Lightning-three-dimensional proxy reflectivity inversion method based on ResNet

Abstract

The invention discloses a ResNet-based lightning-three-dimensional proxy reflectivity inversion method, which belongs to the technical field of application meteorology, and comprises the following steps: obtaining lightning positioning observation data, screening, quality control and classification of lightning positioning observation records, determining grids, counting data at grid points of the grids, and processing the counting data. Calculating a space coefficient, a time coefficient and a space-time coefficient, sorting the lightning positioning observation records of each grid point, processing the sorted lightning positioning observation records and other data to obtain the maximum proxy reflectivity, obtaining the three-dimensional proxy reflectivity through the maximum proxy reflectivity and the characteristics of the atmospheric environment where the three-dimensional proxy reflectivity is located, and calculating the spatial coefficient, the time coefficient and the space-time coefficient of the three-dimensional proxy reflectivity. Therefore, three-dimensional proxy reflectivity inversion is realized. According to the method, inversion of the three-dimensional proxy reflectivity is more accurate, and the problem that the severe convection weather monitoring and early warning capacity is insufficient in a radar detection blind area is effectively solved.

Description

Technical Field

[0001] The invention relates to the field of applied meteorological technology, and in particular to a lightning-three-dimensional proxy reflectivity inversion method based on ResNet. Background Art

[0002] Lightning location observation has the advantages of high resolution, wide detection range, little influence by terrain, and continuous monitoring, and it has a good indication effect on severe convective weather. Although my country has deployed many weather radars, there are still certain radar detection blind spots, especially in mountainous areas that are easily blocked by terrain. The radar detection capability is weak, and the severe convection monitoring and early warning capabilities are also very limited. Lightning location observation plays an important complementary role to radar detection. The lightning location observation can be inverted into a three-dimensional proxy reflectivity through the lightning-three-dimensional proxy reflectivity inversion method, and the three-dimensional proxy reflectivity can be used to realize the monitoring and early warning of the occurrence and development of severe convective weather.

[0003] Previous lightning-3D proxy reflectivity inversion methods only considered the frequency of lightning, without considering other information about lightning activity. With the increasing requirements for the accuracy of severe convection monitoring and early warning, inversion using only the frequency of lightning cannot meet the existing monitoring and early warning needs. At the same time, the traditional maximum proxy reflectivity-3D proxy reflectivity inversion method uses the vertical profile coefficient, a fixed data calculated based on historical observation data. Therefore, the process of inverting the maximum proxy reflectivity to the 3D proxy reflectivity is very fixed and stereotyped. The current atmospheric environment affects the distribution and structural characteristics of the 3D radar reflectivity. The inversion accuracy using only the vertical profile coefficient is obviously insufficient, which in turn affects the forecast accuracy of subsequent numerical models after the assimilation of lightning data. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a lightning-3D proxy reflectivity inversion method based on ResNet. A variety of lightning activity information related to radar reflectivity is introduced, and the atmospheric environment characteristics before lightning occur are added. The inversion is performed with the help of ResNet residual neural network to solve the problem of insufficient monitoring and early warning capabilities of severe convective weather in radar detection blind areas, making monitoring more accurate.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a lightning-3D proxy reflectivity inversion method based on ResNet, comprising the following steps:

[0006] Step 1: Obtain lightning location observation data, including: time, longitude and latitude, altitude, current intensity, lightning type, instrument model, fusion mark, and quality control mark of lightning location observation records;

[0007] Step 2: Filter the lightning location observation records with instrument model DDW1, use the fusion mark and quality control mark to perform quality control on the DDW1 lightning location observation records, and use the lightning type to classify the DDW1 lightning location observation records into cloud-flash observation records and ground-flash observation records;

[0008] Step 3: Select the inversion area according to the lightning activity range, divide the inversion area into grids, calculate the number of lightning location observation records, cloud flash observation records, ground flash observation records, and positive ground flash observation records of each grid point in the grid, and record them as total flash frequency, cloud flash frequency, ground flash frequency, and positive ground flash frequency respectively;

[0009] Step 4: Calculate the cloud-to-ground ratio and positive-to-ground ratio of each grid point with a total flash frequency greater than 0 in the inversion area, where the cloud-to-ground ratio is the ratio of the cloud-to-cloud flash frequency to the total flash frequency, and the positive-to-ground ratio is the ratio of the positive-to-ground flash frequency to the ground-to-ground flash frequency;

[0010] Step 5: Calculate the spatial coefficient r of each grid point in the inversion area where the total flash frequency is greater than 0 c , time coefficient t c and the spatiotemporal coefficient tr c ;

[0011] Step 6: Sort the cloud flash location observation records of each grid point with a total flash frequency greater than 0 in order of the time-space coefficient from small to large. For each grid point, retain the first N1 cloud flash observation records with the smallest time-space coefficient, save the space coefficient, time coefficient, and lightning height. If the cloud flash frequency is less than N1, fill it with the default value;

[0012] Step 7: Sort the ground-to-ground lightning observation records of each grid point with a total lightning frequency greater than 0 in order of the time-space coefficient from small to large. For each grid point, retain the first N2 ground-to-ground lightning observation records with the smallest time-space coefficient, save the space coefficient, time coefficient, and current intensity. If the ground-to-ground lightning frequency is less than N2, fill it with the default value;

[0013] Step 8: Input the total flash frequency, cloud-to-flash ratio, positive ground-to-ground ratio, spatial coefficient, time coefficient, lightning height of the first N1 cloud-to-ground flash observation records with the smallest time-space coefficient, and spatial coefficient, time coefficient, and current intensity of the first N2 ground-to-ground flash observation records with the smallest time-space coefficient into the lightning-maximum proxy reflectivity inversion model based on ResNet to obtain the maximum proxy reflectivity of each grid point among all grid points;

[0014] Step 9: Download FY-4B GIIRS atmospheric vertical detection products and atmospheric instability index products, and use the quality control mark to perform quality control on the products;

[0015] Step 10: Extract the atmospheric environmental characteristics of each grid point in the inversion area where the total flash frequency is greater than 0, that is, extract the temperature and humidity profile, K index, and CAPE of the point closest to the grid point in the past 2 hours. If the temperature and humidity profile, K index, and CAPE of the grid point do not exist or fail to pass quality control, then remove the point. Interpolate the temperature and humidity to the height layer of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0 according to the height information of the FY-4BGIIRS atmospheric vertical detection product, and the height layer is 24 layers;

[0016] Step 11: Input the maximum proxy reflectivity, temperature and humidity profile, K index, and CAPE of each grid point in the lightning activity range into the ResNet-based maximum proxy reflectivity-three-dimensional proxy reflectivity inversion model to obtain the three-dimensional proxy reflectivity of each grid point.

[0017] A further improvement of the technical solution of the present invention is that the quality control method for the DDW1 type lightning location observation record in step 2 is: the fusion mark is 0, the quality control mark is 0, and if the above two conditions are met at the same time, the lightning location observation record is considered to be reliable, and the classification method is: when the lightning type is 0, the lightning location observation record is a cloud flash observation record; when the lightning type is 1, the lightning location observation record is a ground flash observation record.

[0018] A further improvement of the technical solution of the present invention lies in that: Step 3 is specifically: referring to the division method of the longitude and latitude information of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0, the inversion area is divided into corresponding longitude and latitude grids of 0.01°*0.01°, and each grid point is taken as the center of the circle. The number of lightning location observation records, cloud flash observation records, ground flash observation records, and positive ground flash observation records within a search radius of 0.08° from 30 minutes before to 10 minutes after the hour are calculated, and are recorded as the total flash frequency, cloud flash frequency, ground flash frequency, and positive ground flash frequency, respectively.

[0019] A further improvement of the technical solution of the present invention is that the spatial coefficient r in step 5 c , time coefficient t c and the spatiotemporal coefficient tr c The calculation is as follows:

[0020] t c =(t lgt -t radar ) / 1800

[0021]

[0022] In the formula, t lgt ,t radar The lightning occurrence time and the maximum proxy reflectivity time, i.e. the hourly moment, lon lgt ,lonradar are the longitude of the lightning occurrence location and the longitude of the maximum proxy reflectivity location, i.e., the grid point of equal longitude and latitude, lat lgt ,lat radar They are the latitude of the lightning occurrence location and the latitude of the maximum proxy reflectivity location, i.e., the grid point of equal longitude and latitude.

[0023] A further improvement of the technical solution of the present invention is that in step 6 and step 7, N1 is the cloud flash cutoff frequency determined by selecting a percentage based on the cumulative frequency curve of cloud flash frequency, and N2 is the ground flash cutoff frequency determined by selecting a percentage based on the cumulative frequency curve of ground flash frequency, and the selected percentage is 90%, 95% or 99%.

[0024] A further improvement of the technical solution of the present invention is that the specific steps of constructing the lightning-maximum proxy reflectivity inversion model based on ResNet in step 8 are as follows:

[0025] Step 8.1: Collect the lightning location data of the National Lightning Detection System and the combined reflectivity factor products of the Weather Radar Mosaic System V3.0 for a certain period of time, with a horizontal resolution of 0.01°*0.01°; perform quality control and classification according to step 2, taking into account the blind spots of radar detection and the missing or late arrival of radar data, and obtain the complete DDW1 lightning location observation records and maximum reflectivity data sets of the maximum reflectivity mosaic when lightning activity occurs within the radar coverage area through screening;

[0026] Step 8.2: Take each grid point in the collected weather radar mosaic system V3.0 combined reflectivity factor product as the center of the circle and follow steps 3-7 to obtain the total flash frequency, cloud-to-flash ratio, positive ground-to-ground flash ratio, spatial coefficient, time coefficient, lightning height of the first N1 cloud-to-ground flash observation records with the smallest time-space coefficient, spatial coefficient, time coefficient, current intensity, and maximum reflectivity of the first N2 ground-to-ground flash observation records with the smallest time-space coefficient for each grid point in the radar coverage area when lightning activity occurs, and form data set A;

[0027] Step 8.3: Perform feature normalization on dataset A and split dataset A into a training set (70%) and a validation set (30%).

[0028] Step 8.4: Design the input layer Input, receive the total flash frequency, cloud-to-flash ratio, positive ground-to-ground ratio, the spatial coefficient, time coefficient, lightning height of the first N1 cloud-to-ground flash observation records with the smallest time-space coefficient, the spatial coefficient, time coefficient, and current intensity of the first N2 ground-to-ground flash observation records with the smallest time-space coefficient, a total of 3+3N 1+ 3N2 features;

[0029] Step 8.5: Perform feature embedding, using the fully connected layer Linear and dimension transformation Reshape to embed 3+3N in each sample 1+ The 3N2 features are reshaped into a matrix acceptable to ResNet. For more efficient gradient descent and back propagation, the ReLU activation function is added during the reshaping process.

[0030] Step 8.6: Remove the original convolutional layer, retain the residual block and the fully connected layer to form the modified ResNet;

[0031] Step 8.7: Design the output layer, add the average pooling layer AveragePooling and the fully connected layer Linear, and output the maximum reflectivity; the maximum reflectivity is specifically the combined reflectivity of the weather radar puzzle system V3.0 combined reflectivity factor product;

[0032] Step 8.8: Set the training parameters for the model built through steps 8.4-8.7, use the training set to train the model, use the validation set to adjust the hyperparameters, save the optimal model with the minimum mean square error (MSE) or mean absolute error (MAE), use early stopping and model checkpoints to prevent overfitting during training, and obtain the lightning-maximum proxy reflectivity inversion model based on ResNet.

[0033] A further improvement of the technical solution of the present invention is that the quality control method in step 9 is: when the quality control mark is 0 or 1, the product is considered to be reliable.

[0034] A further improvement of the technical solution of the present invention is that the specific steps of constructing the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet in step 11 are as follows:

[0035] Step 11.1: Collect the three-dimensional reflectivity factor products of the weather radar mosaic system V3.0 in the same time period as in step 8.1. The horizontal resolution of the product is 0.01°*0.01°;

[0036] Step 11.2: According to step 8, obtain the maximum reflectivity of each grid point in all grid points when lightning activity occurs within the radar coverage area, and extract the radar reflectivity of each grid point according to the three-dimensional reflectivity factor product information of the weather radar puzzle system V3.0, and obtain the maximum reflectivity and radar reflectivity data set of each grid point in all grid points; collect the FY-4B GIIRS atmospheric vertical detection products and atmospheric instability index products in the same time period of step 8.1, and perform quality control; according to step 10, obtain the temperature and humidity profile, K index, and CAPE of each grid point in the combined reflectivity factor product grid collected when lightning activity occurs within the radar coverage area, interpolate the temperature and humidity profile, and interpolate the temperature and humidity profile into the height layer of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0, and combine the maximum reflectivity and radar reflectivity data sets of all grid points to form data set B;

[0037] Step 11.3: Normalize the features of dataset B and split it into a training set (70%) and a validation set (30%).

[0038] Step 11.4: Design the input layer Input, which receives the maximum reflectivity, interpolated temperature and humidity profile, K index, CAPE, a total of 51 features;

[0039] Step 11.5: Perform feature embedding, use the fully connected layer Linear and dimension transformation Reshape to reshape the 51 features in each sample into a matrix acceptable to ResNet. For more effective gradient descent and back propagation, add the ReLU activation function during the reshaping process;

[0040] Step 11.6: Remove the original convolutional layer, retain the residual block and the fully connected layer to form the modified ResNet;

[0041] Step 11.7: Design the output layer, add the fully connected layer Linear, and output the radar reflectivity. The radar reflectivity is the three-dimensional reflectivity of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0;

[0042] Step 11.8: Set the training parameters for the model built through steps 11.4-11.7, train the model using the training set, adjust the hyperparameters using the validation set, save the optimal model with the minimum arithmetic mean of the mean square error (MSE) or mean absolute error (MAE) of all layers as the standard; use early stopping and model checkpoints to prevent overfitting during training; obtain the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet.

[0043] A further improvement of the technical solution of the present invention is that the three-dimensional reflectivity factor product height layer of the weather radar puzzle system V3.0 in step 10 is 24 layers, the temperature and humidity profile is interpolated to 24 layers in step 11.2, and the radar reflectivity in step 11.7 is 24 layers.

[0044] Due to the adoption of the above technical scheme, the technical progress achieved by the present invention is: by considering the lightning activity information related to radar reflectivity such as cloud-flash ratio, positive ground-flash ratio, time, position, intensity and polarity, and adding the atmospheric environment characteristics before the occurrence of lightning (temperature and humidity profile, instability index (K index, CAPE)), a lightning-three-dimensional proxy reflectivity inversion method is constructed, and the inverted three-dimensional proxy reflectivity is closer to the actual situation, which greatly improves the use of existing lightning positioning observations and Fengyun satellite atmospheric vertical detection products. According to the inversion results, the monitoring and early warning of severe convective weather in the radar detection blind area is more accurate, and the monitoring and early warning capabilities are improved. It can also be applied to numerical prediction models at different horizontal resolutions, especially high-resolution numerical prediction models. By carrying out the joint assimilation of three-dimensional proxy reflectivity (lightning positioning observation) and three-dimensional radar reflectivity (radar observation), the assimilation application of DDW1 type full flash positioning observation in the numerical prediction model is realized, which has important practical value for improving the numerical prediction capability of radar detection blind areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 It is a schematic flow chart of the inversion method of the present invention;

[0047] Figure 2 It is a schematic diagram of calculating the total flash frequency of the present invention;

[0048] Figure 3 It is a flowchart of building a lightning-maximum proxy reflectivity inversion model based on ResNet;

[0049] Figure 4 It is a flowchart for constructing the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below in conjunction with embodiments:

[0051] like Figure 1As shown, it is a flowchart of a lightning-3D proxy reflectivity inversion method based on ResNet. The specific steps are as follows:

[0052] Step 1: Obtain lightning location observation data and download the lightning location data of the National Lightning Detection System, including: time, longitude and latitude, altitude, current intensity, lightning type, instrument model, fusion mark, and quality control mark of lightning location observation records;

[0053] Step 2: Filter the lightning location observation records with instrument model DDW1, use the fusion mark and quality control mark to perform quality control on the DDW1 lightning location observation records, and use the lightning type to classify the DDW1 lightning location observation records into cloud-flash observation records and ground-flash observation records; the quality control method for the DDW1 lightning location observation records is: the fusion mark is 0, the quality control mark is 0, and if the above two conditions are met at the same time, the lightning location observation record is considered reliable. Use the lightning type to classify the DDW1 lightning location observation records (cloud-flash observation records and ground-flash observation records). The classification method for the DDW1 lightning location observation records is: when the lightning type is 0, the lightning location observation record is a cloud-flash observation record; when the lightning type is 1, the lightning location observation record is a ground-flash observation record.

[0054] Step 3: If Figure 2 As shown, the inversion area is selected according to the range of lightning activity, and the inversion area is divided into grids. Specifically, the inversion area is divided into corresponding 0.01°*0.01° longitude and latitude grids with reference to the division method of the longitude and latitude information of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0, so that when the corresponding product is used to train the model later, it can correspond to the longitude and latitude information of the weather radar puzzle system V3.0 combined reflectivity factor product and the weather radar puzzle system V3.0 three-dimensional reflectivity factor product, and the grid points are consistent. With each grid point as the center of the circle, the number of lightning location observation records, cloud flash observation records, ground flash observation records, and positive ground flash observation records within a search radius of 0.08° from 30 minutes before to 10 minutes after the hour are calculated, and recorded as the total flash frequency, cloud flash frequency, ground flash frequency, and positive ground flash frequency respectively. Among them, the positive ground flash frequency is the ground flash frequency with positive current intensity.

[0055] Step 4: Calculate the cloud-to-ground ratio and positive-to-ground ratio of each grid point with a total flash frequency greater than 0 in the inversion area, where the cloud-to-ground ratio is the ratio of the cloud-to-cloud flash frequency to the total flash frequency, and the positive-to-ground ratio is the ratio of the positive-to-ground flash frequency to the ground-to-ground flash frequency;

[0056] Step 5: Calculate the spatial coefficient r of each grid point in the inversion area where the total flash frequency is greater than 0 c , time coefficient t c and the spatiotemporal coefficient tr c ; Space coefficient r c, time coefficient t c and the spatiotemporal coefficient tr c The calculation is as follows:

[0057]

[0058] t c =(t lgt -t radar ) / 1800

[0059]

[0060] In the formula, t lgt ,t radar The lightning occurrence time and the maximum proxy reflectivity time, i.e. the hourly moment, lon lgt ,lon radar are the longitude of the lightning occurrence location and the longitude of the maximum proxy reflectivity location, i.e., the grid point of equal longitude and latitude, lat lgt ,lat radar They are the latitude of the lightning occurrence location and the latitude of the maximum proxy reflectivity location, i.e., the grid point of equal longitude and latitude.

[0061] Step 6: Sort the cloud flash observation records of each grid point with a total flash frequency greater than 0 in order of time-space coefficient from small to large. Each grid point retains the first N1 cloud flash observation records with the smallest time-space coefficient, and saves the spatial coefficient, time coefficient, and lightning height. If the cloud flash frequency is less than N1, fill it with the default value; through the analysis of a large amount of historical observation data, determine the cloud flash cutoff frequency N according to the cumulative frequency curve of the cloud flash frequency at 90% (or 95%, 99%, selected according to the actual situation) 1。

[0062] Step 7: Sort the ground-to-ground lightning observation records of each grid point with a total flash frequency greater than 0 in order of time-space coefficient from small to large. For each grid point, retain the first N2 ground-to-ground lightning observation records with the smallest time-space coefficient, save the space coefficient, time coefficient, and current intensity (positive or negative reflects the polarity of lightning). If the ground-to-ground lightning frequency is less than N2, fill it with the default value; through the analysis of a large amount of historical observation data, determine the ground-to-ground lightning cutoff frequency N according to the ground-to-ground lightning frequency cumulative frequency curve of 90% (or 95%, 99%, which can be selected according to the actual situation) 2。

[0063] Step 8: Input the total flash frequency, cloud-flash ratio, positive ground-flash ratio, spatial coefficient, time coefficient, lightning height of the first N1 cloud-flash observation records with the smallest time-space coefficient, and spatial coefficient, time coefficient, and current intensity of the first N2 ground-flash observation records with the smallest time-space coefficient of each grid point in the lightning activity range into the lightning-maximum proxy reflectivity inversion model based on ResNet to obtain the maximum proxy reflectivity of each grid point in all grid points; Figure 3 As shown, the specific steps of constructing the lightning-maximum proxy reflectivity inversion model based on ResNet are as follows:

[0064] Step 8.1: Collect lightning location data of the National Lightning Detection System and combined reflectivity factor products of the Weather Radar Mosaic System V3.0 for a certain period of time. The horizontal resolution of the product is 0.01°*0.01°. The combined reflectivity factor product contains the combined reflectivity and its location (latitude and longitude) information. The grid can be determined according to the longitude and latitude information, and the combined reflectivity is output as the maximum reflectivity in the subsequent output; perform quality control and classification according to step 2, considering that there are blind spots in radar detection and radar data is missing or late, and obtain the complete DDW1 lightning location observation record and maximum reflectivity data set of the maximum reflectivity puzzle when lightning activity occurs within the radar coverage area through screening;

[0065] Step 8.2: Take each grid point in the grid of the weather radar mosaic system V3.0 combined reflectivity factor product as the center and follow steps 3-7 to obtain the total flash frequency, cloud-to-flash ratio, positive ground-to-ground ratio, spatial coefficient, time coefficient, lightning height of the first N1 cloud-to-ground flash observation records with the smallest time-space coefficient, spatial coefficient, time coefficient, current intensity, and maximum reflectivity of the first N2 ground-to-ground flash observation records with the smallest time-space coefficient for each grid point in the radar coverage area when lightning activity occurs, and form data set A;

[0066] Step 8.3: Perform feature normalization on dataset A and split dataset A into a training set (70%) and a validation set (30%).

[0067] Step 8.4: Design the input layer Input, receive the total flash frequency, cloud-to-flash ratio, positive ground-to-ground ratio, the spatial coefficient, time coefficient, lightning height of the first N1 cloud-to-ground flash observation records with the smallest time-space coefficient, the spatial coefficient, time coefficient, and current intensity of the first N2 ground-to-ground flash observation records with the smallest time-space coefficient, a total of 3+3N 1+ 3N2 features;

[0068] Step 8.5: Perform feature embedding, using the fully connected layer Linear and dimension transformation Reshape to embed 3+3N in each sample 1+ The 3N2 features are reshaped into a matrix acceptable to ResNet. For more efficient gradient descent and back propagation, the ReLU activation function is added during the reshaping process.

[0069] Step 8.6: Remove the original convolutional layer, retain the residual block and the fully connected layer to form the modified ResNet;

[0070] Step 8.7: Design the output layer, add the average pooling layer AveragePooling and the fully connected layer Linear, and output the maximum reflectivity; the maximum reflectivity is specifically the combined reflectivity of the weather radar puzzle system V3.0 combined reflectivity factor product;

[0071] Step 8.8: For the model built in steps 8.4 to 8.7, set training parameters, such as the number of training epochs and batch size, use the training set to train the model, use the validation set to adjust the hyperparameters, save the optimal model with the minimum mean square error (MSE) or mean absolute error (MAE) as the standard; use early stopping and model checkpointing to prevent overfitting during training; obtain the lightning-maximum proxy reflectivity inversion model based on ResNet.

[0072] Step 9: Download the FY-4B GIIRS atmospheric vertical detection products and atmospheric instability index products, and use the quality control mark to perform quality control on the products; the quality control method is: when the quality control mark is 0 or 1, the product is considered reliable.

[0073] Step 10: Extract the atmospheric environmental characteristics of each grid point in the inversion area where the total flash frequency is greater than 0, that is, extract the temperature and humidity profile, K index, and CAPE of the point closest to the grid point in the past 2 hours. If the temperature and humidity profile, K index, and CAPE of the grid point do not exist or fail to pass quality control, then remove the point. Interpolate the temperature and humidity to the height layer of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0 according to the height information of the FY-4BGIIRS atmospheric vertical detection product, and the height layer is 24 layers;

[0074] Step 11: Input the maximum proxy reflectivity, temperature and humidity profile, K index, and CAPE of all grid points within the lightning activity range into the ResNet-based maximum proxy reflectivity-three-dimensional proxy reflectivity inversion model to obtain the three-dimensional proxy reflectivity.

[0075] like Figure 4 As shown, the specific steps of constructing the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet are as follows:

[0076] Step 11.1: Collect the three-dimensional reflectivity factor products of the weather radar mosaic system V3.0 in the same time period as in step 8.1. The horizontal resolution of the product is 0.01°*0.01°;

[0077] Step 11.2: According to step 8, obtain the maximum reflectivity of each grid point in the radar coverage when lightning activity occurs, and extract the radar reflectivity of each grid point according to the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0 to obtain the maximum reflectivity and radar reflectivity data set of all grid points; collect the FY-4B GIIRS atmospheric vertical detection products and atmospheric instability index products in the same time period as step 8.1, and perform quality control; according to step 10, obtain the temperature and humidity profile (24 layers after interpolation), K index, and CAPE of each grid point in the grid of all the combined reflectivity factor products collected when lightning activity occurs within the radar coverage, that is, the maximum reflectivity grid point in the past 2 hours, and combine the maximum reflectivity and radar reflectivity data sets of all grid points to form data set B;

[0078] Step 11.3: Normalize the features of dataset B and split it into a training set (70%) and a validation set (30%).

[0079] Step 11.4: Design the input layer Input, receiving the maximum reflectance, temperature and humidity profile (24 layers), K index, CAPE, a total of 51 features;

[0080] Step 11.5: Perform feature embedding, use the fully connected layer Linear and dimension transformation Reshape to reshape the 51 features in each sample into a matrix acceptable to ResNet. For more effective gradient descent and back propagation, add the ReLU activation function during the reshaping process;

[0081] Step 11.6: Remove the original convolutional layer, retain the residual block and the fully connected layer to form the modified ResNet;

[0082] Step 11.7: Design the output layer, add the fully connected layer Linear, and output the radar reflectivity (24 layers); the radar reflectivity is the three-dimensional reflectivity of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0;

[0083] Step 11.8: Set the training parameters, such as the number of training rounds (epoch) and batch size, for the model built by steps 11.4-11.7, use the training set to train the model, use the validation set to adjust the hyperparameters, and save the optimal model with the minimum arithmetic mean of the mean square error (MSE) or mean absolute error (MAE) of all layers. Use early stopping and model checkpoints to prevent overfitting during training. Finally, the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet is obtained. After obtaining the 3D proxy reflectivity in the inversion area by the above method, the intensity, distribution, and structure of the near-real-time 3D proxy reflectivity are analyzed to carry out severe convective weather monitoring in the radar detection blind area. Analyze the evolution of the 3D proxy reflectivity over the past period of time, and use the extrapolation method to carry out severe convective weather warnings. At the same time, three-dimensional proxy reflectivity assimilation, i.e. lightning location observation data assimilation, can also be carried out in the numerical forecast model, and the formed analysis field can be used as the initial field of the numerical forecast model to predict the occurrence and development of severe convective weather in the future through the numerical forecast model, so as to realize the monitoring, early warning and forecast of severe convective weather.

[0084] The embodiments described above are merely descriptions of preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A lightning-3D proxy reflectivity inversion method based on ResNet, characterized by: The steps include: Step 1: Obtain lightning location observation data, including: time, longitude and latitude, altitude, current intensity, lightning type, instrument model, fusion mark, and quality control mark of lightning location observation records; Step 2: Filter the lightning location observation records with instrument model DDW1, use the fusion mark and quality control mark to perform quality control on the DDW1 lightning location observation records, and use the lightning type to classify the DDW1 lightning location observation records into cloud-flash observation records and ground-flash observation records; Step 3: Select the inversion area according to the lightning activity range, divide the inversion area into grids, calculate the number of lightning location observation records, cloud flash observation records, ground flash observation records, and positive ground flash observation records of each grid point in the grid, and record them as total flash frequency, cloud flash frequency, ground flash frequency, and positive ground flash frequency respectively; Step 4: Calculate the cloud-to-ground ratio and positive-to-ground ratio of each grid point with a total flash frequency greater than 0 in the inversion area, where the cloud-to-ground ratio is the ratio of the cloud-to-cloud flash frequency to the total flash frequency, and the positive-to-ground ratio is the ratio of the positive-to-ground flash frequency to the ground-to-ground flash frequency; Step 5: Calculate the spatial coefficient r of each grid point in the inversion area where the total flash frequency is greater than 0 c , time coefficient t c and the spatiotemporal coefficient tr c ; Step 6: Sort the cloud flash location observation records of each grid point with a total flash frequency greater than 0 in order of the time-space coefficient from small to large. For each grid point, retain the first N1 cloud flash observation records with the smallest time-space coefficient, save the space coefficient, time coefficient, and lightning height. If the cloud flash frequency is less than N1, fill it with the default value; Step 7: Sort the ground-to-ground lightning observation records of each grid point with a total lightning frequency greater than 0 in order of the time-space coefficient from small to large. For each grid point, retain the first N2 ground-to-ground lightning observation records with the smallest time-space coefficient, save the space coefficient, time coefficient, and current intensity. If the ground-to-ground lightning frequency is less than N2, fill it with the default value; Step 8: Input the total flash frequency, cloud-to-flash ratio, positive ground-to-ground ratio, spatial coefficient, time coefficient, lightning height of the first N1 cloud-to-ground flash observation records with the smallest time-space coefficient, and spatial coefficient, time coefficient, and current intensity of the first N2 ground-to-ground flash observation records with the smallest time-space coefficient into the lightning-maximum proxy reflectivity inversion model based on ResNet to obtain the maximum proxy reflectivity of each grid point among all grid points; Step 9: Download FY-4B GIIRS atmospheric vertical detection products and atmospheric instability index products, and use the quality control mark to perform quality control on the products; Step 10: Extract the atmospheric environmental characteristics of each grid point in the inversion area where the total flash frequency is greater than 0, that is, extract the temperature and humidity profile, K index, and CAPE of the point closest to the grid point in the past 2 hours. If the temperature and humidity profile, K index, and CAPE of the grid point do not exist or fail to pass quality control, then remove the point, and interpolate the temperature and humidity to the height layer of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0 according to the height information of the FY-4BGIIRS atmospheric vertical detection product; Step 11: Input the maximum proxy reflectivity, temperature and humidity profile, K index, and CAPE of each grid point in the lightning activity range into the ResNet-based maximum proxy reflectivity-three-dimensional proxy reflectivity inversion model to obtain the three-dimensional proxy reflectivity of each grid point.

2. According to the ResNet-based lightning-3D proxy reflectivity inversion method of claim 1, it is characterized by: The quality control method for DDW1 type lightning location observation records in step 2 is: the fusion mark is 0, the quality control mark is 0, and if the above two conditions are met at the same time, the lightning location observation record is considered reliable. The classification method is: when the lightning type is 0, the lightning location observation record is a cloud-flash observation record; when the lightning type is 1, the lightning location observation record is a ground-flash observation record.

3. The ResNet-based lightning-3D proxy reflectivity inversion method according to claim 2, characterized in that: Step 3 is as follows: referring to the division method of longitude and latitude information of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0, the inversion area is divided into corresponding longitude and latitude grids of 0.01°*0.01°, and each grid point is taken as the center of the circle. The number of lightning location observation records, cloud flash observation records, ground flash observation records, and positive ground flash observation records within a search radius of 0.08° from 30 minutes before to 10 minutes after the hour are calculated, and recorded as the total flash frequency, cloud flash frequency, ground flash frequency, and positive ground flash frequency, respectively.

4. The method for lightning-3D proxy reflectivity inversion based on ResNet according to claim 3, characterized in that: The spatial coefficient r in step 5 c , time coefficient t c and the spatiotemporal coefficient tr c The calculation is as follows: t c =(t lgt -t radar ) / 1800 Where, t lgt ,t radar The lightning occurrence time and the maximum proxy reflectivity time, i.e. the hourly moment, lon lgt ,lon radar are the longitude of the lightning occurrence location and the longitude of the maximum proxy reflectivity location, i.e., the grid point of equal longitude and latitude, lat lgt ,lat radar They are the latitude of the lightning occurrence location and the latitude of the maximum proxy reflectivity location, i.e., the grid point of equal longitude and latitude.

5. The method for lightning-3D proxy reflectivity inversion based on ResNet according to claim 4, characterized in that: In step 6 and step 7, N1 is the cloud-to-ground flash cutoff frequency determined by a selected percentage based on the cumulative frequency curve of cloud-to-ground flash frequency, and N2 is the ground-to-ground flash cutoff frequency determined by a selected percentage based on the cumulative frequency curve of ground-to-ground flash frequency, and the selected percentage is 90%, 95% or 99%.

6. The ResNet-based lightning-3D proxy reflectivity inversion method according to claim 5, characterized in that: The specific steps for constructing the ResNet-based lightning-maximum proxy reflectivity inversion model in step 8 are as follows: Step 8.1: Collect the lightning location data of the National Lightning Detection System and the combined reflectivity factor products of the Weather Radar Mosaic System V3.0 for a certain period of time, with a horizontal resolution of 0.01°*0.01°; perform quality control and classification according to step 2, taking into account the blind spots of radar detection and the missing or late arrival of radar data, and obtain the complete DDW1 lightning location observation records and maximum reflectivity data sets of the maximum reflectivity mosaic when lightning activity occurs within the radar coverage area through screening; Step 8.2: Take each grid point in the collected weather radar mosaic system V3.0 combined reflectivity factor product as the center of the circle and follow steps 3-7 to obtain the total flash frequency, cloud-to-flash ratio, positive ground-to-ground flash ratio, spatial coefficient, time coefficient, lightning height of the first N1 cloud-to-ground flash observation records with the smallest time-space coefficient, spatial coefficient, time coefficient, current intensity, and maximum reflectivity of the first N2 ground-to-ground flash observation records with the smallest time-space coefficient for each grid point in the radar coverage area when lightning activity occurs, and form data set A; Step 8.3: Perform feature normalization on dataset A and split dataset A into a training set (70%) and a validation set (30%). Step 8.4: Design the input layer Input, receive the total flash frequency, cloud-to-flash ratio, positive ground-to-ground ratio, the spatial coefficient, time coefficient, lightning height of the first N1 cloud-to-ground flash observation records with the smallest time-space coefficient, the spatial coefficient, time coefficient, and current intensity of the first N2 ground-to-ground flash observation records with the smallest time-space coefficient, a total of 3+3N 1+ 3N2 features; Step 8.5: Perform feature embedding, using the fully connected layer Linear and dimension transformation Reshape to embed 3+3N in each sample 1+ The 3N2 features are reshaped into a matrix acceptable to ResNet. For more efficient gradient descent and back propagation, the ReLU activation function is added during the reshaping process. Step 8.6: Remove the original convolutional layer, retain the residual block and the fully connected layer to form the modified ResNet; Step 8.7: Design the output layer, add the average pooling layer AveragePooling and the fully connected layer Linear, and output the maximum reflectivity; the maximum reflectivity is specifically the combined reflectivity of the weather radar puzzle system V3.0 combined reflectivity factor product; Step 8.8: Set the training parameters for the model built through steps 8.4-8.7, use the training set to train the model, use the validation set to adjust the hyperparameters, save the optimal model with the minimum mean square error (MSE) or mean absolute error (MAE), use early stopping and model checkpoints to prevent overfitting during training, and obtain the lightning-maximum proxy reflectivity inversion model based on ResNet.

7. The ResNet-based lightning-3D proxy reflectivity inversion method according to claim 6, characterized in that: The quality control method in step 9 is: when the quality control mark is 0 or 1, the product is considered reliable.

8. The ResNet-based lightning-3D proxy reflectivity inversion method according to claim 7, characterized in that: The specific steps of constructing the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet in step 11 are as follows: Step 11.1: Collect the three-dimensional reflectivity factor products of the weather radar mosaic system V3.0 in the same time period as in step 8.

1. The horizontal resolution of the product is 0.01°*0.01°; Step 11.2: According to step 8, obtain the maximum reflectivity of each grid point in all grid points when lightning activity occurs within the radar coverage area, and extract the radar reflectivity of each grid point according to the three-dimensional reflectivity factor product information of the weather radar puzzle system V3.0, and obtain the maximum reflectivity and radar reflectivity data set of each grid point in all grid points; collect the FY-4B GIIRS atmospheric vertical detection products and atmospheric instability index products in the same time period of step 8.1, and perform quality control; according to step 10, obtain the temperature and humidity profile, K index, and CAPE of each grid point in the combined reflectivity factor product grid collected when lightning activity occurs within the radar coverage area, interpolate the temperature and humidity profile, and interpolate the temperature and humidity profile into the height layer of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0, and combine the maximum reflectivity and radar reflectivity data sets of all grid points to form data set B; Step 11.3: Normalize the features of dataset B and split it into a training set (70%) and a validation set (30%). Step 11.4: Design the input layer Input, which receives the maximum reflectivity, interpolated temperature and humidity profile, K index, CAPE, a total of 51 features; Step 11.5: Perform feature embedding, use the fully connected layer Linear and dimension transformation Reshape to reshape the 51 features in each sample into a matrix acceptable to ResNet. For more effective gradient descent and back propagation, add the ReLU activation function during the reshaping process; Step 11.6: Remove the original convolutional layer, retain the residual block and the fully connected layer to form the modified ResNet; Step 11.7: Design the output layer, add the fully connected layer Linear, and output the radar reflectivity. The radar reflectivity is the three-dimensional reflectivity of the three-dimensional reflectivity factor product of the weather radar puzzle system V3.0; Step 11.8: Set the training parameters for the model built through steps 11.4-11.7, train the model using the training set, adjust the hyperparameters using the validation set, save the optimal model with the minimum arithmetic mean of the mean square error (MSE) or mean absolute error (MAE) of all layers as the standard; use early stopping and model checkpoints to prevent overfitting during training; obtain the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet.

9. The ResNet-based lightning-3D proxy reflectivity inversion method according to claim 8, characterized in that: In step 10, the three-dimensional reflectivity factor product height layer of the weather radar puzzle system V3.0 is 24 layers, the temperature and humidity profile is interpolated to 24 layers in step 11.2, and the radar reflectivity is 24 layers in step 11.7.

Images