A method and system for inverting three-dimensional temperature and salinity field of ocean subsurface

Abstract

The present invention belongs to the field of seawater temperature and salinity information technology, and discloses a method and system for inverting a three-dimensional temperature and salinity field of the ocean subsurface. The convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM are integrated, and an attention mechanism is added to construct an ocean subsurface temperature and salinity inversion model (CBA) integrating the convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM; the sea surface temperature difference / sea surface salinity difference DSST / DSSS is applied to the ocean subsurface temperature and salinity inversion model, and the ocean subsurface temperature and salinity are inverted using sea surface satellite remote sensing data; by comparing the model inversion value and the true value of the subsurface seawater temperature and salinity, the RMSE and R² are calculated to evaluate the performance of the ocean subsurface temperature and salinity inversion model. The present invention uses the subsurface seawater temperature and salinity data with a water depth of 30-1000m to verify the effectiveness of the model.

Description

Technical Field

[0001] The invention belongs to the technical field of seawater temperature and salt field information, and in particular relates to a method and system for inverting a three-dimensional temperature and salt field of a subsurface ocean. Background Art

[0002] In the work of monitoring and describing the subsurface thermal structure of the global ocean, studying and accurately estimating the ocean subsurface temperature and salinity is crucial to understanding the mechanisms and processes of the entire ocean and the entire Earth's climate system. It can also provide data support for the construction of mid- and deep-ocean datasets and ocean warming analysis.

[0003] The temperature of sea water in the ocean can reflect marine phenomena, such as the most obvious El Niño phenomenon. Due to the abnormal warming of sea water in the South Pacific, precipitation in the desert areas of South America has increased, while drought has occurred in the humid areas of eastern Australia. Therefore, predicting ocean temperature anomalies is very necessary to provide information support for understanding the three-dimensional physical field inside the ocean.

[0004] Satellites can provide a wide range of ocean remote sensing information, but their limitation is that satellite remote sensing can only observe the ocean surface. In order to obtain more abundant ocean parameters below the sea surface, it is necessary to use satellite remote sensing inversion technology, combined with the parameter characteristics of the ocean surface, such as sea surface temperature (SST), sea surface salinity (SSS), sea surface height (SSH) and sea surface wind field, to invert and calculate the parameter information below the sea surface, so as to improve the utilization value and efficiency of remote sensing data.

[0005] In the existing satellite remote sensing intelligent inversion algorithms, three types of neural networks are commonly used: fully connected neural network (ANN), convolutional neural network (CNN) and long short-term memory network (LSTM); among them, ANN can capture sufficient information, but its calculation amount is relatively large; CNN can capture spatial features while reducing the amount of calculation, but it has the problem of information loss; LSTM can better capture temporal regularity characteristics, but it is lacking in spatial feature capture.

[0006] Through the above analysis, the problems and defects of the existing technology are as follows: Accurately obtaining the temperature and salinity information of the subsurface seawater is of great significance for studying the physical, biological and chemical processes of the ocean. At present, the ocean field observation data is sparse, and the satellite remote sensing data is rich, but it can only obtain sea surface information; it is impossible to further accurately obtain the temperature and salinity data of the subsurface seawater with a depth of 30-1000m. Summary of the invention

[0007] In order to overcome the problems existing in the related art, the disclosed embodiments of the present invention provide a method and system for inverting the three-dimensional temperature-salinity field of the subsurface ocean.

[0008] The technical scheme is as follows: a method for inverting the three-dimensional temperature and salinity field of the ocean subsurface, which takes into account the correlation between the physical movement of the ocean surface and subsurface seawater, and fully integrates the spatial and temporal correlation between the sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface; specifically includes the following steps:

[0009] S1, based on the convolutional neural network CNN, the spatial feature correlation between the sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface is extracted, and based on the bidirectional long short-term memory network BiLSTM, the temporal feature correlation between the sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface is extracted. The convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM are integrated, and the attention mechanism is added to construct an ocean subsurface temperature and salinity inversion model (CBA) integrating the convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM.

[0010] S2, apply the sea surface temperature difference DSST / sea surface salinity difference DSSS to the ocean subsurface temperature and salinity field inversion model, conduct subsurface seawater temperature and salinity inversion experiments, including data set division, model training and model testing, and complete the inversion of the ocean subsurface temperature and salinity field using sea surface satellite remote sensing data;

[0011] S3, by comparing the model inversion values ​​of subsurface seawater temperature and salinity with the true values, RMSE and R 2 Evaluating the performance of the ocean subsurface temperature-salinity field inversion model.

[0012] Before step S1, you need to do:

[0013] Collect sea surface satellite remote sensing data, including sea surface temperature SST, sea surface salinity SSS, sea surface height SSH, sea surface wind SSW and three-dimensional temperature and salinity 3DTS data. The sea surface wind SSW is divided into the east component USSW and the north component VSSW. Use sea surface satellite remote sensing data to invert the ocean subsurface temperature field and salinity field respectively;

[0014] The difference between the sea surface temperature / sea surface salinity and the multi-year monthly average of seawater temperature and salinity at different depths in the corresponding month, namely the sea surface temperature difference DSST / sea surface salinity difference DSSS, is introduced and combined with the above-mentioned sea surface satellite remote sensing data for the inversion of subsurface seawater temperature and salinity field;

[0015] The ocean subsurface temperature and salinity field uses three-dimensional temperature and salinity 3DTS. The 3DTS data includes the seawater temperature and salinity of multiple different depth layers, which is used for the construction and evaluation of the ocean subsurface temperature and salinity field inversion model;

[0016] All data used are monthly average data. The monthly average values ​​of SST and SSH are calculated from daily data respectively, and the SST data are interpolated to a grid of 0.25°×0.25°. To ensure spatial consistency, if a node has an empty variable, the node will be deleted. Before all data are input into the subsurface seawater temperature and salinity inversion model, all data are standardized according to the channel dimension so that the mean of the data set is 0 and the variance is 1, eliminating the quantitative differences between different data.

[0017] In step S1, the ocean subsurface temperature and salinity field inversion model includes two branches, the first branch is used to extract the spatial dimension characteristics of the sea surface remote sensing data, and the second branch is used to extract the temporal dimension characteristics of the sea surface remote sensing data;

[0018] The first branch consists of a convolutional attention module, a convolutional neural network (CNN), and a fully connected neural network. Before the data enters the convolutional layer, it is passed to the CBAM layer to highlight important features and enhance the feature expression capability. After the convolutional layer and the activation layer, the CBAM layer is added to filter and enhance the features extracted by the convolutional layer, so that the feature fusion module can use these feature information to improve the overall performance of the model. After each convolutional layer, a batch normalization layer, an activation layer, and a random dropout layer are introduced. The activation layer uses ReLU as the activation function.

[0019] Furthermore, for a given feature map, the convolutional attention module calculates the attention map in sequence according to two independent dimensions, channel and space, multiplies the attention map with the input feature map, performs adaptive feature refinement, and selects a 11×11 local area centered on the true value as input.

[0020] Furthermore, the second branch consists of a one-to-many attention mechanism, a bidirectional long short-term memory network BiLSTM and a fully connected neural network; a random dropout layer is added after the first BiLSTM layer to alleviate overfitting; a one-to-many attention layer is added after the second BiLSTM layer; the specific steps are:

[0021] The attention score is calculated by performing a dot product operation based on the hidden state of the last time step and the hidden state of all time steps, and normalizing them through softmax. The attention score is obtained by performing a dot product operation based on the hidden state of the last time step and the hidden state of all time steps, and normalizing them through softmax. The hidden state of all time steps is weightedly summed up by the attention score to obtain the context vector of the relevant hidden state, and then the context vector is residually connected with the hidden state of the last time step. The context vector is added to the hidden state of the last time step as a residual connection to obtain the final output.

[0022] In step S1, an ocean subsurface temperature and salinity inversion model integrating a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) network is constructed, including:

[0023] The spatial features are extracted through the convolution layer and CBAM, and the temporal features are extracted through BiLSTM and the attention mechanism, and then flattened and converted to channel dimensions through the fully connected layer; the input data processed by these two branches are spliced ​​into a new comprehensive variable in the channel dimension to achieve the fusion of temporal and spatial information; this new comprehensive variable is passed to the last two fully connected layers, the first fully connected layer is used to expand the channel dimension of the new variable, and the second fully connected layer is used to obtain the final output; information fusion is achieved, avoiding information loss or overfitting that may be caused by direct mapping; the inversion model of the ocean subsurface temperature and salinity field is:

[0024]

[0025] In the formula, Respectively represent the spatial latitude features and the temporal latitude features output by the first branch and the second branch; Represents a splicing operation, and represents the weights and biases of the last two fully connected layers, Represents the final output;

[0026] The performance of the ocean subsurface temperature and salinity inversion model is optimized. The ocean subsurface temperature and salinity inversion model adopts an early stopping algorithm and selects the ocean subsurface temperature and salinity inversion model with the best performance on the validation data set. At the same time, the Adam optimizer is used to adjust the learning process of the ocean subsurface temperature and salinity inversion model.

[0027] In step S2, the sea surface temperature difference DSST / sea surface salinity difference DSSS is applied to the ocean subsurface temperature and salinity field inversion model, including:

[0028] Taking the three-dimensional temperature-salinity data 3DT as the real observation value of subsurface seawater temperature and salinity, the multi-source sea surface satellite remote sensing data of sea surface height SSH, sea surface salinity SSS, sea surface temperature SST, sea surface wind speed east component Uwnd, sea surface wind speed north component Vwnd are applied to the ocean subsurface temperature-salinity field inversion model to conduct subsurface seawater temperature and salinity inversion experiments;

[0029] In the data set division stage, the sea surface satellite remote sensing data and three-dimensional temperature and salinity data are processed, and the ocean subsurface temperature and salinity field inversion model accepts time and space data at the same time, so that it meets the input requirements of the first branch and the second branch respectively;

[0030] Introduce new variables, sea surface temperature difference DSST / sea surface salinity difference DSSS, and use them together with sea surface satellite remote sensing data as input data for the ocean subsurface temperature and salinity field inversion model, and participate in the inversion of subsurface seawater temperature and salinity fields;

[0031] In the training phase of the ocean subsurface temperature and salinity inversion model, the sea surface satellite remote sensing data and the matching three-dimensional temperature and salinity data in the training data set are used to train the ocean subsurface temperature and salinity inversion model. The performance of the ocean subsurface temperature and salinity inversion model is evaluated by calculating the mean square error (MSE) between the inverted values ​​and the true values ​​of seawater temperature and salinity. The Adam optimization algorithm is used to iteratively optimize the parameters of the ocean subsurface temperature and salinity inversion model.

[0032] During the model testing phase, the sea surface satellite remote sensing data of the test data set are input into the trained ocean subsurface temperature and salinity field inversion model.

[0033] Furthermore, in the data set division stage, for the first branch of the ocean subsurface temperature and salinity field inversion model, a 11×11 local area centered on the true value is selected as input. The number of invalid value points in the selected local area should be less than half of the total number and the center point of the area should be a valid value. For the second branch of the ocean subsurface temperature and salinity field inversion model, the BiLSTM time step is set to 10, that is, the sea surface satellite remote sensing data of the first 10 months are used to invert the subsurface seawater temperature field and salinity field of the 10th month.

[0034] The calculation formula of the newly introduced variable sea surface temperature difference DSST / sea surface salinity difference DSSS is as follows:

[0035]

[0036]

[0037]

[0038]

[0039] In the formula, is the sea surface temperature difference, is the sea surface salinity difference, is the ocean subsurface temperature, is the temperature difference, is the ocean subsurface salinity, The salinity difference is For the month, is the total number of years, For the year, is the sea surface temperature, is the sea surface salinity; and As a fixed value, it only needs to be calculated once, and in subsequent experiments, only the specific sea surface temperature / sea surface salinity is used minus / Get DSST / DSSS.

[0040] Another object of the present invention is to provide a system for inverting the three-dimensional temperature-salinity field of the subsurface ocean, which implements the three-dimensional temperature-salinity field inversion method of the subsurface ocean, and the system comprises:

[0041] The module for building the ocean subsurface temperature and salinity inversion model is based on the convolutional neural network (CNN) to extract the spatial feature correlation between the sea surface satellite remote sensing data and the ocean subsurface three-dimensional temperature and salinity field. The module is based on the bidirectional long short-term memory (BiLSTM) network to extract the temporal feature correlation between the sea surface satellite remote sensing data and the ocean subsurface three-dimensional temperature and salinity field. The convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network are integrated, and the attention mechanism is added to build an ocean subsurface temperature and salinity inversion model that integrates the convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network.

[0042] The ocean subsurface temperature and salinity inversion module is used to apply the sea surface temperature difference DSST / sea surface salinity difference DSSS to the ocean subsurface temperature and salinity inversion model, conduct subsurface seawater temperature and salinity inversion experiments, including data set division, model training and model testing, and complete the inversion of the ocean subsurface temperature and salinity field using sea surface satellite remote sensing data;

[0043] Performance verification module, used to calculate RMSE and R by comparing the model inversion values ​​of subsurface sea temperature and salinity with the true values 2 Evaluating the performance of the ocean subsurface temperature-salinity field inversion model.

[0044] Furthermore, the ocean subsurface three-dimensional temperature-salinity field inversion system is mounted on a computer-readable storage medium, and the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the functions of the above-mentioned ocean subsurface three-dimensional temperature-salinity field inversion system can be realized.

[0045] Combined with all the above technical solutions, the beneficial effects of the present invention are as follows: the present invention integrates convolutional neural network (CNN) and bidirectional long short-term memory network (BiLSTM), and adds attention mechanism to construct a method for inverting ocean subsurface temperature and salinity field integrating CNN and BiLSTM, and proposes a new parameter - sea surface temperature difference / sea surface salinity difference (DSST / DSSS) for inverting ocean subsurface temperature and salinity field, so as to realize the inversion of ocean subsurface temperature and salinity field using sea surface satellite remote sensing data. Taking the North Pacific as the test area, the inversion of ocean subsurface temperature and salinity field based on sea surface satellite remote sensing data is carried out. The input data of the inversion model include sea surface satellite remote sensing data such as sea surface height (SSH), sea surface temperature (SST), sea surface salinity (SSS) and sea surface wind speed (SSW) and the newly proposed parameter (sea surface temperature difference / sea surface salinity difference). The present invention uses subsurface sea water temperature and salinity field data with a water depth of 30-1000m to verify the effectiveness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure;

[0047] Figure 1 It is a flow chart of the method for inverting the three-dimensional temperature and salinity field of the ocean subsurface provided by an embodiment of the present invention;

[0048] Figure 2 It is a schematic diagram of the architecture of an ocean subsurface temperature and salinity inversion model that integrates a convolutional neural network and a bidirectional long short-term memory network provided by an embodiment of the present invention;

[0049] Figure 3 is an architecture diagram of a convolutional attention module provided by an embodiment of the present invention;

[0050] Figure 4 is a schematic diagram of the structure of a one-to-many attention module provided by an embodiment of the present invention;

[0051] Figure 5 is a flow chart of sea surface temperature difference / sea surface salinity calculation provided by an embodiment of the present invention;

[0052] Figure 6 is a flowchart of constructing a data set provided by an embodiment of the present invention;

[0053] Figure 7 It is a flow chart of using the CBA model to invert the subsurface ST / SS field using sea surface satellite remote sensing data provided by an embodiment of the present invention;

[0054] Figure 8 The distribution diagram of the coefficient of determination (R²) in the accuracy evaluation of the seawater temperature inversion results at different depths in the North Pacific Ocean for the Case 1-Case 4 experimental cases;

[0055] Fig. 9 This is the root mean square error (RMSE) distribution diagram of the accuracy evaluation of the seawater temperature inversion results at different depths in the North Pacific Ocean for the Case 1-Case 4 experimental cases;

[0056] Fig.10 This is the distribution diagram of the coefficient of determination (R²) in the accuracy evaluation of the inversion results of seawater salinity at different depths in the North Pacific Ocean for the Case 5-Case 8 experiments;

[0057] Fig.11 This is the root mean square error (RMSE) distribution diagram of the accuracy evaluation of the seawater temperature inversion results at different depths in the North Pacific Ocean for the Case 5-Case 8 experimental cases;

[0058] Fig.12 The R² of the seawater temperature inversion results of different models varies with seawater depth;

[0059] Fig.13 The R² of the seawater salinity inversion results of different models varies with seawater depth;

[0060] Fig.14 The RMSE of the seawater temperature inversion results of different models varies with the seawater depth;

[0061] Fig.15 This is a graph showing how the RMSE of seawater salinity inversion results from different models changes with seawater depth. DETAILED DESCRIPTION

[0062] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of 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 violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0063] Embodiment 1, as Figure 1 As shown, an embodiment of the present invention provides a method for inverting the three-dimensional temperature and salinity field of the ocean subsurface, which takes into account the correlation between the physical movement of the ocean surface and subsurface seawater, and fully integrates the spatial and temporal correlation of sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface; specifically, the following steps are included:

[0064] S1, based on the convolutional neural network CNN, the spatial feature correlation between the sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface is extracted, and based on the bidirectional long short-term memory network BiLSTM, the temporal feature correlation between the sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface is extracted. The convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM are integrated, and the attention mechanism is added to construct an ocean subsurface temperature and salinity inversion model (CBA) integrating the convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM.

[0065] S2, apply the sea surface temperature difference DSST / sea surface salinity difference DSSS to the ocean subsurface temperature and salinity field inversion model, conduct subsurface seawater temperature and salinity inversion experiments, including data set division, model training and model testing, and complete the inversion of the ocean subsurface temperature and salinity field using sea surface satellite remote sensing data;

[0066] S3, by comparing the model inversion values ​​of subsurface seawater temperature and salinity with the true values, RMSE and R 2 Evaluating the performance of the ocean subsurface temperature-salinity field inversion model.

[0067] Example 2, as another implementation of the present invention, the method for inverting the three-dimensional temperature-salinity field of the ocean subsurface provided by the embodiment of the present invention is further described in detail by taking the North Pacific Ocean (120°E-90°W, 0°-60°N) as an experimental area as an example.

[0068] (1) Data and experimental area.

[0069] The North Pacific (120°E-90°W, 0°-60°N) was selected as the experimental area. As a key area that has a significant impact on both local and global climate, the North Pacific is an important area for the analysis of marine environment and climate change. The present invention uses sea surface satellite remote sensing data to invert the ocean subsurface temperature field and salinity field of the North Pacific respectively. Sea surface satellite remote sensing data include sea surface temperature (SST), sea surface salinity (SSS), sea surface height (SSH) and sea surface wind (SSW), where the sea surface wind is divided into the east component (USSW) and the north component (VSSW). In addition, the present invention introduces the difference between the sea surface temperature / sea surface salinity and the average seawater temperature and salinity at different depths in the corresponding month, which is called the sea surface temperature difference / sea surface salinity difference (DSST / DSSS), which is combined with the above-mentioned sea surface satellite remote sensing data for the inversion of subsurface seawater temperature and salinity field. The SSH, SSS, SST and three-dimensional temperature and salinity (3DTS) used in this invention are all from the Copernicus Ocean Service (https: / / data.marine.copernicus.eu / ), with a spatial resolution of 0.25°×0.25°. The temporal resolution of SSH data is daily, and SSS and 3DTS are provided in the form of monthly average data. SSH data are generated by the DUACS processing system and contain data from multiple satellite altimeters. SSS data are obtained through a multi-source optimal interpolation algorithm, combining sea surface salinity observation data from multiple satellites, such as NASA's SMAP satellite and ESA's SMOS satellite, which provide satellite sea surface salinity information. 3DTS data has 50 different depth layers. This paper selects some depth layers (30m, 50m, 100m, 125m, 150m, 200m, 250m, 300m, 400m, 500m, 600m, 700m, 800m, 900m and 1000m) for the construction and evaluation of the inversion model. SST data are calculated based on multi-source SST data using the optimal interpolation (OI) method, with a temporal resolution of days and a spatial resolution of 0.05°×0.05°. SSW data are obtained by integrating data from different platforms through precise cross-calibration and processing by a cross-calibration multi-platform to provide consistent and accurate sea surface wind field observations on a global scale. The temporal and spatial resolutions of SSW data are monthly averages and 0.25°×0.25°.

[0070] All data used in this paper are monthly average data, ranging from January 2011 to December 2021. In order to ensure the consistency of data in time and space, the monthly average values ​​of SST and SSH are calculated from daily data, and the SST data are interpolated to a grid of 0.25°×0.25°. In order to achieve spatial consistency, a node is deleted if a variable is empty. Before the data is input into the model, it is standardized according to the channel dimension to ensure that the mean of the data set is 0 and the variance is 1, so as to eliminate the difference in the value between different data.

[0071] (2) Model architecture

[0072] In view of the problem that the existing ocean subsurface temperature and salinity inversion only considers the temporal information or spatial information of sea surface data, the present invention integrates CNN and BiLSTM, adds attention mechanism, uses dual input form, fully integrates the temporal and spatial characteristic information of sea surface data, and constructs an ocean subsurface temperature and salinity inversion model using sea surface satellite remote sensing data. The structure of the ocean subsurface temperature and salinity inversion model is as follows: Figure 2 The schematic diagram of the ocean subsurface temperature and salinity inversion model architecture that integrates convolutional neural networks and bidirectional long short-term memory networks is shown. The ocean subsurface temperature and salinity field model contains two branches. The first branch focuses on extracting the spatial dimension characteristics of sea surface remote sensing data, and the second branch focuses on extracting the temporal dimension characteristics of sea surface remote sensing data. Each branch has a different model structure and input data shape.

[0073] The first branch of the ocean subsurface temperature and salinity inversion model focuses on extracting spatial feature information of sea surface remote sensing data. The branch structure is mainly composed of a convolutional attention module, a convolutional neural network (CNN) and a fully connected neural network. After each convolutional layer, the present invention introduces a batch normalization layer, an activation layer and a random dropout layer to improve the performance and generalization ability of the network. The activation layer uses ReLU as the activation function. Sea surface data has a relatively complex distribution in space. The present invention first passes the processed input data into the convolutional attention module, and then passes it to the convolutional layer for processing. This helps the model to better understand the spatial structure of sea surface data. Sea surface data consists of a variety of observation data, and the convolutional attention mechanism can adaptively extract key information features in multi-channel data, so that the model can better utilize information from different channels. The present invention adds a convolutional attention module after each convolutional layer, and performs feature enhancement on the data processed by the convolutional layer, so that the ocean subsurface temperature and salinity inversion model can better adapt to different sea surface data samples and improve the generalization ability of the model.

[0074] The architecture of the convolutional attention module is as follows Figure 3(Convolutional block attention module (CBAM) structure). The convolutional attention module is a simple and efficient feedforward convolutional neural network attention module. For a given feature map, it calculates the attention map in sequence according to the two independent dimensions of channel and space, and then multiplies the attention map with the input feature map to perform adaptive feature refinement. This enables the ocean subsurface temperature and salinity field inversion model to focus on key variables and information features while ignoring irrelevant information. In the present invention, a local area of ​​11×11 centered on the true value is selected as input. This method can effectively extract the spatial dimension features of the input data. Compared with the traditional back propagation neural network (BPNN) method, this approach increases the amount of information contained in the features.

[0075] The second branch of the ocean subsurface temperature and salinity inversion model focuses on extracting the temporal characteristic information of sea surface data. This branch is mainly composed of a one-to-many attention mechanism, a bidirectional long short-term memory network (BiLSTM), and a fully connected neural network. A random dropout layer is added after the first BiLSTM layer to alleviate the overfitting phenomenon. Since the hidden state of each time step output by the BiLSTM layer contributes differently to the model inversion result, the present invention adds a one-to-many attention layer ( Figure 4 Schematic diagram of the one-to-many attention module structure), so that the ocean subsurface temperature and salinity field inversion model pays more attention to important features. The specific steps are: the attention score is calculated based on the hidden state of the last time step and the hidden state of all time steps, and the hidden state of all time steps is weighted summed by these attention scores to obtain the context vector of the relevant hidden state, and then the context vector is residually connected with the hidden state of the last time step to obtain the final output. Introducing such a one-to-many attention layer can improve the accuracy of the BiLSTM model in the inversion task.

[0076] In the ocean subsurface temperature and salinity inversion model fusion module, the present invention splices the input data processed by the two branches in the channel dimension and combines them into a new comprehensive variable. The splicing step can effectively combine the time and space features extracted from the two branches. Then, the present invention passes this new comprehensive variable to the last two fully connected layers. The fully connected layer is used to further process and fuse the temporal and spatial feature information to achieve deep fusion of information. The ocean subsurface temperature and salinity inversion model obtained by this method can comprehensively consider the information of the time and space dimensions and obtain the final output result. This part can be expressed by the following equation, where Represent the spatial dimension features and temporal dimension features output by the first branch and the second branch respectively.

[0077]

[0078] In order to optimize the performance of the ocean subsurface temperature and salinity inversion model, the ocean subsurface temperature and salinity inversion model adopts an early stopping algorithm and selects the ocean subsurface temperature and salinity inversion model with the best performance on the validation data set. At the same time, the Adam optimizer is used to adjust the learning process of the ocean subsurface temperature and salinity inversion model. Table 1 lists the parameter combinations of the CBA model.

[0079] Table 1 Parameter combinations of CBA model

[0080]

[0081] (3) Experimental setup.

[0082] Taking three-dimensional temperature-salinity data as the real observation values ​​of subsurface seawater temperature and salinity, using multi-source sea surface satellite remote sensing data (SSH, SSS, SST, Uwnd, Vwnd), and adopting the CBA model for inverting subsurface seawater temperature and salinity proposed in the present invention, an inversion experiment of subsurface seawater temperature and salinity was carried out. The experiment was divided into three parts: data set division, model training and model testing.

[0083] In the data set division stage, the sea surface satellite remote sensing data and three-dimensional temperature and salinity data are processed, and the ocean subsurface temperature and salinity field inversion model accepts time and space data at the same time, so that it meets the input requirements of the first branch and the second branch respectively. For the first branch of the ocean subsurface temperature and salinity field inversion model, a 11×11 local area centered on the true value is selected as input. The number of invalid value points in the selected local area should be less than half of the total number and the center point of the area is a valid value. For the second branch of the ocean subsurface temperature and salinity field inversion model, the BiLSTM time step is set to 10, that is, the sea surface satellite remote sensing data of the first 10 months are used to invert the subsurface seawater temperature field and salinity field of the 10th month. The present invention uses the data from January 2011 to December 2020 as the training data set, and the data from January to December 2021 as the test data set to evaluate the accuracy and generalization ability of the trained ocean subsurface temperature and salinity field inversion model.

[0084] The present invention introduces a new variable - Difference of SST / SST, DSST / DSSS, which does not require additional data and can be obtained using the current existing data. It is used together with the sea surface satellite remote sensing data as the input of the ocean subsurface temperature and salinity field inversion model to participate in the inversion of the subsurface seawater temperature field and salinity field to improve the accuracy of the inversion results. Taking temperature as an example, first calculate the multi-year average values ​​of the sea surface temperature and the seawater temperature at different depths from 2011 to 2020 by month, and then subtract the difference between the two from the specific sea surface temperature to obtain the sea surface temperature difference. This variable removes the systematic differences between the sea surface temperature and the seawater temperature at different depths, and to a certain extent reflects the vertical relationship between the sea surface temperature and the subsurface temperature. The sea surface salinity difference is calculated using a similar method. The calculation formulas for DSST and DSSD are as follows. Among them and As a fixed value, it only needs to be calculated once. In subsequent experiments, the present invention only needs to use the specific sea surface temperature / sea surface salinity minus / You can get DSST / DSSS, where Y represents the year, M represents the month, and N represents the total number of years.

[0085]

[0086]

[0087]

[0088]

[0089] In the training phase of the ocean subsurface temperature and salinity inversion model, the sea surface satellite remote sensing data and the matching three-dimensional temperature and salinity data in the training data set are used to train the ocean subsurface temperature and salinity inversion model. The performance of the ocean subsurface temperature and salinity inversion model is evaluated by calculating the mean square error (MSE) between the inverted values ​​and the true values ​​of seawater temperature and salinity. The Adam optimization algorithm is used to iteratively optimize the parameters of the ocean subsurface temperature and salinity inversion model.

[0090] During the model testing phase, the sea surface satellite remote sensing data of the test data set were input into the trained ocean subsurface temperature and salinity inversion model. The RMSE and R² were calculated by comparing the model inversion values ​​of subsurface seawater temperature and salinity with the true values ​​to evaluate the performance of the model.

[0091] Among them, Figure 5The figure shows the flow chart for calculating the sea surface temperature difference / sea surface salinity. First, for the Argo ocean subsurface temperature / salinity and sea surface temperature / salinity data, the multi-year averages of sea surface temperature / salinity and seawater temperature / salinity at different depths from 2011 to 2020 are calculated by month, and the specific sea surface temperature / sea surface salinity is subtracted from the difference between the two to obtain the sea surface temperature difference / sea surface salinity difference.

[0092] like Figure 6 As shown in the figure, the flow chart for constructing the data set is as follows; for the sea surface satellite remote sensing data, the monthly average SST and SSH are calculated based on the daily data, and the SST data is interpolated to a 0.25°×0.25° grid. 30% of the sea surface satellite remote sensing data is used as the input data test set, and 70% of the sea surface satellite remote sensing data is used as the input data training set. For the Argo ocean subsurface temperature / salinity data, 30% of the ocean subsurface temperature / salinity test set and 70% of the ocean subsurface temperature / salinity training set are used.

[0093] like Figure 7 The figure shows a flow chart of using the CBA model to invert the subsurface ST / SS field using sea surface satellite remote sensing data. First, the input data training set of ocean subsurface temperature / salinity and sea surface satellite remote sensing data is input into the CBA model for model training to obtain a trained model; secondly, the input data training set of sea surface satellite remote sensing data is input into the trained model to obtain the ST / SS inversion results; finally, the ST / SS inversion results are compared with the ocean subsurface temperature / salinity test set to obtain the inversion accuracy.

[0094] Embodiment 3, the present invention also provides a system for inverting a three-dimensional temperature-salinity field of a subsurface ocean, comprising:

[0095] The module for building the ocean subsurface temperature and salinity inversion model is based on the convolutional neural network (CNN) to extract the spatial feature correlation between the sea surface satellite remote sensing data and the ocean subsurface three-dimensional temperature and salinity field. The module is based on the bidirectional long short-term memory (BiLSTM) network to extract the temporal feature correlation between the sea surface satellite remote sensing data and the ocean subsurface three-dimensional temperature and salinity field. The convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network are integrated, and the attention mechanism is added to construct an ocean subsurface temperature and salinity inversion model (CBA) integrating the convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network.

[0096] The ocean subsurface temperature and salinity inversion module is used to apply the sea surface temperature difference DSST / sea surface salinity difference DSSS to the ocean subsurface temperature and salinity inversion model, conduct subsurface seawater temperature and salinity inversion experiments, including data set division, model training and model testing, and complete the inversion of the ocean subsurface temperature and salinity field using sea surface satellite remote sensing data;

[0097] Performance verification module, used to calculate RMSE and R by comparing the model inversion values ​​of subsurface sea temperature and salinity with the true values 2 Evaluating the performance of the ocean subsurface temperature-salinity field inversion model.

[0098] In order to verify the effectiveness of the model proposed in the present invention and the new variables added, the present invention sets up several sets of experiments for comparative analysis. The experimental settings are shown in Table 2. Among them, the CNN-CBAM model and the BiLSTM-Attention model correspond to the first branch and the second branch of the CBA model respectively, and their hyperparameters are exactly the same. The root mean square error (RMSE) and the coefficient of determination (R²) are used to evaluate the model inversion results.

[0099] Table 2 Experimental design

[0100]

[0101] The effectiveness of the model proposed in this invention is analyzed and evaluated by comparing the subsurface seawater temperature and salinity inversion results of different experiments. The input data of the inversion model include SSH, SSS, SST, SSWU and SSWV. In addition to the above input data, the sea surface temperature difference / sea surface salinity difference (DSST / DSSS) is added to the inversion model input data. The contribution of DSST / DSSS to the inversion of subsurface seawater temperature and salinity is verified by comparing experiments with different input data.

[0102] Evaluation of different experimental inversion results of seawater temperature and seawater salinity in 15 depth layers from 30 to 1000 m in the North Pacific in 2021 can be found in Figure 8 The distribution diagram of the coefficient of determination (R²) in the accuracy evaluation of the seawater temperature inversion results at different depths in the North Pacific Ocean for the Case 1-Case 4 experimental cases; Fig. 9 This is the root mean square error (RMSE) distribution diagram of the accuracy evaluation of the seawater temperature inversion results at different depths in the North Pacific Ocean for the Case 1-Case 4 experimental cases;

[0103] Fig.10 This is the distribution diagram of the coefficient of determination (R²) in the accuracy evaluation of the inversion results of seawater salinity at different depths in the North Pacific Ocean for the Case 5-Case 8 experiments; Fig.11 This is the root mean square error (RMSE) distribution diagram of the accuracy evaluation of the seawater temperature inversion results at different depths in the North Pacific Ocean for the Case 5-Case 8 experimental cases;

[0104] Figure 8-Figure 11 The RMSE, R², and relative error in are averaged over 12 months. Figure 8-Figure 11It can be seen that the comparison of the results of different inversion models shows that the model proposed in this invention is better than the other models in inversion results for seawater temperature (Case3) and seawater salinity (Case7) at all depths, which indicates that the subsurface seawater temperature and salinity inversion model that considers the temporal and spatial variation characteristics of sea surface satellite remote sensing data is better than the model that only considers temporal or spatial characteristics. The comparison of the inversion results before and after adding sea surface temperature difference / sea surface salinity difference (DSST / DSSS) to the inversion model input data shows that the inversion results of subsurface seawater temperature and salinity (Case4 / Case8) with new parameters added to all depths are better than the inversion results without adding (Case3 / Case7), and the R² of all depths is greater than 0.973 / 0.976, with an average R² of 0.985 / 0.985. The difference in R² between Case4 / Case8 and Case3 / Case7 at a water depth of less than 100m becomes obvious, indicating that adding DSST / DSSS variables can significantly improve the performance of the model.

[0105] In addition, the present invention carried out a comparative analysis of the subsurface seawater temperature and salinity inversion results of the CBA model, the CNN model, and the BiLSTM model. The RMSE and R² of the inversion results of different models vary with depth. Fig.12 The R² of the seawater temperature inversion results of different models varies with seawater depth; Fig.13 The R² of the seawater salinity inversion results of different models varies with seawater depth;

[0106] Fig.14 The RMSE of the seawater temperature inversion results of different models varies with the seawater depth;

[0107] Fig.15 The RMSE of the seawater salinity inversion results of different models varies with the seawater depth;

[0108] from Figure 12-Figure 15 It can be seen that the overall R² of the inversion results of the CBA model is significantly higher than that of other models, and its RMSE is lower than that of other models. This shows that the overall accuracy of the subsurface seawater temperature and salinity inversion model proposed in the present invention is better than that of other models, and the CBA model can effectively consider the temporal and spatial correlation of sea surface satellite remote sensing data, which once again proves that the model proposed in the present invention has good subsurface seawater temperature and salinity inversion capabilities.

[0109] In short, in oceanographic and climatological research, it is crucial to accurately obtain information on the distribution of seawater temperature and salinity. However, current ocean field observations are difficult and costly due to the complex and changeable marine environment and the vast sea area, which leads to the relative scarcity of ocean subsurface data. In comparison, the currently available sea surface satellite remote sensing data is more abundant. Satellite remote sensing technology provides long-term continuous, wide-coverage, and high-resolution observation data, and the amount of such remote sensing data is increasing rapidly. Therefore, using deep learning methods, the present invention can learn the relationship between sea surface data and deep surface data from these sea surface data, and then invert subsurface seawater temperature and salinity. This provides an effective way to understand the dynamic characteristics of ocean temperature and salinity more deeply and comprehensively, thereby making up for the limitations of scarce subsurface data.

[0110] The present invention proposes a CBA model for inverting subsurface seawater temperature and salinity based on sea surface satellite remote sensing data (sea surface temperature, sea surface height, sea surface salinity and sea surface wind speed), and adds an attention mechanism suitable for CNN and BiLSTM network structures to the model, which enhances the model's ability to capture spatial and temporal information and represent complex ocean processes. In order to make full use of the vertical distribution characteristic information of seawater temperature and salinity, the present invention proposes sea surface temperature difference / sea surface salinity difference parameters for subsurface seawater temperature and salinity inversion.

[0111] To verify the performance of the model, the present invention selected the North Pacific as the experimental sea area, and based on the sea surface satellite remote sensing data and three-dimensional seawater temperature and salinity data from January 2011 to December 2020, different experiments were set up to carry out the subsurface seawater temperature and salinity inversion research from January to December 2021. The comparison results with the measured values ​​show that the CBA model has the best performance, with the highest R² and the lowest RMSE at all depths, that is, considering the temporal and spatial characteristics of sea surface variables at the same time is better than considering the temporal or spatial characteristics alone.

[0112] In order to further analyze the inversion performance of the subsurface seawater temperature and salinity of the proposed model, the inversion results of the proposed model for seawater temperature and salinity in the depth range of 30-1000m in 2021 were analyzed, and the performance of the CBA model was evaluated. The results show that the model proposed in the present invention has good stability and time generalization ability. The subsurface temperature field and salinity field inverted by the model were compared with the observed values. The results show that the proposed model can invert the distribution characteristics and most details of the seawater temperature and salinity fields, which are very similar to the observed values.

[0113] In summary, the CBA model proposed in the present invention is very suitable for the inversion of subsurface seawater temperature and salinity. The model takes into account the spatiotemporal relationship of sea surface data and can accurately invert seawater temperature and salinity at different depths and times. In the absence of field observation data but with satellite remote sensing observation data, the inversion data of the model proposed in the present invention is particularly valuable, which is of great significance for filling the gap in subsurface ocean observation data and can provide valuable reference for research in the fields of marine science and climate change.

[0114] The present invention provides a new method for acquiring ocean subsurface three-dimensional temperature and salinity field data based on sea surface satellite remote sensing data. The data generated by this method can be applied to marine scientific research, business monitoring and forecasting, etc. The present invention provides an ocean subsurface three-dimensional temperature and salinity field inversion method that comprehensively utilizes the temporal and spatial correlation between sea surface satellite remote sensing data and the ocean subsurface three-dimensional temperature and salinity field, which is superior to the existing methods that only consider time or space correlation. The present invention provides a new way to acquire ocean subsurface three-dimensional temperature and salinity field data. The existing ocean subsurface three-dimensional temperature and salinity field inversion methods based on sea surface satellite remote sensing data only consider the time or space correlation between the ocean surface and the subsurface, and do not consider the time and space correlation at the same time. The present invention provides an ocean subsurface three-dimensional temperature and salinity field inversion method that fully integrates the spatial and temporal correlation between sea surface satellite remote sensing data and the ocean subsurface three-dimensional temperature and salinity field.

[0115] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for inverting the three-dimensional temperature and salinity field of the ocean subsurface, characterized in that: This method takes into account the correlation between the physical movement of ocean surface and subsurface seawater, and fully integrates the spatial and temporal correlation of sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface. It specifically includes the following steps: S1, based on the convolutional neural network CNN, the spatial feature correlation between the sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface is extracted. Based on the bidirectional long short-term memory network BiLSTM, the temporal feature correlation between the sea surface satellite remote sensing data and the three-dimensional temperature and salinity field of the ocean subsurface is extracted. The convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM are integrated, and the attention mechanism is added to construct an ocean subsurface temperature and salinity inversion model integrating the convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM. S2, apply the sea surface temperature difference DSST / sea surface salinity difference DSSS to the ocean subsurface temperature and salinity field inversion model, conduct subsurface seawater temperature and salinity inversion experiments, including data set division, model training and model testing, and complete the inversion of the ocean subsurface temperature and salinity field using sea surface satellite remote sensing data; S3, by comparing the model inversion values ​​of subsurface seawater temperature and salinity with the true values, RMSE and R 2 Evaluate the performance of the ocean subsurface temperature and salinity inversion model; In step S1, the ocean subsurface temperature and salinity field inversion model includes two branches, the first branch is used to extract the spatial dimension characteristics of the sea surface remote sensing data, and the second branch is used to extract the temporal dimension characteristics of the sea surface remote sensing data; The first branch consists of a convolutional attention module, a convolutional neural network (CNN), and a fully connected neural network. Before the data enters the convolutional layer, it is passed to the CBAM layer to highlight important features and enhance the feature expression capability. The CBAM layer is added after the convolutional layer and the activation layer to filter and enhance the features extracted by the convolutional layer, so that the feature fusion module can use these feature information to improve the overall performance of the model. After each convolutional layer, a batch normalization layer, an activation layer, and a random dropout layer are introduced. The activation layer uses ReLU as the activation function. For a given feature map, the convolutional attention module calculates the attention map in two independent dimensions, channel and space, and multiplies the attention map with the input feature map to perform adaptive feature refinement, selecting a 11×11 local area centered on the true value as input; The second branch consists of a one-to-many attention mechanism, a bidirectional long short-term memory network BiLSTM, and a fully connected neural network; a random dropout layer is added after the first BiLSTM layer to alleviate overfitting; a one-to-many attention layer is added after the second BiLSTM layer; the specific steps are: The attention score is calculated by performing a dot product operation based on the hidden state of the last time step and the hidden state of all time steps, and normalizing them through softmax. The attention score is obtained by performing a dot product operation based on the hidden state of the last time step and the hidden state of all time steps, and normalizing them through softmax. The hidden state of all time steps is weighted summed by the attention score to obtain the context vector of the relevant hidden state, and then the context vector is residually connected with the hidden state of the last time step. The context vector is added to the hidden state of the last time step as a residual connection to obtain the final output. In step S1, an ocean subsurface temperature and salinity inversion model integrating a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) network is constructed, including: The spatial features are extracted through the convolution layer and CBAM, and the temporal features are extracted through BiLSTM and the attention mechanism, and then flattened and converted to channel dimensions through the fully connected layer; the input data processed by these two branches are spliced ​​into a new comprehensive variable in the channel dimension to achieve the fusion of temporal and spatial information; this new comprehensive variable is passed to the last two fully connected layers, the first fully connected layer is used to expand the channel dimension of the new variable, and the second fully connected layer is used to obtain the final output; information fusion is achieved, avoiding information loss or overfitting that may be caused by direct mapping; the inversion model of the ocean subsurface temperature and salinity field is: Output=W2(W1(Concat(F S ,F T ))+b1)+b2 In the formula, F S , F T Represent the spatial dimension features and temporal dimension features output by the first branch and the second branch respectively; Concat() represents the concatenation operation, W1, W2 and b1, b2 represent the weights and biases of the last two fully connected layers, and Output represents the final output; The performance of the ocean subsurface temperature and salinity inversion model is optimized. The ocean subsurface temperature and salinity inversion model adopts an early stopping algorithm and selects the ocean subsurface temperature and salinity inversion model with the best performance on the validation data set. At the same time, the Adam optimizer is used to adjust the learning process of the ocean subsurface temperature and salinity inversion model.

2. The method for inverting the three-dimensional temperature-salinity field of the ocean subsurface according to claim 1, characterized in that: Before step S1, you need to do: Collect sea surface satellite remote sensing data, including sea surface temperature SST, sea surface salinity SSS, sea surface height SSH, sea surface wind SSW and three-dimensional temperature and salinity 3DTS data. The sea surface wind SSW is divided into the east component USSW and the north component VSSW. Use sea surface satellite remote sensing data to invert the ocean subsurface temperature field and salinity field respectively; The difference between the sea surface temperature / sea surface salinity and the multi-year monthly average of seawater temperature and salinity at different depths in the corresponding month, namely the sea surface temperature difference DSST / sea surface salinity difference DSSS, is introduced and combined with the above-mentioned sea surface satellite remote sensing data for the inversion of subsurface seawater temperature and salinity field; The ocean subsurface temperature and salinity field uses three-dimensional temperature and salinity 3DTS. The 3DTS data includes the seawater temperature and salinity of multiple different depth layers, which is used for the construction and evaluation of the ocean subsurface temperature and salinity field inversion model; All data used are monthly average data. The monthly average values ​​of SST and SSH are calculated from daily data respectively, and the SST data are interpolated to a grid of 0.25°×0.25°. To ensure spatial consistency, if a node has an empty variable, the node will be deleted. Before all data are input into the subsurface seawater temperature and salinity inversion model, all data are standardized according to the channel dimension so that the mean of the data set is 0 and the variance is 1, eliminating the quantitative differences between different data.

3. The method for inverting the three-dimensional temperature-salinity field of the ocean subsurface according to claim 1, characterized in that: In step S2, the sea surface temperature difference DSST / sea surface salinity difference DSSS is applied to the ocean subsurface temperature and salinity field inversion model, including: Taking the three-dimensional temperature-salinity data 3DT as the real observation value of subsurface seawater temperature and salinity, the multi-source sea surface satellite remote sensing data of sea surface height SSH, sea surface salinity SSS, sea surface temperature SST, sea surface wind speed east component Uwnd, sea surface wind speed north component Vwnd are applied to the ocean subsurface temperature-salinity field inversion model to conduct subsurface seawater temperature and salinity inversion experiments; In the data set division stage, the sea surface satellite remote sensing data and three-dimensional temperature and salinity data are processed, and the ocean subsurface temperature and salinity field inversion model accepts time and space data at the same time, so that it meets the input requirements of the first branch and the second branch respectively; Introduce new variables, sea surface temperature difference DSST / sea surface salinity difference DSSS, and use them together with sea surface satellite remote sensing data as input data for the ocean subsurface temperature and salinity field inversion model, and participate in the inversion of subsurface seawater temperature and salinity fields; In the training phase of the ocean subsurface temperature and salinity inversion model, the sea surface satellite remote sensing data and the matching three-dimensional temperature and salinity data in the training data set are used to train the ocean subsurface temperature and salinity inversion model. The performance of the ocean subsurface temperature and salinity inversion model is evaluated by calculating the mean square error (MSE) between the inverted values ​​and the true values ​​of seawater temperature and salinity. The Adam optimization algorithm is used to iteratively optimize the parameters of the ocean subsurface temperature and salinity inversion model. During the model testing phase, the sea surface satellite remote sensing data of the test data set are input into the trained ocean subsurface temperature and salinity field inversion model.

4. The method for inverting the three-dimensional temperature-salinity field of the ocean subsurface according to claim 3 is characterized in that: In the data set division stage, for the first branch of the ocean subsurface temperature and salinity field inversion model, a 11×11 local area centered on the true value is selected as input. The number of invalid value points in the selected local area should be less than half of the total number and the center point of the area is a valid value. For the second branch of the ocean subsurface temperature and salinity field inversion model, the BiLSTM time step is set to 10, that is, the sea surface satellite remote sensing data of the first 10 months are used to invert the subsurface seawater temperature field and salinity field of the 10th month. The calculation formula of the newly introduced variable sea surface temperature difference DSST / sea surface salinity difference DSSS is as follows: DSST Y,M =SST Y,M -dif_T M DSSS Y,M =SSS Y,M -dif_S M Where, DSST Y,M is the sea surface temperature difference, DSSS Y，M is the sea surface salinity difference, OST Y,M is the ocean subsurface temperature, dif_T M is the temperature difference, OSS Y,M is the ocean subsurface salinity, dif_S M is the salinity difference, M is the month, N is the total number of years, Y is the year, SST Y,M is the sea surface temperature, SSS Y，M is the sea surface salinity; dif_T M and dif_S M As a fixed value, it only needs to be calculated once, and in subsequent experiments, only the specific sea surface temperature / sea surface salinity minus dif_T is used. M / dif_S M Get DSST / DSSS.

5. A three-dimensional ocean subsurface temperature and salinity field inversion system, characterized in that: The system implements the method for inverting the three-dimensional temperature-salinity field of the ocean subsurface as claimed in any one of claims 1 to 4, and the system comprises: The module for building the ocean subsurface temperature and salinity inversion model is based on the convolutional neural network (CNN) to extract the spatial feature correlation between the sea surface satellite remote sensing data and the ocean subsurface three-dimensional temperature and salinity field. The module is based on the bidirectional long short-term memory (BiLSTM) network to extract the temporal feature correlation between the sea surface satellite remote sensing data and the ocean subsurface three-dimensional temperature and salinity field. The convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network are integrated, and the attention mechanism is added to build an ocean subsurface temperature and salinity inversion model that integrates the convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network. The ocean subsurface temperature and salinity inversion module is used to apply the sea surface temperature difference DSST / sea surface salinity difference DSSS to the ocean subsurface temperature and salinity inversion model, conduct subsurface seawater temperature and salinity inversion experiments, including data set division, model training and model testing, and complete the inversion of the ocean subsurface temperature and salinity field using sea surface satellite remote sensing data; Performance verification module, used to calculate RMSE and R by comparing the model inversion values ​​of subsurface sea temperature and salinity with the true values 2 Evaluating the performance of the ocean subsurface temperature-salinity field inversion model.

6. The ocean subsurface three-dimensional temperature-salinity field inversion system according to claim 5, 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 above-mentioned ocean subsurface three-dimensional temperature-salinity field inversion system can be realized.

Images