A seismic data processing method

Abstract

The application discloses a seismic data processing method, and relates to the technical field of seismic data processing. The method comprises the following steps: acquiring a low-resolution seismic image; performing attention convolution operation on the low-resolution seismic image to perform weighting, and performing nonlinear transformation on the weighted key features to obtain initial features; sequentially performing twice attention convolution operation on the initial features to perform weighting; performing nonlinear transformation on each weighted initial feature and performing a cavity convolution operation to obtain intermediate features; performing an up-sampling operation on the intermediate features, and performing attention convolution operation on the up-sampled intermediate features to perform weighting; and performing nonlinear mapping on the weighted up-sampled intermediate features to obtain a high-resolution seismic image. The method can reduce the complexity to close to linear complexity while maintaining the performance of super-resolution reconstruction, and realizes the dual optimization of efficiency and precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic data processing, and particularly relates to a seismic data processing method. BACKGROUND

[0002] In energy exploration and earth science research, high-resolution processing of seismic data is one of the key technologies to improve the geological structure analysis capability. Seismic data acquisition, processing and interpretation are three interdependent and interdependent processes in seismic exploration. The quality of the acquired data, the advantages and disadvantages of the processing method and the correctness of the processing flow affect the results of seismic data processing, and high-quality data can improve the accuracy and reliability of subsequent seismic interpretation.

[0003] Preprocessing, processing analysis and processing are the three main steps of seismic data processing. The task of the preprocessing stage is to convert the data format, including: gain recovery and decoding, establishing trace header, extracting trace set, etc. The task of the processing analysis stage is to select the best parameters for the final processing through analysis experiments on a specific seismic horizon. The processing analysis stage includes four parts of static correction analysis, denoising analysis, deconvolution analysis and velocity analysis. The processing stage includes: using denoising, deconvolution, dynamic and static correction, velocity analysis, stacking migration, inversion and seismic monitoring processing methods to process low-resolution seismic data to obtain high-resolution seismic images.

[0004] The traditional method relies on large-scale matrix operations (such as iterative inversion or matrix inversion) when improving the resolution of seismic data, which has high complexity and low efficiency. SUMMARY

[0005] Therefore, it is necessary to provide a seismic data processing method aiming at the above technical problems.

[0006] The embodiment of the present application provides a seismic data processing method, which comprises:

[0007] obtaining a low-resolution seismic image;

[0008] performing attention convolution operation on the low-resolution seismic image to weight the key features of the low-resolution seismic image, and performing nonlinear transformation on the weighted key features to obtain initial features;

[0009] sequentially performing twice attention convolution operation on the initial features to weight the key features of the initial features; performing nonlinear transformation on each weighted initial feature, and performing hole convolution operation on the nonlinearly transformed features to insert intervals to obtain intermediate features;

[0010] The intermediate feature is up-sampled, and attention convolution is performed on the up-sampled intermediate feature to weight the key features of the up-sampled intermediate feature; and the weighted up-sampled intermediate feature is nonlinearly mapped to obtain a high-resolution seismic image.

[0011] Optionally, the high-resolution seismic image is obtained by improving the generator; the improved generator comprises: a first attention convolution layer, a Prelu function layer, two improved residual blocks, an up-sampling layer, a third attention convolution layer and a Tanh function layer connected in sequence; each of the two improved residual blocks comprises a second attention convolution layer, a Prelu function layer and a spatial convolution layer connected in sequence.

[0012] The low-resolution seismic image is subjected to attention convolution by the first attention convolution layer to weight the key features of the low-resolution seismic image; and the weighted key features are nonlinearly transformed by the Prelu function layer to obtain initial features.

[0013] The initial features are subjected to two times of attention convolution by the second attention convolution layers in the two improved residual blocks in sequence to weight the key features of the initial features; the weighted initial features are nonlinearly transformed by the Prelu function layers in the two improved residual blocks, and the nonlinearly transformed features are subjected to spatial convolution by the spatial convolution layers in the two improved residual blocks to insert intervals to obtain intermediate features.

[0014] The intermediate features are up-sampled by the up-sampling layer, and the up-sampled intermediate features are subjected to attention convolution by the third attention convolution layer to weight the key features of the up-sampled intermediate features; and the weighted up-sampled intermediate features are nonlinearly mapped by the Tanh function layer to obtain a high-resolution seismic image.

[0015] Optionally, the training method of the generator comprises:

[0016] An improved generative adversarial network is constructed, and the improved generative adversarial network comprises: an improved generator and an improved discriminator; the improved discriminator comprises: a plurality of discriminator blocks, a full connection layer and a Sigmoid function layer connected in sequence; each discriminator block comprises a spatial convolution layer, a Leaky Relu function layer and a two-dimensional dropout layer connected in sequence.

[0017] A real high-resolution seismic image corresponding to the low-resolution seismic image is obtained, and the high-resolution seismic image generated by the improved generator is taken as a reconstructed high-resolution seismic image.

[0018] For each discriminator block, the global features of the reconstructed high-resolution seismic image are extracted through the dilated convolution layer of the discriminator block; the global features of the reconstructed high-resolution seismic image are subjected to the Leaky Relu function operation through the Leaky Relu function layer of the discriminator block to obtain nonlinear global features; the nonlinear global features are randomly discarded through the two-dimensional dropout layer of the discriminator block to obtain the final global features;

[0019] The final global feature is mapped to a probability value through the fully connected layer, and the probability value is operated by the S-type function through the Sigmoid function layer to obtain the probability of the low-resolution seismic image recognition result;

[0020] The loss function between the reconstructed high-resolution seismic image and the real high-resolution seismic image is determined according to the probability of the low-resolution seismic image recognition results. The improved generative adversarial network is trained with the goal of minimizing the loss function to obtain a trained improved generative adversarial network; the trained improved generative adversarial network includes: a trained improved generator and a trained improved discriminator.

[0021] Optionally, obtaining a true high-resolution seismic image corresponding to the low-resolution seismic image specifically includes:

[0022] Obtain original seismic images;

[0023] Perform noise reduction and formatting on raw seismic data;

[0024] The processed original seismic image is extracted based on the following formula to obtain a low-resolution seismic image and a true high-resolution seismic image corresponding to the low-resolution seismic image:

[0025] ;

[0026] in, They are low-resolution seismic images and real high-resolution seismic images, is the inverse Fourier transform, are the low-frequency wavelet in the frequency domain and the high-frequency wavelet in the frequency domain, is the frequency, is the reflection coefficient spectrum.

[0027] Optionally, determining a loss function between the reconstructed high-resolution seismic image and the true high-resolution seismic image based on the probability of the low-resolution seismic image recognition result specifically includes:

[0028] The loss function of the improved generator is determined based on the following formula:

[0029] ;

[0030] ;

[0031] ;

[0032] in, represents the reconstruction loss, represents the perceived loss, represents the adversarial loss, α represents the weight of the perceptual loss, and β represents the weight of the adversarial loss. W and H represents the dimension of a low-resolution noisy seismic image, t It represents the resolution improvement from low-resolution noisy seismic images to real high-resolution seismic images, I LR represents a low-resolution noisy seismic image, I HR represents a real high-resolution seismic image, represents the reconstruction of high-resolution seismic images;

[0033] The loss function of the improved discriminator is determined based on the following formula:

[0034] ;

[0035] in, M represents the number of samples of real high-resolution seismic images, N Indicates the number of samples of low-resolution seismic images, log represents the probability of identifying a low-resolution seismic image as a reconstructed high-resolution seismic image, represents the probability of identifying a low-resolution seismic image as a true high-resolution image.

[0036] Optionally, it also includes introducing learning rate decay and momentum optimization methods when training the improved generative adversarial network to accelerate network convergence and shorten training time.

[0037] The seismic data processing method provided by the embodiment of the present invention has the following beneficial effects compared with the prior art:

[0038] The present invention adopts multi-stage attention convolution to dynamically weight channel features, replacing the explicit iterative inversion process that relies on large-scale matrix inversion in traditional methods; and introduces void convolution to insert intervals to expand the receptive field and enhance the feature representation of the initial features, avoiding the dense matrix multiplication operations performed in traditional methods to obtain global context information; while maintaining super-resolution reconstruction performance, it can reduce the complexity to near-linear complexity, achieving dual optimization of efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1A schematic diagram of a generator network structure of a seismic data processing method provided in one embodiment;

[0040] Figure 2 A schematic diagram of a discriminator network structure of a seismic data processing method provided in one embodiment;

[0041] Figure 3 A schematic diagram of a dilated convolution structure of a seismic data processing method provided in one embodiment;

[0042] Figure 4 A schematic diagram of the structure of an attention module of a seismic data processing method provided in one embodiment;

[0043] Figure 5 A data generation flow chart of a seismic data processing method provided in one embodiment;

[0044] Figure 6 A schematic diagram of raw seismic data of a seismic data processing method provided in one embodiment;

[0045] Figure 7 A schematic diagram of seismic data after Bicubic interpolation processing of a seismic data processing method provided in one embodiment;

[0046] Figure 8 A schematic diagram of seismic data after DCGAN processing according to a seismic data processing method provided in one embodiment;

[0047] Figure 9 A schematic diagram of seismic data processed by a seismic data processing method according to an embodiment of the present invention;

[0048] Figure 10 The figure is a flow chart of a seismic data processing method provided in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention 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 invention and are not intended to limit the present invention.

[0050] This paper aims to provide a generative adversarial network (GAN) based on dilated convolution and attention convolution to improve the resolution of seismic data. The network can be widely used in fields such as oil exploration and geological survey, providing technical support for improving the resolution and imaging quality of seismic data.

[0051] The embodiment of the present invention provides a method for processing seismic data. Figure 10 As shown, the method includes:

[0052] Acquire low-resolution seismic images.

[0053] An attention convolution operation is performed on the low-resolution seismic image to weight the key features of the low-resolution seismic image, and the weighted key features are nonlinearly transformed to obtain the initial features.

[0054] The initial features are sequentially subjected to two attention convolution operations to weight the key features of the initial features. Each weighted initial feature is nonlinearly transformed, and the features after nonlinear transformation are subjected to a dilated convolution operation to insert intervals to obtain intermediate features.

[0055] The intermediate features are upsampled and the attention convolution operation is performed on the upsampled intermediate features to weight the key features of the upsampled intermediate features. The weighted upsampled intermediate features are then nonlinearly mapped to obtain high-resolution seismic images.

[0056] 1. Network architecture design.

[0057] The improved generative adversarial network includes an improved generator and an improved discriminator, which are connected in sequence. The improved generator includes: a first attention convolution layer, a Prelu function layer, two improved residual blocks, an upsampling layer, a third attention convolution layer, and a Tanh function layer, which are connected in sequence. Both improved residual blocks include a second attention convolution layer, a Prelu function layer, and a spatial convolution layer, which are connected in sequence. The improved discriminator includes: multiple discriminator blocks, a fully connected layer, and a Sigmoid function layer, which are connected in sequence. Each discriminator block includes a dilated convolution layer, a leaky Relu function layer, and a two-dimensional dropout layer, which are connected in sequence.

[0058] Among them, the improved generator (structure as Figure 1 (as shown in Figure 2): A deep neural network consisting of dilated and attention convolutional layers is used. Specifically, the convolutional layers of the residual block are replaced with attention convolutional layers, and a dilated convolutional layer is added after the attention convolutional layer to obtain an improved residual block. The improved generator is obtained by replacing the convolutional layers of the generator of the generative adversarial network with attention convolutional layers and replacing the residual blocks of the generator network with improved residual blocks.

[0059] Dilated Convolution increases the receptive field by inserting "holes" (intervals) in the convolution kernel, thereby extracting richer contextual information without increasing the number of parameters.

[0060] Attention Convolution weights seismic data through the self-attention mechanism, enhancing the network's ability to capture important structures.

[0061] Among them, the improved discriminator (structure as Figure 2 As shown in Figure 2, this paper adopts a deep neural network composed of multi-scale dilated convolutions and residual blocks. The specific construction method is to replace the convolutional layers of the discriminator of the generative adversarial network with dilated convolutional layers to obtain an improved discriminator.

[0062] 1.1. Generator and Discriminator.

[0063] The generator is responsible for converting low-resolution seismic images into high-resolution seismic images, while the discriminator is used to distinguish between the images generated by the generator and the real high-resolution seismic images.

[0064] The core of the generator lies in its ability to capture and enhance key features in low-resolution seismic images while maintaining their naturalness and realism. The generator's structure primarily consists of multiple convolutional blocks, including attention convolution and dilated convolution. The attention convolution incorporates a channel-wise attention mechanism that adjusts the feature responses of different channels. The attention convolution uses adaptive average pooling and convolution to generate weight coefficients, thereby reweighting the convolution output. This mechanism highlights important features, suppresses redundant information, improves the quality of generated feature maps, and enhances the representation of key information. The dilated convolution applies a dilation factor, allowing the network to expand its receptive field without increasing the number of parameters. This effectively improves the ability to perceive underlying patterns in seismic data, thereby enhancing the detail and clarity of the generated images.

[0065] In addition to multiple convolutional layers, the generator employs an upsampling layer to increase the spatial size of the feature map, restoring the resolution of the original image and ensuring that the output image is suitable for seismic image analysis. The activation function uses Prelu, which provides a nonlinear transformation to the network, enhancing its expressive power and normalizing the output, contributing to the stability of the optimization process. Finally, the generator generates the final high-resolution image through another attention convolutional layer and a Tanh activation function, ensuring that the output values ​​remain within a reasonable range.

[0066] The design goal of the discriminator is to accurately distinguish between the high-resolution images generated by the generator and the real high-resolution images. In this paper, key technologies such as the empty convolution layer, batch normalization and Dropout, fully connected layer, Sigmoid, Leaky Relu activation function, etc. are used to build the discriminator. By using the empty convolution, the discriminator can expand the receptive field without increasing the complexity, effectively capturing the global features of the image. In order to improve the generalization ability and training stability of the discriminator, batch normalization and Dropout layers are added after each convolution block. The Leaky Relu activation function in the discriminator is used to prevent the dead zone phenomenon of neurons, maintaining a small slope in the negative value interval, thereby improving the training performance of the network. After feature extraction, the discriminator maps the learned features to a probability value through a fully connected layer, representing the probability that the image is real data. The output of the discriminator is passed through the Sigmoid activation function to ensure that the output value is a probability value between 0 and 1.

[0067] Through the carefully designed generator and discriminator, the generative adversarial network model can effectively improve the resolution of seismic data images while maintaining the quality and accuracy of the images.

[0068] 1.2, Dilation Convolution Module and Attention Convolution Module.

[0069] Dilated Convolution is an extension of traditional convolution, which expands the receptive field of the convolution kernel by inserting "holes" (i.e. intervals) between the elements of the convolution kernel without increasing the computational complexity. The schematic diagram of dilated convolution is shown in Figure 3 This method can capture more extensive context information while maintaining the resolution of the feature map, and is particularly suitable for image semantic segmentation, time series modeling, target detection, and seismic data processing under complex geological conditions. The attention convolution module refers to the introduction of attention mechanism in the convolution operation, which enables the network to adaptively focus on different regions or channels in the input feature map. Attention convolution combines the natural and efficient feature extraction capability of convolutional neural networks with the dynamic weighting features of attention mechanism, enhancing the model's expression ability compared to directly using attention mechanism. This combination can achieve good results in seismic data super-resolution reconstruction, as it can capture both global features and local details.

[0070] The attention module in attention convolution is shown in Figure 4 First, the feature is extracted through the main convolution layer, then the weight of each channel is calculated through the attention module, and finally the calculated weight is multiplied element by element with the output of the convolution to realize the re-correction of the channel.

[0071] The attention convolution module is mainly composed of two parts: a standard convolution module and an attention mechanism. First, the standard convolution module extracts local features from the input feature map through convolution operations to capture spatial information in the image. Convolution operations can not only extract important local structural features but also help reduce feature dimensions, thereby improving efficiency. Then the channel attention mechanism introduced in the module describes the features of each channel globally through adaptive average pooling and generates corresponding channel weights through a series of convolution and activation functions. After this series of operations, a weight vector between 0 and 1 is output (one weight for each channel), which represents the importance of different channels. This weight vector is then used to multiply the original convolution output feature map element-wise, thereby achieving channel attention weighting of the original feature map. These weights are used to dynamically adjust the convolution output to focus more on important features, thereby enhancing the network's attention to key information.

[0072] 2. Loss function design

[0073] Adversarial Loss: The least squares loss function is used to measure the similarity between generated samples and real samples.

[0074] Perceptual Loss: Based on a pre-trained seismic data feature extraction network, the difference between generated samples and real samples in the feature space is calculated.

[0075] Reconstruction Loss: The reconstruction ability of the generator is measured by the mean square error between the low-resolution input and the high-resolution output.

[0076] The loss function of the generator is defined as follows:

[0077] ;

[0078] where, represents the reconstruction loss, represents the perceptual loss, represents the adversarial loss, and α and β represent the weights of the perceptual loss and the adversarial loss. Pixel-wise calculation of the reconstruction high-resolution seismic image and the real high-resolution seismic image I HR MSE.

[0079] Mean Squared Error (MSE) is a commonly used loss function in image super-resolution, which helps to obtain a high Peak signal-to-noise ratio (PSNR).

[0080] ;

[0081] wherein, W and H denotes the low-resolution noisy seismic image I LR dimension of, t denotes the resolution improvement from I LR to I HR .

[0082] is defined as the probability of the reconstructed high-resolution seismic image being identified by the discriminator as a real high-resolution seismic image:

[0083] ;

[0084] wherein, N denotes the sample number of low-resolution noisy seismic images I LR , denotes the probability of the reconstructed high-resolution seismic image being considered by the discriminator as a real high-resolution noisy seismic image.

[0085] To obtain a better gradient behavior, we use instead of .

[0086] The goal of the discriminator is to maximize the probability of correctly identifying real high-resolution images I HR and to minimize as much as possible the probability of correctly identifying reconstructed high-resolution seismic images . The loss function of the discriminator is defined as follows:

[0087] ;

[0088] wherein, M denotes the sample number of real high-resolution seismic images, N denotes the sample number of low-resolution seismic images, log denotes the probability of a low-resolution seismic image being identified as a reconstructed high-resolution seismic image, denotes the probability of a low-resolution seismic image being identified as a real high-resolution image.

[0089] 3. Training strategy

[0090] The generator and the discriminator are updated respectively using an alternating optimization strategy to ensure that both reach a balanced state during network training. The improved generator generates reconstructed high-resolution seismic images, while the improved discriminator distinguishes between real data samples (real high-resolution seismic images) and fake data samples (reconstructed high-resolution seismic images). The training process alternately optimizes the improved generator and the improved discriminator through mutual confrontation. Learning rate decay and momentum optimization methods (such as the Adam optimizer) are introduced to accelerate network convergence.

[0091] The implementation process includes:

[0092] 3.1, data preprocessing.

[0093] Obtain the original seismic image, denoise and format the original seismic image, extract the seismic data pairs for training, and construct the training sample set.

[0094] The overall flow of generating data is shown in Figure 5 The original seismic image is input into the reflectivity coefficient model, and the low-frequency wavelet and the high-frequency wavelet in the frequency domain are obtained through Gaussian deformation. The high-frequency wavelet in the frequency domain is processed through the high-resolution seismic data convolution formula to obtain the real high-resolution seismic image (corresponding to the high-resolution data in the figure). The low-frequency wavelet in the frequency domain is processed through the low-resolution seismic data convolution formula, and noise and downsampling preprocessing are added to obtain the low-resolution seismic image (corresponding to the low-resolution data in the figure).

[0095] The high and low resolution seismic data convolution formulas are as follows:

[0096] ;

[0097] wherein, are the low-resolution seismic image and the real high-resolution seismic image, is the inverse Fourier transform, are the low-frequency wavelet in the frequency domain and the high-frequency wavelet in the frequency domain, is the frequency, is the reflectivity spectrum.

[0098] 3.2, model training.

[0099] An improved generative adversarial network is constructed, which includes an improved generator and an improved discriminator. End-to-end joint training is performed based on the designed loss function until the model converges.

[0100] Obtain the real high-resolution seismic image corresponding to the low-resolution seismic image, and take the high-resolution seismic image obtained by the improved generative adversarial network as the reconstructed high-resolution seismic image.

[0101] The high-resolution seismic image generated by the improved generator is used as the reconstructed high-resolution seismic image. Specifically, the process includes: performing an attention convolution operation on the low-resolution seismic image through a first attention convolution layer to weight the key features of the low-resolution seismic image. The weighted key features are then nonlinearly transformed through a Prelu function layer to obtain initial features. The initial features are then sequentially subjected to two attention convolution operations through the second attention convolution layers in two improved residual blocks to weight the key features of the initial features. Each weighted initial feature is then nonlinearly transformed through a Prelu function layer in the two improved residual blocks. Dilated convolution operations are then performed on the nonlinearly transformed features through spatial convolution layers in the two improved residual blocks to insert gaps, thereby obtaining intermediate features. The intermediate features are then upsampled through an upsampling layer, and the upsampled intermediate features are then upsampled through a third attention convolution layer to weight the key features of the upsampled intermediate features. The weighted upsampled intermediate features are then nonlinearly mapped through a Tanh function layer to obtain a high-resolution seismic image.

[0102] For each discriminator block, the dilated convolutional layer of the discriminator block extracts the global features of the reconstructed high-resolution seismic image. The Leaky Relu function layer of the discriminator block performs a Leaky Relu function operation on the global features of the reconstructed high-resolution seismic image to obtain nonlinear global features. The two-dimensional dropout layer of the discriminator block randomly discards the nonlinear global features to obtain the final global features.

[0103] The final global feature is mapped to a probability value through the fully connected layer, and the probability value is operated by the S-type function through the Sigmoid function layer to obtain the probability of the low-resolution seismic image recognition result;

[0104] The loss function between the reconstructed high-resolution seismic image and the true high-resolution seismic image is determined based on the probability of the low-resolution seismic image recognition results. With the goal of minimizing the loss function, the improved generative adversarial network is trained to obtain a trained improved generative adversarial network. The trained improved generative adversarial network includes a trained improved generator and a trained improved discriminator.

[0105] 3.3. Model reasoning

[0106] Use the trained generator to perform high-resolution reconstruction of new low-resolution seismic data.

[0107] After entering the improved generator, the low-resolution seismic image first undergoes an attention convolution layer and a Prelu (Parametric Rectified Linear Unit) function for shallow feature extraction. It then passes through two residual blocks containing attention convolution and dilated convolution to further refine and enhance the feature representation. A batch normalization layer is also used in the residual blocks to prevent overfitting, and the Prelu activation function is used to enhance the network's expressiveness and normalize the output. An upsampling layer restores the seismic data, amplifies the low-resolution feature map, and then passes through another attention convolution layer and a Tanh activation function to generate the final high-resolution image. The Tanh function (hyperbolic tangent function) is a nonlinear activation function with an output range of [-1, 1]. Using the Tanh function for nonlinear mapping can make the generator's output more stable. The Prelu and Tanh functions are implemented as independent layers.

[0108] The discriminator of this network is designed to more accurately distinguish between high-resolution images generated by the generator and real high-resolution images. Here, dilated convolution, leaky ReLU function, two-dimensional dropout (Dropout), fully connected layers, and sigmoid function are used to build the discriminator.

[0109] After the data enters the improved discriminator, dilated convolutions efficiently capture the global features of the image. LeakyRelu activation functions are then used to increase nonlinearity, and dropout layers are used to further reduce the risk of overfitting. The discriminator blocks then continue to downsample and further process the feature maps. These subsequent discriminator blocks help the network extract higher-level features and better focus on the complex structures in the seismic data. To improve the generalization and training stability of the improved discriminator, batch normalization and dropout layers are added to each discriminator block. Finally, the features extracted by the convolutional layer are flattened and fed into a fully connected layer, where a sigmoid function is applied to determine the authenticity of the seismic data. After feature extraction, a fully connected layer maps the learned features to a probability value, representing the probability that the image is real data. The output of the improved discriminator passes through a sigmoid activation function to ensure that the output value is between 0 and 1, representing the probability.

[0110] For example, a comparison between DCGAN (Deep Convolutional Generative Adversarial Networks) and the improved generative adversarial network provided by the present invention is used to illustrate seismic data processing. Figure 6 This is a schematic diagram of the original seismic data. The original seismic data image without any processing shows great fuzziness and noise interference. Figure 7This figure shows seismic data after Bicubic interpolation. Interpolation, as a traditional image processing method, can effectively fill in missing data points. Although interpolation can smooth images, it has limitations in recovering detail and removing noise. It cannot recover the complex structure in seismic data, cannot remove noise, and may even smooth out some details. Figure 8 This is a diagram of seismic data processed by DCGAN. Compared to interpolation, DCGAN can learn the underlying distribution of seismic data through generative adversarial network training, effectively restoring some details and improving image resolution. However, due to the limitations of DCGAN's generative capabilities, the processed seismic data is slightly lacking in detail clarity and noise suppression. Figure 9 This is a schematic diagram of seismic data processed by the present invention. The seismic data image processed by the present invention not only has significantly improved image resolution and clarity, but also more clearly displays seismic waveform details and performs better in noise suppression.

[0111] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A seismic data processing method, characterized in that: include: Acquire low-resolution seismic images; Perform an attention convolution operation on the low-resolution seismic image to weight the key features of the low-resolution seismic image, and perform a nonlinear transformation on the weighted key features to obtain the initial features; Perform two attention convolution operations on the initial features in sequence to weight the key features of the initial features; perform nonlinear transformation on each weighted initial feature, and perform a dilated convolution operation on the features after nonlinear transformation to insert intervals to obtain intermediate features; Perform an upsampling operation on the intermediate features and perform an attention convolution operation on the upsampled intermediate features to weight the key features of the upsampled intermediate features; Perform nonlinear mapping on the weighted upsampled intermediate features to obtain high-resolution seismic images; Wherein, high-resolution seismic images are obtained by an improved generator; the improved generator includes: a first attention convolution layer, a Prelu function layer, two improved residual blocks, an upsampling layer, a third attention convolution layer and a Tanh function layer connected in sequence; the two improved residual blocks each include a second attention convolution layer, a Prelu function layer and a spatial convolution layer connected in sequence; The low-resolution seismic image is subjected to an attention convolution operation through the first attention convolution layer to weight the key features of the low-resolution seismic image; and the weighted key features are subjected to a nonlinear transformation through the Prelu function layer to obtain the initial features; The initial features are sequentially subjected to two attention convolution operations through the second attention convolution layer in the two improved residual blocks to weight the key features of the initial features; each weighted initial feature is nonlinearly transformed through the Prelu function layer in the two improved residual blocks, and the features after nonlinear transformation are subjected to a dilated convolution operation through the spatial convolution layer in the two improved residual blocks to insert intervals and obtain intermediate features; The upsampling layer performs an upsampling operation on the intermediate features, and the third attention convolution layer performs an attention convolution operation on the upsampled intermediate features to weight the key features of the upsampled intermediate features; the Tanh function layer performs nonlinear mapping on the weighted upsampled intermediate features to obtain a high-resolution seismic image; The training method of the generator includes: Constructing an improved generative adversarial network, the improved generative adversarial network comprising: the improved generator and an improved discriminator; the improved discriminator comprising: a plurality of discriminator blocks, a fully connected layer, and a Sigmoid function layer connected in sequence; each discriminator block comprising a hole convolution layer, a leaky ReLU function layer, and a two-dimensional dropout layer connected in sequence; Obtain a real high-resolution seismic image corresponding to the low-resolution seismic image, and use the high-resolution seismic image generated by the improved generator as the reconstructed high-resolution seismic image; For each discriminator block, the global features of the reconstructed high-resolution seismic image are extracted through the dilated convolution layer of the discriminator block; the global features of the reconstructed high-resolution seismic image are subjected to the Leaky Relu function operation through the Leaky Relu function layer of the discriminator block to obtain nonlinear global features; the nonlinear global features are randomly discarded through the two-dimensional dropout layer of the discriminator block to obtain the final global features; The final global feature is mapped to a probability value through the fully connected layer, and the probability value is operated by the S-type function through the Sigmoid function layer to obtain the probability of the low-resolution seismic image recognition result; The loss function between the reconstructed high-resolution seismic image and the true high-resolution seismic image is determined according to the probability of the low-resolution seismic image recognition result. The improved generative adversarial network is trained with the goal of minimizing the loss function to obtain a trained improved generative adversarial network; the trained improved generative adversarial network includes: a trained improved generator and a trained improved discriminator.

2. A seismic data processing method according to claim 1, characterized in that: The step of obtaining a true high-resolution seismic image corresponding to the low-resolution seismic image specifically includes: Obtain original seismic images; Perform noise reduction and formatting on raw seismic data; The processed original seismic image is extracted based on the following formula to obtain a low-resolution seismic image and a true high-resolution seismic image corresponding to the low-resolution seismic image: ； in: They are low-resolution seismic images and real high-resolution seismic images, is the inverse Fourier transform, are the low-frequency wavelet and high-frequency wavelet in the frequency domain, is the frequency, is the reflection coefficient spectrum.

3. A seismic data processing method according to claim 2, characterized in that: The loss function between the reconstructed high-resolution seismic image and the true high-resolution seismic image is determined based on the probability of the low-resolution seismic image recognition result, specifically including: The loss function of the improved generator is determined based on the following formula: ； ； ； in, represents the reconstruction loss, represents the perceived loss, represents the adversarial loss, α represents the weight of the perceptual loss, and β represents the weight of the adversarial loss. W and H represents the dimension of a low-resolution noisy seismic image, t It represents the resolution improvement from low-resolution noisy seismic images to real high-resolution seismic images, I LR represents a low-resolution noisy seismic image, I HR represents a real high-resolution seismic image, represents the reconstruction of high-resolution seismic images; The loss function of the improved discriminator is determined based on the following formula: ； in, M represents the number of samples of real high-resolution seismic images, N Indicates the number of samples of low-resolution seismic images, log represents the probability of identifying a low-resolution seismic image as a reconstructed high-resolution seismic image, represents the probability of identifying a low-resolution seismic image as a true high-resolution image.

4. A seismic data processing method according to claim 1, characterized in that: It also includes introducing learning rate decay and momentum optimization methods when training the improved generative adversarial network to accelerate network convergence and shorten training time.

Images