CN117350171A - Mesoscale vortex three-dimensional subsurface structure inversion method and system based on double-flow model - Google Patents

Abstract

The present invention proposes a method and system for inverting the three-dimensional subsurface structure of mesoscale vortices based on a two-flow model, which involves the intersection of deep learning and ocean inversion. The mesoscale vortex sea surface information to be inverted is collected through satellites; the mesoscale vortex sea is The surface information is input into the trained two-flow model, and the temperature results at different depths of the mesoscale eddy are inverted to obtain the subsurface temperature profile of the mesoscale eddy. Among them, the two-flow model introduces a triplet attention mechanism and adopts a three-branch structure. Fusion of channel attention and spatial attention for cross-dimensional interaction; the present invention uses a dual-flow model to achieve mesoscale vortex subsurface structure inversion, explores the data correlation between sea surface parameters, establishes relationship models between different parameters and subsurface temperature, and fuses Multi-source information feature relationships realize feature fusion, effectively integrate multi-source data, and improve the inversion effect.

Description

Mesoscale eddy three-dimensional subsurface structure inversion method and system based on two-flow model

Technical field

The invention belongs to the intersection of deep learning and ocean inversion, and in particular relates to a mesoscale eddy three-dimensional subsurface structure inversion method and system based on a two-flow model.

Background technique

The statements in this section merely provide background technical information related to the present invention and do not necessarily constitute prior art.

As a common ocean phenomenon, mesoscale eddies widely exist in various sea areas. Its internal structure is a closed vortex, which can affect the vertical structure to a depth of several thousand meters. At the same time, mesoscale eddies contain huge energy and drive the ocean. The transmission of heat, salinity, and nutrients has a non-negligible impact on ocean chemistry and biological environment; the kinetic energy of mesoscale eddies even accounts for 90% of the kinetic energy of large and medium oceans in the entire ocean, and its energy will cause chaos in the underwater three-dimensional sound field. , disrupting underwater communications and detection, and affecting the safety of submarines, fishing boats, etc. Therefore, studying the three-dimensional structure of mesoscale vortices will effectively ensure the navigation safety of underwater submarines and provide strong support for marine environment monitoring, national defense and military, etc.

With the development of satellite remote sensing technology, high-resolution satellite remote sensing technology has been widely used in meteorology and hydrology. At present, the extraction of eddy characteristics on the sea surface mainly relies on high-resolution satellite remote sensing observation data, but this method can only obtain eddies. Surface data cannot directly detect information about the interior of the ocean. Compared with remote sensing observations, although ocean-based in-situ observation data (such as Argo, CTD data, etc.) can detect the profile information of the ocean subsurface, they have problems such as uneven and discontinuous spatiotemporal distribution and low spatial resolution, and cannot meet the requirements of Requirements for the internal dynamic processes of mesoscale vortices; therefore, existing technology cannot use high-resolution satellite data to reflect the internal structural characteristics of the subsurface layer, hindering further research on the three-dimensional structure of mesoscale vortices.

Contents of the invention

In order to overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a mesoscale eddy three-dimensional subsurface structure inversion method and system based on a two-flow model, using sea surface information as input parameters and subsurface temperature data as labels to construct three-dimensional three-dimensional data samples. The library uses a dual-flow model to achieve mesoscale eddy subsurface structure inversion, explores the data correlation between sea surface parameters, establishes relationship models between different parameters and subsurface temperature, integrates multi-source information feature relationships, realizes feature fusion, and effectively integrates Multi-source data improves the inversion effect.

To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

The first aspect of the present invention provides a mesoscale eddy three-dimensional subsurface structure inversion method based on a two-flow model.

The mesoscale eddy three-dimensional subsurface structure inversion method based on the two-flow model includes:

Collect mesoscale eddy sea surface information to be retrieved through satellites, including sea surface temperature and sea surface height;

Input the mesoscale eddy sea surface information into the trained two-flow model, invert the temperature results at different depths of the mesoscale eddy, and obtain the mesoscale eddy subsurface temperature profile;

Among them, the dual-stream model introduces a triplet attention mechanism, and uses a three-branch structure to fuse channel attention and spatial attention for cross-dimensional interaction. The first branch establishes the feature interaction between the H dimension and the W dimension, and the second branch establishes the C The feature interaction between dimension and W dimension, and the third branch establishes the feature interaction between H dimension and C dimension.

Furthermore, the two-stream model adopts an encoding-decoding structure. The encoding stage is used for feature extraction of sea surface temperature and sea surface height, and the decoding stage is used for feature fusion.

Further, in the encoding stage, sea surface temperature and sea surface height are used as inputs to explore the relationship between sea surface temperature, sea surface height and subsurface temperature respectively; while performing feature extraction on sea surface height and sea surface temperature, consider Mesoscale eddies have complex nonlinear characteristics, and there is an inevitable connection between their sea surface height information and sea surface temperature information. In order to integrate the relationship characteristics of SSH and SST, a data feature fusion network is constructed, and jump connections are used to reduce the risk of over-fitting;

In the decoding stage, while fusing the feature output of each layer in the encoding stage, deconvolution is used to restore the resolution.

Further, the first branch introduces the identity residual branch structure, first passes through Z-Pool, and integrates the average pooling and maximum pooling features to achieve dimension scaling;

The second branch rotates the input feature map tensor along the W axis to achieve feature interaction in the C dimension and W dimension;

The third branch rotates the input feature map tensor along the H axis to achieve feature interaction in the H and C dimensions.

Further, the two-flow model is trained using the constructed mesoscale eddy subsurface structure inversion sample library as a training data set;

Among them, the mesoscale eddy time and the sea surface information corresponding to the vortex center coordinate position are used as input parameters, and the mesoscale eddy time and the subsurface temperature data corresponding to the vortex center coordinate position are used as labels to construct the mesoscale eddy subsurface structure. Inversion sample library.

Furthermore, the loss function of the two-stream model adopts MAE loss, which represents the sum of absolute differences between the true value and the predicted value. The formula is:

in, is the predicted value,/> is the real value, n is the number of a set of temperature fields.

Further, the evaluation criteria of the two-stream model are compared using R2, MAE and explained_variance_score. The R2 is used to measure the degree of fit to the predicted data, and the formula is expressed as:

Among them, f is the estimated value, y is the real value, is the average of the observed data;

The explained_variance_score is the explained variance, which is used to measure the degree to which the model explains the fluctuations in the data set and calculate the explainable variance between the true value and the predicted value. The formula is as follows:

Among them, var is the variance, y is the true value, is the predicted value.

A second aspect of the present invention provides a mesoscale eddy three-dimensional subsurface structure inversion system based on a two-flow model.

The mesoscale eddy three-dimensional subsurface structure inversion system based on the two-flow model includes a data acquisition module and a temperature inversion module:

The data collection module is configured to: collect mesoscale eddy sea surface information to be inverted through satellites, including sea surface temperature and sea surface height;

The temperature inversion module is configured to: input the mesoscale eddy sea surface information into the trained two-flow model, invert the temperature results at different depths of the mesoscale eddy, and obtain the mesoscale eddy subsurface temperature profile;

Among them, the dual-stream model introduces a triplet attention mechanism, and uses a three-branch structure to fuse channel attention and spatial attention for cross-dimensional interaction. The first branch establishes the feature interaction between the H dimension and the W dimension, and the second branch establishes the C The feature interaction between dimension and W dimension, and the third branch establishes the feature interaction between H dimension and C dimension.

A third aspect of the present invention provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the mesoscale vortex three-dimensional subsurface structure inversion method based on the two-flow model described in the first aspect of the present invention is implemented. steps in.

A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the process based on the first aspect of the present invention is implemented. Steps in the mesoscale eddy three-dimensional subsurface structure inversion method for a two-flow model.

One or more of the above technical solutions have the following beneficial effects:

This invention uses a dual-stream neural network structure to discover data correlations between sea surface parameters, establish relationship models between different parameters and subsurface temperatures, integrate multi-source information feature relationships, achieve feature fusion, effectively integrate multi-source data, and improve response performance. acting effect.

The present invention introduces a Triplet attention mechanism and uses a three-branch structure to fuse channel attention and spatial attention, which is more conducive to realizing cross-feature dimension interaction, improving the interaction ability of mesoscale vortex space features, and improving the inversion accuracy.

The mesoscale vortex three-dimensional subsurface structure inversion scheme developed in this invention is an exploration of the ocean subsurface field, which is more conducive to improving the work efficiency of experts and researchers in the ocean field and is practical.

Additional advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.

Description of drawings

The description and drawings that constitute a part of the present invention are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

Figure 1 is a method flow chart of the first embodiment;

Figure 2 is a flow chart for obtaining the temperature profile in the first embodiment;

Figure 3 is a structural diagram of the dual-flow model of the first embodiment;

Figure 4 is a structural diagram of Triplet Attention in the first embodiment.

Detailed ways

It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meanings commonly understood by one of ordinary skill in the art to which this application belongs.

It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular forms are also intended to include the plural forms unless the context clearly indicates otherwise. Furthermore, it will be understood that when the terms "comprises" and/or "includes" are used in this specification, they indicate There are features, steps, operations, means, components and/or combinations thereof.

Embodiment 1

An embodiment of the present disclosure provides a mesoscale eddy three-dimensional subsurface structure inversion method based on a two-flow model, which includes the following steps:

Step S1: Collect mesoscale eddy sea surface information to be retrieved through satellites, including sea surface temperature and sea surface height;

Step S2: Input the mesoscale eddy sea surface information into the trained two-flow model, invert the temperature results of the mesoscale eddy at different depths, and obtain the mesoscale eddy subsurface temperature profile;

Among them, the dual-stream model introduces a triplet attention mechanism, and uses a three-branch structure to fuse channel attention and spatial attention for cross-dimensional interaction. The first branch establishes the feature interaction between the H dimension and the W dimension, and the second branch establishes the C The feature interaction between dimension and W dimension, and the third branch establishes the feature interaction between H dimension and C dimension.

As an example, from the perspective of model training, the specific implementation of the mesoscale eddy three-dimensional subsurface structure inversion method based on the two-flow model is explained. The specific process is shown in Figure 1:

Step 1. Construct a sample library for mesoscale eddy surface structure inversion

In this step, data samples are obtained and sea surface data and subsurface data are fused. The sea surface data comes from the Meta3.2dt data set on the AVISO official website. The spatial resolution is 0.25° × 0.25°, and the temporal resolution is the daily average. A certain area (0-30°N, 105-130°E) is selected to focus on acquisition. Key information such as the longitude and latitude coordinates of the vortex center, vortex type and time; the sea surface information comes from the Copernicus Marine Environment Monitoring Service (CMEMS). The selected data product is: Global OceanPhysics Reanalysis, the data ID is GLOBAL_MULTIYEAR_PHY_001_030, and its spatial and temporal resolution The rates are 1/12° × 1/12° and the daily average respectively, and the subsurface temperature profile data are also derived from this data.

Perform data processing on the acquired data. The process is shown in Figure 2. The vortex center position coordinates and time information are obtained from AVISO's mesoscale eddy data, and a cyclone and anticyclone sample library is established. This information is used to correspond to the Copernican Ocean. In the surface information, data alignment is achieved, and a 4×4 grid is selected to obtain the sea surface height, sea surface temperature data and subsurface temperature profile data corresponding to the grid, and a mesoscale eddy surface structure inversion sample library is constructed.

Step 2: Data preprocessing to achieve unified resolution and missing value processing

Perform data preprocessing on the constructed sample library, use linear interpolation method to unify the resolution to 0.25° × 0.25°, and fill in missing values with zeros.

Step 3. Input the sample library into the dual-stream model

The processed sample library is input into the dual-stream neural network structure for model training and verification. The model parameters are updated and optimized to output the model. The sea surface parameter information in the test set is input into the output model for inference, and the three-dimensional temperature inversion of the mesoscale vorticity table is obtained. result.

Specifically, the dual-stream model adopts an encoding-decoding structure. The model structure is shown in Figure 3. The encoding stage is used for the feature extraction module of sea surface height SSH and sea surface temperature SST, and the decoding stage implements feature fusion.

In the encoding stage, the dual-stream data SST and SSH are first input to explore the relationship between sea surface temperature, sea surface height and subsurface temperature respectively. While extracting features of sea surface height and sea surface temperature, considering that mesoscale eddies have complex nonlinear characteristics, there is an inevitable connection between their sea surface height information and sea surface temperature information. In order to integrate the relationship features of SSH and SST, a Data feature fusion network, while using skip connections to reduce the risk of overfitting. In this process, SSH and SST are input into two branches of the neural network as dual streams, and are modeled with the subsurface temperature field respectively. Feature extraction is performed between the same branch of the neural network, and feature interaction is also performed between different branches to achieve data feature fusion.

In the encoding stage, the feature map is made smaller by deepening the network, and data information is extracted at different resolutions. In the decoding stage, while fusing the feature output of each layer in the encoding stage, deconvolution is used to restore the resolution; Triplet attention is introduced in the model. The mechanism adopts a three-branch structure to integrate channel attention and spatial attention to achieve cross-dimensional interaction.

This embodiment uses Triplet attention to achieve cross-dimensional interaction. As shown in Figure 4, the fusion of channel attention and spatial attention is achieved by capturing the interaction between the spatial dimension and the channel dimension of the input tensor.

Specifically, given the input tensor , pass it to the three branches of the Triplet attention module. In the first branch, the identity residual branch structure is introduced. First, through Z-Pool, the average pooling and maximum pooling features are integrated to achieve dimension scaling. After a 7×7 size convolution (Conv), Batch Norm normalization is performed, and the attention weight is generated through the Sigmoid activation function, and the first branch output is obtained after dot multiplication with the identity residual branch; Z-Pool is responsible for the connection The maximum pooling kernel average pooling feature of this dimension reduces the 0th dimension of the tensor to 2 dimensions, retaining the rich features of the actual tensor while reducing its depth to achieve lightweight purposes. The Z-Pool formula can be expressed for:

The Sigmoid activation function formula can be expressed as:

Input the feature map tensor in the second branch Rotate 90° counterclockwise along the W axis to realize the dimension transformation tensor expressed as/> , achieve dimension scaling after Z-Pool, and get/> And after a 7×7 matrix and normalization processing, it is obtained/> The output, using the Sigmoid activation function, is rotated 90° clockwise along the W axis to obtain the output/> . Similarly, in the third branch, the H axis is rotated to establish feature interaction between the H dimension and the C dimension.

Step 4: Training and verification, iteratively optimize the model

In this embodiment, model training and model verification output the subsurface temperature structure. The preprocessed mesoscale eddy subsurface structure inversion sample library is used as a training set to train the two-flow model for subsurface temperature inversion, and the inversion model after training is obtained. In the model training, the experiment set batch_size to 64, epoch to 50, Adam was used as the optimizer, and the learning rate was 1e-3.

During the training process, the loss function uses Mean absolute error (MAE) loss, which represents the sum of absolute differences between the real value and the predicted value. The loss function is used for model optimization during the training process to make the prediction result approach the target result, where as Under the formula, where is the predicted value,/> is the real value, n is the number of a set of temperature fields:

The evaluation criteria are used to evaluate the inversion effect and verify the experimental results; R2, MAE and explained_variance_score are used for comparison. R2 is used to measure the degree of fit to the predicted data. The value is between [0,1], and its formula can be Expressed as:

Among them, f is the estimated value, y is the real value, is the average of the observed data.

explained_variance is the explained variance, which is used to measure the degree of explanation of the fluctuation of the data set by the model. It can calculate the explained variance of the true value and the predicted value. The value is between [0,1]. The formula is as follows, where var is the variance, y is the real value, is the predicted value.

Step 5: Obtain the trained model, input test data, and output the subsurface temperature structure

The model after iterative optimization training is used as an inference model. The sea surface information (SSH, SST) in the test data is input into the inference model to obtain the inverted temperature results of mesoscale eddies at different depths, and the matplotlib library is used for data visualization.

Embodiment 2

An embodiment of the present disclosure provides a mesoscale eddy three-dimensional subsurface structure inversion system based on a two-flow model, including a data acquisition module and a temperature inversion module:

The data collection module is configured to: collect mesoscale eddy sea surface information to be inverted through satellites, including sea surface temperature and sea surface height;

The temperature inversion module is configured to: input the mesoscale eddy sea surface information into the trained two-flow model, invert the temperature results at different depths of the mesoscale eddy, and obtain the mesoscale eddy subsurface temperature profile;

Among them, the dual-stream model introduces a triplet attention mechanism, and uses a three-branch structure to fuse channel attention and spatial attention for cross-dimensional interaction. The first branch establishes the feature interaction between the H dimension and the W dimension, and the second branch establishes the C The feature interaction between dimension and W dimension, and the third branch establishes the feature interaction between H dimension and C dimension.

Embodiment 3

The purpose of this embodiment is to provide a computer-readable storage medium.

A computer-readable storage medium has a computer program stored thereon. When the program is executed by a processor, the steps in the mesoscale vortex three-dimensional subsurface structure inversion method based on the two-flow model described in Embodiment 1 of the present disclosure are implemented.

Embodiment 4

The purpose of this embodiment is to provide electronic equipment.

Electronic equipment, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the mesoscale vortex three-dimensional model based on the two-flow model is implemented as described in Embodiment 1 of the present disclosure. Steps in the subsurface structure inversion method.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims (10)

1. The inversion method of the mesoscale vortex three-dimensional subsurface structure based on the double-flow model is characterized by comprising the following steps of:

acquiring mesoscale vortex sea surface information to be inverted, including sea surface temperature and sea surface height, through a satellite;

inputting the mesoscale vortex sea surface information into a trained double-flow model, and inverting temperature results of the mesoscale vortex at different depths to obtain a mesoscale vortex subsurface temperature profile;

the dual-flow model introduces a Triplet attention attention mechanism, adopts a three-branch structure to fuse channel attention and space attention for cross-dimensional interaction, the first branch establishes characteristic interaction between H dimension and W dimension, the second branch establishes characteristic interaction between C dimension and W dimension, and the third branch establishes characteristic interaction between H dimension and C dimension.

2. The inversion method of a mesoscale vortex three-dimensional subsurface structure based on a double-flow model according to claim 1, wherein the double-flow model adopts an encoding-decoding structure, an encoding stage is used for feature extraction of sea surface temperature and sea surface height, and a decoding stage is used for feature fusion.

3. The inversion method of a mesoscale vortex three-dimensional subsurface structure based on a double-flow model according to claim 2, wherein the encoding stage takes sea surface temperature and sea surface height as input to explore the relationship between the sea surface temperature, the sea surface height and the subsurface temperature respectively; the method comprises the steps of carrying out feature extraction on sea surface height and sea surface temperature, simultaneously considering that a mesoscale vortex has complex nonlinear features, inevitably connecting sea surface height information and sea surface temperature information, constructing a data feature fusion network for fusing SSH and SST relationship features, and reducing the risk of overfitting by using jump connection;

and the decoding stage fuses the characteristic output of each layer of the encoding stage and realizes the recovery of resolution by utilizing deconvolution.

4. The inversion method of a mesoscale vortex three-dimensional subsurface structure based on a double-flow model according to claim 1, wherein the first branch is introduced with an identity residual branch structure, and dimension scaling is realized by fusing average pooling and maximum pooling features through Z-Pool;

the second branch rotates the input feature map tensor along the W axis to realize feature interaction of the C dimension and the W dimension;

and the third branch rotates the input feature map tensor along the H axis to realize feature interaction of the H dimension and the C dimension.

5. The inversion method of a mesoscale vortex three-dimensional subsurface structure based on a double-flow model according to claim 1, wherein the double-flow model is trained by taking a constructed mesoscale vortex subsurface structure inversion sample library as a training data set;

and constructing the mesoscale vortex subsurface structure inversion sample library by taking the sea surface information corresponding to the mesoscale vortex time and the vortex center coordinate position as input parameters and the subsurface temperature data corresponding to the mesoscale vortex time and the vortex center coordinate position as labels.

6. The method for inversion of a mesoscale vortex three-dimensional subsurface structure based on a dual-flow model according to claim 1, wherein the loss function of the dual-flow model adopts MAE loss, which represents the sum of absolute differences between a true value and a predicted value, and the formula is:

，

wherein,for predictive value +.>For a true value, n is the number of temperature fields in a set.

7. The inversion method of a mesoscale vortex three-dimensional subsurface structure based on a double-flow model according to claim 1, wherein the evaluation criteria of the double-flow model are compared by using R2, MAE and displayed_variance_score, wherein the R2 is used for measuring the fitting degree of predicted data, and the formula is as follows:

，

wherein f is an estimated value, y is a true value,is the average value of the observed data;

the expanded_variance_score is an interpretation variance, and is used for measuring the interpretation degree of the model on the fluctuation of the data set, and calculating the interpretable variance of the true value and the predicted value, wherein the formula is as follows:

，

wherein va isr is the variance, y is the true value,is a predicted value.

8. The mesoscale vortex three-dimensional subsurface structure inversion system based on the double-flow model is characterized by comprising a data acquisition module and a temperature inversion module:

a data acquisition module configured to: acquiring mesoscale vortex sea surface information to be inverted, including sea surface temperature and sea surface height, through a satellite;

a temperature inversion module configured to: inputting the mesoscale vortex sea surface information into a trained double-flow model, and inverting temperature results of the mesoscale vortex at different depths to obtain a mesoscale vortex subsurface temperature profile;

the dual-flow model introduces a Triplet attention attention mechanism, adopts a three-branch structure to fuse channel attention and space attention for cross-dimensional interaction, the first branch establishes characteristic interaction between H dimension and W dimension, the second branch establishes characteristic interaction between C dimension and W dimension, and the third branch establishes characteristic interaction between H dimension and C dimension.

9. An electronic device, comprising:

a memory for non-transitory storage of computer readable instructions; and

a processor for executing the computer-readable instructions,

wherein the computer readable instructions, when executed by the processor, perform the method of any of the preceding claims 1-7.

10. A storage medium, characterized by non-transitory storing computer-readable instructions, wherein the instructions of the method of any one of claims 1-7 are performed when the non-transitory computer-readable instructions are executed by a computer.