Seismic data reconstruction method based on VMD filtering and U-net image segmentation

Abstract

The invention discloses a seismic data reconstruction method based on VMD filtering and U-net image segmentation, and relates to the technical field of oil exploration and development. According to the method, the seismic data is decomposed through the VMD method, the difference between the effective signal mode component and the noise mode component in the time-frequency domain is analyzed, the effective signal mode component is optimized on the basis to reconstruct the seismic signal, fine suppression is carried out on local noise, the signal-to-noise ratio of the seismic data can be effectively increased, and the seismic performance of the seismic data is improved. And high-quality training and test data are provided for a subsequent U-net convolutional network model. According to the method, bad traces existing in the seismic data can be effectively extracted, and the missing part of the data can be accurately reconstructed.

Description

Technical Field

[0001] The present invention relates to the technical field of oil exploration and development, in particular to the technical field of seismic data processing, and more specifically to a seismic data reconstruction method based on VMD filtering and U-net image segmentation. Background Art

[0002] At present, due to the limitations of the acquisition environment such as surface topography and prohibited mining areas, the seismic data collected in seismic exploration will inevitably have problems such as bad tracks, missing tracks, and insufficient sampling, resulting in poor continuity of seismic data and failure to meet the requirements of subsequent seismic interpretation work for data accuracy. Therefore, bad track removal and data reconstruction are the basic steps in seismic data processing.

[0003] Traditional seismic data interpolation methods can be divided into three categories: the first category is the data reconstruction method based on seismic wave field migration and anti-migration, but the complexity of the strata has a great influence on the data reconstruction effect; the second category is the seismic data reconstruction method based on prediction filters, which is mainly suitable for regularly sampled seismic data; the third category is the data reconstruction method based on sparse transformation, which uses sparse constraints to reconstruct seismic data.

[0004] With the advancement of computer technology, artificial intelligence and machine learning technology have made great progress. Hinton and others first demonstrated the superiority of deep neural networks in image processing. Subsequently, scholars have proposed more efficient network models. Convolutional neural networks are one of the most widely used neural network models.

[0005] U-Net is a convolutional neural network proposed by Ronneberger et al. for medical image segmentation. The network can extract multi-scale features of images without manually designing feature extractors. It automatically learns data features in a data-driven way and has high generalization and robustness. Seismic data is essentially an image. By constructing and training a U-net convolutional neural network, pixel-level prediction of seismic images can be achieved, thereby reconstructing seismic data with high precision.

[0006] For example, the invention patent with the publication date of November 22, 2019, the publication number of CN110490219A, and the name of "A method for seismic data reconstruction based on U-net network with texture constraints", includes: S1, using the K-means algorithm to segment the training data set to obtain the texture label of the training data; S2, training the U-net network as a texture extractor through the training data set and texture label to obtain optimized texture extractor parameters; S3, connecting the trained texture extractor in series with the reconstruction network, using the optimized texture extractor parameters, extracting the reconstruction label and the texture information of the reconstructed seismic data, and obtaining the texture loss; S4, reconstructing the seismic data after optimizing the reconstruction network through texture loss. This invention combines the characteristic that seismic data has rich texture information, and by strengthening the learning of texture, it can improve the reconstruction accuracy and continuity of seismic event axes in the case of limited samples.

[0007] The above-mentioned prior art reconstructs seismic data based on texture information, and cannot eliminate bad tracks in the seismic data, and cannot precisely suppress local noise, which affects the accuracy of the U-net network model in reconstructing seismic data. Summary of the invention

[0008] In order to overcome the defects and deficiencies in the above-mentioned prior art, the present invention provides a seismic data reconstruction method based on VMD filtering and U-net image segmentation. The present invention decomposes seismic data through the VMD method, analyzes the difference between the effective signal modal component and the noise modal component in the time-frequency domain, and on this basis, selects the effective signal modal component to reconstruct the seismic signal, and finely suppresses the local noise, which can effectively improve the signal-to-noise ratio of seismic data and provide high-quality training and test data for the subsequent U-net convolutional network model. The present invention can effectively propose bad tracks in seismic data and accurately reconstruct the missing parts of the data.

[0009] In order to solve the above problems existing in the prior art, the present invention is implemented through the following technical solutions.

[0010] A seismic data reconstruction method based on VMD filtering and U-net image segmentation, the method comprises the following steps: S1. Acquire post-stack seismic data with bad tracks and missing tracks; S2. Perform spectrum analysis on the post-stack seismic data to be processed to clarify the amplitude and frequency characteristics of the bad track data; S3, counting the root mean square amplitude energy of each track in the seismic data, and comparing it with a set threshold value, if it is greater than the set threshold value, it is defined as a bad track, and the bad track is replaced with a blank track, and the position of the bad track is recorded; S4, using variational mode decomposition to filter the seismic data after bad tracks are removed to improve the signal-to-noise ratio of the seismic data; S5, selecting the filtered seismic data to construct a training sample set and a test sample set; S6. Build a U-net network structure model and initialize the parameters of each convolution layer; S7, training the U-net network model using the training sample set constructed by the filtered seismic data, testing the model effect using the test sample set constructed by the filtered seismic data, and converging the model parameters through multiple iterations; S8. The filtered seismic data obtained after processing in steps S1 to S4 is input into the U-net network model trained in step S7 to reconstruct the post-stack seismic data.

[0011] Furthermore, in step S2, spectral analysis is performed on the post-stack seismic data to be processed, specifically, spectral analysis is performed on the original seismic data set without bad tracks and the same seismic data set with bad tracks in the post-stack seismic data with bad tracks and missing tracks obtained in step S1, respectively, to clarify the frequency and amplitude characteristics of the bad tracks.

[0012] Furthermore, in step S3, the root mean square amplitude energy of each trace in the seismic data is calculated, and the calculation formula is: ,in, is the ith sampling point of a single seismic trace, and N is the total number of sampling points of a single seismic trace.

[0013] Furthermore, in step S3, a bad track threshold is defined according to the spectrum analysis result in step S2. When the RMS energy value of a track is greater than the threshold, the track is determined as a bad track and replaced with a blank track. The bad track position is recorded.

[0014] In step S3, the bad track threshold is defined as 1×10 20 .

[0015] Furthermore, in step S4, the seismic signal is decomposed into k modal components with limited bandwidth by using variational modal decomposition, and each modal component and its center frequency are obtained by the following formula: , In the formula, is the earthquake signal, is the modal component, is the Lagrange multiplication operator, , , are their corresponding Fourier transforms, α is the quadratic multiplication factor, represents the center frequency of the decomposed seismic signal component, Represents the center frequency of the kth seismic signal component obtained by decomposition.

[0016] Furthermore, in step S4, the number of decomposed modal components of the variational modal decomposition is determined by the instantaneous frequency mean of all modal components.

[0017] Specifically, the seismic data is decomposed, the instantaneous frequency means of all modal components are calculated, and the curvature of the instantaneous frequency mean curve is calculated. If the curvature is greater than the set curvature threshold, k-1 is the number of optimal modal components, otherwise k is increased and the decomposition is repeated until the curvature is greater than the set curvature threshold.

[0018] The curvature threshold is set to 0.5.

[0019] Furthermore, in the step S4, after determining the optimal k value, the seismic trace is decomposed to calculate the signal-to-noise ratio of all modal components. ;like , then it is a valid mode, otherwise it is a noise mode.

[0020] Furthermore, in step S4, according to the bad track position recorded in step S3, the corresponding track of the seismic data after the variational mode decomposition filtering is replaced with a blank track.

[0021] Furthermore, in step S5, for the filtered seismic data obtained in step S4, a regular area that does not contain blank traces is selected to construct a seismic sub-volume, and 33% of blank traces are randomly inserted into the seismic sub-volume to obtain data of artificially missing seismic traces and their corresponding complete seismic data.

[0022] Furthermore, in step S5, the seismic sub-volume data set with blank traces randomly inserted is divided into a training set and a test set of the U-net convolutional neural network in a ratio of 3:1.

[0023] Furthermore, in step S6, the Keras deep learning framework is used to build a U-net convolutional neural network model.

[0024] The constructed U-net convolutional neural network model includes two parts: encoding and decoding, which correspond to the upsampling and downsampling processes respectively. Each downsampling operation contains two convolutional layers with a convolution kernel size of 3×3 and a maximum pooling layer with a convolution kernel size of 2×2, and each upsampling operation contains an upsampling layer with a convolution kernel size of 2×2 and two convolutional layers with a convolution kernel size of 3×3.

[0025] Furthermore, in the step S7, the training seismic data set and the test seismic data set input into the U-net network model include Inline line direction amplitude slices, CrossLine line direction amplitude slices and time amplitude slices.

[0026] Compared with the prior art, the beneficial technical effects brought by the present invention are as follows: 1. The seismic data is decomposed by the VMD method, and the differences between the effective signal modal components and the noise modal components in the time and frequency domains are analyzed. On this basis, the effective signal modal components are selected to reconstruct the seismic signal and finely suppress the local noise, which can effectively improve the signal-to-noise ratio of the seismic data. The subsequent U-net convolutional network model provides high-quality training and test data.

[0027] 2. Use the U-net convolutional neural network to build a deep learning model, make full use of the redundancy of seismic data, extract and fuse the multi-dimensional features of seismic data through dense convolutional layers and pooling layers, automatically learn data features in a data-driven manner, realize pixel-level prediction of seismic images, and then complete high-precision reconstruction of seismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0029] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a spectrum analysis diagram of an original seismic data set without bad tracks according to an embodiment of the present invention; Figure 3 A spectrum analysis diagram of the same seismic data set with bad sectors in an embodiment of the present invention; Figure 4 This is a rendering of the original seismic data containing bad tracks in an embodiment of the present invention; Figure 5 This is a diagram showing the effect of removing bad sectors according to an embodiment of the present invention; Figure 6 is a graph showing the relationship between the number k of modal components decomposed by VMD and the mean instantaneous frequency according to an embodiment of the present invention; Figure 7 This is a diagram of the signal-to-noise ratio of seismic signals before and after VMD filtering according to an embodiment of the present invention; Figure 8 Inline direction slices in the training data set and test data set examples of the embodiment of the present invention; Fig. 9 Crossline direction slices in the training data set and test data set examples of the embodiment of the present invention; Fig.10 It is a time slice in the training data set and the test data set instance of the embodiment of the present invention; Fig.11 This is a structural diagram of the U-net convolutional neural network according to an embodiment of the present invention; Fig.12 is the U-net model training error of the embodiment of the present invention; Fig.13 The original seismic profile containing bad tracks in the embodiment of the present invention; Fig.14 A seismic profile comparing the effects of bad track removal and data reconstruction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings of the present invention specification to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] Example 1 As a preferred embodiment of the present invention, refer to the attached Figure 1 As shown, this embodiment discloses a seismic data reconstruction method based on VMD filtering and U-net image segmentation, the method comprising the following steps: S1. Acquire post-stack seismic data with bad tracks and missing tracks; S2. Perform spectrum analysis on the post-stack seismic data to be processed to clarify the amplitude and frequency characteristics of the bad track data; S3, counting the root mean square amplitude energy of each track in the seismic data, and comparing it with a set threshold value, if it is greater than the set threshold value, it is defined as a bad track, and the bad track is replaced with a blank track, and the position of the bad track is recorded; S4, using variational mode decomposition to filter the seismic data after bad tracks are removed to improve the signal-to-noise ratio of the seismic data; S5, selecting the filtered seismic data to construct a training sample set and a test sample set; S6. Build a U-net network structure model and initialize the parameters of each convolution layer; S7, training the U-net network model using the training sample set constructed by the filtered seismic data, testing the model effect using the test sample set constructed by the filtered seismic data, and converging the model parameters through multiple iterations; S8. The filtered seismic data obtained after processing in steps S1 to S4 is input into the U-net network model trained in step S7 to reconstruct the post-stack seismic data.

[0032] Example 2 As another preferred embodiment of the present invention, refer to the attached specification Figure 1, Attachment Figure 2 , Attachment Figure 3 , Attachment Figure 4 and attached Figure 5 As shown, this embodiment discloses a seismic data reconstruction method based on VMD filtering and U-net image segmentation, the method comprising the following steps: S1. Acquire post-stack seismic data with bad tracks and missing tracks; S2. Perform spectrum analysis on the post-stack seismic data to be processed to clarify the amplitude and frequency characteristics of the bad track data; Among them, the spectrum analysis of the post-stack seismic data to be processed is carried out, as shown in the attached Figure 2 and attached Figure 3 As shown, specifically, spectrum analysis is performed on the original seismic data set without bad tracks and the same seismic data set with bad tracks in the post-stack seismic data with bad tracks and missing tracks obtained in step S1, so as to clarify the frequency and amplitude characteristics of the bad tracks.

[0033] The original seismic data set without bad tracks represents the normal part of the post-stack seismic data, and the same seismic data set with bad tracks is the bad track in the post-stack seismic data.

[0034] As attached Figure 2 and attached Figure 3 It can be seen that the magnitude of the bad track amplitude energy can reach 10 37 , far exceeding the effective signal, the frequency domain of the bad track signal overlaps with the effective signal and cannot be effectively distinguished.

[0035] S3, counting the root mean square amplitude energy of each track in the seismic data, and comparing it with a set threshold value, if it is greater than the set threshold value, it is defined as a bad track, and the bad track is replaced with a blank track, and the position of the bad track is recorded; The root mean square amplitude energy of each trace in the statistical seismic data is calculated as follows: ,in, is the i-th sampling point of a single seismic trace, and N is the total number of sampling points of a single seismic trace; At the same time, according to the spectrum analysis results in step S2, the threshold value of bad tracks is defined as 1×10 20 When the RMS energy value of a track is greater than the set threshold, the track is determined as a bad track, and the bad track is replaced by a blank track, and the position of the bad track is recorded at the same time; Figure 4 and attached Figure 5 The following is a comparison of the effects before and after bad sector removal.

[0036] S4, using variational mode decomposition to filter the seismic data after bad tracks are removed to improve the signal-to-noise ratio of the seismic data; S5, selecting the filtered seismic data to construct a training sample set and a test sample set; S6. Build a U-net network structure model and initialize the parameters of each convolution layer; S7, training the U-net network model using the training sample set constructed by the filtered seismic data, testing the model effect using the test sample set constructed by the filtered seismic data, and converging the model parameters through multiple iterations; S8. The filtered seismic data obtained after processing in steps S1 to S4 is input into the U-net network model trained in step S7 to reconstruct the post-stack seismic data.

[0037] Example 3 As another preferred embodiment of the present invention, this embodiment is an elaboration of the specific implementation of step S4 in the above-mentioned embodiments 1 and 2.

[0038] As an implementation method of this embodiment, in step S4, variational mode decomposition is used to filter the seismic data after bad tracks are removed to improve the signal-to-noise ratio of the seismic data.

[0039] As an implementation method of this embodiment, VMD, full name variational mode decomposition, is a nonlinear, non-stationary signal time-frequency analysis method that can adaptively decompose seismic signals into k limited bandwidth modal components according to the time scale characteristics of the signal, and adaptively determine the center frequency and bandwidth information of each modal component. The modal components and their center frequencies are obtained using the following formula: , In the formula, is the earthquake signal, is the modal component, is the Lagrange multiplication operator, , , are their corresponding Fourier transforms, α is the quadratic multiplication factor, represents the center frequency of the decomposed seismic signal component, Represents the center frequency of the kth seismic signal component obtained by decomposition.

[0040] As another implementation of this embodiment, VMD decomposition is similar to a Togan bandpass filter, where the k value controls the number and bandwidth of the filter and is determined by the mean of the instantaneous frequencies of the components, such as Figure 6As shown in the figure, when the k value increases to a certain number, the curve has an obvious lower complete phenomenon. This critical number is the optimal k value for VMD decomposition. Set the initial k=2, decompose the seismic data, calculate the instantaneous frequency mean of all modal components, and calculate the curvature of the instantaneous frequency mean curve. If the curvature is greater than the set curvature threshold, k-1 is the optimal mode number, otherwise increase k and repeat the decomposition until the curvature is greater than the set curvature threshold. The set curvature threshold is generally set to 0.5.

[0041] As another implementation of this embodiment, the effective mode determination criterion is based on the signal-to-noise ratio of each mode. Is it greater than zero? , it means that the effective signal energy in this mode is stronger than the noise, which is an effective mode, otherwise it is a noise mode. Finally, all effective modes are superimposed to reconstruct the seismic signal. At the same time, the reconstructed signal needs to replace the corresponding position with a blank track according to the bad track position located in step S3, as shown in the attached figure. Figure 7 The figure shows the comparison of the signal-to-noise ratio of seismic signals before and after VMD denoising.

[0042] Example 4 As another preferred embodiment of the present invention, this embodiment is an elaboration of the specific implementation of step S5 in the above-mentioned embodiments 1 to 3.

[0043] In this embodiment, S5, selecting the filtered seismic data to construct a training sample set and a test sample set.

[0044] As an implementation method of this embodiment, in step S5, for the filtered seismic data obtained in step S4, a regular area without blank traces is selected to construct a seismic sub-volume, and 33% of blank traces are randomly inserted into the seismic sub-volume to obtain data of artificially missing seismic traces and their corresponding complete seismic data. Data augmentation is achieved by 3D stereo time window sliding sampling, with a window size of 16×16×32, i.e., 16 traces in the Inline direction, 16 traces in the Crossline direction, and 32ms in the time direction. The windows overlap by 50% in the Inline, Crossline and time directions, and a total of 83,664 seismic sub-volume sets are constructed, as shown in the attached figure. Figure 8 , Attachment Fig. 9 and attached Fig.10 As shown in Figure 1, the earthquake sub-volume dataset is divided into a training set for training the U-net network and a test set in a ratio of 3:1.

[0045] Example 5 As another preferred embodiment of the present invention, this embodiment is an elaboration of the specific implementation of step S6 in the above-mentioned embodiments 1 to 4.

[0046] In this embodiment, a U-net network structure model is constructed and the parameters of each convolutional layer are initialized.

[0047] As an implementation method of this embodiment, the Keras deep learning framework is selected to build a U-net convolutional neural network model. The model structure includes two parts: encoding process and decoding process.

[0048] The encoding process includes: S61, the input seismic image must first be combined twice (Conv+ReLU) to obtain a 64-channel seismic feature map; S62. The earthquake characteristic map is combined again by [Maxpool+(Conv+ReLU)×2] 4 times to obtain a new 1024-dimensional earthquake characteristic map.

[0049] The decoding process includes: The 1024-dimensional seismic feature map obtained by S63 and S62 is upsampled to a 64-dimensional seismic image feature map of the original seismic image size through four consecutive [Up-Conv+(Conv+ReLU)×2] combinations. At the same time, the feature maps generated by the encoder and decoder are concatenated and fused using a jump connection between the encoder and decoder. S64, finally, a Conv is used to generate a segmented image from the spliced ​​and fused 64-dimensional seismic feature maps; where Conv represents the convolution layer, ReLU represents the activation layer, Maxpool represents the maximum pooling layer, and Up-Conv represents the transposed convolution layer, as shown in the attached figure. Fig.11 shown.

[0050] S7. Use the training sample set constructed from the seismic data sub-body to train the U-net network model, adjust the model parameters to select the model with the highest accuracy; use the seismic data from the test sample set to test the model accuracy. If the accuracy is higher than the set threshold, the model is used as the final U-net network model. Otherwise, the parameters are reselected. Fig.12 Shown is the training error of the U-net network model.

[0051] S8, the model inputs the filtered seismic data, and uses the trained U-net network model to reconstruct the post-stack seismic data. In this operation, the input data of the trained U-net network model is the filtered seismic data obtained in S4. Fig.13 and Fig.14 Shown is a comparison chart of the effects of the final U-Net network model reconstructing the seismic data.

[0052] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the implementation methods of the present invention, and should be understood that the protection scope of the present invention is not limited to such special statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A seismic data reconstruction method based on VMD filtering and U-net image segmentation, characterized in that: The method comprises the following steps: S1. Acquire post-stack seismic data with bad tracks and missing tracks; S2. Perform spectrum analysis on the post-stack seismic data to be processed to clarify the amplitude and frequency characteristics of the bad track data; S3, counting the root mean square amplitude energy of each track in the seismic data, and comparing it with a set threshold value, if it is greater than the set threshold value, it is defined as a bad track, and the bad track is replaced with a blank track, and the position of the bad track is recorded; S4, using variational mode decomposition to filter the seismic data after bad tracks are removed to improve the signal-to-noise ratio of the seismic data; S5, selecting the filtered seismic data to construct a training sample set and a test sample set; S6. Build a U-net network structure model and initialize the parameters of each convolution layer; S7, training the U-net network model using the training sample set constructed by the filtered seismic data, testing the model effect using the test sample set constructed by the filtered seismic data, and converging the model parameters through multiple iterations; S8. The filtered seismic data obtained after processing in steps S1 to S4 is input into the U-net network model trained in step S7 to reconstruct the post-stack seismic data.

2. The seismic data reconstruction method based on VMD filtering and U-net image segmentation as claimed in claim 1, characterized in that: In step S2, spectral analysis is performed on the post-stack seismic data to be processed. Specifically, spectral analysis is performed on the original seismic data set without bad tracks and the same seismic data set with bad tracks in the post-stack seismic data with bad tracks and missing tracks obtained in step S1, so as to clarify the frequency and amplitude characteristics of the bad tracks.

3. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 1 or 2, characterized in that: In step S3, the root mean square amplitude energy of each trace in the seismic data is calculated, and the calculation formula is: ,in, is the i-th sampling point of a single seismic trace, and N is the total number of sampling points of a single seismic trace.

4. The seismic data reconstruction method based on VMD filtering and U-net image segmentation as claimed in claim 3, characterized in that: In step S3, a bad track threshold is defined according to the spectrum analysis result in step S2. When the RMS energy value of a track is greater than the threshold, the track is determined as a bad track and replaced with a blank track. The position of the bad track is recorded.

5. The seismic data reconstruction method based on VMD filtering and U-net image segmentation as claimed in claim 4, characterized in that: The bad sector threshold is defined as 1×10 20 .

6. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 1 or 2, characterized in that: In step S4, the seismic signal is decomposed into k modal components with limited bandwidth by using variational modal decomposition. Each modal component and its center frequency are obtained by the following formula: ， In the formula, is the earthquake signal, is the modal component, is the Lagrange multiplication operator, , , are their corresponding Fourier transforms, α is the quadratic multiplication factor, represents the center frequency of the decomposed seismic signal component, Represents the center frequency of the kth seismic signal component obtained by decomposition.

7. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 6, characterized in that: In step S4, the number of decomposed modal components of variational modal decomposition is determined by the instantaneous frequency mean of all modal components.

8. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 7, characterized in that: The seismic data is decomposed, and the instantaneous frequency mean of all modal components is calculated. At the same time, the curvature of the instantaneous frequency mean curve is calculated. If the curvature is greater than the set curvature threshold, k-1 is the number of optimal modal components. Otherwise, increase k and repeat the decomposition until the curvature is greater than the set curvature threshold.

9. The seismic data reconstruction method based on VMD filtering and U-net image segmentation as claimed in claim 8, characterized in that: The curvature threshold is set to 0.

5.

10. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 8, characterized in that: In step S4, after determining the optimal k value, the seismic trace is decomposed and the signal-to-noise ratio of all modal components is calculated. ;like , then it is a valid mode, otherwise it is a noise mode.

11. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 1, characterized in that: In step S4, according to the bad track position recorded in step S3, the corresponding track of the seismic data after the variational mode decomposition filtering is replaced with a blank track.

12. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 1 or 11, characterized in that: In step S5, for the filtered seismic data obtained in step S4, a regular area without blank traces is selected to construct a seismic subvolume, and 33% of blank traces are randomly inserted into the seismic subvolume to obtain data with artificially missing seismic traces and their corresponding complete seismic data.

13. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 12, characterized in that: In the step S5, the seismic sub-volume data set with blank traces randomly inserted is divided into a training set and a test set of the U-net convolutional neural network in a ratio of 3:

1.

14. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 1, characterized in that: In step S6, the Keras deep learning framework is used to build a U-net convolutional neural network model.

15. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 14, characterized in that: The constructed U-net convolutional neural network model includes two parts, encoding and decoding, corresponding to the upsampling and downsampling processes respectively; each downsampling operation includes two convolutional layers with a convolution kernel size of 3×3 and a maximum pooling layer with a convolution kernel size of 2×2, and each upsampling operation includes an upsampling layer with a convolution kernel size of 2×2 and two convolutional layers with a convolution kernel size of 3×3.

16. The seismic data reconstruction method based on VMD filtering and U-net image segmentation according to claim 1, characterized in that: In the step S7, the training seismic data set and the test seismic data set input into the U-net network model include Inline line direction amplitude slices, CrossLine line direction amplitude slices and time amplitude slices.

Images