Method and system for reconstructing ocean three-dimensional temperature and salt field by using satellite remote sensing data

Abstract

The invention belongs to the technical field of ocean three-dimensional temperature and salt field data processing, and discloses a method and a system for reconstructing an ocean three-dimensional temperature and salt field by utilizing satellite remote sensing data. The method comprises the following steps: acquiring daily average sea surface satellite remote sensing data of an experimental area, and performing interpolation processing and standardization processing; inputting 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 time convolution network, and outputting an ocean three-dimensional temperature field or salinity field through time convolution, graph learning and graph attention processing; and carrying out spatial distribution consistency verification on the output ocean three-dimensional temperature field or salinity field reconstruction result and ocean reanalysis data serving as real data. According to the method, a sea area is taken as an experimental area, a sea three-dimensional temperature and salt field reconstruction experiment based on sea surface satellite remote sensing data is carried out, and the effectiveness and feasibility of the provided model method are evaluated.

Description

Technical Field

[0001] The present invention belongs to the technical field of marine three-dimensional temperature and salinity field data processing, and particularly relates to a method and system for reconstructing a marine three-dimensional temperature and 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 related to its movement characteristics, ocean dynamics, and ecological processes. Information on the physical characteristics of seawater plays a crucial role in aspects such as marine environmental protection, marine resource development, and ensuring the safety of maritime navigation. Seawater temperature and salinity are basic elements of oceanography and are closely related to almost all ocean phenomena and dynamic processes. The spatio-temporal distribution characteristics and their changes constitute one of the core contents of marine scientific research. Although satellite remote sensing data is rich, it can only obtain sea surface information. Summary of the Invention

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

[0004] The technical solution is as follows: A method for reconstructing a marine three-dimensional temperature and salinity field by using satellite remote sensing data, considering the spatio-temporal correlation between the marine three-dimensional temperature and salinity field and sea surface satellite remote sensing data, uses a graph attention network to establish the spatial correlation between sea surface satellite remote sensing data and the marine three-dimensional temperature and salinity field, and uses a temporal convolutional network to establish the temporal correlation between sea surface satellite remote sensing data and the marine three-dimensional temperature and salinity field; specifically includes the following steps: S1, 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 magnitude differences between different data; convert the grid data in the sea surface satellite remote sensing data into graph node data; wherein, 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 zonal component USSW and meridional component VSSW; S2, Input the processed sea surface satellite remote sensing data into the constructed marine three-dimensional temperature and salinity field reconstruction model based on a graph attention network and a temporal convolutional network, and after temporal convolution, graph learning, and graph attention processing, output the marine three-dimensional temperature field or salinity field; S3, Verify the spatial distribution consistency of the reconstructed result of the output marine three-dimensional temperature field or salinity field with the ocean reanalysis data as the real data.

[0005] In step S1, the daily average sea surface satellite remote sensing data is interpolated to a unified spatial resolution of 0.25 °× 0.25 ° ; After unifying the data space resolution, the data is standardized with a mean of 0 and a variance of 1 to eliminate the magnitude differences between different data.

[0006] In step S2, the ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network consists of multiple stacks of a temporal convolutional module, a graph learning module, and a graph attention module. The input sea surface feature sequence is first changed in the number of channels by a 1×1 convolution, converting the low-channel sea surface features into higher-dimensional feature representations.

[0007] Furthermore, the temporal convolutional module is used to extract the temporal information features of sea surface satellite remote sensing data, including dual-path multi-scale dilated convolution and channel attention. Features are extracted through two independent paths. The first path introduces a global channel attention CAM module after multi-scale dilated convolution to capture global dependencies, and the second path introduces a lightweight channel attention ECA after multi-scale dilated convolution to process local feature dependencies.

[0008] Furthermore, the temporal convolutional module effectively captures multi-scale temporal features using multi-scale convolution with various different convolutional kernels; four convolutional kernels of 1×2, 1×3, 1×5, and 1×7 are selected to extract temporal features at different scales. The output features of each dilated convolution are concatenated in the channel dimension to form a fused multi-scale feature representation, which serves as the output of the multi-scale dilated convolution. The formula for the multi-scale dilated convolution is: ; In the formula, is the feature input of the temporal convolutional network, is the convolution operation, is the convolutional kernel, is the convolutional kernel size, ; is the th element in the convolutional kernel , is the input feature value, is the time length, is the dilation factor, is the stride; ; In the formula, is the output after feature concatenation, is the feature concatenation, is the 1×2 convolutional kernel, is the 1×3 convolutional kernel, is the 1×5 convolutional kernel, is the 1×7 convolutional kernel; In the first path of the temporal convolution, a global channel attention CAM module is introduced to capture global dependencies. Global pooling is performed on each channel generated by the multi-scale dilated convolution to compress the node dimension and temporal dimension of each channel, extract global information, and reduce the computational complexity. The calculation formula is as follows: ; In the formula, is the global spatio-temporal information representation of the -th channel, is the number of nodes, is the -th channel, is the -th node, and is the feature representation of the -th time; In the formula, is the output after passing through 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 to convert the channel weights into a probability distribution, generating channel attention weights. The generated channel attention weights are multiplied by the original features to obtain weighted features , dynamically adjusting the feature intensity of each channel, and Dropout is introduced to prevent overfitting; In the second path of the temporal convolution, lightweight channel attention ECA is introduced. Average pooling is performed on each channel generated by the multi-scale dilated convolution to compress the spatio-temporal feature dimension, and 1D convolution is used to capture the local dependencies between adjacent channels. The calculation formula is as follows: ; In the formula, is the output after 1D convolution, is the weight parameter of the 1D convolution kernel, is the feature vector output after average pooling; The Sigmoid activation function is used to generate channel attention weights from the output of the 1D convolution, dynamically adjusting the channel importance. The generated channel attention is multiplied by the original features to obtain weighted features ; The weighted features generated by the first and second paths of the temporal convolution are respectively operated on by the Sigmoid activation function and the Tanh activation function and then fused to obtain the complete temporal features; the final output of the temporal convolution module is: ; In the formula, 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.

[0009] Furthermore, the graph learning module randomly initializes the embedding vectors of all nodes , and dynamically optimizes and updates them during model training to reflect more real ocean space similarity features, is the number of nodes, is the node embedding dimension, is the real number space; The adaptive adjacency matrix is defined as: ; In the formula, is the Softmax activation function, is the embedding vector, is the LeakyReLU activation function, is the transposed vector of; Use to calculate the dot product similarity between nodes, and use the LeakyReLU activation function to enhance the nonlinear ability of the model and avoid gradient disappearance; convert the similarity into a probability distribution through the Softmax activation function for weighted aggregation of information of adjacent nodes; The obtained adaptive adjacency matrix is a fully connected adjacency matrix. Sparsification processing is introduced to only retain the connection relationships of some nodes with relatively high similarity. For each node, retain the top 10% of nodes with higher connection weights to form the node index set ; Construct a mask matrix M to mark the positions of the retained neighbors of each node. The calculation formula for the elements of M: ; In the formula, is the connection relationship between the th node and the th node, is the index set of the th node; Adaptive adjacency matrix after sparsification processing It is expressed as: ; For the matrix perform edge list extraction of the sparse matrix, converting from the dense representation to the sparse representation; extract all non-zero element coordinates in the matrix as the edge connections, and the node correlation strength between all non-zero elements as the edge weights ; Finally, convert the adaptive adjacency matrix into a vector containing two dimensions of edge connection relationship and edge weight, which is used as the input to the subsequent graph attention module.

[0010] 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 , , where is the number of nodes, is the number of feature channels; introduce the weight matrix , and apply the parameterized shared linear transformation of the weight matrix to each node, and calculate the importance between nodes through the shared attention mechanism , and its calculation formula is: ; In the formula, are the feature vectors of nodes and node respectively; According to the calculated node importance, enhance the non-linear ability of the model through the LeakyReLU function and perform normalization using the Softmax function; combined with the adaptive adjacency matrix generated by the graph learning module, use the edge weight generated by the adaptive adjacency matrix to perform weighted attention calculation to obtain the attention coefficient , and its calculation formula is: ; In the formula, is the edge weight connecting the th node and the th node, is the exponent, is the edge weight connecting the th node and the th node; Obtain the attention coefficient After that, aggregate the information of adjacent nodes through weighted summation and update the feature representation of the current node; introduce the multi-head attention mechanism, where each attention head independently calculates the attention coefficients and node representations, and splice the outputs of multiple heads to enhance the expressive power of the model. The calculation formula is as follows: ; In the formula, is the feature splicing output of the multi-head attention, is the number of multi-head attention heads, is the th attention coefficient of the head, is the th weight matrix of the head; is the splicing of multiple features, is the Sigmoid activation function; After the output result of the graph attention module is subjected to residual connection and normalization, the final output of the graph attention module is obtained. The calculation formula is as follows: ; In the formula, is the final output of the graph attention module, is normalization.

[0011] 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.

[0012] Another object of the present invention is to provide a system for reconstructing the ocean three-dimensional temperature and salinity field using satellite remote sensing data. This system implements the method for reconstructing the ocean three-dimensional temperature and salinity field using satellite remote sensing data. The system includes: The sea surface satellite remote sensing data acquisition and preprocessing module is used to acquire the daily 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 magnitude 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 the zonal component USSW and the meridional 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 convolutional network, and through temporal convolution, graph learning, and graph attention processing, output the ocean three-dimensional temperature field or salinity field; A verification module is used to verify the spatial distribution consistency between the reconstructed results of the three-dimensional ocean temperature field or salinity field output and the ocean reanalysis data used as real data.

[0013] Furthermore, the system is carried on a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the functions in the above system can be realized.

[0014] Combining all the above technical solutions, the beneficial effects of the present invention are as follows: To make full use of rich sea surface remote sensing data and achieve global perception of spatio-temporal information, the present invention combines a graph attention neural network with a temporal convolution module to construct a method for reconstructing the three-dimensional ocean temperature and salinity field based on a graph attention neural network, realizing the reconstruction of the three-dimensional ocean temperature and salinity field through sea surface satellite remote sensing data. The input data of the constructed three-dimensional temperature and salinity field reconstruction model includes 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 the output data is the three-dimensional ocean temperature and salinity field. Taking a certain sea area as the experimental area, an experiment on reconstructing the three-dimensional ocean temperature and salinity field based on sea surface satellite remote sensing data is carried out to evaluate the effectiveness and feasibility of the proposed model method. Description of the Drawings

[0015] The drawings here are incorporated into the specification and form a part of this specification, showing the embodiments consistent with the present disclosure and used together with the specification to explain the principles of the present disclosure; Figure 1 It is a flowchart of the method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data provided by an embodiment of the present invention; Figure 2 It is an overall structure diagram of the three-dimensional ocean temperature and salinity field reconstruction model based on a graph attention network and a temporal convolution network provided by an embodiment of the present invention; Figure 3 It is a structure diagram of the temporal convolution module provided by an embodiment of the present invention; Figure 4 It is a reconstructed result diagram of the seawater temperature at a depth of 30m using the present invention; Figure 5 It is a reconstructed result diagram of the seawater temperature at a depth of 65m using the present invention; Figure 6 It is a reconstructed result diagram of the seawater temperature at a depth of 110m using the present invention; Figure 7 It is a reconstructed result diagram of the seawater temperature at a depth of 222m using the present invention; Figure 8 It is a reconstructed result diagram of the seawater temperature at a depth of 30m of ocean reanalysis data; Figure 9 It is a reconstructed result diagram of the seawater temperature at a depth of 65m of ocean reanalysis data; Figure 10 It is a reconstructed result diagram of seawater temperature at a depth of 110m in ocean reanalysis data; Figure 11 It is a reconstructed result diagram of seawater temperature at a depth of 222m in ocean reanalysis data; Figure 12 It is a reconstructed result diagram of seawater salinity at a depth of 30m using the present invention; Figure 13 It is a reconstructed result diagram of seawater salinity at a depth of 65m using the present invention; Figure 14 It is a reconstructed result diagram of seawater salinity at a depth of 110m using the present invention; Figure 15 It is a reconstructed result diagram of seawater salinity at a depth of 222m using the present invention; Figure 16 It is a reconstructed result diagram of seawater salinity at a depth of 30m in ocean reanalysis data; Figure 17 It is a reconstructed result diagram of seawater salinity at a depth of 65m in ocean reanalysis data; Figure 18 It is a reconstructed result diagram of seawater salinity at a depth of 110m in ocean reanalysis data; Figure 19 It is a reconstructed result diagram of seawater salinity at a depth of 222m in ocean reanalysis data; Figure 20 It is a diagram showing the variation of RMSE of the reconstructed result of the temperature field of the present invention with depth; Figure 21 It is a diagram showing the variation of RMSE of the reconstructed result of the temperature field of the present invention with time; Figure 22 It is a diagram showing the variation of RMSE of the reconstructed result of the salinity field of the present invention with depth; Figure 23 It is a diagram showing the variation of RMSE of the reconstructed result of the salinity field of the present invention with time. Detailed implementation manners

[0016] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following will describe the detailed implementation manners of the present invention in conjunction with the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific implementations disclosed below.

[0017] The innovation of the present invention lies in: The present invention takes into account the spatio-temporal 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 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 sea surface satellite remote sensing data and the three-dimensional ocean temperature and salinity field.

[0018] Example 1, as Figure 1 shown, the method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data provided by an embodiment of the present invention includes: S1. Obtain the daily sea surface satellite remote sensing data of the experimental area, perform interpolation processing, unify the spatial resolution of different data, perform standardization processing to eliminate the magnitude difference between different data; convert the grid data in the sea surface satellite remote sensing data into graph node data; wherein, 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 zonal component (USSW) and meridional component (VSSW); S2. Input the processed sea surface satellite remote sensing data into the three-dimensional ocean temperature and salinity field reconstruction model constructed based on the graph attention network and the temporal convolutional network, and after temporal convolution, graph learning, and graph attention processing, output the three-dimensional ocean temperature field or salinity field; S3. Verify the spatial distribution consistency of the reconstructed result of the output three-dimensional ocean temperature field or salinity field with the ocean reanalysis data as the real data.

[0019] Exemplarily, 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 ocean environment of this sea area make it a key area for studying the interaction between the ocean and climate. The method for reconstructing the three-dimensional ocean temperature and salinity field based on the graph attention neural network uses sea surface satellite remote sensing data to reconstruct the three-dimensional ocean temperature and salinity field of a certain sea area. 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 zonal component (USSW) and meridional component (VSSW).

[0020] The SLA data is altimeter L4 grid data provided by the Copernicus Marine Environment Monitoring Service Center (CMEMS), and the time resolution of this data is daily, and the spatial resolution is 0.25°.

[0021] The SST data is microwave and infrared fusion data provided by the REMSS remote sensing system, and the time resolution of this data is daily, and the spatial resolution is 0.25°.

[0022] SSW data is the sea surface wind field fusion data developed and released by the Remote Sensing System (RSS). The time resolution of this data is 6 hours, and the spatial resolution is 0.25°.

[0023] SSS data is the multi-source sea surface salinity fusion data developed by CNR of Italy and provided by CMEMS. The time resolution of this data is daily, and the spatial resolution is 0.125°.

[0024] All data used in the present invention are daily average data, and the time range is from 2013 to 2018. The data used are processed by interpolation to a unified spatial resolution of 0.25 °× 0.25 ° °. After the data spatial resolution is unified, the data is normalized so that its mean is 0 and variance is 1 to eliminate the magnitude differences between different data. Before the original data is input into the model, the original grid data needs to be converted into graph node data to adapt to the graph neural network model.

[0025] Exemplarily, in step S2, most existing existing ocean three-dimensional temperature and salinity field reconstruction methods only consider the time feature information or spatial feature information of sea surface data and the amount of information is relatively limited. The receptive field of models represented by convolutional neural networks is limited, and it is difficult to perform global perception of spatio-temporal information. The present invention combines the Graph Attention Network (GAT) with the Temporal Convolutional Network (TCN), fully integrates the spatio-temporal information characteristics of sea surface remote sensing data, constructs 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, and realizes the 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 Figure 2 shown.

[0026] Among them, the ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network is mainly composed of multiple layers of stacking of a temporal convolution module, a graph learning module, and a graph attention module. The input data of the ocean three-dimensional temperature and salinity field reconstruction model based on the graph attention network and the temporal convolutional network are sea surface height anomaly (SLA), sea surface temperature (SST), sea surface salinity (SSS), and sea surface wind (SSW), and the output data are the ocean three-dimensional temperature field or salinity field. The input sea surface feature sequence is first changed in the number of channels by 1×1 convolution, and the low-channel sea surface features are transformed into higher-dimensional feature representations.

[0027] Exemplarily, the temporal convolution module focuses on extracting the time information characteristics of sea surface satellite remote sensing data, including multi-scale dilated convolution and channel attention with a dual path. Features are extracted through two independent paths to reduce the model's dependence on a single feature and improve the model's robustness. The structure of the temporal convolution module is as Figure 3As shown. The first path introduces a global channel attention (CAM) module after multi-scale dilated convolution to capture global dependencies, and the second path introduces lightweight channel attention (ECA) after multi-scale dilated convolution to process local feature dependencies and enhance the model's expressive power. The dual-path network fuses mutual features, which can not only capture the complex dependencies of the global temperature-salinity field but also efficiently process local features, thus improving the accuracy and efficiency of three-dimensional temperature-salinity field reconstruction.

[0028] The multi-scale dilated convolution uses causal convolution, making the current output depend on the current and past inputs to satisfy the principle of time dependence. Considering that the receptive field of causal convolution is limited by the convolution kernel size, but constructing long-term memory requires a large receptive field, so dilated causal convolution is adopted, which can expand the receptive field while reducing the model's computational complexity. Ocean data contains various time-scale feature information, and a single convolution kernel is difficult to extract complete time feature modes. Therefore, multi-scale convolution is used to effectively capture multi-scale time features through multiple 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 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 of the multi-scale dilated convolution is: ; In the formula, is the feature input of the time convolutional network, is the convolution operation, is the convolution kernel, is the convolution kernel size, ; is the th element in the convolution kernel , is the input feature value, is the time length, is the dilation factor, is the stride;

[0029] In the formula, is the output after feature concatenation, is the feature concatenation, is the 1×2 convolution kernel, is the 1×3 convolution kernel, is the 1×5 convolution kernel, is the 1×7 convolution kernel; Long - term dependencies can be captured through multi - scale dilated convolutions, but the weight assignment to features is uniform, and it is unable to distinguish relatively important channel features. Therefore, by introducing a channel attention mechanism to learn the importance weights of each channel, the feature channels are weighted, so as to highlight important features and suppress unimportant features, and enhance the model's attention to key information.

[0030] In the first path of the temporal convolution, a global channel attention (CAM) module is introduced to capture global dependencies. Global pooling is performed on each channel generated by the multi - scale dilated convolution to compress the node dimension and temporal dimension of each channel, extract global information, and reduce the computational amount. The calculation formula is: ; In the formula, is the global spatio - temporal information representation of the -th channel, is the number of nodes, is the -th channel's feature representation of the -th node and the -th time; Non - linear transformation is performed on the features after global pooling through two different fully - connected layers to learn the complex dependencies between channels. The first fully - connected layer is responsible for compressing the number of channels, and the second fully - connected layer restores the number of channels to the initial dimension.

[0031] ; In the formula, is the output after passing through 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 to convert the channel weights into a probability distribution, generate channel attention weights, 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 the weighted features , dynamically adjusting the feature intensity of each channel. Dropout is introduced to prevent overfitting.

[0032] In the second path of the temporal convolution, lightweight channel attention (ECA) is introduced. Average pooling is performed on each channel generated by the multi - scale dilated convolution to compress the spatio - temporal feature dimension, and 1D convolution is used to capture the local dependencies of adjacent channels, enhancing the model's expressive ability. The calculation formula is: ; In the formula, is the output after passing through 1D convolution, is the weight parameter of the 1D convolutional kernel, is the feature vector output after average pooling processing; The Sigmoid activation function is used to generate channel attention weights from the 1D convolutional output results, dynamically adjusting channel importance; multiplying the generated channel attention by the original features to obtain weighted features ; The weighted features generated by the first and second paths of the temporal convolution are respectively fused after being operated by the Sigmoid activation function and the Tanh activation function to obtain complete temporal features; the final output of the temporal convolution module is: ; In the formula, 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.

[0033] Exemplarily, the purpose of the graph learning module is to construct an adaptive adjacency matrix that adapts to the complex and changeable ocean, for calculating graph attention weight coefficients. The graph learning module constructs an adjacency matrix reflecting the correlation between nodes by introducing learnable random embedding nodes, continuously improves its own parameters as the model is trained, gradually adapts to the changes in the spatial correlation between nodes during the model learning process, and models the stable long-term spatial correlation to obtain the optimal adaptive adjacency matrix.

[0034] Aiming at the problem that the existing construction of the adjacency matrix 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, dynamically adjusts the node relationships, and more realistically reflects the spatial relationships of the ocean dynamic process. The graph learning module first randomly initializes the embedding vectors of all nodes , which are dynamically optimized and updated as the model is trained to reflect more real ocean spatial similarity features, where is the number of nodes, is the node embedding dimension, is the real number space; The adaptive adjacency matrix is defined as: ; In the formula, is the Softmax activation function, is the embedding vector, is the LeakyReLU activation function, is The transposed vector; Use Calculate the dot product similarity between nodes, use the LeakyReLU activation function to enhance the nonlinear ability of the model and avoid gradient vanishing; convert the similarity into a probability distribution through the Softmax activation function for weighted aggregation of information of adjacent nodes; 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 related or irrelevant edges (such as noise or accidental similarities), which will interfere with the model's capture of key relationships and are prone to introducing over-parameterization problems, resulting in model overfitting. Introduce sparsification processing in the graph learning module, only retain part of the node connection relationships with relatively high similarities, so that the model can focus more on the information aggregation of strongly associated nodes. Retain the top 10% of the nodes with higher connection weights for each node to form a node index set .

[0035] Construct a mask matrix , mark the positions of the retained neighbors of each node, and the calculation formula for the elements of M: ; In the formula, is the connection relationship between the th node and the th node, is the index set of the th node; To save memory and accelerate calculations, the present invention innovatively proposes that the sparsified adaptive adjacency matrix is expressed as: ; For the matrix , perform edge list extraction of the sparse matrix, convert from the dense representation to the sparse representation; extract all non-zero element coordinates in the matrix as edge connections, and the node correlation strength between all non-zero elements as edge weights ; finally, convert the adaptive adjacency matrix into a vector containing two dimensions of edge connection relationships and edge weights for input into the subsequent graph attention module.

[0036] Exemplarily, the graph attention module (i.e., the graph attention network GAT) models the spatial relationship of node feature information, combines the edge connection relationship and edge weight parameters of the adaptive adjacency matrix, and realizes the extraction of information feature interaction in the global space.

[0037] In the graph attention module GAT, the adaptive adjacency matrix is associated with the graph attention coefficient matrix to jointly capture the dynamic spatial correlation between nodes. The input of the graph attention module GAT is the node feature representation generated by temporal convolution. = , , where is the number of nodes, is the number of feature channels. To perform deeper learning on the input features, the weight matrix is introduced, and the parameterized shared linear transformation of the weight matrix is applied to each node. The importance between nodes is calculated through the shared attention mechanism , and its calculation formula is: ; In the formula, are the feature vectors of nodes and node respectively; According to the calculated node importance, the nonlinear ability of the model is enhanced through the LeakyReLU function, and the Softmax function is used for normalization. At the same time, combined with the adaptive adjacency matrix generated by the graph learning module, the edge weights generated by the adaptive adjacency matrix are used for weighted attention calculation, and finally the attention coefficient is obtained, and its calculation formula is: ; In the formula, is the edge weight connecting the -th node and the -th node, is exponent, is the edge weight connecting the -th node and the -th node; After obtaining the attention coefficient , the information of adjacent nodes is aggregated by weighted summation, and the feature representation of the current node is updated. The multi-head attention mechanism is introduced. Each attention head independently calculates the attention coefficient and the node representation, and the outputs of multiple heads are concatenated to enhance the expression ability of the model. The calculation formula is: ; In the formula, is the feature concatenation output of multi-head attention, is the number of multi-head attention heads, is the attention coefficient of the -th head, is the The weight matrix of the head; For splicing multiple features, Is the Sigmoid activation function; To prevent gradient vanishing and enhance the stability and robustness of the model, the output result of GAT is connected by residual and normalized to obtain the final output of the graph attention module. The calculation formula is: ; In the formula, Is the final output of the graph attention module, () is Normalization.

[0038] The three-dimensional temperature-salinity field reconstruction model based on the 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.

[0039] Table 1 Model parameter table

[0040] It can be understood that the present invention innovatively proposes to calculate the attention coefficient using the edge weight w, so that the attention coefficient can fully reflect the relationship and importance of features between different nodes.

[0041] Example 2, the system for reconstructing the three-dimensional ocean temperature-salinity field provided by the embodiment of the present invention includes: The sea surface satellite remote sensing data acquisition and preprocessing module is used to acquire the daily sea surface satellite remote sensing data in the experimental area, perform interpolation processing to unify the spatial resolution of different data, perform standardization processing to eliminate the magnitude difference 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 zonal component (USSW) and meridional component (VSSW); The three-dimensional ocean temperature-salinity field reconstruction model construction and processing module is used to input the processed sea surface satellite remote sensing data into the constructed three-dimensional ocean temperature-salinity field reconstruction model based on the graph attention network and the time convolutional network, and output the ocean three-dimensional temperature field or salinity field after time convolution, graph learning, and graph attention processing; The verification module is used to verify the spatial distribution consistency of the reconstructed results of the output ocean three-dimensional temperature field or salinity field with the ocean reanalysis data as the real data.

[0042] To further illustrate the relevant effects of the embodiments of the present invention, a three-dimensional temperature-salinity field reconstruction experiment in a certain sea area.

[0043] Taking ocean reanalysis data as the true observations, based on multi-source sea surface satellite remote sensing data (SST, SSS, SLA, SSW), the three-dimensional temperature and salinity field reconstruction model based on graph attention neural network constructed by the present invention is adopted to carry out the ocean three-dimensional temperature and salinity field reconstruction experiment in a certain sea area.

[0044] The ocean reanalysis data is GLORYS12 from CMEMS. The time resolution of this data is daily, and the spatial resolution is (1 / 12)°×(1 / 12)°. This 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.

[0045] In the data preprocessing stage, the sea surface satellite remote sensing data and ocean reanalysis data are converted from two-dimensional grid data into node data to meet the input requirements of the graph neural network. To ensure consistent spatial resolution, the bilinear interpolation method is used to unify the data to a resolution of 0.25°×0.25°. The time window is selected as 10, and the daily average sea surface satellite remote sensing data of the first 10 days is used to reconstruct the ocean three-dimensional temperature and salinity field data of the 10th day. The daily average sea surface satellite remote sensing data from 2013 to 2017 is selected as the training data, and the data in 2018 is used as the test data to evaluate the accuracy and generalization of the training model. The data is standardized so that its mean is 0 and variance is 1, expressed as: ; In the formula, is the input data, is the average value of the input data, is the variance of the input data.

[0046] To evaluate the results of the three-dimensional temperature and salinity field reconstruction model based on graph attention neural network constructed by the present invention, taking the reconstruction results of seawater temperature and salinity at different depths (30m, 65m, 110m, 220m) as an example, the spatial distribution comparison between the reconstruction results and the reanalysis data is carried out, as Figures 4 - 19 shown. It can be seen that the spatial distribution of the reconstruction results of seawater temperature and salinity at different depths is relatively consistent with the reanalysis data. The ocean three-dimensional temperature and salinity field achieved by the ocean three-dimensional temperature and salinity field reconstruction model based on graph attention network and temporal convolutional network can accurately reflect the real ocean three-dimensional temperature and salinity field.

[0047] To further verify the stability of the model reconstruction structure with depth and time changes, the RMSE variation trends of the seawater temperature and salinity reconstruction results with depth and time are calculated and analyzed. The RMSE variation trend of the temperature field reconstruction results with depth is as Figure 20 shown in the figure of the RMSE variation of the temperature field reconstruction results with depth. The RMSE variation trend of the temperature field reconstruction results with time is as Figure 21As shown in the figure of the change of RMSE of the temperature field reconstruction result over time. Figure 22 Figure of the change of RMSE of the salinity field reconstruction result with depth; Figure 23 Figure of the change of RMSE of the salinity field reconstruction result over time; From Figure 21 and Figure 23 It can be seen that the overall error of the temperature and salinity reconstruction results changes relatively stably over time. The RMSE values at depths of 30m, 110m, and 222m change smoothly and stably over time. The reconstruction result at a depth of 65m near the thermocline is slightly worse than those at other depths, but still maintains a relatively stable trend over time. The three-dimensional temperature and salinity field reconstruction model based on the graph attention neural network constructed by the present invention can effectively capture the change characteristics of seawater temperature and salinity over time and effectively reflect the real three-dimensional temperature and salinity field of the ocean.

[0048] The above is only a relatively preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be covered within the protection scope 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 the sea surface satellite remote sensing data, uses the 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 the 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. The specific steps include: S1, obtain the daily average sea surface satellite remote sensing data of the experimental area, and perform interpolation processing, 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; Among them, 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 latitudinal component USSW and longitudinal component VSSW; S2, input the processed sea surface satellite remote sensing data into the constructed ocean three-dimensional temperature and salinity field reconstruction model based on graph attention network and temporal convolution network, and output the ocean three-dimensional temperature field or salinity field after temporal convolution, graph learning and graph attention processing; 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.

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 quantitative differences between 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: 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 and convert the low-channel sea surface features into higher-dimensional feature representations.

4. The method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 3, characterized in that: 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. Features are extracted 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.

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 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 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 of the multi-scale dilated convolution is: ； In the formula, 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; ； In the formula, Output after feature concatenation, 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, and global pooling is performed on each channel generated by multi-scale dilated convolution. The node dimension and time dimension of each channel are compressed to extract global information and reduce the amount of calculation. The calculation formula is: ； In the formula, For the The global spatiotemporal information representation of channels, is the number of nodes, For the The first Nodes and The characteristic representation of time; The features after global pooling 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 number compression, and the second fully connected layer restores the number of channels to the initial dimension. ； In the formula, 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 average pool each channel generated by multi-scale dilated convolution to compress the spatiotemporal feature dimensions, and 1D convolution is used to capture the local dependency of adjacent channels. The calculation formula is: ； In the formula, is the output of 1D convolution, is the weight parameter of the 1D convolution kernel, is the feature vector output after average pooling; 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 fused after being operated by the Sigmoid activation function and the Tanh activation function to obtain the complete temporal features; the final output of the temporal convolution module is: ； In the formula, 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.

6. The method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 3, characterized in that: 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; Adaptive adjacency matrix Defined as: ； In the formula, 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, use the LeakyReLU activation function to enhance the nonlinear ability of the model and avoid gradient disappearance; convert the similarity into a probability distribution through the Softmax activation function, which is used to weighted aggregate the information of adjacent nodes; The obtained adaptive adjacency matrix 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 ; Construct a mask matrix M to mark the neighbor positions retained by each node. The element calculation formula of M is: ； In the formula, For the The node and The connection relationship of the nodes, For the The index set of nodes; Adaptive adjacency matrix after sparse processing It is 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 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.

7. The method for reconstructing the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 6, 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 of nodes , and its calculation formula is: ； In the formula, 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: ； In the formula, For the Nodes and The edge weights connecting the nodes, for index, For the Nodes and The edge weights connecting the nodes; 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, each attention head independently calculates the attention coefficient and node representation, and the outputs of multiple heads are spliced ​​to enhance the expression ability of the model. The calculation formula is: ； In the formula, 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: ； In the formula, is the final output of the graph attention module, for Normalization.

8. 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.

9. A system for reconstructing three-dimensional ocean temperature and salinity fields using satellite remote sensing data, characterized in that: The method for reconstructing the three-dimensional temperature and salinity field of the ocean using satellite remote sensing data as described in any one of claims 1 to 8 is implemented, and the system comprises: 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, and 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 after 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.

10. The system for reconstructing three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 9, 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