A method and system for reconstructing three-dimensional ocean temperature and salinity fields using satellite remote sensing data

Abstract

The present invention belongs to the technical field of ocean three-dimensional temperature and salinity field data processing, and discloses a method and system for reconstructing the ocean three-dimensional temperature and salinity field using satellite remote sensing data. The method obtains the daily average sea surface satellite remote sensing data of the experimental area and performs interpolation and normalization processing; the processed sea surface satellite remote sensing data is input into a constructed ocean three-dimensional temperature and salinity field reconstruction model based on a graph attention network and a temporal convolution network, and after temporal convolution, graph learning, and graph attention processing, the ocean three-dimensional temperature field or salinity field is output; the output ocean three-dimensional temperature field or salinity field reconstruction result is verified for spatial distribution consistency with the ocean reanalysis data as real data. The present invention uses a certain sea area as the experimental area to carry out an ocean three-dimensional temperature and salinity field reconstruction experiment based on sea surface satellite remote sensing data to evaluate the effectiveness and feasibility of the proposed model method.

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean three-dimensional temperature-salinity field data processing, and in particular relates to a method and system for reconstructing ocean three-dimensional temperature-salinity field by using satellite remote sensing data. Background Art

[0002] The ocean covers 71% of the Earth's surface. The physical properties of seawater are closely linked to its movement, dynamic oceanography, and ecological processes. This information plays a crucial role in marine environmental protection, resource development, and maritime navigation safety. Seawater temperature and salinity are fundamental elements of ocean hydrology, closely linked to nearly all ocean phenomena and dynamic processes. Their spatial and temporal distribution and variability constitute a core area of ​​marine scientific research. Satellite remote sensing data is abundant, but it only captures information about the ocean surface. Summary of the Invention

[0003] In order to overcome the problems existing in the related art, the embodiments disclosed in the present invention provide a method and system for reconstructing the three-dimensional temperature and salinity field of the ocean using satellite remote sensing data, specifically relating to a method and system for reconstructing the three-dimensional temperature and salinity field of the ocean using satellite remote sensing data based on a graph attention neural network.

[0004] The technical solution is as follows: a method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data, taking into account the spatiotemporal correlation between the three-dimensional ocean temperature and salinity field and sea surface satellite remote sensing data, using a graph attention network to establish the spatial correlation between the sea surface satellite remote sensing data and the three-dimensional ocean temperature and salinity field, and using a temporal convolutional network to establish the temporal correlation between the sea surface satellite remote sensing data and the three-dimensional ocean temperature and salinity field; specifically comprising the following steps:

[0005] S1. Obtain the daily average sea surface satellite remote sensing data for the experimental area and perform interpolation processing to unify the spatial resolution of different data, perform standardization processing, and eliminate the value differences between different data; convert the grid data in the sea surface satellite remote sensing data into graph node data; the sea surface satellite remote sensing data includes: sea surface height anomaly SLA, sea surface temperature SST, sea surface salinity SSS and sea surface wind SSW, and the sea surface wind is divided into the latitudinal component USSW and the longitudinal component VSSW;

[0006] S2, the processed sea surface satellite remote sensing data is input into the constructed ocean 3D temperature and salinity field reconstruction model based on graph attention network and temporal convolution network, and after temporal convolution, graph learning and graph attention processing, the 3D ocean temperature field or salinity field is output;

[0007] S3, verify the spatial distribution consistency of the output ocean three-dimensional temperature field or salinity field reconstruction results with the ocean reanalysis data as the real data.

[0008] In step S1, the daily average sea surface satellite remote sensing data is processed by interpolation to a uniform spatial resolution of 0.25 ° × 0.25 ° After the spatial resolution of the data is unified, the data is standardized with a mean of 0 and a variance of 1 to eliminate the differences in the values ​​of different data.

[0009] In step S2, the ocean three-dimensional temperature and salinity field reconstruction model based on graph attention network and temporal convolutional network consists of a multi-layer stack of temporal convolution module, graph learning module and graph attention module. The input sea surface feature sequence is first convolved by 1×1 to change the number of channels, converting low-channel sea surface features into higher-dimensional feature representations.

[0010] Furthermore, the temporal convolution module is used to extract the temporal information features of sea surface satellite remote sensing data. It includes dual-path multi-scale dilated convolution and channel attention, and extracts features through two independent paths. The first path introduces the global channel attention CAM module after multi-scale dilated convolution to capture global dependencies. The second path introduces lightweight channel attention ECA after multi-scale dilated convolution to process local feature dependencies.

[0011] Furthermore, the temporal convolution module uses multi-scale convolution to effectively capture multi-scale temporal features through a variety of different convolution kernels. Four convolution kernels, 1×2, 1×3, 1×5, and 1×7, are selected to extract temporal features of different scales. The output features of each dilated convolution are concatenated in the channel dimension to form a fused multi-scale feature representation as the output of the multi-scale dilated convolution. The calculation formula for the multi-scale dilated convolution is:

[0012] ;

[0013] Where, is the feature input of the temporal convolutional network, is the convolution operation, is the convolution kernel, is the convolution kernel size, ; is the convolution kernel Middle elements, is the input eigenvalue, is the length of time, is the expansion factor, is the step length;

[0014] ;

[0015] Where, is the output after feature splicing, For feature splicing, is a 1×2 convolution kernel, is a 1×3 convolution kernel, is a 1×5 convolution kernel, is a 1×7 convolution kernel;

[0016] In the first path of temporal convolution, the global channel attention CAM module is introduced to capture global dependencies, globally pool each channel generated by multi-scale dilated convolution, compress the node dimension and time dimension of each channel, extract global information and reduce the amount of calculation. The calculation formula is:

[0017] ;

[0018] Where, For the The global spatiotemporal information representation of channels, is the number of nodes, For the The first channel nodes and The feature representation of time;

[0019] Two different fully connected layers are used to perform nonlinear transformations on the globally pooled features to learn the complex dependencies between channels. The first fully connected layer is responsible for channel compression, and the second fully connected layer restores the number of channels to the original dimension.

[0020] ;

[0021] Where, is the output of two fully connected layers, is the weight matrix of the second fully connected layer, is the ReLU activation function, is the weight matrix of the first fully connected layer;

[0022] The output is normalized using the Softmax activation function, the channel weight is converted into a probability distribution, the channel attention weight is generated, and the generated channel attention weight is multiplied by the original feature to obtain the weighted feature , dynamically adjust the feature strength of each channel and introduce Dropout to prevent overfitting;

[0023] In the second path of temporal convolution, lightweight channel attention (ECA) is introduced to perform average pooling on each channel generated by multi-scale dilated convolution to compress the spatiotemporal feature dimensions, and use 1D convolution to capture the local dependency of adjacent channels. The calculation formula is:

[0024] ;

[0025] Where, is the output of 1D convolution, is the weight parameter of the 1D convolution kernel, is the feature vector output after average pooling processing;

[0026] The Sigmoid activation function is used to generate channel attention weights from the 1D convolution output results to dynamically adjust the channel importance; the generated channel attention is multiplied by the original feature to obtain the weighted feature ;

[0027] The weighted features generated by the first and second paths of the temporal convolution are respectively fused after the Sigmoid activation function and the Tanh activation function to obtain the complete temporal features; the final output of the temporal convolution module is:

[0028] ;

[0029] Where, is the output of the temporal convolution module, is the Sigmoid activation function, is the Hadamard product, is the tanh activation function, are the outputs of the two paths respectively.

[0030] Furthermore, the graph learning module randomly initializes the embedding vectors of all nodes , dynamically optimized and updated with model training, reflecting more realistic ocean space similarity characteristics, is the number of nodes, is the node embedding dimension, is the real number space;

[0031] Adaptive adjacency matrix Defined as:

[0032] ;

[0033] Where, is the Softmax activation function, is the embedding vector, is the LeakyReLU activation function, for The transposed vector of

[0034] use Calculate the dot product similarity between nodes and use the LeakyReLU activation function to enhance the nonlinearity of the model and avoid gradient vanishing. Use the Softmax activation function to convert the similarity into a probability distribution for weighted aggregation of information from adjacent nodes.

[0035] The adaptive adjacency matrix obtained is a fully connected adjacency matrix. Sparse processing is introduced to retain only the connection relationships of nodes with relatively high similarity. For each node, the top 10% of nodes with high connection weights are retained. nodes, forming a node index set ;

[0036] Construct a mask matrix M to mark the neighbor positions retained by each node. The element calculation formula of M is:

[0037] ;

[0038] Where, For the Node and The connection relationship of the nodes, For the The index set of nodes;

[0039] Adaptive adjacency matrix after sparse processing Expressed as:

[0040] ;

[0041] Pair Matrix Extract the edge list of the sparse matrix and convert it from dense representation to sparse representation; extract the matrix All non-zero element coordinates in are connected as edges, and the correlation strength between nodes of all non-zero elements is used as edge weights ; Finally, the adaptive adjacency matrix is ​​converted into a vector containing two dimensions: edge connection relationship and edge weight, which is used as input to the subsequent graph attention module.

[0042] Furthermore, the graph attention module associates the adaptive adjacency matrix with the graph attention coefficient matrix to jointly capture the dynamic spatial correlation between nodes; the input of the graph attention module is the node feature representation generated by temporal convolution , ,in is the number of nodes, is the number of feature channels; introduce the weight matrix , the weight matrix A parameterized shared linear transformation is applied to each node, and a shared attention mechanism is used between nodes. Calculate the importance between nodes , and its calculation formula is:

[0043] ;

[0044] Where, Node and nodes The eigenvector of

[0045] According to the calculated node importance, the nonlinear ability of the model is enhanced by the LeakyReLU function, and the Softmax function is used for normalization; the adaptive adjacency matrix generated by the graph learning module is combined with the edge weights generated by the adaptive adjacency matrix. Perform weighted attention calculation to obtain the attention coefficient , the calculation formula is:

[0046] ;

[0047] Where, For the nodes and The edge weights connecting the nodes, for index, For the nodes and The edge weights connecting the nodes;

[0048] Get the attention coefficient After that, the information of adjacent nodes is aggregated by weighted summation, and the feature representation of the current node is updated. A multi-head attention mechanism is introduced, and each attention head independently calculates the attention coefficient and node representation. The outputs of multiple heads are spliced ​​to enhance the expression ability of the model. The calculation formula is:

[0049] ;

[0050] Where, It is the feature concatenation output of multi-head attention. is the number of multi-head attention heads, For the The attention coefficient of the head, For the The weight matrix of each head; For multiple feature stitching, is the Sigmoid activation function;

[0051] After the output of the graph attention module is residually connected and normalized, the final output of the graph attention module is obtained. The calculation formula is:

[0052] ;

[0053] Where, is the final output of the graph attention module, for Normalization.

[0054] In step S2, the ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network uses the Adam optimizer to adjust the model learning process and optimize the model performance.

[0055] Another object of the present invention is to provide a system for reconstructing a three-dimensional ocean temperature and salinity field using satellite remote sensing data. The system implements the method for reconstructing a three-dimensional ocean temperature and salinity field using satellite remote sensing data. The system comprises:

[0056] The sea surface satellite remote sensing data acquisition and preprocessing module is used to obtain the daily average sea surface satellite remote sensing data of the experimental area, perform interpolation processing to unify the spatial resolution of different data, perform standardization processing to eliminate the value differences between different data, and convert the grid data in the original sea surface satellite remote sensing data into graph node data; the sea surface satellite remote sensing data includes: sea surface height anomaly SLA, sea surface temperature SST, sea surface salinity SSS and sea surface wind SSW, where the sea surface wind is divided into latitudinal component USSW and longitudinal component VSSW;

[0057] The ocean three-dimensional temperature and salinity field reconstruction model construction and processing module is used to input the processed sea surface satellite remote sensing data into the constructed ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolution network, and output the ocean three-dimensional temperature field or salinity field through temporal convolution, graph learning and graph attention processing;

[0058] The verification module is used to verify the spatial distribution consistency of the output ocean three-dimensional temperature field or salinity field reconstruction results with the ocean reanalysis data as the real data.

[0059] Furthermore, the system is mounted on a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the functions of the above system can be realized.

[0060] Combining all the above technical solutions, the present invention has the following beneficial effects: To fully utilize abundant sea surface remote sensing data and achieve global perception of spatiotemporal information, the present invention integrates a graph attention neural network with a temporal convolution module to construct a graph attention neural network-based ocean three-dimensional temperature and salinity reconstruction method, which can reconstruct the ocean three-dimensional temperature and salinity field from sea surface satellite remote sensing data. The constructed three-dimensional temperature and salinity field reconstruction model takes as input sea surface satellite remote sensing data such as sea surface height anomaly (SLA), sea surface temperature (SST), sea surface salinity (SSS), and sea surface wind speed (SSW), and outputs the three-dimensional ocean temperature and salinity field. Experiments were conducted in a certain sea area to reconstruct the ocean three-dimensional temperature and salinity field based on sea surface satellite remote sensing data to evaluate the effectiveness and feasibility of the proposed model method. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure;

[0062] Figure 1 This is a flow chart of a method for reconstructing a three-dimensional ocean temperature and salinity field using satellite remote sensing data provided by an embodiment of the present invention;

[0063] Figure 2 This is the overall structure diagram of the ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network provided by an embodiment of the present invention;

[0064] Figure 3 is a structural diagram of a temporal convolution module provided by an embodiment of the present invention;

[0065] Figure 4 This is a graph of the seawater temperature reconstruction result at a depth of 30 m using the present invention;

[0066] Figure 5 This is a graph of the seawater temperature reconstruction result at a depth of 65 m using the present invention;

[0067] Figure 6 This is a graph showing the reconstruction result of seawater temperature at a depth of 110 m using the present invention;

[0068] Figure 7 This is a result diagram of seawater temperature reconstruction at a depth of 222m using the present invention;

[0069] Figure 8 This is the result of reconstructing the seawater temperature at 30m depth from the ocean reanalysis data;

[0070] Figure 9 This is the result of reconstructing the seawater temperature at a depth of 65m from the ocean reanalysis data;

[0071] Figure 10 This is the result of reconstructing the seawater temperature at a depth of 110m from the ocean reanalysis data;

[0072] Figure 11 This is the result of reconstructing the seawater temperature at a depth of 222m from the ocean reanalysis data;

[0073] Figure 12 This is a diagram of the reconstructed result of seawater salinity at a depth of 30 m using the present invention;

[0074] Figure 13 This is a result diagram of seawater salinity reconstruction at a depth of 65 m using the present invention;

[0075] Figure 14 This is a diagram of the seawater salinity reconstruction result at a depth of 110 m using the present invention;

[0076] Figure 15This is a result diagram of seawater salinity reconstruction at a depth of 222 m using the present invention;

[0077] Figure 16 This is the result of reconstructing the salinity of seawater at a depth of 30m from the ocean reanalysis data;

[0078] Figure 17 This is the result of reconstructing the salinity of seawater at a depth of 65m from the ocean reanalysis data;

[0079] Figure 18 This is the result of reconstructing the salinity of seawater at a depth of 110m from the ocean reanalysis data;

[0080] Figure 19 This is the result of reconstructing the salinity of seawater at a depth of 222m using ocean reanalysis data;

[0081] Figure 20 This is a graph showing the RMSE of the temperature field reconstruction result versus depth.

[0082] Figure 21 This is a graph showing the RMSE of the temperature field reconstruction result over time;

[0083] Figure 22 This is a graph showing the RMSE of the salinity field reconstruction results versus depth;

[0084] Figure 23 This is a graph showing the RMSE of the salinity field reconstruction results over time. DETAILED DESCRIPTION

[0085] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0086] The innovation of the present invention lies in that: the present invention takes into account the spatiotemporal correlation between the three-dimensional ocean temperature and salinity field and the sea surface satellite remote sensing data, uses a graph attention network to establish the spatial correlation between the sea surface satellite remote sensing data and the three-dimensional ocean temperature and salinity field, and uses a temporal convolutional network to establish the temporal correlation between the sea surface satellite remote sensing data and the three-dimensional ocean temperature and salinity field.

[0087] Example 1, as Figure 1 As shown, the method provided by the embodiment of the present invention for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data includes:

[0088] S1. Obtain the daily average sea surface satellite remote sensing data for the experimental area and perform interpolation processing to unify the spatial resolution of different data, perform standardization processing, and eliminate the value differences between different data; convert the grid data in the sea surface satellite remote sensing data into graph node data; the sea surface satellite remote sensing data includes: sea surface height anomaly SLA, sea surface temperature SST, sea surface salinity SSS and sea surface wind SSW, and the sea surface wind is divided into the latitudinal component USSW and the longitudinal component VSSW;

[0089] S2, the processed sea surface satellite remote sensing data is input into the constructed ocean 3D temperature and salinity field reconstruction model based on graph attention network and temporal convolution network, and after temporal convolution, graph learning and graph attention processing, the 3D ocean temperature field or salinity field is output;

[0090] S3, verify the spatial distribution consistency of the output ocean three-dimensional temperature field or salinity field reconstruction results with the ocean reanalysis data as the real data.

[0091] For example, in step S1, a certain sea area (105°E-123°E, 5°N-23°N) is selected as the experimental area. The unique geographical location and complex marine environment of this sea area make it a key area for studying the interaction between ocean and climate. A three-dimensional ocean temperature and salinity field reconstruction method based on a graph attention neural network is used to reconstruct the three-dimensional ocean temperature and salinity field of a certain sea using sea surface satellite remote sensing data. The sea surface satellite remote sensing data includes: sea surface height anomaly (SLA), sea surface temperature (SST), sea surface salinity (SSS), and sea surface wind (SSW), where the sea surface wind is divided into a latitudinal component (USSW) and a meridional component (VSSW).

[0092] The SLA data are L4 grid data of the altimeter provided by the Copernicus Marine Environment Monitoring Service (CMEMS), with a temporal resolution of daily and a spatial resolution of 0.25°.

[0093] The SST data are microwave and infrared fusion data provided by the REMSS remote sensing system, with a temporal resolution of daily and a spatial resolution of 0.25°.

[0094] The SSW data is a fusion of sea surface wind field data developed and released by the Remote Sensing System (RSS). The temporal resolution of the data is 6 hours and the spatial resolution is 0.25°.

[0095] The SSS data is a multi-source sea surface salinity fusion data developed by CNR in Italy and provided by CMEMS. The temporal resolution of the data is daily and the spatial resolution is 0.125°.

[0096] All data used in this paper are daily average data, the time range is 2013-2018, and the data used are processed by interpolation to a unified spatial resolution of 0.25 °× 0.25 ° After unifying the spatial resolution of the data, the data is normalized to a mean of 0 and a variance of 1 to eliminate the differences in values ​​between different data. Before inputting the original data into the model, the original grid data needs to be converted into graph node data to adapt to the graph neural network model.

[0097] For example, in step S2, most of the existing ocean three-dimensional temperature and salinity field reconstruction methods currently only consider the temporal feature information or spatial feature information of sea surface data and the amount of information is relatively limited. Models represented by convolutional neural networks have limited receptive fields and are difficult to perform global perception of spatiotemporal information. The present invention combines the graph attention network (GAT) with the temporal convolutional network (TCN), fully integrating the spatiotemporal information characteristics of sea surface remote sensing data, and constructing an ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network using sea surface satellite remote sensing data, thereby realizing intelligent reconstruction of the ocean three-dimensional temperature and salinity field based on global perception. The overall structure of the ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network is as follows: Figure 2 shown.

[0098] The 3D ocean temperature and salinity reconstruction model based on a graph attention network and a temporal convolutional network is primarily composed of a multi-layer stack of temporal convolutional modules, graph learning modules, and graph attention modules. The input data for this model are sea surface height anomaly (SLA), sea surface temperature (SST), sea surface salinity (SSS), and sea surface wind (SSW), and the output data is a 3D ocean temperature or salinity field. The input sea surface feature sequence is first subjected to a 1×1 convolution to change the number of channels, converting low-channel sea surface features into higher-dimensional feature representations.

[0099] For example, the temporal convolution module focuses on extracting the temporal information features of sea surface satellite remote sensing data, including dual-path multi-scale dilated convolution and channel attention. It extracts features through two independent paths, reduces the model's dependence on a single feature, and improves the model's robustness. The temporal convolution module structure is as follows: Figure 3 As shown in the figure, the first path introduces a global channel attention (CAM) module after multi-scale dilated convolution to capture global dependencies. The second path introduces lightweight channel attention (ECA) after multi-scale dilated convolution to handle local feature dependencies and enhance the model's expressiveness. The dual-pathway network integrates features to capture the complex dependencies of the global temperature-salinity field while efficiently processing local features, thereby improving the accuracy and efficiency of 3D temperature-salinity field reconstruction.

[0100] Multi-scale dilated convolution uses causal convolution to make the current output dependent on the input at the current and past moments to meet the principle of time dependence. Considering that the receptive field of causal convolution is limited by the size of the convolution kernel, but building long-term memory requires a large amount of perception, the use of dilated causal convolution can reduce the computational complexity of the model while expanding the receptive field. Ocean data contains multiple time-scale feature information, and a single convolution kernel is difficult to extract the complete time feature mode. Therefore, multi-scale convolution is used to effectively capture multi-scale time features through a variety of different convolution kernels. Four convolution kernels of 1×2, 1×3, 1×5, and 1×7 are selected to extract time features of different scales, and the output features of each dilated convolution are spliced ​​in the channel dimension to form a fused multi-scale feature representation as the output of the multi-scale dilated convolution. The calculation formula for multi-scale dilated convolution is:

[0101] ;

[0102] Where, is the feature input of the temporal convolutional network, is the convolution operation, is the convolution kernel, is the convolution kernel size, ; is the convolution kernel Middle elements, is the input eigenvalue, is the length of time, is the expansion factor, is the step length;

[0103]

[0104] Where, is the output after feature splicing, For feature splicing, is a 1×2 convolution kernel, is a 1×3 convolution kernel, is a 1×5 convolution kernel, is a 1×7 convolution kernel;

[0105] Multi-scale dilated convolutions can capture long-term dependencies, but they distribute feature weights uniformly, making it difficult to distinguish relatively important channel features. To address this, a channel attention mechanism is introduced to learn the importance weight of each channel and weight feature channels, thereby highlighting important features and suppressing unimportant ones, enhancing the model's focus on key information.

[0106] In the first path of temporal convolution, the global channel attention (CAM) module is introduced to capture global dependencies, globally pool each channel generated by multi-scale dilated convolution, compress the node dimension and time dimension of each channel, extract global information and reduce the amount of calculation. The calculation formula is:

[0107] ;

[0108] Where, For the The global spatiotemporal information representation of channels, is the number of nodes, For the The first channel nodes and The feature representation of time;

[0109] The globally pooled features are transformed nonlinearly through two different fully connected layers to learn the complex dependencies between channels. The first fully connected layer is responsible for channel compression, and the second fully connected layer restores the number of channels to the original dimension.

[0110] ;

[0111] Where, is the output of two fully connected layers, is the weight matrix of the second fully connected layer, is the ReLU activation function, is the weight matrix of the first fully connected layer;

[0112] The output is normalized using the Softmax activation function, the channel weights are converted into probability distributions, and channel attention weights are generated to enhance the contribution of important channels and suppress the influence of unimportant channels. The generated channel attention weights are multiplied by the original features to obtain weighted features. , dynamically adjust the feature strength of each channel. Dropout is introduced to prevent overfitting.

[0113] In the second path of temporal convolution, lightweight channel attention (ECA) is introduced to perform average pooling on each channel generated by multi-scale dilated convolution to compress the spatiotemporal feature dimensions. 1D convolution is used to capture the local dependencies of adjacent channels and enhance the expressiveness of the model. The calculation formula is:

[0114] ;

[0115] Where, is the output of 1D convolution, is the weight parameter of the 1D convolution kernel, is the feature vector output after average pooling processing;

[0116] The Sigmoid activation function is used to generate channel attention weights from the 1D convolution output results to dynamically adjust the channel importance; the generated channel attention is multiplied by the original feature to obtain the weighted feature ;

[0117] The weighted features generated by the first and second paths of the temporal convolution are respectively fused after the Sigmoid activation function and the Tanh activation function to obtain the complete temporal features; the final output of the temporal convolution module is:

[0118] ;

[0119] Where, is the output of the temporal convolution module, is the Sigmoid activation function, is the Hadamard product, is the tanh activation function, are the outputs of the two paths respectively.

[0120] For example, the graph learning module aims to construct an adaptive adjacency matrix that adapts to the complex and changing nature of the ocean, used to calculate graph attention weight coefficients. By introducing learnable random embedding nodes, the module constructs an adjacency matrix that reflects the correlations between nodes. As the model trains, it continuously improves its parameters, gradually adapting to changes in the spatial correlations between nodes during the model learning process. This module models stable long-term spatial correlations to obtain the optimal adaptive adjacency matrix.

[0121] In order to solve the problem that the existing adjacency matrix constructed based on the inherent attributes between nodes may not fully reflect the potential spatial correlation in the data, the present invention designs an adaptive adjacency matrix that can adapt to the changing relationships between nodes during the model training process through learning, dynamically adjust the node relationships, and more realistically reflect the spatial relationships of ocean dynamic processes. The graph learning module first randomly initializes the embedding vectors of all nodes , is dynamically optimized and updated with model training to reflect more realistic ocean spatial similarity characteristics, among which, is the number of nodes, is the node embedding dimension, is the real number space;

[0122] Adaptive adjacency matrix Defined as:

[0123] ;

[0124] Where, is the Softmax activation function, is the embedding vector, is the LeakyReLU activation function, for The transposed vector of

[0125] use Calculate the dot product similarity between nodes and use the LeakyReLU activation function to enhance the nonlinearity of the model and avoid gradient vanishing. Use the Softmax activation function to convert the similarity into a probability distribution for weighted aggregation of information from adjacent nodes.

[0126] The adaptive adjacency matrix obtained by the above method is a fully connected adjacency matrix. The fully connected adjacency matrix may contain a large number of weakly correlated or irrelevant edges (such as noise or accidental similarity). These edges will interfere with the model's capture of key relationships and easily introduce over-parameterization problems, resulting in model overfitting. Sparse processing is introduced in the graph learning module to only retain some node connection relationships with relatively high similarity, so that the model can focus more on information aggregation of strongly correlated nodes. For each node, the top 10% of nodes with high connection weights are retained. nodes, forming a node index set .

[0127] Constructing the mask matrix , mark the neighbor position retained by each node, the element calculation formula of M is:

[0128] ;

[0129] Where, For the Node and The connection relationship of the nodes, For the The index set of nodes;

[0130] In order to save memory and speed up calculation, the present invention innovatively proposes that the adaptive adjacency matrix after sparse processing Expressed as:

[0131] ;

[0132] Pair Matrix Extract the edge list of the sparse matrix and convert it from dense representation to sparse representation; extract the matrix All non-zero element coordinates in are connected as edges, and the correlation strength between nodes of all non-zero elements is used as edge weight ; Finally, the adaptive adjacency matrix is ​​converted into a vector containing two dimensions: edge connection relationship and edge weight, which is used as input to the subsequent graph attention module.

[0133] For example, the graph attention module (i.e., graph attention network GAT) models the spatial relationship of node feature information, and combines the edge connection relationship and edge weight parameters of the adaptive adjacency matrix to realize the interactive extraction of information features in the global space.

[0134] In the graph attention module GAT, the adaptive adjacency matrix is ​​associated with the graph attention coefficient matrix to capture the dynamic spatial correlation between nodes. The input of the graph attention module GAT is the node feature representation generated by temporal convolution. = , ,in, is the number of nodes, is the number of feature channels. In order to learn the input features in a deeper level, the weight matrix is ​​introduced , the weight matrix A parameterized shared linear transformation is applied to each node, and a shared attention mechanism is used between nodes. Calculate the importance between nodes , and its calculation formula is:

[0135] ;

[0136] Where, Node and nodes The eigenvector of

[0137] According to the calculated node importance, the nonlinear ability of the model is enhanced by the LeakyReLU function, and the Softmax function is used for normalization. At the same time, the edge weights generated by the adaptive adjacency matrix generated by the graph learning module are used. Perform weighted attention calculation and finally get the attention coefficient , and its calculation formula is:

[0138] ;

[0139] Where, For the nodes and The edge weights connecting the nodes, for index, For the nodes and The edge weights connecting the nodes;

[0140] Get the attention coefficient After that, the information of adjacent nodes is aggregated by weighted summation, and the feature representation of the current node is updated. A multi-head attention mechanism is introduced, where each attention head independently calculates the attention coefficient and node representation, and the outputs of multiple heads are spliced ​​together to enhance the expressive power of the model. The calculation formula is:

[0141] ;

[0142] Where, It is the feature concatenation output of multi-head attention. is the number of multi-head attention heads, For the The attention coefficient of the head, For the The weight matrix of each head; For multiple feature stitching, is the Sigmoid activation function;

[0143] To prevent gradient vanishing and enhance the stability and robustness of the model, the GAT output is residually connected and normalized to obtain the final output of the graph attention module. The calculation formula is:

[0144] ;

[0145] Where, is the final output of the graph attention module, ()for Normalization.

[0146] The three-dimensional temperature-salinity field reconstruction model based on graph attention neural network uses the Adam optimizer to adjust the model learning process and optimize the model performance. The parameters of the model are shown in Table 1.

[0147] Table 1 Model parameters

[0148]

[0149] It can be understood that the present invention innovatively proposes to use the edge weight w to calculate the attention coefficient, so that the attention coefficient fully reflects the relationship and importance of the features between different nodes.

[0150] In Example 2, a system for reconstructing a three-dimensional ocean temperature and salinity field using satellite remote sensing data provided by an embodiment of the present invention includes:

[0151] The sea surface satellite remote sensing data acquisition and preprocessing module is used to obtain the daily average sea surface satellite remote sensing data of the experimental area, perform interpolation processing to unify the spatial resolution of different data, perform standardization processing to eliminate the value differences between different data, and convert the grid data in the original sea surface satellite remote sensing data into graph node data; the sea surface satellite remote sensing data includes: sea surface height anomaly (SLA), sea surface temperature (SST), sea surface salinity (SSS) and sea surface wind (SSW), among which the sea surface wind is divided into the latitudinal component (USSW) and the longitudinal component (VSSW);

[0152] The ocean three-dimensional temperature and salinity field reconstruction model construction and processing module is used to input the processed sea surface satellite remote sensing data into the constructed ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolution network, and output the ocean three-dimensional temperature field or salinity field through temporal convolution, graph learning and graph attention processing;

[0153] The verification module is used to verify the spatial distribution consistency of the output ocean three-dimensional temperature field or salinity field reconstruction results with the ocean reanalysis data as the real data.

[0154] In order to further illustrate the effects of the embodiments of the present invention, a three-dimensional temperature-salinity field reconstruction experiment in a certain sea area is conducted.

[0155] Taking ocean reanalysis data as real observation values ​​and based on multi-source sea surface satellite remote sensing data (SST, SSS, SLA, SSW), the three-dimensional temperature-salinity field reconstruction model based on graph attention neural network constructed by the present invention was adopted to carry out an experiment on the reconstruction of the three-dimensional temperature-salinity field of a certain sea area.

[0156] The ocean reanalysis data is GLORYS12 from CMEMS, with a daily temporal resolution and a spatial resolution of (1 / 12)°×(1 / 12)°. The data provides physical variables such as seawater temperature and salinity at 50 different depth layers from 0 to 4000m in the global ocean, as well as sea surface height, mixed layer depth, and sea ice parameters.

[0157] During the data preprocessing phase, sea surface satellite remote sensing data and ocean reanalysis data were converted from two-dimensional grid data to node data to meet the input requirements of the graph neural network. To ensure consistent spatial resolution, the data was unified to a resolution of 0.25°×0.25° using a bilinear interpolation method. A time window of 10 was selected, and the three-dimensional ocean temperature and salinity field data for the 10th day were reconstructed using the daily average sea surface satellite remote sensing data from the previous 10 days. The daily average sea surface satellite remote sensing data from 2013 to 2017 was selected as training data, and the data from 2018 was used as test data to evaluate the accuracy and generalization of the training model. The data was standardized to have a mean of 0 and a variance of 1, expressed as:

[0158] ;

[0159] Where, For input data, is the mean value of the input data, is the variance of the input data.

[0160] In order to evaluate the results of the three-dimensional temperature-salinity field reconstruction model based on the graph attention neural network constructed in this paper, the reconstruction results of seawater temperature and salinity at different depths (30m, 65m, 110m, and 220m) were taken as examples to compare the spatial distribution of the reconstruction results with the reanalysis data. Figure 4-Figure 19 As shown in the figure, it can be seen that the reconstruction results of seawater temperature and salinity at different depths are relatively consistent with the spatial distribution of the reanalysis data. The ocean 3D temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network can accurately reflect the real ocean 3D temperature and salinity field.

[0161] To further verify the stability of the model reconstruction structure with depth and time, the RMSE of the seawater temperature and salinity reconstruction results with depth and time were calculated and analyzed. Figure 20 The RMSE of the temperature field reconstruction results changes with depth as shown in the figure. The RMSE of the temperature field reconstruction results changes with time as shown in the figure Figure 21 The RMSE of the temperature field reconstruction results changes with time as shown in the figure. Figure 22 RMSE of salinity field reconstruction results changes with depth; Figure 23 The RMSE of the salinity field reconstruction results changes with time; Figure 21 and Figure 23 As can be seen, the overall error of the temperature and salinity reconstruction results is relatively stable over time. The RMSE values ​​at depths of 30m, 110m, and 222m vary smoothly and steadily over time. The reconstruction results at a depth of 65m, near the thermocline, are slightly worse than those at other depths, but still maintain a relatively stable trend over time. The three-dimensional temperature-salinity field reconstruction model based on the graph attention neural network constructed in this paper can effectively capture the temporal variations of seawater temperature and salinity, effectively reflecting the true three-dimensional ocean temperature-salinity field.

[0162] The above description is only a preferred specific implementation method of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data, characterized in that: This method considers the spatiotemporal correlation between the three-dimensional ocean temperature and salinity field and sea surface satellite remote sensing data, uses a graph attention network to establish the spatial correlation between the three-dimensional ocean temperature and salinity field and a temporal convolutional network to establish the temporal correlation between the three-dimensional ocean temperature and salinity field and the sea surface satellite remote sensing data. The specific steps include: S1. Obtain the daily average sea surface satellite remote sensing data for the experimental area and perform interpolation processing to unify the spatial resolution of different data, perform standardization processing, and eliminate the value differences between different data; convert the grid data in the sea surface satellite remote sensing data into graph node data; the sea surface satellite remote sensing data includes: sea surface height anomaly SLA, sea surface temperature SST, sea surface salinity SSS and sea surface wind SSW, and the sea surface wind is divided into the latitudinal component USSW and the longitudinal component VSSW; S2, the processed sea surface satellite remote sensing data is input into the constructed ocean 3D temperature and salinity field reconstruction model based on graph attention network and temporal convolution network, and after temporal convolution, graph learning and graph attention processing, the 3D ocean temperature field or salinity field is output; S3, verify the spatial distribution consistency of the output ocean three-dimensional temperature field or salinity field reconstruction results with the ocean reanalysis data used as the real data; In step S2, the ocean 3D temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network consists of a multi-layer stack of temporal convolution modules, graph learning modules, and graph attention modules. The input sea surface feature sequence is first convolved with 1×1 to change the number of channels, converting low-channel sea surface features into higher-dimensional feature representations. The temporal convolution module is used to extract the temporal information features of sea surface satellite remote sensing data. It includes dual-path multi-scale dilated convolution and channel attention, and extracts features through two independent paths. The first path introduces the global channel attention CAM module after multi-scale dilated convolution to capture global dependencies. The second path introduces lightweight channel attention ECA after multi-scale dilated convolution to process local feature dependencies.

2. The method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 1, characterized in that: In step S1, the daily average sea surface satellite remote sensing data is processed by interpolation to a uniform spatial resolution of 0.25 °× 0.25 ° After the spatial resolution of the data is unified, the data is standardized with a mean of 0 and a variance of 1 to eliminate the differences in the values ​​of different data.

3. The method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 1, characterized in that: The temporal convolution module uses multi-scale convolution to effectively capture multi-scale temporal features through a variety of different convolution kernels; Four convolution kernels of 1×2, 1×3, 1×5, and 1×7 are selected to extract temporal features of different scales. The output features of each dilated convolution are spliced ​​in the channel dimension to form a fused multi-scale feature representation as the output of the multi-scale dilated convolution. The calculation formula of the multi-scale dilated convolution is: ； Where, is the feature input of the temporal convolutional network, is the convolution operation, is the convolution kernel, is the convolution kernel size,; is the convolution kernel The element in is the input eigenvalue, is the time length, is the expansion factor, and is the step size; ； Where, is the output after feature splicing, For feature splicing, is a 1×2 convolution kernel, is a 1×3 convolution kernel, is a 1×5 convolution kernel, is a 1×7 convolution kernel; In the first path of temporal convolution, the global channel attention CAM module is introduced to capture global dependencies, globally pool each channel generated by multi-scale dilated convolution, compress the node dimension and time dimension of each channel, extract global information and reduce the amount of calculation. The calculation formula is: ； Where, is the global spatiotemporal information representation of the th channel, is the number of nodes, is the feature representation of the th node and th time of the th channel; Two different fully connected layers are used to perform nonlinear transformations on the globally pooled features to learn the complex dependencies between channels. The first fully connected layer is responsible for channel compression, and the second fully connected layer restores the number of channels to the original dimension. ； Where, is the output of two fully connected layers, is the weight matrix of the second fully connected layer, is the ReLU activation function, is the weight matrix of the first fully connected layer; The output is normalized using the Softmax activation function, the channel weight is converted into a probability distribution, the channel attention weight is generated, and the generated channel attention weight is multiplied by the original feature to obtain the weighted feature , dynamically adjust the feature strength of each channel and introduce Dropout to prevent overfitting; In the second path of temporal convolution, lightweight channel attention (ECA) is introduced to perform average pooling on each channel generated by multi-scale dilated convolution to compress the spatiotemporal feature dimensions, and use 1D convolution to capture the local dependency of adjacent channels. The calculation formula is: ； Where, is the output of 1D convolution, is the weight parameter of the 1D convolution kernel, is the feature vector output after average pooling processing; The Sigmoid activation function is used to generate channel attention weights from the 1D convolution output results to dynamically adjust the channel importance; the generated channel attention is multiplied by the original feature to obtain the weighted feature ; The weighted features generated by the first and second paths of the temporal convolution are respectively fused after the Sigmoid activation function and the Tanh activation function to obtain the complete temporal features; the final output of the temporal convolution module is: ； Where, is the output of the temporal convolution module, is the Sigmoid activation function, is the Hadamard product, is the tanh activation function, are the outputs of the two paths respectively.

4. The method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 1, characterized in that: The graph learning module randomly initializes the embedding vectors of all nodes , is dynamically optimized and updated with model training to reflect more realistic ocean space similarity characteristics, is the number of nodes, is the node embedding dimension, and is the real number space; The adaptive adjacency matrix A is defined as: ； Where, is the Softmax activation function, is the embedding vector, is the LeakyReLU activation function, for The transposed vector of use Calculate the dot product similarity between nodes and use the LeakyReLU activation function to enhance the nonlinearity of the model and avoid gradient vanishing. Use the Softmax activation function to convert the similarity into a probability distribution for weighted aggregation of information from adjacent nodes. The adaptive adjacency matrix obtained is a fully connected adjacency matrix. Sparse processing is introduced to retain only the connection relationships of nodes with relatively high similarity. For each node, the top 10% of nodes with high connection weights are retained. nodes, forming a node index set ; Constructing the mask matrix ,, mark the neighbor positions retained by each node, The element calculation formula is: ； Where, is the connection relationship between the th node and the th node, is the index set of the th node; Adaptive adjacency matrix after sparse processing Expressed as: ； Pair Matrix Extract the edge list of the sparse matrix and convert it from dense representation to sparse representation; extract the matrix All non-zero element coordinates in are connected as edges, and the correlation strength between nodes of all non-zero elements is used as edge weights ; Finally, the adaptive adjacency matrix is ​​converted into a vector containing two dimensions: edge connection relationship and edge weight, which is used as input to the subsequent graph attention module.

5. The method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 4, characterized in that: The graph attention module associates the adaptive adjacency matrix with the graph attention coefficient matrix to jointly capture the dynamic spatial correlation between nodes; the input of the graph attention module is the node feature representation generated by temporal convolution , ,in is the number of nodes, is the number of feature channels; introduce the weight matrix , the weight matrix A parameterized shared linear transformation is applied to each node, and a shared attention mechanism is used between nodes. Calculate the importance between nodes , and its calculation formula is: ； Where, Node and nodes The eigenvector of According to the calculated node importance, the nonlinear ability of the model is enhanced by the LeakyReLU function, and the Softmax function is used for normalization; Combined with the adaptive adjacency matrix generated by the graph learning module, the edge weights generated by the adaptive adjacency matrix are used Perform weighted attention calculation to obtain the attention coefficient , the calculation formula is: ； Where, For the nodes and The edge weights connecting the nodes, for index, For the The edge weight connecting the node and the node; Get the attention coefficient After that, the information of adjacent nodes is aggregated by weighted summation, and the feature representation of the current node is updated. A multi-head attention mechanism is introduced, and each attention head independently calculates the attention coefficient and node representation. The outputs of multiple heads are spliced ​​to enhance the expression ability of the model. The calculation formula is: ； Where, It is the feature concatenation output of multi-head attention. is the number of multi-head attention heads, For the The attention coefficient of the head, For the The weight matrix of each head; For multiple feature stitching, is the Sigmoid activation function; After the output of the graph attention module is residually connected and normalized, the final output of the graph attention module is obtained. The calculation formula is: ； Where, is the final output of the graph attention module, for Normalization.

6. The method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 1, characterized in that: In step S2, the ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network uses the Adam optimizer to adjust the model learning process and optimize the model performance.

7. A system for reconstructing three-dimensional ocean temperature and salinity fields using satellite remote sensing data, characterized in that: The method for reconstructing a three-dimensional ocean temperature and salinity field using satellite remote sensing data as described in any one of claims 1 to 6 is implemented, the system comprising: The sea surface satellite remote sensing data acquisition and preprocessing module is used to obtain the daily average sea surface satellite remote sensing data of the experimental area, perform interpolation processing to unify the spatial resolution of different data, perform standardization processing to eliminate the value differences between different data, and convert the grid data in the original sea surface satellite remote sensing data into graph node data; the sea surface satellite remote sensing data includes: sea surface height anomaly SLA, sea surface temperature SST, sea surface salinity SSS and sea surface wind SSW, where the sea surface wind is divided into latitudinal component USSW and longitudinal component VSSW; The ocean three-dimensional temperature and salinity field reconstruction model construction and processing module is used to input the processed sea surface satellite remote sensing data into the constructed ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolution network, and output the ocean three-dimensional temperature field or salinity field through temporal convolution, graph learning and graph attention processing; The verification module is used to verify the spatial distribution consistency of the output ocean three-dimensional temperature field or salinity field reconstruction results with the ocean reanalysis data as the real data.

8. The system for reconstructing three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 7, characterized in that: The system is mounted on a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the functions of the system can be realized.

Images