A method and device for reconstructing surface rainfall distribution based on a spatio-temporal graph neural network

Abstract

The application discloses a kind of face rainfall distribution reconstruction method and device based on space-time graph neural network, wherein the method first constructs space-time graph structure model, the complex spatial correlation between links is established by the adjacency matrix construction method of multi-factor fusion, and the graph structure data object containing space-time dimension is constructed;Further, hierarchical space-time graph neural network is used for feature extraction and fusion, and the space-time depth feature representation of the fused rainfall field is obtained;In the model optimization process, the microwave rain attenuation physical model is introduced as a constraint condition simultaneously, and the physical consistency loss function is used to ensure that the inversion result conforms to the electromagnetic wave propagation law;Based on the grid rainfall intensity and its uncertainty evaluation output by the network, high-resolution surface rainfall distribution products are generated through spatial interpolation and quality control;Finally, a real-time inference system and feedback mechanism are established to dynamically update and output the rainfall field, providing an intelligent technical solution from "line measurement" to "surface reconstruction" for meteorological and hydrological monitoring.

Description

Technical Field

[0001] This invention relates to the field of hydrological and meteorological monitoring technology, and more specifically, to a method and apparatus for reconstructing areal rainfall distribution based on spatiotemporal graph neural networks. Background Technology

[0002] Accurate, high-resolution rainfall monitoring is crucial for flood warning and water resource management. Commercial Microwave Links (CMLs) networks utilize communication signal attenuation to retrieve rainfall, effectively compensating for the shortcomings of traditional rain gauges, such as poor spatial representativeness and near-surface blind spots of weather radar, thanks to their wide coverage and low cost. However, CML observations essentially only provide a "line-average" rainfall intensity along the path from transmitter to receiver, and cannot directly reveal the spatial distribution of areal rainfall within a region.

[0003] Currently, the industry mainly uses spatial interpolation (such as inverse distance weighting and Kriging) or linear tomography techniques to attempt rainfall field reconstruction. However, these methods typically simplify microwave links into independent observation units, ignoring the complex spatial topological relationships and intersections between link paths in dense networks. When dealing with convective rainfall with extremely uneven spatial distribution, they are prone to losing the location of rainfall centers and texture details. Furthermore, most existing inversion algorithms focus on static spatial mapping, lacking in-depth analysis of the spatiotemporal dynamic evolution of rain cluster movement, formation, and dissipation, resulting in significant lag or morphological distortion in the results. In addition, conventional pure data-driven methods often deviate from the constraints of microwave propagation physical mechanisms (such as the power-law relationship of rain decay) during the inversion process, making the model prone to producing "ill-conditioned solutions" that violate physical laws under sparse data or anomalous interference, and failing to provide reliable uncertainty assessments. Therefore, there is an urgent need for a areal rainfall distribution reconstruction technique that can integrate link network topological characteristics, spatiotemporal dynamic laws, and physical model constraints. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and device for reconstructing area rainfall distribution based on spatiotemporal graph neural networks. This method can break through the limitation of traditional microwave links that can only provide path average rainfall intensity ("line average"), and realize the leap from "line" to "area" to generate a high-resolution spatial rainfall intensity distribution map.

[0005] To achieve the above objectives, a first aspect of the present invention provides a method for reconstructing areal rainfall distribution based on a spatiotemporal graph neural network, comprising:

[0006] Based on the spatial distribution characteristics of microwave links and the geographic information of the monitoring area, a spatiotemporal graph structure model containing node characteristics and spatial relationships is constructed. The spatial relationships between links are established through a multi-factor fusion adjacency matrix construction method, and spatiotemporal graph data objects are generated.

[0007] Based on the generated spatiotemporal graph data object, a hierarchical spatiotemporal graph neural network is used for feature extraction and fusion to obtain the spatiotemporal depth feature representation of the fused precipitation field.

[0008] Based on the spatiotemporal depth feature representation of the fused rainfall field, combined with the constraints of the microwave rain attenuation physical model, the rainfall intensity estimation with uncertainty awareness is carried out, and the physical consistency loss function is used to ensure that the inversion result conforms to the electromagnetic wave propagation law, so as to obtain a physically reasonable grid rainfall intensity value.

[0009] Based on the obtained physically reasonable grid rainfall intensity values ​​and their uncertainty assessment, spatial interpolation and product quality control are performed to generate a high-resolution two-dimensional rainfall intensity distribution map with a credibility assessment.

[0010] Based on a high-resolution two-dimensional rainfall intensity distribution map, a real-time inference system and feedback mechanism are established to dynamically update and output the rainfall field.

[0011] In one embodiment, the method further includes:

[0012] The quality of the collected link data is controlled using a pre-set data quality control system.

[0013] Based on the data after quality control, a multi-dimensional feature vector containing physical meaning is constructed for each link, represented as follows:

[0014]

[0015] in, Indicates the first The link is The feature vector of the spatiotemporal graph data object at any given time. Indicates the first The link at time Total path decay, Indicates the link's operating frequency. Indicates the link path length. Indicates the polarization mode. Indicates temperature. Indicates relative humidity. Indicates air pressure. Indicates wind speed. Indicates wind direction. Indicates the latitude of the link's center point. Indicates the longitude of the link's center point. This indicates the average altitude of the link's location;

[0016] Batch normalization based on Z-score is performed on all continuous features, and one-hot encoding is performed on categorical features to form a normalized node feature matrix.

[0017] In one implementation, based on the spatial distribution characteristics of the microwave link and the geographical information of the monitoring area, a spatiotemporal graph structure model including node characteristics and spatial relationships is constructed, including:

[0018] Based on a standardized node feature matrix, the monitoring area is modeled as a dynamically evolving graph structure, where nodes in the graph structure represent microwave links and edges represent spatial relationships between links.

[0019] Based on graph structure, construct including A sequence of spatiotemporal graphs with consecutive time steps is used as a multi-scale adjacency matrix.

[0020] In one implementation, a spatial association between links is established using a multi-factor fusion adjacency matrix construction method to generate a spatiotemporal graph data object, including:

[0021] Based on the defined graph structure node information, a weighted adjacency matrix reflecting various spatial association patterns is constructed, represented as:

[0022]

[0023] In the formula: Represents the weights of the adjacency matrix; This represents the Euclidean distance between the center points of node i and node j. and Represents the link direction vector; and Indicates the link's operating frequency; Represents the weighting coefficients, satisfying ; and These represent the Gaussian kernel bandwidth parameters for distance and frequency, respectively.

[0024] Construct a k-nearest neighbor graph and a Gaussian kernel graph, and combine the k-nearest neighbor graph, Gaussian kernel graph and multi-scale adjacency matrix into a multi-scale graph structure set to handle features at different spatial scales;

[0025] By combining the spatiotemporal feature tensor with a set of multi-scale adjacency matrices, a spatiotemporal graph data object is formed.

[0026] In one implementation, based on the generated spatiotemporal graph data object, a hierarchical spatiotemporal graph neural network is used for feature extraction and fusion to obtain a fused spatiotemporal depth feature representation of the precipitation field, including:

[0027] A multi-layer graph attention network is used as a spatial feature extractor to extract features from the generated spatiotemporal graph data objects to form feature vectors;

[0028] Spatial dependencies are captured using a graph attention mechanism, and graph attention coefficients are calculated.

[0029] By combining graph attention coefficients and stacking multi-layer graph attention networks, information from multi-hop neighbors is aggregated to capture the spatial correlation of rainfall fields and output spatially enhanced features.

[0030] A gated temporal convolutional network with residual connections is used to process the time series features of each node to obtain the time features.

[0031] A cross-modal attention mechanism is used to adaptively fuse spatial enhancement features and temporal features to obtain a fused spatiotemporal depth feature representation of the precipitation field.

[0032] In one implementation, based on the spatiotemporal depth feature representation of the fused rainfall field, and constrained by a microwave rain attenuation physical model, uncertainty-aware rainfall intensity estimation is performed. A physical consistency loss function is used to ensure that the inversion results conform to the electromagnetic wave propagation laws, resulting in physically reasonable grid rainfall intensity values, including:

[0033] A physical model of the attenuation of microwave signals by rainfall is established to describe the quantitative relationship between rainfall intensity and microwave attenuation.

[0034] The physical laws of microwave propagation are introduced as soft constraints into the optimization process to establish a physical model of attenuation-rainfall intensity.

[0035] The physical consistency loss of the physical model of attenuation-rainfall intensity is used to characterize the difference between the model-predicted attenuation and the measured attenuation.

[0036] Based on rainfall intensity estimation and physical consistency loss, a loss function containing multiple optimization objectives is constructed to guide the model to learn a solution that simultaneously satisfies data fitting and physical constraints, resulting in a trained model.

[0037] The trained model is used to estimate the rainfall intensity and its uncertainty at each node location.

[0038] In one implementation, a real-time inference system and feedback mechanism are established based on a high-resolution two-dimensional rainfall intensity distribution map to dynamically update and output the rainfall field, including:

[0039] Based on the trained model, uncertainty is estimated using Monte Carlo Dropout, and Dropout activation is maintained during the inference phase to perform multiple random forward propagations.

[0040] Based on the generated high-resolution two-dimensional rainfall intensity distribution map and its uncertainty assessment, the inverted rainfall field product is assimilated with the numerical weather prediction model. The initial field is optimized by a four-dimensional variational assimilation method. At the same time, a feedback mechanism is established to feed back the assimilation effect assessment and actual observation verification results to the model training process, forming a closed-loop optimization system.

[0041] Based on the same inventive concept, a second aspect of the present invention provides a device for reconstructing areal rainfall distribution based on a spatiotemporal graph neural network, comprising:

[0042] The data acquisition and preprocessing module is used to construct a spatiotemporal graph structure model containing node characteristics and spatial relationships based on the spatial distribution characteristics of microwave links and the geographic information of the monitoring area. It establishes spatial relationships between links through a multi-factor fusion adjacency matrix construction method and generates spatiotemporal graph data objects.

[0043] The spatiotemporal graph construction module is used to extract and fuse features based on the generated spatiotemporal graph data objects using a hierarchical spatiotemporal graph neural network, to obtain the spatiotemporal depth feature representation of the fused precipitation field.

[0044] The model building and training module is used to estimate rainfall intensity based on the spatiotemporal depth feature representation of the fused rainfall field, combined with the constraints of the microwave rain attenuation physical model, and to ensure that the inversion results conform to the electromagnetic wave propagation law through the physical consistency loss function, so as to obtain physically reasonable grid rainfall intensity values.

[0045] The physical constraint inversion module is used to perform spatial interpolation and product quality control based on the obtained physically reasonable grid rainfall intensity values ​​and their uncertainty assessment, and to generate a high-resolution two-dimensional rainfall intensity distribution map with a credibility assessment.

[0046] The product generation and output module is used to establish a real-time inference system and feedback mechanism based on a high-resolution two-dimensional rainfall intensity distribution map, and to dynamically update and output the rainfall field.

[0047] Based on the same inventive concept, a third aspect of the present invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, provides the method for reconstructing areal rainfall distribution based on a spatiotemporal graph neural network as described in the first aspect.

[0048] Based on the same inventive concept, the fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for reconstructing surface rainfall distribution based on spatiotemporal graph neural networks as described in the first aspect.

[0049] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:

[0050] 1. Strong technological breakthrough and significant cost-effectiveness.

[0051] This invention introduces spatiotemporal graph neural networks into the field of microwave link rainfall inversion. Through an end-to-end deep learning framework, it achieves a direct mapping from raw attenuation data to rainfall field distribution, overcoming the dependence of traditional tomographic methods on linearization and iterative solutions. This method can adaptively learn complex spatiotemporal patterns of rainfall fields and maintain high inversion accuracy even under sparse network conditions, significantly improving the practicality and reliability of the technology.

[0052] 2. The system has good compatibility and can form a collaborative monitoring advantage.

[0053] This invention constructs a deep learning inversion framework for physical model constraints and uncertainty perception, addressing the low reliability of purely data-driven methods under physical consistency and sparse networks. The system fully utilizes existing commercial microwave communication network resources, achieving "value-added services" for rainfall monitoring through "intelligent analysis" of communication signals, demonstrating high cost-effectiveness and widespread application value. The reconstructed high-resolution rainfall field can effectively complement traditional observation methods such as radar and rain gauges, constructing an integrated "point-line-surface" intelligent monitoring network. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of the surface rainfall distribution reconstruction method based on spatiotemporal graph neural network in an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of a multi-layer graph attention network architecture in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram illustrating the variation of the attenuation coefficient with rainfall intensity and frequency in an embodiment of the present invention;

[0058] Figure 4 This is a structural diagram of the areal rainfall distribution reconstruction device based on a spatiotemporal graph neural network in an embodiment of the present invention. Detailed Implementation

[0059] This invention first constructs a spatiotemporal graph structure model containing node features and spatial relationships based on the spatial distribution characteristics of a densely deployed microwave link network and the geographical environment of the monitoring area. Complex spatial relationships between links are established through a multi-factor fusion adjacency matrix construction method. On this basis, multi-link temporal attenuation data and meteorological background field parameters are simultaneously collected to construct a graph structure data object containing spatiotemporal dimensions. Then, a hierarchical spatiotemporal graph neural network is used for end-to-end feature learning, capturing spatial dependencies through a graph attention mechanism. A gated temporal convolutional network is combined to model the dynamic evolution of rainfall, achieving in-depth mining of the spatiotemporal features of the rainfall field. During the model optimization process… In this process, a microwave rain attenuation physical model is introduced as a constraint, and a physical consistency loss function is used to ensure that the inversion results conform to the electromagnetic wave propagation law. Based on the gridded rainfall intensity and its uncertainty assessment from the network output, high-resolution areal rainfall distribution products are generated through spatial interpolation and quality control. Finally, an operational mechanism including Monte Carlo uncertainty quantification and product assimilation is established. Multi-dimensional quality assessment is used to ensure the reliability of the inversion results and to realize real-time dynamic updates of the system. The output is a two-dimensional distribution product that can accurately characterize the spatial heterogeneity of rainfall, providing an intelligent technical solution for meteorological and hydrological monitoring from "line measurement" to "area reconstruction".

[0060] Example 1

[0061] This embodiment provides a method for reconstructing areal rainfall distribution based on a spatiotemporal graph neural network. Please refer to [link to relevant documentation]. Figure 1 ,include:

[0062] S1: Based on the spatial distribution characteristics of microwave links and the geographic information of the monitoring area, a spatiotemporal graph structure model containing node characteristics and spatial relationships is constructed. The spatial relationships between links are established through a multi-factor fusion adjacency matrix construction method, and spatiotemporal graph data objects are generated.

[0063] The method also includes:

[0064] The quality of the collected link data is controlled using a pre-set data quality control system.

[0065] Based on the data after quality control, a multi-dimensional feature vector containing physical meaning is constructed for each link, represented as follows:

[0066] (1)

[0067] in, Indicates the first The link is The feature vector of the spatiotemporal graph data object at any given time. Indicates the first The link at time Total path decay, Indicates the link's operating frequency. Indicates the link path length. Indicates the polarization mode. Indicates temperature. Indicates relative humidity. Indicates air pressure. Indicates wind speed. Indicates wind direction. Indicates the latitude of the link's center point. Indicates the longitude of the link's center point. This indicates the average altitude of the link's location;

[0068] Batch normalization based on Z-score is performed on all continuous features, and one-hot encoding is performed on categorical features to form a normalized node feature matrix.

[0069] In the specific implementation process, the spatial location information of all M microwave links within the monitoring area is first acquired, including the link's unique identifier, the precise geographic coordinates of the transmitter and receiver, operating frequency, polarization, timestamp, and received signal level (RSL). Simultaneously, synchronized regional meteorological background field data, including parameters such as temperature, relative humidity, atmospheric pressure, wind speed, and wind direction, is obtained from the meteorological observation network, with the temporal resolution consistent with the microwave link data. Furthermore, point observation data from the ground rain gauge network and quantitative precipitation estimation (QPE) products from the weather radar system are collected as ground truth labels for supervised learning.

[0070] A three-tiered data quality control system is established. At the first tier, an anomaly detection algorithm based on the median absolute deviation (MAD) within a sliding window is used to identify and mark abnormal data links affected by equipment failure or external interference in real time. The anomaly criteria are as follows:

[0071] (2)

[0072] In the formula: This represents the measured value of the received signal level at the a-th time step; represents the median of the RSL within the window; MAD represents the median absolute deviation calculation function; 'a' represents the index of the data point in the time series.

[0073] The second stage involves applying an adaptive Savitzky-Gore filter to the link data that has passed the initial screening for signal smoothing. This filter preserves the signal trend while suppressing noise by performing local polynomial fitting within a sliding window.

[0074] (3)

[0075] In the formula: This represents the a-th data point after filtering, where a is the time series index of the original signal; represents the (b+a)th data point in the original signal; b represents the local offset index within the sliding window (ranging from -(W-1) / 2 to (W-1) / 2). The coefficients at the corresponding offset position b represent the polynomial fitting coefficients, determined by the least squares method; W is the total length of the filter window, which is an odd number and is dynamically adjusted according to the rainfall intensity.

[0076] The third level involves identifying and correcting for attenuation caused by non-precipitation factors based on meteorological background data. This includes calculating the attenuation caused by gas absorption.

[0077] (4)

[0078] In the formula: This indicates the total attenuation of gas absorption; This indicates the decay caused by oxygen. This indicates the attenuation caused by water vapor; and This represents the specific decay coefficient between oxygen and water vapor; and This indicates the density of oxygen and water vapor.

[0079] Pure Rain-Induced Attenuation Component The following formula is used to calculate:

[0080] (5)

[0081] In the formula: This represents the measured total attenuation; L is the link length. This represents the system's inherent attenuation reference value.

[0082] Based on the data after quality control, a multi-dimensional feature vector containing physical meaning is constructed for each link, as shown in formula (1). Batch standardization based on Z-score is performed on all continuous features, and one-hot encoding is performed on categorical features to form a standardized node feature matrix. .

[0083] S1 can be implemented in the following way:

[0084] S1.1: Based on the obtained standardized node feature matrix, the monitoring area is modeled as a dynamically evolving graph structure. , where the set of nodes correspond A microwave link, where X represents the normalized node feature matrix and the edge set. This indicates the spatial relationships between links. It constructs a system containing... A sequence of spatiotemporal graphs for consecutive time steps:

[0085] (6)

[0086] t represents the current time step index; This represents the length of the time window for the input sequence. The graph structure at each time step shares the same set of nodes and edges, but the node features evolve over time, enabling the simultaneous capture of the spatial topology and temporal dynamics of the precipitation field.

[0087] S1.2: Based on the defined graph structure node information, a weighted adjacency matrix reflecting various spatial association patterns is constructed. It considers not only Euclidean distance but also directional similarity and frequency correlation, forming a comprehensive edge weight calculation model.

[0088] (7)

[0089] In the formula: Represents the weights of the adjacency matrix; Let be the Euclidean distance between the center points of node i and node j; and Represents the link direction vector; and Indicates the link's operating frequency; Represents the weighting coefficients, satisfying ; and The Gaussian kernel bandwidth parameters (standard deviation) represent distance and frequency, respectively.

[0090] It should be noted that links and nodes are corresponding entities. Nodes in the model represent links in the physical world, and edges in the model represent the spatial relationships between links.

[0091] S1.3: To handle features at different spatial scales, additional k-nearest neighbor (k-NN) graphs and Gaussian kernel graphs are constructed to form a multi-scale graph structure set.

[0092] S1.4: Based on spatiotemporal feature tensors and multi-scale adjacency matrices Spatiotemporal feature tensor (N represents the total number of nodes; F represents the number of feature dimensions for each node;) (representing the length of the time series) and the set of multi-scale adjacency matrices Combined, they form an enhanced spatiotemporal graph data object.

[0093] To improve the robustness of the model, a variety of data augmentation techniques are applied during the training phase, including random edge dropping (randomly removing a certain proportion of edges), node feature masking (randomly setting some node features to zero), and time series jittering (adding small random noise) to simulate common data missing and noise situations in real-world applications.

[0094] S2: Based on the generated spatiotemporal graph data object, a hierarchical spatiotemporal graph neural network is used for feature extraction and fusion to obtain the spatiotemporal depth feature representation of the fused precipitation field.

[0095] Specifically, S2 includes:

[0096] S2.1: A multi-layer graph attention network is used as a spatial feature extractor to extract features from the generated spatiotemporal graph data objects to form a feature vector;

[0097] S2.2: Capture spatial dependencies through graph attention mechanism and calculate graph attention coefficients;

[0098] S2.3: Combine graph attention coefficients and stack multi-layer graph attention networks to aggregate information from multi-hop neighbors, capture the spatial correlation of rainfall fields, and output spatially enhanced features;

[0099] S2.4: A gated temporal convolutional network with residual connections is used to process the time series features of each node to obtain the time features;

[0100] S2.5: A cross-modal attention mechanism is used to adaptively fuse spatial enhancement features and temporal features to obtain a fused spatiotemporal depth feature representation of the precipitation field.

[0101] In the specific implementation process, based on the spatiotemporal graph data object output in step S2, a multi-layer graph attention network (GAT) is used as the spatial feature extractor, such as... Figure 2 As shown. Each layer contains multiple independent attention heads, each learning node relationships from a different representation subspace. For the l-th layer, the update formula for node i is:

[0102] (8)

[0103] In the formula: This represents the output feature vector of node i in the (l+1)th layer; This indicates vector concatenation; For the number of heads; It is a non-linear activation function; This indicates that node j belongs to the set of neighbors of node i (including node i itself). The normalized attention coefficient for the k-th attention head; Let be the learnable linear transformation weight matrix for the k-th attention head; Let be the input feature vector of node j in the l-th layer.

[0104] For graph attention coefficient calculation:

[0105] (9)

[0106] In the formula: A vector of learnable parameters representing the attention mechanism; It is a learnable linear transformation weight matrix; and Let be the feature vectors of node i and node j.

[0107] By stacking multiple layers of GAT, nodes can aggregate information from multi-hop neighbors, effectively capturing the spatial correlation of rainfall fields.

[0108] Based on the spatial augmentation features output by the GAT network, a gated temporal convolutional network with residual connections (Gated TCN) is used to process the time-series features of each node. The network consists of multiple stacked temporal convolutional blocks, each containing dilated causal convolutions, a gating mechanism, and residual connections. The gated temporal convolution function is as follows:

[0109] (10)

[0110] In the formula: It is the output of the time-controlled convolutional block; This represents the feature sequence input to this convolutional layer; and These are learnable convolutional kernel weights; It is a dilated causal convolution operation; It is element-wise multiplication; It is the sigmoid activation function.

[0111] Based on spatiotemporal features, a cross-modal attention mechanism is designed to achieve adaptive fusion of spatial features (Hspatial) and temporal features (Htemporal). First, the correlation between the two modalities is calculated through cross-attention:

[0112] (11)

[0113] In the formula: C represents the correlation matrix; and , respectively, are the projection weight matrices for the query and the key; T represents the matrix transpose; d represents the dimension of the feature vector.

[0114] Then, feature fusion is performed based on the correlation matrix C:

[0115] (12)

[0116] In the formula: The projection matrix is ​​learnable; These are adaptive weight parameters.

[0117] S3: Based on the spatiotemporal depth feature representation of the fused rainfall field, combined with the constraints of the microwave rain attenuation physical model, the rainfall intensity is estimated with uncertainty awareness, and the physical consistency loss function is used to ensure that the inversion result conforms to the electromagnetic wave propagation law, so as to obtain a physically reasonable grid rainfall intensity value.

[0118] Specifically, S3 includes:

[0119] S3.1: Establish a physical model for the attenuation of microwave signals by rainfall to describe the quantitative relationship between rainfall intensity and microwave attenuation;

[0120] S3.2: Introduce the physical laws of microwave propagation as a soft constraint into the optimization process to establish a physical model of attenuation-rainfall intensity;

[0121] S3.3: Physical consistency loss of the physical model of attenuation-rainfall intensity is used to characterize the difference between the model-predicted attenuation and the measured attenuation;

[0122] S3.4: Based on rainfall intensity estimation and physical consistency loss, a loss function containing multiple optimization objectives is constructed to guide the model to learn a solution that simultaneously satisfies data fitting and physical constraints, thus obtaining the trained model;

[0123] S3.5: Use the trained model to estimate the rainfall intensity and its uncertainty at each node location.

[0124] In the specific implementation process, based on the ITU-R Recommendation P.838-3, a physical model of the attenuation of microwave signals by rainfall was established. The attenuation coefficient varies with rainfall intensity and frequency as follows: Figure 3 As shown. This model describes the quantitative relationship between rainfall intensity and microwave attenuation:

[0125] (13)

[0126] In the formula: This is the attenuation coefficient per unit path length; Rainfall intensity; and For parameters.

[0127] The physical laws governing microwave propagation are introduced as soft constraints into the optimization process to establish a physical model of attenuation versus rainfall intensity:

[0128] (14)

[0129] In the formula: This represents the theoretical total attenuation value of the i-th link, derived by back-calculating the rainfall intensity predicted by the model. Indicates the length of the i-th link; and This represents parameters related to frequency, polarization, and temperature.

[0130] Physical consistency loss is defined as the difference between the model-predicted decay and the measured decay:

[0131] (15)

[0132] In the formula: This represents a loss of physical consistency. Hyperparameters for controlling the strength of physical constraints; This represents the total attenuation measured for the i-th link; Let be the length of the i-th link; Frequency and polarization-related parameters defined for the ITU-R P.838 model; This represents the rainfall intensity predicted by the model.

[0133] Based on rainfall intensity estimation and physical constraints, a loss function with multiple optimization objectives is constructed to guide the model in learning a solution that simultaneously satisfies data fitting and physical constraints. Total loss function:

[0134] (16)

[0135] In the formula: For data fitting loss; For smoothness loss; This represents sparsity loss.

[0136] After obtaining the trained model, the fused features (Hfusion) are input into the model's probability output layer, simultaneously estimating the rainfall intensity and its uncertainty at each node location. A heteroscedastic regression framework is employed, and the output layer produces two values: the mean rainfall intensity. Sum of logarithmic variance :

[0137] (17)

[0138] In the formula: and These represent the learnable weight matrix and bias vector of the output layer, respectively. This represents the feature vector of the i-th node after cross-modal fusion; and Let represent the mean and variance of the predicted distribution, respectively. The final rainfall intensity estimate is modeled as a Gaussian distribution: ,in Let represent the estimated rainfall intensity random variable for the i-th link.

[0139] S4: Based on the obtained physically reasonable grid rainfall intensity values ​​and their uncertainty assessment, spatial interpolation and product quality control are performed to generate a high-resolution two-dimensional rainfall intensity distribution map with a credibility assessment.

[0140] S5: Based on a high-resolution two-dimensional rainfall intensity distribution map, a real-time inference system and feedback mechanism are established to dynamically update and output the rainfall field.

[0141] Specifically, S5 includes:

[0142] Based on the trained model, uncertainty is estimated using Monte Carlo Dropout, and Dropout activation is maintained during the inference phase to perform multiple random forward propagations.

[0143] Based on the generated high-resolution two-dimensional rainfall intensity distribution map and its uncertainty assessment, the inverted rainfall field product is assimilated with the numerical weather prediction model. The initial field is optimized by a four-dimensional variational assimilation method. At the same time, a feedback mechanism is established to feed back the assimilation effect assessment and actual observation verification results to the model training process, forming a closed-loop optimization system.

[0144] In the specific implementation process, based on the model trained in step 4, uncertainty is estimated using Monte Carlo Dropout. Dropout activation is maintained during the inference phase, and multiple random forward propagations are performed. The predicted mean and uncertainty are:

[0145] (18)

[0146] In the formula: This represents the predicted average rainfall intensity. To predict uncertainty (variance); Total number of samples taken in Monte Carlo; This is the current sampling sequence number; This represents the prediction result of the m-th random forward propagation.

[0147] Based on the generated areal rainfall distribution product and its uncertainty assessment, the retrieved rainfall field product is assimilated with the numerical weather prediction model, and the initial field is optimized using a four-dimensional variational assimilation method. Simultaneously, a feedback mechanism is established to feed back the assimilation effect evaluation and actual observation verification results to the model training process, forming a closed-loop optimization system to continuously improve the inversion accuracy and reliability.

[0148] Previous studies on inverting the spatial distribution of rainfall using microwave link networks typically employ spatial interpolation methods based on path-averaged rainfall rates or simple linear regression models. These methods, by their very nature, fail to overcome the inherent limitations of "path integral measurement," treating each link merely as an independent point observation while ignoring the spatial intersections and topological relationships between links in a dense network. Furthermore, traditional methods lack the ability to model the complex spatiotemporal evolution of rainfall fields, failing to effectively capture the spatial heterogeneity and temporal dynamics of rainfall systems. This results in inversion results with problems such as blurred spatial details, inaccurate rainfall center location, and distorted intensity estimation. Moreover, most existing methods adopt a purely data-driven approach, lacking the consideration of microwave propagation physics as a constraint during the inversion process. This makes them less robust in handling ill-conditioned inversion problems, easily leading to physically inconsistent solutions. This invention not only introduces spatiotemporal graph neural network technology to construct a microwave link network as an intelligent detection system with spatiotemporal perception capabilities, but also innovatively embeds the microwave rain attenuation physical model as a constraint into a deep learning framework, establishing a physics-guided spatiotemporal graph neural network inversion model. Simultaneously, it constructs a complete optimization and evaluation system including multi-task loss functions and Monte Carlo uncertainty quantification, forming a full-link technical solution capable of accurately capturing spatial details of rainfall, strictly ensuring physical rationality, and possessing adaptive optimization and intelligent quality control capabilities. This enables reliable, accurate, and intelligent reconstruction of high spatiotemporal resolution surface rainfall distribution.

[0149] Example 2

[0150] Based on the same inventive concept, this embodiment discloses a device for reconstructing areal rainfall distribution based on a spatiotemporal graph neural network. Please refer to [link to relevant documentation]. Figure 4 ,include:

[0151] The data acquisition and preprocessing module 210 is used to construct a spatiotemporal map structure model containing node features and spatial relationships based on the spatial distribution characteristics of microwave links and the geographic information of the monitoring area. It establishes spatial relationships between links through a multi-factor fusion adjacency matrix construction method and generates spatiotemporal map data objects.

[0152] The spatiotemporal graph construction module 220 is used to extract and fuse features based on the generated spatiotemporal graph data object using a hierarchical spatiotemporal graph neural network, and obtain the spatiotemporal depth feature representation of the fused precipitation field.

[0153] The model building and training module 230 is used to estimate the rainfall intensity based on the spatiotemporal depth feature representation of the fused rainfall field, combined with the constraint of the microwave rain attenuation physical model, and to ensure that the inversion results conform to the electromagnetic wave propagation law through the physical consistency loss function, so as to obtain a physically reasonable grid rainfall intensity value.

[0154] The physical constraint inversion module 240 is used to perform spatial interpolation and product quality control based on the obtained physically reasonable grid rainfall intensity values ​​and their uncertainty assessment, and to generate a high-resolution two-dimensional rainfall intensity distribution map with a credibility assessment.

[0155] Product generation and output module 250 is used to establish a real-time inference system and feedback mechanism based on a high-resolution two-dimensional rainfall intensity distribution map, and to dynamically update and output the rainfall field.

[0156] like Figure 4 As shown, the surface rainfall distribution reconstruction device 20 based on spatiotemporal graph neural network includes: a data acquisition and preprocessing module 210, a spatiotemporal graph construction module 220, a model construction and training module 230, a physical constraint inversion module 240, and a product generation and output module 250.

[0157] The data acquisition and preprocessing module 210 is used to acquire multi-source observation data of the microwave link network and perform data quality control and feature engineering processing; the spatiotemporal graph construction module 220 is used to construct a spatiotemporal graph structure based on the preprocessed data, including node feature definition and spatial adjacency matrix calculation; the model construction and training module 230 is used to construct a hierarchical spatiotemporal graph neural network model and perform end-to-end training; the physical constraint inversion module 240 is used to apply physical constraints during model training and inference to ensure that the inversion results conform to the microwave propagation law; and the product generation and output module 250 is used to generate the final areal rainfall distribution product and perform uncertainty quantification and visualization output.

[0158] In this embodiment of the invention, the data acquisition and preprocessing module 210 is further used to: acquire the received signal level, frequency, polarization mode, and geographical location information of the microwave link from the telecommunications operator network; acquire synchronous temperature, humidity, air pressure, wind speed, and wind direction data from the meteorological observation network; acquire true rainfall observation data from the rain gauge and weather radar network; perform outlier detection and signal smoothing processing on the received signal level data, calculate the total path attenuation, and construct a standardized node feature matrix.

[0159] In this embodiment of the invention, the spatiotemporal graph construction module 220 is further configured to: define each microwave link as a graph node, and the node features adopt a preprocessed feature vector; construct a multi-scale weighted adjacency matrix based on the spatial distance, directional similarity and frequency correlation between links; and stack the node features of multiple consecutive time steps to form a spatiotemporal graph sequence.

[0160] In this embodiment of the invention, the model building and training module 230 is further configured to: construct a spatial feature extraction layer containing a graph attention network to capture spatial dependencies between nodes; construct a temporal feature extraction layer containing a gated temporal convolutional network to model rainfall temporal dynamics; and design a cross-modal attention fusion mechanism to adaptively fuse spatiotemporal features.

[0161] In this embodiment of the invention, the physical constraint inversion module 240 is further configured to: establish a microwave rain attenuation physical model based on ITU-R Recommendation P.838; construct a physical consistency loss function to constrain the output of the neural network to conform to a power-law relationship; and combine the physical loss with the data fitting loss and the smoothness loss to form a multi-objective optimization function.

[0162] In this embodiment of the invention, the product generation and output module 250 is further used to: perform gridded rainfall intensity estimation based on the trained model; quantify prediction uncertainty using the Monte Carlo Dropout method; generate continuous areal rainfall distribution maps through spatial interpolation; establish a product quality evaluation system; and output standardized rainfall monitoring products.

[0163] The areal rainfall distribution reconstruction device proposed in this embodiment of the invention achieves fully automated processing from raw microwave link data to high-resolution areal rainfall distribution products through the collaborative work of multiple modules. This device innovatively combines spatiotemporal graph neural networks with the physical mechanism of microwave propagation, fully utilizing the learning ability of deep learning for complex features while ensuring the rationality of the inversion results through physical constraints. Compared with traditional methods, this device can more accurately capture the spatial detail features of the rainfall field, providing reliable products with uncertainty assessment, and providing important technical support for urban flooding early warning, hydrological modeling, and weather forecasting. Since the device in Embodiment 2 of this invention is the same device used in the areal rainfall distribution reconstruction method based on spatiotemporal graph neural networks in Embodiment 1, those skilled in the art can understand the specific structure and variations of this device based on the method described in Embodiment 1, and therefore will not be repeated here. All devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0164] Example 3

[0165] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for reconstructing areal rainfall distribution based on a spatiotemporal graph neural network of Embodiment 1.

[0166] Since the computer-readable storage medium described in Embodiment 3 of this invention is the same computer-readable storage medium used in implementing the method for reconstructing surface rainfall distribution based on a spatiotemporal graph neural network in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer-readable storage medium based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer-readable storage media used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0167] Example 4

[0168] Based on the same inventive concept, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in Embodiment 1.

[0169] Since the computer device described in Embodiment 4 of this invention is the same computer device used to implement the method for reconstructing surface rainfall distribution based on spatiotemporal graph neural networks in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0170] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0171] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0172] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A method for reconstructing areal rainfall distribution based on spatiotemporal graph neural networks, characterized in that, include: Based on the spatial distribution characteristics of microwave links and the geographic information of the monitoring area, a spatiotemporal graph structure model containing node characteristics and spatial relationships is constructed. The spatial relationships between links are established through a multi-factor fusion adjacency matrix construction method, and spatiotemporal graph data objects are generated. Based on the generated spatiotemporal graph data object, a hierarchical spatiotemporal graph neural network is used for feature extraction and fusion to obtain the spatiotemporal depth feature representation of the fused precipitation field. Based on the spatiotemporal depth feature representation of the fused rainfall field, combined with the constraints of the microwave rain attenuation physical model, the rainfall intensity estimation with uncertainty awareness is carried out, and the physical consistency loss function is used to ensure that the inversion result conforms to the electromagnetic wave propagation law, so as to obtain a physically reasonable grid rainfall intensity value. Based on the obtained physically reasonable grid rainfall intensity values ​​and their uncertainty assessment, spatial interpolation and product quality control are performed to generate a high-resolution two-dimensional rainfall intensity distribution map with a credibility assessment. Based on a high-resolution two-dimensional rainfall intensity distribution map, a real-time inference system and feedback mechanism are established to dynamically update and output the rainfall field; The method also includes: The quality of the collected link data is controlled using a pre-set data quality control system. Based on the data after quality control, a multi-dimensional feature vector containing physical meaning is constructed for each link, represented as follows: in, Indicates the first The link is The feature vector of the spatiotemporal graph data object at any given time. Indicates the first The link at time Total path decay, Indicates the link's operating frequency. Indicates the link path length. Indicates the polarization mode. Indicates temperature. Indicates relative humidity. Indicates air pressure. Indicates wind speed. Indicates wind direction. Indicates the latitude of the link's center point. Indicates the longitude of the link's center point. This indicates the average altitude of the location where the link is located; Batch normalization based on Z-score is performed on all continuous features, and one-hot encoding is performed on categorical features to form a normalized node feature matrix; Based on the spatial distribution characteristics of microwave links and the geographical information of the monitoring area, a spatiotemporal graph structure model containing node characteristics and spatial relationships is constructed, including: Based on a standardized node feature matrix, the monitoring area is modeled as a dynamically evolving graph structure, where nodes in the graph structure represent microwave links and edges represent spatial relationships between links. Based on graph structure, construct including A sequence of spatiotemporal graphs with consecutive time steps is used as a multi-scale adjacency matrix.

2. The method for reconstructing areal rainfall distribution based on spatiotemporal graph neural networks as described in claim 1, characterized in that, Spatial relationships between links are established through a multi-factor fusion adjacency matrix construction method, generating spatiotemporal graph data objects, including: Based on the defined graph structure node information, a weighted adjacency matrix reflecting various spatial association patterns is constructed, represented as: In the formula: Represents the weights of the adjacency matrix; This represents the Euclidean distance between the center points of node i and node j; and Represents the link direction vector; and Indicates the link's operating frequency; Represents the weighting coefficients, satisfying ; and These represent the Gaussian kernel bandwidth parameters for distance and frequency, respectively. Construct a k-nearest neighbor graph and a Gaussian kernel graph, and combine the k-nearest neighbor graph, Gaussian kernel graph and multi-scale adjacency matrix into a multi-scale graph structure set to handle features at different spatial scales; By combining the spatiotemporal feature tensor with a set of multi-scale adjacency matrices, a spatiotemporal graph data object is formed.

3. The method for reconstructing areal rainfall distribution based on spatiotemporal graph neural networks as described in claim 1, characterized in that, Based on the generated spatiotemporal graph data object, a hierarchical spatiotemporal graph neural network is used for feature extraction and fusion to obtain the spatiotemporal depth feature representation of the fused precipitation field, including: A multi-layer graph attention network is used as a spatial feature extractor to extract features from the generated spatiotemporal graph data objects to form feature vectors; Spatial dependencies are captured using a graph attention mechanism, and graph attention coefficients are calculated. By combining graph attention coefficients and stacking multi-layer graph attention networks, information from multi-hop neighbors is aggregated to capture the spatial correlation of rainfall fields and output spatially enhanced features. A gated temporal convolutional network with residual connections is used to process the time series features of each node to obtain the time features. A cross-modal attention mechanism is used to adaptively fuse spatial enhancement features and temporal features to obtain a fused spatiotemporal depth feature representation of the precipitation field.

4. The method for reconstructing areal rainfall distribution based on spatiotemporal graph neural networks as described in claim 1, characterized in that, Based on the spatiotemporal depth feature representation of the fused rainfall field, and constrained by the microwave rain attenuation physical model, uncertainty-aware rainfall intensity estimation is performed. A physical consistency loss function is used to ensure that the inversion results conform to the electromagnetic wave propagation laws, resulting in physically reasonable grid rainfall intensity values, including: A physical model of the attenuation of microwave signals by rainfall is established to describe the quantitative relationship between rainfall intensity and microwave attenuation. The physical laws of microwave propagation are introduced as soft constraints into the optimization process to establish a physical model of attenuation-rainfall intensity. The physical consistency loss of the physical model of attenuation-rainfall intensity is used to characterize the difference between the model-predicted attenuation and the measured attenuation. Based on rainfall intensity estimation and physical consistency loss, a loss function containing multiple optimization objectives is constructed to guide the model to learn a solution that simultaneously satisfies data fitting and physical constraints, resulting in a trained model. The trained model is used to estimate the rainfall intensity and its uncertainty at each node location.

5. The method for reconstructing areal rainfall distribution based on spatiotemporal graph neural networks as described in claim 4, characterized in that, Based on a high-resolution two-dimensional rainfall intensity distribution map, a real-time inference system and feedback mechanism are established to dynamically update and output the rainfall field, including: Based on the trained model, uncertainty is estimated using Monte Carlo Dropout, and Dropout activation is maintained during the inference phase to perform multiple random forward propagations. Based on the generated high-resolution two-dimensional rainfall intensity distribution map and its uncertainty assessment, the inverted rainfall field product is assimilated with the numerical weather prediction model. The initial field is optimized by a four-dimensional variational assimilation method. At the same time, a feedback mechanism is established to feed back the assimilation effect assessment and actual observation verification results to the model training process, forming a closed-loop optimization system.

6. A device for reconstructing areal rainfall distribution based on a spatiotemporal graph neural network, characterized in that, Based on the method described in claim 1, it includes: The data acquisition and preprocessing module is used to construct a spatiotemporal graph structure model containing node characteristics and spatial relationships based on the spatial distribution characteristics of microwave links and the geographic information of the monitoring area. It establishes spatial relationships between links through a multi-factor fusion adjacency matrix construction method and generates spatiotemporal graph data objects. The spatiotemporal graph construction module is used to extract and fuse features based on the generated spatiotemporal graph data objects using a hierarchical spatiotemporal graph neural network, to obtain the spatiotemporal depth feature representation of the fused precipitation field. The model building and training module is used to estimate rainfall intensity based on the spatiotemporal depth feature representation of the fused rainfall field, combined with the constraints of the microwave rain attenuation physical model, and to ensure that the inversion results conform to the electromagnetic wave propagation law through the physical consistency loss function, so as to obtain physically reasonable grid rainfall intensity values. The physical constraint inversion module is used to perform spatial interpolation and product quality control based on the obtained physically reasonable grid rainfall intensity values ​​and their uncertainty assessment, and to generate a high-resolution two-dimensional rainfall intensity distribution map with a credibility assessment. The product generation and output module is used to establish a real-time inference system and feedback mechanism based on a high-resolution two-dimensional rainfall intensity distribution map, and to dynamically update and output the rainfall field.

7. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method for reconstructing areal rainfall distribution based on a spatiotemporal graph neural network as described in any one of claims 1 to 5.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for reconstructing areal rainfall distribution based on spatiotemporal graph neural networks as described in any one of claims 1 to 5.

Images