Intelligent completion method for missing data of ocean remote sensing based on graph attention network

Abstract

The present invention provides a method for intelligently completing missing ocean remote sensing data based on a graph attention network, comprising: constructing an ocean remote sensing data graph structure for ocean remote sensing data with a large number of missing values, including the definition of edges, nodes, data matrices, mask matrices, and adjacency matrices; on this basis, constructing a generator and discriminator structure of a generative adversarial network, obtaining the spatial dynamic dependency matrix and the temporal short-term dependency matrix of the missing ocean remote sensing data based on a GAT network, and using the data matrix as inputs to the GRU update gate and reset gate of the internal structure of the generator and discriminator; finally, training the generator and discriminator, and iterating and optimizing using a loss function to obtain the optimal completion effect. The present invention dynamically obtains the temporal short-term dependency and spatial dynamic dependency information of ocean remote sensing data based on the GAT network, can better aggregate ocean remote sensing related information during the completion process, effectively realizes the intelligent completion of missing ocean remote sensing data, and improves the accuracy of the completion.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically relates to a method for intelligently completing missing ocean remote sensing data based on a graph attention network. Background Art

[0002] Marine remote sensing technology, characterized by real-time performance, wide coverage, and low cost, is widely used for real-time monitoring of the marine environment. Furthermore, with the rapid development of stereoscopic observation technologies—air, land, and sea—remote sensing data has grown exponentially, generating massive marine remote sensing datasets that provide a powerful reference for marine environmental disaster forecasting, disaster reduction, and other marine environmental protection efforts. Further application and analysis of remote sensing data (such as red tide warnings and typhoon path predictions) is crucial for protecting the marine environment. However, due to unavoidable factors such as cloud cover, weather, and sensor failures, remote sensing data suffers from random omissions in both spatial and temporal series, significantly reducing the effectiveness of long-term, large-scale, real-time monitoring of the marine environment. Therefore, accurate and efficient intelligent completion of remote sensing data is a crucial prerequisite for implementing marine environmental protection.

[0003] Deep learning models have the powerful ability to fully exploit the hidden information of spatiotemporal dependencies in data, and have attracted significant research interest in the prediction and restoration of spatiotemporal series data, such as traffic flow and meteorology. Compared to research on spatiotemporal traffic flow data, there has been relatively little research on the completion and prediction of marine remote sensing spatiotemporal data. However, due to its importance to marine environmental protection, this research is rapidly increasing. Marine remote sensing data exhibits some spatial and temporal characteristics that differ from traffic flow data. Firstly, spatial correlation is affected by uncontrollable factors, resulting in dynamic dependencies between adjacent nodes, with varying degrees of dependence. For example, wind can alter the spatial correlation between sea temperature nodes; changes in phytoplankton bloom patterns can alter the spatial correlation between chlorophyll concentrations. Secondly, marine remote sensing data can experience significant short-term fluctuations, altering the short-term dependency patterns between time series. It is important to note that, unlike prediction, both historical and future time series information for the location being completed is known during the completion process and should be considered. For short-term temporal dependence, such as frequent red tides in summer, when chlorophyll concentrations are abnormally higher than normal values, the dependence of the completed position on future time series is higher than that on historical time series, and the dependence pattern changes. Therefore, the dynamic temporal dependence in the short term cannot be ignored.

[0004] Therefore, it is urgent to propose an intelligent completion method for marine remote sensing data that can dynamically capture the spatial correlation and short-term temporal dependence of remote sensing data. Summary of the Invention

[0005] The present invention is designed to solve the above-mentioned problems and aims to provide an intelligent marine remote sensing data completion method that can dynamically capture the spatial correlation and short-term temporal dependence of remote sensing data, thereby achieving more ideal data completion. The present invention adopts the following technical solutions:

[0006] The present invention provides a method for intelligently completing missing data in ocean remote sensing based on a graph attention network, which is characterized by comprising the following steps:

[0007] Step S1, for the marine remote sensing data with missing data, constructing the marine remote sensing data graph structure, including the definition of nodes, edges, data matrix, mask matrix and adjacency matrix;

[0008] Step S2, based on the ocean remote sensing data graph structure, using two parallel graph attention networks to respectively learn the spatial position feature information and time series feature information of the nodes and assign weights;

[0009] Step S3, constructing a generator and discriminator structure of a generative adversarial network, and fusing the spatial position feature information, the time series feature information and their weights with the data matrix as inputs of the generator and the discriminator, thereby obtaining a data completion result of the ocean remote sensing data;

[0010] Step S4: training the generator and the discriminator, optimizing using a loss function, and obtaining the optimal data completion result of the ocean remote sensing data using the trained generative adversarial network, wherein at least the data matrix, the mask matrix, and the adjacency matrix are used as inputs of the generator.

[0011] The method for intelligently completing missing ocean remote sensing data based on a graph attention network provided by the present invention may also have the following technical features: in step S1, a plurality of directed weighted graphs are used to represent the spatiotemporal data structure of the ocean remote sensing data, each directed weighted graph represents a timestamp,

[0012] The graph structure of the directed weighted graph is represented as follows:

[0013] G=〈V,E,X,M,A〉

[0014] Where V represents the set of nodes; E represents the set of edges; X∈R L×N is a data matrix, where L is the length of the time series and N is the number of nodes; M∈R L×N is the mask matrix, where the element m i,j ∈{0,1}, its value corresponds to the data matrix X, if the element x in the data matrix i,j Is a missing value, then the corresponding mask m i,j =0, otherwise the corresponding mask is 1; A∈RN×N Represents the adjacency matrix, where element a i,j ∈{0,1}, if a i,j =1 means node v i ,v j There is an edge between them, otherwise it is 0, a i,j The value of node v i ,v j The distance between

[0015] According to the spatial position set of the node Node v i ,v j The distance calculation formula is:

[0016]

[0017] Based on node v i ,v j The distance d i,j Calculate mask a i,j The formula is:

[0018]

[0019] Where, d min Indicates the preset distance threshold.

[0020] The method for intelligently completing missing ocean remote sensing data based on a graph attention network provided by the present invention may also have the following technical features: Step S4 includes the following sub-steps:

[0021] Step S4-1: For the generator, the data matrix X and the random noise matrix Z are used as the input of the generator, where Z:N(0,1), and the mask matrix M and the adjacency matrix A corresponding to the data matrix X are also used as the input of the generator.

[0022] The discriminator takes the output G(X+Z,A,M) of the generator and the corresponding data matrix X as input to determine the probability that the completed value at positions 1-M is close to the true value;

[0023] Step S4-2, setting the loss function, including reconstruction loss L R , generator loss L G and the discriminator loss L D , defined as follows:

[0024] L R =||G(X+Z,A)eM-XeM||2

[0025] L G =λLR -D(G(X+Z,A))

[0026] L D =D(X,A)-D(G(X+Z,A))

[0027] Where G(·) represents the generator, D(·) represents the discriminator, ||·||2 represents the mean square error, and λ is the reconstruction loss hyperparameter;

[0028] Step S4-3, using the Adam optimization function to perform gradient descent training, updating the training parameters of the generator and the discriminator, and obtaining the completion model corresponding to the optimal loss function.

[0029] The method for intelligently completing missing data of ocean remote sensing based on a graph attention network provided by the present invention may also have the following technical features, wherein step S2 includes the following sub-steps:

[0030] Step S2-1, using one of the graph attention networks to obtain the spatial dynamic dependency matrix H of the node s , i.e., the spatial position feature information;

[0031] Step S2-2, using another graph attention network to obtain the short-term dependency matrix H of the node t , that is, the time series feature information.

[0032] The method for intelligent completion of missing data of ocean remote sensing based on graph attention network provided by the present invention may also have such a technical feature, wherein, in step S2-1, all spatial position features within the same timestamp are regarded as a complete graph, in which nodes represent spatial position features, edges represent the spatial proximity between two corresponding spatial position features, and the definition of node v j The set of nodes with connected edges is

[0033] The output representation of the spatial position feature for:

[0034]

[0035]

[0036] in,

[0037]

[0038] Where x i,j is the eigenvalue, Indicates that at timestamp τ i Node v j′ For node v jThe spatial relationship dependency coefficient, LeakyReLU(·) is a nonlinear activation function, is the spatial relationship dependence coefficient The corresponding standardized form.

[0039] The method for intelligent completion of missing data of ocean remote sensing based on graph attention network provided by the present invention may also have the following technical features: wherein, in step S2-2, the multivariate time series is divided into L / n time slices according to the length n, and the time series data in the same group is regarded as a fully connected graph, and the time stamp τ is defined. i The set of temporally related moments is

[0040] The output representation of the time series features for:

[0041]

[0042]

[0043] in,

[0044]

[0045] Where x i,j is the eigenvalue, Indicates that at node v j Timestamp τ i′ For timestamp τ i The short-term dependency coefficient of LeakyReLU(·) is a nonlinear activation function. is the short-term dependence coefficient The corresponding standardized form.

[0046] The method for intelligent completion of missing data of ocean remote sensing based on graph attention network provided by the present invention may also have such a technical feature, wherein the generator and the discriminator both include multiple gated recurrent units, and the gated recurrent units include update gates and reset gates. In step S3, the data matrix X and the spatial dynamic dependency matrix H are s And the short-term dependency matrix H t They serve as the inputs of the update gate and the reset gate.

[0047] The method for intelligently completing missing data in ocean remote sensing based on a graph attention network provided by the present invention may also have such a technical feature, wherein the generative adversarial network is a WGAN.

[0048] Functions and effects of the invention

[0049] The present invention's method for intelligently completing missing ocean remote sensing data based on a graph attention network (GAN) builds on the structure of the ocean remote sensing data graph by using two parallel GANs to learn the time series and spatial location feature information of nodes in the graph, assigning corresponding weights to each. These weights are then fused with the data matrix to serve as inputs to the generator and discriminator of a generative adversarial network (GAN), thereby enabling intelligent completion of missing ocean remote sensing data. Because the present invention addresses the short-term temporal dependence and spatial dynamic dependence of ocean remote sensing data, using a GAN to dynamically acquire the time series and spatial location feature information of the data, it can better aggregate ocean remote sensing-related information during the data completion process, effectively enabling intelligent completion of missing ocean remote sensing data and improving data completion accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Flowchart of a method for intelligently completing missing ocean remote sensing data based on a graph attention network in an embodiment of the present invention;

[0051] Figure 2 is an example graph of a directed weighted graph of ocean remote sensing data in an embodiment of the present invention;

[0052] Figure 3 This is an example diagram for obtaining spatial dynamic dependency and temporal short-term dependency in an embodiment of the present invention;

[0053] Figure 4 1 is a flow chart of the internal structure of a generator of a generative adversarial network according to an embodiment of the present invention;

[0054] Figure 5 This is a diagram of the architecture of a generative adversarial network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following is a detailed description of the method for intelligently completing missing data in ocean remote sensing based on a graph attention network in combination with embodiments and drawings.

[0056] <Example>

[0057] Figure 1 This is a flowchart of the method for intelligently completing missing ocean remote sensing data based on a graph attention network in this embodiment.

[0058] like Figure 1 As shown in the figure, the intelligent completion method for missing ocean remote sensing data based on graph attention network includes the following steps:

[0059] Step S1: constructing a marine remote sensing data graph structure for marine remote sensing data with a large number of missing values.

[0060] The graph structure of ocean remote sensing data is represented as:

[0061] G=〈V,E,X,M,A〉

[0062] As shown in the above formula, the ocean remote sensing data graph structure mainly includes the definitions of nodes V, edges E, data matrix X, mask matrix M and adjacency matrix A.

[0063] The specific construction process is described as follows: consider using multiple directed weighted graphs to represent the spatiotemporal data structure of ocean remote sensing, and assume that the topological structure of the directed weighted graph does not change, and each directed weighted graph represents a timestamp.

[0064] Figure 2 It is an example diagram of the directed weighted graph of ocean remote sensing data in this embodiment.

[0065] The graph structure of a directed weighted graph is as follows Figure 2 As shown in the figure, G τi Represents the timestamp τ i The graph structure can be expressed as:

[0066] G=〈V,E,X,M,A〉

[0067] Where V represents the set of nodes; E represents the set of edges; X∈R L×N is a data matrix, where L is the length of the time series and N is the number of nodes; M∈R L×N is the mask matrix, where the element m i,j ∈{0,1}, its value corresponds to the data matrix X, if the element x in the data matrix i,j Is a missing value, then the corresponding mask m i,j =0, otherwise, the corresponding mask value is 1; A∈R N×N Represents the adjacency matrix, and the element a i,j ∈{0,1}, if a i,j =1 means node v i ,v j There is an edge between them, otherwise it is 0, a i,j The value of node v i ,v j The distance between them is determined by the spatial position set of the nodes Available node v i ,v j The distance calculation formula is:

[0068]

[0069] Therefore, based on the distance d between nodes i,j You can get a i,j :

[0070]

[0071] Where, d min It represents a preset distance threshold. In this embodiment, it is considered that only two nodes with a distance less than this threshold are connected.

[0072] In step S2, based on the above-mentioned ocean remote sensing data graph structure, two parallel graph attention networks (GATs) are used to learn the spatial position feature information and time series feature information of the nodes respectively, and assign corresponding weights to them.

[0073] Step S2 specifically includes the following sub-steps:

[0074] Step S2-1, use one of the graph attention networks to obtain the node's spatial dynamic dependency matrix H s , that is, the above-mentioned spatial position feature information.

[0075] Specifically, all spatial position features within the same timestamp are considered as a complete graph, where nodes (denoted as spatial position feature nodes) represent spatial position feature values and edges represent the spatial proximity between two corresponding spatial position features. i,j That is, at timestamp τ i Node v j The corresponding eigenvalues are spatially affected by the node v at the same time stamp. j The influence of the eigenvalues of other spatially adjacent nodes is defined as the same as that of node v j The set of nodes with connected edges in space is

[0076] Figure 3 This is an example diagram for obtaining spatial dynamic dependency and temporal short-term dependency in this embodiment.

[0077] like Figure 3 As shown, the spatial position feature value x of timestamp τ1 1,4 The output is represented as Affected by the spatial position feature node x that has a connection edge with this spatial position feature node 1,1 、x 1,3 and x 1,5 The three nodes affect each other. Then, the data value x can be obtained by using the GAT network that captures the dynamic dependence of space i,j Output representation after obtaining spatial dynamic dependencies for:

[0078]

[0079]

[0080] Where, Indicates that at timestamp τ i Node v j′ For node v j The spatial relationship dependency coefficient of , LeakyReLU( ) is a nonlinear activation function.

[0081] in:

[0082]

[0083] Where, is the spatial relationship dependence coefficient The corresponding standardized form.

[0084] Based on the above formula, the output representation of the spatial dynamic dependency of all eigenvalues can be calculated, and then the spatial dynamic dependency matrix can be obtained.

[0085] Step S2-2, use another graph attention network to obtain the node's short-term dependency matrix H t , that is, the above-mentioned time series feature information.

[0086] Specifically, the multivariate time series is divided into time slices of fixed length. All timestamps in the same time slice form a fully connected complete graph. The timestamp samples are input into GAT, and GAT assigns different weights to different timestamp samples.

[0087] In this embodiment, the multivariate time series is divided into L / n time slices according to the length n, and the time series data in the same time slice is considered to be a completely connected graph. i,j That is, at timestamp τ i Node v j The corresponding eigenvalue is temporally affected by all other timestamps in the same time slice, that is, all other timestamps in the same time slice are affected by the timestamp τ i Related, defined with timestamp τ i The set of temporally related timestamps is

[0088] like Figure 3 As shown, the length is set to 3 to split the time slice. In the first time slice, the time series feature x of node v2 is 3,1 Affected by other nodes in this time slice, that is, affected by x 1,1 with x 1,2 The data value x can be obtained by using the GAT network that captures the short-term dependence of time. i,j Output representation after obtaining short-term temporal dependencies for:

[0089]

[0090]

[0091] Where, Indicates that at node v j Timestamp τ i′ For timestamp τ i The short-term dependence coefficient of time;

[0092]

[0093] Where, is the short-term dependence coefficient The corresponding standardized form.

[0094] Based on the above formula, the output representation of the short-term time dependency of all eigenvalues can be calculated, and then the short-term time dependency matrix can be obtained.

[0095] In step S3, the generator and discriminator structures of the generative adversarial network are constructed, and the time series feature information, node spatial position feature information and their weights obtained in step S2 are fused with the data matrix as the input of the generator and discriminator, thereby obtaining the data completion results of the ocean remote sensing data.

[0096] Figure 4 This is a flowchart of the internal structure of the generator of the generative adversarial network in this embodiment. The discriminator has the same structure as the generator.

[0097] like Figure 4 As shown, both the generator and the discriminator contain multiple gated recurrent units (GRUs), which can be divided into update gates, reset gates, etc. In step S3, the above data matrix X and spatial dynamic dependency matrix H are s And the short-term dependency matrix H t Together they serve as inputs to the update gate and the reset gate.

[0098] Figure 5 This is a diagram of the architecture of the generative adversarial network model in this embodiment.

[0099] like Figure 5 As shown, in this embodiment, the generative adversarial network is WGAN (Wasserstein GAN). WGAN can effectively solve the problems of GAN model collapse and non-convergence. Therefore, this embodiment uses WGAN for intelligent completion of ocean remote sensing data.

[0100] In step S4, the generator and discriminator of the generative adversarial network are trained, the loss function is used to iteratively optimize, and the trained generative adversarial network is used to obtain the optimal data completion result for the ocean remote sensing data.

[0101] Step S4 specifically includes the following sub-steps:

[0102] Step S4-1: For the generator, the data matrix X and the random noise matrix Z are used as the generator input, where Z:N(0,1). At the same time, the mask matrix M and the adjacency matrix A corresponding to the data matrix X are also used as the generator input to generate training data that is closer to the real data;

[0103] For the discriminator, the output of the generator G(X+Z,A,M) and the corresponding original data matrix X are used as input to determine the probability that the completed value at the 1-M position is close to the true value.

[0104] Step S4-2, setting the loss function, which includes the reconstruction loss L R , generator loss L G and the discriminator loss L D , specifically defined as follows:

[0105] L R =||G(X+Z,A)eM-XeM||2

[0106] L G =λL R -D(G(X+Z,A))

[0107] L D =D(X,A)-D(G(X+Z,A))

[0108] Where G(·) represents the generator, D(·) represents the discriminator, ||·||2 represents the mean squared error, and λ is the reconstruction loss hyperparameter.

[0109] In step S4-3, the Adam optimization function is used for gradient descent training to update the generator and discriminator training parameters to obtain the completion model corresponding to the optimal loss function, thereby enabling intelligent completion of missing data in marine remote sensing data.

[0110] In this embodiment, parts not described in detail are well-known technologies in the art.

[0111] Example Function and Effect

[0112] According to the method for intelligent completion of missing ocean remote sensing data based on a graph attention network provided by this embodiment, based on the definition of the graph structure of ocean remote sensing data, two parallel GAT networks are used to learn the time series feature information and spatial position feature information of the nodes in the graph respectively, and corresponding weights are assigned to them to obtain a short-term temporal dependency matrix and a spatial dynamic dependency matrix. These are then fused with the data matrix and used as the inputs of the GRU update gate and reset gate of the generator and discriminator of the GAN network, thereby realizing intelligent completion of missing ocean remote sensing data. Because the present invention targets the characteristics of short-term temporal dependency and spatial dynamic dependency of ocean remote sensing data, and uses a graph attention network to dynamically obtain the time series feature information and spatial position feature information of ocean remote sensing data, it can better aggregate ocean remote sensing related information during the data completion process, effectively realize intelligent completion of missing ocean remote sensing data, and improve the accuracy of data completion.

[0113] In addition, WGAN is used in the embodiment, which can effectively solve the problems of GAN model collapse and non-convergence.

[0114] The above embodiments are only used to illustrate specific implementations of the present invention, and the present invention is not limited to the description scope of the above embodiments.

Claims

1. A method for intelligent completion of missing ocean remote sensing data based on graph attention network, characterized by: The following steps are involved: Step S1, for the marine remote sensing data with missing data, constructing the marine remote sensing data graph structure, including the definition of nodes, edges, data matrix, mask matrix and adjacency matrix; Step S2, based on the ocean remote sensing data graph structure, using two parallel graph attention networks to respectively learn the spatial position feature information and time series feature information of the nodes and assign weights; Step S3, constructing a generator and discriminator structure of a generative adversarial network, and fusing the spatial position feature information, the time series feature information and their weights with the data matrix as inputs of the generator and the discriminator, thereby obtaining a data completion result of the ocean remote sensing data; Step S4: training the generator and the discriminator, optimizing using a loss function, and obtaining an optimal data completion result for the ocean remote sensing data using the trained generative adversarial network, wherein at least the data matrix, the mask matrix, and the adjacency matrix are used as inputs to the generator. Wherein, in step S1, a plurality of directed weighted graphs are used to represent the spatiotemporal data structure of the ocean remote sensing data, and each directed weighted graph represents a timestamp. The graph structure of the directed weighted graph is represented as follows: G=<V,E,X,M,A> Where V represents the set of nodes; E represents the set of edges; X∈R L×N is a data matrix, where L is the length of the time series and N is the number of nodes; M∈R L×N is the mask matrix, where the element m i,j ∈{0,1}, its value corresponds to the data matrix X, if the element x in the data matrix i,j Is a missing value, then the corresponding mask m i,j =0, otherwise the corresponding mask is 1; A∈R N×N Represents the adjacency matrix, where element a i,j ∈{0,1}, if a i,j =1 means node v i ,v j There is an edge between them, otherwise it is 0, a i,j The value of node v i ,v j The distance between According to the spatial position set of the node Node v i ,v j The distance calculation formula is: Based on node v i ,v j The distance d i,j Calculate mask a i,j The formula is: Where, d min Indicates the preset distance threshold.

2. The method for intelligently completing missing ocean remote sensing data based on graph attention network according to claim 1, Its characteristics are: in, Step S4 includes the following sub-steps: Step S4-1: For the generator, the data matrix X and the random noise matrix Z are used as the input of the generator, where Z~N(0,1), and the mask matrix M and the adjacency matrix A corresponding to the data matrix X are also used as the input of the generator. The discriminator takes the output G(X+Z,A,M) of the generator and the corresponding data matrix X as input to determine the probability that the completed value at positions 1-M is close to the true value; Step S4-2, setting the loss function, including reconstruction loss L R , generator loss L G and the discriminator loss L D , defined as follows: L R =||G(X+Z,A)⊙M-X⊙M||2 L G =λL R -D(G(X+Z,A)) L D =D(X,A)-D(G(X+Z,A)) Where G(·) represents the generator, D(·) represents the discriminator, ||·||2 represents the mean square error, and λ is the reconstruction loss hyperparameter; Step S4-3, using the Adam optimization function to perform gradient descent training, updating the training parameters of the generator and the discriminator, and obtaining the completion model corresponding to the optimal loss function.

3. The method for intelligently completing missing data of ocean remote sensing based on graph attention network according to claim 1, Its characteristics are: in, Step S2 includes the following sub-steps: Step S2-1, using one of the graph attention networks to obtain the spatial dynamic dependency matrix H of the node s , i.e., the spatial position feature information; Step S2-2, using another graph attention network to obtain the short-term dependency matrix H of the node t , that is, the time series feature information.

4. The method for intelligently completing missing ocean remote sensing data based on a graph attention network according to claim 3 is characterized by: in, In step S2-1, all spatial position features within the same timestamp are considered as a complete graph, in which nodes represent the spatial position features, and edges represent the spatial proximity between two corresponding spatial position features. j The set of nodes with connected edges is The output representation of the spatial position feature for: in, Where x i,j is the eigenvalue, Indicates that at timestamp τ i Node v j′ For node v j The spatial relationship dependency coefficient, LeakyReLU(·) is a nonlinear activation function, is the spatial relationship dependence coefficient The corresponding standardized form.

5. The method for intelligently completing missing ocean remote sensing data based on a graph attention network according to claim 3 is characterized by: in, In step S2-2, the multivariate time series is divided into L / n time slices according to the length n, and the time series data in the same group is regarded as a fully connected graph, and the time stamp τ is defined. i The set of temporally related moments is The output representation of the time series features for: in, Where x i,j is the eigenvalue, Indicates that at node v j Timestamp τ i′ For timestamp τ i The short-term dependency coefficient of LeakyReLU(·) is a nonlinear activation function. is the short-term dependence coefficient The corresponding standardized form.

6. The method for intelligently completing missing ocean remote sensing data based on a graph attention network according to claim 3 is characterized by: in, The generator and discriminator both contain multiple gated recurrent units. The gated recurrent unit includes an update gate and a reset gate. In step S3, the data matrix X and the spatial dynamic dependency matrix H are s And the short-term dependency matrix H t They serve as the inputs of the update gate and the reset gate.

7. The method for intelligently completing missing ocean remote sensing data based on a graph attention network according to claim 1 is characterized by: in, The generative adversarial network is WGAN.

Images