Global ocean temperature-salinity profile joint reconstruction method based on deep learning

Abstract

The invention discloses a global ocean temperature-salinity profile joint reconstruction method based on deep learning, and relates to the technical field of ocean information, and the method is technically characterized in that the method comprises the steps: S1, obtaining historical sea surface environment parameter remote sensing data information and underwater temperature-salinity profile data, and constructing a reconstruction model data set; s2, dividing global partitions through K-means clustering according to the temperature-salinity profile characteristics; s3, constructing a regional FFNN temperature-salinity profile reconstruction model according to the partitions, and obtaining a global ocean temperature-salinity profile joint reconstruction model; and S4, obtaining real-time sea surface remote sensing data, and obtaining real-time temperature-salinity profile data based on the global ocean temperature-salinity profile joint reconstruction model. According to the method, the properties of the ocean temperature-salinity profile are combined, the performance is excellent, good applicability is achieved, the global temperature-salinity model is reconstructed based on real-time sea surface remote sensing data, and a new thought and technical support are provided for rapid construction of the global underwater temperature-salinity profile.

Description

Technical Field

[0001] The present invention relates to the field of marine information technology, and in particular to a method for jointly reconstructing global ocean temperature and salinity profiles based on deep learning. Background Art

[0002] The underwater temperature and salinity structure of the ocean is an essential environmental factor for studying ocean processes and climate change, and plays an important role in studying the multidimensional and multiscale dynamic processes in the ocean. However, due to the lack of ocean observation data and the difficulty in obtaining them, the currently measurable cross-sectional data on the internal state of the ocean are relatively sparse, and the temporal and spatial resolution is far from meeting the requirements for the detailed study and application of the underwater temperature and salinity structure, which has restricted my country's scientific understanding and support capabilities for the marine environment in marine economy, marine engineering, shipping and transportation, and resource development and utilization.

[0003] At present, the ocean surface-subsurface reconstruction methods are mainly divided into two types: dynamic reconstruction method and statistical reconstruction method. The dynamic reconstruction method is mainly based on the surface quasigeostrophic (SQG) dynamics theory, which projects the internal solution to the first baroclinic mode and the barotropic mode, and uses the sea surface observation data to constrain the solution to achieve the inversion of subsurface parameters such as sea surface height, sea surface temperature and sea surface salinity. The statistical reconstruction method aims to statistically or empirically combine satellite and field observations to seek the coupling relationship between the sea surface and the deep layer or between different observations without involving dynamic equations, thereby reconstructing the three-dimensional element field. Its main purpose is to solve the problem of effectively using satellite remote sensing sea surface temperature and sea surface height data to quickly estimate the real-time dynamic three-dimensional ocean temperature and salinity field when there is little or no real-time underwater three-dimensional temperature and salinity profile observation data. In recent years, there has been great progress in this area of ​​research. Methods such as multivariate linear regression, multivariate EOF reconstruction, coupled mode reconstruction and gravity empirical mode method have been applied in the reconstruction of three-dimensional temperature and salinity fields.

[0004] However, most existing reconstruction methods tend to use the monthly average gridded temperature-salinity field as a historical training label to establish a monthly average temperature-salinity field reconstruction model. Actual marine scientific research and marine environmental protection require a daily quasi-real-time underwater three-dimensional temperature-salinity field. For the daily reconstruction model, it is necessary to use the measured scattered temperature-salinity profile as a label. However, the profile changes of the measured scattered data are more complex and contain more small and medium-scale information, and the accuracy of the reconstruction method needs to be improved. Most of the previous reconstruction models are based on machine learning methods, and the contribution of deep learning algorithms to the improvement of reconstruction models is less considered. When statistical modeling is performed in global partitions, the properties of temperature-salinity profiles are not considered, and the partition is only performed by geographical location. At the same time, the existing reconstruction method adopts the technical approach of reconstructing and modeling temperature and salinity separately, which is not efficient and inevitably causes the lack of coordination of temperature and salinity elements. Therefore, how to flexibly use deep learning algorithms in the reconstruction model, further adapt to improve and optimize the reconstruction model effect, and provide a global temperature-salinity profile coordinated intelligent reconstruction method with high timeliness, easy execution and high accuracy is a technical problem that technicians in this field are very concerned about. Summary of the invention

[0005] The purpose of the present invention is to solve the above problems and provide a method for joint reconstruction of global ocean temperature and salinity profiles based on deep learning.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] The present invention provides a method for joint reconstruction of global ocean temperature and salinity profiles based on deep learning, comprising the following steps:

[0008] Step S1, obtaining historical sea surface environmental parameter remote sensing data and underwater temperature and salinity profile data, and constructing a reconstruction model data set;

[0009] Step S2, according to the characteristics of temperature and salinity profiles, global regions are delineated by K-means clustering;

[0010] Step S3, constructing a regional FFNN temperature-salinity profile reconstruction model according to the partitions to obtain a global ocean temperature-salinity profile joint reconstruction model;

[0011] Step S4: Acquire real-time sea surface remote sensing data, and obtain real-time temperature-salinity profile data based on the global ocean temperature-salinity profile joint reconstruction model.

[0012] The present invention is further configured as follows: the specific sub-steps of step S1 are:

[0013] Step S1.1, collect historical temperature and salinity profile data as training data and independent verification data for inverted temperature and salinity profiles, calculate the statistical regression relationship between the surface layer and the subsurface layer, and the selected historical temperature and salinity profile data must meet the following four conditions: (1) measure temperature and salinity at the same time; (2) measure depth above 1000m; (3) measure minimum depth less than 5m; (4) the three data quality control symbols Temperature_CODCflag, Depth_WODflag, and Salinity_CODCflag are all "0";

[0014] Step S1.2, use cubic spline interpolation to vertically interpolate the temperature-salinity profile selected in S11 to 20 inhomogeneous layers of [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500, 600, 700, 800, 900, 1000] m;

[0015] Step S1.3, select the delayed quality controlled L4 level 0.25° daily gridded sea surface height anomaly data product released by the Copernicus Marine Environment Monitoring Service;

[0016] Step S1.4, interpolate the sky-averaged SLA data with a spatial resolution of 0.25° to the corresponding position of the historical temperature-salinity profile;

[0017] Step S1.5: Construct a reconstruction model dataset Y. The aggregation method of dataset Y is as follows:

[0018]

[0019] Among them, each sample in the data set Y is recorded as y (i,j)It is represented as the jth element of the ith sample, and the longitude and latitude of all samples are consistent; i = [1, N] represents the 1st to Nth samples, one sample is a row in the matrix Y, each row contains 43 elements, j = [1, 43] represents the 1st to 43rd elements, corresponding to the 43 columns of the matrix Y; the first and second columns are longitude and latitude respectively, and the third column is the input data sea surface height; the ocean temperature and salinity profile data are divided into [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500 , 600, 700, 800, 900, 1000] m; j = [4, 23] represents the temperature of 20 depth layers in turn, and j = [24, 43] represents the salinity of 20 depth layers in turn; the paired data set Y is divided into training set Y1, validation set Y2 and test set Y3 in a ratio of 8:1:1; the specific division method is: select the first 0.8×N rows of samples as training set Y1, the remaining 0.1×N rows of samples as validation set X2, and 0.1×N rows of samples as test set Y3, and there is no intersection between the three;

[0020] Step S1.6: Normalize the input data. Select the MinMax normalization method to normalize the temperature, salinity and sea surface height data in the input data set Y to the interval [0,1]. The normalization formula is as follows:

[0021]

[0022] The input dataset Y represents the dataset constructed in step S1.5, y j Represents an input sea surface feature in the dataset, Y ijmin is the minimum value of the i-th factor in the data set Y, Y ijmax is the maximum value of the i-th factor in the data set Y. The set of factors normalized according to the formula is recorded as Y' j ; Y ijmin With Y ijmax Will be retained. During model validation and testing, the normalization of a batch of input data in the validation set and test set also uses Y ijmin With Y ijmax The calculation is performed because the statistical information contained in the training set is considered to be a known condition, while the statistical information of the test set and the validation set is an unknown condition. The purpose is to simulate the reasoning effect of the model on the unknown data set based on the existing information in most cases.

[0023] The present invention is further configured as follows: the specific sub-steps of step S2 are:

[0024] Step S2.1, construct a clustering data set, collect monthly average 1°×1° temperature and salinity profile data, sample the temperature and salinity data at a depth of [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500, 600, 700, 800, 900, 1000] (unit: meter), calculate the multi-year average value, and obtain the annual average temperature and salinity data set at different longitudes and latitudes. The data set is constructed as a training set X in the following format:

[0025]

[0026] Among them, each element in the data set X is recorded as x i,j , x i,j represents the jth element of the ith sample, and the longitude and latitude of all samples are consistent; i = [1, N] represents the 1st to Nth samples, one sample is a row in the matrix X, each row contains 42 elements, j = [1, 42] represents the 1st to 42nd elements, corresponding to the 42 columns of the matrix X: the first two columns are the longitude and latitude of the input data, the 3rd to 22nd columns are the temperature values ​​of each sample profile in 20 depth layers, and the 23rd to 42nd columns are the salinity values ​​of each sample profile in 20 depth layers;

[0027] Step S2.2, select the K value range: according to the characteristics of ocean water masses, the preset K value range is [2,20];

[0028] Step S2.3: Randomly select K samples from the training data set as the initial cluster centers:

[0029] Step S2.3.1: For each sample X i , calculate the Euclidean distance between it and the K cluster centers respectively, and record its distance to the jth cluster center as d ij , j∈[1,K];

[0030] Step S2.3.2: Compare the distances from the sample to the cluster center and get D i =min{d ij} i , each sample can select a cluster center that is closest to it, and merge the samples with the same nearest cluster center into one cluster, thus obtaining K sample clusters;

[0031] Step S2.3.3, calculate the mean of the samples in each cluster and use the mean as the new cluster center;

[0032] Step S2.3.4: If the new cluster center is different from the original cluster center, repeat steps S2.3.1-S2.3.3 until the cluster center no longer changes or the predetermined number of iterations m is reached;

[0033] Step S2.4, output cluster division, and obtain the clustering results corresponding to different longitude and latitude sea areas as shown in the figure; calculate the silhouette coefficient of the entire data set. The silhouette coefficient is an indicator for evaluating clustering quality. Its value range is between [-1, 1]. The closer the value is to 1, the better the clustering effect is.

[0034] Step S2.5, repeat steps S2.2-S2.4, compare the silhouette coefficients of the clustering results under different K values, obtain the optimal value of K value and its corresponding clustering result. After calculation, when K=7, the silhouette coefficient is the largest, which is 0.327; according to the clustering results, obtain the division of sea areas at different longitudes and latitudes according to the properties of temperature-salinity profiles, and divide the sea areas according to the clustering results.

[0035] The present invention is further configured as follows: the specific sub-steps of step S3 are:

[0036] Step S3.1: Build the FFNN feedforward neural network reconstruction model. Assuming that n feature variables and m samples are input, the input layer X∈R m×n , construct a hidden layer with h neurons, then the weight and bias of the hidden layer can be expressed as W h ∈R n×h 、b h ∈R 1×h , if there are q output label values, the weight and bias parameters of the output layer are W o ∈R h×q 、b O ∈R 1 ×q , the output of the hidden layer and the output of the output layer are expressed as follows:

[0037] H=XW h +b h

[0038] O=HW o +b O

[0039] Build three hidden layers. The first hidden layer is set to h1 neurons, the second hidden layer has h2 neurons, and the third hidden layer has h3 neurons. Each layer is connected according to the above formula; each hidden layer is followed by a ReLU activation layer, the formula is as follows:

[0040] ReLU(input) = max(0, input)

[0041] Build two hidden layers. The first hidden layer has 128 neurons, and the second hidden layer has 64 neurons. Each layer is connected according to the above formula. Each hidden layer is followed by a ReLU activation layer. Adding an activation layer introduces nonlinear transformation to improve the expressiveness of the model.

[0042] Step S3.2: Select mean square error (MSE) as the loss function of the FFNN reconstruction model to measure the difference between the reconstructed underwater temperature and salinity profile and the measured temperature and salinity profile. The formula is as follows:

[0043]

[0044] Among them, Y i Indicates the label of sample i, i.e. the measured temperature-salinity profile data, Y' i Represents the value reconstructed by the model of sample i;

[0045] Step S3.3: According to the partitioning result of S2, the dataset Y constructed in S1 is divided into seven regional datasets, which are respectively put into the FFNN reconstruction model constructed in S3.1 for training, and the hidden layer weights and bias W are fitted. h 、b h , the loss function is MSE, the number of initial iterations is epoch, epoch is a positive integer, the initial learning rate is lr, the initial batch size is batch size, and the optimizer is Adam;

[0046] Step S3.4, save the trained FFNN models and model output parameters of the seven regions, and the seven regional reconstruction models constitute a global ocean temperature and salinity profile joint reconstruction model.

[0047] The present invention is further configured as follows: the specific sub-steps of step S4 are:

[0048] Step S4.1, obtaining real-time sea surface remote sensing temperature and salinity data as model input;

[0049] Step S4.2: Input data into the global ocean temperature-salinity profile joint reconstruction model constructed in S3 to obtain the real-time global ocean temperature-salinity profile data field.

[0050] Compared with the prior art, the beneficial effects of this solution are as follows: Based on satellite remote sensing data and historical temperature-salinity profile data, the present invention divides the global ocean into seven sea areas according to the characteristics of the temperature-salinity profile, and constructs the FFNN temperature-salinity profile reconstruction model by region. This method combines the properties of the ocean temperature-salinity profile, has superior performance, and has good applicability. It realizes the reconstruction of the global temperature-salinity model based on real-time sea surface remote sensing data, and provides new ideas and technical support for the rapid construction of the global underwater temperature-salinity profile. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is an implementation flow chart of jointly reconstructing the global underwater temperature-salinity profile based on deep learning in an embodiment of the present invention;

[0052] Figure 2is a line graph of the silhouette coefficient of the clustering results under different K values ​​in the embodiment of the present invention;

[0053] Figure 3 It is a global sea area division map for temperature-salinity profile characteristic clustering in an embodiment of the present invention;

[0054] Figure 4 It is a structural diagram of the global ocean temperature-salinity profile joint reconstruction model in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0056] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below in conjunction with the embodiments.

[0057] Example:

[0058] A joint reconstruction method of global ocean temperature and salinity profiles based on deep learning, such as Figure 1-4 As shown, the following steps are included:

[0059] Step S1: Obtain historical sea surface environmental parameter remote sensing data and underwater temperature and salinity profile data to construct a reconstruction model data set.

[0060] Step S2: Delineate global regions through K-means clustering based on the characteristics of the temperature-salinity profile.

[0061] Step S3: Construct a regional FFNN temperature-salinity profile reconstruction model according to the partitions to obtain a joint reconstruction model of the global ocean temperature-salinity profile.

[0062] Step S4: Acquire real-time sea surface remote sensing data, and obtain real-time temperature-salinity profile data based on the global ocean temperature-salinity profile joint reconstruction model.

[0063] The specific steps to obtain historical sea surface environmental parameter remote sensing data and underwater temperature and salinity profile data and construct the reconstruction model data set are as follows:

[0064] Step S1.1: Download the historical temperature-salinity profile dataset IAP from the website of the Institute of Atmospheric Physics, Chinese Academy of Sciences (http: / / www.ocean.iap.ac.cn / ftp / cheng / CODC-GOSD_database / ), use the historical temperature-salinity profiles from 2002 to 2021 as training data, calculate the statistical regression relationship of the surface and subsurface layers, and the profiles of 2022 and 2023 are used as independent verification data and test data for inverting temperature-salinity profiles. The selected profile data must meet the following four conditions: (1) temperature and salinity are measured simultaneously; (2) the measurement depth reaches more than 1000m; (3) the minimum measurement depth is less than 5m (4) the three data quality control symbols Temperature_CODCflag, Depth_WODflag, Salinity_CODCflag are all '0' (high quality).

[0065] Step S1.2: Use cubic spline interpolation to vertically interpolate the temperature-salinity profile selected in S1.1 to 20 inhomogeneous layers of [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500, 600, 700, 800, 900, 1000] m.

[0066] Step S1.3: Select the delayed quality-controlled L4-level 0.25° daily gridded sea surface height anomaly (SLA; Sea Level Anomaly) data product released by the Copernicus Marine Environment Monitoring Service (CMEMS), with a time range of 2002-2023.

[0067] Step S1.4: Interpolate the sky-averaged SLA data with a spatial resolution of 0.25° to the corresponding position of the historical temperature-salinity profile.

[0068] Step S1.5: Construct the reconstruction model dataset Y. The aggregation method of dataset Y is as follows:

[0069]

[0070] Among them, each sample in the data set Y is recorded as y (i,j)It is represented as the jth element of the ith sample, and the longitude and latitude of all samples are consistent; i = [1, N] represents the 1st to Nth samples, one sample is a row in the matrix Y, each row contains 43 elements, j = [1, 43] represents the 1st to 43rd elements, corresponding to the 43 columns of the matrix Y; the first and second columns are longitude and latitude respectively, and the third column is the input data sea surface height; the ocean temperature and salinity profile data are divided into [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500 , 600, 700, 800, 900, 1000]m; j = [4, 23] represents the temperature of the 20 depth layers respectively, and j = [24, 43] represents the salinity of the 20 depth layers respectively; the paired data set Y is divided into training set Y1, validation set Y2 and test set Y3 in a ratio of 8:1:1; the specific division method is: select the first 0.8×N rows of samples as training set Y1, the remaining 0.1×N rows of samples as validation set X2, and 0.1×N rows of samples as test set Y3, and there is no intersection between the three.

[0071] Step S1.6: Normalize the input data. Select the MinMax normalization method to normalize the temperature, salinity and sea surface height data in the input data set Y to the interval [0,1]. The normalization formula is as follows:

[0072]

[0073] The input dataset Y represents the dataset constructed in step 1.5, y j Represents an input ocean feature in the dataset, Y ijmin is the minimum value of the i-th factor in the data set Y, Y ijmax is the maximum value of the i-th factor in the data set Y. The set of factors normalized according to the formula is recorded as Y' j ; Y ijmin With Y ijmax Will be retained. During model validation and testing, the normalization of a batch of input data in the validation set and test set also uses Y ijmin With Y ijmax The calculation is performed because the statistical information contained in the training set is considered to be a known condition, while the statistical information of the test set and the validation set is an unknown condition. The purpose is to simulate the reasoning effect of the model on the unknown data set based on the existing information in most cases.

[0074] Step S2: According to the characteristics of the temperature-salinity profile, the global regions are delineated by K-means clustering. The specific steps are as follows:

[0075] Step S2.1. Construct a clustering dataset. Download the monthly average 1°×1° temperature-salinity profile data from 2002 to 2023 from the website of the Institute of Atmospheric Physics, Chinese Academy of Sciences (http: / / www.ocean.iap.ac.cn / ftp / cheng / CODC-GOSD_database / ). The temperature-salinity data are sampled at a depth of [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500, 600, 700, 800, 900, 1000] (unit: meter), and the multi-year average is calculated to obtain the 20-year average temperature-salinity dataset at different longitudes and latitudes. The dataset is constructed as the training set X in the following format:

[0076]

[0077] Among them, each element in the data set X is recorded as x i,j , x i,j represents the jth element of the ith sample, and the longitude and latitude of all samples are consistent; i = [1, N] represents the 1st to Nth samples, one sample is a row in the matrix X, each row contains 42 elements, j = [1, 42] represents the 1st to 42nd elements, corresponding to the 42 columns of the matrix X: the first two columns are the longitude and latitude of the input data, the 3rd to 22nd columns are the temperature values ​​of each sample profile in 20 depth layers, and the 23rd to 42nd columns are the salinity values ​​of each sample profile in 20 depth layers;

[0078] Step S2.2. Select the K value range: according to the characteristics of ocean water masses, the preset K value range is [2,20];

[0079] Step S2.3. Randomly select K samples from the training data set as the initial cluster centers.

[0080] Step S2.3.1. For each sample X i , calculate the Euclidean distance between it and the K cluster centers respectively, and record its distance to the jth cluster center as d ij , j∈[1,K];

[0081] Step S2.3.2. Compare the distances between the sample and the cluster center and get D i =min{d ij} i , each sample can select a cluster center that is closest to it, and merge the samples with the same nearest cluster center into one cluster, thus obtaining K sample clusters;

[0082] Step S2.3.3. Calculate the mean of the samples in each cluster and use the mean as the new cluster center;

[0083] Step S2.3.4. If the new cluster center is different from the original cluster center, repeat steps S2.3.1-S2.3.3 until the cluster center no longer changes or the predetermined number of iterations m is reached.

[0084] Step S2.4. Output the cluster division and obtain the clustering results corresponding to different longitude and latitude sea areas as shown in the figure. Calculate the silhouette coefficient of the entire data set. The silhouette coefficient is an indicator for evaluating clustering quality. Its value range is between [-1,1]. The closer the value is to 1, the better the clustering effect.

[0085] Step S2.5. Repeat steps S2.2-S2.4, compare the silhouette coefficients of clustering results under different K values, and obtain the optimal value of K value and its corresponding clustering result, see Figure 2 After calculation, when K = 7, the silhouette coefficient is the largest, which is 0.327. According to the clustering results, the division of sea areas at different longitudes and latitudes is obtained according to the properties of the temperature and salinity profiles. The sea areas are divided according to the clustering results. Figure 3 .

[0086] Step S3: Construct a regional FFNN temperature-salinity profile reconstruction model according to the partitions to obtain a global ocean temperature-salinity profile joint reconstruction model. The specific steps are as follows:

[0087] Step S3.1: Build the FFNN feedforward neural network reconstruction model. Assuming that n feature variables and m samples are input, the input layer X∈R m×n , construct a hidden layer with h neurons, then the weight and bias of the hidden layer can be expressed as W h ∈R n×h , b h ∈R 1×h , if there are q output label values, the weight and bias parameters of the output layer are W o ∈R h×q , b O ∈R 1 ×q , the output of the hidden layer and the output of the output layer are expressed as follows:

[0088] H=XW h +b h

[0089] O=HW o +b O

[0090] Build three hidden layers. The first hidden layer is set to h1 neurons, the second hidden layer has h2 neurons, and the third hidden layer has h3 neurons. Each layer is connected according to the above formula; each hidden layer is followed by a ReLU activation layer, the formula is as follows:

[0091] ReLU(input) = max(0, input)

[0092] Reference Figure 4 , build two hidden layers. The first hidden layer has 128 neurons, and the second hidden layer has 64 neurons. Each layer is connected according to the above formula; each hidden layer is followed by a ReLU activation layer. Adding an activation layer introduces nonlinear transformation to improve the expressiveness of the model.

[0093] Step S3.2: Select the mean square error (MSE) as the loss function of the FFNN reconstruction model to measure the difference between the reconstructed underwater temperature and salinity profile and the measured temperature and salinity profile. The formula is as follows:

[0094]

[0095] Among them, Y i Indicates the label of sample i, i.e. the measured temperature-salinity profile data, Y' i Represents the value reconstructed by the model for sample i.

[0096] Step S3.3: According to the partitioning result of S2, the dataset Y constructed in S1 is divided into seven regional datasets, which are respectively put into the FFNN reconstruction model built in S3.1 for training, and the hidden layer weights and bias W are fitted. h 、b h , the loss function is MSE, the number of initial iterations is epoch, epoch is a positive integer, the initial learning rate is lr, the initial batch size is batch size, and the optimizer is Adam.

[0097] Step S3.4: Save the trained FFNN models and model output parameters of the seven regions. The seven regional reconstruction models constitute the global ocean temperature and salinity profile joint reconstruction model.

[0098] Step S4: Acquire real-time sea surface remote sensing data, and obtain real-time temperature-salinity profile data based on the global ocean temperature-salinity profile joint reconstruction model. The specific steps are as follows:

[0099] Step S4.1: Obtain real-time sea surface remote sensing temperature and salinity data as model input;

[0100] Step S4.2: Data is input into the global ocean temperature-salinity profile joint reconstruction model constructed in S3 to obtain the real-time global ocean temperature-salinity profile data field.

[0101] The above specific embodiments are merely explanations of the present invention and are not limitations of the present invention. After reading this specification, those skilled in the art may make modifications to the embodiments without any creative contribution as needed. However, such modifications are protected by the patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for joint reconstruction of global ocean temperature and salinity profiles based on deep learning, characterized by: The following steps are involved: Step S1, obtaining historical sea surface environmental parameter remote sensing data and underwater temperature and salinity profile data, and constructing a reconstruction model data set; Step S2, according to the characteristics of temperature and salinity profiles, global regions are delineated by K-means clustering; Step S3, constructing a regional FFNN temperature-salinity profile reconstruction model according to the partitions to obtain a global ocean temperature-salinity profile joint reconstruction model; Step S4: Acquire real-time sea surface remote sensing data, and obtain real-time temperature-salinity profile data based on the global ocean temperature-salinity profile joint reconstruction model.

2. The method for joint reconstruction of global ocean temperature and salinity profiles based on deep learning according to claim 1, characterized in that: The specific sub-steps of step S1 are: Step S1.1, collect historical temperature-salinity profile data as training data, independent verification data and test data for inverting temperature-salinity profiles, calculate the statistical regression relationship between the surface layer and the subsurface layer, and the selected historical temperature-salinity profile data must meet the following four conditions: (1) temperature and salinity are measured simultaneously; (2) the measurement depth reaches more than 1000m; (3) the minimum measurement depth is less than 5m; (4) the three data quality control symbols Temperature_CODCflag, Depth_WODflag, and Salinity_CODCflag are all "0"; Step S1.2, use cubic spline interpolation to vertically interpolate the temperature-salinity profile selected in S1.1 to 20 inhomogeneous layers of [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500, 600, 700, 800, 900, 1000] m; Step S1.3, select the delayed quality controlled L4 level 0.25° daily gridded sea surface height anomaly data product released by the Copernicus Marine Environment Monitoring Service; Step S1.4, interpolate the sky-averaged SLA data with a spatial resolution of 0.25° to the corresponding position of the historical temperature-salinity profile; Step S1.5: Construct a reconstruction model dataset Y. The aggregation method of dataset Y is as follows: Among them, each sample in the data set Y is recorded as y (i,j) It is represented as the jth element of the ith sample, and the longitude and latitude of all samples are consistent; i = [1, N] represents the 1st to Nth samples, one sample is a row in the matrix Y, each row contains 43 elements, j = [1, 43] represents the 1st to 43rd elements, corresponding to the 43 columns of the matrix Y; the first and second columns are longitude and latitude respectively, and the third column is the input data sea surface height; the ocean temperature and salinity profile data are divided into [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500 , 600, 700, 800, 900, 1000] m; j = [4, 23] represents the temperature of 20 depth layers in turn, and j = [24, 43] represents the salinity of 20 depth layers in turn; the paired data set Y is divided into training set Y1, validation set Y2 and test set Y3 in a ratio of 8:1:1; the specific division method is: select the first 0.8×N rows of samples as training set Y1, the remaining 0.1×N rows of samples as validation set X2, and 0.1×N rows of samples as test set Y3, and there is no intersection between the three; Step S1.6: Normalize the input data. Select the MinMax normalization method to normalize the temperature, salinity and sea surface height data in the input data set Y to the interval [0,1]. The normalization formula is as follows: The input dataset Y represents the dataset constructed in step S1.5, y j Represents an input ocean feature in the dataset, Y ijmin is the minimum value of the i-th factor in the data set Y, Y ijmax is the maximum value of the i-th factor in the data set Y. The set of factors normalized according to the formula is recorded as Y' j ; Y ijmin With Y ijmax Will be retained. During model validation and testing, the normalization of a batch of input data in the validation set and test set also uses Y ijmin With Y ijmax The calculation is performed because the statistical information contained in the training set is considered to be a known condition, while the statistical information of the test set and the validation set is an unknown condition. The purpose is to simulate the reasoning effect of the model on the unknown data set based on the existing information in most cases.

3. The method for joint reconstruction of global ocean temperature and salinity profiles based on deep learning according to claim 1, characterized in that: The specific sub-steps of step S2 are: Step S2.1, construct a clustering data set, collect monthly average 1°×1° temperature and salinity profile data, sample the temperature and salinity data at a depth of [5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500, 600, 700, 800, 900, 1000] (unit: meter), calculate the multi-year average value, and obtain the annual average temperature and salinity data set at different longitudes and latitudes. The data set is constructed as a training set X in the following format: Among them, each element in the data set X is recorded as x i,j , x i,j represents the jth element of the ith sample, and the longitude and latitude of all samples are consistent; i = [1, N] represents the 1st to Nth samples, one sample is a row in the matrix X, each row contains 42 elements, j = [1, 42] represents the 1st to 42nd elements, corresponding to the 42 columns of the matrix X: the first two columns are the longitude and latitude of the input data, the 3rd to 22nd columns are the temperature values ​​of each sample profile in 20 depth layers, and the 23rd to 42nd columns are the salinity values ​​of each sample profile in 20 depth layers; Step S2.2, select the K value range: according to the characteristics of ocean water masses, the preset K value range is [2,20]; Step S2.3: Randomly select K samples from the training data set as the initial cluster centers: Step S2.3.1: For each sample X i , calculate the Euclidean distance between it and the K cluster centers respectively, and record its distance to the jth cluster center as d ij , j∈[1,K]; Step S2.3.2: Compare the distances from the sample to the cluster center and get D i =min{d ij } i , each sample can select a cluster center that is closest to it, and merge the samples with the same nearest cluster center into one cluster, thus obtaining K sample clusters; Step S2.3.3, calculate the mean of the samples in each cluster and use the mean as the new cluster center; Step S2.3.4: If the new cluster center is different from the original cluster center, repeat steps S2.3.1-S2.3.3 until the cluster center no longer changes or the predetermined number of iterations m is reached; Step S2.4, output cluster division, and obtain the clustering results corresponding to different longitude and latitude sea areas as shown in the figure; calculate the silhouette coefficient of the entire data set. The silhouette coefficient is an indicator for evaluating clustering quality. Its value range is between [-1, 1]. The closer the value is to 1, the better the clustering effect is. Step S2.5, repeat steps S2.2-S2.4, compare the silhouette coefficients of the clustering results under different K values, obtain the optimal value of K value and its corresponding clustering result. After calculation, when K=7, the silhouette coefficient is the largest, which is 0.327; according to the clustering results, obtain the division of sea areas at different longitudes and latitudes according to the properties of temperature-salinity profiles, and divide the sea areas according to the clustering results.

4. The method for joint reconstruction of global ocean temperature and salinity profiles based on deep learning according to claim 1, characterized in that: The specific sub-steps of step S3 are: Step S3.1: Build the FFNN feedforward neural network reconstruction model. Assuming that n feature variables and m samples are input, the input layer X∈R m×n , construct a hidden layer with h neurons, then the weight and bias of the hidden layer can be expressed as W h ∈R n×h , b h ∈R 1×h , if there are q output label values, the weight and bias parameters of the output layer are W o ∈R h×q , b O ∈R 1×q , the output of the hidden layer and the output of the output layer are expressed as follows: H=XW h +b h O=HW o +b O Build three hidden layers. The first hidden layer is set to h1 neurons, the second hidden layer has h2 neurons, and the third hidden layer has h3 neurons. Each layer is connected according to the above formula; each hidden layer is followed by a ReLU activation layer, the formula is as follows: ReLU(input) = max(0, input) Build two hidden layers. The first hidden layer has 128 neurons, and the second hidden layer has 64 neurons. Each layer is connected according to the above formula. Each hidden layer is followed by a ReLU activation layer. Adding an activation layer introduces nonlinear transformation to improve the expressiveness of the model. Step S3.2: Select mean square error (MSE) as the loss function of the FFNN reconstruction model to measure the difference between the reconstructed underwater temperature and salinity profile and the measured temperature and salinity profile. The formula is as follows: Among them, Y i Indicates the label of sample i, i.e. the measured temperature-salinity profile data, Y' i Represents the value reconstructed by the model of sample i; Step S3.3: According to the partitioning result of S2, the dataset Y constructed in S1 is divided into seven regional datasets, which are respectively put into the FFNN reconstruction model constructed in S3.1 for training, and the hidden layer weights and bias W are fitted. h , b h , the loss function is MSE, the number of initial iterations is epoch, epoch is a positive integer, the initial learning rate is lr, the initial batch size is batchsize, and the optimizer is Adam; Step S3.4, save the trained FFNN models and model output parameters of the seven regions, and the seven regional reconstruction models constitute a global ocean temperature and salinity profile joint reconstruction model.

5. The method for joint reconstruction of global ocean temperature and salinity profiles based on deep learning according to claim 1, characterized in that: The specific sub-steps of step S4 are: Step S4.1, obtaining real-time sea surface remote sensing temperature and salinity data as model input; Step S4.2: Input data into the global ocean temperature-salinity profile joint reconstruction model constructed in S3 to obtain the real-time global ocean temperature-salinity profile data field.

Images