A ResNet-based method for lightning-3D proxy reflectivity inversion

Abstract

This invention discloses a ResNet-based method for retrieving three-dimensional surrogate reflectance from lightning, belonging to the field of applied meteorology. The method includes acquiring lightning location observation data, screening, quality control, and classification of lightning location observation records, determining a grid, statistically analyzing data at grid points, processing the statistical data, calculating spatial, temporal, and spatiotemporal coefficients, sorting the lightning location observation records at each grid point, processing the sorted lightning location observation records and other data to obtain the maximum surrogate reflectance, and obtaining the three-dimensional surrogate reflectance by combining the maximum surrogate reflectance with the atmospheric environmental characteristics, thus achieving three-dimensional surrogate reflectance inversion. This invention provides more accurate three-dimensional surrogate reflectance inversion and effectively solves the problem of insufficient monitoring and early warning capabilities for severe convective weather in radar detection blind zones.

Description

Technical Field

[0001] This invention relates to the field of applied meteorology technology, and in particular to a ResNet-based method for lightning-three-dimensional proxy reflectivity inversion. Background Technology

[0002] Lightning location observation has advantages such as high resolution, wide detection range, minimal impact from terrain, and continuous monitoring, making it highly effective in indicating severe convective weather. Although my country has deployed multiple weather radars, certain radar detection blind spots still exist, especially in mountainous areas prone to terrain obstruction, where radar detection capabilities are weak and the ability to monitor and warn of severe convective weather is very limited. Lightning location observation plays a crucial supplementary role to radar detection. The lightning-3D proxy reflectivity inversion method can be used to invert lightning location observations into 3D proxy reflectivity, enabling the monitoring and early warning of the occurrence and development of severe convective weather.

[0003] Previous lightning-3D surrogate reflectivity inversion methods only considered lightning frequency, neglecting other information about lightning activity. With increasing demands for accuracy in monitoring and early warning of severe convection, relying solely on lightning frequency for inversion is insufficient to meet current monitoring and early warning needs. Furthermore, the traditional maximum surrogate reflectivity-3D surrogate reflectivity inversion method uses fixed data such as vertical profile coefficients derived from historical observations. Therefore, the process of inverting from maximum surrogate reflectivity to 3D surrogate reflectivity is highly rigid and inflexible. However, the current atmospheric environment influences the distribution and structural characteristics of 3D radar reflectivity, making the accuracy of inversion using only vertical profile coefficients significantly insufficient, thus affecting the forecast accuracy of subsequent lightning data assimilation in numerical models. Summary of the Invention

[0004] The technical problem this invention aims to solve is to provide a ResNet-based lightning-3D proxy reflectivity inversion method. It incorporates various lightning activity information related to radar reflectivity, along with atmospheric environmental characteristics prior to lightning strikes. The method utilizes a ResNet residual neural network for inversion to address the insufficient monitoring and early warning capabilities for severe convective weather in radar detection blind zones, thereby improving monitoring accuracy.

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

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

[0007] Step 2: Screen the lightning location observation records of the DDW1 instrument model, use fusion identifier and quality control identifier to perform quality control on the DDW1 lightning location observation records, and classify the DDW1 lightning location observation records by lightning type into cloud lightning observation records and ground lightning observation records;

[0008] Step 3: Select the inversion area based on the range of lightning activity, divide the inversion area into grids, and calculate the number of lightning location observation records, cloud lightning observation records, ground lightning observation records, and direct ground lightning observation records for each grid point. These are recorded as the total lightning frequency, cloud lightning frequency, ground lightning frequency, and direct ground lightning frequency, respectively.

[0009] Step 4: Calculate the cloud-to-ground flash ratio and the positive-ground flash ratio for each grid point in the inversion region where the total flash frequency is greater than 0. The cloud-to-ground flash ratio is the ratio of the cloud flash frequency to the total flash frequency, and the positive-ground flash ratio is the ratio of the positive-ground flash frequency to the ground flash frequency.

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

[0011] Step 6: Sort the cloud lightning location observation records of each grid point with a total flash frequency greater than 0 in ascending order of spatiotemporal coefficient. For each grid point, retain the top N1 cloud lightning observation records with the smallest spatiotemporal coefficient, and save the spatial coefficient, time coefficient, and lightning height. If the cloud lightning frequency is less than N1, fill in the missing values ​​with default values.

[0012] Step 7: Sort the ground flash location observation records of each grid point with a total flash frequency greater than 0 in ascending order of spatiotemporal coefficient. For each grid point, retain the top N2 ground flash observation records with the smallest spatiotemporal coefficient, and save the spatial coefficient, time coefficient, and current intensity. If the ground flash frequency is less than N2, fill in the missing values ​​with default values.

[0013] Step 8: Input the total flash frequency, cloud-to-lightning ratio, ground-to-ground flash ratio, spatial coefficients, temporal coefficients of the N1 cloud-to-lightning observation records with the smallest spatiotemporal coefficients, lightning height, and spatial coefficients, temporal coefficients, and current intensity of the N2 ground-to-ground flash observation records with the smallest spatiotemporal coefficients into the ResNet-based lightning-maximum surrogate reflectivity inversion model to obtain the maximum surrogate reflectivity of each grid point in all grid points;

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

[0015] Step 10: Extract the atmospheric environmental features of each grid point in the grid points with a total flash frequency greater than 0 in the inversion area, that is, extract the temperature and humidity profile, K index, and CAPE of the nearest point 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 the quality control, then remove the point. Based on the height information of the FY-4BGIIRS atmospheric vertical sounding product, interpolate the temperature and humidity to the height layer of the three-dimensional reflectivity factor product of the weather radar mosaic system V3.0. There are 24 height layers.

[0016] Step 11: Input the maximum surrogate reflectance, temperature and humidity profile, K-index, and CAPE of each grid point in the lightning activity range into the ResNet-based maximum surrogate reflectance-3D surrogate reflectance inversion model to obtain the 3D surrogate reflectance 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 lightning location observation record in step 2 is: the fusion identifier is 0, the quality control identifier is 0, and if both of the above two conditions are met, 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 lightning observation record; when the lightning type is 1, the lightning location observation record is a ground lightning observation record.

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

[0019] A further improvement to the technical solution of this invention lies in: the spatial coefficient r in step 5. c Time coefficient t c and spatiotemporal coefficient tr c The calculation method is as follows:

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

[0021]

[0022] In the formula, t lgt t radar These are the time of lightning occurrence and the time of maximum proxy reflectivity, i.e., the hour.lgt lon radar These represent the longitude of the lightning strike location and the location of the maximum proxy reflectivity, i.e., the longitude of the grid points on the same latitude and longitude grid. lgt lat radar These represent the latitude of the lightning strike location and the latitude of the grid point at the location of maximum proxy reflectivity, respectively.

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

[0024] A further improvement to the technical solution of this invention lies in the following steps for constructing the ResNet-based lightning-maximum surrogate reflectivity inversion model in step 8:

[0025] Step 8.1: Collect lightning location data from the National Lightning Detection System and reflectivity factor products from the Weather Radar Mosaic System V3.0 over a certain period of time. The horizontal resolution of the products is 0.01°*0.01°. Perform quality control and classification according to Step 2. Considering that there are blind spots in radar detection and that radar data may be missing or arrive late, the complete DDW1 lightning location observation records and maximum reflectivity datasets with maximum reflectivity mosaics when lightning activity occurs within the radar coverage area are obtained through screening.

[0026] Step 8.2: Using each grid point in the collected weather radar mosaic system V3.0 combined reflectivity factor product grid as the center, follow steps 3-7 to obtain the total flash frequency, cloud-to-ground flash ratio, ground-to-ground flash ratio, spatial coefficient, temporal coefficient, lightning height, spatial coefficient, temporal coefficient, current intensity, and maximum reflectivity of each grid point when lightning activity occurs within the radar coverage area, forming dataset 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 to receive the total flash frequency, cloud-to-flash ratio, ground-to-ground flash ratio, the spatial coefficient, time coefficient, and lightning height of the top N1 cloud-to-flash observation records with the smallest spatiotemporal coefficient, and the spatial coefficient, time coefficient, and current intensity of the top N2 ground-to-ground flash observation records with the smallest spatiotemporal coefficient, for a total of 3+3N. 1+ 3N2 features;

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

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

[0031] Step 8.7: Design the output layer, add an average pooling layer (AveragePooling) and a fully connected linear layer (Linear), and output the maximum reflectivity; the maximum reflectivity is specifically the combined reflectivity of the combined reflectivity factor product of the Weather Radar Mosaic System V3.0.

[0032] Step 8.8: Using the model constructed in Steps 8.4-8.7, set the training parameters, train the model using the training set, and adjust the hyperparameters using the validation set. Save the optimal model with the minimum mean squared error (MSE) or mean absolute error (MAE) as the standard. During training, use early stopping and model checkpoints to prevent overfitting. The resulting ResNet-based lightning-maximum surrogate reflectivity inversion model is obtained.

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

[0034] A further improvement to the technical solution of this invention lies in the following steps for constructing the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet in step 11:

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

[0036] Step 11.2: Following Step 8, obtain the maximum reflectivity of each grid point within the radar coverage area when lightning activity occurs. Extract the radar reflectivity of each grid point based on the 3D reflectivity factor product information of the Weather Radar Mosaic System V3.0 to obtain the maximum reflectivity and radar reflectivity datasets for each grid point. Collect the FY-4B GIIRS atmospheric vertical sounding product and atmospheric instability index product from the same time period in Step 8.1 and perform quality control. Following Step 10, obtain the temperature and humidity profiles, K index, and CAPE of each grid point within the combined reflectivity factor product grid collected when lightning activity occurs within the radar coverage area over the past 2 hours. Interpolate the temperature and humidity profiles and use the interpolated temperature and humidity profiles as the height layer of the 3D reflectivity factor product of the Weather Radar Mosaic System V3.0. Combine the maximum reflectivity and radar reflectivity datasets of all grid points to form dataset B.

[0037] Step 11.3: Perform feature normalization on dataset B, and split dataset B into a training set (70%) and a validation set (30%).

[0038] Step 11.4: Design the input layer to receive 51 features, including maximum reflectivity, interpolated temperature and humidity profile, K-index, and CAPE.

[0039] Step 11.5: Perform feature embedding. Use a fully connected Linear layer and dimensionality transformation Reshape to reshape the 51 features in each sample into a matrix that can be accepted by ResNet. To make gradient descent and backpropagation more effective, the ReLU activation function is added during the reshaping process.

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

[0041] Step 11.7: Design the output layer, add a fully connected Linear layer, and output the radar reflectivity. The radar reflectivity is the three-dimensional reflectivity of the Weather Radar Mosaic System V3.0 three-dimensional reflectivity factor product.

[0042] Step 11.8: Using the model constructed in Steps 11.4-11.7, set the training parameters, train the model using the training set, and adjust the hyperparameters using the validation set. Save the optimal model with the minimum arithmetic mean squared error (MSE) or mean absolute error (MAE) of all layers. During training, use early stopping and model checkpoints to prevent overfitting. Obtain the ResNet-based maximum surrogate reflectivity-3D surrogate reflectivity inversion model.

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

[0044] The technological advancements achieved by this invention, due to the adoption of the aforementioned technical solutions, are as follows: By considering lightning activity information related to radar reflectivity, such as cloud-to-ground lightning ratio, time, location, intensity, and polarity, and incorporating atmospheric environmental characteristics prior to lightning (temperature and humidity profiles, instability indices (K-index, CAPE)), a lightning-three-dimensional proxy reflectivity inversion method is constructed. The inverted three-dimensional proxy reflectivity more closely reflects actual conditions, significantly improving the usability of existing lightning location observations and Fengyun satellite atmospheric vertical sounding products. Based on the inversion results, the monitoring and early warning of severe convective weather in radar detection blind zones are more accurate, enhancing monitoring and early warning capabilities. Furthermore, it can be applied to numerical weather prediction models at different horizontal resolutions, especially high-resolution models. By conducting joint assimilation of three-dimensional proxy reflectivity (lightning location observations) and three-dimensional radar reflectivity (radar observations), the assimilation application of DDW1-type full-flash location observations in numerical weather prediction models is realized, which has significant practical value for improving numerical weather prediction capabilities in radar detection blind zones. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the inversion method of the present invention;

[0047] Figure 2 This is a schematic diagram illustrating the calculation of the total flash frequency of this invention;

[0048] Figure 3 This is a flowchart of the construction process of the lightning-maximum proxy reflectivity inversion model based on ResNet;

[0049] Figure 4 This is a flowchart of the construction process of the maximum proxy reflectivity-3D proxy reflectivity inversion model based on ResNet. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to embodiments:

[0051] like Figure 1The diagram shows a flowchart of a ResNet-based lightning-3D proxy reflectivity inversion method. The specific steps are as follows:

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

[0053] Step 2: Screen lightning location observation records of the DDW1 instrument model. Perform quality control on the DDW1 lightning location observation records using fusion and quality control identifiers. Classify the DDW1 lightning location observation records by lightning type into cloud-to-cloud lightning observation records and ground-to-ground lightning observation records. The quality control method for DDW1 lightning location observation records is: if both the fusion identifier and the quality control identifier are 0, the lightning location observation record is considered reliable. Classify the DDW1 lightning location observation records by lightning type (cloud-to-cloud lightning observation records and ground-to-ground lightning observation records): when the lightning type is 0, the lightning location observation record is a cloud-to-cloud lightning observation record; when the lightning type is 1, the lightning location observation record is a ground-to-ground lightning observation record.

[0054] Step 3: As Figure 2 As shown, an inversion area is selected based on the lightning activity range and divided into grids. Specifically, referring to the latitude and longitude division method of the Weather Radar Mosaic System V3.0 3D reflectivity factor product, the inversion area is divided into corresponding 0.01°*0.01° latitude and longitude grids. This ensures that the model can be matched when using the corresponding product for training later. The latitude and longitude information of the Weather Radar Mosaic System V3.0 combined reflectivity factor product and the Weather Radar Mosaic System V3.0 3D reflectivity factor product are consistent, and the grid points are consistent. Taking each grid point as the center, the number of lightning location observation records, cloud lightning observation records, ground lightning observation records, and positive ground lightning observation records within a 0.08° search radius from 30 minutes before to 10 minutes after the hour are calculated and recorded as total lightning frequency, cloud lightning frequency, ground lightning frequency, and positive ground lightning frequency, respectively. The positive ground lightning frequency is the ground lightning frequency with positive current intensity.

[0055] Step 4: Calculate the cloud-to-ground flash ratio and the positive-ground flash ratio for each grid point in the inversion region where the total flash frequency is greater than 0. The cloud-to-ground flash ratio is the ratio of the cloud flash frequency to the total flash frequency, and the positive-ground flash ratio is the ratio of the positive-ground flash frequency to the ground flash frequency.

[0056] Step 5: Calculate the spatial coefficient r of each grid point in the inversion region where the total flash frequency is greater than 0. c Time coefficient t c and spatiotemporal coefficient tr c Spatial coefficient r cTime coefficient t c and spatiotemporal coefficient tr c The calculation method is as follows:

[0057]

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

[0059]

[0060] In the formula, t lgt t radar These are the time of lightning occurrence and the time of maximum proxy reflectivity, i.e., the hour. lgt lon radar These represent the longitude of the lightning strike location and the location of the maximum proxy reflectivity, i.e., the longitude of the grid points on the same latitude and longitude grid. lgt lat radar These represent the latitude of the lightning strike location and the latitude of the grid point at the location of maximum proxy reflectivity, respectively.

[0061] Step 6: Sort the cloud-to-lightning observation records for each grid point with a total flash frequency greater than 0, based on the spatiotemporal coefficient from smallest to largest. For each grid point, retain the top N1 cloud-to-lightning observation records with the smallest spatiotemporal coefficient, saving the spatial coefficient, temporal coefficient, and lightning height. If the cloud-to-lightning frequency is less than N1, fill in the missing values ​​with default values. Through analysis of a large amount of historical observation data, determine the cloud-to-lightning cutoff frequency N based on 90% (or 95%, 99%, depending on the actual situation) of the cumulative cloud-to-lightning frequency curve. 1。

[0062] Step 7: Sort the ground flash observation records for each grid point with a total flash frequency greater than 0, in ascending order of spatiotemporal coefficient. For each grid point, retain the top N2 ground flash observation records with the smallest spatiotemporal coefficient, and save the spatial coefficient, time coefficient, and current intensity (positive or negative reflects lightning polarity). If the ground flash frequency is less than N2, fill in the missing values ​​with default values. Through analysis of a large amount of historical observation data, determine the ground flash cutoff frequency N based on 90% (or 95%, 99%, depending on the actual situation) of the cumulative ground flash frequency curve. 2。

[0063] Step 8: Input the total flash frequency, cloud-to-lightning ratio, direct-to-ground flash ratio, spatial coefficients and temporal coefficients of the top N1 cloud-to-lightning observation records with the smallest spatiotemporal coefficients, lightning height, and spatial coefficients and temporal coefficients and current intensity of the top N2 ground-to-ground flash observation records with the smallest spatiotemporal coefficients into the ResNet-based lightning-maximum surrogate reflectivity inversion model to obtain the maximum surrogate reflectivity of each grid point; for example... Figure 3 As shown, the specific steps for constructing the ResNet-based lightning-maximum surrogate reflectivity inversion model are as follows:

[0064] Step 8.1: Collect lightning location data from the National Lightning Detection System and combined reflectivity factor products from the Weather Radar Mosaic System V3.0 over a certain period of time. The horizontal resolution of the products is 0.01°*0.01°. The combined reflectivity factor products contain combined reflectivity and its location (latitude and longitude) information. The grid can be determined based on the latitude and longitude information. The combined reflectivity is used as the maximum reflectivity for subsequent output. Perform quality control and classification according to Step 2. Considering that there are blind spots in radar detection and that radar data may be missing or late, the complete DDW1 lightning location observation records and maximum reflectivity datasets with maximum reflectivity mosaics when lightning activity occurs within the radar coverage area are obtained through screening.

[0065] Step 8.2: Using each grid point within the combined reflectivity factor product grid of the collected weather radar mosaic system V3.0 as the center, follow steps 3-7 to obtain the total flash frequency, cloud-to-ground flash ratio, ground-to-ground flash ratio, spatial coefficients, temporal coefficients, lightning height, spatial coefficients, temporal coefficients, current intensity, and maximum reflectivity of each grid point when lightning activity occurs within the radar coverage area, forming dataset 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 to receive the total flash frequency, cloud-to-flash ratio, ground-to-ground flash ratio, the spatial coefficient, time coefficient, and lightning height of the top N1 cloud-to-flash observation records with the smallest spatiotemporal coefficient, and the spatial coefficient, time coefficient, and current intensity of the top N2 ground-to-ground flash observation records with the smallest spatiotemporal coefficient, for a total of 3+3N. 1+ 3N2 features;

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

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

[0070] Step 8.7: Design the output layer, add an average pooling layer (AveragePooling) and a fully connected linear layer (Linear), and output the maximum reflectivity; the maximum reflectivity is specifically the combined reflectivity of the combined reflectivity factor product of the Weather Radar Mosaic System V3.0.

[0071] Step 8.8: Using the model built in Steps 8.4-8.7, set training parameters such as the number of training epochs and batch size. Train the model using the training set and adjust hyperparameters using the validation set. Save the optimal model with the minimum mean squared error (MSE) or mean absolute error (MAE) as the standard. Use early stopping and model checkpointing during training to prevent overfitting. The resulting ResNet-based lightning-maximum surrogate reflectivity inversion model is obtained.

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

[0073] Step 10: Extract the atmospheric environmental features of each grid point in the grid points with a total flash frequency greater than 0 in the inversion area, that is, extract the temperature and humidity profile, K index, and CAPE of the nearest point 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 the quality control, then remove the point. Based on the height information of the FY-4BGIIRS atmospheric vertical sounding product, interpolate the temperature and humidity to the height layer of the three-dimensional reflectivity factor product of the weather radar mosaic system V3.0. There are 24 height layers.

[0074] Step 11: Input the maximum surrogate reflectance, temperature and humidity profile, K-index, and CAPE of all grid points within the lightning activity range into the ResNet-based maximum surrogate reflectance-3D surrogate reflectance inversion model to obtain the 3D surrogate reflectance.

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

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

[0077] Step 11.2: Following Step 8, obtain the maximum reflectivity of each grid point within the radar coverage area when lightning activity occurs. Extract the radar reflectivity of each grid point based on the three-dimensional reflectivity factor product of the Weather Radar Mosaic System V3.0 to obtain the maximum reflectivity and radar reflectivity datasets of all grid points. Collect the FY-4B GIIRS atmospheric vertical sounding product and atmospheric instability index product from the same time period in Step 8.1 and perform quality control. Following Step 10, obtain the temperature and humidity profiles (interpolated to 24 layers), K-index, and CAPE of each grid point (i.e., the maximum reflectivity grid point) within the grid of all collected combined reflectivity factor products when lightning activity occurs within the radar coverage area over the past 2 hours. Combine these with the maximum reflectivity and radar reflectivity datasets of all grid points to form dataset B.

[0078] Step 11.3: Perform feature normalization on dataset B, and split dataset B into a training set (70%) and a validation set (30%).

[0079] Step 11.4: Design the input layer to receive 51 features, including maximum reflectivity, temperature and humidity profile (24 layers), K-index, and CAPE.

[0080] Step 11.5: Perform feature embedding. Use a fully connected Linear layer and dimensionality transformation Reshape to reshape the 51 features in each sample into a matrix that can be accepted by ResNet. To make gradient descent and backpropagation more effective, the ReLU activation function is added during the reshaping process.

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

[0082] Step 11.7: Design the output layer, add a fully connected Linear layer, 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 Mosaic System V3.0;

[0083] Step 11.8: Using the model constructed in Steps 11.4-11.7, set training parameters such as the number of training epochs and batch size. Train the model using the training set and adjust hyperparameters using the validation set. Save the optimal model with the minimum arithmetic mean of mean squared error (MSE) or mean absolute error (MAE) across all layers. Early stopping and model checkpointing are used during training to prevent overfitting. Finally, a ResNet-based maximum surrogate reflectance-3D surrogate reflectance inversion model is obtained. After obtaining the 3D surrogate reflectance within the inversion area using the above method, the intensity, distribution, and structure of the near-real-time 3D surrogate reflectance are analyzed to monitor severe convective weather in radar detection blind zones. The evolution of the 3D surrogate reflectance over a past period is analyzed, and extrapolation methods are used to issue severe convective weather warnings. Simultaneously, three-dimensional proxy reflectivity assimilation, i.e., lightning location observation data assimilation, can be carried out in numerical weather prediction models. The resulting analysis field is used as the initial field of the numerical weather prediction model, which then forecasts the occurrence and development of severe convective weather over a future period. This enables the monitoring, early warning, and forecasting of severe convective weather.

[0084] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A ResNet-based method for lightning-3D proxy reflectivity inversion, characterized in that: Includes the following steps: Step 1: Obtain lightning location observation data, including: time, latitude and longitude, altitude, current intensity, lightning type, instrument model, fusion identifier, and quality control identifier of the lightning location observation record; Step 2: Screen the lightning location observation records of the DDW1 instrument model, use fusion identifier and quality control identifier to perform quality control on the DDW1 lightning location observation records, and classify the DDW1 lightning location observation records by lightning type into cloud lightning observation records and ground lightning observation records; Step 3: Select the inversion area based on the range of lightning activity, divide the inversion area into grids, and calculate the number of lightning location observation records, cloud lightning observation records, ground lightning observation records, and direct ground lightning observation records for each grid point. These are recorded as the total lightning frequency, cloud lightning frequency, ground lightning frequency, and direct ground lightning frequency, respectively. Step 4: Calculate the cloud-to-ground flash ratio and the positive-ground flash ratio for each grid point in the inversion region where the total flash frequency is greater than 0. The cloud-to-ground flash ratio is the ratio of the cloud flash frequency to the total flash frequency, and the positive-ground flash ratio is the ratio of the positive-ground flash frequency to the ground flash frequency. Step 5: Calculate the spatial coefficient r of each grid point in the inversion region where the total flash frequency is greater than 0. c Time coefficient t c and spatiotemporal coefficient tr c ; Step 6: Sort the cloud lightning location observation records of each grid point with a total flash frequency greater than 0 in ascending order of spatiotemporal coefficient. For each grid point, retain the top N1 cloud lightning observation records with the smallest spatiotemporal coefficient, and save the spatial coefficient, time coefficient, and lightning height. If the cloud lightning frequency is less than N1, fill in the missing values ​​with default values. Step 7: Sort the ground flash location observation records of each grid point with a total flash frequency greater than 0 in ascending order of spatiotemporal coefficient. For each grid point, retain the top N2 ground flash observation records with the smallest spatiotemporal coefficient, and save the spatial coefficient, time coefficient, and current intensity. If the ground flash frequency is less than N2, fill in the missing values ​​with default values. Step 8: Input the total flash frequency, cloud-to-lightning ratio, ground-to-ground flash ratio, spatial coefficients, temporal coefficients of the N1 cloud-to-lightning observation records with the smallest spatiotemporal coefficients, lightning height, and spatial coefficients, temporal coefficients, and current intensity of the N2 ground-to-ground flash observation records with the smallest spatiotemporal coefficients into the ResNet-based lightning-maximum surrogate reflectivity inversion model to obtain the maximum surrogate reflectivity of each grid point in all grid points; Step 9: Download the FY-4B GIIRS atmospheric vertical sounding product and atmospheric instability index product, and use the quality control label to perform quality control on the products; Step 10: Extract the atmospheric environmental features of each grid point in the grid points with a total flash frequency greater than 0 in the inversion area, that is, extract the temperature and humidity profile, K index, and CAPE of the nearest point 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 the quality control, then remove the point. Based on the height information of the FY-4BGIIRS atmospheric vertical sounding product, interpolate the temperature and humidity to the height layer of the three-dimensional reflectivity factor product of the weather radar mosaic system V3.

0. Step 11: Input the maximum surrogate reflectance, temperature and humidity profile, K-index, and CAPE of each grid point in the lightning activity range into the ResNet-based maximum surrogate reflectance-3D surrogate reflectance inversion model to obtain the 3D surrogate reflectance of each grid point.

2. The lightning-3D proxy reflectivity inversion method based on ResNet according to claim 1, characterized in that: In step 2, the quality control method for DDW1 lightning location observation records is as follows: if both the fusion identifier and the quality control identifier are 0, the lightning location observation record is considered reliable. The classification method is as follows: when the lightning type is 0, the lightning location observation record is a cloud lightning observation record; when the lightning type is 1, the lightning location observation record is a ground lightning observation record.

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

4. The lightning-3D proxy reflectivity inversion method based on ResNet according to claim 3, characterized in that: In step 5, the spatial coefficient r c Time coefficient t c and spatiotemporal coefficient tr c The calculation method is as follows: t c =(t lgt -t radar ) / 1800 In the formula, t lgt t radar These are the time of lightning occurrence and the time of maximum proxy reflectivity, i.e., the hour. lgt lon radar These represent the longitude of the lightning strike location and the location of the maximum proxy reflectivity, i.e., the longitude of the grid points on the same latitude and longitude grid. lgt lat radar These represent the latitude of the lightning location and the latitude of the grid point at the location of maximum proxy reflectivity, respectively.

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

6. The lightning-3D proxy reflectivity inversion method based on ResNet according to claim 5, characterized in that: The specific steps for constructing the ResNet-based lightning-maximum surrogate reflectivity inversion model in step 8 are as follows: Step 8.1: Collect lightning location data from the National Lightning Detection System and reflectivity factor products from the Weather Radar Mosaic System V3.0 over a certain period of time. The horizontal resolution of the products is 0.01°*0.01°. Perform quality control and classification according to Step 2. Considering that there are blind spots in radar detection and that radar data may be missing or arrive late, the complete DDW1 lightning location observation records and maximum reflectivity datasets with maximum reflectivity mosaics when lightning activity occurs within the radar coverage area are obtained through screening. Step 8.2: Using each grid point in the collected weather radar mosaic system V3.0 combined reflectivity factor product grid as the center, follow steps 3-7 to obtain the total flash frequency, cloud-to-ground flash ratio, ground-to-ground flash ratio, spatial coefficient, temporal coefficient, lightning height, spatial coefficient, temporal coefficient, current intensity, and maximum reflectivity of each grid point when lightning activity occurs within the radar coverage area, forming dataset 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 to receive the total flash frequency, cloud-to-flash ratio, ground-to-ground flash ratio, the spatial coefficient, time coefficient, and lightning height of the top N1 cloud-to-flash observation records with the smallest spatiotemporal coefficient, and the spatial coefficient, time coefficient, and current intensity of the top N2 ground-to-ground flash observation records with the smallest spatiotemporal coefficient, for a total of 3+3N. 1+ 3N2 features; Step 8.5: Perform feature embedding, using a fully connected Linear layer and dimensionality transformation Reshape to embed the 3+3N features in each sample. 1+ The 3N2 features are reshaped into a matrix acceptable to ResNet. To enable more efficient gradient descent and backpropagation, a ReLU activation function is added during the reshaping process. Step 8.6: Remove the original convolutional layers, retain the residual blocks and fully connected layers to form the modified ResNet; Step 8.7: Design the output layer, add an average pooling layer (AveragePooling) and a fully connected linear layer (Linear), and output the maximum reflectivity; the maximum reflectivity is specifically the combined reflectivity of the combined reflectivity factor product of the Weather Radar Mosaic System V3.

0. Step 8.8: Using the model constructed in Steps 8.4-8.7, set the training parameters, train the model using the training set, and adjust the hyperparameters using the validation set. Save the optimal model with the minimum mean squared error (MSE) or mean absolute error (MAE) as the standard. During training, use early stopping and model checkpoints to prevent overfitting. The resulting ResNet-based lightning-maximum surrogate reflectivity inversion model is obtained.

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

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

1. The horizontal resolution of the products is 0.01°*0.01°. Step 11.2: Following Step 8, obtain the maximum reflectivity of each grid point within the radar coverage area when lightning activity occurs. Extract the radar reflectivity of each grid point based on the 3D reflectivity factor product information of the Weather Radar Mosaic System V3.0 to obtain the maximum reflectivity and radar reflectivity datasets for each grid point. Collect the FY-4B GIIRS atmospheric vertical sounding product and atmospheric instability index product from the same time period in Step 8.1 and perform quality control. Following Step 10, obtain the temperature and humidity profiles, K index, and CAPE of each grid point within the combined reflectivity factor product grid collected when lightning activity occurs within the radar coverage area over the past 2 hours. Interpolate the temperature and humidity profiles and use the interpolated temperature and humidity profiles as the height layer of the 3D reflectivity factor product of the Weather Radar Mosaic System V3.

0. Combine the maximum reflectivity and radar reflectivity datasets of all grid points to form dataset B. Step 11.3: Perform feature normalization on dataset B, and split dataset B into a training set (70%) and a validation set (30%). Step 11.4: Design the input layer to receive 51 features, including maximum reflectivity, interpolated temperature and humidity profile, K-index, and CAPE. Step 11.5: Perform feature embedding. Use a fully connected Linear layer and dimensionality transformation Reshape to reshape the 51 features in each sample into a matrix that can be accepted by ResNet. To make gradient descent and backpropagation more effective, the ReLU activation function is added during the reshaping process. Step 11.6: Remove the original convolutional layers, retain the residual blocks and fully connected layers to form the modified ResNet; Step 11.7: Design the output layer, add a fully connected Linear layer, and output the radar reflectivity. The radar reflectivity is the three-dimensional reflectivity of the Weather Radar Mosaic System V3.0 three-dimensional reflectivity factor product. Step 11.8: Using the model constructed in Steps 11.4-11.7, set the training parameters, train the model using the training set, and adjust the hyperparameters using the validation set. Save the optimal model with the minimum arithmetic mean squared error (MSE) or mean absolute error (MAE) of all layers. During training, use early stopping and model checkpoints to prevent overfitting. Obtain the ResNet-based maximum surrogate reflectivity-3D surrogate reflectivity inversion model.

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

Images