A method and system for inverting atmospheric temperature profiles based on graph neural networks

Abstract

This invention discloses a method and system for atmospheric temperature profile inversion based on graph neural networks. The method includes acquiring GIIRS long-wave infrared radiation data, ERA5 temperature profile data, and radiosonde station observation data, and performing preprocessing to obtain GIIRS brightness temperature data, preprocessed ERA5 temperature profile data, and preprocessed radiosonde station observation data, respectively. A spatial graph inversion network is constructed, and the GIIRS brightness temperature data and its coordinates are input into the spatial graph inversion network to obtain deep features including location context information. A graph structure is constructed based on the deep features and K-NN. Based on the deep features and the edges of the graph structure, depth features and an adjacency matrix are obtained. The adjacency matrix and depth features are jointly input into a feature aggregator to obtain the atmospheric temperature profile. This invention fully utilizes the spatial information of the data, giving it a significant advantage in atmospheric temperature profile inversion tasks.

Description

Technical Field

[0001] This invention relates to the field of atmospheric science and technology, specifically to a method and system for inverting atmospheric temperature profiles based on graph neural networks. Background Technology

[0002] Atmospheric temperature, as a key parameter characterizing the dynamic and thermodynamic state of the atmosphere, has significant application value in weather analysis, numerical models, and other fields, and is crucial for improving the accuracy of weather forecasts and monitoring high-impact weather. Therefore, obtaining continuous and accurate atmospheric temperature profiles is of great value.

[0003] With the continuous advancement of satellite and remote sensing technologies, current methods for retrieving atmospheric profiles from spaceborne hyperspectral infrared radiation observations are mainly categorized into three types: physical inversion methods, statistical regression methods, and machine learning methods. These methods are widely used in the field of temperature inversion, but each has its limitations. Physical inversion methods heavily rely on the quality of the background field (or the initial guessed profile) and the accuracy of the radiative transfer model; statistical regression methods are highly dependent on the representativeness and completeness of the training samples, and their generalization ability is relatively limited. Currently, the mainstream machine learning methods include backpropagation (BP) neural network algorithms and 1D-CNN (1-Dimensional Convolutional Neural Network) algorithms. However, these methods are based on a one-dimensional or pixel-by-pixel model, independently processing the radiation observations of each field of view or pixel to retrieve the vertical temperature distribution at that point. This point-by-point processing paradigm essentially treats a physically continuous atmospheric field as a series of discrete, isolated data points, lacking consideration of the valuable spatial context information contained in neighboring pixels that can constrain the inversion results, thus limiting the improvement of inversion accuracy. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an atmospheric temperature profile inversion method and system based on graph neural networks, which improves the inversion accuracy and applicability and has stronger generalization ability, based on the problem of insufficient utilization of spatial information in the existing technology.

[0005] To solve the above technical problems, the present invention adopts the following technical solution:

[0006] A method for inverting atmospheric temperature profiles based on graph neural networks includes the following steps:

[0007] S1. Collect GIIRS long-wave infrared radiation data, ERA5 temperature profile data, and radiosonde station observation data. Preprocess these three types of data to obtain GIIRS brightness temperature data, preprocessed ERA5 temperature profile data, and preprocessed radiosonde station observation data, respectively. Divide the dataset consisting of GIIRS brightness temperature data and preprocessed ERA5 temperature profile data into training dataset, validation dataset, and test dataset according to a set ratio.

[0008] S2. Construct SGRN (Spatial Graph Retrieval Network), which includes a feature extractor, graph structure building unit, dynamic adjacency learner, and feature aggregator; train SGRN using training dataset, validation dataset, and test dataset to obtain the trained SGRN.

[0009] S3. Input the GIIRS brightness temperature data and its coordinates into the trained SGRN. After processing by the feature extractor, deep features including location context information are obtained.

[0010] S4 and the deep features from step S3 are processed by graph structure building units to obtain a graph structure.

[0011] The deep features from steps S5 and S3, along with the edges of the graph structure, are processed by a dynamic adjacency learner to obtain deep features and a mask vector. Based on the mask vector, an adjacency matrix is ​​obtained. The adjacency matrix and deep features are then input into a feature aggregator to obtain the atmospheric temperature profile.

[0012] Furthermore, in step S1, the GIIRS long-wave infrared radiation data is converted into initial GIIRS brightness temperature data using the Planck inverse function. The specific formula is as follows:

[0013] ;

[0014] in, Indicates the blackbody temperature; Indicates wave number; Represents blackbody radiation; This represents the first radiation constant, which is related to the blackbody radiation intensity; This represents the second radiation constant, which is related to the temperature and wavelength of blackbody radiation.

[0015] Spatiotemporal matching was performed on the initial GIIRS brightness temperature data, ERA5 temperature profile data, and radiosonde station observation data. Initial GIIRS brightness temperature data, ERA5 temperature profile data, and radiosonde station observation data with a time difference within a 1-hour window were considered successfully matched temporally. Spatial matching was then performed on the initial GIIRS brightness temperature data and radiosonde station observation data using the nearest neighbor principle, discarding data that were more than 0.2° apart, thus completing spatial matching. Using the initial GIIRS brightness temperature data as a reference, bilinear interpolation was used to interpolate the ERA5 temperature profile data to the center position of each GIIRS pixel, achieving spatial alignment between the ERA5 temperature profile data and the initial GIIRS brightness temperature data, thus completing spatial matching.

[0016] Based on the cloud mask attribute of the GIIRS L2 commercial product, the initial GIIRS brightness temperature data after spatiotemporal matching under clear sky conditions is retained. The initial GIIRS brightness temperature data after spatiotemporal matching under clear sky conditions, the ERA5 temperature profile data after spatiotemporal matching, and the observation data of the radiosonde station after spatiotemporal matching are normalized to obtain GIIRS brightness temperature data, preprocessed ERA5 temperature profile data, and preprocessed radiosonde station observation data.

[0017] The dataset consisting of GIIRS brightness temperature data and preprocessed ERA5 temperature profile data is divided in a ratio of 7:1:2.

[0018] Furthermore, in step S3, the feature extractor includes a spatial coordinate encoder, a three-layer fully connected layer, a ReLU layer, a batch normalization layer, and a Dropout layer.

[0019] Extract the spatial coordinate information from the GIIRS brightness temperature data to obtain the coordinates. .

[0020] The GIIRS brightness temperature data and its coordinates are input into the feature extractor. The spatial coordinate encoder processes the coordinates of the GIIRS brightness temperature data to extract the longitude component of the coordinates. and latitude components Partitioning in logarithmic space The first metric is set. The frequency parameters corresponding to each scale are: The sine and cosine transformations are applied to the components of the coordinates at different scales. The specific formulas are as follows:

[0021] ;

[0022] in, Indicates the first Coordinates at the i-th scale Position encoding of each component; The components representing the coordinates, ; Indicates the minimum wavelength; Indicates the maximum wavelength; It represents the ratio of the maximum wavelength to the minimum wavelength. .

[0023] For coordinates, which include both longitude and latitude information, the positional encoding of the coordinate components is represented as follows:

[0024] ;

[0025] in, Indicates the first Position encoding of coordinate components at each scale; Indicates the first The position code of the first component of the coordinate system at each scale, i.e., the first... Location encoding of longitude components of coordinates at various scales; Indicates the first The position code of the second component of the coordinate system at each scale, i.e., the... Location encoding of the latitude components of the coordinate system at each scale.

[0026] The positional codes of coordinate components at each scale constitute the positional codes of the coordinates, and the specific formula is as follows:

[0027] ;

[0028] in, Indicates the first Position encoding of coordinate components at each scale This indicates a splicing operation. The location code representing the coordinates.

[0029] The positional encoding of the coordinates is mapped to the corresponding feature dimension to obtain the spatial embedding vector, as shown in the following formula:

[0030] ;

[0031] in, Represents a spatial embedding vector. The learnable weight parameters are: The fully connected layer.

[0032] Spatial embedding vectors are concatenated with GIIRS brightness temperature data. The concatenated data is then processed through three fully connected layers, a ReLU layer, a batch normalization layer, and a Dropout layer to obtain deep features that include location context information.

[0033] Furthermore, the number of neurons in the three fully connected layers are 512, 1024, and 512, respectively.

[0034] Furthermore, in step S4, the distance between each pair of deep features is calculated using the semi-sine formula. The initial weights between deep features are obtained by subtracting the inverted distance from 1, and the weights with the largest values ​​are selected first. The initial weights are used as edges of the graph structure, and the deep features are used as nodes of the graph structure to complete the construction of the graph structure.

[0035] Furthermore, in step S5, the dynamic adjacency learner includes three fully connected layers.

[0036] deep features The deep features are obtained by mapping to a high-dimensional latent space through the first fully connected layer, as shown in the following formula:

[0037] ;

[0038] in, Representing depth features, , This represents the total number of deep features. The feature dimension representing deep features The feature dimension representing deep features This indicates the first fully connected layer.

[0039] Assume each graph structure node exists For each node in the graph structure, the depth feature of the node is subtracted from the depth feature of each neighbor node to obtain the relative difference between the two nodes in the feature space. This relative difference is then input into the second fully connected layer to obtain the potential association strength between the two nodes. The specific formula is as follows:

[0040] ;

[0041] in, Indicates the first The graph structure node and its first The potential association strength between neighboring nodes This represents the Sigmoid function. Represents the scoring function. Indicates the first Depth features of nodes in a graph structure Indicates the first The depth features of each neighboring node, .

[0042] The feature score matrix is ​​obtained by concatenating the potential association strengths between graph structure nodes and their neighbor nodes. ,Will The soft neighbor budget is obtained by accumulating the features along the feature dimension, and the specific formula is as follows:

[0043] ;

[0044] in, Indicates the first Soft neighbor budget for a graph structure node Degree function learner, Represents the set of neighboring nodes. Indicates the first The neighboring node and the first Geographic connectivity weights of nodes in a graph structure Indicates the first The neighbor node pairs with the first The degree of information contribution of each node in the graph structure.

[0045] The soft neighbor budget is input into the third fully connected layer to extract the graph structure nodes. The feature scores of each neighboring node are calculated, and these feature scores are sorted in descending order to obtain the mask vector. The specific formula is as follows:

[0046] ;

[0047] in, Indicates the first Each graph structure node mask vector Represents the tangent function. Indicates hyperparameters, Represents an index sequence. .

[0048] Multiply the descendingly sorted feature scores element-wise with the mask vector, and restore the original neighbor order from the result of the multiplication to obtain the nth neighbor order. The adjacency vector of each node in the graph structure Stack the adjacency vectors of all nodes in the graph structure to obtain the adjacency matrix. .

[0049] Adjacency matrix The deep features are input together into a feature aggregator consisting of two SAGEConv layers, based on the adjacency matrix. The depth features are processed by the first layer of SAGEConv, and the output is added to the depth features. Then, the output is processed by the second layer of SAGEConv to obtain the node representation vector. After the dimension of the node representation vector is denormalized, the atmospheric temperature profile is obtained.

[0050] Furthermore, this invention also proposes an atmospheric temperature profile inversion system based on graph neural networks, comprising:

[0051] The data preprocessing module is used to collect GIIRS long-wave infrared radiation data, ERA5 temperature profile data, and radiosonde station observation data. It preprocesses these three types of data to obtain GIIRS brightness temperature data, preprocessed ERA5 temperature profile data, and preprocessed radiosonde station observation data, respectively. The dataset consisting of GIIRS brightness temperature data and preprocessed ERA5 temperature profile data is divided according to a set ratio to obtain training dataset, validation dataset, and test dataset.

[0052] The network training module is used to construct the SGRN, which includes a feature extractor, a graph structure building unit, a dynamic adjacency learner, and a feature aggregator. The SGRN is trained using the training dataset, the validation dataset, and the test dataset to obtain the trained SGRN.

[0053] The deep feature acquisition module is used to input GIIRS brightness temperature data and its coordinates into the trained SGRN. After processing by the feature extractor, deep features including location context information are obtained.

[0054] The graph structure construction module is used to process the deep features of the deep feature acquisition module to obtain a graph structure.

[0055] The atmospheric temperature profile acquisition module is used to process the deep features of the deep feature acquisition module and the edges of the graph structure through a dynamic adjacency learner to obtain deep features and a mask vector; based on the mask vector, the adjacency matrix is ​​obtained; the adjacency matrix and the deep features are input together into the feature aggregator to obtain the atmospheric temperature profile.

[0056] Furthermore, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the atmospheric temperature profile inversion method based on graph neural networks.

[0057] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the described atmospheric temperature profile inversion method based on graph neural networks.

[0058] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0059] 1. This invention constructs a spatial coordinate encoder to display the spatial relationship between GIIRS detectors, and uses graph convolution to update and aggregate the information of GIIRS detectors, thus solving the limitation of insufficient utilization of spatial information in existing technologies.

[0060] 2. The dynamic adjacency learner designed in this invention can dynamically optimize the weight relationship between nodes based on the characteristics of real-time brightness and temperature data, and can learn the degree of each node, thereby constructing a more reasonable node neighbor relationship and having a stronger generalization ability.

[0061] 3. The root mean square error, average error, and average absolute error of the inversion results of this invention are all improved compared with the existing technology, and all show high accuracy and low error. Therefore, this invention has more advantages in atmospheric temperature profile inversion tasks and is applicable to different regions across the country. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the overall implementation of the present invention.

[0063] Figure 2 This is a flowchart of the process for obtaining deep features according to the present invention.

[0064] Figure 3 This is a flowchart illustrating how the spatial embedding vector is obtained in this invention.

[0065] Figure 4 This is a flowchart of the process for obtaining the adjacency matrix in this invention.

[0066] Figure 5 These are vertical profiles of different networks on evaluation metrics in embodiments of the present invention.

[0067] Figure 6 This is a box plot showing the error distribution characteristics of different networks on the key pressure layer in this embodiment of the invention.

[0068] Figure 7 This is a bar chart showing the error distribution of different networks on the evaluation metrics in the embodiments of the present invention.

[0069] Figure 8 This is a vertical profile diagram of different networks on the evaluation index based on ERA5 temperature profile data in an embodiment of the present invention.

[0070] Figure 9 This is a vertical profile diagram of different networks on evaluation indicators based on observation data from radiosonde stations in this embodiment of the invention.

[0071] Figure 10 These are inversion results diagrams of different networks based on different radiosonde sites at different times in embodiments of the present invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0073] To achieve the above objectives, this invention proposes an atmospheric temperature profile inversion method based on graph neural networks, such as... Figure 1 As shown, the specific steps are as follows:

[0074] S1. This invention collected GIIRS long-wave infrared radiation data, ERA5 temperature profile data, and observation data from 120 radiosonde stations from May to August 2024. These three types of data were preprocessed to obtain GIIRS brightness temperature data, preprocessed ERA5 temperature profile data, and preprocessed radiosonde station observation data, respectively. The dataset consisting of GIIRS brightness temperature data and preprocessed ERA5 temperature profile data was divided into training, validation, and test datasets according to a 7:1:2 ratio. Specifically:

[0075] The inverse Planck function is used to transform GIIRS long-wave infrared radiation data into initial GIIRS brightness temperature data with a clearer physical meaning. The specific formula is as follows:

[0076] ;

[0077] in, Indicates the blackbody temperature; Indicates wave number; Represents blackbody radiation; This represents the first radiation constant, which is related to the blackbody radiation intensity. ; This represents the second radiation constant, which is related to the temperature and wavelength of blackbody radiation. .

[0078] Spatiotemporal matching was performed on the initial GIIRS brightness temperature data, ERA5 temperature profile data, and radiosonde station observation data. Initial GIIRS brightness temperature data, ERA5 temperature profile data, and radiosonde station observation data with a time difference within a 1-hour window were considered successfully matched in time. The FY-4B / GIIRS pixel resolution is 12km × 12km. Spatial matching was performed on the initial GIIRS brightness temperature data and radiosonde station observation data using the nearest neighbor principle. To reduce errors caused by spatial matching, data separated by more than 0.2° were removed. A total of 94 radiosonde stations were successfully matched. Using the high-resolution initial GIIRS brightness temperature data as a benchmark, bilinear interpolation was used to accurately interpolate the low-resolution ERA5 temperature profile data to the center position of each GIIRS pixel, achieving spatial alignment between the ERA5 temperature profile data and the initial GIIRS brightness temperature data, thus completing spatial matching.

[0079] Because infrared hyperspectral detectors are extremely sensitive to clouds, and infrared radiation cannot penetrate cloud layers, information below the clouds cannot be seen under cloudy conditions. Therefore, the initial GIIRS brightness temperature data after spatiotemporal matching is filtered. Based on the cloud mask attributes of the GIIRS L2 commercial product (the cloud mask attributes are 0 and 1, where 0 represents clear sky and 1 represents cloudy sky), the initial GIIRS brightness temperature data after spatiotemporal matching under clear sky conditions is retained. The GIIRS long-wave infrared band has 725 spectral channels. There is some information redundancy among these channels, and some channels do not contribute much to the inversion task. By comparing the channels of CrIS and GIIRS, 201 brightness temperature channels are determined to meet the requirements, thereby reducing the computational load and minimizing the negative impact of redundant channels.

[0080] Normalization processing was performed on the initial GIIRS brightness temperature data, the ERA5 temperature profile data, and the radiosonde station observation data after spatiotemporal matching under clear sky conditions to obtain GIIRS brightness temperature data, preprocessed ERA5 temperature profile data, and preprocessed radiosonde station observation data.

[0081] S2. Construct SGRN (Spatial Graph Retrieval Network), which includes a feature extractor, graph structure building unit, dynamic adjacency learner, and feature aggregator; train SGRN using training dataset, validation dataset, and test dataset to obtain the trained SGRN.

[0082] S3. Input the GIIRS brightness temperature data and its coordinates into the trained SGRN. After processing by the feature extractor, deep features including location context information are obtained. This step lays the groundwork for subsequent operations, providing deep features with spatial information for the learning of the graph neural network below. This facilitates the aggregation and updating of features by graph convolution. Specifically:

[0083] like Figure 2 As shown, the feature extractor includes a spatial coordinate encoder, three fully connected layers, a ReLU layer, a batch normalization layer, and a Dropout layer. The core idea of ​​the spatial coordinate encoder is to learn a flexible, high-dimensional spatial context representation for the geographic coordinates of each observation point.

[0084] To integrate spatial information, spatial coordinate information of GIIRS brightness temperature data was extracted to obtain coordinates. .

[0085] like Figure 3 As shown, GIIRS brightness temperature data and its coordinates are input into the feature extractor. The spatial coordinate encoder processes the coordinates of the GIIRS brightness temperature data to extract the longitude component of the coordinates. and latitude components Partitioning in logarithmic space The first metric is set. The frequency parameters corresponding to each scale are: , will the The frequency parameter at each scale is considered as a scaling factor. Due to the existence Each scale corresponds to one of the following: Several scaling factors, which in turn constitute a frequency set. , Multiply the latitude and longitude by the scaling factor respectively to obtain For each coordinate component at a different scale, a sine and cosine transform are performed on that component, using the following formulas:

[0086] ;

[0087] in, Indicates the first Coordinates at the i-th scale Position encoding of each component; The components representing the coordinates, ; Indicates the minimum wavelength; Indicates the maximum wavelength; It represents the ratio of the maximum wavelength to the minimum wavelength. .

[0088] For coordinates, which include both longitude and latitude information, the positional encoding of the coordinate components is represented as follows:

[0089] ;

[0090] in, Indicates the first Position encoding of coordinate components at each scale; Indicates the first The position code of the first component of the coordinate system at each scale, i.e., the first... Location encoding of longitude components of coordinates at various scales; Indicates the first The position code of the second component of the coordinate system at each scale, i.e., the... Location encoding of the latitude components of the coordinate system at each scale.

[0091] The positional codes of coordinate components at each scale constitute the positional codes of the coordinates, and the specific formula is as follows:

[0092] ;

[0093] in, Indicates the first Position encoding of coordinate components at each scale This indicates a splicing operation. The location code representing the coordinates.

[0094] The positional encoding of the coordinates is mapped to an appropriate feature dimension to obtain a spatial embedding vector, as shown in the following formula:

[0095] ;

[0096] in, Represents a spatial embedding vector. The learnable weight parameters are: The fully connected layer.

[0097] Each row of the spatial embedding vector is a high-dimensional feature that can characterize the spatial context of the corresponding brightness temperature data.

[0098] To incorporate location information into the GIIRS brightness temperature data, the spatial embedding vector is concatenated with the GIIRS brightness temperature data. The concatenated data is then processed through three layers of fully connected layers, ReLU layers, batch normalization layers, and Dropout layers to obtain deep features that include location context information.

[0099] The number of neurons in the three fully connected layers are 512, 1024, and 512, respectively.

[0100] S4. In the graph structure building unit, the semi-versus formula is used to calculate the distance between any two deep features. The initial weights between deep features are obtained by subtracting the inverse distance from 1. The first 16 initial weights are selected as edges of the graph structure, and the deep features are used as nodes to complete the graph structure. The construction of.

[0101] S5. The brightness temperature data at different locations depends on the number of neighboring nodes. The deep features from step S3 and the edges of the graph structure are processed by a dynamic adjacency learner to obtain deep features and a smoothed mask vector. Based on the smoothed mask vector, an adjacency matrix is ​​obtained. The adjacency matrix and deep features are input together into a feature aggregator to obtain the atmospheric temperature profile. Specifically:

[0102] like Figure 4 As shown, the dynamic adjacency learner consists of three fully connected layers.

[0103] deep features The deep features are obtained by mapping to a high-dimensional latent space through the first fully connected layer, as shown in the following formula:

[0104] ;

[0105] in, Representing depth features, , This represents the total number of deep features. The feature dimension representing deep features The feature dimension representing deep features This indicates the first fully connected layer.

[0106] Assume each graph structure node exists For each node in the graph structure, the depth feature of the node is subtracted from the depth feature of each neighbor node to obtain the relative difference between the two nodes in the feature space. This relative difference is then input into the second fully connected layer to obtain the potential association strength between the two nodes. The specific formula is as follows:

[0107] ;

[0108] in, Indicates the first The graph structure node and its first The potential association strength between neighboring nodes This represents the Sigmoid function. Represents the scoring function. Indicates the first Depth features of nodes in a graph structure Indicates the first The depth features of each neighboring node, .

[0109] The feature score matrix is ​​obtained by concatenating the potential association strengths between graph structure nodes and their neighbor nodes. ,Will The soft neighbor budget is obtained by accumulating the features along the feature dimension, and the specific formula is as follows:

[0110] ;

[0111] in, Indicates the first Soft neighbor budget for a graph structure node Degree function learner, Represents the set of neighboring nodes. Indicates the first The neighboring node and the first Geographic connectivity weights of nodes in a graph structure Indicates the first The neighbor node pairs with the first The degree of information contribution of each node in the graph structure.

[0112] The soft neighbor budget is input into the third fully connected layer to extract the graph structure nodes. The feature scores of each neighboring node are calculated, and these feature scores are sorted in descending order to obtain a smooth mask vector. The specific formula is as follows:

[0113] ;

[0114] in, Indicates the first Each graph structure node mask vector Represents the tangent function. Represents an index sequence. , This is a hyperparameter that controls the degree of softness or hardness during the selection process; it is set to 1.

[0115] The feature scores, arranged in descending order, are multiplied element-wise by the smoothed mask vector. The result of the multiplication is then restored to the original neighbor order to obtain the nth... The adjacency vector of each node in the graph structure Stack the adjacency vectors of all graph structure nodes to obtain an adaptively selected and reweighted adjacency matrix for downstream graph convolution. .

[0116] Adjacency matrix The deep features are input together into a feature aggregator consisting of two SAGEConv layers, based on the adjacency matrix. The depth features are processed by the first layer of SAGEConv, and the output is added to the depth features. Then, the output is processed by the second layer of SAGEConv to obtain the node representation vector. The dimension of the node representation vector directly corresponds to the number of layers of the temperature profile. After the dimension of the node representation vector is denormalized, the atmospheric temperature profile is obtained.

[0117] This paper compares the performance of the proposed SGRN method with CNN (Convolutional Neural Network), BPNN (Back Propagation Neural Network), and BRNN (Bidirectional Recurrent Neural Network) in atmospheric temperature profile inversion to evaluate the inversion performance of different methods. The evaluation metrics used are RMSE (Root Mean Square Error), MB (Mean Bias), and MAE (Mean Absolute Error). RMSE measures the dispersion of the error between the predicted and true values, effectively reflecting the maximum deviation level; MB measures the systematic bias of the prediction, indicating whether the predicted value is generally too high or too low; MAE reflects the average deviation of the predicted values. The smaller the evaluation metric, the higher the inversion accuracy of the corresponding method.

[0118] Using the preprocessed ERA5 temperature profile data as a reference, the inversion performance of SGRN, CNN, BPNN, and BRNN was tested based on the test dataset. All evaluation metrics were calculated within the vertical range of 100 hPa to 900 hPa, and the specific results are shown in Table 1.

[0119] Table 1. Inversion performance of different networks on evaluation metrics

[0120]

[0121] As can be seen from Table 1, the SGRN proposed in this invention achieves the best performance in all evaluation metrics, significantly outperforming other networks.

[0122] To further explore the inversion performance of SGRN, CNN, BPNN, and BRNN at different height levels, Figure 5 Vertical profiles of RMSE and MB for different networks on the test dataset are shown. Figure 5 (a) shows the vertical profile of the RMSE of different networks on the test dataset. It can be seen that in the upper troposphere from 100 hPa to 500 hPa, the RMSE obtained by SGRN is stable between 0.8 K and 1.2 K, which is significantly better than other networks. In contrast, the RMSE obtained by BPNN and CNN increases sharply in the near-surface layer below 600 hPa, all exceeding 1.72 K. The RMSE obtained by SGRN is controlled within 1.72 K in all layers. Figure 5(b) shows the vertical profiles of the biased averages (MBs) of different networks on the test dataset. It can be seen that the error obtained based on SGRN is very close to zero in most pressure layers. Between 100 hPa and 500 hPa, the MB obtained based on SGRN fluctuates within a very small range of ±0.1 K, indicating almost no bias in this region. Although in the lower layers between 700 hPa and 900 hPa, the negative bias of the MB obtained based on SGRN (approximately -0.2 K to -0.4 K) is slightly larger than that obtained based on BPNN, SGRN still demonstrates superior bias control capability in terms of overall profile stability and fluctuation amplitude. Therefore, from... Figure 5 As can be seen, SGRN exhibits the lowest error across almost all vertical layers, demonstrating a comprehensive accuracy advantage.

[0123] Box plots were used to reveal the error distribution characteristics of SGRN, CNN, BPNN, and BRNN at key pressure layers in more detail. The results are shown in the figure below. Figure 6 As shown, the blue shaded area represents the error tolerance band of ±1 K. Figure 6 (a) is a box plot of the error distribution features of CNN on the key pressure layer. Figure 6 (b) is a box plot showing the error distribution characteristics of BPNN on key pressure layers. Figure 6 (c) is a box plot showing the error distribution characteristics of BRNN on the key pressure layer. Figure 6 (d) is a box plot showing the error distribution characteristics of SGRN in the key pressure layer. From Figure 6 As can be seen from the data, the SGRN box and whisker lengths are the shortest among all the displayed pressure layers, and the SGRN box almost completely falls within the blue shaded area, indicating that the SGRN inversion error distribution is more concentrated and has less dispersion. That is, the middle 50% of the data and the vast majority of the data (5%-95% quantiles) are closely around the zero deviation line, and the accuracy of most inversion results has reached a high standard.

[0124] To further evaluate the performance of different networks in the inversion task at the sample level, a bar chart was introduced to statistically categorize the prediction results of each network, such as... Figure 7As shown in Figure 7(a), the RMSE error histogram for different network inversion performances is divided into three levels (RMSE≤1K, 1...).<RMSE≤2K、RMSE> The error ranges (2K) represent low, medium, and high errors, respectively, which are labeled as low, medium, and high for ease of plotting to evaluate the network's coverage capability within different error ranges. It can be seen that the RMSE obtained based on SGRN accounts for 39.3% of the samples in the <1K range, significantly higher than the 25.1% obtained based on BRNN, 20.2% based on BPNN, and 13.6% based on CNN. Meanwhile, in the high error region exceeding 2K, the RMSE obtained based on SGRN accounts for only 6.2%, significantly lower than other networks. Figure 7(b) shows the MB error of different networks' inversion performance. The bar chart divides the sample MB into three levels (|MB|≤0.5K, 0.5K<|MB|≤1K, |MB|>1K), representing low error, medium error, and high error, respectively. For ease of plotting, these are labeled low, medium, and high. It can be seen that the MB obtained based on SGRN accounts for 63.9% of the samples in the ±0.5K deviation center region, significantly higher than the 52.5% obtained based on BRNN, 49% based on BPNN, and 37.6% based on CNN. Furthermore, in the region with a deviation exceeding ±1K, the MB obtained based on SGRN accounts for only 8.3%, far lower than other networks. Therefore, from... Figure 7 As can be seen, SGRN not only has a lower overall error level, but also maintains a small prediction error on most samples, demonstrating stronger practicality and robustness.

[0125] To evaluate the absolute accuracy of SGRN in real-world observation scenarios, radiosonde data from the same period as the test dataset were introduced as an independent third-party verification benchmark. GIIRSL2, an operational product released by the China Meteorological Satellite Center, was also used for comparison. The specific results are shown in Table 2.

[0126] Table 2. Inversion performance of different networks on evaluation metrics based on real observation scenarios.

[0127]

[0128] As shown in Table 2, on both different evaluation datasets, the RMSE obtained based on SGRN is superior to that obtained by CNN, BPNN, and BRNN, and is basically on par with the performance of the network in the GIIRS L2 operational product. The MB obtained based on SGRN is only superior to the MB obtained based on BPNN when the preprocessed ERA5 temperature profile data is used as the ground truth; otherwise, it is optimal. The MAE obtained based on SGRN is significantly superior to the MAE obtained based on CNN, BPNN, and BRNN, and is comparable to the MAE obtained by the network based on the GIIRS L2 operational product.

[0129] Figure 8 Figure 8(a) shows the vertical profiles of different networks on the evaluation metrics, using preprocessed ERA5 temperature profile data as the true value, within a data subset under the quality control conditions of the GIIRS L2 commercial product. Figure 8(b) shows the vertical profiles of different networks on RMSE, using preprocessed ERA5 temperature profile data as the true value, within the data subset under the quality control conditions of the GIIRS L2 commercial product. Figure 8 As can be seen, only at 700hPa~800hPa, the MB obtained based on SGRN is lower than that obtained based on BRNN. In other cases, it is superior to BPNN and CNN in all aspects. Compared with the network of the GIIRS L2 operational product, the RMSE obtained based on SGRN is more stable overall. The error of some atmospheric layers is also better than the RMSE obtained by the network of the GIIRS L2 operational product. Moreover, the MB obtained based on SGRN has an amplitude of only ±0.1 in 100~600hPa, which indicates that the inversion result of SGRN has higher unbiasedness and can provide a more reliable overall estimate of atmospheric temperature. This is especially important for application scenarios such as climate change research that require accurate bias correction.

[0130] Using the preprocessed radiosonde station observation data as the ground truth, vertical profiles of different networks on evaluation metrics were obtained, such as... Figure 9 As shown. Figure 9(a) shows the vertical profiles of different networks on RMSE, with preprocessed radiosonde observation data as the true value, in a data subset under quality control conditions of GIIRS L2 operational products. Figure 9(b) shows the vertical profiles of different networks on MB, with preprocessed radiosonde observation data as the true value, in a data subset under quality control conditions of GIIRS L2 operational products. Figure 9As can be seen, the inversion results of SGRN are superior to CNN, BPNN, and BRNN in all atmospheric layers, and superior to the network of GIIRS L2 commercial products in some layers. The MB obtained based on SGRN has a smaller oscillation amplitude compared to the MB obtained by other networks, mainly hovering around ±0, indicating that the inversion results of SGRN have no obvious deviation.

[0131] After evaluating the overall inversion result error, the performance of each network is analyzed at a micro level through three typical inversion cases. The results are shown in the figure below. Figure 10 As shown. Figure 10 (a) shows the inversion results of different networks for site 51076 at 01:00 UTC (Universal Time Coordinated) on August 15, 2024. The profile shows a significant tropopause inversion near 200 hPa. Both SGRN and GIIRS L2 commercial products successfully inverted the overall shape of the profile, showing high consistency. Furthermore, in the key inversion layer region, the inverted profile of SGRN is closer to the true value in terms of transition shape and temperature value, showing a stronger ability to capture details. While CNN, BPNN, and BRNN generally followed the trend of the profile, they performed poorly in key areas, showing significant negative bias (lower temperature) in the 200 hPa inversion layer and the troposphere at 600-900 hPa, with CNN showing particularly severe bias. Figure 10 (b) shows the inversion results of different networks at site 54539 when there was no temperature inversion at 13:00 on August 22, 2024. The network of the GIIRS L2 commercial product performed well under these normal conditions, and the inversion results were largely consistent with the true values, with only a slight positive deviation in the low-altitude 700-900 hPa. In the absence of a temperature inversion layer, CNN, BPNN and BRNN still showed unexplained temperature underestimation in the lower stratosphere at 200-300 hPa; while SGRN once again demonstrated strong stability. Figure 10(c) shows the inversion results of different networks for station 50557 at 13:00 UTC on August 26, 2024. CNN, BPNN and BRNN all captured the trend of temperature inversion, but there is an overall temperature underestimation problem. The network of GIIRS L2 commercial product shows the opposite characteristics, and is slow to respond to the fine changes in the inversion layer, resulting in a significant negative bias in the 100-200 hPa high-altitude region, while turning into a positive bias (higher temperature) in the mid-low altitude region of 450-900 hPa. The profile shape is relatively smooth, but key details are lost. SGRN achieves a perfect balance between sensitivity and accuracy. It not only keenly captures all major vertical structure changes, but the inverted temperature value also matches the true value very well, indicating that SGRN can maintain high accuracy and high reliability under various atmospheric conditions.

[0132] This invention also proposes an atmospheric temperature profile inversion system based on graph neural networks, including a data preprocessing module, a network training module, a deep feature acquisition module, a graph structure construction module, an atmospheric temperature profile acquisition module, and a computer program that can run on a processor. It should be noted that each module in the above system corresponds to a specific step of the method provided in this invention embodiment, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention embodiment.

[0133] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that when the processor executes the computer program, it corresponds to the specific steps of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0134] This invention also proposes a computer-readable storage medium storing a computer program. It should be noted that when the computer program is executed by a processor, it corresponds to the specific steps of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0135] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An atmospheric temperature profile inversion method based on a graph neural network, characterized in that, The method comprises the following steps: S1, collecting GIIRS long-wave infrared radiation data, ERA5 temperature profile data and observation data of sounding stations, preprocessing the three types of data to obtain GIIRS brightness temperature data, preprocessed ERA5 temperature profile data and preprocessed observation data of sounding stations; dividing the data set composed of the GIIRS brightness temperature data and the preprocessed ERA5 temperature profile data according to a set proportion to obtain a training data set, a verification data set and a test data set; S2, constructing a spatial graph inversion network, which comprises a feature extractor, a graph structure construction unit, a dynamic adjacency learner and a feature aggregator; training the spatial graph inversion network by using the training data set, the verification data set and the test data set to obtain a trained spatial graph inversion network; S3, inputting the GIIRS brightness temperature data and its coordinates into the trained spatial graph inversion network to obtain deep features comprising position context information through processing of the feature extractor; Specifically, The feature extractor comprises a spatial coordinate encoder, three fully connected layers, a ReLU layer, a batch normalization layer and a Dropout layer; The spatial coordinate information of the GIIRS brightness temperature data is extracted to obtain coordinates P; The GIIRS brightness temperature data and the coordinates are input into a feature extractor, the coordinates of the GIIRS brightness temperature data are processed by using a spatial coordinate encoder, and the longitude component P [1] and the latitude component P [2] S scales are divided in a logarithmic space, and the frequency parameter corresponding to the s th scale is set as The components of the coordinates under different scales are subjected to sine transformation and cosine transformation, and the specific formula is: where CE s,v denotes the position encoding of the v-th component of the coordinate at the s-th scale; P [v] denotes the component of the coordinate, v = 1, 2; β min denotes the minimum wavelength; β max denotes the maximum wavelength; r denotes the ratio of the maximum wavelength and the minimum wavelength, For the coordinates, each coordinate comprises longitude and latitude information, and the position encoding of the coordinate components is represented as: CE s (P,β min ,β max )=[CE s,1 (P,β min ,β max ),CE s,2 (P,β min ,β max )] wherein CE s denotes the position encoding of the coordinate component at the s-th scale; CE s,1 denotes the position encoding of the 1st coordinate component at the s-th scale, i.e. the position encoding of the longitude coordinate component at the s-th scale; CE s,2 denotes the position encoding of the 2nd coordinate component at the s-th scale, i.e. the position encoding of the latitude coordinate component at the s-th scale; The position encoding of the coordinates at S scales comprises the position encoding of the coordinate components, and the specific formula is: CE(P, β min , β max ) = Concat(CE0(P, β min , β max ),..., CE S-1 (P, β min , β max )) where CE S-1 denotes the position encoding of the coordinate component at the S-1th scale, Concat denotes the concatenation operation, and CE denotes the position encoding of the coordinate; The position encoding of the coordinates is mapped to the corresponding feature dimension to obtain a spatial embedding vector, and the specific formula is: wherein SCE emb represents a spatial embedding vector, represents a learnable weight parameter as θ PE a fully connected layer; The spatial embedding vector and the GIIRS brightness temperature data are spliced, and the spliced data is sequentially processed by the three fully connected layers, the ReLU layer, the batch normalization layer and the Dropout layer to obtain deep features comprising position context information; S4, the deep features of step S3 are processed by the graph structure construction unit to obtain a graph structure; S5, the deep features of step S3 and the edges of the graph structure are processed by the dynamic adjacency learner to obtain a depth feature and a mask vector; based on the mask vector, an adjacency matrix is obtained; the adjacency matrix and the depth feature are jointly input into the feature aggregator to obtain an atmospheric temperature profile.

2. The atmospheric temperature profile retrieval method based on a graph neural network according to claim 1, characterized in that, In step S1, the GIIRS long-wave infrared radiation data is converted into initial GIIRS brightness temperature data by using the inverse function of Planck, and the specific formula is: where T represents blackbody temperature; v ′ represents wave number; L represents blackbody radiation; c1 represents first radiation constant; and c2 represents second radiation constant. The initial GIIRS brightness temperature data, the ERA5 temperature profile data and the observation data of the sounding stations are spatiotemporally matched, and the initial GIIRS brightness temperature data, the ERA5 temperature profile data and the observation data of the sounding stations with a time difference within a 1-hour window are regarded as successful time matching; the initial GIIRS brightness temperature data and the observation data of the sounding stations are spatially matched by using the nearest neighbor principle, and data with a distance exceeding 0.2° are removed to complete the spatial matching; the ERA5 temperature profile data is interpolated to the center position of each GIIRS pixel by using the bilinear interpolation method with the initial GIIRS brightness temperature data as the reference to realize the alignment of the ERA5 temperature profile data and the initial GIIRS brightness temperature data in space, and complete the spatial matching; Based on the cloud mask attribute of the GIIRS L2 commercial product, the initial GIIRS brightness temperature data after spatio-temporal matching under clear sky conditions is retained; the initial GIIRS brightness temperature data after spatio-temporal matching under clear sky conditions, the ERA5 temperature profile data after spatio-temporal matching and the observation data of the sounding station after spatio-temporal matching are normalized to obtain the GIIRS brightness temperature data, the preprocessed ERA5 temperature profile data and the preprocessed observation data of the sounding station.

3. The atmospheric temperature profile retrieval method based on a graph neural network according to claim 1, characterized in that, The number of neurons of the three fully connected layers is 512, 1024 and 512 respectively.

4. The atmospheric temperature profile retrieval method based on a graph neural network according to claim 1, characterized in that, In step S4, the distance between each pair of deep features is calculated using the secant formula, the initial weight between the deep features is obtained by subtracting the inverse distance from 1, the initial weight with a size in the top k is selected as the edge of the graph structure, the deep features are taken as the nodes of the graph structure, and the construction of the graph structure is completed.

5. The atmospheric temperature profile retrieval method based on a graph neural network according to claim 4, characterized in that, In step S5, the dynamic adjacency learner includes three fully connected layers; Deep features X ∈ R N×d Deep features are obtained by mapping to a high-dimensional latent space through the first fully connected layer, and the specific formula is: H = f embed (X) wherein H represents the deep feature, H ∈ R N×d′ N represents the total number of deep features, d represents the feature dimension of the deep feature, d ′ represents the feature dimension of the deep feature, f embed represents the first fully connected layer; The deep features of the graph structure node and the deep features of each neighbor node are subtracted to obtain the relative difference of the two nodes in the feature space, and the relative difference is input into the second fully connected layer to obtain the potential association strength between the two nodes, and the specific formula is: e uv_pi = σ(f score (H up -H vi )) where e uv_pi represents the potential association strength between the p-th graph structure node and its i-th neighbor node, σ represents a Sigmoid function, f score represents a scoring function, H up represents the deep feature of the p-th graph structure node, H vi represents the deep feature of the i-th neighbor node, i = 1, 2, …, k; The potential association strength of the graph structure node and its neighbor nodes is spliced to obtain a feature score matrix E, and the E is accumulated in the feature dimension to obtain a soft neighbor budget, and the specific formula is: wherein K p represents the soft neighbor budget of the pth graph structure node, f degree represents the degree function learner, N(i) represents the set of neighbor nodes, A geo [i,p] represents the geographical connection weight between the ith neighbor node and the pth graph structure node, E[i,p] represents the information contribution degree of the ith neighbor node to the pth graph structure node. The soft neighbor budget is input into the third fully connected layer, the feature scores of the k neighbor nodes of the graph structure node are extracted, and the feature scores are arranged in descending order to obtain a mask vector, and the specific formula is: wherein m p represents the p-th graph structure node mask vector, tanh represents the hyperbolic tangent function, w represents a hyperparameter, t represents an index sequence, t = [0, 1, …, k-1]; The feature scores arranged in descending order are multiplied with the mask vector element by element, and the multiplied result is restored to the original neighbor order to obtain an adjacency vector a of the pth graph structure node p ; the adjacency vectors of all graph structure nodes are stacked to obtain an adjacency matrix A ′ ; The adjacency matrix A ′ and the depth feature are jointly input into a feature aggregator including two layers of SAGEConv layers, based on the adjacency matrix A ′ The depth feature is processed by the first layer of SAGEConv, and the output is added to the depth feature and then processed by the second layer of SAGEConv to obtain a node representation vector. After the dimension of the node representation vector is denormalized, the atmospheric temperature profile is obtained.

6. The system applied to the atmospheric temperature profile retrieval method based on the graph neural network of claim 1, characterized in that, It comprises: A data preprocessing module is configured to collect GIIRS long-wave infrared radiation data, ERA5 temperature profile data and observation data of a sounding station, preprocess the three types of data to obtain GIIRS brightness temperature data, preprocessed ERA5 temperature profile data and preprocessed observation data of the sounding station, and divide a data set composed of the GIIRS brightness temperature data and the preprocessed ERA5 temperature profile data into a training data set, a validation data set and a test data set according to a set proportion. A network training module is configured to construct a spatial graph inversion network, which comprises a feature extractor, a graph structure construction unit, a dynamic adjacency learner and a feature aggregator, and train the spatial graph inversion network using the training data set, the validation data set and the test data set to obtain a trained spatial graph inversion network. A deep feature acquisition module is configured to input the GIIRS brightness temperature data and its coordinates into the trained spatial graph inversion network to obtain deep features comprising position context information through processing of the feature extractor. A graph structure construction module is configured to process the deep features of the deep feature acquisition module through the graph structure construction unit to obtain a graph structure. An atmospheric temperature profile acquisition module is configured to process the deep features of the deep feature acquisition module and the edges of the graph structure through the dynamic adjacency learner to obtain deep features and a mask vector. Based on the mask vector, an adjacency matrix is obtained; the adjacency matrix and the deep feature are jointly input into a feature aggregator to obtain an atmospheric temperature profile.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the atmospheric temperature profile retrieval method based on the graph neural network in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium being characterized by, The computer program is run by the processor to execute the atmospheric temperature profile retrieval method based on the graph neural network in any one of claims 1 to 5.

Images