Regional troposphere delay grid rapid generation method and system

Abstract

The invention belongs to the technical field of surveying and mapping, and discloses a method and a system for quickly generating a regional troposphere delay grid. According to the invention, a terrain and GNSS station network adaptive regional troposphere delay grid rapid generation model is constructed. The model comprises a GNSS observation station feature extraction module, a target grid point feature extraction module and a GNSS observation station-to-grid point mapping module. By introducing a graph attention network and a multi-head cross attention mechanism, local spatial correlation of station network distribution in different areas can be dynamically described, and global change features of troposphere delay are captured at the same time; the observation station position and the terrain attribute are further fused through geospatial feature embedding, and feature consistency expression under the terrain difference condition is achieved. According to the invention, while the local space dependence modeling capability is maintained, the capturing capability of global feature association is enhanced, adaptive modeling of different GNSS station network distributions and topographic changes is realized, and the space continuity and prediction precision of troposphere delay are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of surveying and mapping, and relates to a regional troposphere delay grid rapid generation method and system. BACKGROUND

[0002] Troposphere delay is an important error source in Global Navigation Satellite System (GNSS) precise positioning. In high-precision positioning and navigation applications, this delay can cause ranging errors of several meters to tens of meters, seriously affecting positioning accuracy and reliability. Especially in complex terrain and extreme weather conditions, the troposphere delay changes dramatically with space and time, making it difficult to accurately model.

[0003] In order to achieve centimeter-level high-precision positioning, accurate measurement and modeling of troposphere delay error is crucial, which helps to improve the application performance of GNSS in surveying and mapping, unmanned driving, precision agriculture and other fields.

[0004] Related technologies can establish a troposphere delay model by an empirical troposphere delay model and by interpolating / fitting the troposphere delay of GNSS continuous operation reference stations, but these methods have the following defects: (1) The traditional empirical troposphere delay model is not accurate in capturing the change of wet delay, and generally based on standard atmospheric assumptions or empirical formulas, resulting in low precision and difficulty in meeting the needs of high-precision positioning; (2) The troposphere delay model based on GNSS stations usually uses linear / low-order polynomial or Kriging method for interpolation / fitting, which is difficult to capture the nonlinear influence of complex terrain on troposphere delay; (3) The distribution of GNSS stations is uneven, resulting in poor prediction effect of traditional methods in sparse station areas; therefore, it is difficult to provide a high-precision troposphere delay correction model in complex terrain areas and sparse station areas.

[0005] Therefore, it is urgent to propose a regional troposphere delay grid rapid generation method adaptive to terrain and GNSS station network. SUMMARY

[0006] The purpose of the present application is to propose a regional troposphere delay grid rapid generation method to realize adaptive modeling of different GNSS station network distribution and terrain changes, which is beneficial to significantly improve the spatial continuity and prediction accuracy of troposphere delay.

[0007] In order to achieve the above purpose, the present application adopts the following technical solutions: A regional troposphere delay grid rapid generation method, comprising the following steps: Step 1. Obtain the zenith tropospheric delay of GNSS stations and target grid points in the target area and the geographic spatial feature information to construct a training data set for model training constructed in the following step 2. Step 2. Construct a terrain and GNSS station network adaptive regional tropospheric delay grid fast generation model, which includes a GNSS station feature extraction module, a target grid point feature extraction module and a GNSS station to grid point mapping module. The GNSS station feature extraction module is used to realize the high-dimensional representation of the original GNSS station input data including the station coordinates, the station terrain and the station tropospheric delay data to the station features, and obtain the node features. The target grid point feature extraction module is used to realize the geographic spatial feature embedding operation to obtain the grid point features. The GNSS station to grid point mapping module realizes the mapping of the GNSS station to the grid point through the cross attention, the grid point features are used as the query Q to learn the information from the node feature key K / value V, and the zenith tropospheric delay information of the target grid point is generated. Step 3. Train the regional tropospheric delay grid fast generation model based on the training data set, and realize the fast generation of the zenith tropospheric delay information of the target grid point by using the trained regional tropospheric delay grid fast generation model. Wherein during the model training, the zenith tropospheric delay of the GNSS station and the geographic spatial feature information of the GNSS station and the target grid point in the training data set are used as the input of the model, and the zenith tropospheric delay of the target grid point is used as the label.

[0008] In addition, on the basis of the above-mentioned regional tropospheric delay grid fast generation method, the application further proposes a corresponding regional tropospheric delay grid fast generation system, which adopts the following technical scheme: A regional tropospheric delay grid fast generation system, comprising a computer device; wherein the computer device comprises a memory and a processor; the computer program is stored on the memory; when the processor executes the computer program, it is used to realize the steps of the above-mentioned regional tropospheric delay grid fast generation method.

[0009] The application has the following advantages: As described above, this invention discloses a method for rapidly generating regional tropospheric delay grids. This method constructs a rapid generation model for regional tropospheric delay grids to achieve adaptive modeling of GNSS station distribution and terrain features. The model incorporates a graph attention network and a multi-head cross-attention mechanism. The graph attention network, based on a sparse spatial topology constructed from station latitude and longitude, adaptively aggregates node features within a neighborhood, dynamically characterizing the local spatial correlation of station network distributions in different regions. The multi-head cross-attention mechanism establishes a cross-domain feature mapping between the regular grid and station node features, capturing the global variation features of tropospheric delay through parallel attention calculations in multiple subspaces. Geospatial feature embedding further integrates station location and terrain attributes, achieving consistent feature expression under terrain variation conditions. Through this structural design, this invention enhances the ability to capture global feature correlations while maintaining the ability to model local spatial dependencies, achieving adaptive modeling of different GNSS station network distributions and terrain changes, and significantly improving the spatial continuity and prediction accuracy of tropospheric delay. This invention presents a high-precision regional tropospheric delay grid rapid generation model based on GNSS data, which can improve the tropospheric delay accuracy in complex terrain areas and is of great significance for enhancing real-time precision positioning performance in complex terrain. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the method for rapid generation of regional tropospheric delayed grids in an embodiment of the present invention. Figure 2 This is a network structure diagram of the regional tropospheric delayed grid rapid generation model constructed in an embodiment of the present invention; Figure 3 This is a network structure diagram of geospatial feature embedding in an embodiment of the present invention; Figure 4 This is a network structure diagram of the cross-attention mechanism in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 like Figure 1 As shown in the figure, this embodiment describes a method for rapid generation of regional tropospheric delayed grids, which includes the following steps: Step 1. Obtain zenith tropospheric delay and geospatial feature information of GNSS stations and target grid points within the target area to construct a training dataset for model training in Step 2 below.

[0012] The target region can be a selected geographical region or administrative region, such as a certain province or a certain river basin, and the boundary of the target region can be determined by latitude and longitude coordinates, and the embodiments of the present application do not make any limitation in this regard.

[0013] The geographic spatial feature information includes GNSS station coordinates (i.e. latitude and longitude), terrain, and target grid point coordinates (i.e. latitude and longitude), and terrain. The terrain attributes include elevation, slope, slope direction, surface roughness index, and normalized vegetation index NDVI.

[0014] I. The process of obtaining zenith tropospheric delay of GNSS station is as follows: First, the observation equation is constructed according to the GNSS station phase and pseudorange observation value (the process is more conventional, and will not be repeated here), and the linear combination of the observations of two frequencies is used to eliminate the first-order ionospheric delay term; The zenith hydrostatic delay in the zenith tropospheric delay can be accurately calculated by the model; and the zenith wet delay is related to water vapor, which is difficult to be accurately modeled, and is estimated as a parameter, and the zenith wet delay parameter is estimated by using Kalman filtering method; The zenith wet delay and the zenith hydrostatic delay calculated by the model are summed up to obtain the zenith tropospheric delay information of the GNSS station.

[0015] II. The process of obtaining zenith tropospheric delay information of the target grid point is as follows: Based on the EAR5 reanalysis data, the temperature, pressure, specific humidity and other meteorological elements of the ground layer and each isobaric surface are obtained, and the zenith tropospheric delay information of the target grid point is calculated based on the meteorological elements and taken as the label data.

[0016] The zenith tropospheric delay information of the target grid point is a numerical model zenith tropospheric delay grid. The formula for calculating the zenith tropospheric delay grid based on meteorological elements is as follows: ; Wherein, is the zenith tropospheric delay above the grid point surface altitude; is the zenith tropospheric delay between the grid point surface altitude and the altitude of the top isobaric surface; is the zenith hydrostatic delay above the altitude of the top isobaric surface, and since the water vapor content can be ignored at this height, only the zenith hydrostatic delay is calculated and the zenith wet delay is ignored here; is the altitude of the top isobaric surface, is the surface height of the target grid point, is the gas constant of dry air, is the gas constant of wet air; 、 and is the refractive index coefficient; is the atmospheric pressure at the top isobaric surface; is the gravitational acceleration at the top isobaric surface, is the atmospheric temperature, is the air pressure, is the specific humidity, represents the virtual temperature.

[0017] It should be noted that the ERA5 reanalysis data is the latest generation of global meteorological reanalysis data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). EAR5 covers from 1979 to now, with high time (hourly) and high spatial (about 0.25°x0.25°) resolution, containing various meteorological variables such as atmosphere, surface and radiation, etc. The data is generated by assimilating satellite, ground, balloon observations and numerical models, with global consistency and long time series characteristics, widely used in meteorological research, climate analysis, hydrological simulation.

[0018] ERA5 data is divided into ground layer data and isobaric surface layer data. The ground layer mainly contains ground variables such as surface temperature and humidity, while the isobaric surface layer data contains a three-dimensional atmospheric field of 37 standard pressure layers (from 1000 hPa to 1 hPa). The grid time resolution of the zenith tropospheric delay calculated from the above data is 1 hour, and the spatial resolution is 0.25°x0.25°.

[0019] III. The process of obtaining the geographical spatial characteristics of GNSS stations and target grid points is as follows: Based on high-resolution Digital Elevation Model (DEM) data, the elevation, slope, aspect, surface roughness index and normalized vegetation index of GNSS stations and target grid points are calculated.

[0020] The calculation formulas of slope, aspect and surface roughness index are expressed as follows: ; wherein Slope is the slope; Aspect is the aspect; is the slope-related surface roughness; is the rate of change of elevation in the x direction; is the rate of change of elevation in the y direction; i represents the index, corresponding to a single sampling point or grid cell; N represents the total number of sampling points.

[0021] At the same time, the normalized vegetation index NDVI of GNSS stations and target grid points is obtained.

[0022] The normalized vegetation index at the GNSS stations and the target grid points is obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS) normalized vegetation index product.

[0023] After obtaining the above data, a data set containing the zenith tropospheric delay and geographical spatial feature information of the GNSS stations and the target grid points is constructed, which is used as a training data set for the training of the model in the following step 2.

[0024] Step 2. Constructing a terrain-adaptive regional tropospheric delay grid rapid generation model for GNSS station networks, which includes a GNSS station feature extraction module, a target grid point feature extraction module, and a GNSS station to grid point mapping module.

[0025] The input of the model is the zenith tropospheric delay of the GNSS stations and the geographical spatial feature information of the GNSS stations and the target grid points, and the output of the model is the zenith tropospheric delay of the target grid points (numerical mode).

[0026] The GNSS station feature extraction module is used to realize the high-dimensional representation (node feature representation) of the original GNSS station input data including station coordinates, station terrain, and station tropospheric delay data to station features, and obtain node features.

[0027] As shown in Figure 2 , the GNSS station feature extraction module adopts a double-branch feature extraction structure, specifically, the double-branch feature extraction structure includes a first feature extraction branch and a second feature extraction branch.

[0028] The input of the first feature extraction branch is the GNSS station coordinates and terrain attributes.

[0029] The input of the first feature extraction branch is first subjected to graph batching processing, and then the graph batched processed station coordinates and terrain attributes are mapped to a high-dimensional geographical spatial feature vector through geographical spatial feature embedding by a neural network.

[0030] Graph batching processing itself does not directly extract features, but it plays a bridging role in station feature processing, which converts the station (node) features in a batch from a sparse splicing form to a dense aligned tensor.

[0031] The graph batching processing includes determining the batch size and the number of nodes of each sample, finding the number of nodes of the longest sample, cyclically filling each sample node and using a mask to mark valid nodes, and finally generating aligned node features and node masks.

[0032] The graph batching processing step can: A unified structure facilitates batch processing, transforming irregular node data into regular tensors, enabling the model to process the entire batch in parallel on the GPU at once; Provide a mask marker to clearly distinguish which stations are real and which are filled, which can be directly used as the key-padding mask in the subsequent attention layer. No information is lost; the true node features of each sample are preserved intact, with only zeros padded for alignment. It supports dynamic sample node counts, does not require a fixed number of stations for each sample, and offers high data flexibility, making it the fundamental mechanism for achieving GNSS station network self-adaptation.

[0033] Geospatial feature embedding is a process that maps the station coordinates (latitude and longitude) and terrain attributes (elevation, slope, aspect, surface roughness, and NDVI) after batch processing of maps into a highly expressive vector representation through a neural network.

[0034] Among them, station latitude and longitude embedding can encode geographic coordinates into spatial location features, enabling the model to distinguish between "adjacent" and "far away" stations; terrain embedding enables the model to capture the environmental semantics of the surface at that location, such as mountainous areas, plains, NDVI differences, etc.; feature fusion is based on a gating mechanism to fuse station location and terrain information to form a comprehensive geospatial representation.

[0035] When the model migrates to a new region (with different GNSS station distributions and terrain environments), geospatial feature embedding enables the model to construct similar spatial semantics using geographical features, automatically adapting to the new spatial environment and achieving terrain adaptation and GNSS station network adaptation. For example... Figure 3 As shown, the geospatial feature embedding process is as follows: The input for geospatial feature embedding is the station coordinates and terrain attributes after batch processing of the map; where the station coordinates are the station's latitude and longitude, and the station's terrain attributes include elevation, slope, aspect, surface roughness, and NDVI. First, the station coordinates, i.e., the station's latitude and longitude information, are processed through a single-layer fully connected network to extract the station's latitude and longitude features. Then, the extracted station latitude and longitude features are encoded with location codes to obtain the location feature vector. Simultaneously, the terrain attribute information of the measuring station is processed by a multilayer perceptron network to obtain a terrain feature vector; By fusing the station location feature vector and the terrain feature vector based on the gating mechanism, a high-dimensional geospatial feature vector is obtained.

[0036] The second feature extraction branch is used to extract tropospheric delay features from the station. It includes neighborhood feature aggregation based on graph attention network, sparse node linear mapping, and graph batching to generate a high-dimensional tropospheric delay feature representation.

[0037] The input of the second feature extraction branch is the tropospheric delay of the station, i.e., the node, and the sparse spatial graph topology, i.e., the edge index; wherein the edge index is constructed based on the station coordinates, i.e., the station latitude and longitude information, using the k-neighbor method, specifically as follows: For each station, i.e., node, the geographical distance between the node and the remaining nodes, i.e., stations, is calculated, and the k stations closest to the current node are selected as neighbor nodes, and edges are added to each neighbor node to form a connection relationship; Since each node is connected to only k neighbors, rather than establishing a full connection relationship with all nodes, the resulting graph topology is a sparse spatial graph topology. The "sparse" in it refers to the sparse connection relationship between nodes, i.e., the number of edges is much smaller than the number of edges when nodes are fully interconnected, thereby significantly reducing the computational complexity and highlighting the local spatial correlation.

[0038] The information input into the second feature extraction branch is first passed through the graph attention network, wherein the graph attention network is based on the sparse spatial graph topology constructed based on the station latitude and longitude, and adaptively aggregates the node features within the neighborhood range.

[0039] As shown in Figure 2 , the graph attention network further processes on the sparse spatial topology, calculates the feature correlation (such as tropospheric delay feature similarity) between each node and its neighbor nodes to obtain learnable attention weights, realizes adaptive weighted aggregation of neighborhood features, and thus extracts node representations that fuse local spatial dependency relationships.

[0040] Specifically, the specific processing process of the graph attention network is as follows: I. Linear mapping is performed on the node tropospheric delay, as follows: ; Wherein represents each node, represents the original feature of the node, represents the weight, represents the projected feature.

[0041] II. The unnormalized attention score is calculated, as follows: ; Wherein represents the attention score of node to neighbor j; represents the concatenation of the feature vector of node and neighbor j, represents the neighbor set of node , which is determined by the k-neighbor method, is a learnable parameter vector for computing attention scores, is a nonlinear activation function.

[0042] III. Normalizing the attention weights of neighbors as follows: ; where is the normalized attention weight, node the sum of the exponential scores of all neighbors.

[0043] IV. Weighted aggregation of neighbor features as follows: ; where denotes the updated feature of node , denotes the neighbor node feature.

[0044] After aggregation, in order to fuse the multi-head attention output and unify the feature dimension, a sparse node linear mapping is applied to the sparse node feature output by the graph attention network, which maps the high-dimensional feature after multi-head concatenation to a fixed hidden dimension, so as to realize feature dimension reduction and expression compression. Then the feature output after the sparse node linear mapping is processed by the graph batching, which fills the sparse node features of different samples in the batch into a fixed-size dense tensor and generates a node validity mask, so that the subsequent cross-attention module can align the input dimension, to obtain the troposphere delay feature vector.

[0045] Add the high-dimensional geospatial feature vector obtained above to the troposphere delay feature vector to obtain node tokens, which are used as key / value in the cross-attention mechanism below and are queried by grid tokens to realize information propagation from nodes to grid points.

[0046] In the cross-attention mechanism, grid tokens are used as queries to "query" station feature information (node tokens), which contains troposphere delay information.

[0047] The target grid point feature extraction module is used to realize geospatial feature embedding operation to obtain grid point features.

[0048] The target grid point feature extraction only needs to perform a geographic spatial feature embedding operation to obtain a grid point feature representation (grid tokens), wherein the geographic spatial feature embedding is implemented in the same manner as the geographic spatial feature embedding process described above, and will not be described again here.

[0049] The GNSS station to grid point mapping module maps the GNSS station to the grid point through cross-attention, and the grid point feature is taken as the query Q to learn information from the node feature key K / value V, and then the residual connection, layer normalization and feedforward network are used for nonlinear transformation to generate the target grid point zenith troposphere delay information.

[0050] As shown in Figure 4 , the input of the cross-attention mechanism in the present embodiment is the node feature vector and the grid point feature vector obtained by the GNSS station feature extraction module and the target grid point feature extraction module respectively, and the processing process is as follows: I. Generate query (Query), key (Key) and value (Value).

[0051] The grid point feature is taken as the query input, and linear transformation is performed to generate (Query, Q) required for attention calculation, and the node feature is taken as the key input / value input, and linear transformation is performed to generate the key (Key, K) and value (Value, V) required for attention calculation; each attention head has independent Q, K and V linear mapping to capture the feature correlation in different subspaces.

[0052] II. Calculate the attention score.

[0053] First, calculate the basic attention score, that is, the dot product of each head Q and the corresponding K.

[0054] For the hth attention head, given the query matrix and the key matrix , the basic attention score formula is as follows: ; wherein, is the dimension of each attention head, is the basic attention score matrix of the hth attention head.

[0055] Secondly, calculate the structural mask (bias), and the calculation method is the same as that of the basic attention score, but the input is not Q and K, but the original node feature vector and the grid point feature vector, which represents the prior correlation of the grid point to each station node, as a prior attention weight, to help the model pay more attention to the local related nodes, and the formula is expressed as follows: ; wherein, For grid feature representation vectors, The node feature representation vector. This is a structural bias.

[0056] The node mask generated by graph batching is used again to directly overwrite the values ​​of the invalid nodes in the structural bias matrix with negative infinity, so as to ensure that these positions do not participate in attention weighting in subsequent softmax calculations.

[0057] Finally, the structural bias (effective nodes) is added to the basic attention score to obtain the final attention score matrix for each head.

[0058] ; ; in, For the structural bias after padding masking, the first The structural bias or structural mask value of each grid token relative to the j-th node token. The final attention score matrix for the h-th head.

[0059] III. Calculate the multi-head attention weights.

[0060] Perform softmax along the K-dimensional axis on the attention score matrix for each head to obtain the attention weights. This attention weight is used for weighted aggregation of node features, and its specific formula is as follows: ; The role of Softmax is to convert the "Q-K similarity" into a normalized "probability distribution of attention weights".

[0061] IV. Multi-headed attention output.

[0062] Each head uses the attention weight to perform a weighted summation on V to obtain its own head output. The outputs of all attention heads are concatenated and mapped back to the hidden dimension through the output linear layer to form the final multi-head cross-attention output.

[0063] ; .

[0064] in, The weighted output of the h-th head The value of the h-th head. This means concatenating the outputs of all attention heads. The output is a linear transformation matrix, which is the linear mapping result after concatenating all the heads.

[0065] V. Feature enhancement.

[0066] The output of the multi-head attention is processed through a residual connection, layer normalization, and a feed-forward network nonlinear transformation to achieve nonlinear feature enhancement and training stability, and finally generate the zenith tropospheric delay information of the target grid point.

[0067] Step 3. Train the regional tropospheric delay grid fast generation model based on the training data set, and use the trained regional tropospheric delay grid fast generation model to realize the fast generation of the zenith tropospheric delay information of the target grid point.

[0068] During model training, the zenith tropospheric delay of GNSS stations in the training data set and the geographic spatial feature information of GNSS stations and target grid points are used as the input of the model, and the zenith tropospheric delay of the target grid point is used as the label.

[0069] The longitude, latitude, elevation, slope, and slope direction in the geographic spatial feature are constant features, and the normalized vegetation index is updated every 16 days, which can be prepared in advance; the zenith tropospheric delay of GNSS stations can be obtained in real time by step 1, so after the above input data is prepared, the model can quickly generate the zenith tropospheric delay information of the target grid point.

[0070] The training and optimization process of the terrain and GNSS station network adaptive regional tropospheric delay grid fast generation model is as follows: I. Data set division: The entire data set is divided into a training set, a validation set, and a test set, wherein the training set is used for model parameter learning, the validation set is used for adjusting hyperparameters and monitoring model generalization ability, and the test set is used for evaluating model performance.

[0071] II. Set model hyperparameters: Set the model training learning rate to 0.001 and the iteration number to 1000.

[0072] III. Define the loss function and the optimizer: In the above model, the optimizer uses AdamW, and the loss function combines mean square error and KL divergence constraint, and the formula is applied: ; Where is the number of grid points, is the grid point label, is the true value, is the predicted value, is the normalized probability of the true grid at the th grid point, and To predict the probability of the grid at the is a scale factor that controls the contribution of the KL divergence in the loss.

[0073] IV. Loop through the entire training set: Each iteration includes inputting the training data set into the model to calculate the predicted value, calculating the loss according to the predicted value and the true value, backpropagating to calculate the gradient of the loss function on each parameter, updating the model weights according to the gradient, and zeroing the gradient. Accumulate the loss value, repeat the above process until the validation set loss does not decrease in the next several epochs, then stop training in advance.

[0074] When the model is trained, the zenith tropospheric delay of the GNSS station and the geographic spatial feature information of the GNSS station and the target grid point are input into the trained model, and the model generates the zenith tropospheric delay information of the target grid point.

[0075] Specifically, the geographic spatial feature information of the GNSS station and the target grid point, such as latitude, longitude, elevation, slope, slope direction, and normalized vegetation index, is prepared, and the zenith tropospheric delay of the GNSS station is quickly obtained; After the above input data is prepared, it is input into the already trained model, then the zenith tropospheric delay information of the target grid point can be quickly generated.

[0076] Embodiment 2 This embodiment 2 describes a regional tropospheric delay grid fast generation system, which is based on the same inventive concept as the regional tropospheric delay grid fast generation method in embodiment 1 above.

[0077] The regional tropospheric delay grid fast generation system in this embodiment includes a computer device; wherein the computer device includes a memory and a processor; the computer program is stored on the memory; when the processor executes the computer program, it is used to realize the steps of the regional tropospheric delay grid fast generation method described in embodiment 1 above.

[0078] The present application can quickly obtain high-precision tropospheric delay grid of target area, and can more accurately capture small-scale variation characteristics of tropospheric delay in complex terrain area, which has important significance for enhancing real-time precise positioning performance under complex terrain conditions.

[0079] Of course, the above description is only for the preferred embodiments of the present application, and the present application is not limited to the above-mentioned embodiments. It should be noted that any skilled person in the art can make all equivalent substitutions, obvious modifications under the teaching of the present application, which are within the scope of the present application, and should be protected by the present application.​

Claims

1. A method for fast generation of a regional tropospheric delay grid, characterized in that, The method comprises the following steps: Step 1. Obtain the zenith tropospheric delay of GNSS stations and target grid points in the target area and geographical spatial feature information to construct a training data set for model training in the following step 2; Step 2. Construct a terrain and GNSS station network adaptive regional tropospheric delay grid fast generation model, which comprises a GNSS station feature extraction module, a target grid point feature extraction module and a GNSS station to grid point mapping module; The GNSS station feature extraction module is used to realize the high-dimensional representation of the original GNSS station input data including the station coordinates, station terrain and station tropospheric delay data to the station features, and obtain the node features; The target grid point feature extraction module is used to realize the geographical spatial feature embedding operation to obtain the grid point features; The GNSS station to grid point mapping module realizes the mapping of the GNSS station to the grid point through the cross attention, the grid point features are taken as the query Q, the information is learned from the node feature key K / value V, and the zenith tropospheric delay information of the target grid point is generated; Step 3. Train the regional tropospheric delay grid fast generation model based on the training data set, and realize the fast generation of the zenith tropospheric delay information of the target grid point by using the trained regional tropospheric delay grid fast generation model; During the model training, the zenith tropospheric delay of the GNSS stations and the geographical spatial feature information of the GNSS stations and the target grid points in the training data set are taken as the input of the model, and the zenith tropospheric delay of the target grid point is taken as the label.

2. The regional tropospheric delay grid fast generation method according to claim 1, wherein in the step 1, the geographical spatial feature information comprises the GNSS station longitude and latitude, terrain and target grid point longitude and latitude, terrain, wherein the terrain attributes comprise elevation, slope, slope direction, surface roughness index and normalized vegetation index NDVI.

3. The regional tropospheric delay grid fast generation method according to claim 1, wherein in the step 2, the GNSS station feature extraction module adopts a double-branch feature extraction structure; The input of the first feature extraction branch is the GNSS station coordinates and terrain attributes; The input of the first feature extraction branch is firstly subjected to graph batch processing, and then the GNSS station coordinates and terrain attributes subjected to the graph batch processing are mapped to a high-dimensional geographical spatial feature vector through the neural network by the geographical spatial feature embedding; The input of the second feature extraction branch is the station, i.e. the node tropospheric delay and the sparse spatial graph topology, i.e. the edge index; wherein the edge index is constructed based on the station coordinates, i.e. the station longitude and latitude information by using the k-neighbor method; The information input into the second feature extraction branch is firstly subjected to the graph attention network, wherein the graph attention network is based on the sparse spatial graph topology constructed based on the station longitude and latitude to adaptively aggregate the node features within the neighborhood range; After the aggregation, a layer of sparse node linear mapping is applied to the sparse node features output by the graph attention network; Subsequently, the features output after the sparse node linear mapping are subjected to the graph batch processing to obtain the tropospheric delay feature vector; ​ ​ The high-dimensional geographic space feature vector obtained above is added to the tropospheric delay feature vector to obtain a node feature.

4. The regional tropospheric delay grid fast generation method according to claim 3, characterized in that, In step 2, the processing process of the geographic space feature embedding is as follows: The input of the geographic space feature embedding is the station coordinates and terrain attributes after the graph batch processing; the station coordinates are station longitude and latitude, and the station terrain attributes include elevation, slope, aspect, surface roughness and normalized vegetation index NDVI; First, the station coordinate information, i.e. the station longitude and latitude information, is extracted through a single-layer fully connected network to obtain station longitude and latitude features, and then the extracted station longitude and latitude features are added with position encodings to obtain a position feature vector; At the same time, the station terrain attribute information is processed through a multi-layer perception network to obtain a terrain feature vector; Based on the gating mechanism, the station position feature vector and the terrain feature vector are fused to obtain a high-dimensional geographic space feature vector.

5. The regional tropospheric delay grid fast generation method according to claim 3, characterized in that, The edge index is obtained based on the station coordinates, i.e. the longitude and latitude information, by using a k-neighbor method, and specifically: For each station, i.e. node, the geographical distance between the node and the remaining nodes, i.e. stations, is calculated, and the k stations closest to the current node are selected as neighbor nodes, and edges are added to each neighbor node to form a connection relationship; Since each node is connected to only k neighbors, the obtained graph topology is a sparse spatial graph topology.

6. The regional tropospheric delay grid fast generation method according to claim 3, characterized in that, In step 2, the specific processing process of the graph attention network is as follows: I. Linearly mapping the node tropospheric delay, as follows: ； wherein denotes each node, denotes the original features of a node, denotes the weights, denotes the projected features; II. Calculating the unnormalized attention score, as follows: ； wherein representative node attention score for neighbor j; representative node concatenation with neighbor j feature vector, representative node neighbor set of the representative node determined by k- nearest neighbor method, is a learnable parameter vector for computing attention score, is a nonlinear activation function; III. Normalizing the neighbor attention weight, as follows: ； wherein is a normalized attention weight, node all the sum of the exponentiated scores of the neighbors. IV. Weighted aggregation of neighbor features, as follows: ； wherein representative node updated features, representative neighbor node features.

7. The regional tropospheric delay grid fast generation method according to claim 1, characterized in that, In step 2, the processing process of the cross-attention is as follows: I. Taking the grid point feature as the query input, generating the Q required for attention calculation through linear transformation, and taking the node feature as the key input / value input, generating the key K and value V required for attention calculation through linear transformation; Each attention head has independent Q, K and V linear mappings to capture the feature correlation in different subspaces; II. First, the basic attention score is calculated, i.e. the dot product of Q and the corresponding K of each head; Second, the structural bias is calculated, which is calculated in the same way as the basic attention score, but the input is not Q and K, but the original node feature vector and grid point feature vector; Third, the node mask generated by the graph batch processing is used to directly cover the value of the invalid nodes filled in the structural bias matrix as negative infinity to ensure that these positions do not participate in attention weighting in the subsequent softmax calculation; Finally, the structural bias and the basic attention score are added to obtain the final attention score matrix of each head. III. Perform softmax on the attention score matrix of each head along the K dimension to get the attention weight; IV. Each head uses the attention weight to perform weighted summation on V to get the respective head output, concatenates the outputs of all attention heads, and maps back to the hidden dimension through the output linear layer to form the final multi-head cross-attention output; V. The output of the multi-head attention is processed through the residual connection, layer normalization and feedforward network nonlinear transformation to realize nonlinear feature enhancement and training stability, and finally generate the target grid point tropospheric delay information.

8. The regional tropospheric delay grid fast generation method according to claim 1, characterized in that, In step 3, define the loss function as follows : ； wherein is the number of grid points, is the grid point label, is the true value, is the predicted value, is the normalized probability of the true grid at the th grid point, is the probability of the predicted grid at the th grid point, is the scaling factor.

9. The regional tropospheric delay grid fast generation method according to claim 1, characterized in that, In step 3, after the model is trained, the zenith tropospheric delay of the GNSS station and the geographic spatial feature information of the GNSS station and the target grid point are input into the trained model, and the model generates the zenith tropospheric delay information of the target grid point.

10. A regional tropospheric delay grid fast generation system, comprising a computer device; wherein the computer device comprises a memory and a processor; a computer program is stored on the memory; characterized in that, When the processor executes the computer program, the steps for implementing the regional tropospheric delay grid fast generation method according to any one of claims 1 to 9 are implemented.

Images