A lightning strike zone identification method based on radar data

Abstract

The invention discloses a method for identifying lightning strike areas based on radar data. The method comprises the following steps: constructing a two-dimensional radar grid field with a weather radar position as the center and a grid size of 1 square kilometer, converting a plurality of radar data obtained by converting radar base data into Cartesian coordinates and then rasterizing the data on the two-dimensional radar grid field, stacking all rasterized data within a volume scanning cycle into three-dimensional radar grid data; extracting lightning position information within the volume scanning cycle and mapping it to the two-dimensional radar grid field, generating binary classification label data according to whether lightning exists in the grid; stacking the three-dimensional radar grid data and the label data into radar lightning grid field data according to a spatial relationship; dividing a set of radar lightning grid field data into a training set, a validation set and a test set, and training a deep semantic segmentation network model based on SegNet by using the training set and the validation set; and evaluating the ground lightning recognition performance of the model by using the test set. The results show that the model has a high recognition accuracy and can be combined with radar extrapolation technology for lightning forecasting.

Description

Technical Field

[0001] The invention relates to a lightning strike zone positioning technology, and in particular to a lightning strike zone identification method based on radar data. Background Art

[0002] Lightning is a long-distance discharge phenomenon that occurs in the natural atmosphere under thunderstorm weather conditions. Lightning has the characteristics of large current, high voltage, and strong electromagnetic radiation. Lightning can generally be divided into two categories: ground lightning and cloud lightning. Among them, ground lightning refers to the discharge process between the charged center in the cloud and the earth and ground objects. Its harm is not only reflected in the loss caused by strong current during direct strike, but more importantly, the impact of the induction field and radiation field it produces. Therefore, most lightning disasters are caused by ground lightning.

[0003] Since thunderstorm weather has a small spatial and temporal scale, changes quickly, is severe, has a great social impact, and has a complex mechanism of occurrence and development, lightning monitoring and early warning are of great significance. At present, most of the research on lightning strike location technology is achieved through the lightning location system. The lightning location system is a real-time monitoring system that can locate the location of the lightning strike area in real time after the lightning occurs. However, the characteristics of its one-dimensional data determine that the system cannot capture the changes in the generation and disappearance of strong convective bodies, and thus cannot be combined with radar extrapolation technology such as optical flow method to predict the movement and development of lightning, and cannot predict the subsequent development of lightning. China's new generation weather radar CINRAD / SA is one of the most effective means of severe convective weather monitoring. It can detect a wider area and a higher spatial resolution of the convective system structure. Therefore, there are also many methods that use radar data to identify lightning. These methods mainly analyze the radar data within a certain spatial and temporal range when the lightning occurs, extract the threshold parameter, and consider the spatial area corresponding to the radar data as the lightning strike area when the radar data is greater than the threshold. However, these methods only consider the characteristics of the grid points where the radar echo is located, and do not consider the spatial characteristics. Therefore, these methods are sensitive to noise, have a high false alarm rate, and perform poorly in severe convective weather. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a lightning strike zone identification method based on radar data, which has high identification accuracy, can be combined with radar extrapolation technology to perform lightning forecasting, and can be applied to radar detection environments where there is no lightning location system or the detection range of the lightning location system is not covered.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a lightning strike zone identification method based on radar data, characterized by comprising the following steps:

[0006] Step S1: Select lightning location information in lightning location data as a lightning data source; select N different types of radar data obtained by converting radar base data as a radar data source, wherein the number of types of radar data N is determined by the volume scanning mode of the weather radar;

[0007] Step S2: Taking the effective detection distance of the weather radar of 230 km as the spatial range and the volume scanning period of the weather radar as the time range; converting each radar data in each volume scanning period from the polar coordinate format to the Cartesian coordinate format, wherein the coordinate origin of the Cartesian coordinate format is the longitude and latitude position of the weather radar; then constructing a 2W×2W two-dimensional radar grid field centered at the longitude and latitude position of the weather radar and with a grid width of 1 km, wherein W represents the value of the effective detection distance of the weather radar, W=230, the lower left corner coordinates of the two-dimensional radar grid field are (0, 0), and the upper right corner coordinates of the two-dimensional radar grid field are (2W, 2W); then, based on the two-dimensional radar grid field, performing rasterization processing on each Cartesian coordinate format radar data in each volume scanning period to obtain rasterized data corresponding to each Cartesian coordinate format radar data in each volume scanning period; finally, stacking the rasterized data corresponding to all Cartesian coordinate format radar data in each volume scanning period into 2W×2W×N three-dimensional radar grid data according to the spatial correspondence;

[0008] Step S3: Taking the effective detection distance of 230km of the weather radar as the spatial range and the volume scanning cycle of the weather radar as the time range; extracting the lightning position information in each volume scanning cycle within the effective detection distance of the weather radar; then mapping the lightning position information in each volume scanning cycle to the two-dimensional radar grid field; then judging whether there is lightning position information in each grid point in the two-dimensional radar grid field, if there is, marking the grid point as 1, if not, marking the grid point as 0, and generating 2W×2W×1 binary classification label data;

[0009] Step S4: The 2W×2W×N three-dimensional radar grid point data in each volume scanning cycle and the corresponding 2W×2W×1 binary classification label data are used to form a radar lightning grid point field data with a specification of 2W×2W×(N+1) according to the spatial correspondence relationship;

[0010] Step S5: randomly divide the set of radar lightning grid field data within all volume scanning cycles of thunderstorm days in the past years into a training set, a validation set, and a test set; and introduce a deep semantic segmentation network model based on SegNet;

[0011] Step S6: Use the training set to train the model; then use the validation set to validate the model, select the model and weight with the best performance based on the validation results and save them; then use the saved model to identify the test set, and the lightning strike zone recognition results are evaluated by the boundary success index algorithm.

[0012] In the step S1, the lightning location data is obtained through the lightning location system, and the lightning location is the latitude and longitude of the lightning strike area; the radar data is obtained through the weather radar, and the radar data is mainly composed of two types of data: radar multi-elevation plane position scanning echo data and radar product data. The radar multi-elevation plane position scanning echo data includes 9 elevation radar echo data of 0.5 degrees, 1.45 degrees, 2.4 degrees, 3.35 degrees, 4.3 degrees, 6.0 degrees, 9.0 degrees, 14.6 degrees, and 19.5 degrees. The radar product data includes echo top height, combined reflectivity, radial velocity, and vertical liquid water content, that is, the value of N is 13.

[0013] In step S2, the process of rasterization is as follows: taking the two-dimensional radar grid field as the standard, the radar data of each Cartesian coordinate format in each volume scanning cycle is interpolated according to the spatial correspondence to realize rasterization processing.

[0014] The interpolation algorithm used in the interpolation is the inverse distance weighted interpolation method.

[0015] In step S3, the process of mapping lightning position information into a 2W×2W two-dimensional radar grid field is as follows: setting the longitude and latitude position of the weather radar to (X R ,Y R ), set the lightning position information to (X L ,Y L ), set the row and column numbers of the lightning position information mapped to the 2W×2W two-dimensional radar grid field to (X, Y), X = W + (X R -X L )×S X , Y=W+(Y R -Y L )×S Y , where X R Indicates the longitude and latitude of the weather radar. Y R Indicates the latitude and longitude of the weather radar location, X L Indicates the longitude of the lightning location, Y L represents the latitude of the lightning position, X represents the column number of the grid position to which the lightning position information is mapped in the two-dimensional radar grid field, and Y represents the row number of the grid position to which the lightning position information is mapped in the two-dimensional radar grid field. The column and row numbers of the grid position in the grid field start from 0. XIndicates the distance represented by 1° longitude within the effective detection range of the weather radar, S Y Indicates the distance represented by 1° latitude within the effective detection range of the weather radar.

[0016] In step S5, the ratio of the training set, the validation set and the test set is 7:2:1.

[0017] In the step S6, when training the model, four hyperparameters, namely, learning rate, batch size, loss function and early stopping rounds, are set. The learning rate for the first 10 rounds is set to 0.001, and the learning rate for every 5 rounds after 10 rounds is reduced by 10% on the basis of the previous learning rate. The batch size is set to 8, the loss function adopts the gradient coordination mechanism function, and the early stopping rounds are set to 10;

[0018] When training the model, firstly, at the beginning of each round of iterative training, a batch-sized training sample, i.e., radar lightning grid field data, is selected from the training set. Then, all the training samples of this batch-sized size are input into the model to obtain the recognition result of the lightning strike zone in each training sample of this batch-sized size. Then, the loss function, i.e., the gradient coordination mechanism function, is used to calculate the loss value between the recognition result of the lightning strike zone in all the training samples of this batch-sized size and the corresponding true value, i.e., the binary classification label data. Finally, the back propagation algorithm in the stochastic optimizer of the adaptive momentum is used to update the model parameters in the opposite direction of the gradient with the loss value. The model is iteratively trained to make the loss value as low as possible. At the end of each round of iterative training, the model is verified using the validation set, and the average loss value between the recognition results of all the validation samples in the validation set and the corresponding true value, i.e., the binary classification label data, is recorded. If the average loss value of the model on the validation set does not decrease for 10 consecutive rounds during the iterative training process, the training is terminated and the model and weights at the lowest average loss value on the validation set are saved.

[0019] Compared with the prior art, the advantages of the present invention are:

[0020] 1) The method of the present invention constructs radar lightning field data based on the spatiotemporal correlation of radar data and lightning location data, and combines it with deep learning methods to mine the spatiotemporal laws and characteristic attributes hidden in radar data, greatly improving the accuracy of identifying lightning strike areas.

[0021] 2) The method of the present invention can be combined with radar extrapolation techniques such as the cross-correlation method, the monomer centroid method, the optical flow method, and the radar extrapolation algorithm based on deep learning technology to perform lightning forecasting, and can have good application value in the short-term lightning forecasting business, providing high-quality technical support for lightning disaster prevention.

[0022] 3) The method of the present invention can be applied to radar detection environments where there is no lightning location system or the detection range of the lightning location system is not covered. Only radar data is needed to identify the lightning strike area, and the method has good versatility and portability. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a general implementation block diagram of the method of the present invention;

[0024] Figure 2 It is a schematic diagram of a 2W×2W two-dimensional radar grid field with the latitude and longitude of the weather radar as the center and a grid width of 1km. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below with reference to the accompanying drawings.

[0026] The present invention proposes a lightning strike zone identification method based on radar data, and its overall implementation block diagram is as follows: Figure 1 As shown, it includes the following steps:

[0027] Step S1: Select lightning location information in lightning location data as a lightning data source; select N different types of radar data obtained by converting radar base data as a radar data source, wherein the number of types of radar data N is determined by the volume scanning mode of the weather radar (such as China's new generation weather radar CINRAD / SA).

[0028] In this embodiment, in step S1, lightning location data is obtained through an existing lightning location system, and the lightning location is the latitude and longitude of the lightning strike area; radar data is obtained through an existing weather radar, and the radar data is mainly composed of two types of data: radar multi-elevation plane position scanning echo (PPI) data and radar product data. The radar multi-elevation plane position scanning echo data includes radar echo data at nine elevation angles: 0.5 degrees, 1.45 degrees, 2.4 degrees, 3.35 degrees, 4.3 degrees, 6.0 degrees, 9.0 degrees, 14.6 degrees, and 19.5 degrees. The radar product data includes echo top height (ET), combined reflectivity (CR), radial velocity (V), and vertical liquid water content (VIL), that is, the value of N is 13. Convective fluid is the basic condition for lightning to occur. Radar multi-elevation angle PPI data represents the reflectivity factor at different elevation angles, which can show the spatial structure of the convection fluid; echo top height can indicate the strength of vertical airflow in thunderstorm clouds. The higher the cloud top, the more vigorous the convection development. Its height is an important condition for whether strong electrification process can occur in thunderstorm clouds; combined reflectivity reflects the maximum echo intensity, which reflects the vigorous degree of convection development. Studies have found that the spatial distribution range of lightning is consistent with the location of strong echoes; radial velocity has a corresponding relationship with the movement, strengthening and weakening of lightning activities; vertical liquid water content can indicate the water vapor conditions in convective clouds, and suitable water vapor conditions are conducive to the occurrence of lightning. Taking the weather radar in Ningbo as an example, the basic data of the weather radar contains radar echo data at 9 elevation angles, and then the radar PPI data is converted into a variety of radar product data, such as ET and CR.

[0029] Step S2: Taking the effective detection distance of the weather radar of 230 km as the spatial range and the volume scanning period of the weather radar as the time range; convert each radar data in each volume scanning period from the polar coordinate format to the Cartesian coordinate format, wherein the coordinate origin of the Cartesian coordinate format is the longitude and latitude position of the weather radar; then construct a 2W×2W two-dimensional radar grid field with the longitude and latitude position of the weather radar as the center and a grid width of 1 km, such as Figure 2 As shown, W represents the value of the effective detection distance of the weather radar, W=230. In order to facilitate conversion into a computer matrix, the lower left corner coordinate of the two-dimensional radar grid field is defined as (0, 0), and the upper right corner coordinate of the two-dimensional radar grid field is defined as (2W, 2W). Then, based on the two-dimensional radar grid field, the radar data in each Cartesian coordinate format within each volume scanning cycle is rasterized to obtain the rasterized data corresponding to each Cartesian coordinate format of the radar data within each volume scanning cycle. Finally, the rasterized data corresponding to all Cartesian coordinate formats of the radar data within each volume scanning cycle are stacked into 2W×2W×N three-dimensional radar grid data according to the spatial correspondence.

[0030] Here, a two-dimensional radar grid field is formed based on the longitude and latitude of the weather radar. Data with longitude and latitude information can be converted into a grid position through a longitude coefficient and a latitude coefficient. For example, for the Ningbo area, the longitude coefficient is approximately 111.11 km, which means that 1°=111.11 km in the horizontal direction; the latitude coefficient is approximately 96.225 km, which means that 1°=96.225 km in the vertical direction.

[0031] In this embodiment, in step S2, the process of rasterization is: taking each radar data in Cartesian coordinate format in each volume scanning cycle as the standard, interpolating according to spatial correspondence to realize rasterization processing. Here, the interpolation algorithm used for interpolation is the inverse distance weighted interpolation method (IDW, Inverse Distance Weight, also known as the inverse distance multiplication method).

[0032] Step S3: Taking the effective detection distance of 230km of the weather radar as the spatial range and the volume scanning cycle of the weather radar as the time range; extract the lightning location information in each volume scanning cycle within the effective detection distance of the weather radar; then map the lightning location information in each volume scanning cycle to the two-dimensional radar grid field; then determine whether there is lightning location information in each grid point in the two-dimensional radar grid field (i.e., whether lightning has occurred within the 1 square kilometer range represented by the grid point). If so, mark the grid point as 1, if not, mark the grid point as 0, and generate 2W×2W×1 binary classification label data. Here, with the start time and end time of the weather radar volume scanning as the limit, extract the lightning location record within the time period from the lightning location data, i.e., the lightning location information, to achieve time screening, and then take the longitude and latitude position of the weather radar as the center and the effective detection distance of the weather radar as the boundary to perform spatial screening on the lightning location record; the order of time screening and space screening can be reversed.

[0033] In this embodiment, in step S3, the process of mapping lightning position information into a 2W×2W two-dimensional radar grid field is as follows: Figure 2 As shown, the longitude and latitude of the weather radar is set to (X R ,Y R ), set the lightning position information to (X L ,Y L ), set the row and column numbers of the lightning position information mapped to the 2W×2W two-dimensional radar grid field to (X, Y), X = W + (X R -X L )×S X , Y=W+(Y R -Y L )×S Y , where XR Indicates the longitude and latitude of the weather radar. Y R Indicates the latitude and longitude of the weather radar location, X L Indicates the longitude of the lightning location, Y L represents the latitude of the lightning position, X represents the column number of the grid position to which the lightning position information is mapped in the two-dimensional radar grid field, and Y represents the row number of the grid position to which the lightning position information is mapped in the two-dimensional radar grid field. The column and row numbers of the grid position in the grid field start from 0. X Indicates the distance represented by 1° longitude within the effective detection range of the weather radar, S Y Indicates the distance represented by 1° latitude within the effective detection range of the weather radar.

[0034] Step S4: The 2W×2W×N three-dimensional radar grid data in each volume scanning cycle and the corresponding 2W×2W×1 binary classification label data are spatially corresponded to form a radar lightning grid field data with a specification of 2W×2W×(N+1).

[0035] Step S5: randomly divide the set of radar lightning grid field data within all volume scanning cycles of thunderstorm days in the past years (such as the past 10 years) into a training set, a validation set and a test set; and introduce a deep semantic segmentation network model based on SegNet.

[0036] In this embodiment, in step S5, the ratio of the training set, the validation set and the test set is 7:2:1.

[0037] Here, the deep semantic segmentation network model based on SegNet is an existing network model, which is cited from Medi. Neural Networks for Image Semantic Segmentation - From SegNet to U-Net [J]. Electronic Production, 2021(12): 49-52. In the present invention, it is only necessary to adapt the input of the deep semantic segmentation network model based on SegNet.

[0038] Step S6: Use the training set to train the model; then use the validation set to validate the model, select the model and weight with the best performance according to the validation results and save them; then use the saved model to identify the test set, and the lightning strike zone recognition results are evaluated by the boundary success index (CSI) algorithm.

[0039] In this embodiment, in step S6, when training the model, four hyperparameters, namely, learning rate, batch size, loss function and early stopping rounds, are set. The learning rate for the first 10 rounds (EPOCH) is set to 0.001. After 10 rounds, the learning rate for every 5 rounds is reduced by 10% based on the previous learning rate. The batch size is set to 8, and the loss function adopts the gradient harmonizing mechanism (GHM, Gradient Harmonizing Mechanism) function, the early stopping round is set to 10; when training the model, firstly, at the beginning of each round of iterative training, a batch-scale training sample in the training set, i.e., radar lightning grid field data, is selected, and then all the training samples in this batch-scale are input into the model (the working principle of the model is the forward propagation algorithm), and the recognition result of the lightning strike area in each training sample of this batch-scale is obtained, and then the loss function, i.e., the gradient coordination mechanism function, is used to calculate the loss value between the recognition result of the lightning strike area in all training samples of this batch-scale and the corresponding true value, i.e., the binary classification label data, and finally, the back propagation (BP) algorithm in the stochastic optimizer (Adam) of the adaptive momentum is used to update the model parameters in the opposite direction of the gradient with the loss value; the model is iteratively trained to make the loss value as low as possible, and the model is verified by the validation set at the end of each round of iterative training, and the average loss value between the recognition results of all the validation samples in the validation set and the corresponding true value, i.e., the binary classification label data, is recorded. If the average loss value of the model on the validation set does not decrease for 10 consecutive rounds during the iterative training process, the training is terminated and the model and weights when the average loss value on the validation set is the lowest are saved.

[0040] When the method of the present invention is used for real-time recognition, the three-dimensional radar grid point data of 2W×2W×N is obtained in the same manner according to the process of step S1 and step S2, and then the three-dimensional radar grid point data is recognized using the saved model to obtain a lightning strike zone with a specification of 2W×2W×1.

[0041] In order to further illustrate the feasibility and effectiveness of the method of the present invention, experiments were conducted on the method of the present invention.

[0042] After the SegNet-based deep semantic segmentation network model is trained, the test set and the critical success index (CSI) are used to quantitatively verify the performance of the saved model. CSI represents the rate at which the algorithm successfully detects lightning, reflecting the probability that the algorithm estimates lightning or actually detects lightning and correctly identifies it. TP (True Positive): the prediction is 1, the true value is also 1, true positive; FP (False Positive): the prediction is 1, the true value is 0, false positive; FN (False Negative): the prediction is 0, the true value is 1, false negative; TN (True Negative): the prediction is 0, the true value is also 0, true negative.

[0043] The presence or absence of lightning can be divided into positive and negative categories, and the identification results of the lightning strike area can be divided into correct and incorrect categories, forming a total of four results, as shown in Table 1.

[0044] Table 1 Definition of lightning strike zone identification results

[0045]

[0046] At present, the industry generally recognizes that the probability of lightning occurring when the echo intensity reaches 40dBz at a stratification height of -10℃ is very high, so this threshold (1. Li Fang, Huang Xingyou, Wang Zhenhui, Feng Minxue, Jiao Xue. Preliminary study on lightning identification indicators based on radar data in Nanjing [J]. Meteorological Science, 2010, 30(02): 202-207.; 2. Ma Yurong. Research on weather radar identification and nowcasting of thunderstorms [D]. Nanjing University of Information Science and Technology) is used as a comparison algorithm.

[0047] The test set was identified by the method of the present invention and the comparison algorithm respectively, and the lightning strike zone identification result of each radar lightning grid field data in the test set was obtained. The CSI score was calculated for the lightning strike zone identification result. The average CSI score of 2395 test samples showed that the average CSI score obtained by the method of the present invention was 0.02924, and the average CSI score obtained by the comparison algorithm was 0.01306. By comparison, it can be seen that compared with the comparison algorithm, the average CSI score of the method of the present invention was improved by 0.01618, an increase of 123.9%, indicating that the saved model can better identify the lightning strike zone.

Claims

1. A lightning strike zone identification method based on radar data, characterized in that The following steps are involved: Step S1: Selecting lightning location information in lightning location data as a lightning data source; N different types of radar data obtained by converting radar base data are selected as radar data sources, where the number of types of radar data N is determined by the volume scanning mode of the weather radar; Step S2: Taking the effective detection distance of the weather radar of 230 km as the spatial range and the volume scanning period of the weather radar as the time range; converting each radar data in each volume scanning period from the polar coordinate format to the Cartesian coordinate format, wherein the coordinate origin of the Cartesian coordinate format is the longitude and latitude position of the weather radar; then constructing a 2W×2W two-dimensional radar grid field centered at the longitude and latitude position of the weather radar and with a grid width of 1 km, wherein W represents the value of the effective detection distance of the weather radar, W=230, the lower left corner coordinates of the two-dimensional radar grid field are (0, 0), and the upper right corner coordinates of the two-dimensional radar grid field are (2W, 2W); then, based on the two-dimensional radar grid field, performing rasterization processing on each Cartesian coordinate format radar data in each volume scanning period to obtain rasterized data corresponding to each Cartesian coordinate format radar data in each volume scanning period; finally, stacking the rasterized data corresponding to all Cartesian coordinate format radar data in each volume scanning period into 2W×2W×N three-dimensional radar grid data according to the spatial correspondence; Step S3: Taking the effective detection distance of 230km of the weather radar as the spatial range and the volume scanning cycle of the weather radar as the time range; extracting the lightning position information in each volume scanning cycle within the effective detection distance of the weather radar; then mapping the lightning position information in each volume scanning cycle to the two-dimensional radar grid field; then judging whether there is lightning position information in each grid point in the two-dimensional radar grid field, if there is, marking the grid point as 1, if not, marking the grid point as 0, and generating 2W×2W×1 binary classification label data; Step S4: The 2W×2W×N three-dimensional radar grid point data in each volume scanning cycle and the corresponding 2W×2W×1 binary classification label data are used to form a radar lightning grid point field data with a specification of 2W×2W×(N+1) according to the spatial correspondence relationship; Step S5: randomly divide the set of radar lightning grid field data within all volume scanning cycles of thunderstorm days in the past years into a training set, a validation set, and a test set; and introduce a deep semantic segmentation network model based on SegNet; Step S6: Use the training set to train the model; then use the validation set to validate the model, select the model and weight with the best performance based on the validation results and save them; then use the saved model to identify the test set, and the lightning strike zone recognition results are evaluated by the boundary success index algorithm.

2. The method for identifying lightning strike zones based on radar data according to claim 1 is characterized in that In the step S1, the lightning location data is obtained through the lightning location system, and the lightning location is the latitude and longitude of the lightning strike area; the radar data is obtained through the weather radar, and the radar data is mainly composed of two types of data: radar multi-elevation plane position scanning echo data and radar product data. The radar multi-elevation plane position scanning echo data includes 9 elevation radar echo data of 0.5 degrees, 1.45 degrees, 2.4 degrees, 3.35 degrees, 4.3 degrees, 6.0 degrees, 9.0 degrees, 14.6 degrees, and 19.5 degrees. The radar product data includes echo top height, combined reflectivity, radial velocity, and vertical liquid water content, that is, the value of N is 13.

3. A lightning strike zone identification method based on radar data according to claim 1 or 2, characterized in that In step S2, the process of rasterization is as follows: taking the two-dimensional radar grid field as the standard, the radar data of each Cartesian coordinate format in each volume scanning cycle is interpolated according to the spatial correspondence to realize rasterization processing.

4. The method for identifying lightning strike zones based on radar data according to claim 3 is characterized in that The interpolation algorithm used in the interpolation is the inverse distance weighted interpolation method.

5. The method for identifying lightning strike zones based on radar data according to claim 1 is characterized in that In step S3, the process of mapping the lightning position information into the 2W×2W two-dimensional radar grid field is as follows: the longitude and latitude position of the weather radar is set as (X R ,Y R ), set the lightning position information to (X L ,Y L ), set the row and column numbers of the lightning position information mapped to the 2W×2W two-dimensional radar grid field to (X, Y), X = W + (X R -X L )×S X , Y=W+(Y R -Y L )×S Y , where X R Indicates the longitude and latitude of the weather radar. Y R Indicates the latitude and longitude of the weather radar location, X L Indicates the longitude of the lightning location, Y L represents the latitude of the lightning position, X represents the column number of the grid position to which the lightning position information is mapped in the two-dimensional radar grid field, and Y represents the row number of the grid position to which the lightning position information is mapped in the two-dimensional radar grid field. The column and row numbers of the grid position in the grid field start from 0. X Indicates the distance represented by 1° longitude within the effective detection range of the weather radar, S Y Indicates the distance represented by 1° latitude within the effective detection range of the weather radar.

6. The method for identifying lightning strike zones based on radar data according to claim 1, characterized in that In step S5, the ratio of the training set, the validation set and the test set is 7:2:

1.

7. The method for identifying lightning strike zones based on radar data according to claim 1, characterized in that In the step S6, when training the model, four hyperparameters, namely, learning rate, batch size, loss function and early stopping rounds, are set. The learning rate for the first 10 rounds is set to 0.001, and the learning rate for every 5 rounds after 10 rounds is reduced by 10% on the basis of the previous learning rate. The batch size is set to 8, the loss function adopts the gradient coordination mechanism function, and the early stopping rounds are set to 10; When training the model, firstly, at the beginning of each round of iterative training, a batch-sized training sample, i.e., radar lightning grid field data, is selected from the training set. Then, all the training samples of this batch-sized size are input into the model to obtain the recognition result of the lightning strike zone in each training sample of this batch-sized size. Then, the loss function, i.e., the gradient coordination mechanism function, is used to calculate the loss value between the recognition result of the lightning strike zone in all the training samples of this batch-sized size and the corresponding true value, i.e., the binary classification label data. Finally, the back propagation algorithm in the stochastic optimizer of the adaptive momentum is used to update the model parameters in the opposite direction of the gradient with the loss value. The model is iteratively trained to make the loss value as low as possible. At the end of each round of iterative training, the model is verified using the validation set, and the average loss value between the recognition results of all the validation samples in the validation set and the corresponding true value, i.e., the binary classification label data, is recorded. If the average loss value of the model on the validation set does not decrease for 10 consecutive rounds during the iterative training process, the training is terminated and the model and weights at the lowest average loss value on the validation set are saved.

Images