Seismic data resolution enhancement method based on residual channel attention network

Abstract

The invention discloses a seismic data resolution enhancement method based on a residual channel attention network, and the method comprises the steps: constructing a three-dimensional reflectivity model containing wrinkle and fault features, generating a synthetic seismic data volume through combining seismic wavelets of different frequencies, and obtaining a two-dimensional low-resolution sample and high-resolution label data; introducing a depth-dependent amplitude attenuation mechanism and a stratum-oriented coherence attenuation constraint, and forming a final training data set in combination with a plurality of data enhancement means; a residual channel attention network composed of Canny edge detection, shallow feature extraction, a multi-scale residual module, a stacked residual group and a bottleneck attention module is constructed, and an amplitude and gradient joint loss function is used for training; according to the method, the seismic data resolution and anti-noise performance are effectively improved, complex structures such as faults and stratum pinching are clearly depicted, the precision and reliability of geological interpretation are improved, and the method has good engineering application prospects.

Description

Technical Field

[0001] This invention belongs to the field of exploration seismology, and specifically relates to a method for enhancing seismic data resolution based on an improved residual channel attention network. Background Technology

[0002] With the continuous development of exploration seismic technology, seismic data plays an increasingly important role in oil and gas exploration, structural interpretation, and reservoir prediction. However, due to limitations in acquisition conditions, instrument bandwidth, and the complexity of subsurface media, conventional seismic data often suffers from narrow bandwidth, insufficient resolution, and low signal-to-noise ratio, making it difficult to clearly characterize subtle geological structural features such as faults, bedding pinch-outs, and thin interbedded layers. This not only increases the difficulty of interpreting seismic data in complex structural areas but also directly restricts exploration accuracy and the reliability of reservoir prediction. To address the problem of insufficient resolution, traditional methods typically employ techniques such as deconvolution, absorption compensation, and time-frequency analysis to improve the bandwidth and resolution of seismic data. These methods can expand the bandwidth or improve imaging clarity to some extent, but most of these methods are based on linear or empirical models and lack effective modeling of complex nonlinear wavefield characteristics, thus making it difficult to fully recover true subsurface structural information. In recent years, with the development of artificial intelligence and deep learning technologies, more and more research has attempted to apply neural networks to seismic data resolution enhancement tasks. Typical representative methods include networks based on the UNet encoder-decoder structure, which effectively capture multi-scale features of seismic data through a combination of layer-by-layer downsampling and upsampling with skip connections, thereby improving the continuity and thin-layer distinguishability of reflection bedding to some extent; and methods based on generative adversarial networks (GANs), which, through adversarial training between the generator and discriminator, make the generated results closer to real high-resolution seismic data and improve the adaptability to noise and nonlinear features. In addition, some studies have attempted to introduce residual networks (ResNet), attention mechanisms, and other improved structures to further enhance the ability to recover weak signals and characterize complex structures. These methods can improve the resolution and noise resistance of seismic data to some extent, but still have significant shortcomings: firstly, most network structures fail to fully utilize the edge information and tectonic constraints of seismic data, easily producing blurring or artifacts at fault and complex stratigraphic boundaries; secondly, the lack of reasonable modeling of the actual acquisition degradation process during training leads to insufficient generalization ability of the model in practical applications. Therefore, there is an urgent need for a new method that can combine geophysical mechanism constraints with deep learning feature extraction capabilities, highlighting subtle geological structural features while maintaining amplitude authenticity, and possessing good noise resistance and engineering feasibility, thereby providing more reliable technical support for the interpretation of seismic data in complex tectonic regions. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned deficiencies in the existing technology by providing a seismic data resolution enhancement method based on an improved residual channel attention network. This method can effectively improve the resolution and noise resistance of seismic data while maintaining amplitude accuracy, and clearly characterize complex structural features such as faults and interlayer pinch-outs.

[0004] The technical solution of the present invention is as follows:

[0005] A method for enhancing seismic data resolution based on residual channel attention networks includes the following steps:

[0006] Step S1: By summarizing typical geological structures at home and abroad, typical geophysical structure models are drawn, and an initial three-dimensional reflectance model is constructed accordingly. On this basis, the formation process of strata folds is simulated by applying vertical shear transformation, and fault cutting operation based on volume vector field is further introduced to obtain a final three-dimensional reflectance model with fold and fault characteristics that is more in line with actual geological laws.

[0007] Step S2: Convolve the final 3D reflectivity model with different types of seismic wavelets with different frequency characteristics to obtain a 3D synthetic seismic data volume, where the highest frequency data represents the label data and the remaining frequency data are 3D samples; then slice the 3D samples and label data in two dimensions along a fixed direction to obtain a series of 2D low-resolution sample data and high-resolution label data, forming an initial dataset for training the residual channel attention network. The initial dataset includes an initial sample dataset and an initial label dataset.

[0008] Step S3: Introduce a depth-dependent amplitude attenuation mechanism, and simultaneously superimpose a formation-guided coherence attenuation constraint. Combine affine transformation, random pruning, rotation and flipping data augmentation methods to perform pairwise random operations on the initial dataset to obtain the final dataset for training the residual channel attention network. The final data includes the final sample dataset and the final label dataset.

[0009] Step S4: Construct a residual attention network. The residual attention network structure includes a head, a body, and a tail. The head includes a Canny edge detection module and a shallow feature extraction module. The body includes a multi-scale residual module, 10 stacked residual groups, and a bottleneck attention module. The tail includes convolutional layers, upsampling layers, and an output module. The output module is connected sequentially through convolutional layers and upsampling layers to achieve resolution enhancement and high-quality reconstruction of low-resolution sample data. The loss function of this network structure is a joint loss function, which includes amplitude loss and gradient loss.

[0010] Step S5: Apply the Canny operator to each low-resolution sample data one by one to obtain the edge feature data corresponding to the low-resolution sample data; combine the low-resolution sample dataset and the corresponding edge feature dataset into a dual-channel input, and use it as the input of the residual channel attention network for training. With the final label dataset as a constraint, obtain the best-performing residual channel attention network model that can be used to improve the resolution of seismic data through network training and parameter optimization.

[0011] Step S6: Input any low-resolution seismic data into the best-performing residual channel attention network model for processing; through the inference process of the best-performing residual channel attention network, output the processed seismic data with higher resolution, and finally obtain a high-resolution seismic profile.

[0012] The beneficial effects of the seismic data resolution enhancement method based on the improved residual channel attention network of this invention are as follows:

[0013] 1. The method of the present invention can highlight subtle geological structural features while maintaining the authenticity of the amplitude, and achieve a clear characterization of complex structures such as faults and bedding pinch-outs.

[0014] 2. The introduction of physical mechanism-based attenuation modeling and data augmentation operations improves the robustness and generalization ability of the method under actual acquisition conditions.

[0015] 3. An improved residual channel attention network is adopted, which combines edge guidance mechanism, multi-scale residual learning and bottleneck attention module to effectively enhance the network’s feature extraction ability and noise resistance.

[0016] 4. The method of this invention has high computational efficiency and engineering feasibility, and can be widely applied to seismic data interpretation and reservoir prediction in complex structural areas, providing reliable technical support for improving exploration accuracy and efficiency. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the seismic data resolution enhancement method based on an improved residual channel attention network provided by the present invention.

[0018] Figure 2 shows the Marmousi model offset profile provided by the present invention and the high-resolution Marmousi model offset profile predicted by the trained network model: Figure 2(a) is the original offset profile without processing, Figure 2(b) is the high-resolution offset profile predicted by the original residual channel attention network, and Figure 2(c) is the high-resolution offset profile predicted by the improved residual channel attention network.

[0019] Figure 3 is a partial magnification of Figure 2, where Figure 3(a) is a partial magnification of the original unprocessed migration profile, Figure 3(b) is a partial magnification of the high-resolution migration profile predicted using the original residual channel attention network, and Figure 3(c) is a partial magnification of the high-resolution migration profile predicted using the improved residual channel attention network.

[0020] Figure 4 It is the two-dimensional wavenumber spectrum of the Marmousi model offset profile and the profiles before and after the improved method.

[0021] Figure 5 shows the frequency division scanning results of the Marmousi model migration profile after passing through the improved residual channel attention network. Figure 5(a) is 5-15 Hz, Figure 5(b) is 15-25 Hz, Figure 5(c) is 25-35 Hz, and Figure 5(d) is 35-45 Hz.

[0022] Figure 6 shows the inline offset profile of the three-dimensional real data provided by the present invention and the high-resolution offset profile predicted by the trained network model: Figure 6(a) is the original inline offset profile without processing, Figure 6(b) is the high-resolution inline offset profile predicted by the original residual channel attention network, and Figure 6(c) is the high-resolution inline offset profile predicted by the improved residual channel attention network.

[0023] Figure 7 is a partial magnification of Figure 6, where Figure 7(a) is a partial magnification of the original inline direction offset profile without processing, Figure 7(b) is a partial magnification of the high-resolution inline direction offset profile predicted using the original residual channel attention network, and Figure 7(c) is a partial magnification of the high-resolution inline direction offset profile predicted using the improved residual channel attention network.

[0024] Figure 8 shows the crossline direction offset profile of the three-dimensional actual data provided by the present invention and the high-resolution offset profile predicted by the trained network model: Figure 8(a) is the original crossline direction offset profile without processing, Figure 8(b) is the high-resolution crossline direction offset profile predicted by the original residual channel attention network, and Figure 8(c) is the high-resolution crossline direction offset profile predicted by the improved residual channel attention network.

[0025] Figure 9 is a partial magnification of Figure 8, where Figure 8(a) is a partial magnification of the original crossline direction offset profile without processing, Figure 8(b) is a partial magnification of the high-resolution crossline direction offset profile predicted using the original residual channel attention network, and Figure 8(c) is a partial magnification of the high-resolution crossline direction offset profile predicted using the improved residual channel attention network.

[0026] Figure 10 shows the depth profile of the three-dimensional actual data provided by the present invention and the high-resolution depth profile predicted by the trained network model: Figure 10(a) is the original unprocessed depth profile, Figure 10(b) is the high-resolution depth profile predicted by the original residual channel attention network, and Figure 10(c) is the high-resolution depth profile predicted by the improved residual channel attention network.

[0027] Figure 11 is a partial magnification of the depth profile of the three-dimensional actual data provided by the present invention and the high-resolution depth profile predicted by the trained network model: wherein, Figure 11(a) is a partial magnification of the original unprocessed depth profile, Figure 11(b) is a partial magnification of the high-resolution depth profile predicted by the original residual channel attention network, and Figure 11(c) is a partial magnification of the high-resolution depth profile predicted by the improved residual channel attention network. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0029] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0030] A method for enhancing seismic data resolution based on residual channel attention networks includes the following steps:

[0031] Step S1: By summarizing typical geological structures at home and abroad, typical geophysical structure models are drawn, and an initial three-dimensional reflectance model is constructed accordingly. On this basis, the formation process of strata folds is simulated by applying vertical shear transformation, and fault cutting operation based on volume vector field is further introduced to obtain a final three-dimensional reflectance model with fold and fault characteristics that is more in line with actual geological laws.

[0032] Step S2: Convolve the final 3D reflectivity model with different types of seismic wavelets with different frequency characteristics to obtain a 3D synthetic seismic data volume, where the highest frequency data represents the label data and the remaining frequency data are 3D samples; then slice the 3D samples and label data in two dimensions along a fixed direction to obtain a series of 2D low-resolution sample data and high-resolution label data, forming an initial dataset for training the residual channel attention network. The initial dataset includes an initial sample dataset and an initial label dataset.

[0033] Step S3: Introduce a depth-dependent amplitude attenuation mechanism, and simultaneously superimpose a formation-guided coherence attenuation constraint. Combine affine transformation, random pruning, rotation and flipping data augmentation methods to perform pairwise random operations on the initial dataset to obtain the final dataset for training the residual channel attention network. The final data includes the final sample dataset and the final label dataset.

[0034] Step S4: Construct a residual attention network. The residual attention network structure includes a head, a body, and a tail. The head includes a Canny edge detection module and a shallow feature extraction module. The body includes a multi-scale residual module, 10 stacked residual groups, and a bottleneck attention module. The tail includes convolutional layers, upsampling layers, and an output module. The output module is connected sequentially through convolutional layers and upsampling layers to achieve resolution enhancement and high-quality reconstruction of low-resolution sample data. The loss function of this network structure is a joint loss function, which includes amplitude loss and gradient loss.

[0035] Step S5: Apply the Canny operator to each low-resolution sample data one by one to obtain the edge feature data corresponding to the low-resolution sample data; combine the low-resolution sample dataset and the corresponding edge feature dataset into a dual-channel input, and use it as the input of the residual channel attention network for training. With the final label dataset as a constraint, obtain the best-performing residual channel attention network model that can be used to improve the resolution of seismic data through network training and parameter optimization.

[0036] Step S6: Input any low-resolution seismic data into the best-performing residual channel attention network model for processing; through the inference process of the best-performing residual channel attention network, output the processed seismic data with higher resolution, and finally obtain a high-resolution seismic profile.

[0037] Step S1 of this implementation plan includes the following sub-steps:

[0038] Step S11: Construct the initial three-dimensional reflectivity model To incorporate geological structural complexity, a vertical shear transformation is applied to the initial three-dimensional reflectivity model to simulate the tilt and folding of strata. The coordinate transformation relationship is as follows:

[0039]

[0040] in, These are the coordinates in the transformed coordinate system. These are the coordinates in the original coordinate system. The shear coefficient is along the x-direction;

[0041] Step S12: Subsequently, fault structures are introduced into the reflectivity model, and the fault plane equation is set as follows:

[0042]

[0043] in, This represents the transformed reflectance value. These are three coordinates in the spatial coordinate system, corresponding to the coordinates in the transformed coordinate system. Coordinates, coefficients The weighting coefficients in the transformation matrix represent the changes in the reflectivity model along different coordinate directions, with constant terms. This indicates the offset or translation effect that may occur during the transformation process;

[0044] Step S13: If a certain point satisfies If the point is displaced in the depth direction, it is represented as:

[0045]

[0046] in, Indicates the depth after transformation. This represents the original depth value. This refers to the change in depth, usually a correction or compensation term indicating adjustments made to the model after certain physical transformations.

[0047] Step S14: By sequentially applying shear transformation and fault displacement operations to the initial reflectivity model, the final three-dimensional reflectivity model is obtained as follows: The three-dimensional reflectance model, while maintaining the horizontal layered characteristics, introduces tilted and fault structures, which can more realistically reflect underground geological conditions and obtain a final three-dimensional reflectance model that conforms to geological laws.

[0048] Step S2 of this implementation plan includes the following steps:

[0049] Step S21: Let the sub-wave function be... Then synthesize the seismic record It can be represented as:

[0050]

[0051] in, This represents the reflected signal in three-dimensional space and time. This is a reflectivity model, representing the original reflectivity value at a given location and time. It is a time-weighted function used to adjust the influence of the reflected signal in the time dimension;

[0052] Step S22: The wavelet adopts the Ricker wavelet. The mathematical expression of the Ricker wavelet in the time domain is:

[0053]

[0054] in, It is a weighted function that varies with time. For frequency, It is a time variable. This weighting function is used to describe the attenuation characteristics of the reflected wave over time, reflecting the frequency dependence during wave propagation.

[0055] Step S23: High-resolution label data is obtained by convolving a high-frequency wavelet with a reflectivity model, which can contain richer high-frequency components; low-resolution sample data is generated by convolving a low-frequency wavelet with a reflectivity model, which has limited bandwidth and lacks detailed information.

[0056] Step S24: To maintain data consistency, after the 3D seismic analysis is completed, the 3D volume is sliced ​​in two dimensions along a fixed cross-sectional direction to extract two-dimensional data pairs corresponding to the high-resolution volume synthesis rate and low resolution. Each pair of two-dimensional data pairs is denoted as high-resolution label data and low-resolution sample data, respectively. Through the above convolution and slicing operations, paired samples that can be directly used for network training are formed, providing a data foundation for subsequent model resolution enhancement.

[0057] Step S3 of this implementation plan includes the following steps:

[0058] Step S31: Perform a Hilbert transform on the data to obtain the analytic signal.

[0059]

[0060] in, It is the original signal. It is the Hilbert transform of the signal. It is the imaginary unit. The result of this transformation is the analytical expression of the generated signal, which helps to further analyze the frequency and phase characteristics of the signal.

[0061] Step S32: Phase expansion is as follows:

[0062]

[0063] in, To reconstruct the phase of the signal, it is represented by the phase relative to the target signal. Phase difference, This method takes the phase value and obtains the best-fit phase value by optimizing the timing matching in order to further construct the dynamic model between layers.

[0064] Step S33: Phase expansion is obtained To perform dynamic time warping (DTW) and dense mesh interpolation for bedding constraints, the reconstructed bedding set is extracted.

[0065] To simulate the weakening of coherence within the bedding neighborhood, each bedding plane is horizontally extended to obtain a banded domain:

[0066]

[0067] in, Indicates the modal layer region The inner region is a strip-shaped area that is expanded horizontally, and the expanded area is determined by certain parameters. This is defined to provide an optimized spatial region for subsequent model training and signal processing;

[0068] Step S34: Assign an independent layer attenuation factor to each zonal region. Formation of a formation-guided coherence attenuation mask:

[0069]

[0070] in, This represents a formation horizontal attenuation model. It is the attenuation factor;

[0071] Simultaneously, depth-dependent amplitude decay is introduced, employing an exponential model:

[0072]

[0073] in, It is a depth-dependent decay function, representing the decay coefficient that varies with depth. For depth attenuation parameters;

[0074] The combination of the two yields the total attenuation field:

[0075]

[0076] in, Indicates depth and level The combined results of the attenuation models in the direction reflect the attenuation characteristics of geological layers in the vertical and horizontal directions;

[0077] Step S35: After physical attenuation, spatial domain data augmentation is applied synchronously. Assume the augmentation operators are a cascade of atomic transformations, with each operator triggered independently and sharing the same set of parameters to ensure consistent registration.

[0078] Step S35 of this implementation plan includes:

[0079] Step S351: Random rotation: Triggered by probability, the rotation angle is randomly sampled from a uniform distribution.

[0080]

[0081] in, For the rotation angle, and That is the maximum angle of rotation;

[0082] Its corresponding affine matrix is:

[0083]

[0084] This operation is able to Randomly rotate the data within the range to enhance the robustness of orientation invariance;

[0085] Step S352: Random flip: This can be performed in the horizontal or vertical direction, and is defined as flipping along the x-axis. and flipping along the y-axis Where H and W are the height and width of the data, respectively;

[0086] Step S353: Random scaling: triggered by probability, the scaling factor ranges from an interval Random sampling:

[0087]

[0088] in, It is a scaling factor, indicating the scaling range. This represents the amount of change in the scaling range;

[0089] Its corresponding affine matrix is:

[0090]

[0091] in, It is a scaling transformation matrix. It is a scaling factor that determines the scaling of data in the horizontal or vertical direction; this operation can enlarge or reduce data within a certain range, enhancing scale invariance.

[0092] Step S354: Random translation: The translation amounts along the x and y directions are respectively:

[0093]

[0094] in, , These represent the translation amounts in the x and y directions, respectively. , This is a parameter representing the translation range, indicating the maximum translation value in each direction;

[0095] Its affine matrix is:

[0096]

[0097] in, The translation transformation matrix is... , These represent the translation amounts in the x and y directions, respectively;

[0098] Step S355: Random cropping: Randomly select a fixed-size window from the original image and extract the region as a new sample. If it is necessary to keep the input size unchanged, then perform bilinear interpolation resampling on the cropped region.

[0099] Step S356: Elastic Deformation: Local deformation is achieved by applying a smooth random displacement field to each pixel. Let the displacement field be:

[0100]

[0101] in, These represent the position offsets along the x-axis and y-axis, respectively.

[0102] From random noise field through Gaussian kernel Smoothing results in:

[0103]

[0104] in, It is a Gaussian smoothing kernel. It is random noise uniformly sampled from the interval [−1,1], representing a smooth transformation within that interval;

[0105] The final coordinates after deformation are:

[0106]

[0107] in, The new coordinates after deformation Original coordinates , For the corresponding translation offset, this operation can simulate local nonlinear disturbances in complex structures.

[0108] Step S4 of this implementation plan includes the following steps:

[0109] Step S41: Extract edge feature maps from low-resolution seismic data using the Canny operator. This is then multiplied point-by-point with the original earthquake amplitude to ensure consistency between the structural boundary and amplitude information:

[0110]

[0111] Here, ⊙ represents point-by-point multiplication. For the edge operator results,

[0112] The final input is composed of the original low-resolution data and the edge enhancement data:

[0113]

[0114] This formula uses the original image With edge feature map By splicing the data together to form a new input, the network can simultaneously process amplitude information and edge features.

[0115] Step S42: Dual-channel input passes through the shallow feature extraction module Initial features obtained:

[0116]

[0117] in, For the feature extraction module, generate preliminary feature maps. ;

[0118] To enhance the diversity of receptive fields and improve the recovery capability of weak reflection signals, a multi-scale residual module is introduced at the network front end, which uses convolutions with different dilation rates (dilation = 1, 2, 3) to extract features in parallel.

[0119]

[0120] in, Let the i-th feature map be , The LeakyReLU activation is applied, and the three branches are concatenated, convolved, and then connected to the input residual.

[0121]

[0122] in, The final multi-scale feature map combines the preliminary feature map. and feature maps processed by multiple scales In order to fully extract multi-scale information from seismic signals,

[0123] Based on this, 10 residual groups RGs are stacked, each group consisting of multiple residual channel attention blocks RCABs. The RCABs implement channel attention through global average pooling (GAP) and a two-layer fully connected network.

[0124]

[0125] in, It is the input feature map , It is the weight matrix of the fully connected layer. Represents the global average pooling operation, and σ(⋅) represents the Sigmoid function.

[0126] The recursive expression for the residual set is:

[0127]

[0128] in, For multi-scale feature maps, and There are two bottleneck layers, which are used to extract features at different scales, and weighted features are obtained by combining spatial and channel information.

[0129] Step S43: Introduce a bottleneck attention module (BAM) at the end of the network to combine channel attention and spatial attention, and weight the features:

[0130]

[0131] in It is the weighted feature map. For channel attention mask, As a spatial attention mask, this module can focus on fault boundaries and bedding variation areas, suppressing redundant features.

[0132] Finally, the output features are processed by a convolutional layer and a tail upsampling module. The data was processed to obtain enhanced high-resolution seismic data.

[0133]

[0134] in, For tail processing module, For convolutional modules, BAM stands for Bottleneck Attention Module, which is the module that processes the input multi-scale feature maps. Key features are extracted, and the quality of the final result is improved through residual learning.

[0135] Step S44: The residual channel attention network achieves effective recovery of weak signals and accurate characterization of complex geological structures by integrating edge-guided dual-channel input, multi-scale residual learning and bottleneck attention mechanism.

[0136] Step S44 of this implementation plan includes the following steps:

[0137] Step S441: Edge-guided dual-channel input: Given a single-channel seismic slice normalized to [−1,1], first use the Canny operator to extract the structural edges on the slice. To preserve the true amplitude at the edges, multiply the edges point-by-point with the original slice to obtain the edge-enhanced slice.

[0138]

[0139] in, The edge features are calculated using the Canny operator. For the input earthquake data, This indicates element-wise multiplication;

[0140] The network input is then formed by concatenating the data along the channel dimension.

[0141]

[0142] in, It is the original input and edge features The concatenated input feature map, where B is the batch size, and H and W represent the height and width of the feature map, respectively;

[0143] Step S442: Multi-scale Feature Extractor: To simultaneously capture local details and the broader geological background, a multi-scale residual module with three-branch dilated convolutions in parallel is used. Convolutions with different dilation rates are performed on the input features, followed by 1×1 compression and 3×3 refinement, and then LeakyReLU activation to obtain three feature paths:

[0144]

[0145] in, For the i-th feature map, Let be the dilation rate of the i-th convolution. This represents the LeakyReLU activation function;

[0146] The three features are concatenated along the channel dimension and fused using a 1×1 convolution:

[0147]

[0148] in, These are the feature maps obtained from three-way convolution. This means that the three feature maps are concatenated and then fused using a 1×1 convolution.

[0149] Output using residual connection:

[0150]

[0151] in, For the final multi-scale feature map output, residual connections are used to enhance the feature information;

[0152] Step S443: Bottleneck Attention Module (BAM): Simultaneously applies channel attention (CA) and spatial attention (SA) to the features after the residual group to enhance key structural regions and suppress redundant background.

[0153] Channel attention involves first performing global average pooling to obtain channel descriptors. The weights are obtained by 1×1 dimensionality reduction and ReLU, followed by dimensionality increase and Sigmoid:

[0154]

[0155] in, , It is a weight matrix. This indicates that the Sigmoid activation function is used to output the attention weights for each channel;

[0156] Spatial attention: After reducing the number of channels by 1×1, a bottleneck convolution path with dilated convolution is concatenated to expand the receptive field, and then compressed to a single channel by BatchNorm, ReLU, and 1×1, followed by Sigmoid.

[0157]

[0158] in, , These represent 1×1 and 3×3 convolution operations, respectively. BN is the batch normalization operation. It is an activation function. Input feature map;

[0159] Jointly obtained:

[0160]

[0161] in, and These are channel attention and spatial attention modules, and the final output is a feature map weighted by these modules. ;

[0162] Step S444: L1 loss directly measures the point-by-point difference in amplitude between the predicted result and the true label. To further constrain the consistency of seismic data at structural boundaries, gradient consistency loss is introduced to ensure the similarity between the predicted and true data in horizontal and vertical gradient features. The horizontal and vertical gradients are calculated using a finite difference method:

[0163]

[0164] in, and These represent the gradients of the image in the horizontal and vertical directions, respectively. This represents the pixel value of the image at position (x, y). By calculating these gradient differences, we ensure that the model not only preserves amplitude information but also maintains structural consistency.

[0165] We obtain the gradient fields of the predicted data and the true labels respectively, and calculate their differences in the two directions. If we use the L1 form, the gradient loss can be written as:

[0166]

[0167] in, and These are the gradients of the predicted image in the horizontal and vertical directions, respectively. To predict the image, This is a real image.

[0168] If the L2 form is used, the absolute value is replaced with a square term;

[0169] Combining the above two parts, the overall form of the mixed loss is:

[0170]

[0171] Where α and β are automatically adjusted weighting factors;

[0172] Step S445: To avoid manually setting hyperparameters and ensure the dynamic balance of the two types of losses during training, an exponential moving average strategy is introduced to record the historical L1 and gradient loss statistics, and the weights are updated every N steps.

[0173] Specifically: Let the L1 loss and gradient loss of the current step be respectively... The updated moving average is then:

[0174]

[0175] in, The smoothing factor is used to control the degree of influence of historical losses. L1 and gradient loss are respectively, with the weights adjusted as follows:

[0176]

[0177] in , These are the adjustment weights for amplitude loss and gradient loss, respectively. To prevent extremely small constants with a denominator of zero, the loss function adaptively balances the constraints of amplitude and structure at different stages of training, based on the characteristics of the samples and the convergence state: in the early stages, it relies more on L1 loss to ensure the convergence of the overall amplitude, while in the middle and later stages, it gradually increases the weight of gradient loss to highlight detailed structures such as faults and interlayer pinch-outs, thereby obtaining a high-resolution network model that can balance amplitude realism and structural integrity.

[0178] Step S5 of this implementation plan is as follows: During the training process, the loss value between the predicted output and the high-resolution label data is calculated. If the loss function does not converge or its value does not reach the preset threshold, parameter adjustment and iterative training are continued until the loss function converges and the optimal parameters are obtained, thereby obtaining the network model with the best performance.

[0179] Step S6 of this implementation plan includes the following steps:

[0180] Step S61: Normalize the sample data and extract its edge information using the Canny operator. Then, multiply the edge information point by point with the original amplitude to obtain the edge enhancement amplitude. Finally, concatenate the original data and the edge enhancement data along the channel dimension to form a dual-channel input.

[0181]

[0182] in, This indicates a low-resolution input image. The edge-enhanced image obtained through the Canny operator is then stitched together and fed into the network as input.

[0183] Step S62: The dual-channel input is fed into the trained network model f(⋅;Θ∗). Through stepwise feature extraction and enhancement via shallow convolution, multi-scale residual modules, residual channel attention groups, and bottleneck attention modules, the final high-resolution prediction result is obtained.

[0184]

[0185] in, This represents the function obtained by training network Θ, used to map low-resolution input and edge features to high-resolution output. .

[0186] In this process, the network utilizes an edge-guided mechanism to enhance its sensitivity to structural boundaries, and achieves cross-scale feature aggregation and key area highlighting through multi-scale convolution and attention mechanisms, thereby significantly improving the resolution of seismic data while maintaining amplitude accuracy. The output high-resolution seismic data not only possesses stronger structural continuity and fault identification capabilities in the horizontal direction, but also more clearly depicts bedding details and pinch-out features in the vertical direction. Simultaneously, thanks to the introduction of physical attenuation modeling and random noise enhancement during training, the prediction results exhibit strong noise resistance, accurately recovering subsurface structural features even under complex acquisition conditions and low signal-to-noise ratio environments, providing reliable support for subsequent geological interpretation, reservoir prediction, and engineering applications.

[0187] This invention relates to a method for enhancing seismic data resolution based on an improved residual channel attention network. The invention includes: constructing a large number of universally applicable elastic models by analyzing the geophysical elastic characteristics in geological profiles, gravity anomaly profiles, sonic logging curves, seismic imaging profiles, and seismic elastic interpretation profiles; increasing the diversity and number of elastic models through affine transformations; obtaining three-component seismic data by numerically solving the elastic wave equation based on the elastic wave equation of anisotropic media using the finite difference method; obtaining separated P-wave and S-wave data by solving the three-component seismic data using a model-driven P-wave and S-wave separation method; and training the three-component seismic data and P-wave and S-wave data using the WCycleGAN artificial intelligence network as a dataset to obtain an intelligent P-wave and S-wave separation method. This invention introduces artificial intelligence technology into the seismic P-wave and S-wave separation process, establishing a mapping relationship between the three-component seismic data and the P-wave and S-wave data, and extracting P-wave and S-wave data from the three-component seismic data to obtain high-precision P-wave and S-wave data. The P-wave and S-wave data obtained by this method meet the accuracy requirements, have high computational efficiency, do not depend on accurate medium parameters, and can be applied to various complex geological conditions, laying a solid data foundation for inversion imaging of actual seismic data.

[0188] To further illustrate the feasibility and effectiveness of this invention, two examples are given below:

[0189] Example 1:

[0190] Figure 2 shows the Marmousi model migration profile provided by this invention and the high-resolution Marmousi model migration profile predicted by the trained network model: Figure 2(a) is the original migration profile without processing, Figure 2(b) is the high-resolution migration profile predicted using the original residual channel attention network, and Figure 2(c) is the high-resolution migration profile predicted using the improved residual channel attention network. As can be seen from Figures 2(b) and 2(c), the improved method effectively improves the resolution of the migration profile, resulting in clearer and more detailed results. Figure 3 is a magnified view of a portion of Figure 2: Figure 3(a) is a magnified view of the original migration profile without processing, Figure 3(b) is a magnified view of the high-resolution migration profile predicted using the original residual channel attention network, and Figure 3(c) is a magnified view of the high-resolution migration profile predicted using the improved residual channel attention network. The comparison shows that the improved method of this invention can maintain better structural continuity and interlayer clarity at local details, with a resolution significantly superior to the original method. Figure 4 Figure 5 shows the two-dimensional wavenumber spectra of the Marmousi model migration profile and the profiles before and after the improved method. As can be seen from the figure, the improved result retains the high wavenumber components more fully, demonstrating the method's advantage in resolution improvement. Figure 5 shows the frequency-division scanning results of the Marmousi model migration profile after passing through the improved residual channel attention network: Figure 5(a) shows the results in the 5–15 Hz band, Figure 5(b) shows the results in the 15–25 Hz band, Figure 5(c) shows the results in the 25–35 Hz band, and Figure 5(d) shows the results in the 35–45 Hz band. As can be seen from the figure, the improved method can recover signal characteristics well in different frequency bands, especially in the high-frequency band where the resolution improvement is more significant, proving the effectiveness and superiority of the method of this invention.

[0191] Example 2:

[0192] Figure 6 shows the inline offset profile of the actual 3D data and the predicted high-resolution offset profile. Figure 6(a) is the original inline offset profile without processing; Figure 6(b) is the high-resolution profile predicted using the original residual channel attention network; and Figure 6(c) is the high-resolution profile predicted using the improved residual channel attention network. Figure 7 is a magnified view of Figure 6, showing that the improved method has stronger local detail preservation capabilities in the inline direction. Figure 8 shows the crossline offset profile of the actual 3D data and the prediction results. Figure 8(a) is the original profile; Figures 8(b) and 8(c) are the prediction results of the original and improved networks, respectively. The improved method significantly improves the profile resolution. Figure 9 is a magnified view of Figure 8, showing that the improved method more clearly delineates the construction boundaries in the crossline direction. Figure 10 shows the depth profile of the actual 3D data and the prediction results. Figure 10(a) is the original depth profile; Figure 10(b) is the prediction result of the original network; and Figure 10(c) is the prediction result of the improved network. Figure 11 is a partial enlargement of Figure 10, and the results show that the method of the present invention can also enhance the continuity and detail features of the structural boundary in the depth direction.

Claims

1. A method for enhancing seismic data resolution based on residual channel attention networks, characterized in that, Includes the following steps: Step S1: By summarizing typical geological structures at home and abroad, typical geophysical structure models are drawn, and an initial three-dimensional reflectance model is constructed accordingly. On this basis, the formation process of strata folds is simulated by applying vertical shear transformation, and fault cutting operation based on volume vector field is further introduced to obtain a final three-dimensional reflectance model with fold and fault characteristics that is more in line with actual geological laws. Step S2: Convolve the final 3D reflectivity model with different types of seismic wavelets with different frequency characteristics to obtain a 3D synthetic seismic data volume, where the highest frequency data represents the label data and the remaining frequency data are 3D samples; then slice the 3D samples and label data in two dimensions along a fixed direction to obtain a series of 2D low-resolution sample data and high-resolution label data, forming an initial dataset for training the residual channel attention network. The initial dataset includes an initial sample dataset and an initial label dataset. Step S3: Introduce a depth-dependent amplitude attenuation mechanism, and simultaneously superimpose a formation-guided coherence attenuation constraint. Combine affine transformation, random pruning, rotation and flipping data augmentation methods to perform pairwise random operations on the initial dataset to obtain the final dataset for training the residual channel attention network. The final data includes the final sample dataset and the final label dataset. Step S4: Construct a residual attention network. The residual attention network structure includes a head, a body, and a tail. The head includes a Canny edge detection module and a shallow feature extraction module. The body includes a multi-scale residual module, 10 stacked residual groups, and a bottleneck attention module. The tail includes convolutional layers, upsampling layers, and an output module. The output module is connected sequentially through convolutional layers and upsampling layers to achieve resolution enhancement and high-quality reconstruction of low-resolution sample data. The loss function of this network structure is a joint loss function, which includes amplitude loss and gradient loss. Step S5: Apply the Canny operator to each low-resolution sample data one by one to obtain the edge feature data corresponding to the low-resolution sample data; combine the low-resolution sample dataset and the corresponding edge feature dataset into a dual-channel input, and use it as the input of the residual channel attention network for training. With the final label dataset as a constraint, obtain the best-performing residual channel attention network model that can be used to improve the resolution of seismic data through network training and parameter optimization. Step S6: Input any low-resolution seismic data into the best-performing residual channel attention network model for processing; through the inference process of the best-performing residual channel attention network, output the processed seismic data with higher resolution, and finally obtain a high-resolution seismic profile.

2. The seismic data resolution enhancement method based on residual channel attention network according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11: Construct the initial three-dimensional reflectivity model To incorporate geological structural complexity, a vertical shear transformation is applied to the initial three-dimensional reflectivity model to simulate the tilt and folding of strata. The coordinate transformation relationship is as follows: ； in, These are the coordinates in the transformed coordinate system. These are the coordinates in the original coordinate system. The shear coefficient is along the x-direction; Step S12: Subsequently, fault structures are introduced into the reflectivity model, and the fault plane equation is set as follows: ； in, This represents the transformed reflectance value. These are three coordinates in the spatial coordinate system, corresponding to the coordinates in the transformed coordinate system. Coordinates, coefficients The weighting coefficients in the transformation matrix represent the changes in the reflectivity model along different coordinate directions, with constant terms. This indicates the offset or translation effect that may occur during the transformation process; Step S13: If a certain point satisfies If the point is displaced in the depth direction, it is represented as: ； in, Indicates the depth after transformation. This represents the original depth value. This refers to the change in depth, usually a correction or compensation term indicating adjustments made to the model after certain physical transformations. Step S14: By sequentially applying shear transformation and fault displacement operations to the initial reflectivity model, the final three-dimensional reflectivity model is obtained as follows: The three-dimensional reflectance model, by introducing tilting and fault structures while maintaining the horizontal layered characteristics, can more realistically reflect underground geological conditions and obtain a final three-dimensional reflectance model that conforms to geological laws.

3. The seismic data resolution enhancement method based on residual channel attention network according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Let the sub-wave function be... Then synthesize the seismic record It can be represented as: ； in, This represents the reflected signal in three-dimensional space and time. This is a reflectance model, representing the original reflectance value at a given location and time. It is a time-weighted function used to adjust the influence of the reflected signal in the time dimension; Step S22: The wavelet adopts the Ricker wavelet. The mathematical expression of the Ricker wavelet in the time domain is: ； in, It is a weighted function that varies with time. For frequency, It is a time variable. This weighting function is used to describe the attenuation characteristics of the reflected wave over time, reflecting the frequency dependence during wave propagation. Step S23: High-resolution label data is obtained by convolving a high-frequency wavelet with a reflectivity model, which can contain richer high-frequency components; low-resolution sample data is generated by convolving a low-frequency wavelet with a reflectivity model, which has limited bandwidth and lacks detailed information. Step S24: To maintain data consistency, after the 3D seismic analysis is completed, the 3D volume is sliced ​​in two dimensions along a fixed cross-sectional direction to extract two-dimensional data pairs corresponding to the high-resolution volume synthesis rate and low resolution. Each pair of two-dimensional data pairs is denoted as high-resolution label data and low-resolution sample data, respectively. Through the above convolution and slicing operations, paired samples that can be directly used for network training are formed, providing a data foundation for subsequent model resolution enhancement.

4. The seismic data resolution enhancement method based on residual channel attention network according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Perform a Hilbert transform on the data to obtain the analytic signal. ； in, It is the original signal. It is the Hilbert transform of the signal. It is the imaginary unit. The result of this transformation is the analytical expression of the generated signal, which helps to further analyze the frequency and phase characteristics of the signal. Step S32: Phase expansion is as follows: ； in, To reconstruct the phase of the signal, it is represented by the phase relative to the target signal. Phase difference, This method takes the phase value and obtains the best-fit phase value by optimizing the timing matching in order to further construct the dynamic model between layers. Step S33: Phase expansion is obtained To perform dynamic time warping (DTW) and dense mesh interpolation for bedding constraints, the reconstructed bedding set is extracted. ; To simulate the weakening of coherence within the bedding neighborhood, each bedding plane is horizontally extended to obtain a banded domain: ； in, Indicates the modal layer region The inner region is a strip-shaped area that is expanded horizontally, and the expanded area is determined by certain parameters. This is defined to provide an optimized spatial region for subsequent model training and signal processing; Step S34: Assign an independent layer attenuation factor to each zonal region. Formation of a formation-guided coherence attenuation mask: ； in, This represents a formation horizontal attenuation model. It is the attenuation factor; Simultaneously, depth-dependent amplitude decay is introduced, employing an exponential model: ； in, It is a depth-dependent attenuation function, representing the attenuation coefficient that varies with depth. For depth attenuation parameters; The combination of the two yields the total attenuation field: ； in, Indicates depth and level The combined results of the attenuation models in the direction reflect the attenuation characteristics of geological layers in the vertical and horizontal directions; Step S35: After physical attenuation, spatial domain data augmentation is applied synchronously. The augmentation operators are cascaded atomic transformations, with each operator triggered independently and sharing the same set of parameters to ensure consistent registration.

5. The seismic data resolution enhancement method based on residual channel attention network according to claim 4, characterized in that, Step S35 includes: Step S351: Random rotation: triggered by probability, the rotation angle is randomly sampled from a uniform distribution: ； in, For the rotation angle, and That is the maximum angle of rotation; Its corresponding affine matrix is: ； This operation is able to Randomly rotate the data within the range to enhance the robustness of orientation invariance; Step S352: Random flip: This can be performed in the horizontal or vertical direction, and is defined as flipping along the x-axis. and flipping along the y-axis Where H and W are the height and width of the data, respectively; Step S353: Random scaling: triggered by probability, the scaling factor ranges from an interval Random sampling: ； in, It is a scaling factor, indicating the scaling range. This represents the amount of change in the scaling range; Its corresponding affine matrix is: ； in, It is a scaling transformation matrix. It is a scaling factor that determines the scaling of data in the horizontal or vertical direction; this operation can enlarge or reduce data within a certain range, enhancing scale invariance. Step S354: Random translation: The translation amounts along the x and y directions are as follows: ； in, , These represent the translation amounts in the x and y directions, respectively. , This is a parameter representing the translation range, indicating the maximum translation value in each direction; Its affine matrix is: ； in, The translation transformation matrix is... , These represent the translation amounts in the x and y directions, respectively; Step S355: Random cropping: Randomly select a fixed-size window from the original image and extract the region as a new sample. If it is necessary to keep the input size unchanged, then perform bilinear interpolation resampling on the cropped region. Step S356: Elastic Deformation: Local deformation is achieved by applying a smooth random displacement field to each pixel. Let the displacement field be: ； in, These represent the position offsets along the x-axis and y-axis, respectively. From random noise field through Gaussian kernel Smoothing results in: ； in, It is a Gaussian smoothing kernel. It is random noise uniformly sampled from the interval [−1,1], representing a smooth transformation within that interval; The final deformed coordinates are: ； in, The new coordinates after deformation Original coordinates , For the corresponding translation offset, this operation can simulate local nonlinear disturbances in complex structures.

6. The seismic data resolution enhancement method based on residual channel attention network according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Extract edge feature maps from low-resolution seismic data using the Canny operator. This is then multiplied point-by-point with the original earthquake amplitude to ensure consistency between the structural boundary and amplitude information: ； Here, ⊙ represents point-by-point multiplication. For the edge operator results, The final input is composed of the original low-resolution data and the edge enhancement data: ； This formula uses the original image With edge feature map By splicing the data together to form a new input, the network can be ensured to process both amplitude information and edge features simultaneously. Step S42: Dual-channel input passes through the shallow feature extraction module Initial features obtained: ； in, For the feature extraction module, generate preliminary feature maps. ; To enhance the diversity of receptive fields and improve the recovery capability of weak reflection signals, a multi-scale residual module is introduced at the network front end, which uses convolutions with different dilation rates (dilation = 1, 2, 3) to extract features in parallel. ； in, Let the i-th feature map be , The LeakyReLU activation is applied, and the three branches are concatenated, convolved, and then connected to the input residual. ； in, The final multi-scale feature map combines the preliminary feature map. and feature maps processed by multiple scales In order to fully extract multi-scale information from seismic signals, Based on this, 10 residual groups RGs are stacked, each group consisting of multiple residual channel attention blocks RCABs. The RCABs implement channel attention through global average pooling (GAP) and a two-layer fully connected network. ； in, It is the input feature map , It is the weight matrix of the fully connected layer. Represents the global average pooling operation, and σ(⋅) represents the Sigmoid function. The recursive expression for the residual set is: ； in, For multi-scale feature maps, and There are two bottleneck layers, which are used to extract features at different scales, and weighted features are obtained by combining spatial and channel information. Step S43: Introduce a bottleneck attention module (BAM) at the end of the network to combine channel attention and spatial attention, and weight the features: ； in It is the weighted feature map. For channel attention mask, As a spatial attention mask, this module can focus on fault boundaries and bedding variation areas, suppressing redundant features; Finally, the output features are processed by a convolutional layer and a tail upsampling module. The data was processed to obtain enhanced high-resolution seismic data. ； in, For tail processing module, For convolutional modules, BAM stands for Bottleneck Attention Module, which is the module that processes the input multi-scale feature maps. Key features are extracted, and the quality of the final result is improved through residual learning. Step S44: The residual channel attention network achieves effective recovery of weak signals and accurate characterization of complex geological structures by integrating edge-guided dual-channel input, multi-scale residual learning and bottleneck attention mechanism.

7. The seismic data resolution enhancement method based on residual channel attention network according to claim 6, characterized in that, Step S44 includes the following steps: Step S441: Edge-guided dual-channel input: Given a single-channel seismic slice normalized to [−1,1], first use the Canny operator to extract the structural edges on the slice. To preserve the true amplitude at the edges, multiply the edges point-by-point with the original slice to obtain the edge-enhanced slice. ； in, The edge features are calculated using the Canny operator. For the input earthquake data, This indicates element-wise multiplication; The network input is then formed by concatenating the data along the channel dimension. ； in, It is the original input and edge features The concatenated input feature map, where B is the batch size, and H and W represent the height and width of the feature map, respectively; Step S442: Multi-scale Feature Extractor: To simultaneously capture local details and the broader geological background, a multi-scale residual module with three-branch dilated convolutions in parallel is used. Convolutions with different dilation rates are performed on the input features, followed by 1×1 compression and 3×3 refinement, and then LeakyReLU activation to obtain three feature paths: ； in, For the i-th feature map, Let be the dilation rate of the i-th convolution. This represents the LeakyReLU activation function; The three features are concatenated along the channel dimension and fused using a 1×1 convolution: ； in, These are the feature maps obtained from three-way convolution. This means concatenating the three feature maps and then fusing them using a 1×1 convolution. Output using residual connection: ； in, For the final multi-scale feature map output, residual connections are used to enhance the feature information; Step S443: Bottleneck Attention Module (BAM): Simultaneously applies channel attention (CA) and spatial attention (SA) to the features after the residual group to enhance key structural regions and suppress redundant background. Channel attention involves first performing global average pooling to obtain channel descriptors. The weights are obtained by 1×1 dimensionality reduction and ReLU, followed by dimensionality increase and Sigmoid: ； in, , It is a weight matrix. This indicates that the Sigmoid activation function is used to output the attention weights for each channel; Spatial attention: After reducing the number of channels by 1×1, a bottleneck convolution path with dilated convolution is concatenated to expand the receptive field, and then compressed to a single channel by BatchNorm, ReLU, and 1×1, followed by Sigmoid. ； in, , These represent 1×1 and 3×3 convolution operations, respectively. BN is the batch normalization operation. It is an activation function. Input feature map; Combined results: ； in, and These are channel attention and spatial attention modules, and the final output is a feature map weighted by these modules. ; Step S444: L1 loss directly measures the point-by-point difference in amplitude between the predicted result and the true label. To further constrain the consistency of seismic data at structural boundaries, gradient consistency loss is introduced to ensure the similarity between the predicted and true data in horizontal and vertical gradient features. The horizontal and vertical gradients are calculated using a finite difference method: ； in, and These represent the gradients of the image in the horizontal and vertical directions, respectively. This represents the pixel value of the image at position (x, y). By calculating these gradient differences, we ensure that the model not only preserves amplitude information but also maintains structural consistency. We obtain the gradient fields of the predicted data and the true labels respectively, and calculate their differences in the two directions. If we use the L1 form, the gradient loss can be written as: ； in, and These are the gradients of the predicted image in the horizontal and vertical directions, respectively. To predict the image, It is a real image; If the L2 form is used, the absolute value is replaced with a square term; Combining the above two parts, the overall form of the mixed loss is: ； Where α and β are automatically adjusted weighting factors; Step S445: To avoid manually setting hyperparameters and ensure the dynamic balance of the two types of losses during training, an exponential moving average strategy is introduced to record the historical L1 and gradient loss statistics, and the weights are updated every N steps. Specifically: Let the L1 loss and gradient loss of the current step be respectively... The updated moving average is then: ； in, The smoothing factor is used to control the degree of influence of historical losses. L1 and gradient loss are respectively, with the weights adjusted as follows: ； in , These are the adjustment weights for amplitude loss and gradient loss, respectively. To prevent extremely small constants with a denominator of zero, the loss function adaptively balances the constraints of amplitude and structure at different stages of training, based on the characteristics of the samples and the convergence state: in the early stages, it relies more on L1 loss to ensure the convergence of the overall amplitude, while in the middle and later stages, it gradually increases the weight of gradient loss to highlight detailed structures such as faults and interlayer pinch-outs, thereby obtaining a high-resolution network model that can balance amplitude realism and structural integrity.

8. The seismic data resolution enhancement method based on residual channel attention network according to claim 1, characterized in that, Specifically, step S5 involves calculating the loss value between the predicted output and the high-resolution label data during the training process. If the loss function fails to converge or its value does not reach the preset threshold, parameter adjustment and iterative training continue until the loss function converges and the optimal parameters are obtained, thereby obtaining the network model with the best performance.

9. The seismic data resolution enhancement method based on residual channel attention network according to claim 1, characterized in that, Step S6 includes the following steps: Step S61: Normalize the sample data and extract its edge information using the Canny operator. Then, multiply the edge information point by point with the original amplitude to obtain the edge enhancement amplitude. Finally, concatenate the original data and the edge enhancement data along the channel dimension to form a dual-channel input. ； in, This indicates a low-resolution input image. The edge-enhanced image obtained through the Canny operator is then stitched together and fed into the network as input. Step S62: The dual-channel input is fed into the trained network model f(⋅;Θ∗). Through stepwise feature extraction and enhancement via shallow convolution, multi-scale residual modules, residual channel attention groups, and bottleneck attention modules, the final high-resolution prediction result is obtained. ； in, This represents the function obtained by training network Θ, used to map low-resolution input and edge features to high-resolution output. .

Images