Method for tracking and warning of severe convective weather based on radar data and deep learning

Abstract

The application relates to the technical field of meteorological disaster monitoring, and particularly discloses a strong convective weather tracking and early warning method based on radar data and deep learning. The method realizes real-time collection of original radar data through laser radar and spaceborne radar, obtains the original radar data, and carries out cleaning and feature extraction processing to obtain real-time space-time feature data. The deep learning model automatically learns the real-time space-time feature data of the radar data, so that the deep learning model has the ability to automatically identify and track the strong convective system. The deep learning model directly outputs the boundary and intensity of the strong convective system, and directly generates the mask of the strong convective system through one forward propagation, thereby effectively reducing the calculation amount of the radar data. Finally, the future influence area of the strong convective system is predicted based on the evolution characteristics of the strong convective system, early warning information is generated, and the early warning information is output. The whole process from data processing to early warning publishing is free of manual intervention, and the automation degree and efficiency of early warning are improved.

Description

Technical Field

[0001] This invention belongs to the field of meteorological disaster monitoring technology, specifically involving a method for tracking and early warning of severe convective weather based on radar data and deep learning. Background Technology

[0002] Severe convective weather is one of the most sudden and destructive weather types in meteorological monitoring. It mainly includes thunderstorms, strong winds, hail, tornadoes, and short-duration heavy rainfall. It has a short lifespan and a small spatial scale (down to the level of hundreds of meters), posing a major threat to people's property and safety. In order to reduce the damage to property and the harm to life caused by severe convective weather, it is necessary to track severe convective weather.

[0003] Currently, the tracking of severe convective weather typically employs observation methods such as satellites and lidar. Brightness and temperature data obtained from satellite observations are used to identify and track mesoscale convective systems (MCS), thereby predicting the occurrence and movement of severe convective weather. For example, patent application number CN202310589785.X discloses a remote sensing-based method and system for monitoring and warning of severe convective weather over power lines. This method includes: identifying and tracking the occurrence and evolution of mesoscale convective systems using brightness and temperature data obtained from satellite remote sensing; extracting the main evolutionary features of the mesoscale convective system based on its evolutionary process, and predicting the impact area of ​​the mesoscale convective system at the next moment using weighted averaging and Lagrange interpolation based on these main evolutionary features; issuing warnings to transmission towers located within the impact area of ​​the mesoscale convective system based on the prediction results and the locations of each transmission tower on the power line, marking the transmission towers within the impact area, and outputting the warning results.

[0004] However, the above-mentioned methods suffer from low computational efficiency and difficulty in rapidly outputting early warning results when processing large amounts of satellite or radar data. For example, when tracking the evolution of mesoscale convective systems and extracting evolutionary features, the overlapping area method for tracking cold cloud systems requires calculating the overlapping area ratio of all regions at adjacent times, which necessitates traversing the pixels of both regions for each overlapping area calculation, resulting in high complexity. Therefore, we need to propose a severe convective weather tracking and early warning method based on radar data and deep learning to solve the above-mentioned problems. This method can automatically process and analyze large amounts of data through deep learning technology, reducing manual intervention and improving the automation and efficiency of early warning. Summary of the Invention

[0005] The purpose of this invention is to provide a method for tracking and warning of severe convective weather based on radar data and deep learning. This method can automatically process and analyze large amounts of data through deep learning technology, reduce manual intervention, and improve the automation and efficiency of early warning, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for tracking and issuing early warnings of severe convective weather based on radar data and deep learning includes the following steps:

[0008] S1. Real-time acquisition of raw radar data, including radar reflectivity, radial velocity, and velocity spectrum width, using lidar and spaceborne radar;

[0009] S2. Acquire raw radar data and perform cleaning and feature extraction to obtain real-time spatiotemporal feature data;

[0010] S3. Construct a deep learning model suitable for tracking severe convective weather, and train the deep learning model with historical data to enable the deep learning model to automatically identify and track severe convective systems.

[0011] S4. Input the real-time spatiotemporal feature data into the trained deep learning model to automatically identify strong convective systems in the radar data, establish spatiotemporal correlation tracking, and obtain the historical evolution characteristics of strong convective systems.

[0012] S5. Predict the future impact area of ​​strong convective systems based on their historical evolution characteristics;

[0013] S6. Generate and output early warning information based on the predicted impact area of ​​the severe convective system.

[0014] The preferred processing flow for raw radar data is as follows:

[0015] A1. Use the median filtering algorithm to remove noise from the original radar data and fill in the missing values ​​by interpolation;

[0016] A2. Extract the basic features of radar reflectivity factor, radial velocity and velocity spectrum width from the filtered radar data. At the same time, calculate the strong convection related features, including storm top height, vertical liquid water content, mesocyclone characteristic quantity and hail index.

[0017] A3. Normalize the characteristics related to strong convection.

[0018] Preferably, the deep learning model uses a spatiotemporal convolutional neural network combined with an attention mechanism to capture the spatial features and temporal evolution of radar data. The attention mechanism includes spatial attention and channel attention. Spatial attention focuses on areas with strong convection by calculating the spatial weights of the feature map; channel attention automatically learns the importance weights of radar features.

[0019] The preferred training process for a deep learning model is as follows:

[0020] B1. Collect historical radar data containing radar reflectivity factor, radial velocity and velocity spectrum width, as well as manually labeled data of strong convective system boundaries and intensity levels to establish a dataset.

[0021] B2. Divide the dataset into training, validation, and test sets in an 8:1:1 ratio;

[0022] B3. The deep learning model is trained using the cross-entropy loss function based on the training set data, and the Adam optimizer is used to dynamically adjust the learning rate to accelerate convergence.

[0023] B4. Evaluate the performance of the deep learning model based on the validation set data. When the prediction accuracy reaches the preset threshold, the training is completed. Otherwise, continue to train and converge the deep learning model until the deep learning model training is completed.

[0024] The preferred real-time identification process for strong convective systems is as follows:

[0025] C1. Divide the real-time spatiotemporal feature data according to time intervals or spatial ranges to form multiple data segments with independent spatiotemporal attributes;

[0026] C2. Output the pixel-level segmentation bounding box of the strong convection system through forward propagation, and output the intensity score at the same time;

[0027] C3. Use nonmaximum suppression to remove overlapping detection boxes and retain the best detection results.

[0028] The preferred spatiotemporal correlation tracking process is as follows:

[0029] D1. Extract the spatiotemporal features of the centroid coordinates, area, and mean reflectance of the bounding box for each detected strong convective system;

[0030] D2. Calculate the spatial distance and characteristic distance between strong convective systems at adjacent time points, and construct the cost matrix;

[0031] D3. Use the Hungarian algorithm to find the optimal match, establish the continuity relationship of system IDs, and handle the complex cases of splitting and merging.

[0032] Preferably, the cost matrix formula is expressed as follows:

[0033] ,

[0034] ,

[0035] ,

[0036] in, For time t, the first A strong convective system, Let j be the j-th strong convective system at time t+1. and for centroid coordinates and for The coordinates of the centroid, for The average reflectance, for The average reflectance, These are the weighting coefficients used to adjust the importance of spatial distance and feature distance in the cost matrix. for and The cost of matching.

[0037] Preferably, the continuation relationship of system IDs is constructed by solving the cost matrix using the Hungarian algorithm to obtain the optimal matching pair of strong convective systems at adjacent time points. and If matched, then inherit Establish a continuation relationship using the system ID.

[0038] The preferred process for predicting the impact area is as follows:

[0039] E1. Extract the characteristics of velocity, acceleration, and shape changes from the evolution trajectory of a strong convective system obtained by spatiotemporal correlation tracing to obtain historical features;

[0040] E2. Use self-attention mechanisms to predict the location and extent of future moments based on historical features;

[0041] E3. Based on the predicted location and range, conduct an error assessment and select the location and range with the smallest error as the affected area.

[0042] Preferably, when generating early warning information, the predicted impact area is spatially overlaid with the location of meteorological stations to determine which meteorological stations are located within the impact area. Different early warning levels are then defined based on the intensity and impact of the severe convective system. Finally, early warning information containing the impact area, early warning level, and disaster prevention measures is generated based on the early warning level and disseminated to each impact area.

[0043] The severe convective weather tracking and early warning method based on radar data and deep learning proposed in this invention has the following advantages compared with existing technologies:

[0044] 1. This invention automatically learns the real-time spatiotemporal characteristics of radar data through a deep learning model, enabling the deep learning model to automatically identify and track strong convective systems. The deep learning model directly outputs the boundary and intensity of the strong convective system, and generates the mask of the strong convective system directly through a single forward propagation, effectively reducing the computational load on radar data. Finally, based on the evolution characteristics of the strong convective system, the future impact area of ​​the strong convective system is predicted, and early warning information is generated and output according to the impact area. The entire process from data processing to early warning issuance is completed without human intervention, improving the automation and efficiency of early warning. Attached Figure Description

[0045] Figure 1 A flowchart of a severe convective weather tracking and early warning method according to an embodiment of the present invention is shown;

[0046] Figure 2 A flowchart illustrating the processing of raw radar data according to an embodiment of the present invention is shown;

[0047] Figure 3 A flowchart illustrating the training process of a deep learning model according to an embodiment of the present invention is shown.

[0048] Figure 4 A flowchart illustrating real-time identification of a strong convection system according to an embodiment of the present invention is shown. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] This invention provides, for example Figure 1-4 The severe convective weather tracking and early warning method based on radar data and deep learning, as shown, includes the following steps:

[0051] S1. Real-time acquisition of raw radar data, including radar reflectivity, radial velocity, and velocity spectrum width, using lidar and spaceborne radar;

[0052] The spaceborne radar is mounted on a GPM satellite, enabling global observation and providing global background information to help grasp the overall movement trend and development direction of weather systems. The lidar is carried by a UAV located at a meteorological station, providing high-precision three-dimensional spatial information to help accurately track the internal structure and dynamic changes of convective weather. The combination of lidar and spaceborne radar allows for a more comprehensive understanding of the characteristics and evolution patterns of severe convective weather, improving the ability to track and warn of severe convective weather.

[0053] S2. Acquire raw radar data and perform cleaning and feature extraction to obtain real-time spatiotemporal feature data;

[0054] like Figure 2 As shown, the processing flow of raw radar data is as follows:

[0055] A1. The median filtering algorithm is used to filter out noise in the original radar data, and missing values ​​are filled in by interpolation to eliminate the impact of outliers on the analysis of the original radar data.

[0056] A2. Extract the basic features of radar reflectivity factor, radial velocity and velocity spectrum width from the filtered radar data. At the same time, calculate the strong convection related features, including storm top height, vertical liquid water content, mesocyclone characteristic quantity and hail index.

[0057] The basic feature extraction methods are as follows:

[0058] The radar reflectivity factor extraction method is to directly read the reflectivity factor value from the original radar data. The unit is dBZ. The radar reflectivity factor reflects the ability of precipitation particles to scatter electromagnetic waves within a unit volume and is used to determine the intensity of precipitation.

[0059] Radial velocity is calculated based on the Doppler effect, using the frequency shift of the echo signal. The formula for calculating radial velocity is:

[0060] ,in, Radial velocity, For radar wavelength, For Doppler, The distance from the target to the radar is denoted by , and the radial velocity is the radial velocity of the target relative to the radar. A positive value indicates that the target is moving away from the radar, and a negative value indicates that the target is moving closer to the radar.

[0061] The velocity spectral width is the standard deviation obtained from calculating the power spectrum of the echo signal. The formula for calculating the standard deviation is:

[0062] ,in, For the velocity spectrum, For the first One velocity component For average speed, For the first The power corresponding to each velocity, and the velocity spectrum width are used to reflect the intensity of atmospheric turbulence and the size distribution of precipitation particles;

[0063] The storm top height is the vertical development height of a strong convective cloud, reflecting the intensity of convection. The higher the height, the more vigorous the convection. It is calculated by extracting the vertical distribution of reflectivity factors along the vertical direction and identifying the height layer where the vertical reflectivity gradient decreases significantly. The calculation formula is as follows:

[0064] ,in, The height of the storm crest. The vertical height. Let h be a small change in height. This represents the variation of the reflectivity factor with height in the vertical direction. The gradient threshold used to reflect the critical value of reflectivity decreasing with height is usually taken as -5 dBZ / km;

[0065] Vertical liquid water content is the total mass of precipitation particles per unit area, used to assess contact potential and hail risk. It is calculated by converting the radar reflectivity factor to liquid water content using the reflectivity-liquid water content relationship, and then integrating along the vertical direction. The formula is as follows:

[0066] ,in, For vertical liquid water content, The height of the storm crest. The lowest measurable height, Let h be a small change in height. Let Z be the liquid water content at height h. The relationship between reflectivity and liquid water content is: Z = a * L b a=200, b=1.6, Z is the radar reflectivity factor, and L is the liquid water content corresponding to the radar reflectivity factor Z;

[0067] Mesocyclone characteristic quantities reflect the rotating airflow within strong convective systems and are precursory features of severe weather events such as tornadoes. The formula for calculating mesocyclone characteristic quantities is:

[0068] ,in, and The maximum values ​​of positive and negative velocities. The horizontal distance between the speed pairs;

[0069] The hail index is used to assess the probability of hail formation from severe convective systems. The formula for calculating the hail index is:

[0070] ,in, Hail index For vertical liquid water content, The height of the storm crest. The altitude of the 0°C isotherm in the atmosphere;

[0071] A3. Normalize the features related to strong convection to avoid interference from differences in the scale of different features on the training of deep learning models.

[0072] S3. Construct a deep learning model suitable for tracking severe convective weather, and train the deep learning model with historical data to enable the deep learning model to automatically identify and track severe convective systems.

[0073] The deep learning model employs a spatiotemporal convolutional neural network combined with an attention mechanism to capture the spatial features and temporal evolution patterns of radar data. The formula for the spatiotemporal convolutional neural network is as follows:

[0074] ,

[0075] ,

[0076] ,

[0077] ,

[0078] ,

[0079] in, For input data, Let t represent the cell state at time t. The cell state at time t-1. Let be the hidden state at time t. The hidden state at time t-1 For use in controlling the input data at the current time Multiple pieces of information are stored in the cell state. The input gate, For use in determining cell state What historical information has been forgotten? (This is a fragment of a longer text, likely due to a formatting error.) For use in controlling cell state How much information is output as a hidden state? The output gate, and These are the cellular state and the hidden state, respectively. For the sigmoid function, For convolution operations, This is element-wise multiplication. For input data The convolution weight matrix to the input gate, Hidden state Convolution weights to the input gate, For the bias term of the input gate, For the bias term of the forget gate, This is a bias term for the cell state. and These are the convolution weights from the input data and the hidden state to the candidate memory, respectively. The hyperbolic tangent activation function is used. , and These represent the convolution weights from the input data, hidden state, and cell state to the output gate, respectively. This is the bias term for the output gate;

[0080] Attention mechanisms include spatial attention and channel attention. Spatial attention focuses on regions with strong convection by calculating spatial weights in the feature map. The formula for calculating spatial weights is:

[0081] ,in, For spatial weight values, For the sigmoid function, For m*m convolution, X is the feature map. and Max pooling and average pooling for spatial dimensions;

[0082] Channel attention is the importance weight of automatically learned radar features. The formula for calculating the importance weight is:

[0083] ,in, For spatial weight values, For the sigmoid function, For m*m convolution, For channel dimension features, and Max pooling and average pooling are used for spatial dimensions. For a multilayer perceptron, weights are generated through global pooling and nonlinear transformation along the channel dimension. ;

[0084] By introducing spatial attention and channel attention, we can enhance our sensitivity to key areas of strong convection.

[0085] like Figure 3 As shown, the training process for a deep learning model is as follows:

[0086] B1. Collect historical radar data containing radar reflectivity factor, radial velocity and velocity spectrum width, as well as manually labeled data of strong convective system boundaries and intensity levels to establish a dataset.

[0087] B2. Divide the dataset into training, validation, and test sets in an 8:1:1 ratio to ensure that the deep learning model performs stably on unseen data.

[0088] B3. The deep learning model is trained using the cross-entropy loss function based on the training set data, and the Adam optimizer is used to dynamically adjust the learning rate to accelerate convergence. In this field, the use of the cross-entropy loss function and the Adam optimizer to train and optimize deep learning models is common knowledge, so it will not be described in detail here.

[0089] B4. Evaluate the performance of the deep learning model based on the validation set data. When the prediction accuracy reaches the preset threshold (the preset threshold is set according to the actual early warning needs), the training is completed. Otherwise, continue to train and converge the deep learning model until the deep learning model is trained.

[0090] S4. Input the real-time spatiotemporal feature data into the trained deep learning model to automatically identify strong convective systems in the radar data, establish spatiotemporal correlation tracking, and obtain the historical evolution characteristics of strong convective systems.

[0091] Severe convective systems are small- to medium-scale weather systems caused by intense vertical motion in the atmosphere. They are characterized by suddenness, short lifespan, and violent weather phenomena, and are often accompanied by severe weather events such as thunderstorms, strong winds, hail, short-duration heavy rainfall, and even tornadoes. These systems are manifested in radar data as specific reflectivity structures and spatiotemporal evolution characteristics, making them a key focus of meteorological monitoring and forecasting.

[0092] like Figure 4 As shown, the process for real-time identification of strong convective systems is as follows:

[0093] C1. Divide the real-time spatiotemporal feature data according to time intervals or spatial ranges to form multiple data segments with independent spatiotemporal attributes;

[0094] C2. Output the pixel-level segmentation bounding box of the strong convection system through forward propagation, and at the same time output the intensity score, such as the probability of belonging to thunderstorm;

[0095] The forward propagation formula is:

[0096] ,in, The output is the intensity rating value. It is the ReLU activation function. This is the weight matrix used to perform a linear transformation on the input features. For input feature values, This is the bias vector used to adjust the output of the linear transformation;

[0097] C3. Use non-maximum suppression to remove overlapping detection boxes and retain the best detection result; the formula for non-maximum suppression is:

[0098] ,in, To preserve the test results, These are the regions of detected strong convective systems. For intersection, union, and comparison, For threshold;

[0099] The spatiotemporal correlation tracing process is as follows:

[0100] D1. Extract the spatiotemporal features of the centroid coordinates, area, and mean reflectance of the bounding box for each detected strong convective system;

[0101] D2. Calculate the spatial distance and characteristic distance between strong convective systems at adjacent time points (using Euclidean distance for spatial distance and reflectivity difference for characteristic distance), and construct the cost matrix. The formula for the cost matrix is ​​as follows:

[0102] ,

[0103] ,

[0104] ,

[0105] in, For time t, the first A strong convective system, Let j be the j-th strong convective system at time t+1. and for centroid coordinates and for The coordinates of the centroid, for The average reflectance, for The average reflectance, These are the weighting coefficients used to adjust the importance of spatial distance and feature distance in the cost matrix. for and The lower the cost of matching, the stronger the correlation.

[0106] D3. Use the Hungarian algorithm to solve for the optimal matching, establish the continuity relationship of system IDs, and handle the complex cases of splitting and merging.

[0107] The formula for the Hungarian algorithm is:

[0108] ,

[0109] The constraints are: ,

[0110] ,

[0111] ,when Time indicates and Matches, otherwise does not match;

[0112] in, The elements in the cost matrix, i.e. and The matching cost, A matching strategy variable that takes the value 0 or 1, used to indicate whether to... and match, Let be the number of strong convective systems detected at time t. denoted as the number of strong convective systems detected at time t adjacent to time t.

[0113] The continuation relationship of system IDs is constructed by solving the cost matrix using the Hungarian algorithm to obtain the optimal matching pairs of strong convective systems at adjacent time points. and If matched, then inherit Establish a continuation relationship based on the system ID, and the ID allocation rule is as follows:

[0114] like Matched unique At that time, Continue ID;

[0115] like No match found At that time, The ID is temporarily retained or marked as expired;

[0116] like No match found If so, a new ID will be assigned.

[0117] During split processing, if a system It splits into multiple systems at time t. , When ..., the Hungarian algorithm, based on minimum cost, can ... With one of the best matches Related, the rest are not matched Assigning a new ID can be further evaluated by combining it with the splitting rules, which are set as thresholds for the difference between area and reflectivity. and If the feature distances are all less than the threshold, then it is marked as Derivative branches that split off from each other and share the same ID;

[0118] During the merge process, if multiple systems , ...merge into one system at time t In some cases, the Hungarian algorithm fails to directly match a pair of pairs because they do not meet the constraints. Therefore, it calculates... With each Choose the option with the lowest cost. Establish ID continuation, the rest Mark as merged and destroyed, if multiple and If both the spatial distance and feature distance are less than the threshold, then... Consider these The merged result is used to assign a new ID and record its historical origin;

[0119] Real-time identification and tracking of strong convective systems are achieved through deep learning models and spatiotemporal correlation algorithms. Forward propagation is used for feature extraction and classification. The cost matrix is ​​combined with spatial and feature distances to construct matching relationships. The Hungarian algorithm solves the optimal matching to establish ID continuation. The splitting and merging processes need to be combined with threshold judgment and logical rules to ensure the accuracy of system evolution characteristics. These processes together support the full-chain analysis of strong convective systems from detection to historical tracing.

[0120] S5. Predict the future impact area of ​​strong convective systems based on their historical evolution characteristics;

[0121] The process for predicting the impact area is as follows:

[0122] E1. Extract the characteristics of velocity, acceleration, and shape changes from the evolution trajectory of a strong convective system obtained by spatiotemporal correlation tracing to obtain historical features;

[0123] E2. Use self-attention mechanisms to predict the location and extent of future moments based on historical features;

[0124] U, of which These are query, key, and value matrices, respectively. The dimension of the key;

[0125] E3. Based on the predicted location and range, conduct an error assessment, selecting the location and range with the smallest error as the affected area. The formula for calculating the prediction error assessment is as follows:

[0126] ,in, For prediction error, To predict the sample size, For the true value, This is a predicted value;

[0127] S6. Generate and output early warning information based on the predicted impact area of ​​the severe convective system;

[0128] When generating early warning information, the predicted impact area is spatially overlaid with the locations of meteorological stations to determine which stations are located within the impact area. Different warning levels are then assigned based on the intensity and impact of the severe convective system. Finally, an early warning message containing the impact area, warning level, and disaster prevention measures is generated and issued to each impact area. The warning levels are based on the China Meteorological Administration's severe convective weather warning level standards, and are ranked from lowest to highest severity as blue, yellow, orange, and red. Blue indicates that severe convective weather is likely to occur with minor impact; yellow indicates that the probability of severe convective weather is increasing and the impact is becoming more obvious; orange indicates that severe convective weather is about to occur with significant impact and may cause some disasters; and red indicates that severe convective weather has seriously affected the local area and may cause major disasters.

[0129] By automatically learning real-time spatiotemporal characteristics of radar data through deep learning models, the deep learning models are able to automatically identify and track severe convective systems. The deep learning models directly output the boundaries and intensity of severe convective systems and generate the mask of severe convective systems through a single forward propagation, effectively reducing the computational load on radar data. Finally, based on the evolutionary characteristics of severe convective systems, the future impact area of ​​the severe convective systems is predicted, and early warning information is generated and output according to the impact area. The entire process from data processing to early warning issuance is completed without human intervention, improving the automation and efficiency of early warning.

[0130] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for tracking and warning severe convective weather based on radar data and deep learning, characterized by: Includes the following steps: S1. Real-time acquisition of raw radar data, including radar reflectivity, radial velocity, and velocity spectrum width, using lidar and spaceborne radar; the spaceborne radar is configured as a GPM satellite and mounted on the satellite; the lidar is mounted on a drone located at a weather station. S2. Acquire raw radar data and perform cleaning and feature extraction to obtain real-time spatiotemporal feature data; S3. Construct a deep learning model suitable for tracking severe convective weather, and train the deep learning model with historical data to enable the deep learning model to automatically identify and track severe convective systems. The deep learning model uses a spatiotemporal convolutional neural network combined with an attention mechanism to capture the spatial features and temporal evolution of radar data. The attention mechanism includes spatial attention and channel attention. Spatial attention focuses on areas with strong convection by calculating the spatial weights of the feature map. Channel attention is the importance weight of radar features learned automatically; S4. Input the real-time spatiotemporal feature data into the trained deep learning model to automatically identify strong convective systems in the radar data, establish spatiotemporal correlation tracking, and obtain the historical evolution characteristics of strong convective systems. The spatiotemporal correlation tracing process is as follows: D1. Extract the spatiotemporal features of the centroid coordinates, area, and mean reflectance of the bounding box for each detected strong convective system; D2. Calculate the spatial distance and characteristic distance between strong convective systems at adjacent time points, and construct the cost matrix; D3. Use the Hungarian algorithm to solve for the optimal matching, establish the continuity relationship of system IDs, and handle the complex cases of splitting and merging. S5. Predict the future impact area of ​​strong convective systems based on their historical evolution characteristics; The process for predicting the impact area is as follows: E1. Extract the characteristics of velocity, acceleration, and shape changes from the evolution trajectory of a strong convective system obtained by spatiotemporal correlation tracing to obtain historical features; E2. Use self-attention mechanisms to predict the location and extent of future moments based on historical features; E3. Based on the predicted location and range, conduct an error assessment and select the location and range with the smallest error as the affected area; S6. Generate and output early warning information based on the predicted impact area of ​​the severe convective system.

2. The severe convective weather tracking and early warning method based on radar data and deep learning according to claim 1, characterized in that: The processing flow for raw radar data is as follows: A1. Use the median filtering algorithm to remove noise from the original radar data and fill in the missing values ​​by interpolation; A2. Extract the basic features of radar reflectivity factor, radial velocity and velocity spectrum width from the filtered radar data. At the same time, calculate the strong convection related features, including storm top height, vertical liquid water content, mesocyclone characteristic quantity and hail index. A3. Normalize the characteristics related to strong convection.

3. The severe convective weather tracking and early warning method based on radar data and deep learning according to claim 2, characterized in that: The training process for deep learning models is as follows: B1. Collect historical radar data containing radar reflectivity factor, radial velocity and velocity spectrum width, as well as manually labeled data of strong convective system boundaries and intensity levels to establish a dataset. B2. Divide the dataset into training, validation, and test sets in an 8:1:1 ratio; B3. The deep learning model is trained using the cross-entropy loss function based on the training set data, and the Adam optimizer is used to dynamically adjust the learning rate to accelerate convergence. B4. Evaluate the performance of the deep learning model based on the validation set data. When the prediction accuracy reaches the preset threshold, the training is completed. Otherwise, continue to train and converge the deep learning model until the deep learning model training is completed.

4. The severe convective weather tracking and early warning method based on radar data and deep learning according to claim 3, characterized in that: The process for real-time identification of strong convective systems is as follows: C1. Divide the real-time spatiotemporal feature data according to time intervals or spatial ranges to form multiple data segments with independent spatiotemporal attributes; C2. Output the pixel-level segmentation bounding box of the strong convection system through forward propagation, and output the intensity score at the same time; C3. Use nonmaximum suppression to remove overlapping detection boxes and retain the best detection results.

5. The severe convective weather tracking and early warning method based on radar data and deep learning according to claim 4, characterized in that: In step D2, the cost matrix formula is expressed as follows: ， ， ， in, For time t, the first A strong convective system, Let j be the j-th strong convective system at time t+1. and for centroid coordinates and for The coordinates of the centroid, for The average reflectance, for The average reflectance, These are the weighting coefficients used to adjust the importance of spatial distance and feature distance in the cost matrix. for and The cost of matching.

6. The severe convective weather tracking and early warning method based on radar data and deep learning according to claim 5, characterized in that: The continuation relationship of system IDs is constructed by solving the cost matrix using the Hungarian algorithm to obtain the optimal matching pairs of strong convective systems at adjacent time points. and If matched, then inherit Establish a continuation relationship using the system ID.

7. The severe convective weather tracking and early warning method based on radar data and deep learning according to claim 6, characterized in that: When generating early warning information, the predicted impact area is spatially overlaid with the location of meteorological stations to determine which meteorological stations are located within the impact area. Different early warning levels are then assigned based on the intensity and impact of the severe convective system. Finally, early warning information containing the impact area, early warning level, and disaster prevention measures is generated based on the early warning level and disseminated to each impact area.

Images