A small-sample pre-stack seismic reflection pattern analysis method based on large-core attention

Abstract

The present invention discloses a small-sample pre-stack seismic reflection pattern analysis method based on large-kernel attention, which is applied to the field of seismic data processing. Aiming at the problem of insufficient recognition ability of the existing technology for complex geological structures, the present invention constructs an image classification seismic reflection pattern analysis model based on an improved ConvNext framework; by introducing a large-kernel attention (LKA) mechanism, the receptive field of the model is increased, the model's ability to sense global information is enhanced, and the spatial lateral characteristics and reflection information of the strata are fully utilized and learned; then the model in step S1 is trained with the augmented labeled pre-stack seismic data; finally, the unlabeled pre-stack seismic pictures are predicted according to the model trained in step S2, and the results are inserted into the seismic profile according to the xline and inline coordinates of the pictures, and finally the overall profile prediction result is obtained.

Description

Technical Field

[0001] The present invention belongs to the field of seismic data processing, and particularly relates to a pre-stack seismic reflection mode analysis technology. Background Art

[0002] Seismic exploration is a commonly used method in the field of geophysics, providing a necessary basis for the identification of oil and gas reservoirs. In seismic exploration, different underground lithological parameters, different structural features (such as fractures, karst caves), etc. will all bring differences in seismic reflection signals. In order to emphasize the similarity of the same type of seismic reflection signals and the differences between different types of seismic reflection signals, we define the unique features of a class of seismic reflection signals, which are different from other seismic reflection signals, as the mode of this class of seismic reflection signals. By analyzing the seismic reflection mode, the common features of the same type of seismic reflection signals can be mined, different underground geological structures can be predicted, providing a theoretical basis for subsequent geological exploration.

[0003] In the early stage, seismic reflection mode analysis was usually carried out manually, which highly relied on expert knowledge and rich experience, and the results were highly subjective. Moreover, some subtle reflection information in seismic data was difficult to observe and describe manually. In recent years, seismic reflection mode analysis technology has begun to develop towards automation and intelligence, and artificial intelligence methods such as k-means, SVN, and CLS have also gradually been applied to seismic mode analysis, which can more effectively describe the changes in formation lithology and related formation structures, and have the characteristics of rapidity, quantification, and objectivity.

[0004] Previously, limited by the signal-to-noise ratio of seismic signals, the collected multi-covered raw data was usually selected for processing and stacking to form post-stack data, and seismic signal reflection mode analysis was carried out based on single-channel post-stack seismic signals. However, due to the loss of multi-angle information in post-stack seismic signals, they cannot represent the changes of seismic reflection signals with azimuth or offset, and the reflection mode analysis of a single channel cannot make full use of the spatial lateral characteristics of the formation. Therefore, the reflection mode analysis based on post-stack seismic data often has difficulty in accurately predicting underground reservoirs. Nowadays, with the development of seismic signal processing technology, the quality of pre-stack multi-channel seismic signals has been significantly improved. The intelligent analysis method based on pre-stack signals can make more full use of the rich reflection information between formations, effectively reduce the multi-solution of reservoir analysis, and make geological exploration continue to develop towards refinement.

[0005] According to whether well logging labels are utilized during the classification process, seismic signal reflection pattern analysis methods can be divided into two categories: supervised and unsupervised. Currently, traditional seismic reflection pattern analysis based on pre-stack data is basically an unsupervised method. The unsupervised approach mainly focuses on clustering algorithms and feature mapping methods: Clustering algorithms start from the overall data and automatically classify seismic reflection signals into different clusters according to different differential criteria, and each type of signal belongs to the same seismic reflection pattern. Feature mapping methods transform seismic attributes through a certain transformation to facilitate the display of underlying structural features or reservoir characteristics. The supervised approach is to establish constraints based on existing well logging information and obtain an optimal classification model through manual or computer training, thereby realizing seismic reflection pattern recognition. Among them, supervised algorithms can make full use of the accurate information of well logging data and have high credibility, so they have gradually received extensive attention.

[0006] In actual projects, due to the high cost of obtaining well logging data, the quantity is often very scarce and cannot meet the training needs of supervised models. Under such small sample conditions, overfitting is likely to occur during model training, making it difficult to apply to the reflection pattern analysis of a large number of unlabeled pre-stack seismic data. At the same time, current traditional supervised models are based on the idea of traditional image processing, with a small receptive field and unable to capture the connection between different angle gathers in pre-stack seismic data.

[0007] Supervised learning represents a machine learning task of learning internal features from labeled data and inferring target values. Supervised seismic reflection pattern analysis mainly takes the convolutional neural network as the core architecture, establishes a learning task by constructing a network model, and then trains the model parameters with the help of labels (usually well logging data), ultimately enabling the model to predict seismic facies. Among them, according to the task category of the constructed model, supervised seismic pattern analysis can be roughly divided into two methods: image classification and semantic segmentation.

[0008] 1. Seismic pattern analysis based on image classification

[0009] An image classification method refers to calculating the overall category to which an image belongs through an algorithm model. In 2012, Alex Krizhevsky proposed the AlexNet network, which won the first place in the ISLVRC challenge through a more reasonable network construction and greatly increased the attention in the field of image classification. Subsequently, image classification models developed rapidly. In 2014, Simonyan designed the VGG network architecture to achieve a larger receptive field by stacking multiple small convolutional kernels. In 2015, He et al. proposed the residual network architecture and designed the ResNet model, which achieved the fusion between features at different levels by designing residual modules. In 2017, the lightweight network MobileNet proposed by Howard et al. significantly reduced the model parameters while ensuring the training effect and was often used in devices with weak performance such as mobile terminals. In 2021, the Vision Transformer proposed by Alexey et al. introduced the attention mechanism in natural language processing into image processing, divided the image into small patches to calculate self-attention respectively to replace the convolutional network, and significantly improved the classification accuracy.

[0010] In the field of seismic exploration, image classification methods have also been widely used: In 2016, Alaudah et al. proposed a weakly supervised label mapping algorithm that can generate a large amount of training data by inputting a small number of pre-stack labels and use it for seismic pattern analysis. In 2018, Chevitarese et al. proposed that by dividing the pre-stack seismic profile image into multiple small image patches along the length and width and assuming that each patch belongs to the same category, inputting these patches into the classification network for training and prediction and then combining them can obtain the complete seismic profile classification result based on pre-stack data. Dramsch et al. proposed that a sliding window mechanism can be applied to the pre-stack seismic profile to greatly increase the number of cut patches and improve the model training effect. The basic framework of pre-stack seismic pattern analysis based on image classification is as Figure 1 shown.

[0011] 2. Seismic pattern analysis based on semantic segmentation

[0012] In addition to image classification, the method of semantic segmentation can also be used to classify the categories to which images belong, that is, to distinguish different geological targets on the same seismic section. Different from classification, semantic segmentation does not output a single category for each image, but outputs the category to which each pixel on the image belongs. In 2015, Long et al. proposed the FCN model (Full Convolutional Neural Network). By migrating the feature extraction part of the classification task to the segmentation task and adding upsampling to restore the image size, the first semantic segmentation model was constructed. Subsequently, the field of semantic segmentation has developed rapidly and achieved good results in many aspects such as medical images and road scenes. In the same year, Ronneberger et al. proposed the Unet network, which solved the problem of insufficient traditional medical image labels to a certain extent; Chen et al. proposed the Deeplab network, which expanded the receptive field by introducing dilated convolution, obtained global image information, and reduced the loss of image details during downsampling; in 2016, Zhao et al. proposed the PSPNet network, which achieved multi-scale pooling of deep features by introducing a pyramid pooling module, effectively improving the classification effect.

[0013] In the field of geological exploration, the idea of image segmentation has also begun to be widely used: in 2018, Dramsch achieved seismic pattern analysis based on patch-based CNN, and the classification results can be obtained by inputting pre-stack seismic data. In 2019, Alaudah et al. proposed a deconvolution network based on small-scale images and seismic profiles. The model can not only achieve pre-stack seismic pattern analysis, but also complete data augmentation and improve the number of samples. In 2020, Zhang proposed that the Deeplabv3+ network model can be used to extract multi-scale features of pre-stack seismic data through the ASPP pyramid structure, and good results were also obtained. The basic framework of pre-stack seismic pattern analysis based on semantic segmentation is as Figure 2 shown.

[0014] In the reflection pattern analysis method of pre-stack multi-channel seismic data, traditional supervised methods often rely on rich well logging sample data:

[0015] (1) Seismic pattern analysis based on semantic segmentation requires a large number of labeled complete seismic profiles as training data. However, in actual projects, the acquisition of well logging label data is very expensive and difficult. Under such small sample conditions, overfitting often occurs in model training and cannot be extended to large-scale unlabeled data;

[0016] (2) Although the method based on image classification expands the number of samples to a certain extent by cutting labeled seismic images into patches, the degree of label repetition is relatively high, still unable to meet the needs of model training, and it is necessary to assume that each patch belongs to the same seismic pattern, with low accuracy.

[0017] In addition, existing supervised pre-stack models are directly transplanted from traditional image processing models. When processing, pre-stack seismic data is regarded as ordinary images, and it is impossible to specifically learn the rich unique features of pre-stack seismic data. As Figure 3 shown, the multi-angle gathers obtained by stacking are rich in formation reflection information and anisotropic features, but the receptive fields of traditional supervised models are usually small, making it difficult to capture the global deep features between different-angle gathers at a relatively long distance. Summary of the Invention

[0018] To solve the above technical problems, the present invention proposes a small-sample pre-stack seismic reflection pattern analysis method based on large-kernel attention, introducing the large-kernel attention mechanism into the ConvNext model to learn the deep global features of the data, and completing the pre-stack seismic reflection pattern analysis through single-point prediction and stitching.

[0019] The technical solution adopted by the present invention is: a small-sample pre-stack seismic reflection pattern analysis method based on large-kernel attention, including:

[0020] S1. Construct an image classification seismic reflection pattern analysis model based on an improved ConvNext framework;

[0021] S2. Train the model in step S1 with the augmented labeled pre-stack seismic data;

[0022] S3. Predict the unlabeled pre-stack seismic pictures according to the model trained in step S2, and then stitch the single-point prediction results at all positions into a complete seismic profile prediction result.

[0023] The model described in step S1 is specifically: four cascaded improved ConvNext modules are used to expand the number of channels and extract the deep features of pre-stack seismic pictures. A downsampling module is added between two adjacent improved ConvNext modules to change the size of the pre-stack seismic pictures. Finally, the length and width of the input pre-stack seismic pictures are converted to 1×1 through a global pooling layer, and then the number of channels is mapped to the number of categories to be classified through a fully connected layer, obtaining the probabilities that the input pre-stack seismic pictures are respectively judged as each category, and the maximum probability is the final classification result corresponding to the pre-stack seismic pictures.

[0024] The improved ConvNext module specifically adds LayerNorm regularization, GELU activation function, Layer Scale parameter scaling, and Drop Path layer on the basis of the inverted bottleneck structure; the inverted bottleneck structure is a cascaded structure of an Attention module and two Linear fully connected layers; the Attention module replaces the 7×7 convolutional layer in ConvNext with a large kernel attention layer.

[0025] The process of augmenting pre-stack seismic data with labels in step S2 is as follows: obtaining the overall distribution of different seismic facies on the seismic profile through a clustering algorithm, and then correcting and augmenting it in combination with the accurate labels of well logging data; it includes the following steps:

[0026] A1. Clustering the post-stack seismic profile, classifying the seismic profile according to the clustering results, and framing the parts of the clustering results belonging to the same cluster;

[0027] A2. In the clustering results, project the well logging data into the post-stack seismic profile according to the xline and inline positions of the well logging data, and then augment the clustering results in combination with the accurate seismic facies information and expert experience carried by the well logging interpretation data;

[0028] A3. Calculate the xline and inline number position ranges of the corresponding three-dimensional pre-stack data of the augmented labeled data, and project them onto the three-dimensional pre-stack seismic data volume to obtain the pre-stack multi-channel seismic data at the corresponding positions;

[0029] A4. Draw the pre-stack multi-channel seismic data with labels at each xline and inline position into n curves, and fill the positive value regions of the curves to generate pictures to obtain a large number of pre-stack labeled data. The value of n is determined by the number of channels of the pre-stack seismic data, and one channel of seismic data corresponds to one curve.

[0030] The beneficial effects of the present invention are as follows: In terms of the model, the present invention combines the characteristics of pre-stack seismic signals, and on the basis of the traditional ConvNext model, increases the receptive field of the model by introducing the large kernel attention (LKA) mechanism, enhances the model's ability to sense global information, and makes full use of and learns the lateral spatial characteristics and reflection information of the formation. In terms of data processing, combining the clustering algorithm and well logging labels to augment the number of labels is more scientific and reasonable than the traditional cutting method. At the same time, replacing the input from small-scale image patches with the pre-stack multi-channel seismic data pictures corresponding to each xline and inline position not only augments the labeled data to a greater extent but also further improves the model's recognition ability for complex geological structures.

[0031] The present invention uses a large kernel attention mechanism to mine the global features of pre-stack seismic data and combines the ConvNext image classification model to achieve intelligent seismic pattern analysis based on pre-stack data. Compared with the traditional method of classifying pictures by dividing them into small blocks, the method of the present invention has a higher utilization rate of labeled data. Experiments show that:

[0032] The present invention makes full use of the only well logging data, and by combining machine learning clustering and manual correction, greatly expands the number of labels, effectively improving the model training effect.

[0033] The pre-stack seismic pattern analysis result of the present invention is better than the prediction results of traditional models (such as traditional ConvNext models and AlexNet models). Description of the Drawings

[0034] Figure 1 It is a framework diagram of seismic pattern analysis based on image classification;

[0035] Figure 2 It is a framework diagram of seismic pattern analysis based on semantic segmentation;

[0036] Figure 3 It is a pre-stack multi-channel seismic data diagram;

[0037] Figure 4 It is a framework diagram of pre-stack seismic reflection pattern analysis based on large kernel attention and ConvNext model;

[0038] Figure 5 It is a schematic diagram of the large kernel attention mechanism;

[0039] Figure 6 It is an architecture diagram of the Attention module;

[0040] Figure 7 It is an anti-bottleneck structure based on large kernel attention;

[0041] Figure 8 It is a schematic diagram of the LKA-ConvNext Block and DowmSample module structures;

[0042] Figure 9 It is a network framework diagram;

[0043] Figure 10 It is a clustering result diagram of the post-stack seismic profile;

[0044] Figure 11 It is a schematic diagram of the correction of the post-stack seismic profile;

[0045] Figure 12 It is a schematic diagram of pre-stack seismic data;

[0046] Figure 13 It is a pre-stack label filling picture;

[0047] Figure 14 Predict seismic profiles for different models;

[0048] Among them, (a) the prediction result of the model of the present invention; (b) the prediction result of the traditional ConvNext model; (c) the prediction result of the AlexNet model. Detailed implementation manners

[0049] To facilitate those skilled in the art to understand the technical content of the present invention, the following further explains the content of the present invention with reference to the accompanying drawings.

[0050] The present invention proposes an image classification seismic reflection pattern analysis model based on an improved ConvNext framework and a corresponding seismic data processing method. In terms of the model, the present invention combines the characteristics of pre-stack seismic signals, and on the basis of the traditional ConvNext model, by introducing the large kernel attention (LKA) mechanism, the receptive field of the model is increased, and the model's ability to sense global information is enhanced, making full use of and learning the lateral spatial characteristics and reflection information of the formation. In terms of data processing, combined with the clustering algorithm and well logging labels, the number of labels is expanded, which is more scientific and reasonable compared with the traditional cutting method. At the same time, the input is replaced from a small-scale image patch to the pre-stack multi-channel seismic data picture corresponding to each xline and inline position. While expanding the labeled data to a greater extent, the model's ability to identify complex geological structures is further improved. The specific method and flowchart of the present invention are as Figure 4 shown, including the following steps:

[0051] (1) Based on a small amount of well logging label data, perform label data expansion processing;

[0052] (2) Use the expanded pre-stack seismic data with labels as samples to input the model to predict the category;

[0053] (3) Calculate the loss between the true category and the predicted category of the input pre-stack data, and update the model parameters;

[0054] The model loss selects the cross-entropy loss cross-entropy, and the formula is as follows:

[0055] H(p,q) = -∑p(x)logq(x)

[0056] Among them, p is the true label distribution of the pre-stack data, q is the normalized distribution calculated by the model, and x represents the input seismic data picture;

[0057] (4) Repeat steps (2)-(3) until the loss cannot continue to decrease, indicating that the model training is completed;

[0058] (5) Call the trained model to directly predict the unlabeled prestack seismic images, and then insert the results into the seismic profile according to the xline and inline coordinates of the seismic images to finally obtain the overall profile prediction result.

[0059] 1. Network model

[0060] The network model used in the present invention is an improvement based on the ConvNext network. In the present invention, the 7×7 convolutional layer in ConvNext is replaced with a large kernel attention layer (LKA) that has a more significant effect and is more sensitive to global features. The following will introduce it in detail from two aspects: network composition and overall network architecture.

[0061] 1.1 Network composition

[0062] The improved network model of the present invention is based on the ConvNext network architecture and is composed of alternating stacks of LKA-ConvNext Blocks and downsample layers. Each part mainly uses operations such as large kernel attention and inverted bottleneck structure. Next, the large kernel attention module and inverted bottleneck structure will be introduced first, and then each component module in the network model will be introduced in detail.

[0063] 1.1.1 Large kernel attention (LKA, Large kernel attention)

[0064] In the field of computer vision, there are usually two methods to increase the receptive field and capture long-range information. The first method is the self-attention mechanism based on transformers, and the second method is large kernel convolution. The self-attention mechanism was initially designed for one-dimensional language processing tasks, so when processing images, it also treats the two-dimensional structure as a one-dimensional sequence, destroying the key two-dimensional characteristics of the image. And large kernel convolution introduces a large number of parameters and computational amounts, doubling the model training time. Based on the problems of the above two methods, the present invention introduces the large kernel attention mechanism. The schematic diagram of large kernel attention is as Figure 5 shown. Similar to depthwise separable convolution, it decomposes a convolution with a kernel size of k into the sum of three convolutions, namely, a depth convolution with a kernel size of k / d, a dilated convolution with a kernel size of 2d - 1 and a dilation rate of d, and a pointwise convolution with a kernel size of 1×1. It also has characteristics such as spatial adaptability, channel adaptability, and long-range dependence learning ability. Next, based on the large kernel attention mechanism, activation functions and pointwise convolutions are supplemented to construct a complete Attention module, and the structure is as Figure 6 shown.

[0065] 112. Inverted bottleneck

[0066] The inverted bottleneck structure adopts a pattern of first performing dimensionality-increasing convolution and then dimensionality-reducing convolution, believing that this way can enable information to be converted between different-dimensional feature spaces, thereby avoiding information loss caused by dimensionality reduction during dimensionality compression and improving the model effect. The present invention constructs an inverted bottleneck architecture integrating large-kernel attention by cascading an Attention module and two Linear fully-connected layers, where the height and width of the input and output remain unchanged, and only the number of channels changes. The principle is as Figure 7 shown.

[0067] 113. Network module

[0068] The network module of the present invention mainly includes the LKA-ConvNext Block and the downsampling layer (downSample). The LKA-ConvNext Block refers to the original ConvNext model. By adding LayerNorm regularization, GELU activation function, Layer Scale parameter scaling, and Drop Path layer on the basis of the inverted bottleneck structure, its structure is made more perfect to achieve the extraction of different-level features of the image. In the LKA-ConvNext Block, the dimensions of the input and output remain unchanged. The main function of the downsampling layer is to change the image dimensions through a 2×2 convolution with a stride of 2. The architectures of the two modules are as Figure 8 shown.

[0069] 12. Overall network architecture

[0070] The improved overall network architecture used in the present invention is as Figure 9 shown. First, the pre-stack seismic labels are uniformly scaled to a size of 224×224×3. Then, first pass through a 4×4 convolution and a normalization layer, and then sequentially pass through multiple cascaded LKA-ConvNext Blocks to expand the number of channels and extract deep features. Add three downsampling layers in the middle to change the image size. Finally, through global pooling (Global Avg Pooling), convert the length and width of the input to 1×1, and then map the number of channels to the number of classes (classes) to be classified through a fully-connected layer to obtain the probabilities that the input is judged as each category respectively, where the maximum probability is the classification result judged by the network.

[0071] 2. Data augmentation processing

[0072] For the processing of labeled data, the present invention learns the processing process of handwritten digit recognition in the field of computer vision, and converts multi-channel pre-stack seismic data into a picture form to train the network. The core idea is to obtain the overall distribution of different seismic facies on the seismic profile through a clustering algorithm, and then combine the accurate labels of well logging data for manual correction and expansion. The specific process is as follows:

[0073] (1) Cluster the post-stack seismic profile, classify the seismic profile according to the clustering results, and manually frame the parts of the clustering results belonging to the same cluster. The specific different clusters are as Figure 10 shown;

[0074] (2) In the clustering results, the specific categories of different clusters are not yet clear. According to the xline and inline positions of the well logging data, project the well logging data onto the post-stack seismic profile, and then combine the accurate seismic facies information and expert experience carried by the well logging interpretation data to expand the clustering results, such as Figure 11 . The specific situation can be divided into the following two types:

[0075] a. When the well logging is distributed in the clustering results (triangle label), it can be determined that all of this type of clustering results belong to the category described by the well logging, thereby realizing the expansion of labeled data;

[0076] b. When the well logging is distributed outside the background or clustering results (circle label), analyze the geological structure in combination with expert experience, and it can be judged that the blocks belonging to the same category near the well logging, and correct and expand the clustering results;

[0077] (3) Calculate the xline and inline number position ranges of the corresponding 3D pre-stack data of the expanded large batch of labeled data, and project them onto the 3D pre-stack seismic data volume to obtain the pre-stack multi-channel seismic data at the corresponding positions, such as Figure 12 ;

[0078] (4) Draw the labeled pre-stack multi-channel seismic data at each xline and inline position into n curves, and fill the positive value areas of the curves to generate pictures( Figure 13 ), and obtain a large batch of pre-stack labeled data;

[0079] The method of the present invention is applied to the actual pre-stack seismic data in a certain work area in Sichuan to verify the effectiveness of the method. The Inline range of this seismic volume data is 7600 - 7970, and the Xline range is 7700 - 8250, including three seismic facies: background, channel 1, and channel 2. After the above-mentioned labeled data processing steps, a total of 78,594 labeled images are finally obtained, including 33,494 background labels, 19,896 channel 1 labels, and 25,204 channel 2 labels. In this experiment, in addition to the improved LKA-ConvNext model proposed by the present invention, a conventional ConvNext model and a small-scale AlexNet model are also selected for comparative experiments to verify the effect of the improved model proposed by the present invention.

[0080] In this experiment, the common parameters of all network models are: the division ratio of the training set and the validation set of the network is 8:2, the input resolution is 128×128, the loss function is cross-entropy, the optimizer is adam, and the initial learning rate is 0.001.

[0081] The comparison of the training results and time consumption of different models is shown in Table 1. It can be seen that although the small-scale AlexNet model takes less training time, its accuracy is significantly lower than that of the ConvNext model. For the improved model proposed by the present invention, although there is a certain increase in training time, the accuracy has also been further improved compared with the traditional ConvNext model. Considering comprehensively, the effectiveness of the improved model proposed by the present invention can be seen.

[0082] Table 1 Comparison of the accuracy of the test set and training time consumption of different models

[0083]

[0084] Next, the model is applied to the actual unlabeled data to predict the complete seismic profile, and qualitative analysis is carried out according to the prediction results. The prediction results are as Figure 14 . It can be seen that in the application of unlabeled data, the model cannot maintain the high accuracy during training. For the area where channels overlap, the model is difficult to predict. The prediction results of the small-scale AlexNet model are not ideal and cannot accurately predict continuous channels. The improved model of the present invention and the traditional ConvNext model have better prediction effects in these areas, and the improved method of the present invention has a better prediction effect on channel 1 and is more superior in terms of result continuity. This also reflects the advantages of the method proposed by the present invention.

[0085] Those of ordinary skill in the art will realize that the embodiments described herein are provided to assist the reader in understanding the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A small-sample pre-stack seismic reflection pattern analysis method based on large-core attention, characterized in that Including: S1. Construct an image classification seismic reflection pattern analysis model based on the improved ConvNext module. The specific model in step S1 is as follows: Four cascaded improved ConvNext modules are used to expand the number of channels and extract deep features of pre-stack seismic images. A downsampling module is added between two adjacent improved ConvNext modules to change the size of pre-stack seismic images. Finally, through the global pooling layer, the length and width of the input pre-stack seismic images are converted to 1×1, and then through the fully connected layer, the number of channels is mapped to the number of categories to be classified, obtaining the probabilities that the input pre-stack seismic images are respectively judged as each category, and the maximum probability is the final classification result corresponding to the pre-stack seismic images. The improved ConvNext module is specifically based on the inverted bottleneck structure, adding LayerNorm regularization, GELU activation function, Layer Scale parameter scaling, and Drop Path layer. The inverted bottleneck structure is a cascaded structure of an Attention module and two Linear fully connected layers. The Attention module uses a large kernel attention layer to replace the 7×7 convolutional layer in ConvNext. S2. Use the augmented labeled pre-stack seismic data to train the model in step S1. S3. According to the model trained in step S2, predict the unlabeled pre-stack seismic images, and then splice the single-point prediction results at all positions into a complete seismic profile prediction result.

2. The small-sample pre-stack seismic reflection pattern analysis method based on large-core attention according to claim 1, characterized in that The large kernel attention layer is specifically the sum of three convolutions, which include: a depth convolution with a kernel size of k / d, a dilated convolution with a kernel size of 2d - 1 and a dilation rate of d, and a channel convolution with a kernel size of 1×1.

3. A small-sample pre-stack seismic reflection pattern analysis method based on large-core attention according to claim 2, characterized in that The augmentation process of the labeled pre-stack seismic data in step S2 is as follows: Obtain the overall distribution of different seismic facies on the seismic profile through the clustering algorithm, and then combine the accurate labels of well logging data for correction and augmentation. It includes the following steps: A1. Cluster the post-stack seismic profile, classify the seismic profile according to the clustering results, and frame the parts of the clustering results belonging to the same cluster. A2. In the clustering results, project the well logging data into the post-stack seismic profile according to the xline and inline positions of the well logging data, and then combine the accurate seismic facies information and expert experience carried by the well logging interpretation data to augment the clustering results. A3. Calculate the xline and inline number position ranges of the corresponding three-dimensional pre-stack data of the augmented labeled data, and project them onto the three-dimensional pre-stack seismic data volume to obtain the pre-stack multi-channel seismic data at the corresponding positions. A4. Draw the labeled pre-stack multi-channel seismic data at each xline and inline position as n curves, and fill the positive value regions of the curves to generate images, obtaining a large number of pre-stack labeled data.

4. A method for analyzing pre-stack seismic reflection patterns with few samples based on large-core attention according to claim 3, characterized in that The value of n is determined by the number of channels of the pre-stack seismic.

Images