Machine learning tornado identification method driven by storm radar characteristic quantity

Abstract

The invention discloses a storm radar characteristic quantity driven machine learning tornado identification method, and belongs to the field of meteorological disaster weather identification, and the method comprises the following steps: obtaining a storm radar characteristic quantity containing key characteristic information through characteristic extraction and optimization processing based on body sweeping radabundance data; performing feature quantity normalization processing and feature quantity missing value processing on the storm radar feature quantity to obtain an input sequence containing a plurality of elevation angle layer data; and inputting the input sequence into a recognition model fusing Transform and CNN, and outputting a tornado recognition result. According to the method, the technical problems of feature redundancy, insufficient local and global feature fusion, data missing sensitivity and the like in a traditional tornado identification method are effectively solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meteorological disaster weather identification, and particularly relates to a tornado identification method driven by storm radar characteristic quantity and machine learning. BACKGROUND

[0002] Tornado is a funnel-shaped violent rotating air current extending from the bottom of thunderstorm cloud to the ground, and is a small-scale disastrous weather with extremely destructive power. At present, the monitoring and near-forecasting of tornado mainly rely on high-temporal and high-spatial resolution weather radar data. Based on the radar data, various tornado identification technologies have been developed in the meteorological field, mainly including fuzzy logic methods based on low-level mesocyclone, tornado vortex characteristic, tornado debris characteristic and other radar characteristic parameters, and machine learning methods gradually emerging in recent years. The fuzzy logic method relies on artificial expert experience to construct a rule base, and has certain effect in the scene with clear rules and low dimension. The machine learning method, especially the traditional machine learning model, shows greater potential in dealing with the extreme weather identification problem of tornado with high dynamics, multiple variables and small samples.

[0003] However, the existing machine learning identification method still has obvious limitations. On the one hand, there are many redundant characteristic quantities extracted by radar, and directly using all the characteristics will introduce noise, increase the complexity of the model, and affect the generalization ability. On the other hand, the vortex characteristic of tornado has complex spatial structure and multi-level atmospheric vertical structure information, and the traditional machine learning model is difficult to effectively capture the deep spatial correlation and sequence dependence relationship. In addition, the radar data itself has problems such as ground clutter, velocity ambiguity, beam blockage and data missing, which further increases the difficulty of accurately identifying tornado. Therefore, there is an urgent need for a new identification method that can automatically filter key features and effectively fuse local details and global dependence relationships. SUMMARY

[0004] To solve the above technical problems, the application provides a machine learning tornado identification method driven by storm radar characteristic quantity to solve the problems existing in the prior art.

[0005] To achieve the above purpose, the application provides a machine learning tornado identification method driven by storm radar characteristic quantity, comprising: Based on the body scanning radar base data, the storm radar characteristic quantity containing key feature information is obtained through feature extraction and optimization processing; The storm radar characteristic quantity is subjected to feature quantity normalization processing and feature quantity missing value processing to obtain an input sequence containing several elevation angle layer data; The input sequence is input into the identification model of fusion Transformer and CNN to output the tornado identification result.

[0006] Optionally, the process of obtaining storm radar feature quantities containing key feature information through feature extraction and optimization based on the volume scan radar base data includes: A storm cell identification and tracking algorithm was used to identify three-dimensional storm cells from volume scan radar base data; Based on radial velocity data, the azimuth vorticity shear value is calculated using the two-dimensional local linear least squares derivative method. Based on the azimuth vortex shear value, a multi-threshold identification process is used to identify the three-dimensional vortex features to obtain the three-dimensional vortex features. The identified three-dimensional vortex features are spatially correlated with three-dimensional storm cells, and several two-dimensional storm component features of the correlated storm cells are extracted. Based on tornado observation records, a gradient boosting decision tree model is trained. The importance of the two-dimensional storm component features is ranked using the trained gradient boosting decision tree model to obtain storm radar features.

[0007] Optionally, the process of identifying three-dimensional storm cells from volume scanned radar base data using a storm cell identification and tracking algorithm includes: Based on low elevation angle reflectivity factor data, identify two-dimensional storm components that exceed a preset reflectivity threshold and satisfy spatial continuity. Based on the two-dimensional storm components, a three-dimensional storm unit is constructed through vertical correlation of multiple elevation angles; Based on the three-dimensional storm cell, the movement path and life history information of the storm cell are tracked by utilizing the spatial overlap relationship between the continuous volume scan data.

[0008] Optionally, the process of handling missing values ​​of the storm radar features includes: when all features of the current elevation layer in the three-dimensional vortex features have no missing values, but the total number of elevation layers is less than a preset number of layers, a preset negative value is used to fill the missing elevation layers.

[0009] Optionally, the process of inputting the input sequence into the recognition model that fuses Transformer and CNN to output the tornado recognition result includes: Based on the input sequence, a one-dimensional convolutional neural network is used to increase the vector dimension from 14 to 120, resulting in the first dimension-enhanced feature. Based on the enhanced features of the first dimension, a Transformer encoder is used to encode the features to obtain the encoded sequence features. Based on the encoded sequence features, a one-dimensional convolutional neural network is used again to increase the vector dimension from 120 dimensions to 384 dimensions, resulting in the second dimension-enhanced features. Based on the enhanced features of the second dimension, a fully connected layer is used for processing to output the tornado identification results.

[0010] Optionally, the Transformer encoder of the recognition model that integrates Transformer and CNN consists of 6 stacked sub-encoders, and the self-attention module of each sub-encoder adopts a 4-head self-attention mechanism.

[0011] The present invention also provides a machine learning tornado recognition system driven by storm radar features, for implementing the method, the system comprising: The feature extraction and optimization module is used to obtain storm radar feature quantities containing key feature information based on volume scan radar base data through feature extraction and optimization processing. The data preprocessing module is used to perform feature normalization and missing feature value processing on the storm radar features to obtain an input sequence containing several elevation angle layer data. The tornado recognition module is used to input the input sequence into a recognition model that fuses Transformer and CNN to output tornado recognition results.

[0012] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described thereon.

[0013] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0014] The present invention also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the steps of the method.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: By constructing a complete radar data processing and feature optimization workflow, efficient extraction and accurate identification of tornado features were achieved. Ground object suppression and radial velocity deblurring techniques were employed to effectively improve the quality of radar data, laying a reliable foundation for subsequent feature extraction. Through feature extraction and optimization, the most discriminative key features were selected from complex meteorological data, significantly improving feature representation capabilities. Feature normalization and missing value handling ensured the standardization and completeness of input data, enhancing model stability. An innovative recognition model integrating Transformer and CNN was adopted, fully leveraging the advantages of CNN in local feature extraction and Transformer in global correlation modeling, enabling deep feature learning of multi-elevation-angle radar data. This method effectively improves the accuracy and reliability of tornado identification, providing strong technical support for meteorological disaster early warning and possessing significant application value. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a diagram illustrating the TFM+CNN tornado recognition model framework and multi-head attention layer architecture of an embodiment of the present invention. Figure 2 This is a structural diagram of the storm radar feature parameter-driven TFM+CNN tornado identification device according to an embodiment of the present invention. Figure 3 This is a cumulative contribution diagram of principal component feature components in an embodiment of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0019] Example 1.

[0020] like Figure 1 and Figure 2As shown, this embodiment provides a machine learning-based tornado identification method driven by storm radar features. Based on volume scan radar data of tornadoes from 2003 to 2023, this embodiment identifies storms and extracts numerous storm attribute radar features. For tornadoes, it uses a gradient boosting decision tree (XGBoost) to rank the importance of radar features and establishes a tornado identification model using a TFM+CNN model to improve the accuracy of tornado storm identification. Specifically, it includes the following steps: 1. Calculation of storm radar characteristic quantities.

[0021] Based on volume scan radar base data, storm radar feature quantities containing key feature information are obtained through feature extraction and optimization processing. The specific implementation process includes steps a-c.

[0022] Furthermore, the process of obtaining storm radar feature quantities containing key feature information through feature extraction and optimization based on volume scan radar base data includes: identifying three-dimensional storm cells from the volume scan radar base data using storm cell identification and tracking algorithms; calculating azimuth vortex shear values ​​based on radial velocity data using the two-dimensional local linear least squares derivative method; identifying three-dimensional vortex features using a multi-threshold identification process based on the azimuth vortex shear values; spatially associating the identified three-dimensional vortex features with three-dimensional storm cells to extract several two-dimensional storm component feature quantities of the associated storm cells; training a gradient boosting decision tree model based on tornado observation records, and using the trained gradient boosting decision tree model to rank the importance of the two-dimensional storm component feature quantities to obtain storm radar feature quantities.

[0023] a. Storm Cell Identification and Tracking: Using the SCIT method for storm cell identification and tracking, firstly, two-dimensional storm components exceeding a preset reflectivity threshold (e.g., 35 dBZ) and satisfying certain spatial continuity are identified from the low elevation reflectivity factor data of a single radar volume scan. Subsequently, a three-dimensional storm cell is constructed through vertical correlation of multiple elevation layers, and the spatial overlap between continuous volume scan data is utilized to track the storm cell, thereby obtaining information such as the movement path and life history of each storm cell.

[0024] Furthermore, the process of identifying three-dimensional storm cells from volume scan radar base data using storm cell identification and tracking algorithms includes: identifying two-dimensional storm components that exceed a preset reflectivity threshold and satisfy spatial continuity based on low elevation angle reflectivity factor data; constructing three-dimensional storm cells based on the two-dimensional storm components through vertical correlation of multiple elevation angles; and tracking the movement path and life history information of storm cells based on the three-dimensional storm cells by utilizing the spatial overlap relationship between continuous volume scan data.

[0025] b. Three-dimensional vortex feature recognition and storm attribute extraction.

[0026] Azimuth vortex shear calculation: Based on defuzzified radial velocity data, the two-dimensional local linear least squares derivative (LLSD) method is used to estimate the vortex shear value in the azimuth direction point by point. This method obtains the derivative of the radial velocity in the azimuth direction by performing linear least squares fitting within a local region (such as a 5x5 grid window). The results are more robust and have stronger noise resistance than the traditional one-dimensional calculation method.

[0027] 3D Vortex Recognition: Based on the 3D azimuth vortex shear data field calculated above, a multi-threshold recognition process similar in structure to the SCIT algorithm is used for 3DVS recognition. The specific process is as follows: In the 3D mesh, from the highest threshold (70×10) -4 s -1 The scan begins, identifying the core points that satisfy the shear strength requirement. The threshold is then progressively decreased (in order: 60, 50, 40, 30, 20, 10×10). -4 s -1 It merges and expands the vortex regions identified by the previous threshold in space, thereby outlining a complete three-dimensional vortex structure with different intensity levels.

[0028] Feature Correlation and Extraction: The identified 3DVSs are spatially correlated with the storm cells identified by the SCIT algorithm. For each storm cell successfully correlated with a vortex, a series of attribute features are automatically extracted. This method mainly selects 2D storm component features layered by elevation angle, totaling 28 features. These features are mainly based on the parameters describing the 2D vortex characteristics (2DVS) of each layer (such as maximum shear value, vortex vertical thickness, and mid-level vortex intensity).

[0029] c. Feature Engineering Method Based on Combination of PCA and XGBoost Results: To optimize model input and improve tornado recognition performance, a feature engineering strategy is designed. The core of this strategy is to use XGBoost for feature ranking and then use PCA to verify the effectiveness of the ranking results. The specific method is as follows: Step 1: Feature importance ranking and filtering based on XGBoost.

[0030] Model Training and Importance Evaluation: An XGBoost classification model was trained using all 28 two-dimensional storm features as input and tornado observation records as labels. After training, the built-in feature importance metric was extracted. This metric is calculated based on the sum of information gain (Gain) of each feature used for splitting nodes across all decision trees, effectively measuring the contribution of each feature to tornado classification.

[0031] Feature ranking and initial screening: Based on the calculated importance scores, all 28 features are ranked from highest to lowest. Then, the top 14 features are initially selected to form a subset of candidate features. This step aims to filter out half of the features with lower contribution from the original feature set, achieving preliminary dimensionality reduction and noise reduction.

[0032] Step 2: Verify the validity of the feature subset based on PCA.

[0033] PCA Contribution Analysis: To verify the sufficiency of the above candidate feature subset, we performed principal component analysis (PCA) on it. Specifically, we performed PCA analysis only on the subset consisting of the first 14 important features and calculated the variance contribution rate of each principal component.

[0034] Results Interpretation and Final Confirmation: The PCA analysis results show that the cumulative variance contribution rate of the first few principal components corresponding to these 14 features exceeds 90%. This key indicator proves that this 14-dimensional feature subset can explain most of the information variation in the original 28-dimensional features.

[0035] Step 3: The first 14 most important raw features are directly used as input to the TFM+CNN model.

[0036] Relevant information is shown in Table 1, and the calculation formulas for some of the features are as follows: (1) (2) (3) (4) in j, k These are the azimuth and slant range numbers, respectively. Radial velocity (unit: m·s) -1 ), Slope distance (unit: km). Radial velocity is measured from the distance of the container (in km). V represents the azimuth difference (in radians) between adjacent radial directions. max and V min These represent the maximum and minimum radial velocities, respectively.

[0037] The formula for calculating excess rotational kinetic energy is as follows: (5) in air density (unit: kg∙m³) -3 (Using the internationally accepted standard atmospheric density) For thickness, since 2DVS is on a plane, it has no thickness, so we take... , The minimum threshold (10×10) -4 s -1 The two-dimensional shear characteristic radius (in meters). Rotational speed (unit: m·s) -1 ), To maintain the minimum shear threshold for moderate-intensity cyclones (adjustable, default value is 50 × 10⁻⁶). -4 s -1 The unit of excess rotational kinetic energy of a two-dimensional shear characteristic is kg∙m∙s. -2 That is, J∙m -1 .when At that time, x21 is considered to be 0. The larger the value of x21, the greater the rotational kinetic energy that exceeds the minimum mesocyclone requirement, and the more likely it is to form a tornado.

[0038] In polar coordinates, take a (2 N +1)×(2) M +1) A two-dimensional local window of size , where the radial velocity of each azimuth-slant range within the window is . Taking the center of the window as a reference, along the radial direction from the center, j For positive, rotate clockwise along the tangential direction. k If positive, then the azimuth vorticity shear at the center of the window is... The calculation formula is as follows: (6) in, , The slant range (in km) from the center of the local two-dimensional window to the radar. Beamwidth (in radians). As the weight, this method takes If there is cyclonic rotation in the velocity field, then .

[0039] Since the LLSD method uses polar coordinate radial velocity data to calculate azimuth vortex shear, its azimuth resolution varies with distance from the radar. As this changes, the size of the local window also changes accordingly. Changes. Let the size of the mesocyclone nucleus be a constant D (default 1.25 km), then the values ​​of N and M are as follows: (7) (8) Where B refers to the angular width of the electromagnetic wave beam emitted by the radar, W represents the window, i.e., the half-width of the local analysis window used to calculate azimuth vortex shear in the azimuth direction, and (int) represents rounding to the nearest integer. If Then let ;if Then let .

[0040] Table 1

[0041] The storm radar features are normalized and missing values ​​are processed to obtain an input sequence containing several elevation angle layer data. The specific implementation process includes steps 2-3.

[0042] 2. Normalization of Radar Quantities: Since the dynamic range of each radar characteristic value is very large, and some of them are different in magnitude, it is necessary to normalize the radar characteristic values ​​so that their range is controlled between (1.0, 2.0). The calculation method for normalization is shown in formula (9).

[0043] +1.0 (9) in , , , For a certain feature, there are the normalized value, current value, minimum value, and maximum value.

[0044] 3. Handling Missing Feature Values: 3DVS consists of multiple vertically correlated 2DVS. The number of 2DVS is related to the characteristics of the storm itself, and is also affected by factors such as beam blocking, range folding, and the distance of the storm from the radar. Therefore, different 3DVS will contain different numbers of 2DVS and have different elevation angles at the bottom layers, resulting in some parameters of a 3DVS with a fixed number of layers being missing. If a 2DVS in a 3DVS dataset has some missing features, the 3DVS dataset is discarded; this situation is relatively rare. If all features of each 2DVS in a 3DVS dataset are complete, but the number of 2DVS layers is less than the set number of data layers, a very small negative value (such as -1.0E9) is used to fill the missing data layers. This method of handling missing data is selected based on the characteristics of the TFM model. For example, -1.0E9 will result in a weight of 0 after passing through the Softmax function when calculating the attention score, thus completely masking the filling position and causing the model to ignore the non-existent layer. In a tree model, the choice of which branch to take is determined by the information gain of the nodes in the computation tree. The tree model always chooses the branch with the largest information gain, and branches filled with small values ​​are discarded because of their small computational gain. In a TFM model, the small values ​​can be replaced with missing values ​​from the system, such as torch.inf in PyTorch, while missing values ​​from the system are ignored (masked).

[0045] After the above data processing, this method established three types of storm radar feature datasets with a total of 7756 samples. The first type is the radar feature dataset for tornado-affected storms, with 924 samples. The datasets selected are those where the maximum adjacent azimuth velocity difference between the two lowest elevation angles is ≥20 m / s during or half an hour before the tornado occurs. -1 The dataset consists of three categories: 1) 3DVS (3D Response Values), calculated by summing the number of 3DVS values ​​for the same storm across different volume scans; 2) a dataset of radar features without tornadoes that has significant 3DVS, containing 1174 samples; and 3) a dataset of radar features without tornadoes that has no significant 3DVS, containing 5658 samples. Each sample contains 28 features, of which 14 are selected and used by the model.

[0046] The input sequence is fed into a recognition model that integrates Transformer and CNN to output tornado recognition results. The specific implementation process includes steps 4-6.

[0047] Furthermore, the process of inputting the input sequence into the recognition model that integrates Transformer and CNN to output the tornado recognition result includes: based on the input sequence, using a one-dimensional convolutional neural network to increase the vector dimension from 14 dimensions to 120 dimensions to obtain the first dimension-enhanced features; based on the first dimension-enhanced features, using a Transformer encoder to encode the features to obtain the encoded sequence features; based on the encoded sequence features, using a one-dimensional convolutional neural network again to increase the vector dimension from 120 dimensions to 384 dimensions to obtain the second dimension-enhanced features; based on the second dimension-enhanced features, using a fully connected layer for processing to output the tornado recognition result.

[0048] 4. Model Input Processing: The model initially consisted of 14 time-series vectors of elevation angles (the top 14 features in Table 1). Through extensive experimental analysis, it was found that using vector data from 14 elevation angles did not improve the model's prediction accuracy. The main reason is that the higher the elevation angle layer, the more missing values ​​there are, and the less information the model learns. Therefore, the feature values ​​from the lower 6 elevation angles were ultimately selected as the model input.

[0049] 5. Model Structure: A complete TFM architecture consists of two parts: an encoder and a decoder module. Since our task is simply to identify tornadoes, we will only select the encoder module, such as... Figure 1 As shown, its main function is to extract feature information for the decoder, but this information can also be used for classification and recognition. For the encoding module of the TFM model, each vector in the input six elevation angle sequences contains only 14 dimensions of features. Because the number of dimensions is too small, the amount of information contained is limited. If a multi-head self-attention mechanism is used for splitting, the information contained in each head is even more limited. Therefore, a one-dimensional CNN (CONV_1D) is used to increase the vector dimension from 14 to 120 dimensions, with the following parameter settings: batch: 32; Input dimension: (14, 1); Convolutional layer parameters: filters: 120 (target dimension); kernel_size: 3; padding: 'same'; activation: 'relu'; Output dimensions: (1, 120); After the information passes through the TFM encoding module, a one-dimensional CNN is used again to increase the vector dimension from 120 to 384 dimensions, and multiple fully connected layers are added at the end of the model. Since positional information is not included when increasing the vector dimension, sine / cosine functions are used to generate positional information. This positional information has the same dimension as the original feature values, so it is directly added to the original feature values ​​for training. The multi-head attention layer includes the calculation of K, Q, and V. The feature dimensions d of Q, K, and V are divided into h heads, and attention is calculated in parallel and then concatenated. Each head is N×d. k (d) k The output (for a single-head dimension) is concatenated into an N×d matrix, and then mapped back to the model dimension through a linear layer.

[0050] The specific segmentation method for multi-head attention is as follows: 1) Projection: Q, K, and V are generated through linear transformation.

[0051] 2) Reshaping and transposing: Physically divide the tensor of dimension d_model into h tensors of dimension d_k, and achieve parallel layout by transposing.

[0052] 3) Parallel attention computation: Each head computes attention in an independent subspace.

[0053] 4) Reverse reshaping and splicing: splice the outputs of h heads to restore a unified representation of the d_model dimension.

[0054] 5) Linear mapping: Perform the final transformation on the spliced ​​result.

[0055] 6. Model Output Products: Before the model experiment, principal component analysis was performed on datasets with and without tornadoes. It was found that the first 14 dimensions, as shown in Table 1, contributed nearly 100% of the total output, especially the first three dimensions, which contained approximately 73% of the information. Figure 3 As shown, some factors were highly correlated when selecting factors, so only the first 14 features were selected for learning.

[0056] Three binary classification models are used for tornado storm identification. The first binary classification classifies storms into those with tornadoes and those without significant 3DVS (3DV fluctuation). The second binary classification classifies storms into those with tornadoes and those without significant 3DVS. The third binary classification classifies storms into those with tornadoes and those without (including storms with / without significant 3DVS). The first binary classification is easily achieved using conventional parametric thresholding, but the second binary classification is difficult. In practical applications, conventional parametric thresholding can be used to exclude storms without significant 3DVS, and then machine learning can be used for the second binary classification.

[0057] As shown in Table 2, in the datasets with tornadoes and without significant 3DVS, the ratio of positive to negative samples is 1:6, with positive samples being tornadoes. The recognition accuracy on the test set is 0.978. In the datasets with tornadoes and without significant 3DVS, the ratio of positive to negative samples is close to 1:1, with a recognition accuracy of 0.863 on the test set. In the datasets with tornadoes and without tornadoes, the ratio of positive to negative samples is 1:6, with positive samples being tornadoes. The recognition accuracy on the test set is 0.96.

[0058] Table 2

[0059] The TFM parameters in the TFM+CNN model were determined experimentally, as shown in Table 3. The model parameters consisting of 6 sub-encoders and 4 self-attention heads yielded better results. The optimizer used was Adam, and the loss function was a binary cross-entropy loss function. Due to various reasons, some layers in the input layer may have missing data. A mask was used to replace the data in these layers, with the mask dimension matching the data dimension of that layer. If data was missing, it was filled with torch.inf, representing an infinitely large negative number, which was made to approach 0 during probability calculation using softmax. The model then masked this layer during computation. Furthermore, increasing the input layer dimension to 240 dimensions could lead to overfitting during training, as the available feature data was limited. After the input data passed through the model, 5 fully connected layers were added at the end to increase the model depth and improve recognition accuracy. The model was trained for 5000 epochs. After each training epoch, accuracy, precision, recall, and F1 score were calculated on the test data. The threshold for the model to identify the probability of tornado occurrence was determined based on the ratio of positive to negative samples in the dataset. The threshold was 0.86 on the dataset with tornadoes and no significant 3DVS (Dual Variable Spanning) storms, 0.56 on the dataset with tornadoes and significant 3DVS storms, and 0.88 on the dataset with tornadoes and no tornadoes. When the model was trained on these three data categories, the loss function value tended to stabilize after 3500 epochs.

[0060] Table 3

[0061] The advantage of this invention lies in the fact that it constructs a TFM+CNN deep learning model for tornado identification, mainly benefiting from the combination of self-attention mechanism and CNN to achieve "global correlation + local deepening" feature extraction. Experiments show that the most difficult distinction between tornadoes and non-tornado storms with significant 3DVS is the main cause of false alarms. The model demonstrates good early warning capabilities in typical tornado cases.

[0062] Example 2.

[0063] This embodiment also provides a machine learning tornado recognition system driven by storm radar features, for implementing the above method. The system includes: The feature extraction and optimization module is used to obtain storm radar feature quantities containing key feature information based on volume scan radar base data through feature extraction and optimization processing. The data preprocessing module is used to perform feature normalization and missing feature value processing on storm radar features to obtain an input sequence containing several elevation layer data. The tornado recognition module is used to input the input sequence into a recognition model that fuses Transformer and CNN to output tornado recognition results.

[0064] This embodiment also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described thereon.

[0065] This embodiment also provides a storage medium on which a computer program is stored, which, when executed by a processor, implements the method described thereon.

[0066] This embodiment also provides an electronic device, the electronic device including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the steps of the method.

[0067] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A machine learning-based tornado recognition method driven by storm radar features, characterized in that, Includes the following steps: Based on volume scan radar base data, storm radar feature quantities containing key feature information are obtained through feature extraction and optimization processing. The storm radar features are normalized and missing values ​​are processed to obtain an input sequence containing several elevation angle layer data. The input sequence is input into a recognition model that fuses Transformer and CNN to output tornado recognition results.

2. The machine learning tornado recognition method driven by storm radar feature quantities according to claim 1, characterized in that, Based on the volume scan radar base data, the process of obtaining storm radar feature quantities containing key feature information through feature extraction and optimization includes: A storm cell identification and tracking algorithm was used to identify three-dimensional storm cells from volume scan radar base data; Based on radial velocity data, the azimuth vorticity shear value is calculated using the two-dimensional local linear least squares derivative method. Based on the azimuth vortex shear value, a multi-threshold identification process is used to identify the three-dimensional vortex features to obtain the three-dimensional vortex features. The identified three-dimensional vortex features are spatially correlated with three-dimensional storm cells, and several two-dimensional storm component features of the correlated storm cells are extracted. Based on tornado observation records, a gradient boosting decision tree model is trained. The importance of the two-dimensional storm component features is ranked using the trained gradient boosting decision tree model to obtain storm radar features.

3. The machine learning tornado recognition method driven by storm radar features according to claim 2, characterized in that, The process of identifying three-dimensional storm cells from volume scan radar base data using storm cell identification and tracking algorithms includes: Based on low elevation angle reflectivity factor data, identify two-dimensional storm components that exceed a preset reflectivity threshold and satisfy spatial continuity. Based on the two-dimensional storm components, a three-dimensional storm unit is constructed through vertical correlation of multiple elevation angles; Based on the three-dimensional storm cell, the movement path and life history information of the storm cell are tracked by utilizing the spatial overlap relationship between the continuous volume scan data.

4. The machine learning tornado recognition method driven by storm radar feature quantities according to claim 1, characterized in that, The process of handling missing values ​​of the storm radar features includes: when all features of the current elevation layer in the three-dimensional vortex features have no missing values, but the total number of elevation layers is less than a preset number of layers, a preset negative value is used to fill the missing elevation layers.

5. The machine learning tornado recognition method driven by storm radar features according to claim 1, characterized in that, The process of inputting the input sequence into a recognition model that fuses Transformer and CNN to output tornado recognition results includes: Based on the input sequence, a one-dimensional convolutional neural network is used to increase the vector dimension from 14 to 120, resulting in the first dimension-enhanced feature. Based on the enhanced features of the first dimension, a Transformer encoder is used to encode the features to obtain the encoded sequence features. Based on the encoded sequence features, a one-dimensional convolutional neural network is used again to increase the vector dimension from 120 dimensions to 384 dimensions, resulting in the second dimension-enhanced features. Based on the enhanced features of the second dimension, a fully connected layer is used for processing to output the tornado identification results.

6. The machine learning tornado recognition method driven by storm radar feature quantities according to claim 5, characterized in that, The Transformer encoder of the recognition model that integrates Transformer and CNN consists of 6 stacked sub-encoders, and the self-attention module of each sub-encoder adopts a 4-head self-attention mechanism.

7. A machine learning-based tornado recognition system driven by storm radar features, characterized in that, The system for implementing the method of claim 1, wherein the system comprises: The feature extraction and optimization module is used to obtain storm radar feature quantities containing key feature information based on volume scan radar base data through feature extraction and optimization processing. The data preprocessing module is used to perform feature normalization and missing feature value processing on the storm radar features to obtain an input sequence containing several elevation angle layer data. The tornado recognition module is used to input the input sequence into a recognition model that fuses Transformer and CNN to output tornado recognition results.

8. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in claim 1.

9. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 1.

10. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the steps of the method of claim 1.

Images