A method for tornado identification and path prediction based on X-band dual-polarization radar data

Abstract

The present invention discloses a method for tornado identification and path prediction based on X-band dual-polarization radar data. The method mainly utilizes six observation quantities of the two lowest elevation angles of the radar, combines the deep learning model Unet, introduces the attention mechanism and position coding information, optimizes the data transmission process, and constructs a weakly supervised lightweight recognition model TD-Unet. The model extracts features from multi-elevation radar data to generate a tornado recognition probability and a geographic orientation map. In the recognition result, the geographic coordinate sequence of the tornado is further extracted, and the particle filter algorithm is optimized based on the tornado motion characteristics to construct a path prediction algorithm, and the path prediction is realized through the extracted real-time tornado coordinate sequence. The present invention can improve the accuracy and efficiency of tornado identification, provide a scientific and feasible method for path prediction, and provide strong technical support for meteorological warning and disaster prevention and control.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence-driven meteorological monitoring and disaster warning, and in particular relates to a tornado identification and path prediction method based on X-band dual-polarization radar data. Background Art

[0002] Tornadoes are extremely destructive natural disasters. Their powerful whirlwinds and extremely high wind speeds can cause huge casualties and property losses in a very short period of time. As a localized, small-scale, sudden severe convective weather, early identification and warning of tornadoes are crucial for early evacuation or rescue. However, due to their suddenness and complexity, accurate detection and forecasting are extremely challenging.

[0003] Traditional tornado detection methods mainly rely on weather radars to identify tornadoes by analyzing weather patterns and wind speed data. However, these methods often have problems such as low detection accuracy and poor timeliness, making it difficult to identify the formation of tornadoes in a timely and accurate manner.

[0004] With the development of weather radar technology, dual-polarization radar technology has been widely used in meteorological observations, especially in the identification and monitoring of tornadoes, showing unique advantages. By simultaneously measuring horizontal and vertical polarization signals, dual-polarization radar can provide richer meteorological information and effectively distinguish different types of precipitation and meteorological phenomena, thereby improving the detection accuracy of tornadoes. Compared with traditional single-polarization radars, dual-polarization radars can more accurately identify the characteristics of tornadoes, such as the intensity, shape, and movement trajectory of storms, providing more reliable data support for real-time monitoring and early warning, greatly enhancing the responsiveness and accuracy of tornado early warning systems.

[0005] Although significant progress has been made in applying machine learning methods to identify tornadoes, these methods require extensive preprocessing of radar data, which may result in the loss of key information inherent in radar data. In contrast, deep learning methods have powerful end-to-end learning capabilities and can directly process raw data without manual feature extraction. In recent years, deep learning technology has made significant progress in image recognition, natural language processing and other fields, and has gradually been applied to weather forecasting and disaster detection. Unlike traditional methods, these methods can extract tornado features directly from radar data without relying heavily on specific physical assumptions. By learning from a large amount of historical data, deep learning models can spontaneously extract and identify complex weather patterns, and are efficient and accurate.

[0006] In the scenario of tornado identification and prediction, there is an urgent need to design an efficient and accurate identification model to accurately identify the occurrence and location of tornadoes through in-depth analysis of radar data, and to predict the path based on the identification results. This can not only effectively improve the monitoring accuracy of tornadoes, but also warn of their possible paths in advance, providing more accurate decision-making support for scientific prevention and control. Through the real-time early warning system, an alarm can be issued in the early stages of a tornado, reducing the threat of disasters to society and people's lives and property. In addition, the efficiency and accuracy of the model provide strong technical support for disaster prevention and mitigation work, ensuring that it can fully play a role in emergency response and rescue decision-making, thereby ensuring the safety of the people and social stability. Summary of the invention

[0007] The object of the present invention is to provide a method for tornado identification and path prediction based on X-band dual-polarization radar data, thereby overcoming the defects in the above-mentioned prior art.

[0008] To achieve the above object, the present invention provides a method for tornado identification and path prediction based on X-band dual-polarization radar data, comprising the following steps:

[0009] Step 1) Use X-band radar data to build a tornado identification deep learning dataset:

[0010] First, with the tornado as the center, a fixed-size data block (W, H) around the tornado is intercepted from the six basic data (reflectivity, radial velocity, velocity spectrum width, differential reflectivity, differential phase shift, and correlation coefficient) at the two lowest elevation angles of the radar to form a positive sample. The sample shape is (12, W, H). Each observation is then normalized separately, and the negative sample is the data from the random weather process that does not contain a tornado. The number of positive and negative samples should be close to a 1:1 ratio. Then create an index file for the positive and negative samples and mark them. If a tornado occurs in the data block, the sample is marked as 1, and if not, it is marked as 0. Finally, it is divided into training set and test set in a ratio of 4:1, and the ratio of positive and negative samples in the training set and test set is also 1:1;

[0011] Step 2) Construct a deep learning model for tornado identification:

[0012] Using the U-Net model as the basic framework, the encoder-decoder structure extracts features and restores spatial resolution; on this basis, a weakly supervised lightweight tornado recognition model, TD-Unet (TornadoDetection Unet), was further improved and constructed to specifically identify tornadoes; first, the introduction of weakly supervised learning significantly reduces the labeling cost of positive sample data, while also reducing the impact of errors or deviations that may occur during manual labeling on model training. Secondly, the lightweight design enables the model to be efficiently deployed and recognized in real time in a radar network composed of multiple X-band radars. The specific method of weakly supervised learning for tornado recognition is that the model is trained only by the label of whether a tornado exists, without the need to label the specific location, shape or size of the tornado, so that the model can autonomously learn the characteristics of the tornado and obtain the location where the tornado occurs;

[0013] Step 3) The X-band dual-polarization tornado data set obtained in step 1) is divided into a training set and a test set in a ratio of 4:1, and the training set is input into the constructed TD-Unet model for training; the performance of the trained model is evaluated on the test set using a binary classification confusion matrix to obtain an objective evaluation index;

[0014] Step 4) Extract the tornado coordinates from the recognition results, and convert the coordinates recognized at multiple time steps into a tornado coordinate sequence; since there are always false positives in the recognition results that need to be eliminated, and there are also multiple tornadoes where the coordinates of different tornadoes need to be predicted separately, this provides data guarantee for the next step of path prediction;

[0015] Filter false positives from the output of the previous deep learning model and merge adjacent coordinate points to obtain the coordinates of the current tornado vortex center; the coordinates of the recognition results of multiple time steps form a historical tornado coordinate sequence;

[0016] Step 5) Tornado path prediction:

[0017] The tornado path prediction problem is classified as a two-dimensional particle motion prediction problem. The tornado coordinate sequences that are filtered in the previous step and attributed to multiple different tornado processes are used as the input of the tornado path prediction algorithm to obtain the path direction of the tornado in the next 10 steps.

[0018] The specific algorithm of tornado path prediction algorithm is:

[0019] The particle filtering algorithm PF (Particle Filtering) is used to solve the nonlinear problem of tornado motion prediction. Based on the motion trajectory of observation points with noise in the past, the dynamic state of the nonlinear system is estimated according to the customized tornado dynamic model. After multiple iterations, the state at multiple moments in the future is obtained to achieve path prediction at multiple moments in the future.

[0020] Preferably, in the above technical solution, the specific steps of constructing the TD-Unet deep learning model in step 2 are as follows:

[0021] Step 2.1) Improve the CBAM (Convolutional Block Attention Module) attention module to highlight the key features of the data;

[0022] The specific method is to improve a plug-and-play attention module iCBAM based on the characteristics of tornado radar data; the channel attention mechanism in the module can highlight the key features of different observations and improve the ability to extract information; the spatial attention mechanism enables the model to focus on the features near the maximum value of the observation, adjust the importance of the features of each spatial position, so that the model can focus on the key information of the tornado occurrence area; effectively enhance the recognition effect of the tornado vortex center;

[0023] ,

[0024] ,

[0025] , where sigmod represents the sigmod activation function, MLP represents the fully connected layer, AvgPool and MaxPool represent the average pooling layer and the maximum pooling layer respectively. Represents a convolution operation with a window width and height of 23 by 23, is the channel attention feature extraction process, The spatial attention mechanism extraction process is the channel attention feature extraction process;

[0026] Step 2.2) Combine the iCBAM module with the CoordConv2D convolution to improve the Unet left encoding module into a TD-block convolution module to enhance feature extraction and spatial information;

[0027] The specific method is to combine the iCBAM module in the encoding block, first through the CoordConv2D convolution, and then through the residual connection to extract the tornado features; the position code is calculated according to the distance of the radar station: , ,

[0028] in, Indicates the calculation of the values ​​of each grid point in the 0th layer position code. represents the calculation of the first layer position encoding, and represents the radar site coordinates, Indicates the maximum scanning distance of the radar. is the proportionality factor so that the calculation result is within a reasonable range;

[0029] To enhance the spatial location information, the CoordConv2D convolution module is used to combine the data feature information and the position coding information. The data information and the position coding information are processed separately through two branches. The data feature information is reduced in dimension through the maximum pooling operation MaxPool to highlight the key features of the data. The position coding information is reduced in dimension through the average pooling operation AvgPool to reduce the loss of position coding information.

[0030] The specific feature extraction process of the TD-block convolution module is as follows:

[0031] , ,

[0032] ,in, represents the input feature map of TD-block, represents the input position code, represents the output feature map, represents the activation function, represents the position coordinate convolution, represents the iCBAM feature extraction process;

[0033] Step 2.3) Improve the data transmission mechanism of the U-Net decoding part to reduce redundant information and noise transmission; the U-Net model directly transmits the data processed by the encoder to the decoder through jump connections, and adds a CBAM attention module in the jump connection to enhance the selective transmission of features, so that the decoder can automatically focus on the most important part when processing features from the encoder, thereby suppressing interference from irrelevant areas; in each layer of jump connections, the contribution of features at different positions is dynamically adjusted by calculating weighted coefficients. In the decoding part, the data feature information is restored to the original resolution by upsampling the encoded feature map, and the correspondence between geographic coordinates and data is restored. In order to avoid the loss of detail information in the small-scale position encoding during the upsampling process, the position encoding in the decoding stage transmits this information through the jump connection of the left encoder, thereby ensuring that accurate spatial position information can be maintained when the image is restored.

[0034] The original U-Net model directly passes the data processed by the encoder to the decoder through a jump connection; however, the improved encoder part has been enhanced in information extraction capability. Directly passing unprocessed raw data through a jump connection may result in the transmission of noise and redundant information, thereby affecting the performance of the decoder; the present invention enhances the selective transmission of features by adding a CBAM attention module to the jump connection, so that the decoder can automatically focus on the most important part when processing features from the encoder, thereby suppressing interference from irrelevant areas. In each layer of jump connection, the contribution of features at different positions is dynamically adjusted by calculating weighted coefficients. In this way, the decoder can more accurately focus on the key features contained in the encoder when reconstructing the image. In the decoding part, the encoded feature map is restored to the original resolution by upsampling, and the correspondence between geographic coordinates and data is restored. In order to avoid the loss of detail information in the small-scale position coding during the upsampling process, the upsampling in the decoding stage does not directly use the position coding information, but transmits this information through the jump connection of the left encoder, thereby ensuring that accurate spatial position information can be maintained when the image is restored.

[0035] Preferably, in the above technical solution, step 4 comprises the following specific steps:

[0036] Step 4.1) First, filter the connected blocks of the recognition results. The shape and size of the false positive points are smaller than those of the hit points. Filter the false positive points according to the shape and size of the connected blocks.

[0037] Step 4.2) Secondly, using the identification result of each grid point as a weight, the weighted coordinates of the entire connected block are calculated as the coordinates of the tornado represented by the connected block; ,in, Indicates the first The weight of the grid points, and Respectively represent The latitude and longitude of each grid point, and Respectively represent the weighted latitude and longitude coordinates of the entire connected block;

[0038] Step 4.3) Due to the existence of multiple tornadoes, it is necessary to attribute the tornado coordinates identified at multiple time steps to multiple tornado processes to obtain the coordinate sequences of multiple tornadoes;

[0039] The specific algorithm flow is as follows: traverse all tornado coordinates identified from a single base data file. When there is no tornado process currently recorded, a tornado process is created and the coordinates are added to the tornado process, and the tornado process is marked. The marked tornado process will no longer add new coordinates in this attribution process, and all marks will be cancelled before the next attribution process; when a tornado process exists, the distance between the coordinate and the last coordinate of all tornado processes is calculated. If the closest distance is less than a threshold, the tornado process with the closest distance is selected, the coordinates are added to the tornado process, and the tornado process is marked; if the closest distance is greater than a threshold, a new tornado process is created, the coordinates are added to the tornado process, and the tornado process is marked; at this point, the tornado attribution process of a single time is completed, and multiple tornado process coordinate sequences are obtained;

[0040] The whole process is expressed as:

[0041] initialization: (No tornado process);

[0042] For each coordinate Traversal:

[0043] if Is empty: Create a new tornado process and mark it; Add new procedures to the collection;

[0044] if Not null: and is not marked;

[0045] Select the minimum distance;

[0046] if : The closest tornado process will join in , and mark ;

[0047] if : Create a new tornado process and mark it; Add new processes to the collection; unmark all tornado processes in preparation for the next time coordinate attribution.

[0048] Preferably, in the above technical solution, the algorithm of step 5 is specifically:

[0049] Step 5.1) The basic steps of the particle filter algorithm PF (Particle Filter) are:

[0050] Initialization: Before starting the particle filter algorithm, the system state needs to be initialized; use the prior probability distribution to initialize the state of the particles;

[0051] Prediction: For each particle, its state at the next moment needs to be predicted based on the dynamic model of the system; this process can be expressed as: ,in Indicates Particles at time The predicted state of The dynamic model of the system is the state transition equation. Indicates the control input ;

[0052] Status update: based on observed values and the predicted state of each particle To update the particle weight;

[0053] For each particle, its state is updated according to the observed data; this process can be expressed as:

[0054] Calculate how well each particle matches the actual observation: ,

[0055] here It is a particle The observation probability of

[0056] Then, normalize the weights of all particles so that the total weight sums to 1: ,in, is the normalized particle weight;

[0057] Resampling: Resample the particles according to the weight size, remove the particles with smaller weights, retain the particles that contribute significantly to the results, and reduce the consumption of computing resources; the resampling process can be expressed as: ,in, Represents the time after resampling No. Particle state;

[0058] State prediction: The system state is estimated by resampling the particle state; this process can be expressed as: ,in, Indicates time System state estimation;

[0059] Repeat the above process to obtain the future multi-step state prediction;

[0060] Step 5.2) The specific state transition equation of the tornado path prediction system is: , in which the control input speed is added , acceleration ; Specifically represents the particle at time Two-dimensional vector velocity and acceleration; significantly improve the prediction accuracy and stability;

[0061] The specific calculation is as follows: The speed expression is as follows: , ;

[0062] The speed is calculated as the weighted sum of the nearest coordinate difference, the mean of the first and last coordinate differences, and the speed at the previous moment; it helps to smooth speed changes when the path is long, improve the accuracy of dynamic system prediction, smooth the estimate at the current moment, reduce the impact of noise on the prediction results, and enhance the ability to capture global trends.

[0063] Calculation of acceleration: , ; The average of historical acceleration is used as the acceleration at the current moment. By introducing acceleration, the model can capture the acceleration or deceleration trend of tornado path changes; and effectively deal with turns and irregular movements in path prediction.

[0064] Step 5.3) Calculate the path prediction probability field based on the particle position after state prediction:

[0065] Main elements:

[0066] Particle position: The state of a particle is represented by a set of positions (longitude and latitude), denoted by ,in It is The longitude of a particle, is its latitude;

[0067] Gaussian kernel function: used to describe the influence of each particle on a certain position; use Gaussian kernel function to calculate the contribution of particles to a certain position; the general form of Gaussian kernel function is: ,in is the calculated target location (latitude and longitude); is the current position of the particle; is the bandwidth parameter, which controls the extent of the Gaussian kernel expansion;

[0068] Probability field calculation: For each position on the grid, calculate the contribution of all particles to that position and add them up; this is because each particle affects the probability of the surrounding area, together forming a probability field;

[0069] There are N particles and the grid locations where the probability field is to be calculated , then the probability value at each grid point is It can be calculated by the following steps: The formula description of the probability field:

[0070] Calculate the contribution of each particle: For grid points The contribution of is based on the Gaussian kernel function: ,here, It is Particle pair position Contribution is the weight calculated by the Gaussian kernel;

[0071] Total probability field: for all grid points , calculate their probability values: , that is, the grid point The probability value of all particles is the sum of its contributions;

[0072] Normalized probability field: Normalize the probability value of each grid point so that the value ranges between 0 and 1 based on the original distribution; ; As the number of iterations increases, the bandwidth of the Gaussian kernel function is gradually increased. This will achieve a more focused prediction result in the near distance and a more divergent prediction result in the far distance, thus enhancing the early warning effect.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] The weakly supervised tornado recognition algorithm based on X-band radar base data proposed in this invention is

[0075] At the algorithm implementation level, the use of weakly supervised learning significantly reduces the difficulty of building a tornado deep learning dataset. By reducing the reliance on high-quality labeled data, weakly supervised learning not only optimizes the data labeling process, but also enables the model to effectively learn from limited label information in the absence of sufficient labeled samples. This method effectively solves the bottleneck problem of data acquisition and labeling in tornado detection, and provides greater flexibility and scalability for the construction of large-scale datasets and model training.

[0076] In terms of recognition effect, compared with the traditional tornado recognition algorithm, this model significantly improves the recognition accuracy and greatly reduces the false alarm rate. In addition, the optimized algorithm architecture and feature extraction method effectively improve the robustness of the system in complex environments, ensuring that a high level of recognition performance can be maintained under various meteorological conditions.

[0077] In the improved recognition model TD-Unet for tornadoes: by improving the CNN attention mechanism based on the characteristics of tornado radar data, and combining the spatial position encoding information to build the TD-block module, the model's feature extraction and spatial information are enhanced; the data transmission mechanism of the U-Net decoding part is improved to reduce redundant information and noise transmission. These improvements enable the model to show stronger recognition accuracy and lower false alarm rate in practical applications, greatly improving the reliability and practicality of the early warning system.

[0078] Experiments were conducted on radar data from Guangdong, and the model was also trained on the Tornet tornado public dataset and evaluated on the test set. The experimental results showed that the tornado recognition model had a recall rate of 85% with a false alarm ratio of 14.3%.

[0079] At the implementation level, the calculation parameters of this model only occupy 4M. Due to the small amount of model calculation, it significantly reduces resource consumption and provides strong support for the synchronous deployment of multiple radar systems.

[0080] In practical applications, this model can not only provide more accurate tornado identification results, but also meet the equipment's stringent requirements for low latency and low resource consumption, significantly improving the real-time processing capabilities and deployment flexibility of multi-radar systems, and further promoting the application of radar-based tornado warning systems.

[0081] The tornado path prediction algorithm based on particle filtering algorithm proposed in the present invention:

[0082] In order to better adapt to the movement characteristics of tornadoes, the present invention improves the basic particle filter algorithm and introduces speed and acceleration control quantities in its state transition. This improvement improves the algorithm's adaptability to the tortuosity of the tornado path, because the path of a tornado is usually not a straight line, but has twists and turns, turns, and even sudden changes in direction. By introducing speed and acceleration, the particle filter algorithm can take into account the inertia of the tornado's movement, ensuring that its movement direction and speed will not change dramatically in a short period of time, thereby more accurately tracking and predicting the path of the tornado.

[0083] In addition, the improved algorithm has also been improved in terms of the smoothness of motion prediction. Since the changes in the tornado's path may be sudden, relying solely on position coordinates for prediction may result in an uneven path or even unreasonable jumps. After introducing speed and acceleration, the algorithm can more smoothly describe the movement of the tornado, avoid abnormal fluctuations in path prediction, and ensure a more natural and reasonable trajectory.

[0084] Finally, the introduction of speed and acceleration also enhances the algorithm's adaptability to multi-scale motion. The movement of a tornado may show characteristics of different scales, including small fluctuations in a short period of time (such as acceleration fluctuations caused by changes in wind speed) and long-term stable movement (such as a tornado steadily advancing in a certain direction). The combination of speed and acceleration helps the algorithm capture these multi-scale motion patterns, enabling it to show better adaptability and prediction accuracy in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a flowchart of the overall steps;

[0086] Figure 2 This is the network structure diagram of the tornado identification model TD-Unet;

[0087] Figure 3 This is the TD-block structure diagram;

[0088] Figure 4 This is the structure diagram of the improved spatial attention module;

[0089] Figure 5-Figure 7 It is Tornet public data reflectivity, radial velocity and identification result map;

[0090] Figure 8-Figure 10 This is the X-band radar reflectivity, radial velocity and identification result diagram of Guangdong;

[0091] Fig.11 This is the latitude and longitude coordinate result map extracted from the base data recognition result;

[0092] Fig.12 is the prediction graph based on the observations of the first 5 steps;

[0093] Fig.13 is the prediction graph based on the observations of the first 9 steps;

[0094] Fig.14 is the prediction graph based on the previous 11 observations. DETAILED DESCRIPTION

[0095] The specific embodiments of the present invention are described in detail below, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

[0096] Unless explicitly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising”, etc., will be understood to include the stated elements or components but not to exclude other elements or components.

[0097] Example 1

[0098] The steps of the present invention are as follows Figure 1 As shown in the figure, six data types, namely, reflectivity factor (DBZ), radial velocity (VEL), phase difference (KDP), correlation coefficient (RHOHV), differential reflectivity (ZDR), and spectral width (WIDTH) at two low elevation angles are obtained from the X-band meteorological radar base data. A tornado identification data set is made, a tornado identification model is constructed and trained, and the X-band radar base data is input into the trained identification model to obtain the tornado identification results and identification azimuth map of a single time. The identification result coordinates of multiple times are extracted to obtain multiple tornado coordinates of each time, and the coordinates of adjacent tornadoes are merged into the same tornado process as the coordinate sequence of a tornado. A tornado path prediction algorithm based on the particle filter algorithm is constructed, and the obtained multiple tornado coordinate sequences are respectively input into the tornado path prediction algorithm for prediction and adjustment to obtain the tornado path prediction probability field.

[0099] Six types of data, including reflectivity factor (DBZ) at two low elevation angles, radial velocity (VEL), phase difference ratio (KDP), correlation coefficient (RHOHV), differential reflectivity (ZDR), and spectrum width (WIDTH), are obtained from the X-band meteorological radar to construct a deep learning tornado identification dataset.

[0100] Improve the CNN attention mechanism and use the residual structure and position encoding information to construct the TD convolution block; Figure 3 As shown, the feature extraction process of each individual TD convolution block is as follows: , , ,in, represents the input feature map of TD-block, represents the input position code, represents the output feature map, represents the activation function, CoordConv represents the position coordinate convolution, Represents the improved CNN attention mechanism feature extraction;

[0101] The improved CNN convolutional attention mechanism feature extraction process is as follows: , , ,in is the channel attention feature extraction process, Extraction process for the spatial attention mechanism.

[0102] The improved spatial attention mechanism for tornado radar data features is as follows: Figure 4 As shown, the model can focus on the features near the maximum value of the observed quantity, effectively enhancing the recognition effect of the tornado vortex center.

[0103] Based on the Unet model and combined with TD-block, a tornado recognition model TD-Unet is constructed. Figure 2 As shown. The encoding module in the original Unet model is replaced with the TD-block module constructed by the present invention to optimize the information transmission process: different data use different downsampling processes, feature information uses maxpool, and position encoding information uses avgpool. The information transmission process to the decoder is improved, and the redundant information transmission is reduced by adding the CBAM attention mechanism. In the encoding part, the position encoding information is directly transmitted through the jump connection of the encoding part without going through the upsampling process.

[0104] The six observation data at two low elevation angles of the X-band radar base data are input into the tornado recognition model to obtain the final recognition result. The experimental verification results are as follows: Figure 5-Figure 7 As shown, Figure 5 and Figure 6 The reflectivity data and radial velocity diagram of a tornado are shown respectively. The tornado occurred in the center of the diagram. The identification results are as follows: Figure 7 As shown in the figure, the tornado weather event was accurately identified and located. Combining the comparison of various indicators of the recognition model experimental results in Table 1, the model recall rate (Recall) of the present invention is 0.85, and the false alarm rate (False Alarm Ratio) is 0.143, which are better than other Unet models. Experiments were conducted on Guangdong radar data, such as Figure 8 , Fig. 9 and Fig.10 As shown, Figure 8 and Fig. 9 The reflectivity data and radial velocity diagram of the tornado weather on June 16, 22 are shown respectively. Fig.10 To demonstrate the model’s identification results, two tornadoes in the north and northwest directions were successfully hit.

[0105] Table 1 Recognition model experimental results indicators

[0106] .

[0107] The identification results are further processed to prepare for path prediction. The false positive results are filtered out by the size of the connected blocks in the identification results, and the longitude and latitude coordinates of all tornadoes at a single time are obtained by weighted average coordinates. The partial tornado coordinates extracted at each time of the tornado process on June 16, 22 are as follows: Fig.11 As shown in the figure, multiple tornado coordinates were identified for 5 consecutive minutes. The coordinates of these adjacent tornadoes are merged into one tornado process according to the latitude and longitude distance to obtain the coordinate sequence of multiple tornadoes. The specific calculation of the threshold is: , where the value represents the 1-minute time resolution identification method. Under normal circumstances, a tornado moves an average of 15 meters per second, and 1 meter corresponds approximately to 0.00001 in longitude and latitude.

[0108] Finally, each tornado coordinate sequence is input into the tornado path prediction algorithm based on the particle filter algorithm to obtain the path prediction of the tornado in the future. Figure 12-14 As shown in the figure, the green path represents the historical path that the tornado has traveled, and is also the input of the tornado path prediction algorithm based on the particle filter algorithm. In order to reflect the prediction effect, the black complete path composed of all the extracted tornado coordinates is drawn in advance. The red probability field is the prediction output after the tornado path prediction algorithm based on the particle filter algorithm is transformed.

[0109] The following are prior art documents related to this application:

[0110] 1. (CN202210298767.1) / Growth of hail embryos and forecasting method of hailfall stage during hailstorm in complex terrain / Applicant: Chengdu University of Information Technology / Application date: 2022.03.23

[0111] 2. (CN202311005168.7) / A dual-polarization radar raindrop spectrum parameter inversion method / Applicant: Suzhou University of Science and Technology / Application date: 2023.08.10

[0112] 3. (CN201910811238.5) / A hailstorm early warning method in complex terrain areas / Applicant: Chengdu University of Information Technology / Application date: 2019.08.30

[0113] 4. Yu Xiaoding, Zheng Yongguang. Research and operational progress of severe convective weather in contemporary China[J]. Acta Meteorologica Sinica, 2020, 78(3): 391-418.

[0114] 5. ZHOU Yong, YANG Bo, TANG Wei, LI Yue'an, WANG Yuedong, LI Zechun. Overview of development of public safety-oriented meteorological monitoring, forecasting and early warning software[J]. China Safety Science Journal, 2022, 32(9): 76-85.

[0115] 6. Yang Hequn, Wang Xiaofeng, Gao Yanqing, Lu Yiwen, Ma Bingxin, Wang Xinyao. Analysis of the demand for satellite big data in numerical weather forecasting[J]. Big Data, 2022, 8(2): 89-102.

[0116] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims

1. A method for tornado identification and path prediction based on X-band dual-polarization radar data, characterized in that: The following steps are involved: Step 1) Use X-band radar data to build a tornado identification deep learning dataset: First, with the tornado as the center, a fixed-size data block (W, H) around the tornado is intercepted from the six basic data at the two lowest elevation angles of the meteorological radar base data. The above 12 data blocks are combined into a positive sample with a sample shape of (12, W, H). Each observation is then normalized separately, and the negative sample intercepts the data of the random weather process that does not contain the tornado. Step 2) Construct a deep learning model for tornado identification: Based on the U-Net model, the encoder-decoder structure extracts features and restores spatial resolution. On this basis, the improved CBAM module, CoordConv2D convolution, and spatial position encoding are combined to build the TD-block module through jump connections to replace the original encoding module in Unet. The CBAM module is added to the jump connection of Unet to build a weakly supervised lightweight tornado recognition model TD-Unet, which is used to specifically identify tornadoes. The specific method of weakly supervised learning for tornado recognition is that the model is trained by the label of whether there is a tornado, without marking the specific location, shape or size of the tornado, so that the model can autonomously learn the characteristics of the tornado and obtain the location where the tornado occurs. Step 3) The data set obtained in step 1) is divided into a training set and a test set in a ratio of 4:1, and the training set is input into the constructed TD-Unet model for training; the performance of the trained model is evaluated on the test set using a binary classification confusion matrix to obtain an objective evaluation index; Step 4) extract the tornado coordinates from the recognition results, and convert the coordinates recognized at multiple time steps into a tornado coordinate sequence; eliminate the false positives in the recognition results, and in the case of multiple tornadoes, predict the coordinates of different tornadoes separately, and merge adjacent coordinate points to obtain the coordinates of the current tornado vortex center; the coordinates of the recognition results of multiple time steps constitute a historical tornado coordinate sequence; Step 5) Tornado path prediction: The tornado path prediction problem is classified as a two-dimensional particle motion prediction problem. The tornado coordinate sequences that are filtered in the previous step and attributed to multiple different tornado processes are used as the input of the tornado path prediction algorithm to obtain the path direction of the tornado in the next 10 steps. The specific algorithm of tornado path prediction algorithm is: The particle filtering algorithm is used to solve the nonlinear problem of tornado motion prediction, and the state transfer equation is improved. According to the historical motion state of the tornado, the speed and acceleration at the current moment are obtained. The speed and acceleration are used as parameters of the state transfer equation to obtain a custom tornado dynamic model. Based on the motion trajectory of the observation point with noise in the past, the dynamic change state of the nonlinear system is estimated according to the custom tornado dynamic model. After multiple iterations, the state at multiple moments in the future is obtained to achieve path prediction at multiple moments in the future.

2. The method for tornado identification and path prediction based on X-band dual-polarization radar data according to claim 1 is characterized in that: In step 1, the ratio of the number of positive and negative samples is 1:1; then create an index file for the positive and negative samples and mark them. If a tornado occurs in the data block, the sample is marked as 1, otherwise it is marked as 0; finally, the data block is divided into training set and test set in a ratio of 4:1, and the ratio of positive and negative samples in the training set and test set is also 1:

1.

3. The method for tornado identification and path prediction based on X-band dual-polarization radar data according to claim 1 is characterized in that: The specific steps for building the TD-Unet deep learning model in step 2 are as follows: Step 2.1) Improve the CBAM attention module to highlight the key features of the data; The specific method is to optimize the spatial attention module in CBAM to obtain the iCBAM module based on the characteristics of tornado radar data; the channel attention mechanism in the module can highlight the key features of different observations and improve the information extraction ability; the spatial attention mechanism enables the model to focus on the features near the maximum value of the observation, adjust the importance of the features of each spatial position, so that the model can focus on the key information of the tornado occurrence area; effectively enhance the recognition effect of the tornado vortex center; ， ， , where sigmod represents the sigmod activation function, MLP represents the fully connected layer, AvgPool and MaxPool represent the average pooling layer and the maximum pooling layer respectively. Represents a convolution operation with a window width and height of 23 by 23, is the channel attention feature extraction process, The spatial attention mechanism extraction process is the channel attention feature extraction process; Step 2.2) Combine the iCBAM module with the CoordConv2D convolution to improve the Unet left encoding module into a TD-block convolution module to enhance feature extraction and spatial information; The specific method is to combine the iCBAM module in the encoding block, first through the CoordConv2D convolution, and then through the residual connection to extract the tornado features; the position code is calculated according to the distance of the radar station: , , in, Indicates the calculation of the values ​​of each grid point in the 0th layer position code. represents the calculation of the first layer position encoding, and represents the radar site coordinates, Indicates the maximum scanning distance of the radar. is the proportionality factor so that the calculation result is within a reasonable range; To enhance the spatial location information, the CoordConv2D convolution module is used to combine the data feature information and the position coding information. The data information and the position coding information are processed separately through two branches. The data feature information is reduced in dimension through the maximum pooling operation MaxPool to highlight the key features of the data. The position coding information is reduced in dimension through the average pooling operation AvgPool to reduce the loss of position coding information. The specific feature extraction process of the TD-block convolution module is as follows: , , ,in, represents the input feature map of TD-block, represents the input position code, represents the output feature map, represents the activation function, represents the position coordinate convolution, represents the iCBAM feature extraction process; Step 2.3) Improve the data transmission mechanism of the U-Net decoding part to reduce redundant information and noise transmission; the U-Net model directly transmits the data processed by the encoder to the decoder through skip connections, and adds a CBAM attention module in the skip connection to enhance the selective transmission of features, so that the decoder can automatically focus on the most important part when processing features from the encoder; in each layer of skip connection, the contribution of features at different positions is dynamically adjusted by calculating weighted coefficients; in the decoding part, the data feature information is restored to the original resolution of the encoded feature map through upsampling, and the correspondence between geographic coordinates and data is restored; and the position code in the decoding stage is directly transmitted through the skip connection of the left encoder.

4. The method for tornado identification and path prediction based on X-band dual-polarization radar data according to claim 1, characterized in that: Step 4 The specific steps are: Step 4.1) First, filter the connected blocks of the recognition results. The shape and size of the false positive points are smaller than those of the hit points. Filter the false positive points according to the shape and size of the connected blocks. Step 4.2) Secondly, using the identification result of each grid point as a weight, the weighted coordinates of the entire connected block are calculated as the coordinates of the tornado represented by the connected block; ;in, Indicates the first The weight of the grid points, and Respectively represent The latitude and longitude of each grid point, and Respectively represent the weighted latitude and longitude coordinates of the entire connected block; Step 4.3) Due to the existence of multiple tornadoes, it is necessary to attribute the tornado coordinates identified at multiple time steps to multiple tornado processes to obtain the coordinate sequences of multiple tornadoes; The specific algorithm flow is as follows: traverse all tornado coordinates identified from a single base data file. When there is no tornado process currently recorded, a tornado process is created and the coordinates are added to the tornado process, and the tornado process is marked. The marked tornado process will no longer add new coordinates in this attribution process, and all marks will be cancelled before the next attribution process; when a tornado process exists, the distance between the coordinate and the last coordinate of all tornado processes is calculated. If the closest distance is less than a threshold, the tornado process with the closest distance is selected, the coordinates are added to the tornado process, and the tornado process is marked; if the closest distance is greater than a threshold, a new tornado process is created, the coordinates are added to the tornado process, and the tornado process is marked; at this point, the tornado attribution process of a single time is completed, and multiple tornado process coordinate sequences are obtained; The whole process is expressed as: initialization: , no tornado process; For each coordinate Traversal: if Is empty: Create a new tornado process and mark it; Add new procedures to the collection; If not empty: and is not marked; Select the minimum distance; if : The closest tornado process will join in , and mark ; if : Create a new tornado process and mark it; Add new processes to the collection; unmark all tornado processes in preparation for the next time coordinate attribution.

5. The method for tornado identification and path prediction based on X-band dual-polarization radar data according to claim 1, characterized in that: The specific algorithm of step 5 is: Step 5.1) The basic steps of particle filtering algorithm are: Initialization: Before starting the particle filter algorithm, the system state needs to be initialized; use the prior probability distribution to initialize the state of the particles; Prediction: For each particle, its state at the next moment needs to be predicted based on the dynamic model of the system; this process can be expressed as: ,in Indicates Particles at time The predicted state of The dynamic model of the system is the state transition equation. Indicates the control input quantity; Status update: based on observed values and the predicted state of each particle To update the particle weight; For each particle, its state needs to be updated based on the observed data; This process can be expressed as: Calculate how well each particle matches the actual observation: ,here It is a particle The observation probability of Then, normalize the weights of all particles so that the total weight sums to 1: ,in, is the normalized particle weight; Resampling: Resample the particles by weight, remove particles with smaller weights, and retain particles that contribute significantly to the results; the resampling process can be expressed as: ,in, Represents the time after resampling No. Particle state; State prediction: The system state is estimated by resampling the particle state; this process can be expressed as: ,in, Indicates time The system state estimation is repeated to obtain the future multi-step state prediction; Step 5.2) The specific state transition equation of the tornado path prediction system is: , in which the control input speed is added , acceleration , specifically represents the particle at time Two-dimensional vector velocity and acceleration; The specific calculation is as follows: The speed expression is as follows: , ; The speed is calculated by the weighted sum of the nearest coordinate difference, the mean of the first and last coordinate differences, and the speed at the previous moment; The calculation expression of acceleration is as follows: , ; The average of historical acceleration is used as the acceleration at the current moment. By introducing acceleration, the model can capture the acceleration or deceleration trend of the tornado path change; Step 5.3) Calculate the path prediction probability field based on the particle position after state prediction: Main elements: Particle position: The state of a particle is represented by a set of positions, denoted by ,in It is The longitude of a particle, is its latitude; Gaussian kernel function: used to describe the influence of each particle on a certain position; use Gaussian kernel function to calculate the contribution of particles to a certain position; the general form of Gaussian kernel function is: ,in is the calculated target location (latitude and longitude); is the current position of the particle; is the bandwidth parameter, which controls the extent of the Gaussian kernel expansion; Probability field calculation: For each position on the grid, calculate the contribution of all particles to that position and add them up to form a probability field; There are N particles and the grid locations where the probability field is to be calculated , then the probability value at each grid point is It can be calculated by the following steps: The formula description of the probability field: Calculate the contribution of each particle: For grid points The contribution of is based on the Gaussian kernel function: ,here, It is Particle pair position Contribution are the weights calculated by the Gaussian kernel; Total probability field: for all grid points , calculate their probability values: , that is, the grid point The probability value of all particles is the sum of its contributions; Normalized probability field: Normalize the probability value of each grid point so that the value ranges from 0 to 1 based on the original distribution; ; As the number of iterations increases, the bandwidth of the Gaussian kernel function is gradually increased.

Images