Seismic horizon identification method based on data enhancement scse_vgg16-unet++

Abstract

The present application is suitable for the field of seismic exploration technology, and provides a scSE_VGG16-UNet++ seismic horizon identification method based on data enhancement, comprising the following steps: constructing a UNet basic network with VGG16 as the backbone; introducing a dense connection and a deep supervision mechanism to improve the VGG16-UNet++ network; in view of the characteristics of seismic data, using multiple horizontal transformations and noise addition methods for data enhancement; designing a scSE attention mechanism module to simultaneously enhance the expression ability of feature channels and spatial dimensions; embedding the scSE attention mechanism module into the skip connection of the VGG16-UNet++, forming a final identification model; using cross-entropy loss and Adam optimizer to train the network; finally evaluating the identification effect through the confusion matrix, mIoU and PA indicators. The present application effectively solves the problem of discontinuous and mislayering in seismic horizon identification caused by data scarcity and similar features, and significantly improves the identification accuracy, continuity and model generalization ability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of seismic exploration, and particularly relates to a scSE_VGG16-UNet++ seismic horizon identification method based on data enhancement. BACKGROUND

[0002] Seismic exploration is the core means of oil and gas resource exploration, and its key link lies in accurately identifying the reflection horizon representing the geological interface from the seismic data, i.e., seismic horizon interpretation. The accurate identification of the horizon directly determines the reliability of the underground structure analysis and is crucial for oil and gas reservoir prediction.

[0003] Traditional horizon identification methods mainly include manual picking and automatic identification. The automatic identification technology includes methods based on the coherent algorithm and the image processing algorithm. These methods have improved the efficiency, but generally rely on preset rules and local features, and have significant limitations in accuracy, continuity and adaptability when dealing with complex geological conditions such as low signal-to-noise ratio, discontinuous reflection or the presence of faults.

[0004] In recent years, artificial intelligence technology represented by deep learning, especially convolutional neural networks and their variants, has provided a new way for intelligent horizon identification due to its strong feature learning ability, effectively improving the identification efficiency and accuracy. However, existing intelligent methods still face challenges such as insufficient deep feature extraction capability, weak generalization due to sensitivity to data similarity, and poor robustness in identifying complex structures such as discontinuity and mislayering. Therefore, it is urgent to develop an intelligent identification method with better robustness and generalization ability to further improve the automation level and reliability of seismic horizon interpretation. SUMMARY

[0005] The purpose of the embodiment of the application is to provide a scSE_VGG16-UNet++ seismic horizon identification method based on data enhancement, aiming to solve the problems raised in the background technology.

[0006] The embodiment of the application is implemented in the following way: the scSE_VGG16-UNet++ seismic horizon identification method based on data enhancement includes the following steps:

[0007] Step 1: replace the backbone feature extraction network in the UNet encoder with the first 5 convolution blocks of the VGG16 network to form a VGG16-UNet network;

[0008] Step 2: introduce dense connection and deep supervision mechanism on the basis of the VGG16-UNet network to construct a VGG16-UNet++ network model;

[0009] Step 3: perform synchronous enhancement processing on the original seismic profile data and the corresponding horizon label data;

[0010] Step 4: constructing an scSE attention mechanism module containing a channel attention module and a spatial attention module at the same time;

[0011] Step 5: embedding the scSE attention mechanism module in the skip connection between the encoder and the decoder of the VGG16-UNet++ network to form an scSE_VGG16-UNet++ network model, and training the model using enhanced data;

[0012] Step 6: training the scSE_VGG16-UNet++ network model using a cross-entropy loss function and an Adam optimization algorithm;

[0013] Step 7: evaluating the trained model using the mean intersection over union and the pixel accuracy.

[0014] The scSE_VGG16-UNet++ seismic layer identification method based on data enhancement provided by the embodiment of the application has the following beneficial effects:

[0015] (1) The application constructs a VGG16-UNet architecture based on the traditional UNet and in combination with the deep feature extraction capability of VGG16, and first applies it to seismic layer identification. The experimental results show that the VGG16-UNet can effectively realize seismic layer identification.

[0016] (2) In view of the high cost of seismic data labeling and the similar sample features of a single work area, the specific data enhancement technology is used to effectively enrich the features of the seismic profile according to the strong lateral characteristics of the seismic layer. The experimental results show that the data enhancement technology can not only effectively alleviate the discontinuity problem of layer identification caused by the limitation of data features, but also improve the recognition ability of the network architecture in noisy data.

[0017] (3) By introducing the scSE attention mechanism module at the skip connection of the VGG16-UNet, the network architecture can pay attention to the boundary information of the layer, and suppress the interference of noise and non-target regions. The evaluation index in the comparative experiment shows that the introduction of the scSE attention mechanism module effectively improves the accuracy and continuity of the recognition result.

[0018] (4) On the basis of the scSE_VGG16-UNet method, the dense connection and deep supervision mechanism of UNet++ are introduced, combined with data enhancement technology, and a scSE_VGG16-UNet++ seismic horizon identification method using data enhancement is proposed. Whether in the quantitative indicators such as PA and mIoU, or in the test of theoretical model and actual data, the continuity and accuracy of the identification result are better than those of VGG16-UNet and the intermediate improved model, and it also shows strong identification ability in the noisy theoretical model data, which proves the application potential of the model in intelligent identification of target horizon in complex seismic data, and has strong generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flow chart of the scSE_VGG16-UNet++ seismic horizon identification method based on data enhancement provided for the embodiments of the application is shown in the figure.

[0020] Figure 2 The VGG16-UNet++ architecture diagram is shown in the figure.

[0021] Figure 3 The channel dimension attention mechanism module is shown in the figure.

[0022] Figure 4 The spatial dimension attention mechanism module is shown in the figure.

[0023] Figure 5 The scSE attention mechanism module is shown in the figure.

[0024] Figure 6 The scSE_VGG16-UNet++ architecture diagram is shown in the figure.

[0025] Figure 7 The theoretical model data and its label data are shown in the figure (wherein a is a Synmarine model, b is a Synmarine model label, c is a Sigmoid model, and d is a Sigmoid model label).

[0026] Figure 8 The actual data and its label data are shown in the figure (wherein a is A work area seismic data, b is the corresponding label of A work area, c is F3 work area data, and d is the corresponding label of F3 work area).

[0027] Figure 9 The data enhancement processing is shown in the figure (wherein a is the original data, b is horizontal flip, c is horizontal translation, d is horizontal stretching, e is horizontal compression, f is horizontal shear transformation, g is Gaussian noise, and h is Gaussian blur).

[0028] Figure 10Synmarine model test results (wherein a is a Sigmoid model, b is labeled data, c is a VGG16-UNet recognition result, d is a data enhancement + VGG16-UNet recognition result, e is a data enhancement + scSE_VGG16-UNet recognition result, and f is a data enhancement + scSE_VGG16-UNet++ recognition result);

[0029] Figure 11 Synmarine model test results (wherein a is a Sigmoid model, b is labeled data, c is a VGG16-UNet recognition result, d is a data enhancement + VGG16-UNet recognition result, e is a data enhancement + scSE_VGG16-UNet recognition result, and f is a data enhancement + scSE_VGG16-UNet++ recognition result);

[0030] Figure 12 Synmarine model test results (wherein a is a Sigmoid model, b is labeled data, c is a VGG16-UNet recognition result, d is a data enhancement + VGG16-UNet recognition result, e is a data enhancement + scSE_VGG16-UNet recognition result, and f is a data enhancement + scSE_VGG16-UNet++ recognition result);

[0031] Figure 13 Synmarine model test results (wherein a is a Sigmoid model, b is labeled data, c is a VGG16-UNet recognition result, d is a data enhancement + VGG16-UNet recognition result, e is a data enhancement + scSE_VGG16-UNet recognition result, and f is a data enhancement + scSE_VGG16-UNet++ recognition result);

[0032] Figure 14 Synmarine model test results (wherein a is a Sigmoid model, b is labeled data, c is a VGG16-UNet recognition result, d is a data enhancement + VGG16-UNet recognition result, e is a data enhancement + scSE_VGG16-UNet recognition result, and f is a data enhancement + scSE_VGG16-UNet++ recognition result);

[0033] Figure 15 Synmarine model test results (wherein a is a Sigmoid model, b is labeled data, c is a VGG16-UNet recognition result, d is a data enhancement + VGG16-UNet recognition result, e is a data enhancement + scSE_VGG16-UNet recognition result, and f is a data enhancement + scSE_VGG16-UNet++ recognition result). DETAILED DESCRIPTION

[0034] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0035] The specific implementation of the present application will be described in detail below in combination with specific embodiments.

[0036] As shown in the figure, the data enhancement-based scSE_VGG16-UNet++ seismic horizon identification method provided by an embodiment of the present application includes the following steps: Figure 1 Step 1: constructing a VGG16-UNet basic network model for seismic horizon identification based on a classic UNet architecture;

[0037] The classic UNet architecture is composed of four types of layers: convolutional layers, pooling layers, dropout layers and deconvolutional layers. The VGG16-UNet combines the VGG16 and UNet architectures, and this combination provides an efficient solution in image segmentation tasks, especially in horizon efficient identification problems.

[0038] The VGG16-UNet is a modification of the backbone feature extraction network in the UNet, and the network structure is applied to seismic horizon identification for the first time in the present method. The UNet is divided into an encoder and a decoder, wherein the encoder contains a backbone feature extraction network, which is modified according to the VGG16 architecture. The backbone network is modified to the first five blocks of the VGG16 architecture, wherein each block contains 64, 128, 256, 512 and 512 convolutional layers, respectively. The convolutional step length of each layer is 1, and the padding is also 1. A maximum pooling layer is provided at the end of each block, and the pooling size is 2. Such a setting helps to gradually extract high-level features of the input image and reduce its spatial dimension. The decoder part is based on the UNet, which effectively recovers and refines the image size and details by gradually upsampling and fusing the features of the encoder stage, so as to realize accurate image segmentation.

[0039]

[0040] ​Compared with UNet, the encoder part of VGG16-UNet increases three convolution layers and their corresponding three ReLU activation functions, which can better identify the target layer information. Meanwhile, in the fifth convolution block of VGG16-UNet, 512 convolution layers are used instead of 1024 convolution layers in UNet, which makes the network have fewer training parameters as a whole. Compared with VGG16, VGG16-UNet removes the fully connected layer in VGG16 to optimize the performance of image segmentation. This change not only reduces the number of parameters and computational complexity of the network architecture, but also better preserves the spatial information of the feature map, making the network architecture adapt to different resolutions of input. By combining the deep feature extraction capability of VGG16 and the efficient segmentation mechanism of UNet, VGG16-UNet achieves higher accuracy.

[0041] Step 2: Based on the VGG16-UNet base network model, introduce dense connection and deep supervision mechanism to build VGG16-UNet++ network model;

[0042] Although VGG16-UNet achieves efficient segmentation, it also has some defects such as unknown optimal depth of network, and jump connection can only fuse features at the same scale. Figure 2 As shown in the figure, VGG16-UNet++ uses dense connection and deep supervision structure. Compared with UNet, which starts upsampling after downsampling to the deepest layer, VGG16-UNet++ gets a feature prediction through upsampling after each downsampling, and finally fuses the prediction results of different scales to output the final prediction result. VGG16-UNet++ not only integrates features at different levels to improve the accuracy of the result, but also can flexibly select the appropriate network depth, greatly reducing the parameter amount within the acceptable range of accuracy.

[0043] Step 3: Perform synchronous enhancement processing on the original seismic profile data and corresponding layer label data;

[0044] Unlike general semantic segmentation tasks, which have at least several thousand training sets with different features, seismic data labeling is time-consuming and varies greatly between different regions. In the same block, only a few sections of training data can be obtained. Therefore, data enhancement techniques are used to expand the processed theoretical and actual data to further enhance the network's feature extraction capability. Unlike traditional data enhancement methods such as flipping and cropping, considering the strong lateral characteristics of seismic layer, six enhancement methods are used, including horizontal flipping, horizontal translation, horizontal scaling, horizontal shear transformation, Gaussian noise addition, and Gaussian blur, to simulate geological changes and noise interference and enrich the features of training data.

[0045] Step 4: Constructing the scSE attention mechanism module

[0046] To further enhance the network's ability to express the characteristics of the layer, an attention mechanism is introduced in both the spatial and channel dimensions of the feature map to construct the scSE attention mechanism module. The scSE attention mechanism module improves the expressiveness of the network by focusing on both the spatial and channel dimensions simultaneously. This mechanism includes two main parts: channel attention (Channel Squeeze & Excitation, cSE) and spatial attention (Spatial Squeeze & Excitation, sSE). The cSE module highlights important channels by compressing the spatial dimension, while the sSE module highlights important spatial regions by compressing the channel dimension.

[0047] The channel dimension attention module (i.e., the cSE module) is a type of attention mechanism, as shown in Figure 3 , which is mainly used to strengthen important feature channels in the network. Let the input feature map be:

[0048]

[0049] where is a real set, is the spatial height, is the spatial width, is the number of channels.

[0050] The input feature map is considered as a combination of channels , where is the global average pooling for the th channel , and the feature map containing global information is compressed into in the spatial dimension, obtaining the channel description vector :

[0051]

[0052] where is a scalar representing the channel description vector; , ;

[0053] After passing through two fully connected layers and a ReLU activation function, it is transformed into :

[0054]

[0055] where is the channel-wise feature map after the channel-wise SE module, and are the weight matrices of two fully connected layers, respectively, is the ReLU activation function. After the normalization by the sigmoid activation function, we get and then get the feature map re-scaled in the channel dimension :

[0056]

[0057] where is the output of the cSE module, is the sigmoid activation function, represents the importance of the th channel;

[0058] The spatial attention module (i.e., sSE module) is the part of the attention mechanism that specifically processes spatial features, as shown in Figure 4 . The input feature map is considered as another form of slicing, i.e., sliced by spatial positions into , and is applied with a convolution kernel . The output of the convolution is , where corresponds to the spatial position , and , ; is the convolution kernel weight, is the projection tensor of the convolution output, is the two-dimensional discrete convolution operation.

[0059] Each element in the projection represents the linear combination of all channels at the spatial position , After the normalization by the sigmoid activation function, we get and then get the feature map re-scaled in the spatial dimension :

[0060]

[0061] where is the output of the sSE module, represents the importance of the spatial position ;

[0062] The two attention modules are combined to form an scSE attention mechanism module, as shown in the structure of Figure 5 The combination can simultaneously recalibrate the channel dimension and the spatial dimension to obtain the recalibrated feature map :

[0063]

[0064] wherein, is the output of the scSE module.

[0065] In this way, the scSE attention mechanism module effectively enhances the expression ability of the features and enhances the model's ability to extract key features in the image.

[0066] Step 5: Embed the scSE attention mechanism module constructed in step 4 into the level jump connection between the encoder and the decoder of the VGG16-UNet++ network model constructed in step 2. Through this integration, the network can adaptively strengthen the key features related to the layer position and suppress irrelevant information when the features are fused, forming the final scSE_VGG16-UNet++ recognition model. Then, combine the data enhancement results of step 3 to train the model;

[0067] Step 6: Set the loss function and optimization algorithm to complete the network training;

[0068] Use the cross-entropy loss function to improve the classification accuracy of model training, and use the Adam optimization algorithm to adaptively adjust the learning rate of the model parameters. Train the network model constructed in step 5 end-to-end.

[0069] For classification problems, the cross-entropy loss can be defined as:

[0070]

[0071] wherein, is the number of samples, is the number of classes, indicates whether the i-th sample belongs to class c (0 or 1), and indicates the probability that the i-th sample belongs to class c. Through this loss function, the probability difference between the model prediction and the true label can be reduced, thereby improving the classification accuracy of the model.

[0072] Step 7: Evaluate the layer position recognition results using evaluation indicators.

[0073] After the network training is completed, as shown in Table 1, the confusion matrix is used to construct the evaluation indicators to quantitatively evaluate the layer position recognition results.

[0074] Table 1 Confusion matrix

[0075]

[0076] In Table 1, TP represents correctly predicting the target horizon as the target horizon, FP represents incorrectly predicting the non-target horizon as the target horizon, FN represents incorrectly predicting the target horizon as the non-target horizon, and TN represents correctly predicting the non-target horizon as the non-target horizon.

[0077] The mean intersection over union (mIoU) and the pixel accuracy (PA) can be calculated through the confusion matrix to comprehensively evaluate the horizon recognition result.

[0078] Mean intersection over union The calculation formula is as follows:

[0079]

[0080] Pixel accuracy refers to the proportion of pixels with correct prediction categories in the total pixels, and the calculation formula is as follows:

[0081] .

[0082] In the embodiments of the present application, in order to verify the effectiveness of the method, theoretical models and actual data are used for experimental verification.

[0083] (1) Data set construction: Before constructing the training set, the data is normalized to eliminate the difference in data feature size, so that the model can extract data features.

[0084] Theoretical data uses the Synmarine model containing multiple hyperbolic events and the Sigmoid model, as shown in Figure 7 24 groups of 256x256 two-dimensional data are cut from the model.

[0085] The actual data comes from the A area 3D seismic data and the Netherlands F3 area 3D seismic data. The A area 3D seismic data is 201x201x201 3D data with a sampling frequency of 4ms. The Netherlands F3 area has 401 main lines and 701 contact lines, with a time range of 1848ms and a sampling rate of 4ms. In order to facilitate subsequent network training, the present method uses a non-overlapping cutting method from the two actual data, as shown in Figure 8 901 groups of 256x256 two-dimensional data are cut, and are divided into training set, validation set and test set in the ratio of 8:1:1.

[0086] Data augmentation is performed on the theoretical model and actual data, such as Figure 9 As shown in FIG. 1, six augmentations are performed, including horizontal flip, horizontal translation, horizontal scaling, horizontal shear transformation, Gaussian noise addition, and Gaussian blur.

[0087] (2) Experimental setup: the network model is built and trained using the Pytorch framework;

[0088] First, the key parameters of the model are configured. In the training process, the cross-entropy loss function is used to improve the classification accuracy of the model, and the Adam optimizer is used to adjust the parameters. The initial learning rate is set to 0.001, and the learning rate is automatically adjusted according to the training progress to adapt to the needs of different training stages. The batch size is set to 8, and a total of 100 rounds of training are performed. The hardware configuration of the experiment is shown in Table 2.

[0089] Table 2 Experimental environment

[0090]

[0091] (3) Comparative experiment

[0092] To verify the performance of the method, the quantitative evaluation indicators of data augmentation technology, VGG16 module, scSE attention mechanism module, and UNet++ on the theoretical model and actual data are analyzed comprehensively, as shown in Table 3, which proves the effectiveness of the method in layer recognition.

[0093] Table 3 Evaluation indicators of each network architecture (%)

[0094]

[0095] As shown in Table 3, the PA and mIoU of the scSE_VGG16-UNet++ method using data augmentation are higher than those of VGG16-UNet, data enhanced VGG16-UNet, and data enhanced scSE_VGG16-UNet. The PA is increased by 5.04%, 4.78%, and 3.81%, respectively. The mIoU is increased by 68.71%, 57.86%, and 56.30%, respectively. Therefore, the scSE_VGG16-UNet++ method using data augmentation has stronger recognition ability.

[0096] (4) Theoretical model test: first, the Synmarine model and Sigmoid model without noise are tested by VGG16-UNet, data augmentation + VGG16-UNet, data augmentation + scSE_VGG16-UNet, and data augmentation + scSE_VGG16-UNet++, respectively, as shown in Figure 10 and Figure 11 .

[0097] Six target horizons are included in the Synmarine model (e.g. Figure 10 a and Figure 10 b), and the identification results of the four methods are relatively continuous for the three target horizons H1, H3 and H5, but the identification results of the first three methods are poor in continuity for the three target horizons H2, H4 and H6 (e.g. Figure 10 c, Figure 10 d and Figure 10 e), while the identification results of the present method are continuous and smooth for the six target horizons (e.g. Figure 11 f).

[0098] The Sigmoid model is composed of an unconformity and its underlying strata, and the target horizons are relatively complex (e.g. Figure 11 a and Figure 11 b). For the tilted strata of 0-100ms and 400-600ms, the identification results of the first three methods all have discontinuous phenomena (e.g. Figure 11 c, Figure 11 d and Figure 11 e), and the VGG16-UNet has a relatively obvious discontinuous phenomenon at the position indicated by the arrow (e.g. Figure 11 c). In the syncline area of the underlying strata of the tilted strata in the square box area of 0.5-1.5km and the anticline area of the ellipse and the rounded square box area of 1.5-25km, the identification results of the first three methods all have relatively obvious discontinuous phenomena (e.g. Figure 11 c, Figure 11 d and Figure 11 e), and the identification result of the data enhancement + scSE_VGG16-UNet has the most serious discontinuous phenomenon (e.g. Figure 11 e). At the same time, in the fault area in the rounded rectangle, the first three methods all have the phenomenon of mislayer identification (e.g. Figure 11 c, Figure 11 d and Figure 12 e). The identification results of the present method in the tilted strata area and the underlying strata area are continuous and clear, and do not have the phenomenon of mislayer identification (e.g. Figure 12 f).

[0099] After adding Gaussian noise with a variance of 0.01 to the Synmarine model (e.g. Figure 13 a and Figure 13 b) and the Sigmoid model (e.g. Figure 12 a and Figure 13 b), the four methods are tested respectively, and the test results are shown in Figure 12 and Figure 12 .

[0100] For the Synmarine model with noise, the four methods are also relatively continuous in the identification results of the three target horizons H1, H3 and H5. However, the identification results of the three target horizons H2, H4 and H6 by the first three methods all have discontinuous phenomena (such as Figure 12 c、 Figure 12 d and Figure 12 e), among which VGG16-UNet fails to identify the H4 horizon ( Figure 13 c). However, the six target horizons identified by the method are continuous and smooth ( Figure 13 f).

[0101] For the Sigmoid model with noise, the identification results of the first three methods all have discontinuous phenomena (such as Figure 13 c、 Figure 13 d and Figure 13 e) for the 0-100ms and 400-600ms inclined strata. The identification result of data augmentation + scSE_VGG16-UNet has a relatively obvious discontinuous phenomenon at the position indicated by the arrow ( Figure 13 e). In the syncline area in the 0.5-1.5km underlying strata of the inclined strata, the ellipse and the rounded square area in the anticline of 1.5-25km, the identification results of the first three methods have relatively obvious discontinuous phenomena and layer errors (such as Figure 13 c、 Figure 13 d and Figure 14 e), while the identification result of the method is more continuous, more in line with the horizon trend, and does not have the layer error identification phenomenon ( Figure 15 f).

[0102] Based on the above test analysis of the theoretical model, the use of data augmentation technology can effectively enrich the characteristics of the seismic profile, and the method combining scSE attention and UNet++ can better identify continuous horizons, and has strong anti-noise ability.

[0103] (5) Actual work area application: To test the application effect of the above method in the actual work area, two actual data of A work area and F3 work area in the Netherlands (such as Figure 14 a and Figure 15 a) are selected for testing. The test results are shown in Figure 14 and Figure 14 .

[0104] The test profile of A work area has three target horizons ( Figure 14 b), and the identification results of the four methods for the two horizons H1 and H2 are relatively continuous. However, for H3, the identification result of VGG16-UNet has a relatively obvious discontinuous phenomenon ( Figure 14 c), and the identification result of data augmentation + VGG16-UNet is not continuous in 0-20 channels (Figure 14 d), although data augmentation + scSE_VGG16-UNet and data augmentation + scSE_VGG16-UNet++ have better continuity in the recognition results of the three target horizons, the recognition result of data augmentation + scSE_VGG16-UNet++ is more consistent with the target horizon (e.g. Figure 15 e and Figure 15 f).

[0105] The F3 work area test profile also contains three target horizons (H1, H2 and H3) Figure 15 b). The recognition results of VGG16-UNet for H1, H2 and H3 are not continuous, especially the recognition result of H1 has a relatively obvious discontinuity phenomenon (c). Figure 15 c). The recognition result of data augmentation + VGG16-UNet for H2 is better than that of VGG16-UNet (d) after enriching the profile features by data augmentation. Figure 15 d). The recognition results of data augmentation + scSE_VGG16-UNet for H1 and H2 are relatively continuous, but the recognition result of H3 is not consistent with the trend of the same phase axis (e). ​ e). The recognition result of data augmentation + scSE_VGG16-UNet++ is better than that of the above three methods, and the recognition result is more consistent with the trend of the same phase axis (f). ​

[0106] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.​

Claims

1. A method for seismic horizon identification based on data enhancement scSE_VGG16-UNet++, characterized in that, The method comprises the following steps: Step 1: replace the backbone feature extraction network in the UNet encoder with the first 5 convolutional blocks of the VGG16 network to form a VGG16-UNet network; Step 2: introduce a dense connection and a deep supervision mechanism on the basis of the VGG16-UNet network to construct a VGG16-UNet++ network model; Step 3: perform synchronous enhancement processing on the original seismic profile data and the corresponding horizon label data; Step 4: construct a scSE attention mechanism module containing a channel attention module and a spatial attention module; Step 5: embed the scSE attention mechanism module in the skip connection between the encoder and the decoder of the VGG16-UNet++ network to form a scSE_VGG16-UNet++ network model, and train the model using the enhanced data; Step 6: train the scSE_VGG16-UNet++ network model using a cross-entropy loss function and an Adam optimization algorithm; Step 7: evaluate the horizon identification results of the trained model using the mean intersection over union and the pixel accuracy; The step 4 comprises the following specific steps: For the attention module of the channel dimension, let the input feature map be: ； wherein, is a real set, is a spatial height, is a spatial width, is a number of channels; input feature map is regarded as a channel , wherein , a global average pooling is performed on the first channel to compress the feature map containing global information in the spatial dimension into , to obtain a channel description vector : ； wherein is a scalar, representing a channel description vector; , ; After transformation by two fully connected layers and ReLU activation function : ； wherein, is the transformed channel description vector, and are the weight matrices of two fully connected layers, respectively, is the ReLU activation function; after normalization by the sigmoid activation function, we get and then get the feature map re-scaled in the channel dimension : ； wherein, is the output of the cSE module, is the sigmoid activation function, denotes the importance of the th channel; For the attention module of spatial dimension, the input feature map is regarded as another form of slicing, i.e. being divided into ; is a two-dimensional discrete convolution operation;​​​​​​​​​ Each element in the projection Indicates spatial location All passages linear combination, After normalization using the sigmoid activation function, the result is... This leads to the understanding of spatial dimensions. Recalibrated feature map : ； wherein, is the output of the sSE module, denotes the importance of the spatial position at which the sSE module is applied. Combining the two attention modules above forms an scSE attention mechanism module, which can simultaneously perform rescaling in the channel dimension and the spatial dimension to obtain a rescaled feature map :​ ； wherein, is the output of the scSE module.

2. The data augmentation based scSE_VGG16-UNet++ seismic horizon identification method according to claim 1, characterized in that In the step 1, the first 5 convolutional blocks of the VGG16 network successively contain convolutional layers with 64, 128, 256, 512, and 512 convolutional layers, and each convolutional block is followed by a maximum pooling layer; the VGG16-UNet network removes the fully connected layer in the VGG16 network.

3. The data augmentation based scSE_VGG16-UNet++ seismic horizon identification method according to claim 1, characterized in that, In the step 2, the VGG16-UNet++ network model performs upsampling prediction after each downsampling, and fuses the feature prediction results of different scales through dense connection.

4. The data augmentation based scSE_VGG16-UNet++ seismic horizon identification method according to claim 1, characterized in that, In the step 3, the data enhancement processing includes the following modes: horizontal flip, horizontal translation, horizontal scaling, horizontal shear transformation, Gaussian noise addition, and Gaussian blur.

5. The data augmentation based scSE_VGG16-UNet++ seismic horizon identification method according to claim 1, characterized in that, In the step 5, the scSE attention mechanism module is added to each skip connection of the VGG16-UNet++ network in parallel to recalibrate the features before the encoder features and the decoder features are fused.

6. The data augmentation based scSE_VGG16-UNet++ seismic horizon identification method according to claim 1, characterized in that, In the step 6, the cross-entropy loss function is used to improve the classification accuracy of model training, and the Adam optimization algorithm is used to adaptively adjust the learning rate of the model parameters to train the network model constructed in step 5 end-to-end; For classification problems, the cross-entropy loss is defined as: ； wherein, is the number of samples, is the number of classes, denotes an indication whether the ith sample belongs to class c, denotes a probability whether the ith sample belongs to class c; by this loss function the probability difference between the model prediction and the true label is reduced, improving the classification accuracy of the model.

7. The data augmentation based scSE_VGG16-UNet++ seismic horizon identification method according to claim 1, characterized in that, In the step 7, the average intersection over union The calculation formula is as follows: ； Wherein, TP represents correctly predicting the target horizon as the target horizon, FP represents incorrectly predicting the non-target horizon as the target horizon, FN represents incorrectly predicting the target horizon as the non-target horizon, and TN represents correctly predicting the non-target horizon as the non-target horizon; is referred to as the prediction class correct pixel ratio, and is calculated as follows: 。

Images