Mesoscale vortex temperature salt structure reconstruction method based on ocean reanalysis data

Abstract

The invention provides a mesoscale vortex temperature salt structure reconstruction method based on ocean reanalysis data, and belongs to the field of mesoscale vortex three-dimensional reconstruction. The method comprises the following steps: firstly, identifying mesoscale vortexes by adopting a hybrid identification algorithm of the mesoscale vortexes, then extracting a data set based on an identification result and ocean reanalysis, converting a three-dimensional reconstruction problem of a mesoscale vortex temperature salt structure into a prediction problem of a layer depth sequence, and realizing construction of a mesoscale vortex reconstruction model by virtue of deep learning; meanwhile, mesoscale vortex vertical and horizontal structures are introduced, the mesoscale vortex vertical and horizontal structures are taken as a constraining physical mechanism to be incorporated into a model input module, and an attention mechanism is correspondingly added to assist weight output of the model, so that three-dimensional reconstruction of the mesoscale vortex temperature salt structure is realized. According to the method, the complete mesoscale vortex can be better understood from the perspective of vertical water transportation, one-sided summarization of characteristic properties at the vortex center is made up, and the model shows good adaptability and generalization ability on different data sets.

Description

Technical Field

[0001] The invention belongs to the technical field of three-dimensional reconstruction of mesoscale vortices, and relates to a method for reconstructing the temperature-salinity structure of a mesoscale vortex based on ocean reanalysis data. Background Art

[0002] Mesoscale eddies have a profound impact on the temperature and salinity structure of the ocean on a global scale. Currently, most scholars combine the sea surface eddy information with its captured (matched) in situ observation data (mostly Argo) for data fusion, and the reconstruction results are mostly profiles or outlines. Although excellent experimental results have been achieved, the description of the mesoscale eddies in a higher dimension by the three-dimensional structure has been ignored. At the same time, there is still a gap in the application of deep learning in the three-dimensional reconstruction of mesoscale eddies. Based on this, the present invention proposes a neural network based on slice-based layer depth sequence prediction to realize the application of deep learning in the field of mesoscale eddy data mining and three-dimensional reconstruction. Summary of the invention

[0003] In order to overcome the urgent problem of how to use limited mesoscale eddy environment data to reconstruct the refined sound field under offshore conditions in the middle and far seas, the present invention proposes a mesoscale eddy temperature-salinity structure reconstruction method based on ocean reanalysis data. The method first uses a hybrid recognition algorithm of mesoscale eddies to identify mesoscale eddies, and then, based on the recognition results and the ocean reanalysis extraction data set, the three-dimensional reconstruction problem of the mesoscale eddy temperature-salinity structure is converted into a layer depth sequence prediction problem, and the construction of the mesoscale eddy reconstruction model is realized with the help of deep learning technology; at the same time, the vertical and horizontal structures of the mesoscale eddies are introduced, and they are incorporated into the model input module as constrained physical mechanisms, and the attention mechanism is added accordingly to assist the weighted output of the model, thereby realizing the three-dimensional reconstruction of the mesoscale eddy temperature-salinity structure.

[0004] The technical solutions proposed by the present invention for the above-mentioned technical problems are as follows:

[0005] A method for reconstructing mesoscale eddy temperature-salinity structure based on ocean reanalysis data comprises the following steps:

[0006] S1. Collect ocean observation data and model forecast data to form the original data set;

[0007] S2. Based on the original data set, a mesoscale eddy mixing identification algorithm is used to obtain a mesoscale eddy identification feature data set;

[0008] S3, extracting a layer depth sequence based on the mesoscale eddy recognition feature data set;

[0009] S4, performing data cleaning on the obtained layer depth sequence to obtain a mesoscale vortex layer depth sequence;

[0010] S5. Convert the three-dimensional reconstruction problem of mesoscale eddy temperature and salt structure into the prediction problem of layer depth sequence;

[0011] S6. Build and train a mesoscale eddy reconstruction model based on layer-depth sequence deep learning;

[0012] S7. Based on the trained mesoscale eddy reconstruction model, three-dimensional reconstruction of the mesoscale eddy temperature and salinity structure is performed.

[0013] Furthermore, the specific process of S2 is as follows:

[0014] S2.1, the data in the original data set are converted from the sea surface height field to the geostrophic flow field. The formula is as follows:

[0015]

[0016] Where u' and v' are the meridional and zonal geostrophic current velocities, h' is the sea surface height, f is the Coriolis force parameter, and g is the gravitational acceleration;

[0017] S2.2, using flow field geometry methods to identify mesoscale vortices and vortex centers;

[0018] S2.3, in order to minimize the possibility of non-closed vortices before the subsequent intersection detection vortex, the sea surface height data is used to search for closed contours to obtain mesoscale vortices;

[0019] S2.4, compare the mesoscale vortex information obtained by S2.2 and S2.3 respectively. Only when the intersection area of ​​the two mesoscale vortices exceeds 50% of the area of ​​a single mesoscale vortex determined by each method and the distance between the vortex centers does not exceed 1 / 12°, the identification results of the two methods are considered valid; under this condition, the vortex center identified by S2.2 is preferentially used as the actual center point, and finally a mesoscale vortex identification feature data set with sea surface vortex identification contour and center information is formed.

[0020] Furthermore, in S3, the specific process of extracting the layer depth sequence is as follows:

[0021] The sea surface characteristic information obtained from the mesoscale eddy identification feature dataset was used to perform in-situ projection on the JCOPE2M ocean reanalysis temperature and salinity data. The general impact depth of the mesoscale eddy 0-1000 meters was evenly divided into 10 typical layer depth sequences at intervals of 100 meters and extracted according to the projection position.

[0022] Furthermore, in S4, due to the diverse data sources, in order to ensure the quality of the data, the layer depth sequence data obtained in S3 is cleaned. The specific method is as follows:

[0023] First, take the layer depth temperature and salinity data in S3 (without distinguishing between cold and warm vortices), and calculate all the grid points x of the two-dimensional temperature and salinity field corresponding to each layer depth i The mean μ and covariance matrix S are then used to calculate the error distance D M , the formula is as follows:

[0024]

[0025]

[0026]

[0027] When D M When it is greater than three times the standard deviation of the average error distance, the grid point data is considered to be a bad point. When the number of bad points in the two-dimensional temperature-salinity field exceeds 1% of the total number of grid points, the layer depth temperature-salinity data formed in S3 and the sequences in the group where the corresponding vortex is located are considered to be unusable; after cleaning, the mesoscale vortex layer depth sequence is obtained.

[0028] Furthermore, in S5, the three-dimensional reconstruction problem of the mesoscale vortex is transformed into a prediction problem of the layer depth sequence, as follows:

[0029] After data cleaning, a layer depth sequence consisting of M rows and N columns is obtained to identify the mesoscale eddy region R. Each segment in the sequence contains the temperature and salinity information of the mesoscale eddy at different layer depths. Therefore, the temperature and salinity field of any segment in the spatial region can be expressed by the tensor X t ∈R M×N×V Indicates, where V represents the predicted data changes with the layer depth; the three-dimensional reconstruction problem of mesoscale vortex is transformed into: given a known temperature (salt) field X at n layer depths t-n+1 ,X t-n+2 ,…,X t To predict k continuous temperature (salt) fields according to layer depth Sequence prediction problem:

[0030]

[0031] Among them, P(·) represents conditional probability. Considering that in actual measurement, the sea surface temperature (salinity) field data can only be obtained long-term and stably through satellites, and other acquisition methods are unstable and difficult to obtain continuously, n=1 is set. It should also be pointed out that the significant impact depth of mesoscale eddies is generally within 1000 meters, so the prediction sequence length k is as small as possible while meeting the layer depth of 1000 meters. This is because sequence prediction is often accompanied by the problem of error accumulation. Reducing the prediction length can improve the effectiveness of the prediction, k=5-20, preferably 5-10.

[0032] Furthermore, in S6, the construction process of the mesoscale eddy reconstruction model is as follows:

[0033] The mesoscale eddy reconstruction model includes a priority stacking and extraction module, a spatiotemporal feature extraction module and an attention mechanism module; the output of the priority stacking and extraction module is used as the input of the spatiotemporal feature extraction module, and the output of the spatiotemporal feature extraction module is used as the input of the attention mechanism module;

[0034] The priority stacking and extraction module is used to extract the original temperature and salinity data (i.e., the mesoscale vortex layer depth sequence obtained by S4), temperature and salinity anomaly data and pressure anomaly data; the spatiotemporal feature extraction module uses the same network architecture as the PredRNN model to extract sequence features with continuous meaning through the transmission between cells with the same step length; the attention mechanism module is used to self-weight the input multidimensional data, weight and quantify the multidimensional features, and output more realistic synthetic data.

[0035] Furthermore, the method for extracting the temperature-salinity anomaly data is: using the mesoscale eddy layer depth sequence obtained by S4 to subtract the corresponding WOA area and layer depth climatological data, to obtain the temperature-salinity anomaly layer depth sequence.

[0036] Furthermore, the pressure anomaly data includes a three-dimensional vertical pressure anomaly structure and a three-dimensional horizontal pressure anomaly structure, and the extraction method is as follows:

[0037] Under the quasi-geostrophic vertical mode equation shown in formula (6), given the tensile coordinate transformation as shown in formula (7), formula (6) is transformed into a constant coefficient differential equation that is independent of local stratification and rotation, and the normalized three-dimensional vertical pressure anomaly structure is obtained, as shown in formula (8):

[0038]

[0039]

[0040] H(x)=sin Z n (8)

[0041] Among them, f 0 is the Coriolis force; N is the buoyancy frequency; λ is the eigenvalue corresponding to the vertical mode; H is the vertical component of the three-dimensional pressure anomaly structure of the mesoscale vortex; R is the vortex radius, which is defined as the equivalent radius for irregular vortices;

[0042] In addition, it is assumed that different layers have pressure anomaly structures that conduct downward, and formula (9) is used as a constraint condition on the horizontal pressure anomaly structure:

[0043]

[0044] Among them, r n is the normalized distance from the vortex center, and R(·) is the three-dimensional horizontal pressure anomaly function.

[0045] Furthermore, the method for weighted quantization of the attention mechanism module is as follows:

[0046] The attention mechanism module takes the output X of the spatiotemporal feature extraction module as its input, integrates the dimensions through the Permute operator, passes through the attention network and performs the Softmax operation to obtain the weight of each vector, and then adds the output of each iteration of the spatiotemporal feature extraction module to its corresponding weight and reverse dot product to obtain a weighted vector, and finally maps it to the output result through the fully connected layer. The calculation formula is as follows:

[0047] Out = sum(softmax(tanh(X·W s +b))×X) (10)

[0048] Among them, W s is the weight matrix and b is the bias term.

[0049] Furthermore, in the S6, during the training of the mesoscale eddy reconstruction model, the L1+L2 loss function is used, and the Adam optimizer is used for training; the hyperparameters are set as follows: the learning rate is 0.0002, the batch_size is 8, the corresponding epoch is increased to 100000, the filter_size is 5, and the stride is 1; for the β in the Adam optimizer 1 and β 2 Parameters, set β 1 = 0.9 and β 2 =0.999.

[0050] Beneficial effects of the invention: The invention innovatively proposes a concept of a mesoscale vortex layer depth sequence, and based on this perspective, performs sequence prediction with the PredRNN sequence prediction model to obtain temperature and salinity information at the depth of each layer, and generates a relatively complete three-dimensional structure of the mesoscale vortex. Compared with the current mainstream statistical characteristics and the generated field under the pure fluid mechanics background, the method of the invention can better understand the complete mesoscale vortex from the perspective of vertical water transport, and makes up for the one-sided generalization of the characteristic properties at the center of the vortex; considering the physical constrainability of the model and the emphasis on expression, a physical mechanism layer and a corresponding attention mechanism are introduced in the input layer, which has a better effect on the characteristic expression of the mesoscale vortex, and the model shows good adaptability and generalization ability on different data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is the technical roadmap of the present invention.

[0052] Figure 2 It is an overview of the vortex sea surface eddy information (base map sea surface temperature and superimposed sea surface geostrophic flow field) obtained by hybrid identification algorithm and projected onto the anticyclonic vortex extracted from the ocean reanalysis data and the four-dimensional schematic diagram of the cyclonic vortex.

[0053] Figure 3 It is a diagram of the data cleaning process (including training, testing, and validation data sets).

[0054] Figure 4 It is a schematic diagram of the overall flow of the model. DETAILED DESCRIPTION

[0055] The specific implementation of the present invention is described in detail below in conjunction with the technical scheme and the accompanying drawings.

[0056] like Figure 1 As shown, an embodiment of the present invention provides a method for reconstructing a mesoscale eddy temperature-salinity structure based on ocean reanalysis data, comprising the following steps:

[0057] S1. Collect ocean observation data and model forecast data to form the original data set; the details are as follows:

[0058] The sea surface height data, temperature and salinity reanalysis data and some in situ observation data are collected from January 1, 2007 to December 31, 2020; the reanalysis data include JCOPE2M ocean reanalysis data, HYCOM ocean reanalysis data and Argo floating data.

[0059] S2. Based on the original data set, a mesoscale eddy mixing identification algorithm is used to obtain a mesoscale eddy identification feature data set; including:

[0060] S2.1, the data in the original data set are converted from the sea surface height field to the geostrophic flow field. The formula is as follows:

[0061]

[0062] Where u' and v' are the meridional and zonal geostrophic current velocities, h' is the sea surface height, f is the Coriolis force parameter, and g is the gravitational acceleration;

[0063] S2.2, Identification of mesoscale vortices and vortex centers using flow field geometry methods

[0064] The geometric characteristics of a mesoscale vortex are defined as a nearly circular region in which the velocity vector rotates symmetrically around the center of the vortex. When the velocity vector field in a region exhibits rotational flow characteristics, the region is considered to be a potential vortex. In this region, if there is a velocity extreme point and the velocity vectors around this point show a symmetrical distribution pattern, and these velocity vectors rotate around the center point in a clockwise or counterclockwise direction, then this structure is identified as a mesoscale vortex.

[0065] S2.3, in order to minimize the possibility of non-closed vortices before the subsequent intersection detection vortex, the sea surface height data is used to search for closed contours to obtain mesoscale vortices;

[0066] S2.4, compare the mesoscale vortex information obtained by S2.2 and S2.3 respectively. Only when the intersection area of ​​the two mesoscale vortices exceeds 50% of the area of ​​a single mesoscale vortex determined by each method and the distance between the vortex centers does not exceed 1 / 12°, the identification results of the two methods are considered valid; under this condition, the vortex center identified by S2.2 is preferentially used as the actual center point, and finally a mesoscale vortex identification feature data set with sea surface vortex identification contour and center information is formed.

[0067] Experiments based on the above-mentioned mesoscale eddy mixing identification algorithm have shown a significant improvement in identification accuracy, which indicates that both hybrid algorithms have achieved self-elimination of some errors based on the original algorithms; since the mesoscale eddy mixing identification algorithm is the basis for subsequent extraction of deep thermohaline data and deep learning training, the identification algorithm improves data reliability and quality.

[0068] S3, based on the mesoscale eddy recognition feature data set, extracting a layer depth sequence; specifically:

[0069] The sea surface feature information in the mesoscale eddy identification feature dataset was used to perform in-situ projection in the JCOPE2M ocean reanalysis temperature and salinity data. The mesoscale eddy general impact depth of 0-1000 meters was evenly divided into 10 typical layer depth sequences at 100-meter intervals and extracted according to the projection position, such as Figure 2 In this embodiment, 10 layers of temperature and salinity data of each vortex are grouped together, and a total of 550,000 groups of anticyclonic vortex layer depth data samples and 430,000 groups of cyclonic vortex layer depth samples are obtained.

[0070] S4, ocean observation data and model forecast data have the advantages of wide coverage and clear grid points, but they also often have uncontrollable abnormal data (such as Figure 2The missing values ​​in the four-dimensional extracted data of the mesoscale eddy on the right side are due to the seabed topography. Since the data sources used in this embodiment are diverse, in order to ensure the quality of the data when forming the deep learning sample data set, the obtained layer depth sequence is cleaned to obtain the mesoscale eddy layer depth sequence; the specific method is:

[0071] First, take the layer depth temperature and salinity data in S3 (without distinguishing between cold and warm vortices), and calculate all the grid points x of the two-dimensional temperature and salinity field corresponding to each layer depth i The mean μ and covariance matrix S are then used to calculate the error distance d M , the formula is as follows:

[0072]

[0073]

[0074]

[0075] When D M When it is greater than three times the standard deviation of the average error distance, the grid point data is considered to be a bad point. When the number of bad points in the two-dimensional temperature-salinity field exceeds 1% of the total number of grid points, the layer-depth temperature-salinity data formed in S3 and the sequences in the group where the corresponding vortex is located are considered to be unavailable. Figure 3 As shown; after cleaning, the mesoscale vortex layer depth sequence is obtained.

[0076] S5. The purpose of this embodiment is to achieve three-dimensional reconstruction of mesoscale vortices based on the prediction of layer depth sequence, so the demand for time dimension can be converted into layer depth dimension. After S4 data cleaning, a layer depth sequence consisting of M rows and N columns is obtained to identify the mesoscale vortex region R. Each segment in the sequence contains the temperature and salt information of the mesoscale vortex at different layer depths. Therefore, the temperature and salt field of any segment in the spatial region can be obtained by the tensor X t ∈R M×N×V Indicates, where V represents the predicted data changes with the layer depth; the three-dimensional reconstruction problem of mesoscale vortex is transformed into: given a known temperature (salt) field X at n layer depths t-n+1 ,X t-n+2 ,…,X t To predict k continuous temperature (salt) fields according to layer depth Sequence prediction problem:

[0077]

[0078] Among them, P(·) represents conditional probability. Considering that in actual measurement, sea surface temperature (salinity) field data can only be obtained long-term and stably through satellites, and other acquisition methods are unstable and difficult to obtain continuously, this embodiment sets n=1. It should also be pointed out that the significant impact depth of mesoscale vortices is generally within 1000 meters, so the prediction sequence length k is as small as possible while meeting the layer depth of 1000 meters. This is because sequence prediction is often accompanied by the problem of error accumulation. Reducing the prediction length can improve the effectiveness of the prediction. This embodiment sets k=5, 10, and 20, corresponding to the extraction of layer depth sequences at intervals of 200 meters, 100 meters, and 50 meters, respectively. It is worth emphasizing here that k=5, 10 is generally considered to be a medium-short prediction, and k=20 is a long prediction. In general, the spatiotemporal prediction model is more suitable for short-term prediction effects, but in order to compare the prediction effects of the model later, the long prediction is included in the control experimental group.

[0079] S6. Construct and train a mesoscale eddy reconstruction model based on deep learning of layer-depth sequences; Figure 4 As shown, the mesoscale eddy reconstruction model includes a priority stacking and extraction module, a spatiotemporal feature extraction module and an attention mechanism module; the output of the priority stacking and extraction module serves as the input of the spatiotemporal feature extraction module, and the output of the spatiotemporal feature extraction module serves as the input of the attention mechanism module.

[0080] The priority stacking and extraction module is used to extract the original temperature and salinity data (i.e., the mesoscale vortex layer depth sequence obtained by S4), temperature and salinity anomaly data and pressure anomaly data; the spatiotemporal feature extraction module uses the same network architecture as the PredRNN model to extract sequence features with continuous meaning through the transmission between cells with the same step length; the attention mechanism module is used to self-weight the input multidimensional data, weight and quantify the multidimensional features, and output more realistic synthetic data.

[0081] The method for extracting the temperature-salinity anomaly data is as follows: the mesoscale vortex layer-depth sequence obtained by S4 is used to subtract the corresponding climatological data of the WOA area and layer depth to obtain the temperature-salinity anomaly layer-depth sequence.

[0082] The pressure anomaly data includes a three-dimensional vertical pressure anomaly structure and a three-dimensional horizontal pressure anomaly structure, and the extraction method is as follows:

[0083] Under the quasi-geostrophic vertical mode equation shown in formula (6), given the tensile coordinate transformation as shown in formula (7), formula (6) is transformed into a constant coefficient differential equation that is independent of local stratification and rotation, and the normalized three-dimensional vertical pressure anomaly structure is obtained, as shown in formula (8):

[0084]

[0085]

[0086] H(z)=sin Z n (8)

[0087] Among them, f 0 is the Coriolis force; N is the buoyancy frequency; λ is the eigenvalue corresponding to the vertical mode; H is the vertical component of the three-dimensional pressure anomaly structure of the mesoscale vortex; R is the vortex radius, which is defined as the equivalent radius for irregular vortices;

[0088] In addition, it is assumed that different layers have pressure anomaly structures that conduct downward, and formula (9) is used as a constraint condition on the horizontal pressure anomaly structure:

[0089]

[0090] Among them, r n is the normalized distance from the vortex center, and R(·) is the three-dimensional horizontal pressure anomaly function.

[0091] The method for weighted quantization of the attention mechanism module is as follows: the attention mechanism module takes the output X of the spatiotemporal feature extraction module as its input, integrates the dimensions through the Permute operator, passes through the attention network and performs the Softmax operation to obtain the weight of each vector, and then adds the output of each iteration of the spatiotemporal feature extraction module to its corresponding weight and reverse dot product to obtain a weighted vector, and finally maps it to the output result through the fully connected layer. The calculation formula is as follows:

[0092] Out = sum(softmax(tanh(X·W s +b))×X) (10)

[0093] Among them, W s is the weight matrix and b is the bias term.

[0094] In the training process of the mesoscale eddy reconstruction model, the L1+L2 loss function is used, and the Adam optimizer is used for training; the hyperparameters are set as follows: the learning rate is 0.0002, the batch_size is 8, the corresponding epoch is increased to 100000, the filter_size is 5, and the stride is 1; for the β in the Adam optimizer 1 and β 2 Parameters, set β 1 = 0.9 and β 2 =0.999.

[0095] S7. Evaluation and judgment method adapted to mesoscale eddy reconstruction:

[0096] For the prediction of layer depth sequence, the prediction accuracy needs to be evaluated from two aspects: the first is the error evaluation of each layer of the divided layer depth sequence, and the second is the prediction accuracy evaluation of the entire sequence. Although the two are errors of the same source, they require different calculation methods for evaluation.

[0097] First, for the evaluation of the temperature-salinity error of a single layer, the common L1 or L2 error can be used. In this embodiment, MSE (Mean Square Error) is selected as the preferred calculation method, see formula (11), where n is the number of all grid points to be calculated, i, j are the relative horizontal and vertical coordinates of the grid points where the calculated values ​​are located, is the prediction result of position (i, j), X i,j Then it is the measured result at the (i, j) position.

[0098]

[0099] Then, the overall sequence prediction effect cannot be simply evaluated by numerical error, but needs to be judged by correlation analysis between the predicted sequence temperature and salinity structure characteristics and the measured data. This will involve a very serious problem: what kind of measured data to use. The model sequence evaluation is based on two types of data: ship-borne measured profiles and Argo float data.

[0100] The profile data still uses MSE as the error calculation method, while for the Argo float data, the temperature-salinity trend correlation of the custom layered depth is used as the judgment criterion. Here, the Pearson correlation coefficient is used. Before calculating the Pearson correlation coefficient, the predicted data and the measured data need to be normalized (Min-max). The relevant formula is as follows:

[0101]

[0102]

[0103] Where x is the full length of any sequence, x i and i is the i-th pair of observations in the sample data, and are the means of the sequences x and y respectively.

[0104] The above is a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data, characterized in that: The following steps are involved: S1. Collect ocean observation data and model forecast data to form the original data set; S2. Based on the original data set, a mesoscale eddy mixing identification algorithm is used to obtain a mesoscale eddy identification feature data set; S3, extracting a layer depth sequence based on the mesoscale eddy recognition feature data set; S4, performing data cleaning on the obtained layer depth sequence to obtain a mesoscale vortex layer depth sequence; S5. Convert the three-dimensional reconstruction problem of mesoscale eddy temperature and salt structure into the prediction problem of layer depth sequence; S6. Build and train a mesoscale eddy reconstruction model based on layer-depth sequence deep learning; S7. Based on the trained mesoscale eddy reconstruction model, three-dimensional reconstruction of the mesoscale eddy temperature and salinity structure is performed.

2. The method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 1, characterized in that: The specific process of S2 is: S2.1, the data in the original data set are converted from the sea surface height field to the geostrophic flow field. The formula is as follows: Where u' and v' are the meridional and zonal geostrophic current velocities, h' is the sea surface height, f is the Coriolis force parameter, and g is the gravitational acceleration; S2.2, using flow field geometry methods to identify mesoscale vortices and vortex centers; S2.3, using sea surface height data to search for closed contours and obtain mesoscale vortices; S2.4, compare the mesoscale vortex information obtained by S2.2 and S2.3 respectively. The identification results of the two methods are valid only when the intersection area of ​​the two mesoscale vortices exceeds 50% of the area of ​​a single mesoscale vortex determined by each method and the distance between the vortex centers does not exceed 1 / 12°. Under this condition, the vortex center identified by S2.2 is preferentially used as the actual center point, and finally a mesoscale vortex identification feature data set with sea surface vortex identification contour and center information is formed.

3. A method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 1 or 2, characterized in that: In S3, the specific process of extracting the layer depth sequence is as follows: The sea surface characteristic information obtained from the mesoscale eddy identification feature dataset was used to perform in-situ projection on the JCOPE2M ocean reanalysis temperature and salinity data. The general impact depth of the mesoscale eddy 0-1000 meters was evenly divided into 10 typical layer depth sequences at intervals of 100 meters and extracted according to the projection position.

4. The method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 3 is characterized in that: In S4, the layer depth sequence data obtained in S3 is cleaned, and the specific method is as follows: First, take the layer depth temperature and salinity data in S3 and calculate all grid points x of the two-dimensional temperature and salinity field corresponding to each layer depth. i The mean μ and covariance matrix S are then used to calculate the error distance D M , the formula is as follows: When D M When it is greater than three times the standard deviation of the average error distance, the grid point data is a bad point. When the number of bad points in the two-dimensional temperature-salinity field exceeds 1% of the total number of grid points, the layer depth temperature-salinity data formed in S3 and the sequence in the group where the corresponding vortex is located are unavailable; after cleaning, the mesoscale vortex layer depth sequence is obtained.

5. A method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 1, 2 or 4, characterized in that: In S5, the three-dimensional reconstruction problem of the mesoscale vortex is transformed into a prediction problem of the layer depth sequence, as follows: After data cleaning, a layer depth sequence consisting of M rows and N columns is obtained to identify the mesoscale eddy region R. Each segment in the sequence contains the temperature and salinity information of the mesoscale eddy at different layer depths. Therefore, the temperature and salinity field of any segment in the spatial region can be expressed by the tensor X t ∈R M×N×V Represented by, where V represents the predicted data changes with layer depth; the three-dimensional reconstruction problem of mesoscale eddies is transformed into: given n layer depths with known temperature / salinity fields X t-n+1 ,X t-n+2 ,…,X t To predict k continuous temperature / salinity fields according to layer depth Sequence prediction problem: Wherein, P(·) represents conditional probability; n=1; k=5-20, preferably 5-10.

6. The method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 5, characterized in that: In S6, the construction process of the mesoscale eddy reconstruction model is as follows: The mesoscale eddy reconstruction model includes a priority stacking and extraction module, a spatiotemporal feature extraction module and an attention mechanism module; the output of the priority stacking and extraction module is used as the input of the spatiotemporal feature extraction module, and the output of the spatiotemporal feature extraction module is used as the input of the attention mechanism module; The priority stacking and extraction module is used to extract the original temperature and salinity data, i.e., the mesoscale vortex layer depth sequence, temperature and salinity anomaly data and pressure anomaly data obtained by S4; the spatiotemporal feature extraction module uses the same network architecture as the PredRNN model to extract sequence features with continuous meaning through the transmission between cells with the same step length; the attention mechanism module is used to self-weight the input multidimensional data, weight and quantify the multidimensional features, and output more realistic synthetic data.

7. The method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 6, characterized in that: The method for extracting the temperature-salinity anomaly data is as follows: using the mesoscale vortex layer-depth sequence obtained by S4 to subtract the corresponding WOA area and layer-depth climatological data, to obtain the temperature-salinity anomaly layer-depth sequence.

8. The method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 6, characterized in that: The pressure anomaly data includes a three-dimensional vertical pressure anomaly structure and a three-dimensional horizontal pressure anomaly structure, and the extraction method is: Under the quasi-geostrophic vertical mode equation shown in formula (6), given the tensile coordinate transformation as shown in formula (7), formula (6) is transformed into a constant coefficient differential equation that is independent of local stratification and rotation, and the normalized three-dimensional vertical pressure anomaly structure is obtained, as shown in formula (8): H(z)=sin Z n (8) Where f0 is the Coriolis force; N is the buoyancy frequency; λ is the eigenvalue corresponding to the vertical mode; H is the vertical component of the three-dimensional pressure anomaly structure of the mesoscale vortex; R is the vortex radius, which is defined as the equivalent radius for irregular vortices; In addition, it is assumed that different layers have downward-conducting pressure anomaly structures, and formula (9) is used as a constraint condition on the horizontal pressure anomaly structure: Among them, r n is the normalized distance from the vortex center, and R(·) is the three-dimensional horizontal pressure anomaly function.

9. The method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 6, characterized in that: The method for weighted quantization of the attention mechanism module is as follows: The attention mechanism module takes the output X of the spatiotemporal feature extraction module as its input, integrates the dimensions through the Permute operator, passes through the attention network and performs the Softmax operation to obtain the weight of each vector, and then adds the output of each iteration of the spatiotemporal feature extraction module to its corresponding weight and reverse dot product to obtain a weighted vector, and finally maps it to the output result through the fully connected layer. The calculation formula is as follows: Out=sum(softmax(tanh(X·W s +b))×X) (10) Among them, W s is the weight matrix and b is the bias term.

10. The method for reconstructing mesoscale eddy temperature and salinity structure based on ocean reanalysis data according to claim 6, characterized in that: In S6, during the training of the mesoscale eddy reconstruction model, the L1+L2 loss function was used, and the Adam optimizer was used for training; the hyperparameters were set as follows: the learning rate was 0.0002, the batch_size was 8, the corresponding epoch increased to 100000, the filter_size was 5, and the stride was 1; for the β1 and β2 parameters in the Adam optimizer, β1=0.9 and β2=0.999 were set.

Images