Deep learning based ghost wave suppression in imaging domain

Abstract

The present application relates to the technical field of offshore seismic data ghost wave suppression processing, and more particularly to an imaging domain ghost wave suppression method based on deep learning. It comprises the following steps: S1. constructing a seismic training sample set, artificially synthesizing a fault velocity model, and synthesizing a ghost wave training sample set based on the fault velocity model using a reverse time migration (RTM) algorithm; S2. constructing a U-net neural network; S3. training the U-net neural network using the synthesized ghost wave training sample set; and S4. inputting target work area seismic data into the U-Net neural network for intelligent ghost wave suppression processing to obtain ghost wave suppressed seismic data. The present application can more effectively and stably remove or suppress ghost waves and reduce their interference with effective reflection signals.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine seismic data ghost wave suppression processing, and more particularly to an imaging domain ghost wave suppression method based on deep learning. BACKGROUND

[0002] In the process of wide frequency marine towed streamer seismic data acquisition, in order to avoid the influence of complex sea surface caused by noise, the source and receiver are usually placed below a certain depth of the sea surface. Due to the huge wave impedance difference between sea water and air, a strong reflection interface (reflection coefficient is about -1) is formed between them, which constitutes a free surface. This makes the receiver below the sea surface receive not only the reflection wave from the underground interface, but also the free surface reflection wave at the source end and the receiver end. Because the depth of the source and receiver in the same work area is fixed, in the seismic shot record or processed seismic image, this free surface reflection wave will usually follow the underground interface reflection wave, and therefore is called "ghost wave" figuratively. The ghost wave and the effective reflection wave will interfere with each other, resulting in a longer continuation time of the effective reflection wave, which greatly reduces the vertical resolution of the seismic image. In addition, the ghost wave has a wave-trapping effect, which will make the spectrum of the seismic signal periodically appear zero points, resulting in the loss of some specific frequency components (especially high and low frequency information), thereby narrowing the effective frequency band of the seismic data and bringing great difficulties to subsequent seismic inversion and data interpretation. Therefore, removing or suppressing ghost wave is an important step in marine seismic data processing.

[0003] The cost of seismic data acquisition is relatively high, and in the actual exploration process, the cost needs to be controlled, and repeated trial and error or repeated acquisition is not possible, so it is necessary to develop corresponding processing technology to maximize the accuracy of existing seismic data. Due to the wave-trapping effect of the ghost wave and the complexity of the theoretical derivation, the conventional ghost wave suppression technology has problems of numerical instability and low calculation efficiency in the calculation process. In addition, the actual exploration environment is complex and diverse, and the traditional method has certain limitations. In order to overcome these problems, new methods and technologies need to be continuously researched and developed to better cope with the challenges in marine exploration. SUMMARY

[0004] The purpose of the present application is to overcome the deficiencies in the prior art in effectively and stably suppressing ghost waves, and to provide an imaging domain ghost wave suppression method based on deep learning, which can more effectively and stably remove or suppress ghost waves and reduce their interference with effective reflection signals.

[0005] To solve the above technical problems, the technical solution adopted by the present application is:

[0006] Provided is an imaging domain ghost wave suppression method based on deep learning, comprising the following steps:

[0007] S1. Construct a seismic training sample set, artificially synthesize a fault velocity model, and synthesize a ghost wave training sample set based on the fault velocity model using a reverse time migration (RTM) algorithm;

[0008] S2. Construct a U-net neural network;

[0009] S3. Train the U-net neural network using the synthesized ghost wave training sample set;

[0010] S4. Input target work area seismic data into the U-Net neural network for intelligent ghost wave suppression processing to obtain ghost wave suppressed seismic data.

[0011] According to the above technical means, the imaging domain ghost wave suppression method based on deep learning provided by the application finds and extracts the nonlinear mapping relationship between the seismic images containing ghost waves and the seismic images not containing ghost waves through a deep neural network, so as to remove or suppress the ghost waves in the imaging domain, which can more effectively and stably remove or suppress the ghost waves and reduce the interference of the ghost waves on effective reflection signals.

[0012] Preferably, the step S1 comprises:

[0013] S11. Generating a horizontal layered velocity model V1(X,Z): a one-dimensional velocity sequence V(Z) is randomly generated, wherein the number of strata, thickness, and velocity are randomly selected from a reasonable range determined according to the actual work area geological conditions, and the one-dimensional velocity sequence V(Z) is extended along the horizontal direction to a two-dimensional horizontal layered velocity model V1(X,Z);

[0014] S12. Generating a tilted fold model On the basis of the horizontal layered velocity model V1(X,Z), a vertical displacement field M(X,Z) is used to add stratum dip angle and fold structure to the horizontal layered velocity model V1(X,Z), to obtain a tilted fold model

[0015]

[0016] S13. Generating a real fault velocity model A fault displacement field is used to add a fault structure to the tilted fold model to obtain a real fault velocity model containing stratum dip angle, fold, and fault structure

[0017] S14. Generating an RTM image: an RTM algorithm is used to convert the real fault velocity model into a seismic image to obtain seismic images with and without ghost waves;

[0018] S15. Generating a seismic training sample set.

[0019] Preferably, step S12 includes: the vertical displacement field M(X,Z) consists of two parts, M1(X,Z) and M2(X,Z).

[0020] M(X,Z)=M1(X,Z)+M2(X,Z)(1)

[0021] In formula (1), M1(X,Z) is a linear equation used to simulate the dip angle of the formation, and its expression is:

[0022] M1(X,Z)=tanα·(XX c (2)

[0023] Where α represents the dip angle of the formation; V1(X c Z) represents the middle path of the velocity model;

[0024] M2(X,Z) is composed of N one-dimensional Gaussian functions and is used to simulate underground fold structures. Its expression is as follows:

[0025]

[0026] Among them, Z max b represents the maximum depth of the velocity model; k c k and σ k These represent the amplitude, center position, and half-length of the k-th Gaussian function, respectively, and their magnitudes are all randomly generated. Multiple different fold structures are added to the velocity model by superimposing and summing N different Gaussian functions.

[0027] Using the aforementioned vertical displacement field M(X,Z), we can add stratigraphic dip angles and fold structures to the horizontal layered velocity model V1(X,Z) to generate the tilted fold model V2(X,Z):

[0028]

[0029] Preferably, step S13 includes:

[0030] S131. Determine the fault reference line: Use the position of the center point of the reference line. A fault reference line is defined with the fault dip angle θ; where the center point position P and the fault dip angle θ are both randomly generated within a set range;

[0031] S132. Coordinate system transformation: using the randomly determined center point of the reference line. And the fault dip angle θ will be used to establish the global coordinate system Transform to the local coordinate system (x,z):

[0032]

[0033] where R is a rotation matrix:

[0034]

[0035] After coordinate transformation, P is the origin of the local coordinate system; the positive direction of the x-axis points to the dip direction; the z-axis is the normal of the reference line, and its positive direction points to the hanging wall of the fault; the fault reference line is the z-axis of the local coordinate system;

[0036] S133. Generating a fault curve: a series of perturbation points are randomly generated near the fault reference line z = 0, and then the fault curve z = f(x) is constructed by interpolation using these perturbation points; f

[0037] S134. Calculating the fault displacement field D(x,z): first, the dip displacement field Dx(x,z) in two-dimensional space is calculated, and then the normal displacement field Dz(x,z) is calculated; the fault displacement field D(x,z) is a two-dimensional vector field defined in the local coordinate system (x,z), which consists of the dip displacement field Dx(x,z) and the normal displacement field Dz(x,z):

[0038] D(x,z) = [D x (x,z)u x ,D z (x,z)u z ](7)

[0039] where u x and u z represent the unit vectors in the x and z directions, respectively;

[0040] S135. Simulating fault structures using the fault displacement field: according to the calculated fault displacement field D(x,z), the tilted fold model is added with fault structures to obtain the fault model

[0041]

[0042] where T represents the transpose symbol.

[0043] Preferably, the step S134 comprises:

[0044] First, the projection displacement field d(x; z = 0) of the dip displacement field Dx(x,z) on the reference line is calculated:

[0045]

[0046] where d max represents the maximum displacement of the center point, i.e., the origin of the local coordinate system; r(x) is the normalized radius: ​

[0047]

[0048] wherein, l x represents the half-length of the fault line;

[0049] Then, the fault displacement of each point in the two-dimensional space is calculated by using the projection displacement field d(x; z=0) on the fault reference line; the boundary γ is defined so that Dx(x,z) is 0 when it reaches or exceeds the boundary γ;

[0050] For the hanging wall f(x)≤z≤f(x)+γ, the calculation expression of Dx(x,z) is:

[0051] D x (x,z) = λ · d(x; z = 0) · ζ(x,z) (11)

[0052] wherein, λ represents the percentage of the hanging wall displacement amount in the projection displacement field d(x; z=0), which is randomly selected between [0, 1]; ζ(x,z) is a nonlinear attenuation function, which is used to attenuate the dip displacement field Dx(x,z) along the normal direction away from the fault curve, and the attenuation is 0 when the traction radius γ is reached; the nonlinear attenuation function ζ(x,z) is expressed as:

[0053]

[0054] For the footwall f(x)-γ≤z≤f(x), the calculation expression of Dx(x,z) is:

[0055] D x (x,z) = (λ-1) · d(x; z = 0) · ζ(x,z) (13)

[0056] After the dip displacement field Dx(x,z) in the two-dimensional space is calculated, the normal displacement field Dz(x,z) is directly obtained by using the following formula:

[0057] D z (x,z) = f(x+D x (x,z))-f(x) (14).

[0058] Preferably, by using steps S131-S135, a fault can be simulated in the inclined fold model V2(X,Z); steps S131-S135 are repeated to generate a fault velocity model containing multiple faults in the same velocity model

[0059]

[0060] Preferably, step S14 includes: first, defining a marine seismic data observation system; randomly selecting one from conventional horizontal towed cable, top-and-bottom cable, variable-depth cable, and deep-towed types as the seismic data acquisition method, with the sinking depth of the seismic source and detector randomly selected within a set range; then, using boundary conditions with and without free surfaces respectively, for... RTM imaging was performed to obtain seismic images with and without ghost waves.

[0061] Preferably, step S15 includes: randomly selecting an image of size 128×128 pixels at the same location from the calculated RTM images with and without ghost waves, and using the image with ghost waves as the earthquake training sample and the image without ghost waves as the label image; repeatedly randomly selecting multiple 128×128 training samples and labels from the two RTM images to obtain a small earthquake training sample set.

[0062] Preferably, in step S2, the data processing process of the U-net neural network includes: after the seismic sample with ghost waves is input into the U-net, it first undergoes a series of downsampling steps in the encoder path to help the U-net capture hierarchical features; wherein, the seismic image with an input size of 128×128 is gradually reduced to 64×64, 32×32, 16×16, 8×8 and 4×4, and the number of channels gradually increases to 128, 256, 512, 1024, 2048 and 4096; subsequently, a series of upsampling processes are performed in the decoder path to restore the spatial size; at the same time, skip connections combine high-level semantic features in the decoder with low-level details in the encoder, enabling the U-net neural network to effectively aggregate coarse global contextual information and detailed local features; finally, a 1×1 convolutional layer outputs a suppressed ghost wave image of the same size as the input.

[0063] Then, the mean squared error, i.e., the L2 norm, is used as the loss function:

[0064]

[0065] Where n represents the total number of pixels in the input seismic image; p i This represents the suppressed ghost wave image output by the U-net neural network; i This represents a labeled image that does not contain ghost waves.

[0066] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] The imaging domain ghost wave suppression method based on deep learning provided by the present application has the advantage of learning complex features and patterns from a large amount of training data by using a deep learning model, and thus can be more advantageous in dealing with complex situations. Moreover, the present application has strong flexibility and can establish a training sample set containing different ghost wave characteristics for different work areas and exploration environments to adapt to actual complex geological conditions. In addition, the present application can automatically perform ghost wave suppression without the need for ghost wave delay time estimation, reducing the need for human intervention and improving efficiency. In particular, the present application has the potential to be directly applied to three-dimensional situations, while the prior art is limited due to the sparse Crossline acquisition system. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 FIG. 1 is a flowchart of the imaging domain ghost wave suppression method based on deep learning of the present application;

[0070] Figure 2 FIG. 2 is a flowchart of constructing a seismic training sample set in the imaging domain ghost wave suppression method based on deep learning of the present application. DETAILED DESCRIPTION

[0071] The present application will be further described below in conjunction with specific embodiments.

[0072] Embodiment One

[0073] This embodiment is the first embodiment of the imaging domain ghost wave suppression method based on deep learning, which includes the following steps:

[0074] S1. Construct a seismic training sample set, artificially synthesize a fault velocity model, and synthesize a ghost wave training sample set based on the fault velocity model using the reverse time migration (RTM) algorithm;

[0075] S2. Construct a U-net neural network;

[0076] S3. Train the U-net neural network using the synthesized ghost wave training sample set;

[0077] S4. Input the target work area seismic data into the U-Net neural network for intelligent ghost wave suppression processing to obtain ghost wave suppressed seismic data.

[0078] The step S1 includes:

[0079] S11. Generate a horizontal layered velocity model V1(X,Z): randomly generate a one-dimensional velocity sequence V(Z), wherein the number of layers, thickness, and velocity are randomly selected from a reasonable range determined by the actual work area geological conditions, and the one-dimensional velocity sequence V(Z) is expanded along the horizontal direction to a two-dimensional horizontal layered velocity model V1(X,Z);

[0080] S12. Generating the tilted-fold model V2(X,Z): on the basis of the horizontal layered velocity model V1(X,Z), the stratum dip angle and fold structure are added to it by using the vertical displacement field M(X,Z), to obtain the tilted-fold model

[0081]

[0082] In the step S12, the vertical displacement field M(X,Z) is composed of two parts M1(X,Z) and M2(X,Z):

[0083] M(X,Z) = M1(X,Z) + M2(X,Z) (1)

[0084] In the formula (1), M1(X,Z) is a linear equation, which is used to simulate the stratum dip angle, and its expression is as follows:

[0085] M1(X,Z) = tan a · (X - X c )(2)

[0086] Wherein, a represents the stratum dip angle; V1(X c ,Z) represents the middle trace of the velocity model;

[0087] M2(X,Z) is composed of N one-dimensional Gaussian functions, which is used to simulate the underground fold structure, and its expression is as follows:

[0088]

[0089] Wherein, Z max represents the maximum depth of the velocity model; b k , c k and σ k respectively represent the amplitude value, center position and half length of the kth Gaussian function, and their sizes are generated randomly; by superimposing and summing N different Gaussian functions, multiple different fold structures are added in the velocity model;

[0090] By using the above vertical displacement field M(X,Z), the stratum dip angle and fold structure can be added to the horizontal layered velocity model V1(X,Z), to generate the tilted-fold model

[0091]

[0092] S13. Generating the real fault velocity model By using the fault displacement field, the fault structure is added to the tilted-fold model , to obtain the real fault velocity model containing the stratum dip angle, fold and fault structure The step S13 includes:

[0093] S131. Determine the fault reference line: utilize the reference line center point position and the fault dip angle θ to define a fault reference line; wherein the center point position P and the fault dip angle θ are randomly generated within a set range;

[0094] S132. Coordinate system conversion: utilize the randomly determined reference line center point position and the fault dip angle θ to convert the global coordinate system to the local coordinate system (x, z):

[0095]

[0096] wherein R is the rotation matrix:

[0097]

[0098] After the coordinate system conversion, the P point is the origin of the local coordinate system; the positive direction of the x-axis points to the dip direction; the z-axis is the normal line of the reference line, and the positive direction points to the hanging wall part of the fault; the fault reference line is the z-axis of the local coordinate system;

[0099] S133. Generate the fault curve: randomly generate a series of perturbation points near the fault reference line z = 0, and then use these perturbation points to interpolate to construct the fault curve z f = f(x);

[0100] S134. Calculate the fault displacement field D(x, z): the fault displacement field D(x, z) is a two-dimensional vector field defined in the local coordinate system (x, z), which consists of two parts: the dip direction displacement field Dx(x, z) and the normal direction displacement field Dz(x, z):

[0101] D(x, z) = [D x (x, z)u x , D z (x, z)u z ](7)

[0102] wherein u x and u z represent the unit vectors in the x and z directions, respectively;

[0103] Since this embodiment only considers the seamless sliding of the upper and lower blocks along the fault curve, the normal direction displacement field Dz(x, z) can be calculated according to the dip direction displacement field Dx(x, z) and the fault curve z f = f(x). Therefore, in order to facilitate the calculation of the two-dimensional fault displacement field D(x, z), the projection displacement field d(x; z = 0) of the dip direction displacement field Dx(x, z) on the reference line is calculated first:

[0104]

[0105] where d max represents the maximum displacement of the center point, i.e., the origin of the local coordinate system; r(x) is the normalized radius:

[0106]

[0107] where l x represents the half length of the fault line;

[0108] Then, the fault displacement of each point in the two-dimensional space is calculated using the projection displacement field d(x; z = 0) on the fault reference line; since the actual fault displacement is maximum at the fault core (at the fault curve) and is 0 at a certain distance (traction radius) away from the core, a boundary γ is defined such that Dx(x, z) is 0 when it reaches or exceeds the boundary γ;

[0109] For the hanging wall f(x)≤z≤f(x)+γ, the calculation expression of Dx(x, z) is:

[0110] D x (x, z) = λ · d(x; z = 0) · ζ(x, z) (11)

[0111] where λ represents the percentage of the hanging wall displacement to the projection displacement field d(x; z = 0), which is randomly selected between [0, 1]; ζ(x, z) is a nonlinear decay function, which decays the dip displacement field Dx(x, z) along the normal direction away from the fault curve, and is 0 when the traction radius γ is reached; the nonlinear decay function ζ(x, z) is expressed as:

[0112]

[0113] For the footwall f(x)-γ≤z≤f(x), the calculation expression of Dx(x, z) is:

[0114] D x (x, z) = (λ-1) · d(x; z = 0) · ζ(x, z) (13)

[0115] After the dip displacement field Dx(x, z) in the two-dimensional space is calculated, the normal displacement field Dz(x, z) is directly obtained using the following formula:

[0116] D z (x, z) = f(x+D x (x, z))-f(x) (14);

[0117] S135. Simulating fault structures using the fault displacement field: according to the calculated fault displacement field D(x, z), the tilted fold model Add a fault structure to obtain a fault model

[0118]

[0119] In the formula, T represents a transposition symbol.

[0120] In the present embodiment, by using steps S131-S135, a fault can be simulated in the inclined fold model ; by repeating steps S131-S135 in the same velocity model, a fault velocity model

[0121] S14. Generating an RTM image: converting the real fault velocity model into a seismic image using an RTM algorithm to obtain a seismic image with and without a ghost wave; first, define a marine seismic data observation system; randomly select one from among conventional horizontal streamer, up-down streamer, variable-depth streamer, and deep-towed streamer as the seismic data acquisition method, and randomly select the source and receiver sinking depths within the set range; then, perform RTM imaging on V3(X,Z) using boundary conditions with and without a free surface to obtain a seismic image with and without a ghost wave.

[0122] S15. Generating a seismic training sample set: randomly selecting images of the same size of 128x128 pixels from the calculated RTM images with and without a ghost wave, and taking the image with a ghost wave as a seismic training sample and the image without a ghost wave as a label image; repeating the random selection of multiple 128x128 training samples and labels in the two RTM images to obtain a small seismic training sample set.

[0123] Repeating the above steps S11-S15 can construct a diversified seismic training sample set containing rich structural information and different ghost wave characteristics for neural network training.

[0124] Compared with the prior art, the ghost wave suppression method based on a deep learning algorithm provided in the present embodiment is a data-driven method, which uses a deep learning model to learn complex features and patterns from a large amount of training data, and thus can have an advantage in dealing with complex situations. Moreover, the present application has strong flexibility and can establish a training sample set containing different ghost wave characteristics for different work areas and exploration environments to adapt to actual complex geological conditions. In addition, the present application can automatically suppress ghost waves without the need for ghost delay time estimation, reducing the need for manual intervention and improving efficiency. In particular, the present application has the potential to be directly applied to three-dimensional situations, while the prior art is limited by the sparse Crossline acquisition system.

[0125] Example Two

[0126] This embodiment is a second embodiment of the ghost wave suppression method based on deep learning algorithm, which is similar to example one, the difference is that in this embodiment, the processing process of U-net neural network is provided; in the neural network training stage, a U-net neural network is first constructed. Compared with the existing U-net, the U-net used in this embodiment contains more convolution layers, more feature maps output by each layer, and more down-sampling and up-sampling layers. After the seismic sample with ghost wave is input into the U-net, it first experiences a series of down-sampling steps in the encoder path to help the U-net capture hierarchical features. Among them, the seismic image with an input size of 128x128 is gradually reduced to 64x64, 32x32, 16x16, 8x8 and 4x4, and the number of channels is gradually increased to 128, 256, 512, 1024, 2048 and 4096. Then, a series of up-sampling processes are performed in the decoder path to restore the spatial size. At the same time, the skip connection combines the high-level semantic features in the decoder with the low-level details in the encoder, so that the U-net neural network can effectively aggregate rough global context information and detailed local features. Finally, a 1x1 convolution layer outputs a suppressed ghost wave image with the same size as the input.

[0127] Then the mean squared error (MSE), i.e. L2 norm, is used as the loss function:

[0128]

[0129] Where n represents the total number of pixel points in the input seismic image; p i represents the suppressed ghost wave image output by the U-net neural network; l i represents the label image without ghost wave.

[0130] The MSE loss function is optimized by using the Adam algorithm, where the initial learning rate is 10 -5 , the batch size is set to 10, and the training round is set to 100.

[0131] Example Three

[0132] This embodiment provides a computer device, which includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method of example one or example two when executing the computer program.

[0133] In the specific contents of the foregoing specific embodiments, each technical feature can be combined arbitrarily without contradiction. In order to make the description simple, all possible combinations of the foregoing technical features are not described, but as long as the combinations of the technical features do not contradict, they should be considered as the scope of the present disclosure.

[0134] Obviously, the above embodiments of the present application are merely exemplary and are not intended to limit the implementation modes of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and also impossible to exhaust all the implementation modes. Any modification, equivalent replacement and improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A deep learning based imaging domain ghost wave suppression method, characterized in that, The method comprises the following steps: S1. Constructing a seismic training sample set, artificially synthesizing a fault velocity model, and synthesizing a ghost wave training sample set based on the fault velocity model using a reverse time migration (RTM) algorithm; S11. Generating a horizontal layered velocity model V 1( X , Z ) : Randomly generating a one-dimensional velocity sequence V ( Z ) where the number of layers, thickness and velocity are randomly selected from a reasonable range determined by the actual geological conditions of the work area, and the one-dimensional velocity sequence V ( Z ) is extended along the horizontal direction to a two-dimensional horizontal layered velocity model V 1( X , Z ) S12. Generating the tilted-fold model V 2( X , Z ) : on the basis of the horizontal layered velocity model V 1( X , Z ), add stratigraphic dip and fold structure to it using the vertical displacement field M ( X , Z ) to obtain the tilted-fold model V 2( X , Z ) ; S13. Generate a real faulted velocity model V 3( X , Z ) : add fault structures in a tilted fold model V 2( X , Z ) using fault displacement field to get a real faulted velocity model V 3( X , Z ) containing stratigraphic dip, fold and fault structures; S14. Generate RTM images: Convert the true tomographic velocity model V 3( X , Z ) to seismic images using the RTM algorithm, resulting in seismic images with and without ghosts; S15. Generating a seismic training sample set; S2. Constructing a U-net neural network; The process of processing data of the U-net neural network comprises: after a seismic sample with a ghost wave is input into the U-net, first, a series of down-sampling steps are experienced in an encoder path to help the U-net capture hierarchical features; wherein a seismic image with an input size of 128 × 128 is gradually reduced to 64 × 64, 32 × 32, 16 × 16, 8 × 8 and 4 × 4, and the number of channels is gradually increased to 128, 256, 512, 1024, 2048 and 4096; then, a series of up-sampling processes are performed in a decoder path to restore the spatial size; at the same time, a high-level semantic feature in the decoder is combined with a low-level detail in the encoder through a skip connection, so that the U-net neural network can effectively aggregate coarse global context information and detailed local features; finally, a 1 × 1 convolution layer outputs a suppressed ghost wave image with the same size as the input; The mean squared error, i.e. L 2-norm, is then used as the loss function: （15） wherein, n represents the total number of pixel points in the input seismic image; p i represents the suppressed ghost image output by the U-net neural network; l i represents the label image without ghost. S3. Training the U-net neural network using the synthesized ghost wave training sample set; S4. Inputting target work area seismic data into the U-Net neural network for intelligent ghost wave suppression processing to obtain ghost wave suppressed seismic data.

2. The deep learning based imaging domain ghost wave suppression method of claim 1, wherein, The step S12 comprises vertically displacing the field M ( X , Z ) consists of two parts: M 1( X , Z ) and M 2( X , Z ) （1） In equation (1), M 1( X , Z ) is a linear equation for modeling the formation dip angle, which is expressed as: （2） wherein, α denotes the formation dip angle; M 2( X , Z ) are composed of N one-dimensional Gaussian functions, which are used to simulate the underground fold structure, and the expression is as follows: （3） wherein, Z max represents the maximum depth of the velocity model; b k , c k and σ k respectively represent the amplitude value, the center position and the half-length of the first k Gaussian function, the sizes of which are generated randomly; by superimposing and summing up N different Gaussian functions, a plurality of different fold structures are added in the velocity model; The vertical displacement field M ( X , Z ) is added to the horizontal layered velocity model V 1( X , Z ) to generate a tilted and folded model V 2( X , Z ): V 2 （4）。 3. The deep learning based imaging domain ghost wave suppression method of claim 2, wherein, The step S13 comprises: S131. Determine a fault reference line: with a reference line center point position and a fault dip angle θ define a fault reference line; wherein the center point position P and the fault dip angle θ are randomly generated within a set range; S132. Coordinate system conversion: using randomly determined reference line center point position and fault dip θ convert global coordinate system to local coordinate system x , z ) （5） wherein R is a rotation matrix: （6） After coordinate system conversion, P The point is the origin of the local coordinate system. x The axis positive direction points to the tendency. z The axis is the normal line of the fault reference line, and its positive direction points to the hanging wall part of the fault. S133. Generate fault curve: generate a series of perturbation points randomly around the fault reference line z = 0, then use these perturbation points to interpolate a fault curve z f = f ( x ) S134. Calculate the fault displacement field D( x , z ). First, calculate the dip displacement field Dx ( x , z ) in two-dimensional space, then calculate the normal displacement field Dz ( x , z ). The fault displacement field D( x , z ) is a two-dimensional vector field defined in the local coordinate system x , z ), which consists of two parts: the dip displacement field Dx ( x , z ) and the normal displacement field Dz ( x , z ). （7） wherein u x and u z respectively represent x and z unit vectors in the direction of S135. Simulating fault structures using fault displacement field: according to the calculated fault displacement field D( x , z ), add fault structures to the tilted fold model V 2( X , Z ) to obtain a fault model V 3( X , Z ): V 3 （8） In the formula, T represents a transposition symbol.

4. The deep learning based imaging domain ghost wave suppression method of claim 3, wherein, The step S134 comprises: First, the tendency displacement field is computed Dx ( x , z ) the projected displacement field on the reference line d ( x ; z = 0): （9） in, d max This represents the maximum displacement of the center point, i.e., the origin of the local coordinate system. r ( x ) is the normalized radius: （10） wherein, l x represents half the length of the fault line; Then, the fault displacement field on the fault reference line is used to calculate the fault displacement of each point in the two-dimensional space d ( x ; z = 0); define the boundary γ , so that Dx ( x , z ) is 0 when it reaches or exceeds the boundary γ . For the upper disk f ( x ) ≤ z ≤ f ( x ) + γ , Dx ( x , z ) of the calculation expression is: （11） where, λ represents the percentage of the upper disc displacement amount to the projection displacement field d ( x ; z = 0), and is randomly selected between [0, 1]; ζ ( x , z ) is a nonlinear decay function, which functions to decay the tendency displacement field Dx ( x , z ) along the normal direction away from the fault curve, and when the decay reaches the pull radius γ , the decay is 0; the nonlinear decay function ζ ( x , z ) is represented as: （12） For the lower disk f ( x ) γ ≤ z ≤ f ( x ), Dx ( x , z ) （13） Tendency displacement field in two-dimensional space Dx ( x , z ) is calculated, the normal displacement field Dz ( x , z ) is directly obtained using the following equation: （14）。 5. The deep learning based imaging domain ghost wave suppression method according to any one of claims 1 to 4, characterized in that, The step S14 includes: firstly, defining an offshore seismic data observation system; randomly selecting one from conventional horizontal streamer, up-down streamer, variable depth streamer and deep tow as a seismic data acquisition mode, and randomly selecting the sink depth of the seismic source and the geophone within a set range; then, respectively using the boundary conditions with and without the free surface to perform RTM imaging on V 3( X , Z ) to obtain the seismic images with and without the ghost wave.

6. The deep learning based imaging domain ghost wave suppression method according to any one of claims 1 to 4, characterized in that, The step S15 comprises: in the calculated RTM images with and without a ghost wave, an image with a size of 128 × 128 pixels at the same position is randomly selected, and the image containing a ghost wave is taken as a seismic training sample, and the image without a ghost wave is taken as a label image; a plurality of 128 × 128 training samples and labels are randomly selected in the two RTM images to obtain a small seismic training sample set.

7. A computer device comprising a memory and a processor, said memory storing a computer program, characterized in that, The processor executes the computer program to realize the steps of the method in any one of claims 1 to 6.

Images