dice index. The recognition system includes:

[0059] An original data processing module for processing the original seismic data obtained from field exploration to obtain a set of seismic profile images;

[0060] An image set construction module for dividing the set of seismic profile images, constructing a sample set of seismic profile images, and dividing the sample set of seismic profile images into a training set and a validation set;

[0061] A model construction module for constructing a fully convolutional DeepLabV3+ neural network model;

[0062] A neural network model training module for inputting the training set into the fully convolutional neural network model for training to obtain a trained fully convolutional neural network model;

[0063] An input training module for inputting the seismic profile image to be recognized into the trained fully convolutional neural network model to obtain a rough result image of seismic facies segmentation of the profile image;

[0064] A fine segmentation fusion result calculation module for obtaining the fine segmentation fusion result processed by the fully connected CRF model according to the rough segmentation result output by the fully convolutional neural network;

[0065] A fusion result division module for dividing the connected regions that need to be reclassified according to the fine segmentation fusion result.

[0066] A seismic facies identification method and system based on the Dice index provided by the present invention, the identification method comprising: Step 1, after processing the original seismic data obtained from field exploration, obtaining a seismic profile image set; Step 2, dividing the seismic profile image set to construct a seismic profile image set sample, and dividing the seismic profile image set sample into a training set and a validation set; Step 3, constructing a fully convolutional DeepLabV3+ neural network model; Step 4, inputting the training set into the fully convolutional neural network model for training to obtain a trained fully convolutional neural network model; Step 5, inputting the seismic profile image to be identified into the trained fully convolutional neural network model to obtain a rough result image of seismic facies segmentation of the profile image; Step 6, obtaining a fine segmentation fusion result processed by a fully connected CRF model according to the rough segmentation result output by the fully convolutional neural network; Step 7, dividing the connected regions that need to be reclassified according to the fine segmentation fusion result. Solve the problems that the seismic facies classification boundary is not fine enough, the segmentation accuracy is not high, and the interference caused by profile image noise. The present invention strongly supports the segmentation of seismic facies profile images, which is of great significance for helping geological exploration staff improve work efficiency. Solve the problems that the seismic facies classification boundary is not fine enough, the segmentation accuracy is not high, and the interference caused by profile image noise. The present invention strongly supports the segmentation of seismic facies profile images, which is helpful for geological exploration staff to improve work efficiency.

[0067] The above description is only an overview of the technical solution of the present invention. In order to be able to understand the technical means of the present invention more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention are specifically exemplified below. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0069] Figure 1 It is a flowchart of a seismic facies identification method based on the Dice index provided by an embodiment of the present invention;

[0070] Figure 2 It is a practical operation flowchart provided by an embodiment of the present invention;

[0071] Figure 3 It is a structural diagram of a fully convolutional neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0073] The terms "including" and "having" and any variations thereof in the description of the embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusion. For example, including a series of steps or units.

[0074] The technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

[0075] As Figure 1 shown, a seismic facies identification method based on the Dice index and the maximum probability map is as follows:

[0076] Step 1: After decoding the original seismic data obtained from field exploration, edit the seismic trace data, fill zero for abnormal traces with low signal-to-noise ratio, and correct the working traces with polarity inversion. After gain recovery, extract the CDP gather. Remove the interference noise through deconvolution and filtering. After static and dynamic corrections, common depth point stacking and migration processing, a seismic profile image set is obtained.

[0077] Step 2: Divide a part of the seismic profile image set that has been manually interpreted and labeled, construct a seismic profile image set sample, and divide the sample image data set into a training set and a validation set.

[0078] Step 3: Construct a fully convolutional DeepLabV3+ neural network model, introduce a spatial pyramid pooling module with dilated convolution, and use dilated convolutions with different dilation coefficients to further extract multi-scale feature information, so that the model can perceive rich semantic information at different scales. Then use the pre-trained model to initialize the model weights.

[0079] Step 4: Input the training set into the fully convolutional neural network model for training. Then, during the learning process, use a hybrid loss function L Dice composed of the Dice index loss function L CE and the multi-class cross-entropy loss function L lossTo describe the closeness between the classification results and the actual strata. Based on the multi-class cross-entropy loss function, the Dice index loss function is introduced, enabling the model to better evaluate the similarity of the classification results even under extremely imbalanced classification samples, thereby improving the robustness of the model. Then, the validation set is used to adjust the model parameters during the training process to obtain the trained fully convolutional neural network model.

[0080] In step 4, the hybrid loss function L loss has the following specific calculation formula:

[0081]

[0082] where λ CE represents the weight of the multi-class cross-entropy loss function; λ Dice represents the weight of the Dice index loss function; C represents the total number of seismic facies categories; N represents the total number of pixel points contained in the input seismic section; y c,n represents the label data, that is, whether the nth pixel point on the actual seismic section is the cth seismic facies. If so, it is 1; otherwise, it is 0; z c,n represents the output result of the neural network, and its physical meaning can be understood as the probability that this point is predicted as the cth seismic facies; p c,n represents the result after normalizing z c,n using the Softmax function; ε represents the smoothing coefficient.

[0083] Step 5: Input the seismic section image to be recognized into the trained fully convolutional neural network model to obtain the rough result image of seismic facies segmentation of the section image.

[0084] Step 6: According to the rough segmentation result output by the fully convolutional neural network in step 5, initialize the energy function in the fully connected conditional random fields (CRF) model to obtain the original attribution label probability value of the pixel points. Subsequently, calculate the fully connected CRF model, continuously iterate and correct the maximum probability map in the seismic section segmentation image to obtain the fine segmentation fusion result after being processed by the fully connected CRF.

[0085] The steps for calculating the fully connected CRF model include the following:

[0086] Step 6.1: Construct the energy function E(x|I), which is composed of the unary potential function θ i (x i ) and the binary potential function θ ij (x i ,x j ).

[0087] In step 6.1, the calculation formula for the energy function \(E(x|I)\) of the fully connected CRF for the label assignment scheme \(x\) is defined as:

[0088]

[0089] \(\theta\) i (x i ) = -\(\log\)P(x i )

[0090]

[0091] where \(I\) represents the current input pixel vector; \(x\) represents the label predicted for the current observed pixel; \(\theta\) i (x i ) represents the unary potential function, which is defined as the state feature function at the observation position \(x\) i ; \(x\) i represents the current input pixel; P(x i ) is the probability value of the seismic facies classification of this pixel;

[0092] \(\theta\) ij (x i , x j ) represents the binary potential function, which is defined as the transition feature function at different observation positions; \(\mu(x\) i , x j ) represents the label compatibility term. When the predicted classification labels of the seismic facies of \(x\) i and \(x\) j are inconsistent, this value is 1; otherwise, it is 0; \(\omega\) 1 represents the weight parameter of the surface kernel; \(\omega\) 2 represents the weight parameter of the smoothing kernel; \(p\) i represents the position encoding of pixel \(i\); \(p\) i represents the position encoding of pixel \(j\); \(I_i\) represents the color encoding of pixel \(i\); \(I\) j represents the color encoding of pixel \(j\); \(\sigma\) α represents the variance coefficient of the position vector \(p\); \(\sigma\) β represents the variance coefficient of the color vector \(I\); \(\sigma\) γ represents the variance coefficient of the smoothing term.

[0093] In step 6.2, normalize the energy function value to obtain the probability value of the pixel belonging to the seismic facies classification label, and take the label corresponding to this pixel with the maximum probability until the iteration of the probability values of all pixel points belonging to their respective classification labels reaches more than 90% and then exit the iteration;

[0094] In step 6.2: The calculation formula for the posterior probability value of the pixel point is:

[0095]

[0096] Among them, Z(I) represents the normalization factor, and x represents the label predicted by the current observed pixel.

[0097] After step 6, a seismic facies segmentation map with refined boundaries will be obtained.

[0098] Step 7: For the seismic facies segmentation map with refined boundaries, optimize the seismic facies classification in the small connected regions on the seismic facies segmentation image to obtain a more refined seismic facies classification result.

[0099] The steps for optimizing the seismic facies classification of small connected regions in the image include:

[0100] Step 7.1: Based on the respective seismic facies classifications in the seismic facies segmentation map with refined boundaries, count the connected region information in the whole map, sort the connected regions according to the area size of the connected regions, and divide the connected regions that need to be reclassified.

[0101] Step 7.2: Statistically analyze the seismic facies classification categories of the eight neighborhoods of the boundaries of the connected regions that need to be reclassified respectively, and use the voting system to update the pixel positions at the current iteration with the seismic facies classifications outside the neighborhood. If the statistical results show that two types of votes are the same, the processing will be postponed. Among them, if a single pixel is not updated, randomly select the seismic facies classification with equal votes as the final optimized classification category for the pixel position.

[0102] Embodiment 1

[0103] The flow of the seismic facies recognition method based on the Dice index and the maximum probability atlas of the present invention is as Figure 1 shown. The actual operation process in this embodiment is as Figure 2 shown. In this actual operation process, it is necessary to preprocess the field seismic tapes, including operations such as demultiplexing, excision, amplitude recovery, deconvolution, filtering, static and dynamic correction, and migration processing; continuously update and iterate the parameters of the fully convolutional neural network model according to the marked profiles; load the trained model parameters, perform rough seismic facies segmentation, CRF boundary optimization, and small neighborhood optimization on the profile images, and finally obtain a refined seismic facies classification result.

[0104] A seismic facies recognition method based on the Dice index and the maximum probability atlas, the steps are as follows:

[0105] First, after demultiplexing the original seismic data obtained from field exploration, edit the seismic trace data, fill zeros for abnormal traces with low signal-to-noise ratio, and correct the working traces with reversed polarities. After gain recovery, extract the CDP gather. Remove the interference noise through deconvolution and filtering. After static and dynamic correction, common depth point stacking, and migration processing, a set of seismic profile images is obtained.

[0106] Partition a set of seismic profile images that have been partially manually interpreted and labeled to construct a sample set of seismic profile images, and divide the sample image data set into a training set and a validation set.

[0107] Then construct a fully convolutional neural network model ( Figure 3 ), introduce a spatial pyramid pooling module with dilated convolutions, and use dilated convolutions with different dilation coefficients to further extract multi-scale feature information, enabling the model to perceive rich semantic information at different scales. Then initialize the weights using a pre-trained model.

[0108] Input the training set into the fully convolutional neural network model for training. Then, during the learning process, use a hybrid loss function \(L\) Dice composed of the Dice coefficient loss function \(L\) CE and the multi-class cross-entropy loss function \(L\) loss to describe the closeness between the classification result and the true formation. Based on the multi-class cross-entropy loss function, introduce the Dice coefficient loss function, enabling the model to still evaluate the similarity of the classification result well even under the condition of extremely unbalanced classification samples, thereby improving the robustness of the model. Then adjust the model parameters during the training process using the validation set to obtain a trained fully convolutional neural network model.

[0109] In step 4, the calculation formula of the hybrid loss function \(L\) loss is specifically as follows:

[0110]

[0111] where \(\lambda\) CE represents the weight of the multi-class cross-entropy loss function; \(\lambda\) Dice represents the weight of the Dice coefficient loss function; \(C\) represents the total number of seismic facies classes; \(N\) represents the total number of pixel points contained in the input seismic profile; \(y\) c,n represents the label data, that is, whether the \(n\)th pixel point on the actual seismic profile is the \(c\)th seismic facies class. If so, it is 1; if not, it is 0; \(z\) c,n represents the output result of the neural network, and its physical meaning can be understood as the probability that this point is predicted to be the \(c\)th seismic facies class; \(p\) c,n represents the result after normalizing \(z\) c,n using the Softmax function; \(\epsilon\) represents the smoothing coefficient.

[0112] Then, input the seismic profile image to be recognized into the trained fully convolutional neural network model to obtain a rough result image of seismic facies segmentation for the profile image.

[0113] According to the rough segmentation result output by the fully convolutional neural network above, initialize the energy function in the Conditional Random Fields (CRF) model to obtain the original belonging label probability value of the pixel points. Subsequently, calculate the fully connected CRF model, continuously iterate and correct the maximum probability map in the seismic profile segmentation image, and obtain the fine segmentation fusion result after the fully connected CRF processing.

[0114] The steps of calculating the fully connected CRF model include the following:

[0115] First, construct the energy function E(x|I), which is composed of the unary potential function θ i (x i ) and the binary potential function θ ij (x i , x j ). First, determine the predicted value of the model for the pixel, and then determine the predicted seismic facies category of the pixel value. Then, encode the actual category and the pixel predicted value between two adjacent pixels, and calculate the dependence potential value. When the actual categories of adjacent pixels are the same, their dependence potential is higher. Since this CRF is fully connected, information can be shared even between pixels with a large Euclidean distance.

[0116] Among them, the calculation formula for the energy function E(x|I) of the fully connected CRF for the label assignment scheme x is defined as:

[0117]

[0118] θ i (x i ) = -logP(x i )

[0119]

[0120] In the formula, I represents the current input pixel vector; x represents the label predicted by the current observed pixel; θ i (x i ) represents the unary potential function, which is a state feature function defined at the observation position x i ; x i represents the current input pixel; P(x i ) is the probability value of the seismic facies classification of this pixel;

[0121] θ ij (x i , x j ) represents the binary potential function, which is a transition feature function defined at different observation positions; μ(x i , x j ) represents the label compatibility term. When x i and xj When the seismic facies prediction classification labels are inconsistent, the value is 1; otherwise, it is 0; ω 1 represents the weight parameter of the surface kernel; ω 2 represents the weight parameter of the smoothing kernel; p i represents the position encoding of pixel i; p i represents the position encoding of pixel j; Ii represents the color encoding of pixel i; I j represents the color encoding of pixel j; σ α represents the variance coefficient of the position vector p; σ β represents the variance coefficient of the color vector I; σ γ represents the variance coefficient of the smoothing term. All three variance coefficients belong to hyperparameters and are used to adjust the weights of their respective Gaussian distributions.

[0122] Then, normalize the energy function value to obtain the probability value of the pixel belonging to the seismic facies classification label, and take the label corresponding to the pixel with the maximum probability until the iteration of the probability values of all pixel points belonging to their respective classification labels reaches more than 90% and then exit the iteration;

[0123] Among them: The posterior probability value of the pixel point satisfies the Gibbs distribution, and its calculation formula is:

[0124]

[0125] Among them, Z(I) represents the normalization factor, and x represents the label predicted by the current observed pixel.

[0126] After calculating through the CRF model, obtain the seismic facies segmentation map with refined boundaries. The seismic facies classification in the small connected regions on this seismic facies segmentation image will be optimized.

[0127] The optimization steps for the seismic facies classification of the small connected regions in the image are as follows:

[0128] First, based on the respective seismic facies classifications in the seismic facies segmentation map with refined boundaries, count the connected region information in the whole map, sort the connected regions according to the area size of the connected regions, and divide the connected regions that need to be reclassified.

[0129] Then, count the seismic facies classification categories of the eight neighborhoods of the boundaries of the respective connected regions that need to be reclassified, and use the voting system to update the pixel positions of the current iteration with the seismic facies classifications outside the neighborhood. If the statistical results show that the votes of two categories are the same, postpone the processing. Among them, if a single pixel is not updated, randomly select the seismic facies classification with equal votes as the final optimized classification category for the position of the covered pixel point.

[0130] The above embodiments preprocess the original observation data obtained from field seismic surveys to obtain source seismic profiles. The fully convolutional neural network is trained for seismic facies segmentation of profile images. The source seismic profiles are input into the fully convolutional neural network, and profile images with low seismic facies boundary segmentation accuracy are obtained. The post-processing optimization is performed on the rough segmentation results, including: iteratively optimizing the segmentation boundary using a fully connected CRF and optimizing the classification of small connected components using small connected component statistical analysis. This effectively solves the problems of insufficient fineness of seismic facies classification boundaries, low segmentation accuracy, and interference caused by profile image noise. The present invention strongly supports the segmentation of seismic facies profile images and is of great significance for helping geological exploration workers improve work efficiency.

[0131] Beneficial effects: The present invention pre-trains the fully convolutional neural network for seismic facies segmentation of profile images, inputs the seismic profiles obtained through observation and processing into the fully convolutional neural network, and obtains profile images with low seismic facies segmentation accuracy. The post-processing optimization is performed on the rough segmentation results, including: iteratively optimizing the segmentation boundary using a fully connected CRF and optimizing the classification of small connected components using small connected component statistical analysis. This effectively solves the problems of insufficient fineness of seismic facies classification boundaries, low segmentation accuracy, and interference effects caused by profile image noise. The present invention strongly supports the segmentation of seismic facies profile images and is of great significance for helping geological exploration workers improve work efficiency.

[0132] The above specific embodiments further elaborate on the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A seismic phase identification method based on Dice index, characterized in that: The identification method comprises: Step 1, after processing the original seismic data obtained from field exploration, a set of seismic profile images is obtained; Step 2, dividing the seismic profile image set to construct a seismic profile image set sample, and dividing the seismic profile image set sample into a training set and a validation set; Step 3, build a fully convolutional DeepLabV3+ neural network model; Step 4, input the training set into the fully convolutional neural network model for training, and obtain a trained fully convolutional neural network model; Step 5, inputting the seismic section image to be identified into the trained full convolutional neural network model to obtain a rough result image of seismic phase segmentation of the section image; Step 6: According to the coarse segmentation result output by the full convolutional neural network, the fine segmentation fusion result after being processed by the fully connected CRF model is obtained; Step 7: divide the connected domains that need to be reclassified according to the fine segmentation and fusion results.

2. A seismic phase identification method based on Dice index according to claim 1, characterized in that: The step 1, after processing the original seismic data obtained from field detection, to obtain a seismic profile image set specifically comprises: Decode the original seismic data obtained from field exploration to obtain seismic trace data; Editing and processing the seismic trace data; After gain recovery, extract CDP gathers; Remove interference noise; The seismic profile image set is obtained by using the correction migration process.

3. A seismic phase identification method based on Dice index according to claim 2, characterized in that: The editing process of the seismic trace data specifically includes: Zero-filling is performed on seismic traces whose signal-to-noise ratio is lower than the threshold; Correct the working channel with reversed polarity.

4. The seismic phase identification method based on Dice index according to claim 2, characterized in that: The method of adopting correction and offset processing to obtain a seismic profile image set specifically includes: adopting static and dynamic correction, common depth point superposition and offset processing to obtain a seismic profile image set.

5. The seismic phase identification method based on Dice index according to claim 2, characterized in that: The removing of interference noise specifically includes: removing the interference noise by using deconvolution and filtering.

6. The seismic phase identification method based on Dice index according to claim 1, characterized in that: The step 3, constructing a fully convolutional DeepLabV3+ neural network model specifically includes: Build a fully convolutional DeepLabV3+ neural network model and introduce a spatial pyramid pooling module with dilated convolutions; Use dilated convolution with different dilation coefficients; The pre-trained model is used to initialize the model weights.

7. The seismic phase identification method based on Dice index according to claim 1, characterized in that: In step 4, the training set is input into the fully convolutional neural network model for training to obtain a trained fully convolutional neural network model: Input the training set into the fully convolutional neural network model for training; In the learning process, the Dice exponential loss function L is used. Dice and the multivariate cross entropy loss function L CE The mixed loss function L loss To describe the closeness between the classification results and the real strata; On the basis of the multivariate cross entropy loss function, the Dice exponential loss function is introduced, so that the fully convolutional neural network model can still better evaluate the similarity of classification results under the condition of extremely unbalanced classification samples; The validation set is used to adjust the model parameters during the training process to obtain a trained fully convolutional neural network model.

8. A seismic phase identification method based on Dice index according to claim 7, characterized in that: The hybrid loss function L loss The calculation formula is: Among them, λ CE represents the weight of the multivariate cross entropy loss function; λ Dice represents the weight of the Dice index loss function; C represents the total number of seismic phases; N represents the total number of pixels contained in the input seismic profile; y c,n Indicates label data, whether the nth pixel point on the actual seismic profile is the cth seismic phase, if yes, it is 1, otherwise it is 0; z c,n represents the output result of the neural network, and its physical meaning is the possibility that the corresponding point is predicted to be the c-th type of earthquake phase; p c,n Indicates z c,n The result after Softmax function normalization; ε represents the smoothing coefficient.

9. The seismic phase identification method based on Dice index according to claim 1, characterized in that: The step 6, according to the coarse segmentation result output by the full convolutional neural network, obtains the fine segmentation fusion result after being processed by the fully connected CRF model, specifically includes: According to the coarse segmentation results output by the fully convolutional neural network; Initialize the energy function in the fully connected conditional random fields (CRF) model to obtain the original attribute label probability value of the pixel point; The fully connected CRF model is calculated, and the maximum probability spectrum in the seismic profile segmentation image is iteratively corrected to obtain the fine segmentation fusion result after being processed by the fully connected CRF model.

10. A seismic phase identification method based on Dice index according to claim 9, characterized in that: The steps of calculating the fully connected CRF model specifically include: Step 6.1, construct the energy function E(x|I), which is composed of the unary potential function θ i (x i ) and the binary potential function θ ij (x i ,x j )constitute; In step 6.1, the calculation formula of the energy function E(x|I) of the fully connected CRF for the label assignment scheme x is defined as: i i (x i )=-logP(x i ) Where I represents the current input pixel vector; x represents the label predicted by the current observed pixel; θ i (x i ) represents a univariate potential function, defined at the observation position x i The state characteristic function of x i Represents the current input pixel; P(x i ) the probability value of the seismic phase classification of the pixel; θ ij (x i ,x j ) represents the binary potential function, which is defined as the transfer characteristic function at different observation positions; μ(x i ,x j ) indicates a label compatible item, when x i and x j When the earthquake phase prediction classification labels are inconsistent, the value is 1, otherwise it is 0; ω1 represents the weight parameter of the surface kernel; ω2 represents the weight parameter of the smoothing kernel; p i represents the position code of pixel i; p i represents the position code of pixel j; Ii represents the color code of pixel i; I j represents the color code of pixel j; σ α represents the variance coefficient of the position vector p; σ β represents the variance coefficient of the color vector I; σ γ represents the variance coefficient of the smooth term; Step 6.2, normalize the energy function value to obtain the probability value of the pixel belonging to the seismic phase classification label, and select the label corresponding to the pixel according to the maximum probability, until the probability value of all pixels belonging to their respective classification labels reaches more than 90% and then exit the iteration; In step 6.2, the posterior probability value of the pixel is calculated as: Among them, Z(I) represents the normalization factor, and x represents the label predicted by the current observed pixel; The seismic phase segmentation map with refined boundaries is obtained.

11. A seismic phase identification method based on Dice index according to claim 1, characterized in that: The step 7 of dividing the connected domains to be reclassified according to the fine segmentation and fusion results also includes: The eight-domain seismic phase classification categories are counted for the boundaries of the connected domains that need to be reclassified, and the pixel position of the current iteration is updated with the seismic phase classification outside the domain by voting. If the statistical results show that the number of votes for two categories is consistent, the processing is postponed; Among them, if a single pixel is not updated, the equal-vote seismic phase classification is randomly selected as the final optimized classification category of the cover pixel position.

12. A seismic phase identification system based on Dice index, using a seismic phase identification method based on Dice index as described in any one of claims 1 to 11, characterized in that: The identification system comprises: The raw data processing module is used to process the raw seismic data obtained from field detection to obtain a set of seismic profile images; An image set construction module, used for dividing the seismic profile image set, constructing seismic profile image set samples, and dividing the seismic profile image set samples into a training set and a verification set; Model building module, used to build a fully convolutional DeepLabV3+ neural network model; The neural network model training module is used to input the training set into the full convolutional neural network model for training to obtain a trained full convolutional neural network model; An input training module is used to input the seismic section image to be identified into the trained full convolutional neural network model to obtain a rough result image of seismic phase segmentation of the section image; The fine segmentation fusion result calculation module is used to obtain the fine segmentation fusion result after being processed by the fully connected CRF model according to the coarse segmentation result output by the full convolutional neural network; The fusion result division module is used to divide the connected domains that need to be reclassified according to the fine segmentation fusion results.

Images