A method, system, electronic device and medium for reconstructing three-dimensional temperature and salinity fields in the ocean

Abstract

The application discloses a marine three-dimensional temperature and salinity field reconstruction method, system, electronic equipment and medium, comprising: constructing a target field and constructing an observation field; adding a plurality of levels of Gaussian noise to each sample data in the standardized target field in turn to construct a noise field sample data set; constructing a marine three-dimensional temperature and salinity field reconstruction model comprising a convolution local dependence modeling module, a patch division module, a multi-head attention global dependence modeling module, a patch restoration module and a convolution reconstruction module; training the marine three-dimensional temperature and salinity field reconstruction model using the noise field sample data set to obtain a trained marine three-dimensional temperature and salinity field reconstruction model; using the observation field to guide the trained marine three-dimensional temperature and salinity field reconstruction model to denoise a pure Gaussian noise field to obtain a marine three-dimensional temperature and salinity field reconstruction result, and performing inverse standardization on the marine three-dimensional temperature and salinity field reconstruction result to obtain a reconstructed marine three-dimensional temperature and salinity field. The application can improve the accuracy of marine three-dimensional temperature and salinity field reconstruction.

Description

Technical Field

[0001] This application relates to the field of marine data processing technology, and in particular to a method, system, electronic device and medium for reconstructing a three-dimensional marine temperature and salinity field. Background Technology

[0002] Ocean temperature and salinity are fundamental climate variables for understanding and recording changes in the marine environment. Observing their three-dimensional field is crucial for monitoring and studying the laws of ocean physics. At the same time, it has valuable significance for improving my country's marine resource management, response to extreme events, ecological sustainable development, and marine environmental protection in military activities.

[0003] While numerical models combined with assimilation modules can also provide high-precision three-dimensional ocean fields, they rely on the systematic assimilation of ocean conditions over the past few decades, primarily focusing on the study of long-term historical ocean conditions. Therefore, the initial deployment and operational use of data assimilation and numerical models depend heavily on computational resources, resulting in low update timeliness, typically lagging by several months to several years. In contrast, ocean data reconstruction focuses on serving oceanographic support. Utilizing near-real-time satellite observations, it rapidly reconstructs the three-dimensional ocean field, offering operational advantages such as low operating costs and high update timeliness. This allows it to meet the needs of many time-sensitive applications, such as real-time underwater acoustic field calculations, three-dimensional vortex and "sea cliff" identification, thereby supporting underwater vehicle path planning, underwater communication, and mission-aided decision-making.

[0004] Existing deep learning-based reconstruction techniques have limitations. Firstly, reconstructing high spatiotemporal resolution three-dimensional ocean features relies on high spatiotemporal resolution subsurface observation data as a reference for underwater projection. However, due to the relatively coarse spatiotemporal resolution of the analytical field from field measurements, most current three-dimensional reconstruction studies in the open ocean can only achieve a monthly average reconstruction of 1°. Although high spatiotemporal resolution (e.g., daily, 1 / 12°) ocean reanalysis products can be used as a reference, the simulation of eddy positions and structures by these products may not be entirely accurate, and their precision at certain depths and regions may not be entirely reliable compared to field observations. Secondly, statistical models lack physical constraints and cannot learn the inherent spatiotemporal evolution laws and physical constraints within the reanalysis products. Therefore, the accuracy of existing reconstruction techniques for ocean three-dimensional temperature and salinity field reconstruction is relatively low. Summary of the Invention

[0005] This application aims to propose a method, system, electronic device, and medium for reconstructing a three-dimensional marine temperature and salinity field, which can improve the accuracy of marine three-dimensional temperature and salinity field reconstruction.

[0006] In a first aspect, embodiments of this application provide a method for reconstructing a three-dimensional marine temperature-salinity field, the method comprising:

[0007] Acquire gridded satellite sea surface temperature and salinity product data, field-measured ocean temperature and salinity profile data, and three-dimensional temperature and salinity fields from ocean reanalysis products;

[0008] The three-dimensional temperature field and three-dimensional salinity field in the aforementioned marine reanalysis product are merged to construct the target field;

[0009] The satellite sea surface temperature and salinity product data and the field-measured ocean temperature and salinity profile data are preprocessed to obtain preprocessed satellite sea surface temperature and salinity product data and preprocessed field-measured ocean temperature and salinity profile data.

[0010] The preprocessed satellite sea surface temperature and salinity product data are combined with the preprocessed field-measured ocean temperature and salinity profile data to construct an observation field;

[0011] The sample data in the target field are standardized to obtain a standardized target field. Multiple levels of Gaussian noise are added to each sample data in the standardized target field to construct a noise field sample dataset.

[0012] A three-dimensional marine temperature and salinity field reconstruction model is constructed, which includes a convolutional local dependency modeling module, a patch partitioning module, a multi-head attention global dependency modeling module, a patch restoration module, and a convolutional reconstruction module. The convolutional local dependency modeling module includes two convolutional layers. The first convolutional layer includes two convolutions and an activation layer, and the second convolutional layer includes one convolution and an activation layer. The multi-head attention global dependency modeling module includes multiple residual-connected multi-head attention global dependency sub-modules. Each multi-head attention global dependency sub-module includes multiple layer normalization layers and multiple fully connected layers.

[0013] The ocean three-dimensional temperature and salinity field reconstruction model is trained using the noise field sample dataset to obtain a trained ocean three-dimensional temperature and salinity field reconstruction model.

[0014] The observation field is used to guide the trained three-dimensional ocean temperature and salinity field reconstruction model to denoise the pure Gaussian noise field, thereby obtaining the three-dimensional ocean temperature and salinity field reconstruction result. The three-dimensional ocean temperature and salinity field reconstruction result is then de-normalized to obtain the reconstructed three-dimensional ocean temperature and salinity field.

[0015] In some embodiments, the preprocessing of the satellite sea surface temperature and salinity product data and the field-measured ocean temperature and salinity profile data to obtain preprocessed satellite sea surface temperature and salinity product data and preprocessed field-measured ocean temperature and salinity profile data includes:

[0016] Bilinear interpolation is used to interpolate the satellite sea surface temperature and salinity product data into a grid with the same horizontal resolution as the target field, resulting in an interpolated satellite temperature grid field and an interpolated satellite salinity grid field. The interpolated satellite temperature grid field and the interpolated satellite salinity grid field are used as the preprocessed satellite sea surface temperature and salinity product data.

[0017] Two three-dimensional grid fields with the same size as the target field are set up. The measured ocean temperature and salinity profile data are interpolated into the two three-dimensional grid fields by three-dimensional nearest neighbor interpolation, respectively, to obtain the measured temperature three-dimensional grid field and the measured salinity three-dimensional grid field. The measured temperature three-dimensional grid field and the measured salinity three-dimensional grid field are used as the preprocessed measured ocean temperature and salinity profile data.

[0018] In some embodiments, merging the preprocessed satellite sea surface temperature and salinity product data with the preprocessed field-measured ocean temperature and salinity profile data to construct an observation field includes:

[0019] The interpolated satellite temperature grid field is added to the first layer data of the temperature measurement three-dimensional grid field on the same day. If several grids in the first layer of the temperature measurement three-dimensional grid field have both satellite temperature observation values ​​and on-site temperature measurement values, the average value of the satellite temperature observation values ​​and the on-site temperature measurement values ​​is taken as the temperature value of several grids in the first layer of the temperature measurement three-dimensional grid field, thus obtaining the merged temperature observation data.

[0020] The interpolated satellite salinity grid field is added to the first layer data of the salinity measurement three-dimensional grid field on the same day. If several grids in the first layer of the salinity measurement three-dimensional grid field have both satellite salinity observation values ​​and on-site salinity measurement values, the average value of the satellite salinity observation values ​​and the on-site salinity measurement values ​​is taken as the salinity value of several grids in the first layer of the salinity measurement three-dimensional grid field, thus obtaining the salinity merged observation data.

[0021] The combined temperature and salinity observation data are merged along the channel dimension to obtain the observation field.

[0022] In some embodiments, the noise field sample dataset includes temperature noise field data and salinity noise field data. The step of using the noise field sample dataset to train the ocean three-dimensional temperature-salinity field reconstruction model to obtain a trained ocean three-dimensional temperature-salinity field reconstruction model includes:

[0023] The temperature noise field data and the salinity noise field data are input into the three-dimensional ocean temperature and salinity field reconstruction model to obtain the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field.

[0024] A loss function is constructed based on the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field.

[0025] The loss function is used to train the three-dimensional marine temperature and salinity field reconstruction model until the loss function converges or reaches a preset number of iterations, thus obtaining the trained three-dimensional marine temperature and salinity field reconstruction model.

[0026] In some embodiments, inputting the temperature noise field data and the salinity noise field data into the three-dimensional ocean temperature-salinity field reconstruction model to obtain the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field includes:

[0027] The temperature noise field data and the salinity noise field data are input into the convolutional local dependency modeling module to obtain the target local dependency representation.

[0028] The target local dependency representation is input into the patch partitioning module to obtain the flattened patch field;

[0029] The flattened patch field is input into the multi-head attention global dependency modeling module to obtain the target multi-head attention global dependency field.

[0030] The target multi-head attention global dependency field is input into the patch restoration module to obtain the restored feature field;

[0031] The restored feature field is input into the convolutional reconstruction module to obtain the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field.

[0032] In some implementations, the step of inputting the temperature noise field data and the salinity noise field data into the convolutional local dependency modeling module to obtain the target local dependency representation includes:

[0033] The temperature noise field data and the salinity noise field data are input into the first convolutional layer. The temperature noise field data and the salinity noise field data are convolved by the two convolutions respectively to obtain the convolved temperature noise field and the convolved salinity noise field.

[0034] The convolutional temperature noise field and the convolutional salinity noise field are input into the activation layer to obtain a preliminary local dependency characterization.

[0035] The preliminary local dependency representation is input into the second convolutional layer, and the preliminary local dependency representation is processed by convolution and activation layers to obtain the target local dependency representation.

[0036] In some implementations, inputting the flattened patch field into the multi-head attention global dependency modeling module to obtain the target multi-head attention global dependency field includes:

[0037] The flattened patch field is input into the first multi-head attention global dependency submodule. The flattened patch field is processed by multiple first-layer normalization layers and multiple first-layer fully connected layers to obtain the first multi-head attention global dependency field.

[0038] The flattened patch field and the first multi-head attention global dependency field are residually connected to obtain the first residual result;

[0039] The first residual result is input into the second multi-head attention global dependency submodule, and the first residual result is processed through multiple second-layer normalization layers and multiple second-layer fully connected layers to obtain the second multi-head attention global dependency field.

[0040] The first multi-head attention global dependency field and the second multi-head attention global dependency field are residually connected to obtain the second residual result;

[0041] The second residual result is input into the third multi-head attention global dependency submodule for processing. The process continues until all multi-head attention global dependency submodules have finished processing, at which point the target multi-head attention global dependency field is obtained.

[0042] Secondly, embodiments of this application also provide a three-dimensional marine temperature and salinity field reconstruction system, the system comprising:

[0043] The data acquisition unit is used to acquire gridded satellite sea surface temperature and salinity product data, field-measured ocean temperature and salinity profile data, and three-dimensional temperature field and three-dimensional salinity field in ocean reanalysis products;

[0044] The first merging unit is used to merge the three-dimensional temperature field and the three-dimensional salinity field in the marine reanalysis product to construct the target field;

[0045] The data preprocessing unit is used to preprocess the satellite sea surface temperature and salinity product data and the field-measured ocean temperature and salinity profile data to obtain preprocessed satellite sea surface temperature and salinity product data and preprocessed field-measured ocean temperature and salinity profile data.

[0046] The second merging unit is used to merge the preprocessed satellite sea surface temperature and salinity product data with the preprocessed field-measured ocean temperature and salinity profile data to construct an observation field.

[0047] The noise addition unit is used to standardize the sample data in the target field to obtain a standardized target field, and to add multiple levels of Gaussian noise to each sample data in the standardized target field in sequence to construct a noise field sample dataset.

[0048] The model building unit is used to construct a three-dimensional marine temperature and salinity field reconstruction model, which includes a convolutional local dependency modeling module, a patch partitioning module, a multi-head attention global dependency modeling module, a patch restoration module, and a convolutional reconstruction module. The convolutional local dependency modeling module includes two convolutional layers. The first convolutional layer includes two convolutions and an activation layer. The second convolutional layer includes one convolution and an activation layer. The multi-head attention global dependency modeling module includes multiple residual-connected multi-head attention global dependency sub-modules. Each multi-head attention global dependency sub-module includes multiple layer normalization layers and multiple fully connected layers.

[0049] The model training unit is used to train the ocean three-dimensional temperature and salinity field reconstruction model using the noise field sample dataset to obtain the trained ocean three-dimensional temperature and salinity field reconstruction model.

[0050] The temperature and salinity field reconstruction unit is used to guide the trained ocean three-dimensional temperature and salinity field reconstruction model with the observation field to denoise the pure Gaussian noise field, obtain the ocean three-dimensional temperature and salinity field reconstruction result, and perform denormalization on the ocean three-dimensional temperature and salinity field reconstruction result to obtain the reconstructed ocean three-dimensional temperature and salinity field.

[0051] Thirdly, embodiments of this application also provide an electronic device, including at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, the instructions being executed by the at least one control processor to enable the at least one control processor to perform a marine three-dimensional temperature and salinity field reconstruction method as described above.

[0052] Fourthly, embodiments of this application also provide a computer-readable storage medium storing computer-executable instructions for causing a computer to execute a marine three-dimensional temperature and salinity field reconstruction method as described above.

[0053] Compared with the prior art, this application has the following beneficial effects:

[0054] (1) By merging the preprocessed satellite sea surface temperature and salinity product data with the preprocessed field measured ocean temperature and salinity profile data, it is possible to effectively reveal the small and medium scale ocean structure and make up for the problems of insufficient coverage and poor spatiotemporal continuity of traditional field observations.

[0055] (2) In the three-dimensional temperature and salinity field reconstruction model of the ocean, the convolutional local dependency modeling module can capture the local features of the input variables, improve the sensitivity of the three-dimensional temperature and salinity field reconstruction model to local information, and perform preliminary fusion of temperature and salinity feature information; the patch partitioning module can flatten the features to meet the input size requirements of the subsequent multi-head attention global dependency modeling module; the multi-head attention global dependency modeling module can better capture the global dependency relationship between features; the patch restoration module restores the flattened features to their original shape to adapt to the convolution operation; the convolutional reconstruction module maps the output results from the feature space back to the dimension of the temperature and salinity fractional field, and further strengthens the local features, thereby improving the accuracy of the three-dimensional temperature and salinity field reconstruction of the ocean.

[0056] (3) By merging the three-dimensional temperature field and the three-dimensional salinity field in the ocean reanalysis product to construct the target field, the sample data in the target field are standardized and noise is added. Then, the ocean three-dimensional temperature and salinity field reconstruction model is trained using the noise field sample dataset. The ocean three-dimensional temperature and salinity field reconstruction model is then guided by the observation field to denoise the pure Gaussian noise field. This makes the reconstructed three-dimensional temperature and salinity field not only meet the observation constraints, but also conform to the inherent spatiotemporal evolution law of the ocean. This overcomes the shortcomings of traditional statistical models that lack physical constraints, and the generated results are closer to the real ocean state, improving the reliability and scientific nature of the reconstruction, thereby further improving the accuracy of the ocean three-dimensional temperature and salinity field reconstruction. Attached Figure Description

[0057] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0058] Figure 1 This is a flowchart illustrating an embodiment of the marine three-dimensional temperature and salinity field reconstruction method provided in this application;

[0059] Figure 2 This is a schematic diagram of the overall SRM model in the best embodiment of the marine three-dimensional temperature and salinity field reconstruction method provided in this application;

[0060] Figure 3 This is a schematic diagram of an embodiment of the marine three-dimensional temperature and salinity field reconstruction system provided in this application;

[0061] Figure 4 This is a schematic diagram of the structure of an embodiment of the electronic device provided in this application. Detailed Implementation

[0062] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0063] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0064] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0065] In the description of this application, it should be noted that, unless otherwise explicitly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0066] To address the issue of low accuracy in reconstructing three-dimensional temperature and salinity fields of the ocean in related technologies, this application proposes a method, system, electronic device, and medium for reconstructing three-dimensional temperature and salinity fields of the ocean.

[0067] Reference Figure 1 This application provides a schematic flowchart of a method for reconstructing a three-dimensional marine temperature and salinity field. This method is applied to an electronic device, which may be a server or a mobile terminal, etc. Figure 1 As shown, the ocean three-dimensional temperature and salinity field reconstruction method may include the following steps S101 to S108.

[0068] Step S101: Obtain gridded satellite sea surface temperature and salinity product data, field-measured ocean temperature and salinity profile data, and three-dimensional temperature field and three-dimensional salinity field from ocean reanalysis products.

[0069] Specifically, the three-dimensional temperature field and three-dimensional salinity field from the ocean reanalysis product are obtained from open-source data download channels, denoted as... and Horizontal coverage area The grid resolution is ,Include Each depth layer has a complete three-dimensional dimension of [number]. The time resolution is one session per day, denoted as A sample is used for unsupervised autoregressive training of an unconditionally constrained SRM model (i.e., a three-dimensional ocean temperature and salinity field reconstruction model). Gridded satellite sea surface temperature and salinity products, as well as field-measured ocean temperature and salinity profile data, are obtained as constraints to guide the denoising process of the trained SRM model and generate the three-dimensional ocean temperature and salinity field for the target date.

[0070] Step S102: Merge the three-dimensional temperature field and the three-dimensional salinity field in the ocean reanalysis product to construct the target field.

[0071] Specifically, the three-dimensional temperature field in the ocean reanalysis product and three-dimensional salinity field The two are merged into the target field in the channel dimension. The size is .

[0072] Step S103: Preprocess the satellite sea surface temperature and salinity product data and the field measured ocean temperature and salinity profile data to obtain preprocessed satellite sea surface temperature and salinity product data and preprocessed field measured ocean temperature and salinity profile data. This may include the following steps S1031 to S1032.

[0073] Step S1031: Use bilinear interpolation to interpolate the satellite sea surface temperature and salinity product data into a grid with the same horizontal resolution as the target field, to obtain the interpolated satellite temperature grid field and the interpolated satellite salinity grid field. Use the interpolated satellite temperature grid field and the interpolated satellite salinity grid field as the preprocessed satellite sea surface temperature and salinity product data.

[0074] Specifically, a bilinear interpolation method is used to interpolate the gridded satellite sea surface temperature and salinity products obtained in step S11 into a grid field with the same horizontal resolution as the target field, unifying the spatial grid resolution to [value missing]. .

[0075] Step S1032: Set up two three-dimensional grid fields with the same size as the target field. Use three-dimensional nearest neighbor interpolation to interpolate the measured ocean temperature and salinity profile data into the two three-dimensional grid fields respectively, to obtain the measured temperature three-dimensional grid field and the measured salinity three-dimensional grid field. Use the measured temperature three-dimensional grid field and the measured salinity three-dimensional grid field as the preprocessed measured ocean temperature and salinity profile data.

[0076] Specifically, the three-dimensional nearest neighbor interpolation method is used, setting two interpolation methods with the same size as the target field. ) three-dimensional mesh field and This is used to unify field-measured temperature and salinity profile data into a regular grid field. Specifically:

[0077] 1. Transform the three-dimensional mesh field and The value of each grid position in the 3D mesh field is set to 0; and The three-dimensional temperature field in the ocean reanalysis product corresponding to a certain grid. and three-dimensional salinity field If the value of the grid is NaN, i.e., land, then the three-dimensional grid field will be... and The grid was synchronously set to NaN; each observation point in the field-measured temperature and salinity profiles was incorporated into the three-dimensional grid field. and The nearest grid position in the middle.

[0078] 2. Extract all temperature and salinity measured profiles for the same date, and iterate through the measured values ​​in each temperature and salinity profile. Place the measured temperature value into a three-dimensional mesh field. The grid with the closest spatial location is used to place the measured salinity value in the three-dimensional grid field. The grid with the closest spatial location in the middle; after traversal, if the 3D grid field... and If a grid has two or more measured points, then the value of that grid is recorded as the average of all the measured values ​​therein.

[0079] Step S104: Merge the preprocessed satellite sea surface temperature and salinity product data with the preprocessed field measured ocean temperature and salinity profile data to construct an observation field.

[0080] In this embodiment, the interpolated satellite temperature grid field is added to the first layer of the measured temperature 3D grid field for the same day. If several grids in the first layer of the measured temperature 3D grid field have both satellite temperature observations and on-site measured temperature values, the average of these two values ​​is used as the temperature value for those grids in the first layer of the measured temperature 3D grid field, resulting in merged temperature observation data. Similarly, the interpolated satellite salinity grid field is added to the first layer of the measured salinity 3D grid field for the same day. If several grids in the first layer of the measured salinity 3D grid field have both satellite salinity observations and on-site measured salinity values, the average of these two values ​​is used as the salinity value for those grids in the first layer of the measured salinity 3D grid field, resulting in merged salinity observation data. The merged temperature and salinity observation data are then merged along the channel dimension to obtain the observation field. Thus, using the observation field as an observation constraint to guide the denoising of the ocean 3D temperature and salinity field reconstruction model can improve the accuracy of the ocean 3D temperature and salinity field reconstruction.

[0081] Specifically, the interpolated satellite temperature grid field is compared with the measured three-dimensional temperature grid field for the same date. The first layer is added together, if If a certain grid in the first layer contains both satellite temperature observations and on-site temperature measurements, then the temperature value of that grid is set as the average of the satellite temperature observations and the on-site temperature measurements. The resulting interpolated satellite salinity grid field is then compared with the salinity measurement three-dimensional grid field for the same date. The first layer is added together, if If a grid in the first layer contains both satellite salinity observations and in-situ measured salinity values, then the salinity value of that grid is set as the average of the satellite and in-situ measured values. At this point, a complete three-dimensional grid field recording both satellite sea surface observations and in-situ measurements is obtained. and As an observation field (i.e., observation constraint), the two are merged in the channel dimension and denoted as the observation field. The size is .

[0082] Step S105: Standardize the sample data in the target field to obtain the standardized target field. Add multiple levels of Gaussian noise to each sample data in the standardized target field to construct a noise field sample dataset.

[0083] Specifically, calculate all of them separately. Target field of each sample The mean values ​​of the temperature and salinity variables, i.e., the temperature mean. and average salinity and their respective variances, namely temperature variance. and salinity variance ; Since the vertical temperature is not averaged, the above mean and variance are both for a dimension of . A one-dimensional vector. The expression is as follows:

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] Then, for the target field The temperature and salinity variables are standardized to obtain the target field of each sample. The standardized temperature variable and standardized salinity variable The size is still Since both the mean and variance are of size ; This is a one-dimensional vector, therefore equivalent to independently standardizing each depth layer. The expression is as follows:

[0089] ;

[0090] .

[0091] Step S106: Construct a three-dimensional marine temperature and salinity field reconstruction model that includes a convolutional local dependency modeling module, a patch partitioning module, a multi-head attention global dependency modeling module, a patch restoration module, and a convolutional reconstruction module.

[0092] Specifically, (1) the convolutional local dependency modeling module contains two convolutional layers; the first layer has two convolutions with a kernel size of 5×5, which respectively target kernels of size 5×5. The input model SRM's random Gaussian noise field Temperature noise field and salinity noise field Perform independent calculations to initially extract the local dependencies between the two types of variables, and map the last dimension to... Subsequently, the two initial variable fields for extracting local dependencies have sizes of [size missing]. Then it goes through a GELU activation layer, and then is merged in the last feature dimension to obtain a size of The first layer represents the initial local dependency; the second layer has a convolutional layer and a GELU activation layer with a 3×3 kernel, which fuses the local dependencies of the two classes of variables to obtain the local dependency representation. , will feature Dimensional compression to The obtained target local dependency representation The size is The expression for the convolutional local dependency modeling module is as follows:

[0093] ;

[0094] (2) Build the patch partitioning module; first, use a size of The window moves up and down in increments of 1. The step size for moving left and right is The scanning size from top to bottom and from left to right is Target local dependency representation Divide it into Each size is The patch; to ensure Given an integer, the requirement is... and Divisible by each and After scanning is complete, all patches are stitched together to a size of [size missing]. patch field Secondly, each patch is flattened into a one-dimensional sequence, and the length of the flattened sequence is... The patch field after being flattened The size is .

[0095] (3) Construct a multi-head attention global dependency modeling module, which includes multiple multi-head attention global dependency sub-modules. Each multi-head attention global dependency sub-module includes multiple layer normalization layers and multiple fully connected layers. Its expression is as follows:

[0096] ;

[0097] Flattened patch field First, it goes through a layer normalization layer. along Normalize the dimension to obtain the normalized patch field. The expression is as follows:

[0098] ;

[0099] Then, Input three fully connected layers respectively Three coding fields were obtained. The expression is as follows:

[0100] ;

[0101] ;

[0102] ;

[0103] Next, we calculate the multi-head self-attention and perform parallel calculations. Self-attention field The calculation process for the self-attention field of one of the heads is as follows: and Perform matrix multiplication, then scale the matrices back to their original size. Times, then along the feature dimension conduct Function activation, and with Perform matrix multiplication to obtain a matrix of size . Self-attention field Each self-attention field The expression is as follows:

[0104] ;

[0105] in, Indicates the first The value matrix of a self-attention field Indicates the first A query matrix with a self-attention field. Indicates the first A bond matrix of a self-attention field.

[0106] Merge in the last feature dimension Self-attention field The resulting size is Multi-head self-attention field Then it goes through a normalization layer and a fully connected layer. The compressed feature dimension is The output size is Multi-head attention global dependency field The expression is as follows:

[0107] ;

[0108] The multi-head attention global dependency submodules are sequentially residual-connected, and each multi-head attention global dependency field is denoted as... All sizes Among them, residual connection refers to the... The input to the multi-head attention global dependency submodule is It should be noted that the input to the first multi-head attention global dependency submodule is a patch field. Without residual connections, the input to the second multi-head attention global dependency submodule is... .

[0109] (4) Build the patch restoration module; this module is the reverse process of the patch partitioning module described above. First, the dimensions of each patch are... Expand, then place each patch back into its original position before partitioning, thus making the size [size missing]. The A multi-head attention global dependency field (i.e., the target multi-head attention global dependency field), restored to a size of Feature field .

[0110] (5) Convolutional reconstruction module is built to reconstruct the feature field after reconstruction. Refactor, refactor into Fractional field of size The first layer of the last dimension corresponds to the fractional field of the three-dimensional temperature field. The second layer of the last dimension corresponds to the fractional field of the three-dimensional salinity field. The convolutional reconstruction module contains one convolution, one GELU activation layer, and one convolution, as shown in the following expression:

[0111] .

[0112] (6) Connect the above modules in sequence to construct a complete fractional-based three-dimensional ocean temperature-salinity field reconstruction model (SRM), the expression of which is as follows:

[0113] ;

[0114] ;

[0115] in, The functional expression (i.e., fractional function) of the SRM model outputs random Gaussian noise. Denoising is directed towards the true data, i.e., the fractional field. .

[0116] Step S107: Use the noise field sample dataset to train the three-dimensional ocean temperature and salinity field reconstruction model to obtain the trained three-dimensional ocean temperature and salinity field reconstruction model, which may include steps S1071 to S1073.

[0117] Step S1071: Input the temperature noise field data and salinity noise field data into the three-dimensional ocean temperature and salinity field reconstruction model to obtain the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field.

[0118] Specifically, temperature noise field data and salinity noise field data are input into the convolutional local dependency modeling module to obtain the target local dependency representation; the target local dependency representation is input into the patch partitioning module to obtain the flattened patch field; the flattened patch field is input into the multi-head attention global dependency modeling module to obtain the target multi-head attention global dependency field; the target multi-head attention global dependency field is input into the patch restoration module to obtain the restored feature field; the restored feature field is input into the convolutional reconstruction module to obtain the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field. Wherein:

[0119] The convolutional local dependency modeling module comprises two convolutional layers. The first convolutional layer contains two convolutions and one activation layer, while the second convolutional layer contains one convolution and one activation layer. Temperature and salinity noise field data are input into the first convolutional layer, where they are processed by two convolutions to obtain convolutional temperature and salinity noise fields, respectively. These convolutional temperature and salinity noise fields are then input into the activation layer to obtain a preliminary local dependency representation. Finally, this preliminary local dependency representation is input into the second convolutional layer, where it is processed by convolutions and the activation layer to obtain the target local dependency representation.

[0120] The multi-head attention global dependency modeling module includes multiple multi-head attention global dependency sub-modules with residual connections. Each multi-head attention global dependency sub-module includes multiple normalization layers and multiple fully connected layers. The flattened patch field is input into the first multi-head attention global dependency sub-module, where it is processed by multiple first-level normalization layers and multiple first-level fully connected layers to obtain the first multi-head attention global dependency field. The flattened patch field and the first multi-head attention global dependency field are then residually connected to obtain the first residual result. The first residual result is input into the second multi-head attention global dependency sub-module, where it is processed by multiple second-level normalization layers and multiple second-level fully connected layers to obtain the second multi-head attention global dependency field. The first and second multi-head attention global dependency fields are then residually connected to obtain the second residual result. The second residual result is input into the third multi-head attention global dependency sub-module for processing, and this process continues until all multi-head attention global dependency sub-modules have completed their processing, resulting in the target multi-head attention global dependency field.

[0121] The first and second normalization layers mentioned above are one of multiple normalization layers, located in different multi-head attention global dependency submodules. The structures of multiple normalization layers can be the same.

[0122] The first and second fully connected layers mentioned above are one of multiple fully connected layers, located in different multi-head attention global dependency submodules, and the structures of multiple fully connected layers can be the same.

[0123] Step S1072: Construct a loss function based on the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field.

[0124] Specifically, determine the training objective of the SRM model; since the SRM model does not directly learn the standardized target field. The joint distribution of all variables Instead, learn fractions The score is the gradient field of the logarithm of the data distribution, indicating the direction in which any starting point in the data space should move to get closer to the high-probability region of the true data distribution; the model's learning objective, i.e., the loss function, can be expressed as:

[0125] ;

[0126] Based on this, the fractional function learned by the SRM model is denoted as... , These are the learnable parameters in the model, including the weights and biases of each network layer; where, This represents all possible noise intensity levels. Take the expected value. Indicates that the pair follows a distribution of Sample expectation, This indicates that each pair follows a distribution. of Sample expectation, Representing the square of the L2 norm, equivalent to The sum of squares of the components of the tensor field in the tensor field; Represents a normalized target field with added noise of varying intensities. ; The distribution is represented The score; Indicates the noise level of the disturbance. The process index represents the change in noise intensity, standardizing the continuous noise addition process to a standard range. Indicates pure noise. Equivalent to a random Gaussian noise field , Indicates clean data. Equivalent to standardized real data ; Follows a standard Gaussian distribution , It is the identity matrix. It is a dimension. This is equivalent to random noise sampled from a standard normal distribution, with a noise level of... ;because ,so Its probability density function is:

[0127] ;

[0128] Taking the logarithm, we get:

[0129] ;

[0130] in, Since it is a constant, we take it with respect to... The gradient is obtained as follows:

[0131] ;

[0132] because , Simplified to Based on this, To further simplify:

[0133] ;

[0134] Therefore, the SRM model is a fit to a fractional function. Its learning involves adding noise. The reverse is equivalent to learning how to remove noise, and then gradually recovering the real data from the noise.

[0135] Step S1073: Train the three-dimensional marine temperature and salinity field reconstruction model using a loss function until the loss function converges or the preset number of iterations is reached, and obtain the trained three-dimensional marine temperature and salinity field reconstruction model.

[0136] Step-by-step training of the SRM model; training Rounds (i.e., the preset number of iterations) are randomly selected from the standardized reanalysis product dataset in each round. indivual The samples form a whole batch sample; for indivual Samples are added sequentially Gaussian noise at various levels This is equivalent to starting from continuous Take out evenly in segments Each value, thus each All samples are available Noise field sample ; An automatic optimizer (such as Adam, AdamW, and SGD) is used for training sequentially using all batches of samples, referring to the loss function. The SRM model is trained to gradually fit and obtain an accurate fractional function. ;

[0137] Ideally, standardized real data should be used. The trained SRM model fits a score function. It can output standardized real data. Score field That is, the direction of denoising, which enables a random Gaussian noise field to be denoised. Denoising to standardize real data .

[0138] Step S108: Use the observation field to guide the trained three-dimensional marine temperature and salinity field reconstruction model to denoise the pure Gaussian noise field, obtain the three-dimensional marine temperature and salinity field reconstruction result, and perform inverse normalization on the three-dimensional marine temperature and salinity field reconstruction result to obtain the reconstructed three-dimensional marine temperature and salinity field.

[0139] 1. Construct the sampling equation, i.e., the denoising equation:

[0140] ;

[0141] in, The fractional function fitted to the trained SRM model. This indicates an observation field based on sparse satellite and field measurement data. The constraints imposed on the denoising process are determined by both through random... Changing hyperparameters and Balancing can be viewed as moving step sizes in two directions. This is equivalent to moving a certain distance along the fractional direction. This represents the Wiener process, which is equivalent to Gaussian noise simulating disturbances that may exist in real-world situations, such as measurement errors.

[0142] in:

[0143] ;

[0144] Equivalent to based on the current observation field The location of the observation point in the noisy target field Select the value at the corresponding position, and then calculate the derivative of the error between the two, which is the direction in which the error increases the fastest. Therefore This indicates that the constraint denoising process is directed towards the observation field. The nearest target field moves a certain distance.

[0145] 2. By means of arrive Iteratively solve the denoising equation in step 1; when hour, Equivalent to a random pure Gaussian noise field , with the original real data It's no longer relevant, when hour, Equivalent to clean, standardized real data after noise reduction. Therefore, based on the denoising equations described above, the model SRM no longer depends on the standardized target field provided by the reanalysis product. Instead of relying on relevant noise information, this information is based on observations from satellite observations and on-site measurements. The constraint is to directly sample real data from random pure Gaussian noise.

[0146] 3. Standardized real data Reconstructed temperature field and salinity field By performing destandardization and regressing to the original distribution range, we obtain the true data field under the true distribution. ; Calculate the temperature field within it and salinity field The expression is as follows:

[0147] ;

[0148] .

[0149] To facilitate understanding by those skilled in the art, a set of preferred embodiments is provided below:

[0150] This embodiment provides a method for reconstructing a three-dimensional marine temperature-salinity field, which specifically includes the following steps:

[0151] S1. Acquire and process satellite sea surface temperature and salinity products and field-measured ocean temperature and salinity profile data, and acquire the three-dimensional field of temperature and salinity from the ocean reanalysis products used to train the model. Step S1 specifically includes the following steps:

[0152] S11. Download the global ocean physical reanalysis product GLORYS12V1 from the open-source data service center, and obtain the three-dimensional temperature field and three-dimensional salinity field from the GLORYS12V1 product, denoted as... and Horizontally covering longitude range arrive Latitude range arrive The grid resolution is ,Right now spatial range ,Right now ,Include Each depth layer has a complete three-dimensional dimension of [number]. The time range is from January 2, 2010 to December 30, 2023, a total of 5111 days, with a time resolution of one field per day. Therefore, there are... A sample is used to train an unconditionally constrained SRM model using unsupervised autoregression; the two are merged into the target field along the channel dimension. The size is The optimal interpolation products for sea surface temperature from the L4-level OISST gridded satellite (version 5.1) were obtained from the Remote Sensing Systems Centre. The resolution is daily, approximately 1 / 10° × 1 / 10°, and the time range is from January 2, 2010 to December 30, 2023, totaling 5111 samples. The sea surface salinity products from the L4-level CCI SSS gridded satellite (version 5.5) were obtained from the European Space Agency. The resolution is daily, approximately 1 / 4° × 1 / 4°, and the time range is from January 2, 2010 to December 30, 2023, totaling 5111 samples. Ocean temperature and salinity profile data were obtained from the in-situ observational temperature and salinity profile database (EN4.2.2 global quality control) released by the UK Met Office Hadley Centre. The data was obtained with a daily temporal resolution, covering the period from January 2, 2010 to December 30, 2023, totaling 5111 samples. However, the horizontal distribution of the profiles was randomized each day. These gridded satellite sea surface temperature and salinity products, along with the in-situ measured ocean temperature and salinity profile data, served as constraints to guide the denoising process of a pre-trained SRM model, generating a three-dimensional ocean temperature and salinity field for the target date.

[0153] S12. Using bilinear interpolation, the L4-level gridded satellite sea surface temperature and salinity products obtained in step S11 are interpolated into a grid field with the same horizontal resolution as the target field, unifying the spatial grid resolution to [value missing]. .

[0154] S13. Using the three-dimensional nearest neighbor interpolation method, two interpolation methods with the same size as the target field are set ( ) three-dimensional mesh field and This is used to unify the field-measured temperature and salinity profile data into a regular grid field. Step S13 specifically includes the following steps:

[0155] S131, The three-dimensional mesh field and The value of each grid position in the 3D mesh field is set to 0; and The three-dimensional temperature field in the ocean reanalysis product corresponding to a certain grid. and three-dimensional salinity field If the value of the grid is NaN, i.e., land, then the three-dimensional grid field will be... and The grid was synchronously set to NaN; each observation point in the field-measured temperature and salinity profiles was incorporated into the three-dimensional grid field. and The nearest position grid in the middle.

[0156] S132. Extract all temperature and salinity measured profiles for the same date, iterate through the measured values ​​in each temperature and salinity profile, and place the measured temperature value in a three-dimensional mesh field. The grid with the closest spatial location is used to place the measured salinity value in the three-dimensional grid field. The grid with the closest spatial location; after traversal, if the 3D grid field... and If a grid has two or more measured points, then the value of that grid is recorded as the average of all the measured values ​​therein.

[0157] S14. Merge satellite observation data and field measured profile data. Combine the interpolated satellite temperature grid field obtained in S12 with the field measured 3D temperature grid field for the same date. The first layer is added together, if If a certain grid in the first layer contains both satellite temperature observations and on-site temperature measurements, then the temperature value of that grid is set as the average of the satellite temperature observations and on-site temperature measurements (i.e., temperature merged observation data). The interpolated satellite salinity grid field obtained in S12 is then compared with the salinity measured three-dimensional grid field for the same date. The first layer is added together, if If a certain grid in the first layer contains both satellite salinity observations and field-measured salinity values, then the salinity value of that grid is set as the average of the satellite salinity observations and field-measured salinity values ​​(i.e., salinity merged observation data). At this point, a complete three-dimensional grid field recording both satellite sea surface observations and field measurements is obtained. (i.e., temperature combined observation data) and (i.e., salinity merged observation data) is used as an observation, and the two are merged along the channel dimension, denoted as the observation field. The size is .

[0158] The fractional marine 3D temperature and salinity field reconstruction model (SRM) can reconstruct a complete marine 3D temperature and salinity field based on limited satellite observations of the sea surface and sparsely distributed field measurements. Essentially, it aims to solve the following inverse problem:

[0159] ;

[0160] in, It indicates the observation field, which is limited to sea surface satellite observations and sparsely distributed field measurements; This represents the complete three-dimensional target field to be reconstructed. This represents Gaussian-distributed noise introduced by uncertainties such as measurement errors; The observation operator represents the ability to observe from a complete three-dimensional target field. The obtained observations are used; typically, deep learning networks learn inverse mappings. The mapping from sparse observations to a complete three-dimensional field is realized, as shown in the following expression:

[0161] ;

[0162] However, the actual spatial distribution of field observations varies randomly and is not a complete, fixed-location grid field, making direct fitting of the inverse mapping impossible. It cannot flexibly adapt to observational information whose distribution location is constantly changing. Therefore, the SRM model does not directly fit the inverse mapping. Instead of an end-to-end model, the approach is to reconstruct a score-based generative model. First, the deep learning network structure of the SRM model is built; details of the model structure are provided in step S2. Then, through self-supervised training, the prior distribution of the physical laws of the product's 3D field is learned and reanalyzed, enabling the model to operate solely based on random Gaussian noise fields. The three-dimensional temperature and salinity field containing physical laws is randomly generated without any constraints. The training process is detailed in step S3. Finally, observation constraints are introduced to restrict the denoising results of each step of SRM so that they are consistent with the observation values. In this way, a specific three-dimensional temperature and salinity field of the ocean is gradually generated, which satisfies both the observation constraints and contains physical laws. The denoising process is detailed in step S4.

[0163] S2. Design and construct a fractional-based oceanic 3D temperature-salinity field reconstruction model (SRM). This model is essentially a fractional generative model, belonging to the diffusion model category. It primarily learns the fractions of the reanalysis product, i.e., the gradient of the logarithm of the data distribution of the reanalysis product, thereby understanding the prior distribution of the intrinsic spatiotemporal evolution laws and physical constraints within the reanalysis product. Step S2 specifically includes the following steps:

[0164] like Figure 2As shown, the complete fractional marine 3D temperature-salinity field reconstruction model (SRM) sequentially includes a convolutional local dependency modeling module, a patch partitioning module, a multi-head attention global dependency modeling module, a patch restoration module, and a convolutional reconstruction module. The convolutional local dependency modeling module captures local features of the input variables, improving the model's sensitivity to local information, and performs preliminary fusion of temperature and salinity features. The patch partitioning module flattens the features to meet the input size requirements of the subsequent multi-head attention global dependency modeling module. The multi-head attention global dependency modeling module better captures the global dependencies between features. The patch restoration module restores the flattened features to their original shape to adapt to the convolution operation. The convolutional reconstruction module maps the output from the feature space back to the dimension of the temperature and salinity fractional field, further enhancing the local features.

[0165] S21. Convolutional local dependency modeling module for fractional ocean three-dimensional temperature and salinity field reconstruction model (SRM).

[0166] The convolutional local dependency modeling module contains two convolutional layers; the first layer has two convolutions. The convolution kernel size is 5×5, and the kernels are respectively for sizes of The input model SRM's random Gaussian noise field Temperature noise field and salinity noise field Perform independent computations to initially extract the local dependencies of the two types of variables (including the temperature noise field and the salinity noise field after convolution), and map the last dimension to... Subsequently, the sizes of the two initial local dependency extraction variable fields are: Then it goes through a GELU activation layer, and then is merged in the last feature dimension to obtain a size of The initial local dependency representation. The second layer has a convolution. A local dependency representation that fuses the local dependencies of two classes of variables, consisting of a GELU activation layer and a 3×3 kernel. Compress the feature dimension to The obtained target local dependency representation The size is The expression for the convolutional local dependency modeling module is as follows:

[0167] .

[0168] S22. Build the patch partitioning module. First, use a module with a size of... ,Right now The window moves up and down in increments of 1. The step size for moving left and right is The scanning size from top to bottom and from left to right is Target local dependency representation Divide it into Each size is The patch; because and Divisible by: and This ensured The value is an integer; after scanning is complete, all patches are stitched together to form a shape of size [size missing]. Patch field Secondly, each patch is flattened into a one-dimensional sequence, and the length of the flattened sequence is... The patch field after being flattened The size is .

[0169] S23. Construct a multi-head attention global dependency modeling module, which includes multiple multi-head attention global dependency sub-modules. Each multi-head attention global dependency sub-module includes multiple layer normalization layers and multiple fully connected layers. Its expression is as follows:

[0170] ;

[0171] Flattened patch field First, it goes through a layer normalization layer. along Normalize the dimension to obtain the normalized patch field. The expression is as follows:

[0172] ;

[0173] Then, Input three fully connected layers respectively Three coding fields were obtained. The expression is as follows:

[0174] ;

[0175] ;

[0176] ;

[0177] Next, we calculate the multi-head self-attention and perform parallel calculations. Self-attention field The calculation process for the self-attention field of one of the heads is as follows: and Perform matrix multiplication, then scale the matrices back to their original size. Then, multiply by 100% and then proceed along the feature dimension. Function activation, and with Perform matrix multiplication to obtain a matrix of size . Self-attention field Each self-attention field The expression is as follows:

[0178] ;

[0179] Merge in the last feature dimension Self-attention field The resulting size is Multi-head self-attention field Then it goes through a normalization layer and a fully connected layer. The compressed feature dimension is The output size is Multi-head attention global dependency field The expression is as follows:

[0180] ;

[0181] The multi-head attention global dependency submodules are sequentially residual-connected, and each multi-head attention global dependency field is denoted as... All sizes Among them, residual connection refers to the... The input to the multi-head attention global dependency submodule is It should be noted that the input to the first multi-head attention global dependency submodule is a patch field. Without residual connections, the input to the second multi-head attention global dependency submodule is... .

[0182] S24. Build the patch restoration module; this module is the reverse process of the patch partitioning module mentioned in step S23; first, divide the dimensions of each patch... Expand, then place each patch back into its original position before partitioning, thus making the size [size missing]. The A multi-head attention global dependency field (i.e., the target multi-head attention global dependency field), restored to a size of Feature field .

[0183] S25. Build a convolutional reconstruction module to process the reconstructed feature field. Reconstruct the structure to a size of Score field The first layer of the last dimension corresponds to the fractional field of the three-dimensional temperature field. The second layer of the last dimension corresponds to the fractional field of the three-dimensional salinity field. The convolutional reconstruction module contains one convolution, one GELU activation layer, and one convolution, as shown in the following expression:

[0184] .

[0185] S26. Connect the above modules in sequence to construct a complete fractional-based three-dimensional ocean temperature-salinity field reconstruction model (SRM), the expression of which is as follows:

[0186] ;

[0187] ;

[0188] in, The functional expression representing the SRM model, i.e., the fractional function, outputs random Gaussian noise. Denoising is directed towards the true data, i.e., the fractional field. Step S3 will detail the fractional function. And how to train the SRM model to accurately fit the fractional function. Step S4 will detail how to fit the trained model SRM, i.e., the fitted fractional function. Provided noise reduction direction Step by step, random Gaussian noise The noise reduction is based on real data.

[0189] S3. Using the ocean reanalysis product obtained in step S1, perform self-supervised training on the SRM model to learn the prior distribution of physical laws in the reanalysis product, enabling it to generate a three-dimensional temperature-salinity field containing physical laws without unconditional constraints based on a three-dimensional random Gaussian noise field. Step S3 specifically includes the following steps:

[0190] S31, will target the field Standardizing the temperature and salinity variables can remove variables with huge differences in distribution dimensions and ranges, unify the data distribution with a mean of 0 and a variance of 1, and improve the convergence speed and stability of the model.

[0191] First, calculate all of them separately. Target field of each sample The mean values ​​of the temperature and salinity variables. and and their respective variances and Since the vertical temperature is not averaged, the above mean and variance are both for a dimension of [missing value]. A one-dimensional vector. The expression is as follows:

[0192] ;

[0193] ;

[0194] ;

[0195] ;

[0196] Then, for the target field The temperature and salinity variables are standardized to obtain the target field of each sample after standardization. The standardized temperature and salinity variables and The size is still Since both the mean and variance are of size ; This is a one-dimensional vector, therefore equivalent to independently standardizing each depth layer; the expression is as follows:

[0197] ;

[0198] .

[0199] S32. Determine the training objective of the SRM model. Since the SRM model does not directly learn the standardized target field... The joint distribution of all variables Instead, learn fractions The score is the gradient field of the logarithm of the data distribution, indicating the direction in which any starting point in the data space should move to get closer to the high-probability region of the true data distribution; the model's learning objective, i.e., the loss function, can be expressed as:

[0200] ;

[0201] Based on this, the fractional function learned by the SRM model is denoted as... , These are the learnable parameters in the model, including the weights and biases of each network layer; where, This represents all possible noise intensity levels. Take the expected value. Indicates that the pair follows a distribution of Sample expectation, This indicates that each pair follows a distribution. of Sample expectation, Representing the square of the L2 norm, equivalent to The sum of squares of the components of the tensor field in the tensor field; Represents a normalized target field with added noise of varying intensities. ; The distribution is represented The score; Indicates the noise level of the disturbance. The process index represents the change in noise intensity, standardizing the continuous noise addition process to a standard range. Indicates pure noise. Equivalent to a random Gaussian noise field , Indicates clean data. Equivalent to standardized real data ; Follows a standard Gaussian distribution , It is the identity matrix. It is a dimension. This is equivalent to random noise sampled from a standard normal distribution, with a noise level of... ;because ,so Its probability density function is:

[0202] ;

[0203] Taking the logarithm, we get:

[0204] ;

[0205] in, Since it is a constant, we take it with respect to... The gradient is obtained as follows:

[0206] ;

[0207] because , Simplified to Based on this, To further simplify:

[0208] ;

[0209] Therefore, the SRM model is a fit to a fractional function. Its learning involves adding noise. The reverse is equivalent to learning how to remove noise, and then gradually recovering the real data from the noise.

[0210] S33, the step-by-step training model SRM, is trained using the open-source deep learning framework PyTorch; training Rounds are randomly selected from the standardized reanalysis product composition dataset in each round. indivual The samples form a whole batch sample; for indivual Samples are added sequentially Gaussian noise at various levels This is equivalent to starting from continuous Take out evenly in segments Each value, thus each All samples are available Noise field sample ; An automatic optimizer (such as Adam, AdamW, and SGD) is used to sequentially utilize all batches of samples (including noise field samples). and original sample data (Used as training labels) for training, referring to the loss function. The SRM model is trained to gradually fit and obtain an accurate fractional function. .

[0211] Ideally, standardized real data should be used. The trained SRM model fits a score function. It can output standardized real data. Score field That is, the direction of denoising, which enables a random Gaussian noise field to be denoised. Denoising to obtain standardized real data For details of the noise reduction process, please refer to step S4.

[0212] S4. Using the satellite sea surface temperature and salinity products obtained in step S1, as well as the field-measured ocean temperature and salinity profile data, as constraints, guide the denoising direction of the SRM model trained in S3, ensuring consistency with the observed values. Furthermore, since the SRM denoising process conforms to the prior distribution of the physical laws of the three-dimensional ocean temperature and salinity field, a specific three-dimensional ocean temperature and salinity field is gradually generated, satisfying both observational constraints and embodying physical laws. Step S4 specifically includes the following steps:

[0213] S41. Construct the sampling equation, that is, establish the denoising equation:

[0214] ;

[0215] in, The fractional function fitted to the trained SRM model. This indicates an observation field based on sparse satellite and field measurement data. The constraints imposed on the denoising process are determined by both through random... Changing hyperparameters and Balancing can be viewed as moving step sizes in two directions. This is equivalent to moving a certain distance along the fractional direction. This represents the Wiener process, which is equivalent to Gaussian noise simulating disturbances that may exist in real-world situations, such as measurement errors.

[0216] in:

[0217] ;

[0218] Equivalent to based on the current observation field The location of the observation point in the noisy target field Select the value at the corresponding position, and then calculate the derivative of the error between the two, which is the direction in which the error increases the fastest. Therefore This indicates that the constraint denoising process is directed towards the observation field. The nearest target field moves a certain distance.

[0219] S42, via from arrive Iteratively solve the denoising equation in step S41; when hour, Equivalent to a random pure Gaussian noise field , with the original real data It's no longer relevant, when hour, Equivalent to clean, standardized real data after noise reduction. Therefore, based on the denoising equations described above, the model SRM no longer depends on the standardized target field provided by the reanalysis product. Instead of relying on relevant noise information, this information is based on observations from satellite observations and on-site measurements. The constraint is to directly sample real data from random pure Gaussian noise.

[0220] S43. Standardized real data Reconstructed temperature field and salinity field By performing destandardization and regressing to the original distribution range, we obtain the true data field under the true distribution. ; Calculate the temperature field within it and salinity field The expression is as follows:

[0221] ;

[0222] .

[0223] Compared with the prior art, the technical solution of this embodiment has the following beneficial effects:

[0224] (1) Achieving high spatiotemporal resolution three-dimensional reconstruction. Existing field observation data is sparse and has low resolution (e.g., monthly average, 1°). However, this embodiment is based on a fractional generation model, which can integrate satellite sea surface observations and field profile data to reconstruct a high spatiotemporal resolution (e.g., daily average, 1 / 12° or higher) ocean three-dimensional temperature and salinity field. This allows the reconstructed three-dimensional temperature and salinity field to meet observation constraints, effectively reveal small- and medium-scale ocean structures, and make up for the problems of insufficient coverage and poor spatiotemporal continuity of traditional field observations.

[0225] (2) Incorporating physical constraints to enhance the rationality of the reconstructed field. This embodiment uses unsupervised learning to reanalyze the prior distribution of physical laws in the product, so that the reconstructed three-dimensional temperature and salinity field not only meets the observation constraints, but also conforms to the inherent spatiotemporal evolution law of the ocean. This overcomes the shortcomings of traditional statistical models that lack physical constraints, and the generated results are closer to the real ocean state, thus improving the reliability and scientific nature of the reconstruction.

[0226] (3) Flexible integration of multi-source observation data to enhance adaptability. The model in this embodiment adopts the observation operator to guide the denoising process, which can dynamically integrate satellite sea surface temperature and salinity products as well as spatially randomly distributed field measured profile data. This avoids dependence on fixed grid observations, improves the adaptability of the reconstruction process to sparse and irregular observations, and ensures that the generated field (i.e. the reconstructed three-dimensional temperature and salinity field) is consistent with the measured values ​​(i.e., field measured ocean temperature and salinity profile data).

[0227] (4) Low operating cost and high update speed, supporting real-time application. Compared with numerical models and assimilation methods, which require a lot of computing resources and have a slow update speed (months to years), the reconstruction method based on the generative model in this embodiment has low operating cost and fast update speed. It can generate three-dimensional temperature and salinity fields in near real time, meeting the military and civilian needs with high timeliness requirements such as underwater sound field calculation and vortex identification.

[0228] Reference Figure 3 This application also provides a three-dimensional marine temperature and salinity field reconstruction system, which includes:

[0229] The data acquisition unit 301 is used to acquire gridded satellite sea surface temperature and salinity product data, field measured ocean temperature and salinity profile data, and three-dimensional temperature field and three-dimensional salinity field in ocean reanalysis products.

[0230] The first merging unit 302 is used to merge the three-dimensional temperature field and the three-dimensional salinity field in the marine reanalysis product to construct the target field;

[0231] The data preprocessing unit 303 is used to preprocess satellite sea surface temperature and salinity product data and field-measured ocean temperature and salinity profile data to obtain preprocessed satellite sea surface temperature and salinity product data and preprocessed field-measured ocean temperature and salinity profile data.

[0232] The second merging unit 304 is used to merge the preprocessed satellite sea surface temperature and salinity product data with the preprocessed field measured ocean temperature and salinity profile data to construct an observation field.

[0233] The noise addition unit 305 is used to standardize the sample data in the target field to obtain the standardized target field. Multiple levels of Gaussian noise are added to each sample data in the standardized target field in sequence to construct a noise field sample dataset.

[0234] Model building unit 306 is used to build a marine three-dimensional temperature and salinity field reconstruction model that includes a convolutional local dependency modeling module, a patch partitioning module, a multi-head attention global dependency modeling module, a patch restoration module, and a convolutional reconstruction module.

[0235] Model training unit 307 is used to train the three-dimensional temperature and salinity field reconstruction model of the ocean using a noise field sample dataset, so as to obtain a trained three-dimensional temperature and salinity field reconstruction model of the ocean.

[0236] The temperature and salinity field reconstruction unit 308 is used to denoise the pure Gaussian noise field by using the ocean three-dimensional temperature and salinity field reconstruction model trained by the observation field, to obtain the ocean three-dimensional temperature and salinity field reconstruction result, and to denormalize the ocean three-dimensional temperature and salinity field reconstruction result to obtain the reconstructed ocean three-dimensional temperature and salinity field.

[0237] It should be noted that since the marine three-dimensional temperature and salinity field reconstruction system in this embodiment is based on the same inventive concept as the marine three-dimensional temperature and salinity field reconstruction method described above, the corresponding content in the method embodiment is also applicable to this system embodiment, and will not be described in detail here.

[0238] Reference Figure 4 This application also provides an electronic device, which includes:

[0239] At least one memory;

[0240] At least one processor;

[0241] At least one program;

[0242] The program is stored in memory, and the processor executes at least one program to implement the above-described method for reconstructing the three-dimensional marine temperature and salinity field according to this disclosure.

[0243] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0244] The electronic devices according to embodiments of this application will now be described in detail.

[0245] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0246] The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to implement the marine three-dimensional temperature and salinity field reconstruction method of the embodiments of this disclosure.

[0247] The input / output interface 1800 is used to implement information input and output.

[0248] The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0249] Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900);

[0250] The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0251] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the above-described marine three-dimensional temperature and salinity field reconstruction method.

[0252] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0253] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0254] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0255] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0256] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0257] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0258] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0259] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0260] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0261] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0262] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. The embodiments of this application have been described in detail above with reference to the accompanying drawings, but this application is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of this application.

[0263] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A method for reconstructing a three-dimensional marine temperature-salinity field, characterized in that, The method includes: Acquire gridded satellite sea surface temperature and salinity product data, field-measured ocean temperature and salinity profile data, and three-dimensional temperature and salinity fields from ocean reanalysis products; The three-dimensional temperature field and three-dimensional salinity field in the aforementioned marine reanalysis product are merged to construct the target field; The satellite sea surface temperature and salinity product data and the field-measured ocean temperature and salinity profile data are preprocessed to obtain preprocessed satellite sea surface temperature and salinity product data and preprocessed field-measured ocean temperature and salinity profile data. The preprocessed satellite sea surface temperature and salinity product data are combined with the preprocessed field-measured ocean temperature and salinity profile data to construct an observation field; The sample data in the target field are standardized to obtain a standardized target field. Multiple levels of Gaussian noise are added to each sample data in the standardized target field to construct a noise field sample dataset. A three-dimensional marine temperature and salinity field reconstruction model is constructed, comprising a convolutional local dependency modeling module, a patch partitioning module, a multi-head attention global dependency modeling module, a patch restoration module, and a convolutional reconstruction module. The convolutional local dependency modeling module includes two convolutional layers: the first layer contains two convolutions and one activation layer, and the second layer contains one convolution and one activation layer. The multi-head attention global dependency modeling module includes multiple residual-connected multi-head attention global dependency sub-modules. Each multi-head attention global dependency sub-module includes multiple layer normalization layers and multiple fully connected layers, including: The temperature noise field data and salinity noise field data in the noise field sample dataset are input into the convolutional local dependency modeling module to obtain the target local dependency representation. The target local dependency representation is input into the patch partitioning module to obtain the flattened patch field; The flattened patch field is input into the multi-head attention global dependency modeling module to obtain the target multi-head attention global dependency field. The target multi-head attention global dependency field is input into the patch restoration module to obtain the restored feature field; The restored feature field is input into the convolutional reconstruction module to obtain the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field. The ocean three-dimensional temperature and salinity field reconstruction model is trained using the noise field sample dataset to obtain a trained ocean three-dimensional temperature and salinity field reconstruction model. The observation field is used to guide the trained three-dimensional ocean temperature and salinity field reconstruction model to denoise the pure Gaussian noise field, thereby obtaining the three-dimensional ocean temperature and salinity field reconstruction result. The three-dimensional ocean temperature and salinity field reconstruction result is then de-normalized to obtain the reconstructed three-dimensional ocean temperature and salinity field.

2. The method for reconstructing a three-dimensional marine temperature and salinity field according to claim 1, characterized in that, The preprocessing of the satellite sea surface temperature and salinity product data and the field-measured ocean temperature and salinity profile data to obtain preprocessed satellite sea surface temperature and salinity product data and preprocessed field-measured ocean temperature and salinity profile data includes: Bilinear interpolation is used to interpolate the satellite sea surface temperature and salinity product data into a grid with the same horizontal resolution as the target field, resulting in an interpolated satellite temperature grid field and an interpolated satellite salinity grid field. The interpolated satellite temperature grid field and the interpolated satellite salinity grid field are used as the preprocessed satellite sea surface temperature and salinity product data. Two three-dimensional grid fields with the same size as the target field are set up. The measured ocean temperature and salinity profile data are interpolated into the two three-dimensional grid fields by three-dimensional nearest neighbor interpolation, respectively, to obtain the measured temperature three-dimensional grid field and the measured salinity three-dimensional grid field. The measured temperature three-dimensional grid field and the measured salinity three-dimensional grid field are used as the preprocessed measured ocean temperature and salinity profile data.

3. The method for reconstructing a three-dimensional marine temperature and salinity field according to claim 2, characterized in that, The process of merging the preprocessed satellite sea surface temperature and salinity product data with the preprocessed field-measured ocean temperature and salinity profile data to construct an observation field includes: The interpolated satellite temperature grid field is added to the first layer data of the temperature measurement three-dimensional grid field on the same day. If several grids in the first layer of the temperature measurement three-dimensional grid field have both satellite temperature observation values ​​and on-site temperature measurement values, the average value of the satellite temperature observation values ​​and the on-site temperature measurement values ​​is taken as the temperature value of several grids in the first layer of the temperature measurement three-dimensional grid field, thus obtaining the merged temperature observation data. The interpolated satellite salinity grid field is added to the first layer data of the salinity measurement three-dimensional grid field on the same day. If several grids in the first layer of the salinity measurement three-dimensional grid field have both satellite salinity observation values ​​and on-site salinity measurement values, the average value of the satellite salinity observation values ​​and the on-site salinity measurement values ​​is taken as the salinity value of several grids in the first layer of the salinity measurement three-dimensional grid field, thus obtaining the salinity merged observation data. The combined temperature and salinity observation data are merged along the channel dimension to obtain the observation field.

4. The method for reconstructing a three-dimensional marine temperature and salinity field according to claim 1, characterized in that, The step of training the ocean three-dimensional temperature and salinity field reconstruction model using the noise field sample dataset to obtain the trained ocean three-dimensional temperature and salinity field reconstruction model includes: The temperature noise field data and the salinity noise field data are input into the three-dimensional ocean temperature and salinity field reconstruction model to obtain the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field. A loss function is constructed based on the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field. The loss function is used to train the three-dimensional marine temperature and salinity field reconstruction model until the loss function converges or reaches a preset number of iterations, thus obtaining the trained three-dimensional marine temperature and salinity field reconstruction model.

5. The method for reconstructing a three-dimensional marine temperature and salinity field according to claim 1, characterized in that, The step of inputting the temperature noise field data and the salinity noise field data into the convolutional local dependency modeling module to obtain the target local dependency representation includes: The temperature noise field data and the salinity noise field data are input into the first convolutional layer. The temperature noise field data and the salinity noise field data are convolved by the two convolutions respectively to obtain the convolved temperature noise field and the convolved salinity noise field. The convolutional temperature noise field and the convolutional salinity noise field are input into the activation layer to obtain a preliminary local dependency characterization. The preliminary local dependency representation is input into the second convolutional layer, and the preliminary local dependency representation is processed by convolution and activation layers to obtain the target local dependency representation.

6. The method for reconstructing a three-dimensional marine temperature and salinity field according to claim 1, characterized in that, The step of inputting the flattened patch field into the multi-head attention global dependency modeling module to obtain the target multi-head attention global dependency field includes: The flattened patch field is input into the first multi-head attention global dependency submodule. The flattened patch field is processed by multiple first-layer normalization layers and multiple first-layer fully connected layers to obtain the first multi-head attention global dependency field. The flattened patch field and the first multi-head attention global dependency field are residually connected to obtain the first residual result; The first residual result is input into the second multi-head attention global dependency submodule, and the first residual result is processed through multiple second-layer normalization layers and multiple second-layer fully connected layers to obtain the second multi-head attention global dependency field. The first multi-head attention global dependency field and the second multi-head attention global dependency field are residually connected to obtain the second residual result; The second residual result is input into the third multi-head attention global dependency submodule for processing. The process continues until all multi-head attention global dependency submodules have finished processing, at which point the target multi-head attention global dependency field is obtained.

7. A three-dimensional marine temperature and salinity field reconstruction system, characterized in that, The system includes: The data acquisition unit is used to acquire gridded satellite sea surface temperature and salinity product data, field-measured ocean temperature and salinity profile data, and three-dimensional temperature field and three-dimensional salinity field in ocean reanalysis products; The first merging unit is used to merge the three-dimensional temperature field and the three-dimensional salinity field in the marine reanalysis product to construct the target field; The data preprocessing unit is used to preprocess the satellite sea surface temperature and salinity product data and the field-measured ocean temperature and salinity profile data to obtain preprocessed satellite sea surface temperature and salinity product data and preprocessed field-measured ocean temperature and salinity profile data. The second merging unit is used to merge the preprocessed satellite sea surface temperature and salinity product data with the preprocessed field-measured ocean temperature and salinity profile data to construct an observation field. The noise addition unit is used to standardize the sample data in the target field to obtain a standardized target field, and to add multiple levels of Gaussian noise to each sample data in the standardized target field in sequence to construct a noise field sample dataset. The model building unit is used to construct a three-dimensional marine temperature and salinity field reconstruction model, which includes a convolutional local dependency modeling module, a patch partitioning module, a multi-head attention global dependency modeling module, a patch restoration module, and a convolutional reconstruction module. The convolutional local dependency modeling module includes two convolutional layers: the first layer contains two convolutions and one activation layer, and the second layer contains one convolution and one activation layer. The multi-head attention global dependency modeling module includes multiple residual-connected multi-head attention global dependency sub-modules. Each multi-head attention global dependency sub-module includes multiple layer normalization layers and multiple fully connected layers, including: The temperature noise field data and salinity noise field data in the noise field sample dataset are input into the convolutional local dependency modeling module to obtain the target local dependency representation. The target local dependency representation is input into the patch partitioning module to obtain the flattened patch field; The flattened patch field is input into the multi-head attention global dependency modeling module to obtain the target multi-head attention global dependency field. The target multi-head attention global dependency field is input into the patch restoration module to obtain the restored feature field; The restored feature field is input into the convolutional reconstruction module to obtain the fractional fields of the three-dimensional temperature field and the three-dimensional salinity field. The model training unit is used to train the ocean three-dimensional temperature and salinity field reconstruction model using the noise field sample dataset to obtain the trained ocean three-dimensional temperature and salinity field reconstruction model. The temperature and salinity field reconstruction unit is used to guide the trained ocean three-dimensional temperature and salinity field reconstruction model with the observation field to denoise the pure Gaussian noise field, obtain the ocean three-dimensional temperature and salinity field reconstruction result, and perform denormalization on the ocean three-dimensional temperature and salinity field reconstruction result to obtain the reconstructed ocean three-dimensional temperature and salinity field.

8. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform the marine three-dimensional temperature and salinity field reconstruction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the marine three-dimensional temperature and salinity field reconstruction method as described in any one of claims 1 to 6.

Images