A seismic data reconstruction method based on cross-domain transfer learning and double-path encoding

Abstract

The present application belongs to the technical field of seismic exploration and artificial intelligence, and relates to a seismic data reconstruction method based on cross-domain transfer learning and double-path coding, which solves the problem that the existing deep learning-based seismic data reconstruction method is highly dependent on large-scale labeled data. The method includes two stages of natural image pre-training and seismic data fine-tuning; first, a natural image dataset simulating the missing features of seismic data channels is constructed, and then the open source mask self-supervised model weight is loaded for reconstruction pre-training to realize cross-domain feature adaptation between natural images and seismic data; subsequently, in the seismic data transfer fine-tuning stage, a small amount of paired samples of the target work area are used, and relying on a specially designed double-path coding architecture, the model's representation ability for seismic data characteristics is further strengthened. The present application effectively breaks through the sample size limitation in the sparse sampling scene, and provides an efficient and reliable technical solution for high-precision and high-integrity seismic data reconstruction under the condition of few samples.

Description

Technical Field

[0001] This invention belongs to the fields of seismic exploration and artificial intelligence technology, specifically relating to a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding. Background Technology

[0002] Oil and gas resources are among the core energy sources of modern society, occupying an irreplaceable and vital position in the global energy supply system. Whether for everyday fuel use or as raw material for industrial production, a stable supply of oil and gas resources is a necessary prerequisite for the smooth operation of various economic activities. Given the significant regional clustering of oil and gas resources, the complexity of their exploration and extraction directly impacts the stable operation of the international energy market.

[0003] Seismic exploration is an indispensable core technology in oil and gas resource exploration. This field currently faces numerous challenges, particularly the issues of missing seismic data and insufficient acquisition density. These problems often stem from factors such as complex terrain limitations, constraints on exploration equipment performance, and economic cost control, ultimately resulting in flawed seismic data that cannot comprehensively and accurately reflect subsurface geological structural characteristics. Low-density data acquisition not only reduces the analytical accuracy of seismic profiles but also increases the difficulty of subsequent data processing and geological interpretation, potentially leading to biased exploration conclusions. Particularly in exploration scenarios targeting deep or complex structural oil and gas reservoirs, missing data and uneven acquisition distribution severely impact imaging quality and reduce the accuracy of reservoir location. Therefore, utilizing intelligent technologies to efficiently process seismic data, improve data reliability, and reduce exploration costs has become a key research focus in the field of seismic exploration.

[0004] For decades, seismic data reconstruction technology has been a research hotspot in both academia and industry. Traditional seismic data reconstruction methods mainly fall into the following categories: acoustic equation modeling methods, predictive filtering methods, rank-reduction algorithms, sparse transform-based methods, and dictionary learning methods. Rank-reduction algorithms include singular spectrum analysis, matrix completion techniques, principal component analysis, and Cadzow filtering. Sparse transform-based methods include wavelet transform, Radon transform, Seislet transform, Curvelet transform, and Shearlet transform. Although these classical theory-based reconstruction methods can achieve certain results in most application scenarios, they are generally limited by specific prior assumptions, such as the linear characteristics of seismic phase axes, accurate subsurface velocity field models, and the sparsity of seismic signals. Another common problem with these methods is their strong dependence on empirical parameter adjustments; unreasonable parameter settings can easily lead to a decline in method performance. Furthermore, high time costs are also a key factor restricting their application, especially when processing large-scale seismic data.

[0005] In recent years, deep learning methods have received widespread attention in the field of seismic data reconstruction, demonstrating promising application potential. In the early stages of research, researchers primarily employed Convolutional Neural Network (CNN) architectures (such as U-Net and Res-Net) and Generative Adversarial Networks (GANs) for seismic data reconstruction. However, as research into deep learning reconstruction methods deepened, scholars gradually discovered that due to the inherent local perception characteristic of CNNs, they struggled to effectively capture global features in the data, which are crucial for improving reconstruction performance. This limitation restricted further breakthroughs in the performance of deep learning methods for seismic data reconstruction. Therefore, researchers began to explore applying Transformer models to seismic data reconstruction tasks. These Transformer-based reconstruction methods, by introducing a multi-head self-attention (MSA) mechanism to capture global features, achieve superior reconstruction results compared to traditional CNN methods. However, the introduction of the MSA mechanism incurs significant computational overhead, making it difficult to meet the efficiency requirements of practical seismic exploration data processing. Meanwhile, seismic exploration data itself suffers from sample scarcity and high acquisition costs. Furthermore, seismic data from different regions and under different geological conditions exhibit significant heterogeneity, leading to poor generalization ability and unstable performance of deep learning models trained on limited seismic data. This further exacerbates the challenges of implementing seismic data reconstruction technology. It is worth noting that in fields such as computer vision, pre-trained weights obtained from training on large-scale natural images have been widely applied to various transfer learning tasks. They can leverage the general features learned from natural images to provide effective initialization for training models for specific tasks. Natural images and seismic images share underlying visual features, both containing hierarchically extractable local spatial structures, texture distribution patterns, and neighboring pixel correlation features. This provides a core foundation for the cross-domain transfer of pre-trained weights from natural images. Based on this, a dual-path encoding model architecture integrating Transformer and convolutional networks is developed. This architecture combines the global correlation capture capability of Transformer with the local fine feature extraction advantages of convolutional networks, enabling the simultaneous capture of global spatial patterns and local detailed features of seismic data. This significantly improves the model's reconstruction accuracy and is of significant research importance. Summary of the Invention

[0006] The purpose of this invention is to provide a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, so as to solve the problem of improving the integrity and reconstruction effect of the reconstructed seismic data.

[0007] This invention is achieved through the following technical solution: A seismic data reconstruction method based on cross-domain transfer learning and dual-path coding includes the following steps: S1. Construct a dataset for reconstructing models from natural image data and seismic data; S2. Construct a seismic data reconstruction network based on cross-domain transfer learning and dual-path coding to recover missing traces in seismic data; S3. Construct a loss function for reconstructing dual-path coded seismic data based on cross-domain transfer learning; S4. Model Training: The seismic data reconstruction network constructed in step S2 is trained using a cross-domain transfer learning paradigm of natural image pre-training and seismic data fine-tuning. The global and local complementary features of the seismic data are extracted and fused through a dual-path coding architecture to reconstruct the missing information. The loss function set in step S3 is optimized based on the reconstruction results. After completing the predetermined number of training rounds, the model parameters gradually converge to obtain the optimal model parameters, and the training task is completed. S5. Network Model Evaluation: Test the network model for seismic data reconstruction that was trained in step S4. If the performance evaluation index obtained during the test reaches the preset threshold requirement, the trained model can be identified as the optimal seismic data reconstruction network model. If the test index fails to meet the set standard, return to step S4 and retrain the model by adjusting the training parameters to further improve the model performance. S6. Model Deployment and Application: Input the real seismic profile to be reconstructed, which has been processed for missing random traces, into the pre-trained optimal seismic data reconstruction model, and output the reconstructed seismic profile data after model processing.

[0008] Further, step S1 specifically includes the following steps: S11. Data source determination: The three sets of 3D seismic data (Kerry, Waihapa, and Parihaka) and the open-source dataset COCO2017 (a natural image dataset) were downloaded from the SEG wiki platform and used as the raw data for the dataset construction. S12. Seismic data preprocessing: The Kerry, Waihapa, and Parihaka 3D seismic data are processed into 2D slices using a sliding window. Random trace deletion is performed on each slice to construct 1000 seismic data pairs with one-to-one correspondence between missing and complete data. These 2D seismic slices are divided into training, validation, and test sets in an 8:1:1 ratio. S13. Natural Image Dataset Processing: For 76,000 natural image data in the COCO dataset, simulate random trace deletion operation of earthquake data to generate missing samples. Perform random trace deletion operation on each image to construct natural image data pairs with one-to-one correspondence between missing and complete data. Divide these earthquake-related missing natural images into training set, validation set and test set in a ratio of 8:1:1.

[0009] Furthermore, in step S2, the seismic data reconstruction network includes a global feature encoder for extracting global dependency features, a local feature encoder for extracting local detailed structural features, and a decoder for generating data based on the above features.

[0010] Furthermore, step S2 specifically includes the following steps: S21. Construct a global feature encoder consisting of a patch embedding layer and 12 multi-head coding blocks. The input feature map is first converted into sequence features and positional encoding is added by the patch embedding layer. Then, it is passed through the multi-head coding blocks in sequence. In the multi-head coding blocks, the input is first normalized by a layer and then fed into a multi-head self-attention sub-layer. The residual is then added to the input. After being normalized by a layer again, it is fed into a feedforward neural network sub-layer and the residual is added again. Finally, the output sequence features are reshaped and converted into a feature map with a size of 14×14 and 768 channels. S22. Construct a local feature encoder containing 4 feature coding blocks to extract multi-scale features step by step. The input feature map of each feature extraction block is processed by 2 3×3 convolutional layers, 2 batch normalization layers and 2 ReLU activation functions, and then downsampled by a max pooling layer. Finally, the local feature encoder outputs 4 feature maps of different scales, which are used for skip connection fusion in the subsequent decoder. S23. Construct a decoder containing 4 feature reconstruction blocks. Each feature reconstruction block first concatenates the local and global feature channels through a feature fusion layer, then processes it through two 3×3 convolutional layers and a ReLU activation function, then through one 3×3 convolutional layer and a ReLU activation function, and finally obtains the reconstructed seismic data through bilinear reconstruction upsampling.

[0011] Furthermore, in step S3, the loss function is L2 loss, i.e., the root mean square error loss function, whose mathematical expression is: ; In the formula, These are the earthquake data after model reconstruction. It is the training set and the corresponding complete seismic data. and These represent the number of time sampling points and the number of seismic traces in the seismic profile, respectively. i , j These are the indexes for the corresponding data time dimension and the Dao collection dimension, respectively. Loss It is the root mean square error loss function, used to measure the error between the reconstructed data and the real data.

[0012] Furthermore, in step S4, the training process of the network model is divided into two stages: pre-training with natural images and fine-tuning with seismic images. S41, Natural Image Pre-training Stage: The pre-trained weights obtained by the mask autoencoder on the ImageNet dataset containing 1.2 million natural images are loaded into the global feature encoder. The natural image data obtained in step S13 is used to carry out pre-training for a seismic-like natural image reconstruction task, thus completing the cross-domain feature adaptation between natural image reconstruction and seismic data reconstruction. S42, Seismic Image Fine-tuning Stage: The weights of the global feature encoder obtained in the natural image pre-training stage of step S41 are loaded into the current global feature encoder, and combined with the local feature encoder, the seismic data produced in step S12 is used for dual-path encoding joint fine-tuning training, so as to fully capture and fuse the global structural information and local texture details of the seismic data, and realize high-precision reconstruction of the seismic data.

[0013] Further, step S5, model testing includes the following steps: S51. Input the paired seismic profiles with missing and complete sections from the test set into the trained seismic data reconstruction network based on cross-domain transfer learning and dual-path coding for testing. S52. When the evaluation indicators of the test results, such as peak signal-to-noise ratio and structural similarity, do not meet the set indicator thresholds, return to step S4, optimize and adjust the parameters, and retrain the seismic data reconstruction network based on cross-domain transfer learning and dual-path coding. S53. When the evaluation indicators of the test results, such as peak signal-to-noise ratio and structural similarity, meet the set indicator thresholds, training is stopped; then, the trained seismic data reconstruction network model based on cross-domain transfer learning and dual-path coding is applied to real seismic profiles.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention loads the weights of a masked autoencoder pre-trained on the ImageNet dataset as the initial weights of the global encoder. This masked autoencoder, pre-trained on large-scale natural image data, already possesses rich general feature extraction capabilities, providing excellent initial parameters for the global encoder, thereby improving the stability and reliability of feature extraction; 2. This invention achieves cross-domain feature adaptation between natural image reconstruction and seismic data reconstruction through pre-training of a seismic-like natural image reconstruction task. It effectively reduces the difference between the natural image and seismic data corresponding to the pre-trained weights of the mask autoencoder, improves the problem of poor adaptability of cross-domain pre-training weight reuse in the prior art, and enables the general feature extraction capability of the pre-trained model to be effectively transferred in the seismic data reconstruction task, thereby improving the model's learning effect on seismic data features. 3. This invention employs a dual-encoder collaborative architecture, including a global feature encoder, a local feature encoder, and a decoder, forming a "dual-encoder-single-decoder" feature processing and reconstruction system. The global feature encoder extracts global and deep features from seismic data, while the local feature encoder captures local details and spatial texture features. The features extracted by both are fused and reconstructed by the decoder, enabling the model to simultaneously consider both global structure and local details, effectively improving the accuracy and completeness of seismic data reconstruction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram of the seismic data reconstruction network structure based on cross-domain transfer learning and dual-path coding; Figure 2 A flowchart of the seismic data reconstruction method; Figure 3 The images show a seismic profile with missing data from the Parihaka test set, a seismic profile reconstructed using a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, a corresponding complete seismic profile, and a single-channel amplitude comparison image. (a) shows a seismic profile with missing data from the Parihaka test set, (b) shows a seismic profile reconstructed using a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, (c) shows a corresponding complete seismic profile, and (d) shows a single-channel amplitude comparison image. Figure 4 The images show a seismic profile with missing data from the Kerry test set, a seismic profile reconstructed using a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, a corresponding complete seismic profile, and a single-channel amplitude comparison image. (a) shows a seismic profile with missing data from the Kerry test set, (b) shows a seismic profile reconstructed using a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, (c) shows a corresponding complete seismic profile, and (d) shows a single-channel amplitude comparison image. Figure 5The images show a seismic profile with missing data from the Waihapa test set, a seismic profile reconstructed using a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, a corresponding complete seismic profile, and a single-channel amplitude comparison image. (a) shows a seismic profile with missing data from the Waihapa test set, (b) shows a seismic profile reconstructed using a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, (c) shows a corresponding complete seismic profile, and (d) shows a single-channel amplitude comparison image. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0018] This invention proposes a two-stage model training strategy: pre-training based on natural images for earthquake reconstruction and targeted fine-tuning based on earthquake data. Specifically, pre-trained weights from natural images are first introduced into the earthquake data reconstruction model. Task-guided pre-training is conducted by simulating missing seismic traces in natural images, allowing the model to learn the completion logic and feature mapping rules of missing earthquake data in advance. Subsequently, targeted fine-tuning is performed using a small amount of earthquake data, enabling cross-domain transfer and reuse of pre-trained weights. This significantly reduces model training and computational costs, effectively compensates for the scarcity of earthquake data, and reduces the model's dependence on annotated earthquake data. Meanwhile, Transformer models focus on global feature modeling in earthquake data reconstruction, easily neglecting local fine structures such as earthquake reflection layer details and minor fault textures, leading to insufficient reconstruction accuracy. Convolutional networks, on the other hand, are limited to local feature extraction and lack effective capture of large-scale spatial correlation features in earthquake data, making them unsuitable for reconstruction needs in complex geological scenarios. Based on this, this invention proposes a dual-path encoding model architecture that integrates Transformer and convolutional networks, such as... Figure 1 As shown, by employing a dual-encoding structure, the global correlation capture capability of the Transformer and the local fine feature extraction advantage of the convolutional network are respectively leveraged to achieve simultaneous capture of global spatial patterns and local detailed features of seismic data, thereby significantly improving the model's reconstruction accuracy. This invention was developed and debugged in a Python environment using the Anaconda platform, running on Windows 11, with an NVIDIA GeForce RTX 5060 graphics card and PyTorch 2.7.0 as the deep learning framework.

[0019] This invention relates to a seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, such as... Figure 2 As shown, it includes the following steps: S1. Constructing datasets for reconstructing models from natural and seismic data: Specifically, constructing a pre-training dataset of natural images that supports cross-domain transfer learning, as well as a fine-tuning, training, validation, and testing dataset of seismic data; S2. Construct a seismic data reconstruction network based on cross-domain transfer learning and dual-path coding to recover missing traces in seismic data; S3. Construct a loss function for dual-path coding seismic data reconstruction based on cross-domain transfer learning to evaluate the performance of the seismic data reconstruction network based on cross-domain transfer learning and dual-path coding. S4. Model Training: Train the network for cross-domain transfer learning using natural data: Train the seismic data reconstruction network based on cross-domain transfer learning and dual-path coding constructed in step S2. This training process employs a cross-domain transfer learning paradigm of natural image pre-training followed by seismic data fine-tuning. The dual-path coding architecture extracts and fuses complementary global and local features from the seismic data to reconstruct missing information. Based on the reconstruction results, the loss function set in step S3 is continuously optimized. After completing the predetermined number of training rounds, the model parameters gradually converge, ultimately yielding the optimal model parameters, at which point the training task is complete.

[0020] S5. Network Model Evaluation: The dual-channel encoded seismic data reconstruction network is trained using inherited weights. The model's performance is assessed to determine if it meets the threshold requirements. If not, the network model and parameters are optimized and retrained. Specifically, the seismic data reconstruction network model trained in step S4 is tested. If the performance evaluation metrics obtained during testing meet the preset threshold requirements, the trained model is confirmed as the optimal seismic data reconstruction network model. If the test metrics fail to meet the set standards, the process returns to step S4, and the model is retrained by adjusting the training parameters to further improve its performance.

[0021] S6. Model Deployment and Application: If the threshold requirements are met, the model will be applied to actual seismic data. Specifically, the real seismic profile to be reconstructed, after random trace missing processing, will be input into the pre-trained optimal seismic data reconstruction model. After model processing, the reconstructed seismic profile data will be output.

[0022] Step S1 specifically includes the following steps: S11. Data Source Determination: Three sets of 3D seismic data (Kerry, Waihapa, and Parihaka) were downloaded from the SEG wiki platform, along with the open-source dataset COCO2017 as a natural image dataset, which served as the raw data for constructing the dataset.

[0023] S12. Seismic Data Preprocessing: The Kerry, Waihapa, and Parihaka 3D seismic data were processed into 2D slices using a sliding window method. Random trace deletion was performed on each slice to construct 1000 seismic data pairs with one-to-one correspondence between missing and complete data. These 2D seismic slices were then divided into training, validation, and test sets in an 8:1:1 ratio.

[0024] S13. Natural Image Dataset Processing: For 76,000 natural images in the COCO dataset, simulate random trace deletion operations to generate missing samples. Perform random trace deletion on each image to construct natural image data pairs with one-to-one correspondences between missing and complete data. Divide these earthquake-related missing natural images into training, validation, and test sets in an 8:1:1 ratio.

[0025] Step S2, the complete architecture of the network model specifically includes: a global feature encoder for extracting global dependency features, a local feature encoder for extracting local detailed structural features, and a decoder for generating data based on the above features.

[0026] S21. Global Feature Encoder: Composed of a patch embedding layer and 12 multi-head encoding blocks. The input feature map is first converted into sequence features and positional encoding is added by the patch embedding layer. Then it passes through the multi-head encoding blocks in sequence. In the multi-head encoding blocks, the input is first normalized by a layer and then fed into a multi-head self-attention sub-layer. The residual is then added to the input. After being normalized by a layer again, it is fed into a feedforward neural network sub-layer and the residual is added again. Finally, the output sequence features are reshaped and converted into a feature map with a size of 14×14 and 768 channels.

[0027] S22, Local Feature Encoder: Contains 4 feature encoding blocks for progressively extracting multi-scale features. The input feature map of each feature extraction block is processed by 2 3×3 convolutional layers, 2 batch normalization layers and 2 ReLU activation functions, and then downsampled by a max pooling layer. Finally, the local feature encoder outputs 4 feature maps of different scales for subsequent skip connection fusion in the decoder.

[0028] S23. Decoder: Contains 4 feature reconstruction blocks. Each feature reconstruction block first concatenates local features with global feature channels through a feature fusion layer, then processes them through two 3×3 convolutional layers and a ReLU activation function, then through one 3×3 convolutional layer and a ReLU activation function, and finally obtains the reconstructed seismic data through bilinear reconstruction upsampling.

[0029] Step S3, the loss function is L2 loss, i.e., root mean square error loss function, whose mathematical expression is: ; in, These are the earthquake data after model reconstruction. It is the training set and the corresponding complete seismic data. and These represent the number of time sampling points and the number of seismic traces in the seismic profile, respectively. i , j These are the indexes for the corresponding data time dimension and the Dao collection dimension, respectively. Loss It is the root mean square error loss function, used to measure the error between the reconstructed data and the real data.

[0030] In step S4, the training process of the network model is divided into two stages: pre-training with natural images and fine-tuning with seismic images. S41, Natural Image Pre-training Stage: The pre-trained weights obtained by the mask autoencoder on the ImageNet dataset containing approximately 1.2 million natural images are loaded into the global feature encoder. Pre-training is carried out using the natural image data obtained in step S13 to perform a seismic-like natural image reconstruction task, thus completing the cross-domain feature adaptation between natural image reconstruction and seismic data reconstruction.

[0031] S42, Seismic Image Fine-tuning Stage: The weights of the global feature encoder obtained in the natural image pre-training stage of step S41 are loaded into the current global feature encoder, and combined with the local feature encoder, the seismic data produced in step S12 is used for dual-path encoding joint fine-tuning training, so as to fully capture and fuse the global structural information and local texture details of the seismic data, and realize high-precision reconstruction of the seismic data.

[0032] In step S5, model testing includes the following steps: S51. Input the paired seismic profiles with missing and complete sections from the test set into the trained seismic data reconstruction network based on cross-domain transfer learning and dual-path coding for testing. S52. When the evaluation indicators of the test results, such as peak signal-to-noise ratio and structural similarity, do not meet the set indicator thresholds, return to step S4, optimize and adjust the parameters, and retrain the seismic data reconstruction network based on cross-domain transfer learning and dual-path coding. S53. When the evaluation indicators of the test results, such as peak signal-to-noise ratio and structural similarity, meet the set indicator thresholds, training is stopped. Then, the trained seismic data reconstruction network model based on cross-domain transfer learning and dual-path coding is applied to the real seismic profile.

[0033] Due to environmental conditions and budget constraints, seismic data often faces problems such as missing data and sparse distribution, severely affecting the accuracy of subsequent data processing and geological interpretation. Existing deep learning-based seismic data reconstruction methods suffer from technical bottlenecks such as difficulty in effectively reusing high-quality cross-domain features and a lack of efficient feature extraction architecture. To address this, this invention proposes a seismic data reconstruction network based on cross-domain transfer learning and dual-path coding. Pre-training is performed using natural image data for seismic-like natural image reconstruction tasks, narrowing the gap between general image domain features and seismic reconstruction tasks, effectively reusing pre-trained feature extraction capabilities, and achieving high-precision restoration of missing traces. In specific implementation, a dataset with one-to-one correspondence between missing and complete natural images, and between missing and complete seismic profiles, is constructed for training, validation, and testing of the aforementioned seismic data reconstruction network based on cross-domain transfer learning and dual-path coding. When the evaluation index corresponding to the test results reaches a preset threshold, the trained seismic data reconstruction network model based on cross-domain transfer learning and dual-path coding is determined as the optimal seismic data reconstruction model. Example

[0034] A seismic data reconstruction method based on cross-domain transfer learning and dual-path coding includes the following steps: S1. Construct a pre-training dataset of natural images that supports cross-domain transfer learning, as well as a training, validation, and testing dataset for fine-tuning earthquake data; First, we downloaded three sets of 3D seismic data (Kerry, Waihapa, and Parihaka) from the SEG wiki platform, along with the open-source dataset COCO2017 as a natural image dataset, as the raw data for constructing the dataset.

[0035] Secondly, a sliding window was used to perform two-dimensional slicing on the 3D seismic data of Kerry, Waihapa, and Parihaka. Random trace deletion was performed on each slice to construct 1000 seismic data pairs with one-to-one correspondence between missing and complete data. For the 76,000 natural image data in the COCO dataset, a random trace deletion operation was simulated to generate missing samples. Random trace deletion was then performed on each image to construct natural image data pairs with one-to-one correspondence between missing and complete data.

[0036] Finally, the two-dimensional seismic slices and seismic-like missing natural images were divided into training, validation, and test sets in an 8:1:1 ratio.

[0037] S2. Construct a seismic data reconstruction network based on cross-domain transfer learning and dual-path coding to recover missing traces in the data; The complete architecture of the network model specifically includes a global feature encoder for extracting globally dependent features, a local feature encoder for extracting local detailed structural features, and a decoder for generating data based on the above features. The weights of the global feature encoder are frozen during the seismic image fine-tuning stage to achieve lightweight fine-tuning, and the adaptation to the seismic scene can be completed by training only the local feature encoder and decoder.

[0038] The aforementioned global feature encoder comprises a patch embedding layer and 12 multi-head encoding blocks. The input feature map is first converted into sequence features and positional encoding is added by the patch embedding layer. Then, it passes through the multi-head encoding blocks sequentially. In the multi-head encoding blocks, the input is first normalized by a layer and then fed into a multi-head self-attention sub-layer. The residual is then added to the input, and then normalized by another layer and fed into a feedforward neural network sub-layer. The residual is added again. Finally, the output sequence features are reshaped and converted into a feature map of size 14×14 with 768 channels.

[0039] The aforementioned local feature encoder contains four feature encoding blocks, which are used to extract multi-scale features step by step. The input feature map of each feature extraction block is processed by two 3×3 convolutional layers, two batch normalization layers and two ReLU activation functions, and then downsampled by a max pooling layer. Finally, the local feature encoder outputs four feature maps of different scales, which are used for skip connection fusion in the subsequent decoder.

[0040] The decoder consists of four feature reconstruction blocks. Each feature reconstruction block first concatenates local features with global feature channels through a feature fusion layer, then processes the data through two 3×3 convolutional layers and a ReLU activation function, followed by one 3×3 convolutional layer and a ReLU activation function. Finally, it undergoes bilinear reconstruction upsampling to obtain the reconstructed seismic data.

[0041] S3. Design a loss function to evaluate the performance of a seismic data reconstruction network based on cross-domain transfer learning and dual-path coding.

[0042] Specifically, the loss function is the L2 loss, i.e., the root mean square error loss function, and its mathematical expression is: ; in, These are the earthquake data after model reconstruction. It is the training set and the corresponding complete seismic data. and These represent the number of time sampling points and the number of seismic traces in the seismic profile, respectively. i , j These are the indexes for the corresponding data time dimension and the data set dimension, respectively. Loss It is the root mean square error loss function, used to measure the error between the reconstructed data and the real data.

[0043] S4. Model Training Process: The seismic data reconstruction network based on cross-domain transfer learning and dual-path coding constructed in step S2 is trained. This training process adopts a cross-domain transfer learning paradigm of natural image pre-training followed by seismic data fine-tuning. The dual-path coding architecture extracts and fuses complementary global and local features from the seismic data to reconstruct missing information. The loss function set in step S3 is continuously optimized based on the reconstruction results. After completing the predetermined number of training rounds, the model parameters gradually converge, ultimately yielding the optimal model parameters, at which point the training task is complete.

[0044] In the natural image pre-training stage, the pre-trained weights obtained by the mask autoencoder on the ImageNet dataset containing approximately 1.2 million natural images are loaded into the global feature encoder. The natural image data obtained in step S13 is used to carry out pre-training for a seismic-like natural image reconstruction task, thus completing the cross-domain feature adaptation between natural image reconstruction and seismic data reconstruction.

[0045] In the seismic image fine-tuning stage, the weights of the global feature encoder obtained in the natural image pre-training stage of step S41 are loaded into the current global feature encoder. Combined with the local feature encoder, the seismic data produced in step S12 is used for dual-path encoding joint fine-tuning training, thereby fully capturing and fusing the global structural information and local texture details of the seismic data to achieve high-precision reconstruction of the seismic data.

[0046] S5. Network Model Evaluation: Test the seismic data reconstruction network model trained in step S4. If the performance evaluation index obtained during the test meets the preset threshold requirements, the trained model can be identified as the optimal seismic data reconstruction network model; if the test index fails to meet the set standard, return to step S4, and retrain the model by adjusting the training parameters to further improve the model performance.

[0047] Specifically, firstly, pairs of seismic profiles with and without missing sections are input into a pre-trained seismic data reconstruction network based on cross-domain transfer learning and dual-path coding for testing. If the evaluation metrics of the test results—peak signal-to-noise ratio (PSNR) and structural similarity—do not meet the set thresholds, the process returns to step S4 to optimize and adjust the parameters before retraining the seismic data reconstruction network based on cross-domain transfer learning and dual-path coding. When the evaluation metrics of the test results—PSNR and structural similarity—meet the set thresholds, training stops, and the trained seismic data reconstruction network model based on cross-domain transfer learning and dual-path coding is applied to real seismic profiles.

[0048] S6. Model Deployment and Application: Input the real seismic profile to be reconstructed, which has been processed for missing random traces, into the pre-trained optimal seismic data reconstruction model, and output the reconstructed seismic profile data after model processing.

[0049] like Figure 3 (a)- Figure 3 As shown in (d), there are seismic profiles with missing data in the Parihaka test set, reconstructed seismic profiles obtained by the cross-domain transfer learning and dual-path coding seismic data reconstruction method of this invention, corresponding complete real seismic profiles, and corresponding single-channel amplitude comparison results.

[0050] like Figure 4 (a)- Figure 4 As shown in (d), there are seismic profiles with missing data in the Kerry test set, reconstructed seismic profiles obtained by the cross-domain transfer learning and dual-path coding seismic data reconstruction method of this invention, corresponding complete real seismic profiles, and corresponding single-channel amplitude comparison results.

[0051] like Figure 5 (a)- Figure 5 As shown in (d), there are seismic profiles with missing data in the Waihapa test set, reconstructed seismic profiles obtained by the cross-domain transfer learning and dual-path coding seismic data reconstruction method of this invention, corresponding complete real seismic profiles, and corresponding single-channel amplitude comparison results.

[0052] It will be understood by those skilled in the art 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 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 seismic data reconstruction method based on cross-domain transfer learning and dual-path coding, characterized in that, Includes the following steps: S1. Construct a dataset for reconstructing models from natural image data and seismic data; S2. Construct a seismic data reconstruction network based on cross-domain transfer learning and dual-path coding to recover missing traces in seismic data; S3. Construct a loss function for reconstructing dual-path coded seismic data based on cross-domain transfer learning; S4. Model Training: The seismic data reconstruction network constructed in step S2 is trained using a cross-domain transfer learning paradigm of natural image pre-training and seismic data fine-tuning. The global and local complementary features of the seismic data are extracted and fused through a dual-path coding architecture to reconstruct the missing information. The loss function set in step S3 is optimized based on the reconstruction results. After completing the predetermined number of training rounds, the model parameters gradually converge to obtain the optimal model parameters, and the training task is completed. S5. Network Model Evaluation: Test the network model for seismic data reconstruction that was trained in step S4. If the performance evaluation index obtained during the test reaches the preset threshold requirement, the trained model can be identified as the optimal seismic data reconstruction network model; if the test index fails to meet the set standard, return to step S4, and retrain the model by adjusting the training parameters to further improve the model performance. S6. Model Deployment and Application: Input the real seismic profile to be reconstructed, which has been processed for missing random traces, into the pre-trained optimal seismic data reconstruction model, and output the reconstructed seismic profile data after model processing. Step S1 specifically includes the following steps: S11. Data source determination: The three sets of 3D seismic data (Kerry, Waihapa, and Parihaka) and the open-source dataset COCO2017 (a natural image dataset) were downloaded from the SEG wiki platform and used as the raw data for the dataset construction. S12. Seismic data preprocessing: The Kerry, Waihapa, and Parihaka 3D seismic data are processed into 2D slices using a sliding window. Random trace deletion is performed on each slice to construct 1000 seismic data pairs with one-to-one correspondence between missing and complete data. These 2D seismic slices are divided into training, validation, and test sets in an 8:1:1 ratio. S13. Natural Image Dataset Processing: For 76,000 natural image data in the COCO dataset, simulate random deletion of earthquake data to generate missing samples. Perform random deletion on each image to construct natural image data pairs with one-to-one correspondence between missing and complete data. Divide these earthquake-related missing natural images into training, validation, and test sets in an 8:1:1 ratio. In step S4, the training process of the network model is divided into two stages: pre-training with natural images and fine-tuning with seismic images. S41, Natural Image Pre-training Stage: The pre-trained weights obtained by the mask autoencoder on the ImageNet dataset containing 1.2 million natural images are loaded into the global feature encoder. The natural image data obtained in step S13 is used to carry out pre-training for a seismic-like natural image reconstruction task, thus completing the cross-domain feature adaptation between natural image reconstruction and seismic data reconstruction. S42, Seismic Image Fine-tuning Stage: The weights of the global feature encoder obtained in the natural image pre-training stage of step S41 are loaded into the current global feature encoder, and combined with the local feature encoder, the seismic data produced in step S12 is used for dual-path encoding joint fine-tuning training, so as to fully capture and fuse the global structural information and local texture details of the seismic data, and realize high-precision reconstruction of the seismic data.

2. The seismic data reconstruction method based on cross-domain transfer learning and dual-path coding according to claim 1, characterized in that, In step S2, the seismic data reconstruction network includes a global feature encoder for extracting global dependency features, a local feature encoder for extracting local detailed structural features, and a decoder for generating data based on the above features.

3. The seismic data reconstruction method based on cross-domain transfer learning and dual-path coding according to claim 2, characterized in that, Step S2 specifically includes the following steps: S21. Construct a global feature encoder consisting of a patch embedding layer and 12 multi-head coding blocks. The input feature map is first converted into sequence features and positional encoding is added by the patch embedding layer. Then, it is passed through the multi-head coding blocks in sequence. In the multi-head coding blocks, the input is first normalized by a layer and then fed into a multi-head self-attention sub-layer. The residual is then added to the input. After being normalized by a layer again, it is fed into a feedforward neural network sub-layer and the residual is added again. Finally, the output sequence features are reshaped and converted into a feature map with a size of 14×14 and 768 channels. S22. Construct a local feature encoder containing 4 feature coding blocks to extract multi-scale features step by step. The input feature map of each feature extraction block is processed by 2 3×3 convolutional layers, 2 batch normalization layers and 2 ReLU activation functions, and then downsampled by a max pooling layer. Finally, the local feature encoder outputs 4 feature maps of different scales, which are used for skip connection fusion in the subsequent decoder. S23. Construct a decoder containing 4 feature reconstruction blocks. Each feature reconstruction block first concatenates the local and global feature channels through a feature fusion layer, then processes it through two 3×3 convolutional layers and a ReLU activation function, then through one 3×3 convolutional layer and a ReLU activation function, and finally obtains the reconstructed seismic data through bilinear reconstruction upsampling.

4. The seismic data reconstruction method based on cross-domain transfer learning and dual-path coding according to claim 1, characterized in that, In step S3, the loss function is L2 loss, which is the root mean square error loss function, and its mathematical expression is: ； In the formula, These are the earthquake data after model reconstruction. It is training focus and The corresponding complete earthquake data, and These represent the number of time sampling points and the number of seismic traces in the seismic profile, respectively. i , j These are the indexes for the corresponding data time dimension and the Dao collection dimension, respectively. Loss It is the root mean square error loss function, used to measure the error between the reconstructed data and the real data.

5. The seismic data reconstruction method based on cross-domain transfer learning and dual-path coding according to claim 1, characterized in that, Step S5, model testing includes the following steps: S51. Input the paired seismic profiles with missing and complete sections from the test set into the trained seismic data reconstruction network based on cross-domain transfer learning and dual-path coding for testing. S52. When the evaluation indicators of the test results, such as peak signal-to-noise ratio and structural similarity, do not meet the set indicator thresholds, return to step S4, optimize and adjust the parameters, and retrain the seismic data reconstruction network based on cross-domain transfer learning and dual-path coding. S53. When the evaluation indicators of the test results, such as peak signal-to-noise ratio and structural similarity, meet the set indicator thresholds, training is stopped; then, the trained seismic data reconstruction network model based on cross-domain transfer learning and dual-path coding is applied to real seismic profiles.

Images