A method and system for reconstructing ocean near-surface flow field

Abstract

The present application relates to a method and system for reconstructing a near-surface flow field of an ocean. The method calculates the linear similarity between a sea surface height data sequence of a target grid and a sea surface height data sequence of a reference grid, calculates a first distance corresponding to a reference grid closest to the target grid according to the linear similarity, calculates an average value of the first distances corresponding to all target grids, takes the average value of the grid in the latitudinal direction as a latitudinal spatial decorrelation scale, and takes the average value of the grid in the longitudinal direction as a longitudinal spatial decorrelation scale; and forms a window by taking the grid where the position point is located as the center point and taking a rectangle constructed by the latitudinal spatial decorrelation scale and the longitudinal spatial decorrelation scale as a physical constraint to constrain the spatial range of each reconstructed field. Compared with a purely data-driven deep learning model, the method takes into account the actual physical conditions of the sea surface, can improve the generalization ability of the reconstruction model, and improve the accuracy of reconstructing the near-surface flow field of the ocean.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of ocean data processing technology, and in particular to a method and system for reconstructing an ocean near-surface flow field. Background Art

[0002] The near-surface flow field of the ocean (referred to as the sea surface flow field) is important reference information for marine climate research, marine disaster prevention and mitigation, and maritime military activities.

[0003] In order to obtain the global sea surface flow field, it has become an important means to indirectly reconstruct the large-area sea surface flow field through satellite remote sensing data. At present, there are three main forms of reconstructing the sea surface flow field using satellite remote sensing sea surface temperature. One is to invert the sea surface flow field by moving the temperature of the final infrared image. This method is called the maximum correlation coefficient method. Considering that the surface depth of the sea surface skin temperature is about a few microns, this means that the sea-air interaction can affect the ocean structure. It is more appropriate to use brightness temperature, but the sea surface "skin effect" will affect the inversion results and cause errors. The second method of sea surface temperature inversion is the dynamic method, that is, solving the surface heat equation. The advantage is that it can comprehensively utilize data from different sources. How to effectively fuse the data is a new challenge. The third method is to extract the sea surface flow field from a single sea surface temperature image, such as the effective surface quasi-geostrophic method, the multi-fractal structure inversion method, or the finite size Lyapunov index method. However, there are two defects in using sea surface temperature to reconstruct the sea surface flow field. First, the sea surface temperature itself will be affected by the atmosphere (especially clouds), which will introduce errors and then be transmitted to the sea surface flow field. Second, changes in sea surface temperature are not necessarily the result of the sea surface flow field. Strong sea-air interactions and upwelling / downwelling currents may cause changes in the subtle structure of the sea surface temperature.

[0004] In recent years, deep learning algorithms have been widely used in the field of marine science. Compared with the indirect reconstruction of large-area sea surface flow fields based on satellite remote sensing data, the sea surface flow fields reconstructed based on deep learning methods have great advantages. However, the goal is mainly to build a reconstruction model of deep learning driven by pure data. It is not enough to rely solely on the mining of the changing laws of pure ocean flow field data. There are defects such as poor generalization ability and poor reconstruction accuracy. Summary of the invention

[0005] The following is a summary of the subject matter described in detail herein. This summary is not intended to limit the scope of the claims.

[0006] The main purpose of the embodiments of the present disclosure is to propose a method and system for reconstructing the near-surface flow field of the ocean, which can improve the accuracy of reconstructing the near-surface flow field.

[0007] A first aspect of an embodiment of the present application provides a method for reconstructing an ocean near-surface flow field, the method comprising:

[0008] Obtain the average sea surface height data product;

[0009] Determine a first linear similarity between a sea surface height data sequence corresponding to a target grid and a sea surface height data sequence corresponding to a first benchmark grid; the target grid is any one of a plurality of grids of the daily average sea surface height data product, the first benchmark grid is different from the target grid, and the first benchmark grid is any one of the plurality of grids at the same latitude as the target grid, the sea surface height data sequence corresponding to the target grid is a corresponding sea surface height data sequence of the target grid in the daily average sea surface height data product, and the sea surface height data sequence corresponding to the first benchmark grid is a corresponding sea surface height data sequence of the first benchmark grid in the daily average sea surface height data product;

[0010] Determine a first reference grid whose first linear similarity is less than a preset first threshold and is closest to the target grid, and determine a first distance between the target grid and the first reference grid that is closest to the target grid;

[0011] Calculating a first average value of the first distances corresponding to all the target grids;

[0012] Determine a second linear similarity between a sea surface height data sequence corresponding to the target grid and a sea surface height data sequence corresponding to a second reference grid; the second reference grid is any one of the plurality of grids that is at the same longitude as the target grid, and the sea surface height data sequence corresponding to the second reference grid is a corresponding sea surface height data sequence of the second reference grid in the sky average sea surface height data product;

[0013] Determine a second reference grid whose second linear similarity is less than a preset second threshold and is closest to the target grid, and determine a second distance between the target grid and the second reference grid that is closest to the target grid;

[0014] Calculating a second average value of the second distances corresponding to all the target grids;

[0015] Obtain satellite remote sensing data and measured ocean near-surface flow field data for multiple preset locations;

[0016] Taking the grid where the position point is located as the center and the rectangular area of ​​the first average value and the second average value as the window, intercepting the data of the input variable of the position point with a size of the window from the satellite remote sensing data as the input sample, and taking the measured ocean near-surface flow field data of the position point as the label sample;

[0017] Using the multiple groups of input samples and the label samples as training data, training the neural network according to the training data, and obtaining a reconstructed model after the training is completed;

[0018] The target ocean near-surface flow field data to be reconstructed is input into the reconstruction model to obtain the target ocean near-surface flow field reconstructed by the reconstruction model.

[0019] The method for reconstructing the near-surface flow field of the ocean provided in this application has at least the following beneficial effects:

[0020] This embodiment calculates the linear similarity between the sea surface height data sequence corresponding to the target grid and the sea surface height data sequence corresponding to the benchmark grid in the daily average sea surface height data product, calculates the first distance corresponding to the benchmark grid closest to the target grid according to the linear similarity, calculates the average value of the first distances corresponding to all the target grids, takes the average value of the grid in the latitudinal direction as the latitudinal spatial decorrelation scale, and takes the average value of the grid in the longitudinal direction as the longitudinal spatial decorrelation scale; and forms a window by taking the grid where the position point is located as the center point and taking the rectangle constructed by the latitudinal spatial decorrelation scale and the longitudinal spatial decorrelation scale as the physical constraint to constrain the spatial range of each reconstructed field, thereby improving the interpretability of the neural network. Compared with a purely data-driven deep learning model, the actual physical conditions of the sea surface are taken into account, and the generalization ability of the reconstruction model can be improved, and the accuracy of reconstructing the near-surface flow field of the ocean can be improved.

[0021] In some embodiments, the calculation formula for determining the first linear similarity is:

[0022] ;

[0023] ;

[0024] ;

[0025] in, is the sea surface height data sequence corresponding to the target grid The sea surface height data sequence corresponding to the first benchmark grid The first linear similarity between is the sea surface height data series exist The sea surface height data at the moment, is the sea surface height data series exist The sea surface height data at the moment, is the time span of the daily mean sea surface height data product.

[0026] In some embodiments, the calculation formula for determining the second linear similarity is:

[0027] ;

[0028] ;

[0029] ;

[0030] in, is the sea surface height data sequence corresponding to the target grid The sea surface height data sequence corresponding to the second benchmark grid The second linear similarity between is the sea surface height data series exist The sea surface height data at the moment, is the sea surface height data series exist The sea surface height data at the moment, is the time span of the daily mean sea surface height data product.

[0031] In some embodiments, the input variables include at least two of absolute dynamic altitude, absolute dynamic altitude latitudinal gradient, absolute dynamic altitude meridional gradient, satellite sea surface temperature, sea surface temperature meridional gradient, sea surface temperature latitudinal gradient, sea surface wind speed, sea surface meridional wind speed, sea surface latitudinal wind speed, global longitude grid field, global latitude grid field and Coriolis parameters;

[0032] The measured ocean near-surface flow field data include measured sea surface latitudinal flow velocity and sea surface meridional flow velocity.

[0033] In some embodiments, the neural network includes a feature encoding module, a multi-frequency attention module, and a feature reconstruction module;

[0034] The neural network is trained according to the training data, and a reconstructed model is obtained after the training is completed, including:

[0035] The neural network is trained multiple times according to the training data until the neural network converges to obtain a reconstructed model; each training process includes:

[0036] Inputting the training data into the feature encoding module to perform feature extraction on the data of each input variable in the training data to obtain a feature map corresponding to the data of each input variable, and fusing the feature maps corresponding to the data of all the input variables to obtain a first fused feature map;

[0037] Inputting the first fused feature map into the multi-frequency attention module to divide the first fused feature map into sub-feature maps of multiple frequency components, and independently capturing the attention feature map corresponding to the sub-feature map of each frequency component;

[0038] Inputting the attention feature maps corresponding to the sub-feature maps of all frequency components into the feature reconstruction module, so as to obtain a second fused feature map by fusing the attention feature maps corresponding to the sub-feature maps of all frequency components, and obtaining a reconstructed ocean near-surface flow field according to the second fused feature map;

[0039] Calculating the loss based on the reconstructed ocean surface flow field and the label samples;

[0040] Parameter gradients of the feature encoding module, the multi-frequency attention module and the feature reconstruction module are updated according to the loss.

[0041] In some embodiments, the formula for calculating the first fusion feature map is:

[0042] ;

[0043] in, to The data of the first input variable to The data of the input variables, is the number of input variables, , The mapping function of the convolution layer is 1×1 convolution kernel. for The mapping function of the activation function, is the mapping function of the fusion operation, It is the first fused feature map.

[0044] In some embodiments, the multi-frequency attention module includes L branch networks, and each of the branch networks includes M multi-head self-attention modules connected in sequence; wherein L is the same as the number of the frequency components; and M is a positive integer greater than 1;

[0045] The first fused feature map is divided into sub-feature maps of multiple frequency components, and an attention feature map corresponding to the sub-feature map of each frequency component is independently captured, including:

[0046] Dividing the first fused feature map into a plurality of sub-feature maps of different frequency components based on wavelet transform;

[0047] Inputting the sub-feature map of the target frequency component into a corresponding one of the branch networks, so as to perform feature learning on the sub-feature map of the target frequency component through M sequentially connected multi-head self-attention modules, and obtaining a feature map corresponding to the sub-feature map of the target frequency component output by the M-th multi-head self-attention module; the sub-feature map of the target frequency component is a sub-feature map of any one frequency component among the sub-feature maps of multiple different frequency components;

[0048] The sub-feature map of the target frequency component and the feature map corresponding to the sub-feature map of the target frequency component are residually connected to obtain an attention feature map corresponding to the sub-feature map of the target frequency component.

[0049] A second aspect of the embodiment of the present application provides a system for reconstructing a near-surface flow field of an ocean, the system comprising:

[0050] A data acquisition unit, used to obtain the sky average sea surface height data product;

[0051] a latitudinal decorrelation scale acquisition unit, configured to determine a first linear similarity between a sea surface height data sequence corresponding to a target grid and a sea surface height data sequence corresponding to a first benchmark grid; the target grid is any one of a plurality of grids of the sky-average sea surface height data product, the first benchmark grid is different from the target grid, and the first benchmark grid is any one of the plurality of grids at the same latitude as the target grid, the sea surface height data sequence corresponding to the target grid is the corresponding sea surface height data sequence of the target grid in the sky-average sea surface height data product, and the sea surface height data sequence corresponding to the first benchmark grid is the corresponding sea surface height data sequence of the first benchmark grid in the sky-average sea surface height data product;

[0052] and, used to determine a first reference grid whose first linear similarity is less than a preset first threshold and is closest to the target grid, and determine a first distance between the target grid and the first reference grid that is closest to the target grid;

[0053] and, used for calculating a first average value of the first distances corresponding to all the target grids;

[0054] a longitudinal decorrelation scale acquisition unit, configured to determine a second linear similarity between a sea surface height data sequence corresponding to the target grid and a sea surface height data sequence corresponding to a second reference grid; the second reference grid is any one of the plurality of grids that is at the same longitude as the target grid, and the sea surface height data sequence corresponding to the second reference grid is a corresponding sea surface height data sequence of the second reference grid in the sky average sea surface height data product;

[0055] And, used to determine a second reference grid whose second linear similarity is less than a preset second threshold and is closest to the target grid, and determine a second distance between the target grid and the second reference grid that is closest to the target grid;

[0056] and, used for calculating a second average value of the second distances corresponding to all the target grids;

[0057] A model training unit, used to obtain satellite remote sensing data of multiple preset locations and measured ocean near-surface flow field data;

[0058] And, using the grid where the position point is located as the center and the rectangular area of ​​the first average value and the second average value as a window, intercepting the data of the input variable of the position point with a size of the window from the satellite remote sensing data as input samples, and using the measured ocean near-surface flow field data of the position point as label samples;

[0059] And, used to use the multiple groups of the input samples and the label samples as training data, train the neural network according to the training data, and obtain a reconstructed model after the training is completed;

[0060] The model application unit is used to input the target ocean near-surface flow field data to be reconstructed into the reconstruction model to obtain the target ocean near-surface flow field reconstructed by the reconstruction model.

[0061] The third aspect of an embodiment of the present application proposes an electronic device, comprising at least one controller and a memory for communicating with the at least one controller; the memory stores instructions that can be executed by the at least one controller, and the instructions are executed by the at least one controller so that the at least one controller can execute the ocean near-surface flow field reconstruction method as described above.

[0062] A fourth aspect of an embodiment of the present application proposes a computer-readable storage medium, which stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the above-mentioned method for reconstructing the near-surface flow field of the ocean.

[0063] It can be understood that the beneficial effects of the second to fourth aspects compared with the related art are the same as the beneficial effects of the first aspect compared with the related art. Please refer to the relevant description in the first aspect, and no further details will be given here. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0065] Figure 1 is a schematic diagram of an embodiment of a method for reconstructing a near-surface flow field of an ocean provided by the present application;

[0066] Figure 2 It is a structural diagram of a feature coding module provided by an embodiment of the present application;

[0067] Figure 3 is a schematic diagram of the structure of a multi-frequency attention module provided by an embodiment of the present application;

[0068] Figure 4 It is a structural diagram of a feature reconstruction module provided by an embodiment of the present application;

[0069] Figure 5 is a schematic diagram of the structure of a reconstruction model provided by an embodiment of the present application;

[0070] Figure 6 is a schematic diagram of an embodiment of an ocean near-surface flow field reconstruction system provided by the present application;

[0071] Figure 7 It is a schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0073] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0074] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0075] In some embodiments of the present application, this embodiment provides a method for reconstructing a near-surface flow field of an ocean, the method comprising:

[0076] Step S110, obtaining the sky average sea level height data product;

[0077] Step S120, determining a first linear similarity between a sea surface height data sequence corresponding to a target grid and a sea surface height data sequence corresponding to a first benchmark grid; the target grid is any one of a plurality of grids of a sky-average sea surface height data product, the first benchmark grid is different from the target grid, and the first benchmark grid is any one of a plurality of grids at the same latitude as the target grid, the sea surface height data sequence corresponding to the target grid is a corresponding sea surface height data sequence of the target grid in the sky-average sea surface height data product, and the sea surface height data sequence corresponding to the first benchmark grid is a corresponding sea surface height data sequence of the first benchmark grid in the sky-average sea surface height data product;

[0078] Step S130, determining a first reference grid whose first linear similarity is less than a preset first threshold and which is closest to the target grid, and determining a first distance between the target grid and the first reference grid that is closest to the target grid;

[0079] Step S140, calculating a first average value of first distances corresponding to all target grids;

[0080] Step S150, determining a second linear similarity between a sea surface height data sequence corresponding to the target grid and a sea surface height data sequence corresponding to a second reference grid; the second reference grid is any one of the plurality of grids that is at the same longitude as the target grid, and the sea surface height data sequence corresponding to the second reference grid is a corresponding sea surface height data sequence of the second reference grid in the sky average sea surface height data product;

[0081] Step S160, determining a second reference grid whose second linear similarity is less than a preset second threshold and is closest to the target grid, and determining a second distance between the target grid and the second reference grid that is closest to the target grid;

[0082] Step S170, calculating a second average value of the second distances corresponding to all target grids;

[0083] Step S180, obtaining satellite remote sensing data and measured ocean near-surface flow field data of a plurality of preset location points;

[0084] Step S190, taking the grid where the location point is located as the center and the rectangular area of ​​the first average value and the second average value as the window, intercepting the data of the input variable of the location point of the window size from the satellite remote sensing data as the input sample, and taking the measured ocean near-surface flow field data of the location point as the label sample;

[0085] Step S200, using multiple groups of input samples and label samples as training data, training the neural network according to the training data, and obtaining a reconstructed model after the training is completed;

[0086] Step S210, inputting the target ocean near-surface flow field data to be reconstructed into the reconstruction model to obtain the target ocean near-surface flow field reconstructed by the reconstruction model.

[0087] In this step S110, the daily average sea level height data product is obtained. The daily average sea level height data product can come from the satellite sea level height L4 product of CMEMS. For details, please refer to the subsequent embodiment description.

[0088] The purpose of steps S120 to S140 is to obtain the latitudinal spatial decorrelation scale. The purpose of steps S150 to S170 is to obtain the longitudinal spatial decorrelation scale.

[0089] In step S120, a first linear similarity between a sea surface height data sequence corresponding to the target grid and a sea surface height data sequence corresponding to the first reference grid is calculated.

[0090] In step S130, a first threshold is first preset, then a first reference grid whose first linear similarity is less than the first threshold and whose distance to the target grid is closest is determined, and finally a first distance between the target grid and the first reference grid closest to the target grid is calculated.

[0091] In step S140 , since each grid can be used as a target grid, the first distance corresponding to each target grid is counted, and finally the first average value of all the first distances is calculated.

[0092] In step S150, a second linear similarity between the sea surface height data sequence corresponding to the target grid and the sea surface height data sequence corresponding to the second reference grid is calculated.

[0093] In step S160, a second threshold is first preset, then a second reference grid whose second linear similarity is less than the second threshold and whose distance to the target grid is closest is determined, and finally a first distance between the target grid and the second reference grid closest to the target grid is calculated.

[0094] In step S170 , because each grid can be used as a target grid, the second distance corresponding to each target grid is counted, and finally the second average value of all the second distances is calculated.

[0095] The purpose of steps S180 to S200 is to train the reconstruction model.

[0096] In step S180, satellite remote sensing data and measured ocean near-surface flow field data of a plurality of preset location points are obtained, and then input sample data is obtained from the satellite remote sensing data, and the measured ocean near-surface flow field data is the label data.

[0097] In step S190, with the grid where the location point is located as the center and the rectangular area of ​​the first average value and the second average value as the window, the data of the input variables of the location point of the window size are intercepted from the satellite remote sensing data as the input samples, and the measured ocean near-surface flow field data of the location point are used as label samples.

[0098] This embodiment calculates the linear similarity between the sea surface height data sequence corresponding to the target grid and the sea surface height data sequence corresponding to the benchmark grid in the daily average sea surface height data product, calculates the first distance corresponding to the benchmark grid closest to the target grid according to the linear similarity, calculates the average value of the first distances corresponding to all target grids, takes the average value of the grid in the latitudinal direction as the latitudinal spatial decorrelation scale, and takes the average value of the grid in the longitudinal direction as the longitudinal spatial decorrelation scale. By forming a window with the grid where the position point is located as the center point, the latitudinal spatial decorrelation scale and the rectangle in the longitudinal space as physical constraints, the spatial range of each reconstructed field is constrained, which can improve the interpretability of the neural network. Compared with a purely data-driven deep learning model, the actual physical conditions of the sea surface are taken into account, the generalization ability of the reconstruction model can be improved, and the accuracy of reconstructing the near-surface flow field of the ocean can be improved.

[0099] In some embodiments of the present application, the input variables include absolute dynamic altitude, absolute dynamic altitude latitudinal gradient, absolute dynamic altitude meridional gradient, satellite sea surface temperature, sea surface temperature meridional gradient, sea surface temperature latitudinal gradient, sea surface wind speed, sea surface meridional wind speed, sea surface latitudinal wind speed, global longitude grid field, global latitude grid field and at least two of Coriolis parameters.

[0100] The measured ocean near-surface flow field data include the measured sea surface latitudinal current velocity and sea surface meridional current velocity.

[0101] In this embodiment, compared with a single input variable, at least two of the above variables are integrated to reconstruct the flow field, which overcomes the limitations of the traditional use of variables such as sea surface height or sea surface temperature alone, integrates the thermodynamic and kinetic relationships between multiple variables, and improves the reconstruction accuracy.

[0102] In some embodiments of the present application, the neural network includes a feature encoding module, a multi-frequency attention module, and a feature reconstruction module;

[0103] The neural network is trained according to the training data, and a reconstructed model is obtained after the training is completed, including:

[0104] The neural network is trained multiple times according to the training data until the neural network converges and a reconstructed model is obtained; each training process includes:

[0105] Step S310, inputting the training data into the feature encoding module to perform feature extraction on the data of each input variable in the training data, obtaining a feature map corresponding to the data of each input variable, and fusing the feature maps corresponding to the data of all input variables to obtain a first fused feature map;

[0106] Step S320, inputting the first fused feature map into a multi-frequency attention module to divide the first fused feature map into sub-feature maps of multiple frequency components, and independently capturing an attention feature map corresponding to the sub-feature map of each frequency component;

[0107] Step S330, inputting the attention feature graphs corresponding to the sub-feature graphs of all frequency components into the feature reconstruction module, so as to obtain a second fused feature graph by fusing the attention feature graphs corresponding to the sub-feature graphs of all frequency components, and obtaining a reconstructed ocean near-surface flow field according to the second fused feature graph;

[0108] Step S340, calculating the loss according to the reconstructed ocean near-surface flow field and the label samples;

[0109] Step S350, performing parameter gradient update on the feature encoding module, the multi-frequency attention module and the feature reconstruction module according to the loss.

[0110] The feature encoding module is used to extract features from the input samples in the training data (i.e., data corresponding to multiple input variables) and fuse the feature maps between multiple input variables to achieve interactive sharing of thermodynamic and kinetic information between multiple input variables. The multi-frequency attention module is used to divide the first fused feature map into sub-feature maps of multiple frequency components and independently capture the attention features in the sub-feature map of each frequency component; finally, the feature reconstruction module is used to fuse the attention feature maps corresponding to the sub-feature maps of all frequency components and reconstruct the near-surface process of the ocean based on the fused feature maps.

[0111] The steps of this embodiment propose a multi-frequency attention mechanism, which can directly divide the feature map into feature maps corresponding to different frequency components, and enhance the attention processing of information at different frequencies respectively. This not only shortens the sequence length of each processing, but also reduces the mixing of signals of different scales, thereby improving the accuracy of reconstruction of the reconstruction model.

[0112] In some embodiments of the present application, the multi-frequency attention module includes L branch networks, and each branch network includes M multi-head self-attention modules connected in sequence; wherein L is the same as the number of frequency components; M is a positive integer greater than 1;

[0113] The first fused feature map is divided into sub-feature maps of multiple frequency components, and the attention feature map corresponding to the sub-feature map of each frequency component is independently captured, including:

[0114] Dividing the first fused feature map into a plurality of sub-feature maps of different frequency components based on wavelet transform;

[0115] The sub-feature map of the target frequency component is input into a corresponding branch network, so as to perform feature learning on the sub-feature map of the target frequency component through M sequentially connected multi-head self-attention modules, and obtain a feature map corresponding to the sub-feature map of the target frequency component output by the M-th multi-head self-attention module; the sub-feature map of the target frequency component is a sub-feature map of any frequency component among the sub-feature maps of multiple different frequency components;

[0116] The sub-feature map of the target frequency component and the feature map corresponding to the sub-feature map of the target frequency component are residually connected to obtain the attention feature map corresponding to the sub-feature map of the target frequency component.

[0117] In this embodiment, the multi-frequency attention module includes L branch networks, and each branch network includes M multi-head self-attention modules (MHSA) connected in sequence, and the input end of the first MHSA is connected to the output end of the Mth MHSA from the residual connection to solve the gradient disappearance and gradient explosion problems. The branch network is divided into sub-feature maps of different frequencies by wavelet transform decomposition.

[0118] like Figures 2 to 5 The present application embodiment provides a method for reconstructing the near-surface flow field of an ocean, and the method includes the following process:

[0119] Step S910, calculating the spatial decorrelation scale of the sea level and defining the spatial field range of the reconstructed flow field. Specifically, the following steps are included:

[0120] Step S9110, obtaining the daily average sea level height data product.

[0121] The L4-level daily average sea surface height data product from CMEMS covers 30 years of daily average data from 1993 to 2022, a total of 10,957 days, with a spatial resolution of , covering the global ocean area, so the spatial range is That is, the grid fields of each day are combined to construct a three-dimensional matrix with a scale of [longitude, latitude, time] , the symbol size is ,Right now , .

[0122] Step S9120, in the plane formed by longitude and latitude, traverse each grid one by one in the order from top to bottom and from left to right, that is, the grid taken out each time is called the target grid, and the matrix where the target grid is located is taken out The length of the spatial position is The sea surface height data series , and then calculate the latitudinal spatial decorrelation scale and the longitudinal spatial decorrelation scale of the target grid in turn.

[0123] Preferably, step S9120 specifically includes the following steps:

[0124] Step S9121, calculating the latitudinal spatial decorrelation scale.

[0125] Calculate the sea surface height data sequence of the target grid one by one The linear correlation of the sequence of each grid with the same latitude, each time a grid with the same latitude is taken out is called the first reference grid, and the calculation formula is:

[0126] (1);

[0127] (2);

[0128] (3);

[0129] in, Indicates the sea surface height data sequence of the first reference grid taken out at a certain time, Indicates the serial number, total grids at the same latitude, including the target grid itself; Indicates the length is At a certain moment in the sea surface height data series, represents the sea surface height value at a certain moment in the sequence of the target grid, Represents the sea surface height data sequence of the target grid The average of all sea surface height values; Indicates the sea surface height value at a certain moment in the sea surface height data sequence of the first reference grid, Represents the sea surface height data sequence of the first benchmark grid The average of all sea surface height values; Represents the sea surface height data sequence of the target grid The sea surface height data sequence of the first benchmark grid The first linear similarity between .

[0130] Then, we calculate Store a list of length 1440 Correspondence The position of the serial number. After the traversal is completed, the latitudinal linear similarity list of the target grid is obtained. , let the latitudinal spatial decorrelation scale threshold (i.e., the first threshold) be , record the target grid and The distance is the farthest and the linear correlation is greater than The distance of the first grid , which is the latitudinal spatial decorrelation scale of the target grid; finally, the traversed target grid , recorded in size The latitudinal spatial decorrelation scale matrix At the corresponding spatial position in .

[0131] Step S9122, calculate the longitudinal spatial decorrelation scale, and calculate the sequence of target grids one by one The linear correlation of the sequence of each grid with the same longitude, each time a grid with the same longitude is taken out is called the second reference grid, and the calculation formula is:

[0132] (4);

[0133] (5);

[0134] (6);

[0135] in, Indicates the sea surface height data sequence of the second reference grid taken out at a certain time, Indicates the serial number, total grids of the same longitude, including the target grid itself; Indicates a moment in a sea surface height data sequence with a length of 10597. Indicates the sea surface height value at a certain moment in the sequence under the second reference grid, Represents the sea surface height data sequence of the second benchmark grid The average of all sea surface height values; Represents the sea surface height data sequence of the target grid The sea surface height data series of the second benchmark grid The second linear similarity between .

[0136] Then, we calculate The length of the storage is List Correspondence The position of the sequence number. After the traversal is completed, the linear similarity list of the target grid is obtained. , let the longitudinal spatial decorrelation scale threshold (i.e., the second threshold) be , record the target grid and The distance is the farthest and the linear correlation is greater than The distance of the second grid , which is the longitudinal spatial decorrelation scale of the target grid; finally, the target grid traversed at this time , recorded in size The longitudinal spatial decorrelation scale matrix At the corresponding spatial position in .

[0137] Step S9123, after completing the traversal of all grids, the latitudinal spatial decorrelation scale matrix is ​​obtained , and the longitudinal spatial decorrelation scale matrix , calculate the mean of the two matrices and obtain the latitudinal spatial decorrelation scale (i.e. the first average value) , and the meridional spatial decorrelation scale (i.e., the second mean) , so as to determine the spatial field range of the reconstructed flow field, that is, with each target grid as the center and the latitudinal side length as =4, the longitudinal side length is The rectangular area.

[0138] Step S920, obtaining satellite remote sensing data and measured ocean near-surface flow field data from open source channels as input and label data to construct training data for the reconstruction model.

[0139] The specific steps include:

[0140] Step S9210, obtain two variables as labels, which are the sea surface zonal velocity and sea surface meridional velocity of the foreign Global Drifting Buoy Program (GDP).

[0141] The label data comes from the Global Drifting Buoy Program, a branch of the Global Ocean Observing System (GOOS) and the Global Climate Observing System (GCOS) of the National Oceanic and Atmospheric Administration (NOAA) of the United States. Each buoy is stored as a NETCDF file with a time resolution of 6 hours. First, the buoy data is preprocessed, including the following steps:

[0142] (1) Read the drifting buoy data information, including buoy number, longitude, longitude standard error, latitude, latitude standard error, time, zonal current velocity, meridional current velocity, sea surface temperature, sea surface temperature error standard deviation, date of the first high-quality buoy data, date of the last high-quality buoy data, buoy, and date when the buoy traction device was lost.

[0143] (2) Keep the data from the first high-quality buoy data date to the date before the buoy traction device was lost, and discard the data after the date when the buoy traction device was lost.

[0144] (3) Determine the missing values ​​of buoy data and discard “-999999” or “NaN” in the buoy data.

[0145] (4) To remove outliers from buoy data, three times the standard deviation of the longitude standard error, latitude standard error and sea surface temperature standard error were used as thresholds, and buoy data outside the thresholds were removed.

[0146] (5) Daily average of buoy data.

[0147] (6) Finally, sort the buoy data by time to obtain the total There are three sequences of LON, LAT and TIME, each with a length of , which records the spatial position and time of data observation at each location point.

[0148] Step S9220, 12 input variables are obtained as input, which are 4 satellite data: satellite altimeter absolute dynamic altitude, satellite sea surface temperature, sea surface meridional wind, sea surface zonal wind; 5 satellite derived data: absolute dynamic altitude zonal gradient, absolute dynamic altitude meridional gradient, satellite sea surface temperature zonal gradient, satellite sea surface temperature meridional gradient, satellite sea surface wind speed; global longitude grid field and global latitude grid field, and 1 global Coriolis force parameter. The time resolution of the above variables is consistent, all from 1993 to 2022, spanning 30 years of daily average data, a total of 10957 days, that is, each input variable has 10957 samples, with consistent coverage, covering the global ocean.

[0149] The specific description and sources of the four satellite data are as follows: The sea surface temperature data comes from the L4-level OISST satellite sea surface temperature optimal interpolation product provided by the National Centers for Environmental Information (NOAA / NCEI) of the National Oceanic and Atmospheric Administration of the United States, with a spatial resolution of approximately The temporal resolution is the average of the sky; the sea surface wind (longitudinal and latitudinal) data also comes from the L4 CCMP satellite wind vector analysis product provided by RemoteSensing Systems. The sea surface 10m longitudinal and latitudinal wind speed data are selected, and the spatial resolution is about The time resolution is 6 hours average. Therefore, in order to unify the time resolution, the data of 4 times a day are averaged to obtain the average meridian and zonal wind data. The absolute dynamic altitude data of the satellite altimeter fusion product comes from the fusion product provided by AVISO of France. The integration product integrates altimeter data from multiple satellites, and the spatial resolution is about , the time resolution is the average of the days.

[0150] The absolute dynamic altitude latitudinal gradient, absolute dynamic altitude meridional gradient, satellite sea surface temperature latitudinal gradient, and satellite sea surface temperature meridional gradient are calculated using first-order central differences. No. The latitudinal gradient of the grid points ( ) and the meridional gradient ( The first-order central difference formula of ) is as follows:

[0151] (7);

[0152] (8);

[0153] in, and Respectively The grid spacing of grid cells.

[0154] Sea surface wind speed The calculation formula is:

[0155] (9);

[0156] in, and They represent the zonal wind speed and the meridional wind speed on the sea surface, respectively.

[0157] Coriolis parameters is a parameter related to the rotation of the earth, it is the latitude Function:

[0158] (10);

[0159] In the above formula is the angular velocity of the Earth's rotation.

[0160] Step S9230, spatially pair the data of the input variables obtained in step S9220, and use the bilinear interpolation method to unify all variables to a spatial resolution of , and then merge the two-dimensional grids of all variables in chronological order so that each variable is a matrix of size [latitude, longitude, time]. Since the size of each input variable is unified, the size is consistent.

[0161] Step S9240, spatially match the input variables with unified spatial resolution obtained in step S9230 with the label data obtained in step S9210. Since the label data is scattered data, first determine the horizontal spatial position of each location point in the three-dimensional input variable matrix based on the LON and LAT sequences of the location point, and determine the vertical moment position of each location point in the three-dimensional input variable matrix based on the TIME sequence of the location point.

[0162] Step S9250, traverse the position points determined in step S9230 one by one, and according to the spatial position of each position point, take the grid where the position point is located as the center and the latitudinal side length is =4 and the longitudinal length is The rectangular area is the window, and the data of the corresponding windows of multiple input variables are intercepted and merged to obtain a window of size A single input sample of , where "1" represents one sample and 12 represents 12 variables, is the width and length; in addition, the sea surface meridional velocity and the sea surface zonal velocity at this point are combined into a sequence of length 2 as a single label sample, and an input sample is combined with a label sample to obtain an "input-label" sample pair; after the traversal is completed, input-label pairs.

[0163] Step S9260, randomly select one of the The input-label sample pairs are the training set , "input-label" sample pairs are the validation set , The input-label sample pairs are the test set .

[0164] Step S9270, the training set obtained from step S9250 or validation set or test set Randomly pick samples, get an input batch , where the input variables are normalized to interval, and obtain a batch of data after normalization , the normalized formula expression is as follows:

[0165] (11);

[0166] in, Represents a variable in a batch A data sample set, For the training set The minimum value of a variable in For the training set The maximum value of a variable in the batch, the data of a variable after normalization is recorded as .

[0167] Step S930, build a deep learning neural network FFRN, the expression of FFRN is as follows:

[0168] (12);

[0169] in, Represents a deep learning-based neural network FFRN, with normalized multivariate input batches as input , the output is the reconstructed flow velocity data , Indicates the first The weights and biases of the layers.

[0170] The specific steps are as follows:

[0171] Step S9310: construct a feature encoding module (FF) to perform feature encoding and fusion on all input variables to obtain preliminary features with reconstruction capabilities. Each input variable contains 1 convolution layer with 1×1 convolution kernel, 1 layer The activation layer combines the features of all input variables after the convolution layer and the activation layer in the channel dimension, and then passes through 1×1 convolution kernel, 1 layer The activation layer aims to learn the feature representation of all input variables in high-dimensional space and improve the expression ability of FFRN; its output feature map is , the expression is:

[0172] (13);

[0173] in, Represents the output feature map of the feature encoding module (i.e., the first fusion feature), with a size of ; represents the mapping function of the feature encoding module, represents the relu activation function, represents the mapping function of the convolution layer with a 1×1 convolution kernel for each variable separately, Represents the mapping function of the convolution layer with a 1×1 convolution kernel used solely for the merged variable features, Indicates merging features in the channel dimension; arrive The data for an input variable in the input batch.

[0174] Step S9320: construct a multi-frequency attention module (MFA). The expression of the multi-frequency attention module is:

[0175] (14);

[0176] in, represents the output feature map of the multi-frequency attention module, Represents the mapping function of the multi-frequency attention module.

[0177] Preferably, step S9320 specifically includes the following steps:

[0178] Step S9321: Output feature map of feature encoding module Perform a wavelet transform decomposition to obtain four components: low-low frequency, high-low frequency, low-high frequency, and high-high frequency. The wavelet transform decomposition is solved by numerical approximation, and the formula is:

[0179] (15);

[0180] (16);

[0181] (17);

[0182] (18);

[0183] (19);

[0184] (20);

[0185] (twenty one);

[0186] (twenty two);

[0187] (twenty three);

[0188] (twenty four);

[0189] Among them, since the input variable corresponds to a 4-dimensional tensor, It means taking out the features consisting of even-numbered rows in the longitude direction in the feature map. It means taking out the features consisting of odd rows in the longitude direction in the feature map. It means taking out the features consisting of even columns in the latitude direction in the feature map. It means taking out the features consisting of odd columns in the latitude direction in the feature map; is the low-frequency feature map, is the high and low frequency feature map, is the low-frequency feature map, is a high-frequency feature map, and the sizes of the four are .

[0190] The multi-frequency attention module consists of multiple branch networks, and the number of branch networks is consistent with the number of frequency components.

[0191] Step S9322, respectively , , and Construct identical but independent attention-capturing branch networks, each consisting of sequential connections indivual Multi-head self-attention module.

[0192] First, flatten the input feature map and swap the last two dimensions to get a size of Three-dimensional features; then, after the layer normalization layer layernorm, along In the dimension, for each length The sequence is normalized; then, it passes through three identical fully connected layers MLP to obtain three independent tensors , the second dimension of these three tensors is expanded times, size is ; Then, Divide into segment, becomes a 4-dimensional size of , , Tensor; Next, and Perform matrix multiplication and reduce each value of the calculation result to the original value times, then, along Softmax activation is performed on the dimension and Perform matrix multiplication and restore the size to ; Finally, Segments merged, restored to size After that, it passes through a fully connected layer MLP and compresses the second dimension to times, the resulting size is The multi-head attention sub-network implementation process is expressed as follows to process For example:

[0193] (25);

[0194] (26);

[0195] (27);

[0196] (28);

[0197] (29);

[0198] (30);

[0199] in, It means flattening a two-dimensional matrix into a one-dimensional vector. Representation layer normalization layer, is the fully connected layer, Indicates that , , Divide into part, Indicates that the preceding The feature maps obtained by processing the segments are concatenated in the channel dimension. Represents the mapping function of the MHSA multi-head self-attention module, Indicates The specific expression of the Softmax activation function for the obtained feature map is:

[0200] (31);

[0201] in, represents the input features of the activation function, Represents the value of each grid on this feature.

[0202] After being processed by the subsequent three multi-head self-attention modules, a feature processed by a branch network is obtained, and the size remains constant.

[0203] Finally, in order to reduce information loss and reduce training difficulty, the feature map of the input branch is residually connected with the feature map processed by M MHSA multi-head self-attention modules, that is, the feature map of the input branch is added to the feature map processed by M MHSA multi-head self-attention modules.

[0204] Step S9330, construct a feature reconstruction module (Feature Reconstruction, FR), the expression of the feature reconstruction module is:

[0205] (32);

[0206] The output of the feature reconstruction module is a reconstructed velocity sequence with a length of 2, which records the values ​​of the sea surface meridional velocity and the sea surface zonal velocity. Represents the mapping function of the feature reconstruction module.

[0207] Preferably, step S9330 specifically includes the following steps:

[0208] Step S9331, obtain the features after each branch network processing , , and Then, the results of the four residual connections are combined in the last dimension, with a size of [16,8,256]. Then, they are normalized once and input into a fully connected layer MLP, the last dimension is compressed, the feature dimension is fused, and a fused feature map with a size of [16,8,2] (i.e., the second fused feature map) is obtained. Finally, after passing through a fully connected layer MLP again, the second dimension is compressed to 1, and the size is [16,1,2]. The redundant dimensions, i.e., the dimensions with a length of 1, are discarded, and the result remains unchanged, and the reconstructed flow velocity is obtained, with a size of [16,2].

[0209] Step S9340, the feature encoding module, four multi-frequency attention modules, and one feature reconstruction module are connected in sequence to obtain a neural network FFRN based on deep learning, and the expression includes:

[0210] (33);

[0211] Preferably, step S9340 specifically includes the following steps:

[0212] Connect one feature encoding module, four multi-frequency attention modules, and one feature reconstruction module in sequence to obtain the deep learning neural network FFRN, which is expressed as follows:

[0213] (34);

[0214] in, It is the reconstructed ocean surface velocity data with a size of [16,2].

[0215] Step S940, using the data set obtained in step S910, according to the spatial range defined in step S920, the network FFRN obtained in step S930 is alternately trained and verified.

[0216] The specific steps are as follows:

[0217] Step S9410, select Huber loss as the loss function for training the neural network FFRN. The expression of Huber loss is as follows:

[0218] (35);

[0219] in, Represents the reconstructed ocean surface velocity data of label data, size and output The sizes are consistent, both , Each grid representing the reconstructed ocean surface velocity data and label data, The value set to 0.1 is a constant hyperparameter that determines the segmentation position of the Huber loss.

[0220] Step S9420, based on the training set With validation set Input the neural network FFRN obtained in step S930 in batches for training and fitting The parameters of each layer in .

[0221] Preferably, step S9420 specifically includes the following steps:

[0222] Step S9421, set the basic hyperparameters for training the neural network FFRN under the open source deep learning framework, and then use the training set The data is input into the neural network FFRN obtained in step S920 in batches for training. Therefore, the training set contain samples, so they can be divided into batches, is the number of random samples taken from the training set ( ), the output results of each batch will calculate the sequential loss , and uses the back propagation optimizer AdamW in the open source deep learning framework to back propagate the gradient of the loss and optimize the parameters , when all batches are trained for one round, it is called an epoch, and the total training epochs, at the end of each epoch, the validation set will be input The data is used to calculate the loss of a validation set, but back propagation is not performed. The purpose is to observe the difference between the loss of the validation set and the loss of the training set, and to judge the parameter fitting effect and generalization performance of the neural network FFRN.

[0223] Step S9422: After training, keep the parameters , and obtain the trained deep learning neural network FFRN as the reconstruction model.

[0224] Step S950, using the trained reconstruction model to reconstruct the ocean near-surface flow field, that is, to obtain high-precision ocean near-surface flow field data.

[0225] The specific steps are as follows:

[0226] Step S9510: Test set The data was normalized using the MinMax method to interval.

[0227] In step S9520, the test set data obtained by the normalization step is applied to the reconstructed model trained in batches, and then the deviation-corrected data is obtained.

[0228] This method has at least the following beneficial effects:

[0229] (1) This method introduces the spatial decorrelation scale of sea surface height as a physical constraint, which improves the interpretability of the neural network based on deep learning. Compared with the purely data-driven deep learning model, considering the actual physical conditions can improve the generalization ability of the model.

[0230] (2) The multi-frequency attention mechanism based on wavelet transform proposed in this method can process information at four frequencies separately, reduce the confusion of signals of different scales, reduce the difficulty of training, and at the same time shorten the length of the sequence processed each time, reducing the computational load of global attention.

[0231] (3) This method integrates multiple satellite remote sensing variables to reconstruct the flow field, overcoming the limitations of the traditional use of variables such as sea surface height or sea surface temperature alone. It integrates the thermodynamic and kinetic relationships between multiple variables, improves the reconstruction accuracy, and enriches the technical means of flow velocity reconstruction.

[0232] (4) This method can realize the reconstruction of near-surface flow velocity in global oceans, has global applicability and global generalization, and the trained model is easy to deploy, has low operating costs, and is easy to promote and apply.

[0233] like Figure 6 One embodiment of the present application provides a system for reconstructing a near-surface flow field of an ocean, the system comprising:

[0234] The data acquisition unit 1100 is used to obtain the sky average sea surface height data product;

[0235] The latitudinal decorrelation scale acquisition unit 1200 is used to determine a first linear similarity between a sea surface height data sequence corresponding to a target grid and a sea surface height data sequence corresponding to a first benchmark grid; the target grid is any one of a plurality of grids of a sky-average sea surface height data product, the first benchmark grid is different from the target grid, and the first benchmark grid is any one of a plurality of grids at the same latitude as the target grid, the sea surface height data sequence corresponding to the target grid is a corresponding sea surface height data sequence of the target grid in the sky-average sea surface height data product, and the sea surface height data sequence corresponding to the first benchmark grid is a corresponding sea surface height data sequence of the first benchmark grid in the sky-average sea surface height data product;

[0236] And, used to determine a first reference grid whose first linear similarity is less than a preset first threshold and is closest to the target grid, and determine a first distance between the target grid and the first reference grid that is closest to the target grid;

[0237] and, for calculating a first average value of first distances corresponding to all target grids;

[0238] The longitude decorrelation scale acquisition unit 1300 is used to determine the second linear similarity between the sea surface height data sequence corresponding to the target grid and the sea surface height data sequence corresponding to the second reference grid; the second reference grid is any one of the multiple grids that is at the same longitude as the target grid, and the sea surface height data sequence corresponding to the second reference grid is the corresponding sea surface height data sequence of the second reference grid in the sky average sea surface height data product;

[0239] And, used to determine a second reference grid whose second linear similarity is less than a preset second threshold and is closest to the target grid, and determine a second distance between the target grid and the second reference grid that is closest to the target grid;

[0240] and, for calculating a second average value of second distances corresponding to all target grids;

[0241] The model training unit 1400 is used to obtain satellite remote sensing data and measured ocean near-surface flow field data of a plurality of preset location points;

[0242] And, it is used to take the grid where the location point is located as the center, take the rectangular area of ​​the first average value and the second average value as the window, intercept the data of the input variable of the location point of the window size from the satellite remote sensing data as the input sample, and take the measured ocean near-surface flow field data of the location point as the label sample;

[0243] and, using multiple groups of input samples and label samples as training data, training the neural network according to the training data, and obtaining a reconstructed model after the training is completed;

[0244] The model application unit 1500 is used to input the target ocean near-surface flow field data to be reconstructed into the reconstruction model to obtain the target ocean near-surface flow field reconstructed by the reconstruction model.

[0245] It should be noted that the ocean near-surface flow field reconstruction system provided in this embodiment and the above-mentioned ocean near-surface flow field reconstruction method are based on the same inventive concept. Therefore, the relevant contents of the above-mentioned ocean near-surface flow field reconstruction method are also applicable to the contents of the ocean near-surface flow field reconstruction system. Therefore, they will not be repeated here.

[0246] like Figure 7 The embodiment of the present application further provides an electronic device, the electronic device comprising:

[0247] at least one memory;

[0248] at least one processor;

[0249] at least one program;

[0250] The program is stored in the memory, and the processor executes at least one program to implement the above-mentioned ocean near-surface flow field reconstruction method implemented in the present disclosure.

[0251] The electronic device may be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (PDA), a vehicle-mounted computer, etc.

[0252] The electronic device according to the embodiment of the present application is described in detail below.

[0253] The processor 1600 may be implemented by a general-purpose central processing unit (CPU), a microprocessor, an 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 the present invention.

[0254] The memory 1700 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1700 can store an operating system and other application programs. When the technical solution provided in the embodiments of this specification is implemented by software or firmware, the relevant program code is stored in the memory 1700, and the processor 1600 calls and executes the ocean near-surface flow field reconstruction method of the embodiment of the present invention.

[0255] Input / output interface 1800, used to implement information input and output;

[0256] Communication interface 1900, used to realize communication interaction between the device and other devices, which can be realized through wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);

[0257] Bus 2000 , which transmits information between various components of the device (e.g., processor 1600 , memory 1700 , input / output interface 1800 , and communication interface 1900 );

[0258] The processor 1600 , the memory 1700 , the input / output interface 1800 , and the communication interface 1900 are connected to each other in communication within the device via the bus 2000 .

[0259] An embodiment of the present invention further provides a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the above-mentioned ocean near-surface flow field reconstruction method.

[0260] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0261] The embodiments described in the present invention are intended to more clearly illustrate the technical solutions of the embodiments of the present invention and do not constitute a limitation on the technical solutions provided by the embodiments of the present invention. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present invention are also applicable to similar technical problems.

[0262] Those skilled in the art will appreciate that the technical solutions shown in the figures do not limit the embodiments of the present invention and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0263] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0264] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0265] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0266] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0267] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0268] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0269] If the integrated unit is implemented in the form of 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 the present application is essentially 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, and the computer software product is stored in a storage medium, including multiple instructions to enable an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of each embodiment method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store programs.

[0270] The above is a specific description of the preferred implementation of the embodiments of the present application, but the embodiments of the present application are not limited to the above-mentioned implementation methods. Technical personnel familiar with the field can also make various equivalent modifications or substitutions without violating the spirit of the embodiments of the present application. These equivalent modifications or substitutions are all included in the scope defined by the claims of the embodiments of the present application.

Claims

1. A method for reconstructing the near-surface flow field of an ocean, characterized in that: The method comprises: Obtain the average sea surface height data product; Determine a first linear similarity between a sea surface height data sequence corresponding to a target grid and a sea surface height data sequence corresponding to a first benchmark grid; the target grid is any one of a plurality of grids of the daily average sea surface height data product, the first benchmark grid is different from the target grid, and the first benchmark grid is any one of the plurality of grids at the same latitude as the target grid, the sea surface height data sequence corresponding to the target grid is a corresponding sea surface height data sequence of the target grid in the daily average sea surface height data product, and the sea surface height data sequence corresponding to the first benchmark grid is a corresponding sea surface height data sequence of the first benchmark grid in the daily average sea surface height data product; Determine a first reference grid whose first linear similarity is less than a preset first threshold and is closest to the target grid, and determine a first distance between the target grid and the first reference grid that is closest to the target grid; Calculating a first average value of the first distances corresponding to all the target grids; Determine a second linear similarity between a sea surface height data sequence corresponding to the target grid and a sea surface height data sequence corresponding to a second reference grid; the second reference grid is any one of the plurality of grids that is at the same longitude as the target grid, and the sea surface height data sequence corresponding to the second reference grid is a corresponding sea surface height data sequence of the second reference grid in the sky average sea surface height data product; Determine a second reference grid whose second linear similarity is less than a preset second threshold and is closest to the target grid, and determine a second distance between the target grid and the second reference grid that is closest to the target grid; Calculating a second average value of the second distances corresponding to all the target grids; Obtain satellite remote sensing data and measured ocean near-surface flow field data for multiple preset locations; Taking the grid where the position point is located as the center and the rectangular area of ​​the first average value and the second average value as the window, intercepting the data of the input variable of the position point with a size of the window from the satellite remote sensing data as the input sample, and taking the measured ocean near-surface flow field data of the position point as the label sample; Using the multiple groups of input samples and the label samples as training data, training the neural network according to the training data, and obtaining a reconstructed model after the training is completed; The target ocean near-surface flow field data to be reconstructed is input into the reconstruction model to obtain the target ocean near-surface flow field reconstructed by the reconstruction model.

2. The method for reconstructing the near-surface flow field of the ocean according to claim 1, characterized in that: The calculation formula for determining the first linear similarity is: ； ； ； in, is the sea surface height data sequence corresponding to the target grid The sea surface height data sequence corresponding to the first benchmark grid The first linear similarity between is the sea surface height data series exist The sea surface height data at the moment, is the sea surface height data series exist The sea surface height data at the moment, is the time span of the daily mean sea surface height data product.

3. The method for reconstructing the near-surface flow field of the ocean according to claim 1, characterized in that: The calculation formula for determining the second linear similarity is: ； ； ； in, is the sea surface height data sequence corresponding to the target grid The sea surface height data sequence corresponding to the second benchmark grid The second linear similarity between is the sea surface height data series exist The sea surface height data at the moment, is the sea surface height data series exist The sea surface height data at the moment, is the time span of the daily mean sea surface height data product.

4. The method for reconstructing the near-surface flow field of the ocean according to claim 1, characterized in that: The input variables include at least two of absolute dynamic altitude, absolute dynamic altitude latitudinal gradient, absolute dynamic altitude meridional gradient, satellite sea surface temperature, sea surface temperature meridional gradient, sea surface temperature latitudinal gradient, sea surface wind speed, sea surface meridional wind speed, sea surface latitudinal wind speed, global longitude grid field, global latitude grid field and Coriolis parameters; The measured ocean near-surface flow field data include measured sea surface latitudinal flow velocity and sea surface meridional flow velocity.

5. The method for reconstructing the near-surface flow field of the ocean according to claim 4, characterized in that: The neural network includes a feature encoding module, a multi-frequency attention module and a feature reconstruction module; The neural network is trained according to the training data, and a reconstructed model is obtained after the training is completed, including: The neural network is trained multiple times according to the training data until the neural network converges to obtain a reconstructed model; each training process includes: Inputting the training data into the feature encoding module to perform feature extraction on the data of each input variable in the training data to obtain a feature map corresponding to the data of each input variable, and fusing the feature maps corresponding to the data of all the input variables to obtain a first fused feature map; Inputting the first fused feature map into the multi-frequency attention module to divide the first fused feature map into sub-feature maps of multiple frequency components, and independently capturing the attention feature map corresponding to the sub-feature map of each frequency component; Inputting the attention feature maps corresponding to the sub-feature maps of all frequency components into the feature reconstruction module, so as to obtain a second fused feature map by fusing the attention feature maps corresponding to the sub-feature maps of all frequency components, and obtaining a reconstructed ocean near-surface flow field according to the second fused feature map; Calculating the loss based on the reconstructed ocean surface flow field and the label samples; Parameter gradients of the feature encoding module, the multi-frequency attention module and the feature reconstruction module are updated according to the loss.

6. The method for reconstructing the ocean surface flow field according to claim 5, characterized in that: The formula for calculating the first fusion feature map is: ； in, to The data of the first input variable to The data of the input variables, is the number of input variables, , The mapping function of the convolution layer is 1×1 convolution kernel. for The mapping function of the activation function, is the mapping function of the fusion operation, It is the first fused feature map.

7. The method for reconstructing the ocean surface flow field according to claim 5, characterized in that: The multi-frequency attention module includes L branch networks, and each of the branch networks includes M multi-head self-attention modules connected in sequence; wherein L is the same as the number of the frequency components; and M is a positive integer greater than 1; The first fused feature map is divided into sub-feature maps of multiple frequency components, and an attention feature map corresponding to the sub-feature map of each frequency component is independently captured, including: Dividing the first fused feature map into a plurality of sub-feature maps of different frequency components based on wavelet transform; Inputting the sub-feature map of the target frequency component into a corresponding one of the branch networks, so as to perform feature learning on the sub-feature map of the target frequency component through M sequentially connected multi-head self-attention modules, and obtaining a feature map corresponding to the sub-feature map of the target frequency component output by the M-th multi-head self-attention module; the sub-feature map of the target frequency component is a sub-feature map of any one frequency component among the sub-feature maps of multiple different frequency components; The sub-feature map of the target frequency component and the feature map corresponding to the sub-feature map of the target frequency component are residually connected to obtain an attention feature map corresponding to the sub-feature map of the target frequency component.

8. An ocean surface flow field reconstruction system, characterized in that: The system comprises: A data acquisition unit, used to obtain the sky average sea surface height data product; a latitudinal decorrelation scale acquisition unit, configured to determine a first linear similarity between a sea surface height data sequence corresponding to a target grid and a sea surface height data sequence corresponding to a first benchmark grid; the target grid is any one of a plurality of grids of the sky-average sea surface height data product, the first benchmark grid is different from the target grid, and the first benchmark grid is any one of the plurality of grids at the same latitude as the target grid, the sea surface height data sequence corresponding to the target grid is the corresponding sea surface height data sequence of the target grid in the sky-average sea surface height data product, and the sea surface height data sequence corresponding to the first benchmark grid is the corresponding sea surface height data sequence of the first benchmark grid in the sky-average sea surface height data product; and, used to determine a first reference grid whose first linear similarity is less than a preset first threshold and is closest to the target grid, and determine a first distance between the target grid and the first reference grid that is closest to the target grid; and, used for calculating a first average value of the first distances corresponding to all the target grids; a longitudinal decorrelation scale acquisition unit, configured to determine a second linear similarity between a sea surface height data sequence corresponding to the target grid and a sea surface height data sequence corresponding to a second reference grid; the second reference grid is any one of the plurality of grids that is at the same longitude as the target grid, and the sea surface height data sequence corresponding to the second reference grid is a corresponding sea surface height data sequence of the second reference grid in the sky average sea surface height data product; And, used to determine a second reference grid whose second linear similarity is less than a preset second threshold and is closest to the target grid, and determine a second distance between the target grid and the second reference grid that is closest to the target grid; and, used for calculating a second average value of the second distances corresponding to all the target grids; A model training unit, used to obtain satellite remote sensing data of multiple preset locations and measured ocean near-surface flow field data; And, using the grid where the position point is located as the center and the rectangular area of ​​the first average value and the second average value as a window, intercepting the data of the input variable of the position point with a size of the window from the satellite remote sensing data as input samples, and using the measured ocean near-surface flow field data of the position point as label samples; And, used to use the multiple groups of the input samples and the label samples as training data, train the neural network according to the training data, and obtain a reconstructed model after the training is completed; The model application unit is used to input the target ocean near-surface flow field data to be reconstructed into the reconstruction model to obtain the target ocean near-surface flow field reconstructed by the reconstruction model.

9. An electronic device, characterized in that: include: at least one control processor and a memory for communicatively coupling with the at least one control processor; The memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute the ocean near-surface flow field reconstruction method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the ocean near-surface flow field reconstruction method according to any one of claims 1 to 7.

Images