U-Net seismic data denoising method based on multi-scale fusion boundary feature enhancement

Abstract

The invention discloses a U-Net seismic data denoising method based on multi-scale fusion boundary feature enhancement, and the method specifically comprises the steps: 1, preliminarily constructing a seismic data training data set A which is not subjected to noise enhancement, and carrying out the preprocessing of the preliminarily obtained seismic data training data set A which is not subjected to noise enhancement, obtaining a seismic data training data set B which is finally subjected to noise enhancement; step 2, building a U-Net network framework with enhanced multi-scale fusion boundary features; step 3, performing parameter configuration and optimization setting on the U-Net network with enhanced multi-scale fusion boundary features constructed in the step 2; and step 4, training and optimizing the U-Net network enhanced by multi-scale fusion boundary features. The method can retain more weak signals and detail information while suppressing seismic noise.

Description

Technical Field

[0001] This invention belongs to the technical field of geophysical exploration signal processing methods, specifically involving a U-Net seismic data denoising method that enhances boundary features through multi-scale fusion. Background Art

[0002] With the continuous advancement of seismic exploration technology, the demand for high-precision, high signal-to-noise ratio (SNR) seismic data is increasing. However, in actual seismic data acquisition, the raw seismic records typically contain a large amount of random noise, coherent noise, and complex noise interference due to environmental noise, instrument noise, and human interference. This severely reduces the data's SNR and affects the accuracy of subsequent seismic imaging, inversion, and geological interpretation. Therefore, how to efficiently suppress noise while preserving effective information of the seismic signal (such as stratigraphic boundaries, faults, and weak reflections) has become a key task in seismic data processing. Based on the differences in the time-frequency characteristics of seismic signals and noise, seismic data denoising methods are mainly divided into two categories: one is denoising methods based on time-domain features, and the other is denoising methods based on transform-domain features. Each of these methods has its own limitations in application. While time-domain feature-based denoising methods can effectively handle local noise interference, they struggle to distinguish noise that overlaps with the effective signal frequency band. For example, median filtering and mean filtering are highly effective at handling isolated outliers, but their effectiveness in removing continuous noise or noise that overlaps with the effective signal spectrum is limited.

[0003] While transform-domain-based denoising methods can better separate the frequency band characteristics of signal and noise, they also have significant limitations. For example, wavelet transform and curvelet transform methods can lose high-frequency components of the effective signal if the threshold is not set properly, and these high-frequency components often contain important stratigraphic interface information. Furthermore, these methods are highly dependent on the choice of transform basis functions, requiring specific optimization for different seismic data characteristics.

[0004] Although these traditional methods have achieved some success in seismic data denoising, they still face several major problems: most methods require manual setting of key parameters such as thresholds, basis functions, or the number of decomposition layers, resulting in poor robustness and limited adaptability; some algorithms (such as low-rank matrix factorization) involve large-scale matrix operations, leading to low computational efficiency and difficulty in meeting the real-time processing requirements of modern seismic exploration; under conditions of strong noise or low signal-to-noise ratio, traditional methods often struggle to balance denoising effectiveness with signal preservation, easily blurring weak signals or damaging important stratigraphic boundary features. Summary of the Invention

[0005] The purpose of this invention is to provide a U-Net seismic data denoising method that enhances boundary features through multi-scale fusion. This method can retain more weak signals and detailed information while suppressing seismic noise.

[0006] The technical solution adopted in this invention is a U-Net seismic data denoising method with multi-scale fusion and boundary feature enhancement, specifically as follows: Step 1: Initially construct a training dataset A of seismic data without noise enhancement, and perform various preprocessing operations on the initial training dataset A to obtain the final training dataset B of seismic data after noise enhancement. Step 2: Construct a U-Net network framework for multi-scale fusion and boundary feature enhancement; Step 3: Configure and optimize the parameters of the U-Net network constructed in Step 2 to enhance the multi-scale fusion boundary features; Step 4: Training and optimization of the U-Net network with multi-scale fusion of boundary features.

[0007] The invention is further characterized in that: Step 1 specifically includes the following steps: Step 1.1: The initial training dataset A of un-noise-enhanced seismic data includes noisy seismic data and corresponding clean, noise-free seismic data. The clean, noise-free seismic data is obtained through forward modeling of the standard Marmousi model. Specifically, the forward modeling method of the acoustic wave equation is used, and Ricker wavelet is used to numerically simulate the standard Marmousi model. A clean, noise-free synthetic shot collection record, each record being [size missing]. Random selection Each shot set record is used as the basic training data, among which ,right Each shot collection record uses Sliding window segmentation, step size is , co-generated Clean, noise-free seismic data; noisy seismic data is obtained by... Noise samples were extracted from actual seismic records from different work areas and superimposed with clean samples to obtain P noisy seismic data sets. A noisy seismic data set and its corresponding clean, noise-free seismic data set were termed a training set, resulting in a total of P training sets. The training data sets constitute the earthquake data training dataset A. Step 1.2: For the seismic data training dataset A obtained in Step 1.1, a multi-scale noise enhancement strategy is adopted. Specifically, firstly, the original noisy data is enhanced by a linear transformation with a fixed amplification factor. Then, these enhanced noisy data are recombined with clean, noise-free seismic data to obtain datasets with different noise intensities. The training dataset consists of sets of data, each maintaining a strict correspondence between noisy and clean data, forming a noise-enhanced earthquake data training dataset B.

[0008] In step 1.2, the magnification is set to 10x, 50x, 100x, and 200x.

[0009] Step 2 specifically includes the following steps: Step 2.1: The U-Net architecture is adopted as the basic framework. This framework includes a symmetrical encoder and decoder structure, and establishes a complete skip connection between each sampling stage of the encoder and decoder to ensure the full fusion of features at different scales. The encoder consists of four downsampling stages. Each downsampling stage includes a multi-scale dual-layer convolutional block and a convolutional downsampling layer, which extracts and compresses features through progressive processing. The multi-scale dual-layer convolutional block contains six convolutional kernels and is divided into two processing units. Each processing unit has three convolutional operations of different sizes set in parallel. The first processing unit is equipped with one 3×3 convolutional kernel, one 5×5 convolutional kernel, and one 7×7 convolutional kernel. The second processing unit is also equipped with one 3×3 convolutional kernel, one 5×5 convolutional kernel, and one 7×7 convolutional kernel. The decoder section contains four upsampling stages. Each upsampling stage includes a multi-scale two-layer convolutional block and a transposed convolutional upsampling layer. The feature map is magnified by the transposed convolutional upsampling layer. The multi-scale two-layer convolutional block here is the same as that used in the encoder. Skip connection part: The output features of each downsampling stage in the encoder are aligned with the corresponding upsampling stage features in the decoder through the skip connection layer, and then spliced ​​and fused in the channel dimension. Simultaneously, a multi-scale dual-layer convolutional block connects the last downsampling stage of the encoder with the first upsampling stage of the decoder. After simultaneous processing by the encoder and decoder, the multi-scale features output by the main network are obtained. In step 2.2, based on the multi-scale U-Net constructed in step 2.1, a boundary enhancement module and a selection fusion attention mechanism are further integrated. At this time, each downsampling stage contains a multi-scale dual-layer convolutional block, an attention mechanism, and a convolutional downsampling layer. Specifically, a boundary enhancement operation is added outside the multi-scale U-Net constructed in step 2.1 to integrate the data obtained from different downsampling stages: the boundary enhancement operation extracts gradient features and fuses features from the multi-scale features output by each downsampling stage of the encoder using three multi-scale dual-layer convolutional blocks, and uses the Sigmoid activation function to focus on strengthening the boundary feature expression in the seismic data to obtain the boundary enhancement features; the selection fusion attention mechanism is embedded after each multi-scale dual-layer convolutional block of the encoder. Through the synergistic effect of selecting kernel attention and efficiently extracting attention, the network's ability to focus on effective signal features is significantly improved. Step 2.3: After fusing the multi-scale features output from the main network in Step 2.1 with the boundary enhancement features obtained in Step 2.2, the final network output, i.e., the denoised seismic data, is obtained. At this point, the U-Net network framework for multi-scale fusion and boundary feature enhancement is complete. Step 3 specifically includes the following steps: Step 3.1: Input the noise-enhanced earthquake data training dataset B into the U-Net network for multi-scale fusion boundary feature enhancement built in Step 2 in batches; The multi-scale fusion boundary feature enhancement U-Net network built in steps 3.2 and 2 is mainly for seismic data denoising. Therefore, the loss function for network training is set to the MSE loss function, the specific expression of which is as follows: (1) in, Indicates coordinates as Denoising data predicted by the network. Indicates coordinates as The data is noise-free, where C and T are the number of traces and sampling points recorded in the synthetic shot gather obtained from the forward modeling simulation. Step 3.3: Optimize the MSE loss function set in Step 3.2 using an optimization algorithm. The optimization algorithm uses the Adam adaptive moment estimator, which achieves stable and efficient optimization of the loss function in Step 3.2 by dynamically adjusting the learning rate of each parameter.

[0010] In step 3.1, the network hyperparameter initialization adopts an adaptive strategy, with the base learning rate set to 0.001. The learning rate for each iteration is set to 90%-95% of the previous iteration's learning rate. Each iteration represents the network learning the data in the training set once. The batch size is set to 8 based on the GPU memory capacity.

[0011] In step 3.1, the number of training cycles is set to 50 epochs.

[0012] Step 4 specifically includes the following steps: Step 4.1: Based on the hyperparameters configured in Step 3, such as learning rate, learning period, loss function, and adaptive moment estimation optimizer Adam, perform end-to-end training on the U-Net network with multi-scale fusion boundary feature enhancement. Step 4.2: After the network training in Step 4.1 is completed, the model with the lowest MSE loss function is loaded as the optimal model into the constructed multi-scale fusion boundary feature enhancement U-Net network.

[0013] The beneficial effects of the present invention are: In recent years, deep learning technology has brought new breakthroughs to seismic data denoising due to its powerful feature learning capabilities and end-to-end processing advantages. In particular, deep learning methods such as convolutional neural networks can automatically learn complex features in seismic data, avoiding the tedious parameter tuning process of traditional methods, and demonstrating significant advantages in handling non-stationary noise and preserving signal details. Among them, the U-Net network, with its unique encoder-decoder structure and skip connection mechanism, stands out in seismic data denoising tasks. This invention proposes an end-to-end intelligent seismic data denoising method. By constructing an improved U-Net network that integrates multi-scale feature extraction and boundary enhancement mechanisms, it achieves the suppression of complex noise and the recovery of weak signals in seismic data. This method adopts an automated training process, achieving stable denoising results without manual parameter tuning, and exhibits excellent performance in maintaining the clarity of seismic data and the continuity of weak reflection signals. Attached Figure Description

[0014] Figure 1 This is a flowchart of the U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement according to the present invention; Figure 2 This is the network model constructed in the U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement of the present invention; Figure 3 This is a structural diagram of a multi-scale, two-layer convolutional block; Figure 4 This is a diagram of the multi-choice fusion attention mechanism structure; Figure 5This is a comparison of the denoising results of the proposed method with other methods when testing synthetic seismic records with a peak signal-to-noise ratio of 12dB. Figure 5 In the diagram, (a) represents a synthetic seismic record; (b) represents the denoising result after wavelet transform; (c) represents the denoising result after using a denoising convolutional neural network (DnCNN); (d) represents the denoising result after using a U-Net; (e) represents the denoising result after using a U++ network (U-Net++); (f) represents the denoising result after using a visual transducer (VIT); (g) represents the denoising result after using a convolutional visual transducer U-Net (SCU-Net); and (h) represents the denoising result after using the method proposed in this invention. Figure 6 This is a comparison chart of the noise removal results of the method proposed in this invention and other methods when testing synthetic seismic records with a peak signal-to-noise ratio of 12dB. Figure 6 In the diagram, (a) represents a pure noise recording; (b) represents the noise-filtered result after using wavelet transform; (c) represents the noise-filtered result after using a denoising convolutional neural network (DnCNN); (d) represents the noise-filtered result after using a U-Net; (e) represents the noise-filtered result after using a U++ network (U-Net++); (f) represents the noise-filtered result after using a visual transducer (VIT); (g) represents the noise-filtered result after using a convolutional visual transducer U-Net (SCU-Net); and (h) represents the noise-filtered result after using the method proposed in this invention. Figure 7 This is a comparison chart of the denoising results of the method proposed in this invention with other methods when testing actual earthquake records; Figure 7 In the table, (a) represents the actual earthquake record; (b) represents the denoising result after wavelet transform; (c) represents the denoising result after using a denoising convolutional neural network (DnCNN); (d) represents the denoising result after using a U-Net; (e) represents the denoising result after using a U++ network (U-Net++); (f) represents the denoising result after using a visual transformer (VIT); (g) represents the denoising result after using a convolutional visual transformer U-Net (SCU-Net); and (h) represents the denoising result after using the method proposed in this invention. In the diagram, 1. Input data, 2. Output data, 3. Multi-scale two-layer convolutional block, 4. Selection and fusion attention mechanism, 5. Convolutional downsampling layer, 6. Skip connection layer, 7. Transposed convolutional upsampling layer, 8. Multi-scale two-layer convolutional input data, 9. Multi-scale two-layer convolutional output data, 10. Size 3 A convolution kernel of size 3, 11, with a size of 5. A 5-kernel convolution with a size of 12 and a value of 7. 7. Convolutional kernel, 13. Select fusion attention mechanism for input data, 14. Select kernel attention mechanism, 15. Efficiently extract attention mechanism, 16. Concatenate convolutional layers, 17. Select fusion attention mechanism for output data. Detailed Implementation

[0015] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0016] This invention provides a method for denoising U-Net seismic data by enhancing boundary features across multiple scales, such as... Figure 1 As shown, please follow these steps: Step 1: Initially construct a training dataset A of seismic data without noise enhancement, and perform various preprocessing operations on the initial training dataset A to obtain the final training dataset B of seismic data after noise enhancement. Step 1 specifically includes the following steps: Step 1.1: The initial training dataset A of un-noise-enhanced seismic data includes noisy seismic data and corresponding clean, noise-free seismic data. The clean, noise-free seismic data is obtained through forward modeling of the standard Marmousi model. Specifically, the forward modeling method of the acoustic wave equation is used, and Ricker wavelet is used to numerically simulate the standard Marmousi model. A clean, noise-free synthetic shot collection record, each record being [size missing]. (Number of time sampling points) (Number of paths), randomly selected Each shot set record is used as the basic training data, among which ,right Each shot collection record uses Sliding window segmentation, step size is , co-generated Clean, noise-free seismic data; noisy seismic data is obtained by... Noise samples were extracted from actual seismic records from different work areas and superimposed with clean samples to obtain P noisy seismic data sets. A noisy seismic data set and its corresponding clean, noise-free seismic data set were termed a training set, resulting in a total of P training sets. The training data sets constitute the earthquake data training dataset A. Step 1.2: For the seismic data training dataset A obtained in Step 1.1, a multi-scale noise enhancement strategy is adopted. Specifically, firstly, the original noisy data is enhanced by a linear transformation with a fixed amplification factor, set at 10x, 50x, 100x, and 200x. Then, these enhanced noisy data are recombinated with clean, noise-free seismic data to obtain datasets with different noise intensities. The training dataset consists of sets of data, each maintaining a strict correspondence between noisy and clean data, forming a noise-enhanced earthquake data training dataset B.

[0017] Step 2: Construct a multi-scale fusion boundary feature enhancement U-Net network framework: Using the U-Net network as the main framework, replace the convolutional blocks with multi-scale double-layer convolutions, and add boundary enhancement and selection fusion attention mechanisms to improve the U-Net. Simultaneously, use a residual learning strategy in the network to complete the network framework construction, as follows: Figure 2 As shown; Step 2 specifically includes the following steps: Step 2.1: The U-Net architecture is adopted as the basic framework. This framework includes a symmetrical encoder and decoder structure, and establishes a complete skip connection between each sampling stage of the encoder and decoder to ensure the full fusion of features at different scales. The encoder consists of four downsampling stages, each containing a multi-scale dual-layer convolutional block and a convolutional downsampling layer. Feature extraction and compression are achieved through progressive processing. The multi-scale dual-layer convolutional block contains six convolutional kernels, divided into two processing units. Each processing unit has three convolutional operations of different sizes running in parallel. The first processing unit is equipped with one 3×3, one 5×5, and one 7×7 convolutional kernel, and the second processing unit is similarly equipped with one 3×3, one 5×5, and one 7×7 convolutional kernel. The different sized convolutional kernels are used to capture feature information from different receptive fields. The captured information is then spliced ​​and fused, and the channels are adjusted to ensure consistency between input and output. Rectified linear LU activation and batch normalization are used to ensure training stability. The decoder section consists of four upsampling stages. Each upsampling stage includes a multi-scale dual-layer convolutional block and a transposed convolutional upsampling layer. The transposed convolutional upsampling layer amplifies the feature map. The multi-scale dual-layer convolutional block used here is the same as that used in the encoder, which uses 3, 5, and 7-scale convolutional kernels to capture features from different receptive fields. After splicing and fusion and channel adjustment, the Rectified Linear Activation Function (ReLU) and batch normalization are used to ensure training stability. Skip connection part: The output features of each downsampling stage in the encoder are aligned with the corresponding upsampling stage features in the decoder through the skip connection layer, and then spliced ​​and fused in the channel dimension. Meanwhile, a multi-scale two-layer convolutional block is connected between the last downsampling stage of the encoder and the first upsampling stage of the decoder. After being processed by the encoder and decoder at the same time, the multi-scale features output by the main network are obtained. Step 2.2: Based on the multi-scale U-Net constructed in Step 2.1, a boundary enhancement module and a selection fusion attention mechanism are further integrated. At this point, each downsampling stage includes a multi-scale two-layer convolutional block, an attention mechanism, and a convolutional downsampling layer. Specifically, a boundary enhancement operation is added outside the multi-scale U-Net constructed in Step 2.1 to integrate the data obtained from different downsampling stages: The boundary enhancement operation extracts and fuses gradient features from the multi-scale features output by each downsampling stage of the encoder using three multi-scale two-layer convolutional blocks, and uses the Sigmoid activation function to focus on enhancing the boundary feature expression in the seismic data to obtain boundary enhancement features; The selection fusion attention mechanism is embedded after each multi-scale two-layer convolutional block of the encoder. Through the synergistic effect of selecting kernel attention and efficiently extracting attention, the network's ability to focus on effective signal features is significantly improved.

[0018] Step 2.3: After fusing the multi-scale features output by the main network in Step 2.1 with the boundary enhancement features obtained in Step 2.2, the final network output, i.e. the denoised seismic data, is obtained. At this point, the U-Net network framework for multi-scale fusion and boundary feature enhancement is completed.

[0019] Step 3: Setting training parameters for the U-Net network with multi-scale fusion boundary feature enhancement. This involves configuring and optimizing the parameters of the U-Net network with multi-scale fusion boundary feature enhancement constructed in Step 2, including network hyperparameter initialization, loss function definition, and optimization algorithm selection. Step 3 specifically includes the following steps: Step 3.1: Input the noise-enhanced earthquake data training dataset B into the U-Net network with multi-scale fusion boundary feature enhancement built in Step 2 in batches. The network hyperparameter initialization adopts an adaptive strategy, the base learning rate is set to 0.001, and the learning rate of each iteration is set to 90%-95% of the learning rate of the previous iteration. One iteration means that the network learns the data in the training set once. The batch size is set to 8 according to the GPU memory capacity, and the number of training cycles is set to 50 epochs. The multi-scale fusion boundary feature enhancement U-Net network built in steps 3.2 and 2 is mainly for seismic data denoising. Therefore, the loss function for network training is set to the MSE loss function, the specific expression of which is as follows: (1) in, Indicates coordinates as Denoising data predicted by the network. Indicates coordinates as The data consists of noise-free labeled data, where C and T represent the number of traces and sampling points recorded in the synthetic shot gather obtained from forward modeling. By minimizing the MSE, the network can progressively optimize the weights, making the prediction results approach the ideal denoising effect, thereby effectively improving the model's performance.

[0020] Step 3.3: Optimize the MSE loss function set in Step 3.2 using an optimization algorithm. The optimization algorithm uses the Adam adaptive moment estimator, which achieves stable and efficient optimization of the loss function in Step 3.2 by dynamically adjusting the learning rate of each parameter.

[0021] Step 4: Training and optimization of U-Net network with multi-scale fusion of boundary features. Based on the network parameters, loss function and optimization algorithm configured in Step 3, the constructed network is trained iteratively. The network weights are updated through the backpropagation algorithm so that the loss function value gradually converges, and finally the optimized seismic data denoising model is obtained. Step 4 specifically includes the following steps: Step 4.1: Based on the hyperparameters configured in Step 3, such as learning rate, learning period, loss function, and adaptive moment estimation optimizer Adam, perform end-to-end training on the U-Net network with multi-scale fusion boundary feature enhancement. Step 4.2: After the network training in Step 4.1 is completed, the model with the lowest MSE loss function is loaded as the optimal model into the constructed multi-scale fusion boundary feature enhancement U-Net network.

[0022] The performance evaluation and verification of the U-Net network enhanced by multi-scale fusion of boundary features were carried out. The denoising performance was verified based on the network model in step 4. Peak signal-to-noise ratio and structural similarity index were used as quantitative evaluation indicators. At the same time, the visual evaluation of seismic signal feature preservation was combined to comprehensively analyze the denoising effect and signal fidelity performance of the network model.

[0023] The specific steps include: Step S1: Regenerate the earthquake data test dataset for testing. This dataset is not included in the earthquake data training dataset B. Apply a non-overlapping sliding window (size...) to the synthetic single-shot seismic records obtained through forward modeling of the standard Marmousi model. Cut into Each sample is superimposed. Actual noise of various intensities, generated in total The earthquake data test dataset consists of 10 test samples; Step S2: Input the earthquake data test dataset obtained in step S1 into the U-Net network for multi-scale fusion boundary feature enhancement that loads the optimal model in step 4. After network processing, the denoised earthquake data is obtained. To quantitatively evaluate the denoising performance of the network, two objective metrics, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), were used to analyze the test results in step S2. A higher PSNR value indicates better denoising performance; the SSIM index ranges from [0,1], with a higher value indicating that the image is closer to the original image. The combined use of these two metrics allows for a comprehensive evaluation of the network's overall performance in noise suppression and signal preservation.

[0024] Example 1 This invention presents a multi-scale fusion boundary feature enhancement method for denoising U-Net seismic data. The experimental platform used is PyTorch, and the installation environment is a computer equipped with an Nvidia GeForce GTX 3060. Figure 1 As shown, the specific steps are as follows: Step 1: Initial Construction of Seismic Data Training Dataset A (without noise enhancement) The initial seismic data training dataset A is constructed, and various preprocessing operations are performed on the raw seismic data in the initial un-noise-enhanced training dataset A to obtain the final noise-enhanced seismic data training dataset B. Step 1 specifically includes the following steps: Step 1.1: The initial training dataset A of un-noise-enhanced seismic data includes noisy seismic data and corresponding clean, noise-free seismic data. The clean, noise-free seismic data is obtained through forward modeling of the standard Marmousi model. Specifically, the forward modeling method of the acoustic equation is used, and Ricker wavelet is used to numerically simulate the standard Marmousi model to generate 100 clean, noise-free synthetic shot gather records, each with a size of 2000. 200 (number of time sampling points) (Number of shots), randomly select 50 shot collection records as the training data, of which 50 100, using 64 for 50 gun collection records A 64-sliding window block with a step size of 32 was used to generate 18,600 clean and noise-free seismic data points. The noisy seismic data was generated by extracting noise samples from actual seismic records from 20 different work areas and overlaying them with clean samples. Each noisy seismic data point and its corresponding clean and noise-free seismic data point were called a set of training data, resulting in 18,600 sets of training data, which constitute the seismic data training dataset A. Step 1.2: For the earthquake data training dataset A obtained in Step 1.1, a multi-scale noise enhancement strategy is adopted. Specifically, the original noisy data is first enhanced by a linear transformation with a fixed amplification factor of 10x, 50x, 100x, and 200x. Then, these enhanced noisy data are recombined with clean, noise-free earthquake data to obtain 18,600 sets of training data with different noise intensities. Each set of data maintains a strict correspondence between noisy and clean data, forming the noise-enhanced earthquake data training dataset B.

[0025] Step 2: Construct a U-Net network framework with multi-scale fusion and boundary feature enhancement: Use the U-Net network as the main framework, replace the convolutional blocks with multi-scale double-layer convolutions, and add boundary enhancement and selection fusion attention mechanisms to the network to improve U-Net. At the same time, use residual learning strategies in the network to complete the construction of the network framework. Step 2 specifically includes the following steps: The multi-scale fusion boundary feature enhancement U-Net network model built in this example is as follows: Figure 2As shown, the network mainly includes an encoder structure and a corresponding decoder structure, a skip connection structure between the encoder and decoder, and a boundary enhancement part. The encoder structure, located on the left side of the network, consists of four downsampling stages, each of which includes a multi-scale dual-layer convolutional block 3, a selection fusion attention mechanism 4, and a downsampling layer 5. The decoder part, located on the right side of the network, consists of four upsampling stages, each of which includes a multi-scale dual-layer convolutional block 3 and an upsampling layer 7. A multi-scale dual-layer convolutional block 3 connects the last downsampling stage of the encoder to the first upsampling stage of the decoder. The boundary enhancement part, located below the encoder part of the network, mainly includes three multi-scale dual-layer convolutional blocks 3. The network input data 1 is processed four times in the encoder structure through a multi-scale dual-layer convolutional block 3, a selection fusion attention mechanism 4, and a downsampling layer 5. After the four serial processing, it is processed again through a multi-scale dual-layer convolutional block 3. In the decoder structure, the output of the previous multi-scale dual-layer convolutional block 3 is used as the input. In each upsampling stage, the input of the multi-scale dual-layer convolutional block 3 is integrated with the information obtained by the selection fusion attention mechanism 4 in the downsampling stage of the coding layer through a skip connection layer 6. After being processed four times in the multi-scale convolutional block 3 and upsampling layer 7, the boundary enhancement part is the superposition of the data processed by the multi-scale convolutional block 3 in each downsampling stage of the coding layer as the input of the boundary enhancement part. After being processed three more times by the multi-scale convolutional block 3, the boundary enhancement detail information is obtained. The boundary detail information is superimposed with the data processed four times in the multi-scale convolutional block 3 and upsampling layer 7 in the decoding layer to obtain the network output data 2.

[0026] The specific structure of the multi-scale dual-layer convolutional block 3 is as follows: it contains six convolutional kernels, divided into two processing units, each with three convolutional operations of different sizes running in parallel. The first processing unit is equipped with one 3×3 convolutional kernel 10, one 5×5 convolutional kernel 11, and one 7×7 convolutional kernel 12. Similarly, the second processing unit is equipped with one 3×3 convolutional kernel 10, one 5×5 convolutional kernel 11, and one 7×7 convolutional kernel 12. Figure 3 As shown; its data processing flow is as follows: the multi-scale dual-layer convolutional input data 8 first enters the first layer processing unit, and simultaneously undergoes convolution operations with three types of convolution kernels (3×3, 5×5, and 7×7) to extract features, obtaining three sets of intermediate features; then these three sets of intermediate features are input into the second layer processing unit, and again undergo secondary convolution processing through convolution kernels of corresponding sizes; finally, the three sets of features after two layers of processing are fused to obtain the multi-scale dual-layer convolutional output data 9. The specific structure of the selected fusion attention mechanism 4 is as follows: it contains two parallel attention processing units and a concatenation convolutional layer, wherein the two attention processing units are the kernel selection attention mechanism 14 and the efficient extraction attention mechanism 15, and the concatenation convolutional layer is operation 16, as shown. Figure 4 As shown, the data processing flow is as follows: Selected fusion attention input data 13 is simultaneously fed into the selection kernel attention mechanism 14 and the efficient extraction attention mechanism 15. The two attention mechanisms respectively focus on and extract features from the input data, resulting in two sets of features with different attention weights. Subsequently, these two sets of features are fed into the concatenation convolutional layer 16, where the concatenation operation is first performed to integrate the feature information, and then further fusion processing is carried out through convolution operations to finally obtain the selected fusion attention output data 17.

[0027] Step 2.1: The U-Net architecture is adopted as the basic framework. This framework includes a symmetrical encoder and decoder structure, and establishes complete skip connections between each sampling stage of the encoder and decoder to ensure the full fusion of features at different scales. The encoder consists of four downsampling stages, each containing a multi-scale dual-layer convolutional block and a convolutional downsampling layer. Feature extraction and compression are achieved through progressive processing. The multi-scale dual-layer convolutional block contains six convolutional kernels, divided into two processing units. Each processing unit has three convolutional operations of different sizes set in parallel. The first processing unit is equipped with one 3×3 convolutional kernel 10, one 5×5 convolutional kernel 11, and one 7×7 convolutional kernel 12. The second processing unit is also equipped with one 3×3 convolutional kernel 10, one 5×5 convolutional kernel 11, and one 7×7 convolutional kernel 12. The convolutional kernels of different sizes capture feature information under different receptive fields. The captured information is spliced ​​and fused and the channels are adjusted to ensure the consistency of input and output relationship. At the same time, the rectified linear activation function (ReLU) and batch normalization are used to ensure the stability of training. The decoder section comprises four upsampling stages. Each upsampling stage includes a multi-scale dual-layer convolutional block and a transposed convolutional upsampling layer. The transposed convolutional upsampling layer amplifies the feature maps. The multi-scale dual-layer convolutional block here is consistent with that used in the encoder, employing 3, 5, and 7-scale convolutional kernels to capture features from different receptive fields. After concatenation and channel adjustment, Rectified Linear Activation (ReLU) and batch normalization ensure training stability. The skip connection section aligns the output features of each downsampling stage in the encoder with the corresponding upsampling stage features in the decoder through skip connection layers, and then concatenates and fuses them along the channel dimension. Simultaneously, a multi-scale dual-layer convolutional block connects the last downsampling stage of the encoder with the first upsampling stage of the decoder. After simultaneous processing by the encoder and decoder, the multi-scale features output by the main network are obtained. Step 2.2: Based on the multi-scale U-Net constructed in Step 2.1, a boundary enhancement module and a selection fusion attention mechanism are further integrated. At this point, each downsampling stage includes a multi-scale two-layer convolutional block, an attention mechanism, and a convolutional downsampling layer. Specifically, a boundary enhancement operation is added outside the multi-scale U-Net constructed in Step 2.1 to integrate the data obtained from different downsampling stages: The boundary enhancement module extracts gradient features and fuses features from the multi-scale features output by the encoder at each stage, and uses the Sigmoid activation function to focus on enhancing the boundary feature expression in the seismic data to obtain boundary enhancement features; The selection fusion attention mechanism is embedded after each multi-scale two-layer convolutional block of the encoder. Through the synergistic effect of selecting kernel attention and efficiently extracting attention, the network's ability to focus on effective signal features is significantly improved.

[0028] Step 2.3: Fuse the multi-scale features output from the main network in Step 2.1 with the boundary enhancement features obtained in Step 2.2 to obtain the final network output, i.e., the denoised seismic data. At this point, the U-Net network framework for multi-scale fusion boundary feature enhancement is complete. Step 3: Set the training parameters for the U-Net network for multi-scale fusion boundary feature enhancement. Configure and optimize the parameters of the U-Net network for multi-scale fusion boundary feature enhancement constructed in Step 2, including network hyperparameter initialization, loss function definition, and optimization algorithm selection. Step 3 specifically includes the following steps: Step 3.1: Input the noise-enhanced earthquake data training dataset B into the U-Net network with multi-scale fusion boundary feature enhancement built in Step 2 in batches. The network hyperparameter initialization adopts an adaptive strategy, the base learning rate is set to 0.001, and the learning rate of each iteration is set to 90%-95% of the learning rate of the previous iteration. One iteration means that the network learns the data in the training set once. The batch size is set to 8 according to the GPU memory capacity, and the number of training cycles is set to 50 epochs. The multi-scale fusion boundary feature enhancement U-Net network built in steps 3.2 and 2 is mainly for seismic data denoising. Therefore, the loss function for network training is set to the MSE loss function, the specific expression of which is as follows: (1) in, Indicates coordinates as Denoising data predicted by the network. Indicates coordinates as The data consists of noise-free labeled data, where C and T represent the number of traces and sampling points recorded in the synthetic shot gather obtained from forward modeling. By minimizing the MSE, the network can progressively optimize the weights, making the prediction results approach the ideal denoising effect, thereby effectively improving the model's performance.

[0029] Step 3.3: Optimize the MSE loss function set in Step 3.2 using an optimization algorithm. The optimization algorithm uses the Adam adaptive moment estimator, which achieves stable and efficient optimization of the loss function in Step 3.2 by dynamically adjusting the learning rate of each parameter.

[0030] Step 4: Training and optimization of U-Net network with multi-scale fusion of boundary features. Based on the network parameters, loss function and optimization algorithm configured in Step 3, the constructed network is trained iteratively. The network weights are updated through the backpropagation algorithm so that the loss function value gradually converges, and finally the optimized seismic data denoising model is obtained. Step 4 specifically includes the following steps: Step 4.1: Based on the hyperparameters configured in Step 3, such as learning rate, learning period, loss function, and adaptive moment estimation optimizer Adam, perform end-to-end training on the U-Net network with multi-scale fusion boundary feature enhancement. Step 4.2: After the network training in Step 4.1 is completed, the model with the lowest MSE loss function is loaded as the optimal model into the constructed multi-scale fusion boundary feature enhancement U-Net network.

[0031] The performance evaluation and verification of the U-Net network enhanced by multi-scale fusion of boundary features were carried out. The denoising performance was verified based on the network model in step 4. Peak signal-to-noise ratio and structural similarity index were used as quantitative evaluation indicators. At the same time, the visual evaluation of seismic signal feature preservation was combined to comprehensively analyze the denoising effect and signal fidelity performance of the network model.

[0032] The specific steps include: Step S1: Regenerate the earthquake data test dataset for testing. This dataset is not included in the earthquake data training dataset B. Apply a non-overlapping sliding window (size 64) to the synthetic single-shot seismic records obtained through forward modeling of the standard Marmousi model. 64) The sample was divided into 128 samples, and four levels of actual noise were superimposed on each sample to generate a total of 512 test samples to form a seismic data test dataset; Step S2: Input the earthquake data test dataset obtained in step S1 into the U-Net network for multi-scale fusion boundary feature enhancement that loads the optimal model in step 4. After network processing, the denoised earthquake data is obtained. To quantitatively evaluate the denoising performance of the network, two objective metrics, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), were used to analyze the test results in step S2. A higher PSNR value indicates better denoising performance; the SSIM index ranges from [0,1], with a higher value indicating that the image is closer to the original image. The combined use of these two metrics allows for a comprehensive evaluation of the network's overall performance in noise suppression and signal preservation.

[0033] Experimental results: The specific implementation process of this invention was applied to synthetic seismic data and actual seismic data. The effectiveness of the proposed method was demonstrated through comparative experiments with different methods. Table 1 shows the comparison results of PSNR (dB) of the data after denoising at different noise intensities using different methods, and Table 2 shows the comparison results of SSIM of the data after denoising at different noise intensities using different methods. It can be seen that the PSNR and SSIM indices of the method proposed in this invention are higher than other methods, demonstrating better denoising capabilities. For visualization of the test results, Figure 5 (include Figure 5 (a) Figure 5 (b) Figure 5 (c) Figure 5 (d) Figure 5 (e) Figure 5 (f) Figure 5 (g) Figure 5 (h) Figure 6 (include Figure 6 (a) Figure 6 (b) Figure 6 (c) Figure 6 (d) Figure 6 (e) Figure 6 (f) Figure 6 (g) Figure 6 (h) and Figure 7 (include Figure 7 (a) Figure 7 (b) Figure 7 (c) Figure 7 (d) Figure 7 (e) Figure 7 (f) Figure 7 (g) Figure 7 (h) shows the denoising results of different methods on synthetic seismic records, the results of noise filtering, and the denoising results of actual seismic records at a peak signal-to-noise ratio of 12 dB. From Figure 5 It can be seen from this: Figure 5(b) shows that the deep learning method significantly outperforms the WT denoising method, especially in preserving seismic details. Furthermore, we can note that the continuity of the seismic signal recovered using this invention is superior to that using DnCNN, U-Net, U-Net++, VIT, and SCU-Net, such as... Figure 5 As shown by the black arrows in (c)-(h), the present invention can effectively suppress earthquake noise while preserving weak signals therein; Figure 6 As can be seen, DnCNN, U-Net, U-Net++, VIT, and SCU-Net all leave some residual effective signals. Figure 6 The green arrows in (b)-(g) provide the corresponding annotations. In contrast, this invention demonstrates superior performance in preserving seismic signals, with no significant signal leakage observed. Figure 7 It can be seen from this that: Figure 7 As shown in (a), the actual earthquake record was severely polluted by noise. From... Figure 7 (b) and Figure 7 As can be seen from (f), due to severe noise pollution, the WT method and VIT are insufficient for noise removal. Figure 7 (c) and Figure 7 (d) It can be seen that DnCNN and U-Net can remove a certain amount of noise, but most of the noise still exists. Figure 7 (e) and Figure 7 The results in (g) show that U-net++ and SCU-Net disrupt the structure of the actual seismic record during the denoising process. Figure 7 As shown in (h), the present invention is superior to the competing methods. Its denoising effect in actual data is still the best in the comparative experiment. It effectively suppresses seismic noise while preserving weak signals.

[0034] Table 1. Noise removal results at different PSNRs / dB

[0035] Table 2. Noise reduction results under different SSIMs

[0036] Through the above tests on synthetic and real data, the present invention can better suppress seismic noise compared to the comparative methods, while preserving weak signals, as shown in the indicators in Tables 1 and 2. Compared with other methods, the present invention maintains the best and most robust PSNR and SSIM values ​​under all tested noise conditions. A multi-scale strategy is used to capture features at different scales, a fusion attention mechanism is selected to reduce feature information loss, and boundary enhancement is used to restore details and signal boundaries. Combining these advantages, the denoising method proposed in this invention has better denoising performance and can be applied in practice.

[0037] Example 2 A multi-scale fusion boundary feature enhancement method for U-Net seismic data denoising is as follows: Step 1: Initially construct a training dataset A of seismic data without noise enhancement, and perform various preprocessing operations on the initial training dataset A to obtain the final training dataset B of seismic data after noise enhancement. Step 2: Construct a U-Net network framework for multi-scale fusion and boundary feature enhancement; Step 3: Configure and optimize the parameters of the U-Net network constructed in Step 2 to enhance the multi-scale fusion boundary features; Step 4: Training and optimization of the U-Net network with multi-scale fusion of boundary features.

[0038] Example 3 A multi-scale fusion boundary feature enhancement method for U-Net seismic data denoising is as follows: Step 1: Initially construct a training dataset A of seismic data without noise enhancement, and perform various preprocessing operations on the initial training dataset A to obtain the final training dataset B of seismic data after noise enhancement. Step 1 specifically includes the following steps: Step 1.1: The initial training dataset A of un-noise-enhanced seismic data includes noisy seismic data and corresponding clean, noise-free seismic data. The clean, noise-free seismic data is obtained through forward modeling of the standard Marmousi model. Specifically, the forward modeling method of the acoustic wave equation is used, and Ricker wavelet is used to numerically simulate the standard Marmousi model. A clean, noise-free synthetic shot collection record, each record being [size missing]. Random selection Each shot set record is used as the basic training data, among which ,right Each shot collection record uses Sliding window segmentation, step size is , co-generated Clean, noise-free seismic data; noisy seismic data is obtained by... Noise samples were extracted from actual seismic records from different work areas and superimposed with clean samples to obtain P noisy seismic data sets. A noisy seismic data set and its corresponding clean, noise-free seismic data set were termed a training set, resulting in a total of P training sets. The training data sets constitute the earthquake data training dataset A. Step 1.2: For the seismic data training dataset A obtained in Step 1.1, a multi-scale noise enhancement strategy is adopted. Specifically, firstly, the original noisy data is enhanced by a linear transformation with a fixed amplification factor. Then, these enhanced noisy data are recombined with clean, noise-free seismic data to obtain datasets with different noise intensities. The training dataset consists of sets of data, each maintaining a strict correspondence between noisy and clean data, forming a noise-enhanced earthquake data training dataset B.

[0039] Step 2: Construct a U-Net network framework for multi-scale fusion and boundary feature enhancement; Step 3: Configure and optimize the parameters of the U-Net network constructed in Step 2 to enhance the multi-scale fusion boundary features; Step 4: Training and optimization of the U-Net network with multi-scale fusion of boundary features.

[0040] Example 4 A multi-scale fusion boundary feature enhancement method for U-Net seismic data denoising is as follows: Step 1: Initially construct a training dataset A of seismic data without noise enhancement, and perform various preprocessing operations on the initial training dataset A to obtain the final training dataset B of seismic data after noise enhancement. Step 1 specifically includes the following steps: Step 1.1: The initial training dataset A of un-noise-enhanced seismic data includes noisy seismic data and corresponding clean, noise-free seismic data. The clean, noise-free seismic data is obtained through forward modeling of the standard Marmousi model. Specifically, the forward modeling method of the acoustic wave equation is used, and Ricker wavelet is used to numerically simulate the standard Marmousi model. A clean, noise-free synthetic shot collection record, each record being [size missing]. Random selection Each shot set record is used as the basic training data, among which ,right Each shot collection record uses Sliding window segmentation, step size is , co-generated Clean, noise-free seismic data; noisy seismic data is obtained by... Noise samples were extracted from actual seismic records from different work areas and superimposed with clean samples to obtain P noisy seismic data sets. A noisy seismic data set and its corresponding clean, noise-free seismic data set were termed a training set, resulting in a total of P training sets. The training data sets constitute the earthquake data training dataset A. Step 1.2: For the seismic data training dataset A obtained in Step 1.1, a multi-scale noise enhancement strategy is adopted. Specifically, firstly, the original noisy data is enhanced by a linear transformation with a fixed amplification factor. Then, these enhanced noisy data are recombined with clean, noise-free seismic data to obtain datasets with different noise intensities. The training dataset consists of sets of data, each maintaining a strict correspondence between noisy and clean data, forming a noise-enhanced earthquake data training dataset B. In step 1.2, the magnification factors are set to 10x, 50x, 100x, and 200x.

[0041] Step 2: Construct a U-Net network framework for multi-scale fusion and boundary feature enhancement; Step 3: Configure and optimize the parameters of the U-Net network constructed in Step 2 to enhance the multi-scale fusion boundary features; Step 4: Training and optimization of the U-Net network with multi-scale fusion of boundary features.

[0042] Example 5 The difference from Example 4 is that step 3 specifically includes the following steps: Step 3.1: Input the noise-enhanced earthquake data training dataset B into the U-Net network for multi-scale fusion boundary feature enhancement built in Step 2 in batches; The multi-scale fusion boundary feature enhancement U-Net network built in steps 3.2 and 2 is mainly for seismic data denoising. Therefore, the loss function for network training is set to the MSE loss function, the specific expression of which is as follows: (1) in, Indicates coordinates as Denoising data predicted by the network. Indicates coordinates as The data is noise-free, where C and T are the number of traces and sampling points recorded in the synthetic shot gather obtained from the forward modeling simulation. Step 3.3: Optimize the MSE loss function set in Step 3.2 using an optimization algorithm. The optimization algorithm uses the Adam adaptive moment estimator, which achieves stable and efficient optimization of the loss function in Step 3.2 by dynamically adjusting the learning rate of each parameter.

[0043] Step 4: Training and optimization of the U-Net network with multi-scale fusion of boundary features.

[0044] Example 6 The difference from Example 5 is as follows: Step 3.1: Input the noise-enhanced earthquake data training dataset B into the U-Net network with multi-scale fusion boundary feature enhancement built in Step 2 in batches.

Claims

1. A U-Net seismic data denoising method with multi-scale fusion and boundary feature enhancement, characterized in that, Specifically: Step 1: Initially construct a training dataset A of seismic data without noise enhancement, and perform various preprocessing operations on the initial training dataset A to obtain the final training dataset B of seismic data after noise enhancement. Step 2: Construct a U-Net network framework for multi-scale fusion and boundary feature enhancement; Step 3: Configure and optimize the parameters of the U-Net network constructed in Step 2 to enhance the multi-scale fusion boundary features; Step 4: Training and optimization of the U-Net network with multi-scale fusion of boundary features.

2. The U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1: The initial training dataset A of un-noise-enhanced seismic data includes noisy seismic data and corresponding clean, noise-free seismic data. The clean, noise-free seismic data is obtained through forward modeling of the standard Marmousi model. Specifically, the forward modeling method of the acoustic wave equation is used, and Ricker wavelet is used to numerically simulate the standard Marmousi model. A clean, noise-free synthetic shot collection record, each record being [size missing]. Random selection Each shot set record is used as the basic training data, among which ,right Each shot collection record uses Sliding window segmentation, step size is , co-generated Clean, noise-free seismic data; noisy seismic data is obtained by... Noise samples were extracted from actual seismic records from different work areas and superimposed with clean samples to obtain P noisy seismic data sets. A noisy seismic data set and its corresponding clean, noise-free seismic data set were termed a training set, resulting in a total of P training sets. The training data sets constitute the earthquake data training dataset A. Step 1.2: For the seismic data training dataset A obtained in Step 1.1, a multi-scale noise enhancement strategy is adopted. Specifically, firstly, the original noisy data is enhanced by a linear transformation with a fixed amplification factor. Then, these enhanced noisy data are recombined with clean, noise-free seismic data to obtain datasets with different noise intensities. The training dataset consists of sets of data, each maintaining a strict correspondence between noisy and clean data, forming a noise-enhanced earthquake data training dataset B.

3. The U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement according to claim 1, characterized in that, In step 1.2, the magnification is set to 10x, 50x, 100x, and 200x.

4. The U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: The U-Net architecture is adopted as the basic framework. This framework includes a symmetrical encoder and decoder structure, and establishes a complete skip connection between each sampling stage of the encoder and decoder to ensure the full fusion of features at different scales. The encoder consists of four downsampling stages. Each downsampling stage includes a multi-scale dual-layer convolutional block and a convolutional downsampling layer, which extracts and compresses features through progressive processing. The multi-scale dual-layer convolutional block contains six convolutional kernels and is divided into two processing units. Each processing unit has three convolutional operations of different sizes set in parallel. The first processing unit is equipped with one 3×3 convolutional kernel, one 5×5 convolutional kernel, and one 7×7 convolutional kernel. The second processing unit is also equipped with one 3×3 convolutional kernel, one 5×5 convolutional kernel, and one 7×7 convolutional kernel. The decoder section contains four upsampling stages. Each upsampling stage includes a multi-scale two-layer convolutional block and a transposed convolutional upsampling layer. The feature map is magnified by the transposed convolutional upsampling layer. The multi-scale two-layer convolutional block here is the same as that used in the encoder. Skip connection part: The output features of each downsampling stage in the encoder are aligned with the corresponding upsampling stage features in the decoder through the skip connection layer, and then spliced ​​and fused in the channel dimension. Meanwhile, a multi-scale two-layer convolutional block is connected between the last downsampling stage of the encoder and the first upsampling stage of the decoder. After being processed by the encoder and decoder at the same time, the multi-scale features output by the main network are obtained. Step 2.2: Based on the multi-scale U-Net constructed in Step 2.1, a boundary enhancement module and a selection fusion attention mechanism are further integrated. At this point, each downsampling stage includes a multi-scale two-layer convolutional block, an attention mechanism, and a convolutional downsampling layer. Specifically, a boundary enhancement operation is added outside the multi-scale U-Net constructed in Step 2.1 to integrate the data obtained from different downsampling stages: The boundary enhancement operation extracts and fuses gradient features from the multi-scale features output by each downsampling stage of the encoder using three multi-scale two-layer convolutional blocks, and uses the Sigmoid activation function to focus on enhancing the boundary feature expression in the seismic data to obtain boundary enhancement features; The selection fusion attention mechanism is embedded after each multi-scale two-layer convolutional block of the encoder. Through the synergistic effect of selecting kernel attention and efficiently extracting attention, the network's ability to focus on effective signal features is significantly improved. Step 2.3: After fusing the multi-scale features output by the main network in Step 2.1 with the boundary enhancement features obtained in Step 2.2, the final network output, i.e. the denoised seismic data, is obtained. At this point, the U-Net network framework for multi-scale fusion and boundary feature enhancement is completed.

5. The U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement according to claim 1, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Input the noise-enhanced earthquake data training dataset B into the U-Net network for multi-scale fusion boundary feature enhancement built in Step 2 in batches; The multi-scale fusion boundary feature enhancement U-Net network built in steps 3.2 and 2 is mainly for seismic data denoising. Therefore, the loss function for network training is set to the MSE loss function, the specific expression of which is as follows: (1) in, Indicates coordinates as Denoising data predicted by the network. Indicates coordinates as The data is noise-free, where C and T are the number of traces and sampling points recorded in the synthetic shot gather obtained from the forward modeling simulation. Step 3.3: Optimize the MSE loss function set in Step 3.2 using an optimization algorithm. The optimization algorithm uses the Adam adaptive moment estimator, which achieves stable and efficient optimization of the loss function in Step 3.2 by dynamically adjusting the learning rate of each parameter.

6. The U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement according to claim 5, characterized in that, In step 3.1, the network hyperparameter initialization adopts an adaptive strategy, with the base learning rate set to 0.

001. The learning rate for each iteration is set to 90%-95% of the previous iteration's learning rate. Each iteration represents the network learning the data in the training set once. The batch size is set to 8 based on the GPU memory capacity.

7. The U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement according to claim 5, characterized in that, In step 3.1, the number of training cycles is set to 50 epochs.

8. The U-Net seismic data denoising method with multi-scale fusion boundary feature enhancement according to claim 1, characterized in that, Step 4 specifically includes the following steps: Step 4.1: Based on the hyperparameters configured in Step 3, such as learning rate, learning period, loss function, and adaptive moment estimation optimizer Adam, perform end-to-end training on the U-Net network with multi-scale fusion boundary feature enhancement. Step 4.2: After the network training in Step 4.1 is completed, the model with the lowest MSE loss function is loaded as the optimal model into the constructed multi-scale fusion boundary feature enhancement U-Net network.

Images