Method for solving boundary effect in deep learning seismic data denoising and storage medium

Abstract

The present application relates to the technical field of seismic data denoising processing, in particular to a deep learning seismic data denoising boundary effect solving method and a storage medium. The deep learning seismic data denoising boundary effect solving method comprises the following steps: S1, using clean seismic data to make a training set and a test set; S2, using the training set and the test set to train and iterate to obtain a network model; S3, using the network model in step S2 to remove seismic data noise to obtain predicted data; S4, cutting the predicted data in step S3, splicing the cut data to obtain complete predicted data, and subtracting the complete predicted data from noisy seismic data to obtain denoised seismic data. The present application can perfectly solve the boundary effect problem, and there is no loss of boundary information, and there is no splicing trace.

Description

Technical Field

[0001] This invention relates to the field of seismic data denoising technology, specifically to a method for addressing boundary effects in deep learning-based seismic data denoising and a storage medium. Background Technology

[0002] Seismic data denoising is a crucial step in seismic exploration, significantly improving seismic imaging quality and resolution. In recent years, deep learning within the field of artificial intelligence has experienced rapid development both domestically and internationally. Due to its powerful feature extraction capabilities and entirely data-driven nature, it has garnered considerable attention in seismic data processing. For example, the existing technology CN114200520B discloses a seismic data denoising method that acquires noisy seismic data including random noise, surface wave noise, and linear noise. The noisy seismic data is then input into a random noise denoising model to obtain first denoised seismic data; the first denoised seismic data is input into a surface wave denoising model to obtain second denoised seismic data; and the second denoised seismic data is input into a linear noise denoising model to obtain third denoised seismic data. This method denoises noisy seismic data using random noise denoising, surface wave denoising, and linear noise denoising models.

[0003] In addition, seismic data denoising technology based on deep learning neural networks has also developed and been applied rapidly, achieving certain results. For example, the prior art CN114114421B discloses a deep learning-based guided self-learning seismic data denoising method and apparatus. The method includes: acquiring guided seismic data and raw seismic data; the guided seismic data is noise-free seismic data synthesized based on forward modeling, used as the guided learning target data; using the raw seismic data and the guided seismic data as training inputs to train a guided self-learning seismic data denoising model until the loss function value and guidance evaluation function of the guided self-learning seismic data denoising model meet preset requirements; using the trained model when the loss function value and guidance evaluation function meet the preset requirements as the seismic data denoising model, and performing denoising processing on the seismic data to be denoised.

[0004] In deep learning-based denoising methods for seismic data, a common problem encountered is the boundary effect. When a trained network is tested for denoising, noticeable stitching artifacts appear when the generated small data patches are combined into a complete dataset. The root cause of the boundary effect lies in the definition of convolution operations. Convolution operations obtain the output feature map by performing multiplication and accumulation operations on the input data through a sliding window. When the convolution kernel slides to the boundary, the window may exceed the image's boundaries, preventing multiplication and accumulation operations and thus producing boundary effects. Padding strategies are among the most common methods for addressing boundary effects. Zero padding fills the boundary with zero pixels, ensuring the convolution kernel's sliding window doesn't exceed the image's boundaries. This method is simple and easy to implement but introduces some noise. Copy padding copies the boundary pixels around the boundary. This method preserves the features at the boundary but may introduce redundant information. Summary of the Invention

[0005] To address the aforementioned technical problems in existing technologies, this invention provides a solution and storage medium for boundary effect in deep learning seismic data denoising. By filling the input complete data with zero values ​​of a specific size in the width and height directions, and sliding and cutting it into small blocks at a certain step size, the data is input into a trained denoising neural network for processing, and then spliced ​​after cropping, the boundary effect problem can be solved.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides a solution for boundary effect denoising in deep learning seismic data, comprising:

[0008] S1. Use clean seismic data to create training and test sets;

[0009] S2. Using the training set and test set, train iteratively to obtain the network model;

[0010] S3. Use the network model from step S2 to remove noise from the seismic data and obtain the predicted data;

[0011] S4. Cut the predicted data in step S3, and stitch the cut data together to obtain the complete predicted data. Subtract the complete predicted data from the noisy earthquake data to obtain the denoised earthquake data.

[0012] Furthermore, in step S1, training and test sets are created using clean seismic data. Specifically, this includes: adding Gaussian noise to the clean seismic data and cropping it into small data blocks of a set size as a feature dataset; cropping the clean seismic data into corresponding small data blocks of a set size as a label dataset; and combining the feature dataset and the label dataset, and dividing them into training and test sets.

[0013] Furthermore, the size is set in pixels, to 96×96.

[0014] Furthermore, the training set accounts for 90% and the test set accounts for 10%.

[0015] Furthermore, both the feature dataset and the label dataset are stored in .npy format.

[0016] Furthermore, the training iterations in step S2 use the DNCNN network structure.

[0017] Furthermore, the feature map size in the DNCNN network structure is set to 96×96.

[0018] Further, in step S3, the network model from step S2 is used to remove noise from the seismic data to obtain the predicted data. The specific method is as follows: the noisy seismic data is filled with zero values ​​on the height side and the width side, and the filled noisy seismic data is marked as the first filled noisy seismic data; then the first filled noisy seismic data is filled with zero values ​​on the other side of the height and the other side of the width, so that the height and width of the first filled noisy seismic data after being filled again are divisible by 96, and the data after being filled again is marked as the second filled noisy seismic data.

[0019] The second noisy seismic data is divided into 96×96 blocks, and the blocks are input into the network model in step S2 to obtain the prediction data.

[0020] Furthermore, the noisy seismic data is padded with zero values ​​on both the height and width sides, and the padded noisy seismic data is marked as the first padded noisy seismic data; the specific method is as follows:

[0021] Extend the noisy seismic data of size h×w by 16 sample points above and to the left, and the first noisy seismic data will be (h+16)×(w+16).

[0022] Furthermore, in step S4, the predicted data is segmented, specifically by dividing the 96×96 small data block in step S3 into 64×64 small data blocks around its perimeter.

[0023] Secondly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned solution for boundary effect in deep learning seismic data denoising.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] This invention provides a solution to the boundary effect problem in deep learning-based seismic data denoising. The solution involves padding the input complete data with zero values ​​of a specific size along one side (width and height), then sliding and cutting it into small blocks with a certain step size. These blocks are then fed into a trained denoising neural network for processing, followed by cropping and stitching to resolve the boundary effect issue. This method perfectly solves the boundary effect problem without losing boundary information. Compared to pure zero-value padding, it does not introduce noise and leaves no stitching traces, providing technical support for the production application of deep learning-based seismic data denoising technology. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a structural diagram of the DNCNN neural network.

[0028] Figure 3 This is a schematic diagram of the test data before and after adding noise.

[0029] Figure 4 A diagram showing the filling of zero values.

[0030] Figure 5 This is a diagram illustrating the cutting of the predicted data around its perimeter.

[0031] Figure 6 This is a schematic diagram of the predicted data segmentation in this embodiment.

[0032] Figure 7 This refers to the noisy data in this embodiment.

[0033] Figure 8 This refers to the predicted noise in this embodiment.

[0034] Figure 9 This is the denoised data in this embodiment.

[0035] Figure 10 An example diagram illustrating how the prediction data is segmented using the pure zero-value filling method.

[0036] Figure 11 The predicted noise is handled by the pure zero-value filling method.

[0037] Figure 12 This refers to the denoised data processed using the pure zero-value filling method. Detailed Implementation

[0038] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0039] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0040] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0041] Firstly, this invention provides a solution for boundary effect denoising in deep learning seismic data, such as... Figure 1 As shown, it includes:

[0042] S1. Create training and test sets using clean seismic data; specifically including:

[0043] Clean seismic data was modified by adding Gaussian noise and cropping it into 96×96 blocks to form the feature dataset. Similarly, clean seismic data was cropped into corresponding 96×96 blocks to form the label dataset. The feature and label datasets were combined, both stored in .npy format, and divided into training and testing sets. The training set comprised 90% of the dataset, and the testing set comprised 10%.

[0044] S2. Using the training and test sets, iteratively train the network model to obtain the network model; the training iteration uses the DNCNN network structure, as shown in the diagram. Figure 2 As shown, specific details will not be repeated. The feature map size in the DNCNN network structure is set to 96×96. The training set and test set from step S1 are input into the neural network with a feature map size of 96*96, and the optimal network model is obtained through iterative training.

[0045] S3. Use the network model from step S2 to remove noise from the seismic data and obtain the predicted data; the specific method is as follows:

[0046] The noisy seismic data of size h×w is filled with zero values ​​on the height side and the width side, and extended by 16 samples on the top and left sides. The filled noisy seismic data is marked as the first filled noisy seismic data, which is (h+16)×(w+16).

[0047] To meet the requirement of cutting data into small blocks (96×96), the lower right boundary of the data also needs to be filled with zero values. Therefore, the first-filled noisy seismic data is filled with zero values ​​on the other side of the height and the other side of the width, so that the height and width of the first-filled noisy seismic data after refilling are divisible by 96. The data after refilling is marked as the second-filled noisy seismic data; the second-filled noisy seismic data is (h+16+m)×(w+16+n), where m and n must satisfy (h+16+m) and (w+16+n) being divisible by 96. The zero-value filling diagram is shown below. Figure 4 As shown.

[0048] The second noisy seismic data (h+16+m)×(w+16+n) is slidably cut into small blocks of 96×96 with a width step size of 64 and a height step size of 64. The small blocks of data are then input into the network model in step S2 to obtain the prediction data.

[0049] S4. Divide the 96×96 predicted data from step S3 into smaller 64×64 blocks around its perimeter, such as... Figure 5 As shown, the 64×64 small data blocks after being cut are sequentially spliced ​​to obtain the complete h×w size prediction data (i.e. prediction noise). The h×w size noisy seismic data is subtracted from the h×w size complete prediction data to obtain the h×w size denoised seismic data.

[0050] Secondly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned solution for boundary effect in deep learning seismic data denoising.

[0051] Example

[0052] The denoising process using the boundary effect solution for deep learning seismic data denoising provided by this invention is as follows:

[0053] Step 1: Method for creating training and test sets:

[0054] Seismic shot gather data were obtained by simulation using forward modeling software. Trace head information was removed, and Gaussian noise of different amplitudes was randomly added to the clean seismic data. The data was then cropped into small blocks of 96×96 (number of traces × number of sampling points) as the feature dataset. The clean seismic data were cropped into small blocks of 96×96 as the label dataset. The data was stored in npy format, totaling 43,710 blocks, of which 39,339 blocks were the training set and 4,371 blocks were the test set.

[0055] Step 2: Train the network to obtain the optimal network model:

[0056] This example uses the DNCNN network, but it can also be used in other denoising networks such as U-NET and GA in other embodiments. The feature map size in DNCNN is set to 96×96.

[0057] Input the training set and test set from step 1 into the constructed DNCNN network, and obtain the optimal network model through iterative training.

[0058] Step 3: Denoising the network model:

[0059] Noise in seismic data was removed using a saved optimal network model. The test data consisted of pre-stack depth migration data (with trace head information removed) processed from shot gather records. To ensure test effectiveness, the data was segmented to obtain seismic data of size 1201×601 (trace number × sampling point), and Gaussian random noise of a certain amplitude was added as noisy seismic data. Figure 3 As shown.

[0060] To meet the sliding cutting conditions, the left and top sides of the seismic data are first filled with 0 values, and the height and width sides are extended by 16 sample points. At this time, the input noisy seismic data becomes 1217×617.

[0061] The neural network input size is 96×96. To meet the integer cutting condition, the right and bottom of the noisy seismic data (1217×617) need to be padded with 0 values. (1217+m) and (617+n) must be divisible by 96. We calculate m=31 and n=55. At this time, the size of the noisy seismic data is 1248×672. We slide and cut the noisy seismic data into small blocks of 96×96 with a width step of 64 and a height step of 64, obtaining a total of 209 data blocks of 96×96. We input the small blocks of data into the optimal network model in step 2 to obtain the prediction data.

[0062] Step 4: Obtain the complete denoised data:

[0063] The 96×96 small predicted data block in step 3 is divided into 64×64 small data blocks around its perimeter, as follows: Figure 6 As shown, sequentially concatenating 64×64 small data blocks yields the complete 1201×601 prediction data (i.e., the predicted noise), as shown. Figure 8 As shown, subtracting the predicted data from the noisy seismic data yields the original 1201×601 denoised seismic data without any stitching artifacts. Figure 7 The data shown is noisy. Figure 9 This is for denoised data.

[0064] Pure zero-value filling exhibits a noticeable boundary effect after splicing, such as... Figures 10 to 12As shown, this invention effectively solves the boundary effect problem in deep learning seismic data denoising.

[0065] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for addressing boundary effects in seismic data denoising using deep learning, characterized in that, include: S1. Use clean seismic data to create training and test sets; S2. Using the training set and test set, train iteratively to obtain the network model; S3. Use the network model from step S2 to remove noise from the seismic data and obtain the predicted data; S4. Cut the predicted data in step S3, and stitch the cut data together to obtain the complete predicted data. Subtract the complete predicted data from the noisy earthquake data to obtain the denoised earthquake data.

2. The method for solving boundary effect denoising in deep learning seismic data according to claim 1, characterized in that, Step S1 involves creating training and testing sets using clean seismic data. Specifically, this includes: adding Gaussian noise to the clean seismic data and cropping it into small data blocks of a set size to serve as a feature dataset; cropping the clean seismic data into corresponding small data blocks of a set size to serve as a label dataset; and combining the feature dataset and the label dataset to divide them into training and testing sets.

3. The method for solving boundary effect in deep learning seismic data denoising according to claim 2, characterized in that, Set the size in pixels, to 96×96.

4. The method for solving boundary effect in deep learning seismic data denoising according to claim 2, characterized in that, The training set accounts for 90% and the test set accounts for 10%.

5. The method for solving boundary effect denoising in deep learning seismic data according to claim 1, characterized in that, The training iteration in step S2 uses the DNCNN network structure.

6. The method for solving boundary effect in deep learning seismic data denoising according to claim 5, characterized in that, The feature map size in the DNCNN network structure is set to 96×96.

7. The method for solving boundary effect denoising in deep learning seismic data according to claim 1, characterized in that, In step S3, the network model from step S2 is used to remove noise from the seismic data to obtain the predicted data. The specific method is as follows: the noisy seismic data is filled with zero values ​​on the height side and the width side, and the filled noisy seismic data is marked as the first filled noisy seismic data; then the first filled noisy seismic data is filled with zero values ​​on the other side of the height and the other side of the width, so that the height and width of the first filled noisy seismic data can be divided by 96 after being filled again, and the data after being filled again is marked as the second filled noisy seismic data. The second noisy seismic data is divided into 96×96 blocks, and the blocks are input into the network model in step S2 to obtain the prediction data.

8. The method for solving boundary effect in deep learning seismic data denoising according to claim 7, characterized in that, The noisy seismic data is padded with zero values ​​on both the height and width sides, and the padded noisy seismic data is marked as the first padded noisy seismic data; the specific method is as follows: Extend the noisy seismic data of size h×w by 16 sample points above and to the left, and the first noisy seismic data will be (h+16)×(w+16).

9. The method for solving boundary effect in deep learning seismic data denoising according to claim 7, characterized in that, In step S4, the predicted data is cut into smaller blocks. Specifically, the 96×96 data block in step S3 is cut into 64×64 data blocks around its perimeter.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the boundary effect solution for deep learning seismic data denoising as described in any one of claims 1-9.

Images