Post-stack seismic profile super-resolution reconstruction method

Abstract

The invention relates to a post-stack seismic section super-resolution reconstruction method comprising the steps that a generative network and an adversarial network are constructed, the generative network is used for improving the resolution of a post-stack seismic section, and the adversarial network is used for judging whether an input image is an original high-resolution seismic section or a high-resolution seismic section reconstructed through the generative network. A generative network composed of a multi-scale residual module, a double-attention module, a residual D module and the like is designed, and the ability of the network in the aspect of extracting important feature information is improved; a Markov discriminator-based adversarial network is also designed, so that the discrimination capability of the network on seismic images is improved; a conventional loss function of the generative adversarial network is improved by introducing total variation loss and edge information retention loss functions. According to the method, the resolution of the post-stack seismic section can be effectively improved, meanwhile, noise signals are better removed, more geological structures and features are reserved, and the sensing quality of the seismic image is improved.

Description

Technical Field

[0001] The present invention belongs to the field of seismic exploration and artificial intelligence technology, and specifically relates to a post-stack seismic profile super-resolution reconstruction method based on a multi-scale edge information recovery generative adversarial network. Background Art

[0002] Oil and gas resources occupy a central position in the global energy system and have a decisive impact on key economic infrastructure such as industrial production, transportation and energy demand. The strong regional distribution and the links involved from exploration, mining, processing to terminal consumption have a significant impact on the international political and economic landscape. A stable supply of oil and gas resources is a basic condition for maintaining national security and sustained economic growth. Faced with the increasing difficulty of oil and gas resource development and the rising demand for exploration of deep, marginal and complex geological structures of oil and gas reservoirs, technological innovation to improve development efficiency and economy has become the main challenge facing the energy industry. Therefore, the use of scientific and efficient strategies for the development and utilization of oil and gas resources is not only related to the development of the national economy, but also a key factor in achieving global energy security and sustainable development goals.

[0003] Seismic exploration is one of the most critical technical means in the exploration and development of oil and gas resources. With the gradual reduction of easily explorable resources around the world, the deepening of exploration targets and the exploration of unconventional oil and gas resources, seismic exploration faces many challenges. Among them, the high cost of data acquisition, the quality of seismic data, and the difficulty of imaging and interpretation are the main problems that need to be solved in current seismic exploration. Therefore, exploring intelligent seismic data processing methods to improve the signal-to-noise ratio and resolution of seismic data, reduce exploration costs, and improve interpretation accuracy has become an important task facing the field of oil and gas exploration.

[0004] Post-stack seismic profile super-resolution reconstruction technology is an effective way to improve the signal-to-noise ratio and resolution of seismic data, thereby making the imaging of underground structures clearer and more accurate. Existing post-stack seismic profile super-resolution reconstruction methods are mainly divided into four methods: deconvolution-based methods, absorption compensation-based methods, spectrum broadening-based methods, and deep learning-based methods.

[0005] Among them, deconvolution technology aims to extract or restore the original signal from the output signal, and improve the resolution by compressing the seismic waveform and expanding its spectral range. Billings et al. proposed a method for frequency domain deconvolution of airborne gamma-ray survey data using Wiener filters, and reconstructed the signal through slight manual adjustment of the geographic detector model and noise level. Zhang et al. proposed a seismic signal deconvolution algorithm, which effectively improved the reflection coefficient accuracy of low signal-to-noise ratio seismic signals through multi-resolution analysis and binary wavelet transform. Although the deconvolution-based method can improve the spatial resolution of seismic data, it will also enhance the background noise. Especially in seismic data with low signal-to-noise ratio, this noise amplification phenomenon is more obvious, affecting the clarity of geological structure and the accuracy of interpretation.

[0006] The absorption compensation method is to recover and enhance the high-frequency components of seismic signals by compensating for the energy loss and frequency attenuation during the propagation of seismic waves, thereby improving the spatial resolution of seismic profiles. Braga et al. proposed an inverse Q filtering mathematical model based on continuous wavelet transform, which effectively corrected the amplitude and phase distortion caused by propagation in attenuated and dispersive media and improved the resolution of seismic profiles. The absorption compensation method relies on an accurate medium absorption model and the processing process is relatively complicated.

[0007] The method based on the spread spectrum takes time-frequency decomposition as the core, processes high-frequency and low-frequency components in the time-frequency domain, realizes wavelet compression and supports broadband. This method effectively expands the frequency band range of seismic data by adjusting the time-frequency characteristics, thereby improving the resolution of seismic profiles. Alaei et al. effectively improved the vertical resolution of seismic images of thin-layer oil reservoirs by combining Gabor deconvolution and wavelet scaling technology, and achieved clear imaging of thin-layer structures. When the method based on the spread spectrum processes and expands the frequency band range of seismic data through time-frequency decomposition, especially in the process of enhancing high-frequency components to improve the resolution of seismic profiles, it may over-amplify the noise or non-geological signals in the data, resulting in artifacts in the reconstructed seismic profiles.

[0008] The post-stack seismic profile super-resolution reconstruction method based on deep learning can learn the mapping relationship between complex stratigraphic features and super-resolution reconstruction of seismic profiles, thereby providing more accurate and clear reconstruction results. Li et al. used a deep convolutional neural network to simultaneously achieve super-resolution reconstruction and denoising of seismic images. This method significantly improves the perceptual quality of seismic images by combining L1 loss and multi-scale structural similarity loss as the training loss function. Zeng et al. proposed a deep feature mining network to achieve super-resolution reconstruction of seismic data. This method effectively restores the detailed features of seismic images by integrating multi-scale feature extraction and deep feature mining technology.

[0009] Although the deep learning-based seismic profile super-resolution reconstruction method has better super-resolution reconstruction effect than the other three methods, the reconstructed seismic profile still has the problem of missing edge texture details and difficulty in recovering weak amplitude signals. Based on this, it is urgent to develop an innovative technical solution to effectively solve the problem of difficulty in recovering weak amplitude signals caused by low resolution and noise interference, the presence of noise signals and unclear stratigraphic structures such as faults. Summary of the invention

[0010] The purpose of the present invention is to provide a post-stack seismic profile super-resolution reconstruction method based on a multi-scale edge information recovery generative adversarial network, so as to solve the problems that the weak amplitude signal appearing in the high-resolution seismic profile obtained after reconstruction by the prior art is difficult to recover, and there are noise signals and unclear stratigraphic structures such as faults.

[0011] The objective of the present invention is achieved through the following technical solutions:

[0012] A post-stack seismic profile super-resolution reconstruction method comprises the following steps:

[0013] S1. Construct the data set required for training, verification and testing of post-stack seismic profile super-resolution reconstruction method;

[0014] S2, constructing a generative network in a multi-scale edge information recovery generative adversarial network to improve the resolution of post-stack seismic profiles while removing noise signals;

[0015] S3, constructing an adversarial network in a multi-scale edge information recovery generative adversarial network to determine whether the input image is a high-resolution post-stack seismic profile in the data set or a high-resolution post-stack seismic profile reconstructed by the generative network;

[0016] S4. Construct a loss function to measure the post-stack seismic profile super-resolution reconstruction performance of the constructed multi-scale edge information recovery generative adversarial network;

[0017] S5, network model training: training the multi-scale edge information recovery generative adversarial network composed of step S2 and step S3, through continuous cyclic adversarial training between the generative network and the adversarial network, according to the verification result of the generative network output, optimizing the loss function constructed in step S4, and finally obtaining the optimal model parameters of the generative network and the adversarial network, and the network model training is completed;

[0018] S6, network model testing: testing the multi-scale edge information restoration generative adversarial network model obtained after training in step S5, if the evaluation index of the test result meets the set index threshold, the multi-scale edge information restoration generative adversarial network model obtained after training in step S5 is used as the optimal super-resolution reconstruction model; otherwise, returning to step S5, and continuing to train the model by adjusting the network training parameters;

[0019] S7. Application of network model: Input the real low-resolution noisy post-stack seismic profile into the optimal super-resolution reconstruction model to obtain the reconstructed high-resolution post-stack seismic profile.

[0020] Furthermore, step S1 specifically includes the following steps:

[0021] S11, defining seismic data parameters to create a one-dimensional subsurface reflection model;

[0022] S12, by adopting vertical shear and folding operations, the complexity of geological structure is simulated, and then the formation tilt is introduced through plane shear operations to reflect the heterogeneity of the formation;

[0023] S13, by adding faults, further simulating the fracture and displacement phenomena of underground strata;

[0024] S14, generating synthetic three-dimensional low-resolution and high-resolution seismic profiles by convolving the model with low-frequency and high-frequency Ricker wavelets, respectively;

[0025] S15, adding colored noise to 3D low-resolution seismic profiles to increase the real complexity of the data;

[0026] S16. A center cropping method is used to obtain two-dimensional seismic low-resolution and high-resolution sections from three-dimensional seismic sections, and a two-fold downsampling operation is performed on the obtained two-dimensional low-resolution seismic sections to complete the entire process of generating seismic section data.

[0027] Further, step S2 is specifically as follows: the generative network for super-resolution reconstruction of post-stack seismic profiles is mainly composed of an encoder and a decoder. The encoder reduces the size of the feature map by continuous downsampling operations; the decoder reconstructs high-quality high-resolution seismic profiles through upsampling operations; the encoder consists of a 3×3 convolutional layer with a LeakyReLU activation function, two multi-scale residual modules, two dual attention modules, two residual D modules and three cascaded dilated convolutional layers; the low-resolution seismic profiles in the training set are first input into a 3×3 convolutional layer with a LeakyReLU activation function; then, the feature map output by the convolutional layer is sequentially input into the multi-scale residual module and the dual attention module; next, the output feature map is input into the residual D module for downsampling operations; the multi-scale residual module, dual attention module, and residual D module of the next scale have the same principle and processing order as before; finally, the output feature map is input into three cascaded dilated convolutions; the decoder The encoder is mainly composed of two upsampling layers, three dual attention modules, three multi-scale residual modules, a double pixel reorganization layer, two 1×1 convolution layers, and a 3×3 convolution with Tanh activation function; the feature map output by the encoder is first input into the first upsampling layer in the decoder; then, the output feature map is sequentially input into the dual attention module and the multi-scale residual module; next, the feature map output by the multi-scale residual module and the feature map output by the first 3×3 convolution in the encoder are subjected to channel concatenation operation for feature fusion; after that, the 1×1 convolution layer is used to adjust the number of channels of the feature map; the next scale upsampling layer, dual attention module, multi-scale residual module, channel concatenation operation and 1×1 convolution layer have the same principle and processing order as before; then, the dual attention module and the multi-scale residual module are used for further feature extraction and integration; finally, a high-quality high-resolution post-stack seismic profile is output through a double pixel reorganization operation and a 3×3 convolution layer with Tanh activation function;

[0028] Furthermore, the multi-scale residual module consists of two branches, and the input feature maps of the two branches are the same; the first branch consists of a jump connection with a 1×1 convolutional layer; the second branch consists of three parallel dilated convolution blocks, an independent LeakyReLU activation function and a 3×3 convolutional layer with a LeakyReLU activation function; wherein the dilated convolution block consists of three cascaded dilated convolution layers with the same dilation rate; the dilation rates of the three dilated convolution blocks are 1, 2, and 3, respectively; the output feature maps of the three parallel dilated convolution blocks are fused through an element-wise addition operation; then, the fused output feature map is input into an independent LeakyReLU activation function and a 3×3 convolutional layer with a LeakyReLU activation function to obtain the output feature map of the second branch; finally, the output feature maps of the two branches are obtained through an element-wise addition operation to obtain the output feature map of the multi-scale residual module.

[0029] Furthermore, the dual attention module is composed of a cascade of two modules, a channel attention module and a spatial attention module; the channel attention module is composed of two branches, and the input feature maps of the two branches are the same; the first branch is a skip connection to prevent information loss; the second branch is composed of a cascade of a global average pooling layer, a 1×1 convolution layer and a Sigmoid activation function; the global average pooling adjusts the size of the feature map to 1×1; the 1×1 convolution layer is used to extract feature information; the Sigmoid activation function is used to obtain the channel attention weight; the feature map output by the first branch and the channel attention weight output by the second branch are obtained by element-wise multiplication to obtain a channel attention feature map; the channel attention feature map is the input of the spatial attention module. The spatial attention module consists of two branches; the first branch is a jump connection used to prevent feature information loss; the second branch consists of three small branches, a 1×1 convolution layer, a 3×3 convolution layer and a Sigmoid activation function; among them, the first small branch is a channel global average pooling layer for aggregating global geological structure information; the second small branch is a channel global maximum pooling layer for aggregating detailed geological structure information; the third small branch is composed of a downsampling layer, three cascaded 3×3 convolution layers and an upsampling layer cascade; the output feature maps of the three small branches are feature fused through channel splicing operations; the fused output feature map is adjusted through a 1×1 convolution layer and a 3×3 convolution layer to adjust the number of channels and extract deep feature information; then the output feature map after feature extraction is activated by the Sigmoid function to obtain the spatial attention weight; the output feature map of the first branch and the spatial attention weight output of the second branch obtain the output feature map of the dual attention module.

[0030] Furthermore, the residual D module consists of two branches, and the input feature maps of the two branches are the same; the first branch consists of a 1×1 convolution layer, a 3×3 convolution layer with a step size of 2, and a 1×1 convolution layer cascaded; the feature map size output by the first branch is half the size of the input feature map; the second branch consists of an average pooling layer and a 1×1 convolution layer cascaded; the feature map size output by the second branch is half the size of the input feature map; the output feature maps of the first branch and the second branch are obtained by element-wise addition operation to obtain the output feature map of the residual D module.

[0031] Furthermore, in step S3, the adversarial network is a network based on a Markov discriminator, consisting of three 3×3 convolutions with a stride of 2 and a LeakyReLU activation function and an independent 3×3 convolution with a stride of 2; the output feature map size of the three 3×3 convolutions with a stride of 2 is reduced to one eighth of the input post-stack seismic profile size; the output feature map size of the last independent 3×3 convolution with a stride of 2 is reduced to one sixteenth of the input post-stack seismic profile size, and the number of channels of the output feature map is adjusted to 1; the output feature map of the last independent 3×3 convolution with a stride of 2 is the discriminant score matrix of whether the input image of the adversarial network is a high-resolution post-stack seismic profile in the data set or a high-resolution post-stack seismic profile obtained after the generation network reconstruction.

[0032] Furthermore, in step S4, the loss function is a loss function that introduces a total variation loss function and an edge information preservation loss function on the conventional loss function of the generative adversarial network, and its mathematical expression is:

[0033] Loss = 0.01L adv +2×10 -8 L tv +L mse +L edge

[0034] Among them, L adv represents the adversarial loss function, L tv represents the total variation loss function, L mse represents the mean square error loss function, L edge Represents the edge information preservation loss function;

[0035] Adversarial loss function L adv The mathematical expression is:

[0036] L adv =E[logD(y)]+E{log[1-D(G(x))]}

[0037] Where x is the input low-resolution post-stack seismic profile, y is the high-resolution post-stack seismic profile in the training set, G() is the generative network, D() is the adversarial network, G(x) is the reconstructed high-resolution post-stack seismic profile, and E represents the averaging operation;

[0038] Where W and H are the dimensions of the post-stack section, It is a pre-trained VGG network;

[0039] The total variational loss L tv The mathematical expression is:

[0040]

[0041] Among them, β is set to 2;

[0042] Mean square error loss L mse The mathematical expression is:

[0043]

[0044] Edge-preserving loss function L edge The mathematical expression is:

[0045]

[0046] Among them, Sobel() represents the edge detection operation;

[0047] Furthermore, in step S6, the network model test specifically includes the following steps:

[0048] S61. Input the paired low-resolution and high-resolution post-stack seismic profiles in the test set into the trained multi-scale edge information recovery generative adversarial network for testing. When the evaluation index of the test result is:

[0049] S62: When the peak signal-to-noise ratio and the structural similarity do not meet the set index threshold, the process returns to step S5, optimizes and adjusts the parameters, and then retrains the multi-scale edge information recovery generative adversarial network.

[0050] S63. When the evaluation indicators of the test results: peak signal-to-noise ratio and structural similarity meet the set indicator thresholds, the training is stopped, and then the trained multi-scale edge information recovery generative adversarial network is applied to the real post-stack seismic profile.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. The present invention designs a multi-scale residual module, which can extract more features of post-stack seismic profiles from receptive fields of different scales and prevent the gradient disappearance phenomenon in the reconstruction process;

[0053] 2. The present invention designs a dual attention module, which gives greater weight to important geological data features, so that the network can better extract important feature information and prevent interference from noise signals;

[0054] 3. The present invention adopts the residual D module to replace the traditional pooling layer for downsampling operation, so as to better prevent the loss of feature information;

[0055] 4. The present invention proposes a codec network combining a multi-scale residual module, a dual attention module and a residual D module as a generating network, which can effectively improve the signal-to-noise ratio of the seismic profile while improving the resolution of the post-stack seismic profile;

[0056] 5. The present invention designs an adversarial network based on Markov discriminator to replace the traditional fully connected layer adversarial network, which is beneficial to reduce the amount of network calculation and improve the network's discrimination ability by distinguishing the blocks of the input image, thereby assisting in improving the super-resolution reconstruction performance of the generated network;

[0057] 6. The present invention proposes an improved loss function, which can enable the reconstructed high-resolution post-stack seismic profile to retain more stratigraphic structures, restore weak amplitude signals, make the stratigraphic structure clearer, and facilitate interpretation. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0059] Figure 1 Schematic diagram of the generation network structure in the post-stack seismic profile super-resolution reconstruction method;

[0060] Figure 2 Schematic diagram of the multi-scale residual module structure;

[0061] Figure 3 Schematic diagram of the dual attention module structure;

[0062] Figure 4 Schematic diagram of the residual D module;

[0063] Figure 5 Schematic diagram of the adversarial network structure in the post-stack seismic profile super-resolution reconstruction method;

[0064] Figure 6 Flow chart of the post-stack seismic profile super-resolution reconstruction method;

[0065] Figure 7 Paired low-resolution post-stack seismic profiles in the test set and high-resolution post-stack seismic profiles reconstructed using a post-stack seismic profile super-resolution reconstruction method based on a multi-scale edge information recovery generative adversarial network;

[0066] Figure 8 Single-channel amplitude comparison curves of the reconstructed high-resolution post-stack seismic profile using a post-stack seismic profile super-resolution reconstruction method based on a multi-scale edge information recovery generative adversarial network and the high-resolution post-stack seismic profile in the test set;

[0067] Fig. 9 Seismic profiles of the Netherlands survey area and high-resolution seismic profiles reconstructed using a post-stack seismic profile super-resolution reconstruction method based on a multi-scale edge information recovery generative adversarial network;

[0068] Fig.10 Single-channel amplitude comparison curves of the seismic profile in the Netherlands survey area and the high-resolution seismic profile reconstructed by the post-stack seismic profile super-resolution reconstruction method based on multi-scale edge information recovery generative adversarial network. DETAILED DESCRIPTION

[0069] The present invention will be further described below in conjunction with embodiments:

[0070] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0071] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0072] The present invention is implemented on the Python language compiler Anaconda platform, the operating system is Ubuntu20.04, the GPU is NVIDIA GeForce RTX 3090, and the deep learning framework is Pytorch. The core is to propose a post-stack seismic profile super-resolution reconstruction method based on a multi-scale edge information recovery generative adversarial network. In the process of seismic exploration data acquisition, there are many challenges, including that the high-frequency components of seismic fluctuations are attenuated by the medium during propagation, and low-frequency data cause the vertical resolution of post-stack data to decrease. Secondly, affected by the acquisition cost and terrain factors, sparse seismic signal acquisition reduces the lateral resolution of post-stack data. Finally, there are various noise interferences in seismic data, such as surface vibration, instrument noise, etc., which will reduce the signal-to-noise ratio of the data. Therefore, the present invention designs a multi-scale edge information recovery generative adversarial network for post-stack seismic profile super-resolution reconstruction, which effectively removes noise signals in seismic profiles while improving the resolution of post-stack seismic profiles. The invented network is trained, verified and tested by constructing paired low-resolution and high-resolution post-stack seismic profiles. When the evaluation index of the test result meets the set index threshold, the trained multi-scale edge information recovery generative adversarial network model is used as the optimal post-stack seismic profile super-resolution reconstruction model.

[0073] The post-stack seismic profile super-resolution reconstruction method of the present invention comprises the following steps:

[0074] S1. Construct the dataset required for training, validation and testing of post-stack seismic profile super-resolution reconstruction method.

[0075] First, seismic data parameters are defined to create a 1D subsurface reflection model;

[0076] Secondly, vertical shear and folding operations are used to simulate the complexity of geological structures, and then the formation tilt is introduced through plane shear operations to reflect the heterogeneity of the formation;

[0077] Then, by adding faults, we further simulate the fracture and displacement phenomena of the underground strata;

[0078] Next, synthetic three-dimensional low-resolution and high-resolution seismic sections are generated by convolving the model with low- and high-frequency Ricker wavelets, respectively;

[0079] Then, colored noise is added to the three-dimensional low-resolution seismic sections to increase the true complexity of the data;

[0080] Finally, the center cropping method is used to obtain two-dimensional seismic low-resolution and high-resolution sections from the three-dimensional seismic section, and the obtained two-dimensional low-resolution seismic section is downsampled by a factor of two, thus completing the entire process of seismic section data generation.

[0081] Through this process, the present invention generates a total of 3000 pairs of low-resolution-high-resolution post-stack seismic profiles, the size of the high-resolution post-stack seismic profiles is 256×256, and the size of the low-resolution post-stack seismic profiles is 128×128; among them, 2100 pairs of seismic profiles are used for training, 300 pairs of seismic profiles are used for verification, and the remaining 600 pairs of seismic profiles are used for testing.

[0082] S2. Construct a generative network in a multi-scale edge information recovery generative adversarial network, which is mainly used to effectively remove noise signals while improving the resolution of post-stack seismic profiles.

[0083] The generative network for post-stack seismic profile super-resolution reconstruction mainly consists of two parts: encoder and decoder, as shown in the figure. The encoder reduces the size of the feature map by continuous downsampling operations, which can better extract the low-frequency information in the seismic profile. The decoder reconstructs high-quality high-resolution seismic profiles through upsampling operations, which can better extract and restore the details and high-frequency information of the seismic profile. The encoder consists of a 3×3 convolutional layer with a LeakyReLU activation function, two multi-scale residual modules, two dual attention modules, two residual D modules, and three cascaded dilated convolutional layers. The low-resolution seismic profiles in the training set are first input into a 3×3 convolutional layer with a LeakyReLU activation function to extract local feature information.

[0084] Then, the feature map output by the convolutional layer is input into the multi-scale residual module and the dual attention module in sequence. The multi-scale feature extraction module is used to adjust the number of channels and extract multi-scale feature information, and the dual attention module enables the network to better extract important feature information by giving greater weight to important features.

[0085] Next, the output feature map is input into the residual D module for downsampling, which can effectively prevent the information loss caused by downsampling in the traditional pooling layer. The next-scale multi-scale residual module, dual attention module, and residual D module have the same principle and processing order as before.

[0086] Finally, the output feature map is input into three cascaded dilated convolutional layers to further expand the receptive field, which helps to better extract low-frequency feature information. The decoder mainly consists of two upsampling layers, three dual attention modules, three multi-scale residual modules, a double pixel reorganization layer, two 1×1 convolutional layers, and a 3×3 convolution with Tanh activation function. The feature map output by the encoder is first input into the first upsampling layer in the decoder. Then, the output feature map is sequentially input into the dual attention module and the multi-scale residual module for extracting important features and multi-scale feature information. Next, the feature map output by the multi-scale residual module is fused with the feature map output by the first 3×3 convolution in the encoder through a channel splicing operation. After that, a 1×1 convolutional layer is used to adjust the number of channels of the feature map. The next scale upsampling layer, dual attention module, multi-scale residual module, channel splicing operation, and 1×1 convolutional layer have the same principle and processing order as before. Then, the dual attention module and the multi-scale residual module are used for further feature extraction. Finally, a high-quality, high-resolution post-stack seismic profile is output through a double pixel binning operation and a 3×3 convolutional layer with a Tanh activation function.

[0087] Among them, the multi-scale residual module in the generation network specifically includes:

[0088] The multi-scale residual module in the generative network for post-stack seismic section super-resolution reconstruction consists of two branches, and the input feature maps of the two branches are the same, such as Figure 2 As shown. The first branch consists of a skip connection with a 1×1 convolutional layer to retain the feature information of the input; the second branch consists of three parallel dilated convolution blocks, an independent LeakyReLU activation function, and a 3×3 convolutional layer with a LeakyReLU activation function. Among them, the dilated convolution block is composed of a dilated convolution and two independent 3×3 convolutional layers. The dilated rates of the dilated convolutions in the three dilated convolution blocks are 1, 2, and 3, respectively, so that feature information can be effectively extracted under different receptive fields. The output feature maps of the three parallel dilated convolution blocks are fused by element-wise addition. Then, the fused output feature map is input into an independent LeakyReLU activation function and a 3×3 convolutional layer with a LeakyReLU activation function. The output feature map is the output feature map of the second branch. Finally, the output feature maps of the two branches are added together to obtain the output feature map of the multi-scale residual module. The size of the output feature map of the multi-scale residual module is the same as the size of its input feature map.

[0089] The dual attention module in the generative network specifically includes:

[0090] The dual attention module in the generative network for post-stack seismic profile super-resolution reconstruction is composed of a cascade of two modules: a channel attention module and a spatial attention module. The channel attention module is composed of two branches, and the input feature maps of the two branches are the same, such as Figure 3 As shown in the figure. The first branch is a skip connection to prevent information loss; the second branch consists of a global average pooling layer, a 1×1 convolution layer, and a sigmoid activation function cascade. Global average pooling adjusts the size of the feature map to 1×1; the 1×1 convolution layer is used to extract feature information; the sigmoid activation function is used to obtain the channel attention weight; the feature map output by the first branch and the channel attention weight output by the second branch are obtained by element-wise multiplication to obtain the channel attention feature map; the channel attention feature map is the input of the spatial attention module. The spatial attention module consists of two branches; the first branch is a skip connection to prevent feature information loss; the second branch consists of three small branches, a 1×1 convolution layer, a 3×3 convolution layer, and a sigmoid activation function. Among them, the first small branch is a channel global average pooling layer for aggregating global geological structure information; the second small branch is a channel global maximum pooling layer for aggregating detailed geological structure information; the third small branch is composed of a downsampling layer, three cascaded 3×3 convolution layers, and an upsampling layer cascade. The output feature maps of the three small branches are fused through channel concatenation; the fused output feature map is adjusted through a 1×1 convolution layer and a 3×3 convolution layer to adjust the number of channels and extract deep feature information. Then, the output feature map after feature extraction is activated by the Sigmoid function to obtain the spatial attention weight. The output feature map of the first branch and the spatial attention weight of the second branch output are used to obtain the output feature map of the dual attention module. The calculation process of the dual attention module can be expressed as:

[0091] DAM=S[C(x)]

[0092] Among them, DAM represents the dual attention module, C(x) represents the channel attention module, S(x) represents the spatial attention module, and x represents the input of the dual attention module; the channel attention module C(x) can be expressed as:

[0093]

[0094] Among them, δ represents the Sigmoid activation function, Conv 1×1 and Conv 3×3 They represent a 1×1 convolution layer and a 3×3 convolution layer respectively; the channel attention module S(x) can be expressed as:

[0095]

[0096] Among them, y represents the input of the spatial attention module, that is, y = C(x), GMP represents the channel-dimensional global maximum pooling layer, GAP represents the channel-dimensional global average pooling layer, Down represents the downsampling operation, Up represents the upsampling operation, Conv 3×3×3 Represents three cascaded 3×3 convolutional layers.

[0097] The residual D module in the generative network specifically includes:

[0098] The residual D module in the generative network for post-stack seismic section super-resolution reconstruction is composed of two branches, and the input feature maps of the two branches are the same, such as Figure 4 As shown in Figure 1. The first branch consists of a 1×1 convolutional layer, a 3×3 convolutional layer with a stride of 2, and a 1×1 convolutional layer in cascade. Among them, the two 1×1 convolutional layers are used to adjust the number of channels, and the 3×3 convolutional layer with a stride of 2 is used to reduce the size of the feature map to half of the original size; the feature map size output by the first branch is half of the input feature map size; the second branch consists of an average pooling layer and a 1×1 convolutional layer in cascade. Among them, the average pooling layer is used to downsample the input feature map by two times, and the 1×1 convolutional layer is used to adjust the number of channels; the feature map size output by the second branch is half of the input feature map; the output feature map of the first branch and the second branch is obtained by element-wise addition to obtain the output feature map of the residual D module.

[0099] S3. Construct an adversarial network in the multi-scale edge information recovery generative adversarial network, which is mainly used to determine whether the input image is a high-resolution post-stack seismic profile in the dataset or a high-resolution post-stack seismic profile reconstructed by the generative network.

[0100] Specifically, the adversarial network for super-resolution reconstruction of post-stack seismic profiles is a Markov discriminator-based network consisting of three 3×3 convolutions with a stride of 2 and a LeakyReLU activation function and an independent 3×3 convolution with a stride of 2 ( Figure 5 ); the output feature map size of the three 3×3 convolutions with a step size of 2 is reduced to one eighth of the input post-stack seismic profile size; the output feature map size of the last independent 3×3 convolution with a step size of 2 is reduced to one sixteenth of the input post-stack seismic profile size, and the number of channels of the output feature map is adjusted to 1; the output feature map of the last independent 3×3 convolution with a step size of 2 is the discriminant score matrix of whether the input image of the adversarial network is a high-resolution post-stack seismic profile in the dataset or a high-resolution post-stack seismic profile obtained after the generative network reconstruction.

[0101] S4. Construct a loss function to measure the post-stack seismic profile super-resolution reconstruction performance of the constructed multi-scale edge information recovery generative adversarial network.

[0102] Specifically, the loss function for super-resolution reconstruction of post-stack seismic profiles introduces a total variation loss function and an edge information preservation loss function on the conventional loss function of the generative adversarial network, and its mathematical expression is:

[0103] Loss = 0.01L adv +2×10 -8 L tv +L mse +L edge

[0104] Among them, L adv Represents the adversarial loss function, L tv represents the total variation loss function, L mse represents the mean square error loss function, L edge Represents the edge information preservation loss function;

[0105] Adversarial loss function L adv The mathematical expression is:

[0106] L adv =E[logD(y)]+E{log[1-D(G(x))]}

[0107] Where x is the input low-resolution post-stack seismic profile, y is the high-resolution post-stack seismic profile in the training set, G() is the generative network, D() is the adversarial network, G(x) is the reconstructed high-resolution post-stack seismic profile, and E represents the averaging operation;

[0108] Total variational loss function L tv Reduce the noise signal through smoothing operation; the total variation loss function L tv The mathematical expression is:

[0109]

[0110] Among them, β is set to 2;

[0111] Mean square error loss function L mse Used to measure the detail differences between the reconstructed high-quality post-stack seismic profile and the high-quality post-stack seismic profile in the test set; mean square error loss function L mse The mathematical expression is:

[0112]

[0113] Edge-preserving loss function L edge Used to measure the edge texture difference between the reconstructed high-quality post-stack seismic profile and the high-quality post-stack seismic profile in the test set; the edge preservation loss function L edge The mathematical expression is:

[0114]

[0115] Among them, Sobel() represents the edge detection operation.

[0116] S5, network model training: train the multi-scale edge information recovery generative adversarial network composed of step S2 and step S3. Through continuous cyclic adversarial training between the generative network and the adversarial network, according to the verification results of the generative network output, optimize the loss function constructed in step S4, and finally obtain the optimal model parameters of the generative network and the adversarial network, and the network model training is completed.

[0117] S6. Network model testing: The multi-scale edge information restoration generative adversarial network model obtained after the training in step S5 is tested. If the evaluation index of the test result meets the set index threshold, the multi-scale edge information restoration generative adversarial network model obtained after the training in step S5 is used as the optimal super-resolution reconstruction model; otherwise, return to step S5 and continue to train the model by adjusting the network training parameters.

[0118] Specifically, first, the paired low-resolution-high-resolution post-stack seismic profiles in the test set are input into the trained multi-scale edge information recovery generative adversarial network for testing. When the evaluation indicators of the test results: peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) do not meet the set indicator thresholds, the process returns to step S5, optimizes and adjusts the parameters and retrains the multi-scale edge information recovery generative adversarial network. When the evaluation indicators of the test results: peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) meet the set indicator thresholds, the training is stopped, and then the trained multi-scale edge information recovery generative adversarial network is applied to the real post-stack seismic profile, such as Figure 6 shown.

[0119] S7. Application of network model: Input the real low-resolution noisy post-stack seismic profile into the optimal super-resolution reconstruction model to obtain the reconstructed high-resolution post-stack seismic profile.

[0120] Example 1

[0121] This embodiment provides an application of super-resolution reconstruction and denoising of low-resolution noisy post-stack seismic profiles using a post-stack seismic profile super-resolution reconstruction method based on a multi-scale edge information recovery generative adversarial network, specifically as follows:

[0122] The present invention inputs the low-resolution noisy post-stack seismic profile in the constructed test set into the trained multi-scale edge information recovery generative adversarial network of the invention, and the output result is as follows: Figure 7As shown in Figure 1, the reconstructed high-resolution post-stack seismic profile output retains a richer stratigraphic structure and the faults are displayed more clearly. Figure 8 As shown in 9, the high-resolution post-stack seismic profile reconstructed by the present invention is very close to the single-channel amplitude of the high-resolution post-stack seismic profile in the original test set, which proves that the present invention has effective super-resolution reconstruction performance of post-stack seismic profile. The trained multi-scale edge information recovery generative adversarial network model of the present invention is applied to the real seismic data of the Netherlands survey area. As shown in 9, the high-resolution post-stack seismic profile reconstructed by the present invention better recovers the weak amplitude signal, and the stratigraphic structure is clearer and the detailed features are more obvious. Fig.10 As shown, the high-resolution post-stack seismic profile reconstructed by the present invention has the same single-channel amplitude value as the post-stack seismic profile in the Netherlands survey area, and the present invention improves the main frequency of the signal, which further proves that the present invention has superior post-stack seismic profile super-resolution reconstruction performance.

[0123] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A post-stack seismic profile super-resolution reconstruction method, characterized in that: The following steps are involved: S1. Construct the data set required for training, verification and testing of post-stack seismic profile super-resolution reconstruction method; S2, constructing a generative network in a multi-scale edge information recovery generative adversarial network to improve the resolution of post-stack seismic profiles while removing noise signals; S3, constructing an adversarial network in a multi-scale edge information recovery generative adversarial network to determine whether the input image is a high-resolution post-stack seismic profile in the data set or a high-resolution post-stack seismic profile reconstructed by the generative network; S4. Construct a loss function to measure the post-stack seismic profile super-resolution reconstruction performance of the constructed multi-scale edge information recovery generative adversarial network; S5, network model training: training the multi-scale edge information recovery generative adversarial network composed of step S2 and step S3, through continuous cyclic adversarial training between the generative network and the adversarial network, according to the verification result of the generative network output, optimizing the loss function constructed in step S4, and finally obtaining the optimal model parameters of the generative network and the adversarial network, and the network model training is completed; S6, network model testing: testing the multi-scale edge information restoration generative adversarial network model obtained after training in step S5, if the evaluation index of the test result meets the set index threshold, the multi-scale edge information restoration generative adversarial network model obtained after training in step S5 is used as the optimal super-resolution reconstruction model; otherwise, returning to step S5, and continuing to train the model by adjusting the network training parameters; S7. Application of network model: Input the real low-resolution noisy post-stack seismic profile into the optimal super-resolution reconstruction model to obtain the reconstructed high-resolution post-stack seismic profile.

2. The post-stack seismic profile super-resolution reconstruction method according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11, defining seismic data parameters to create a one-dimensional subsurface reflection model; S12, by adopting vertical shear and folding operations, the complexity of geological structure is simulated, and then the formation tilt is introduced through plane shear operations to reflect the heterogeneity of the formation; S13, by adding faults, further simulating the fracture and displacement phenomena of underground strata; S14, generating synthetic three-dimensional low-resolution and high-resolution seismic profiles by convolving the model with low-frequency and high-frequency Ricker wavelets, respectively; S15, adding colored noise to 3D low-resolution seismic profiles to increase the real complexity of the data; S16. A center cropping method is used to obtain two-dimensional seismic low-resolution and high-resolution sections from three-dimensional seismic sections, and a two-fold downsampling operation is performed on the obtained two-dimensional low-resolution seismic sections to complete the entire process of generating seismic section data.

3. The post-stack seismic profile super-resolution reconstruction method according to claim 1, characterized in that: Step S2, specifically: the generative network for post-stack seismic profile super-resolution reconstruction is mainly composed of an encoder and a decoder. The encoder reduces the size of the feature map by continuous downsampling operations; the decoder reconstructs high-quality high-resolution seismic profiles by upsampling operations; the encoder is composed of a 3×3 convolutional layer with a LeakyReLU activation function, two multi-scale residual modules, two dual attention modules, two residual D modules and three cascaded hole convolutional layers; the low-resolution seismic profile in the training set is first input into a 3×3 convolutional layer with a LeakyReLU activation function; then, the feature map output by the convolutional layer is sequentially input into the multi-scale residual module and the dual attention module; Next, the output feature map is input into the residual D module for downsampling; the multi-scale residual module, dual attention module, and residual D module of the next scale have the same principle and processing order as before; finally, the output feature map is input into three cascaded dilated convolutions; The decoder mainly consists of two upsampling layers, three dual attention modules, three multi-scale residual modules, a double pixel reorganization layer, two 1×1 convolution layers, and a 3×3 convolution with Tanh activation function; the feature map output by the encoder is first input into the first upsampling layer in the decoder; then, the output feature map is sequentially input into the dual attention module and the multi-scale residual module; Next, the feature map output by the multi-scale residual module is fused with the feature map output by the first 3×3 convolution in the encoder through a channel concatenation operation; After that, a 1×1 convolutional layer is used to adjust the number of channels of the feature map; the next scale upsampling layer, dual attention module, multi-scale residual module, channel splicing operation and 1×1 convolutional layer have the same principle and processing order as before; Then, the dual attention module and the multi-scale residual module are used for further feature extraction and integration; finally, a double pixel binning operation and a 3×3 convolutional layer with a Tanh activation function are performed to output high-quality high-resolution post-stack seismic profiles.

4. The post-stack seismic profile super-resolution reconstruction method according to claim 3, characterized in that: The multi-scale residual module consists of two branches, and the input feature maps of the two branches are the same; the first branch consists of a jump connection with a 1×1 convolutional layer; the second branch consists of three parallel dilated convolution blocks, an independent LeakyReLU activation function and a 3×3 convolutional layer with a LeakyReLU activation function; wherein the dilated convolution block consists of three cascaded dilated convolution layers with the same dilation rate; the dilation rates of the three dilated convolution blocks are 1, 2, and 3, respectively; the output feature maps of the three parallel dilated convolution blocks are fused by element-wise addition; then, the fused output feature map is input into an independent LeakyReLU activation function and a 3×3 convolutional layer with a LeakyReLU activation function to obtain the output feature map of the second branch; finally, the output feature maps of the two branches are obtained by element-wise addition to obtain the output feature map of the multi-scale residual module.

5. The post-stack seismic profile super-resolution reconstruction method according to claim 3, characterized in that: The dual attention module is composed of a cascade of two modules: a channel attention module and a spatial attention module; the channel attention module is composed of two branches, and the input feature maps of the two branches are the same; the first branch is a skip connection to prevent information loss; the second branch is composed of a cascade of a global average pooling layer, a 1×1 convolution layer and a Sigmoid activation function; global average pooling adjusts the size of the feature map to 1×1; the 1×1 convolution layer is used to extract feature information; the Sigmoid activation function is used to obtain the channel attention weight; the feature map output by the first branch and the channel attention weight output by the second branch are obtained by element-wise multiplication to obtain a channel attention feature map; the channel attention feature map is the input of the spatial attention module. The spatial attention module consists of two branches; the first branch is a jump connection used to prevent feature information loss; the second branch consists of three small branches, a 1×1 convolution layer, a 3×3 convolution layer and a Sigmoid activation function; among them, the first small branch is a channel global average pooling layer for aggregating global geological structure information; the second small branch is a channel global maximum pooling layer for aggregating detailed geological structure information; the third small branch is composed of a downsampling layer, three cascaded 3×3 convolution layers and an upsampling layer cascade; the output feature maps of the three small branches are feature fused through channel splicing operations; the fused output feature map is adjusted through a 1×1 convolution layer and a 3×3 convolution layer to adjust the number of channels and extract deep feature information; then the output feature map after feature extraction is activated by the Sigmoid function to obtain the spatial attention weight; the output feature map of the first branch and the spatial attention weight output of the second branch obtain the output feature map of the dual attention module.

6. The post-stack seismic profile super-resolution reconstruction method according to claim 3, characterized in that: The residual D module consists of two branches, and the input feature maps of the two branches are the same; the first branch is composed of a 1×1 convolution layer, a 3×3 convolution layer with a step size of 2, and a 1×1 convolution layer cascaded; the feature map size output by the first branch is half the size of the input feature map; the second branch is composed of an average pooling layer and a 1×1 convolution layer cascaded; the feature map size output by the second branch is half the size of the input feature map; the output feature maps of the first branch and the second branch are obtained by element-wise addition operation to obtain the output feature map of the residual D module.

7. The post-stack seismic profile super-resolution reconstruction method according to claim 1, characterized in that: In step S3, the adversarial network is a network based on a Markov discriminator, which consists of three 3×3 convolutions with a stride of 2 and a LeakyReLU activation function and an independent 3×3 convolution with a stride of 2; the output feature map size of the three 3×3 convolutions with a stride of 2 is reduced to one eighth of the input post-stack seismic profile size; the output feature map size of the last independent 3×3 convolution with a stride of 2 is reduced to one sixteenth of the input post-stack seismic profile size, and the number of channels of the output feature map is adjusted to 1; the output feature map of the last independent 3×3 convolution with a stride of 2 is the discriminant score matrix of whether the input image of the adversarial network is a high-resolution post-stack seismic profile in the data set or a high-resolution post-stack seismic profile obtained after the generation network reconstruction.

8. The post-stack seismic profile super-resolution reconstruction method according to claim 1, characterized in that: In step S4, the loss function is a loss function that introduces a total variation loss function and an edge information preservation loss function on the conventional loss function of the generative adversarial network. Its mathematical expression is: Loss=0.01L adv +2×10 -8 L tv +L mse +L edge Among them, L adv represents the adversarial loss function, L tv represents the total variation loss function, L mse represents the mean square error loss function, L edge Represents the edge information preservation loss function; Adversarial loss function L adv The mathematical expression is: L adv =E[logD(y)]+E{log[1-D(G(x))]} Where x is the input low-resolution post-stack seismic profile, y is the high-resolution post-stack seismic profile in the training set, G() is the generative network, D() is the adversarial network, G(x) is the reconstructed high-resolution post-stack seismic profile, and E represents the averaging operation; The total variational loss L tv The mathematical expression is: Among them, β is set to 2; Mean square error loss L mse The mathematical expression is: Edge-preserving loss function L edge The mathematical expression is: Among them, Sobel() represents the edge detection operation.

9. The post-stack seismic profile super-resolution reconstruction method according to claim 1, characterized in that: In step S6, the network model test specifically includes the following steps: S61. Input the paired low-resolution and high-resolution post-stack seismic profiles in the test set into the trained multi-scale edge information recovery generative adversarial network for testing. When the evaluation index of the test result is: S62: When the peak signal-to-noise ratio and the structural similarity do not meet the set index threshold, the process returns to step S5, optimizes and adjusts the parameters, and then retrains the multi-scale edge information recovery generative adversarial network. S63. When the evaluation indicators of the test results: peak signal-to-noise ratio and structural similarity meet the set indicator thresholds, the training is stopped, and then the trained multi-scale edge information recovery generative adversarial network is applied to the real post-stack seismic profile.

Images