A 3D convolutional neural network thunderstorm recognition method and system based on volumetric radar data

Abstract

The present invention discloses a 3D convolutional neural network thunderstorm recognition method and system based on volume radar data, including: obtaining original radar volume scan data, converting it into radar volume data files as training data, and constructing a volume radar data set; obtaining ground lightning location data corresponding to the time and region of the original radar volume scan data, and marking the volume radar data set according to the latitude and longitude coordinates of the ground lightning location data as the label of the volume radar data set; constructing a thunderstorm recognition model, using the volume radar data set and its labels to train the network weights of the thunderstorm recognition model; inputting the radar volume scan data to be recognized into the trained thunderstorm recognition model, identifying the probability of thunderstorm occurrence in different regions, and generating thunderstorm recognition results. The present invention retains and utilizes the inter-layer information of the original radar data, and improves the accuracy of thunderstorm recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of thunderstorm identification, and in particular relates to a 3D convolutional neural network thunderstorm identification method and system based on volume radar data. Background Art

[0002] Thunderstorms are a strong weather phenomenon accompanied by lightning and rainfall, which is characterized by suddenness and localization, and usually includes thunder, lightning, and heavy rainfall. In addition, thunderstorms often bring some other extreme weather phenomena, such as convective strong winds, hail, short-term heavy rainfall, tornadoes, etc. Thunderstorms and the extreme weather that accompanies them often cause casualties of people and animals, damage to buildings, and even explosions and forest fires, causing serious losses of life or property. Therefore, accurate and efficient thunderstorm nowcasting is very important for doing a good job in lightning protection and disaster reduction.

[0003] Doppler weather radar detection data has good temporal and spatial resolution and is the main tool for detecting and nowcasting severe convective weather such as thunderstorms. The traditional radar-based thunderstorm identification method is mainly based on the threshold method, which identifies areas where the data meets the pre-set threshold as thunderstorms. However, the traditional identification method has certain limitations, resulting in low identification accuracy. In addition, the traditional identification method cannot make use of the meteorological data accumulated to date.

[0004] Deep learning algorithms have very powerful nonlinear mapping capabilities. They can effectively utilize the historical data of various observation sites, learn massive meteorological data through algorithms, and deeply mine the intrinsic characteristics and connections between data. They are an important supplement to traditional thunderstorm identification and prediction methods. The original radar data is layered structured data and three-dimensional data containing height information. The SWAN meteorological radar uses multiple elevation angles for scanning in each scanning cycle, records the radar echo data at different elevation angles, and finally obtains a radar data file with a hierarchical structure. The current thunderstorm identification methods based on deep learning algorithms are all based on two-dimensional combined reflectivity images. These combined reflectivity images are obtained from three-dimensional radar data through mapping algorithms, but in the mapping process of three-dimensional radar data, the two-dimensional combined reflectivity images lose the inter-layer information at different heights in the original data.

[0005] Currently, 3D convolutional neural networks are often used to process video and medical data, but radar volume scan data is a binary data format that is completely different from video and medical data. Common convolutional network models cannot directly process radar volume scan data. Summary of the invention

[0006] The present invention provides a 3D convolutional neural network thunderstorm recognition method and system based on volume radar data, which converts original radar echo data into volume data, and combines the ground lightning location data to produce a volume radar data set, thereby retaining and utilizing the inter-layer information of the original radar data; the present invention improves the 3D convolutional neural network in view of the deficiency that the convolutional network cannot extract the global features of the volume radar data, so that the 3D convolutional neural network can be applied to the volume radar data set, thereby improving the accuracy of thunderstorm recognition.

[0007] In a first aspect, the present invention discloses a 3D convolutional neural network thunderstorm recognition method based on volume radar data, the method comprising the following steps:

[0008] Step A, obtaining original radar volume scan data, converting the original radar volume scan data into radar reflectivity images of corresponding layers by linear interpolation, graying the radar reflectivity images of all layers, combining the grayed images together in order of layers and saving them as radar volume data files, using the radar volume data files as training data, and constructing a volume radar data set;

[0009] Step B, obtaining ground-to-ground lightning location data corresponding to the time and region of the original radar volume scan data, and marking the volume radar data set according to the latitude and longitude coordinates of the ground-to-ground lightning location data as a label of the volume radar data set;

[0010] Step C, constructing a thunderstorm recognition model based on a 3D convolutional neural network, and using a volumetric radar dataset and its labels to train the network weights of the thunderstorm recognition model;

[0011] Step D: inputting the radar volume scan data to be identified into the trained thunderstorm identification model, identifying the probability of thunderstorm occurrence in different areas, and generating thunderstorm identification results.

[0012] Step B further comprises:

[0013] The corresponding area is rasterized, and the location of the ground-to-ground lightning is located in the corresponding grid according to the latitude and longitude coordinates recorded in the ground-to-ground lightning location data, and isolated ground-to-ground lightning locations are removed;

[0014] The grids within a certain distance from the ground-to-ground lightning area with a distribution density greater than the preset density threshold are marked as thunderstorm-affected areas. Combined with the volumetric radar dataset, the annotations are extended to three-dimensional space, and labels are set for the data in the thunderstorm-affected area to complete the generation of dataset labels.

[0015] As a preferred example, in step C, the thunderstorm recognition model includes an input module, a large kernel feature extraction module, an encoder module, a decoder module and an output module;

[0016] The input module is for the first Group Data , randomly extract a size The offspring As the input of the large kernel feature extraction module, the sub-body Number of channels ;

[0017] The large kernel feature extraction module calculates the sub-volume The eigenvector of , and use it as input to the encoder module:

[0018] ;

[0019] In the formula, Represents a 3D convolution operation with a kernel size of 7, a stride of 2, and a padding of 3; represents the group normalization operation, the eigenvector The size is , the number of channels is expanded to ;

[0020] The encoder module includes four layers of encoders connected in sequence, each layer of encoders includes a convolution module and a global feature mining module connected to each other; wherein the first The calculation formula of the layer encoder is:

[0021] ;

[0022] ;

[0023] in, is the output of the previous encoder, is the convolution module, It is The output of the convolutional module in the layer encoder, represents the output function of the global feature mining module based on self-attention mechanism and convolution, It is The output of the layer encoder;

[0024] The decoder module includes five layers of decoders connected in sequence. Each layer of decoders takes the output of the corresponding layer of encoders and the output of the previous layer of decoders as input. Each layer of decoders contains a convolution module. The features output by the last layer of decoders are sequentially passed through The convolutional layers of size and function to obtain the final output of the thunderstorm identification model.

[0025] As a preferred example, the convolution modules include a first convolution block, a second convolution block and a downsampling block connected in sequence; the first convolution block and the second convolution block each include a depth convolution unit, a first point-by-point convolution channel, a first group of normalization units, a first activation unit, a second convolution channel and a first fusion unit connected in sequence;

[0026] The calculation formula of the first convolution block is:

[0027] ;

[0028] ;

[0029] ;

[0030] The calculation formula of the second convolution block is:

[0031] ;

[0032] ;

[0033] ;

[0034] in, and are the outputs of the depthwise convolutional units of the first and second convolutional blocks, respectively. and is the output of the first convolutional block and the first point-wise convolution channel of the second convolutional block, , , , They represent the group normalization function, Activation function, depthwise convolution function and pointwise convolution function;

[0035] The downsampling block outputs the feature map of the second convolution block Downsample to get the output of the convolution module ; Among them, the downsampling block converts the feature map into The length, width and height are reduced to half of the original ones, and the number of channels is doubled.

[0036] As a preferred example, the global feature mining module includes a second group of normalization units, a multi-head self-attention unit, a first convolution unit, a splicing unit, a second convolution unit, a second activation unit and a second fusion unit;

[0037] The second group normalization unit performs group normalization on the output features of the convolution module, and the group normalization results are respectively input into the multi-head self-attention unit and the first convolution unit to extract local features and global features :

[0038] ;

[0039] ;

[0040] In the formula Represented as a multi-head self-attention mechanism;

[0041] The splicing unit converts the local features and global features After splicing, it is input to the second convolution unit for a 3D convolution operation with a convolution kernel size of 1 to complete the local feature and global features The fusion of , the formula is:

[0042] ;

[0043] The fused feature vector The output feature of the second activation unit is input into the second activation unit, and the second fusion unit is used to fuse the output feature of the second activation unit with the input feature of the global feature mining module to obtain the output feature vector of the global feature mining module. :

[0044] ;

[0045] In the formula, is the activation function.

[0046] In a second aspect, the present invention discloses a 3D convolutional neural network thunderstorm recognition system based on volume radar data, the system comprising:

[0047] The volume radar data set processing module is used to obtain the original radar volume scan data, convert the original radar volume scan data into radar reflectivity images of the corresponding number of layers through linear interpolation, grayscale the radar reflectivity images of all layers, combine the grayscale images together according to the order of the layers and save them as radar volume data files, use the radar volume data files as training data, and construct a volume radar data set;

[0048] The label processing module is used to obtain the ground-to-ground lightning location data corresponding to the time and region of the original radar volume scan data, and to mark the volume radar data set according to the latitude and longitude coordinates of the ground-to-ground lightning location data as the label of the volume radar data set;

[0049] A thunderstorm recognition model building module is used to build a thunderstorm recognition model based on a 3D convolutional neural network and train the network weights of the thunderstorm recognition model using a volumetric radar dataset and its labels;

[0050] The thunderstorm identification model is used to identify the input radar volume scan data to be identified to obtain the probability of thunderstorm occurrence in different areas and generate a thunderstorm identification result.

[0051] As a preferred example, the label processing module includes a grid positioning unit, a positioning data screening unit, a positioning data recording unit and a volume marking unit;

[0052] The grid positioning unit is used to grid the corresponding area;

[0053] The positioning data screening unit is used to remove isolated ground lightning positioning; as a preferred example,

[0054] The positioning data recording unit locates the location where the ground lightning occurs to the corresponding grid according to the latitude and longitude coordinates recorded in the ground lightning positioning data;

[0055] The volume labeling unit marks the ground-to-ground lightning area with a distribution density greater than a preset density threshold as a high-risk thunderstorm area, and marks the surrounding areas of the high-risk thunderstorm area as medium-risk thunderstorm areas and low-risk thunderstorm areas according to the distance, and completes the generation of labels in combination with the volume radar data set.

[0056] As a preferred example, the thunderstorm recognition model includes an input module, a large kernel feature extraction module, an encoder module, a decoder module and an output module;

[0057] The input module is for the first Group Data , randomly extract a size The offspring As the input of the large kernel feature extraction module, the sub-body Number of channels ;

[0058] The large kernel feature extraction module calculates the sub-volume The eigenvector of , and use it as input to the encoder module:

[0059] ;

[0060] In the formula, Represents a 3D convolution operation with a kernel size of 7, a stride of 2, and a padding of 3; represents the group normalization operation, the eigenvector The size is , the number of channels is expanded to ;

[0061] The encoder module includes four layers of encoders connected in sequence, each layer of encoders includes a convolution module and a global feature mining module connected to each other; wherein the first The calculation formula of the layer encoder is:

[0062] ;

[0063] ;

[0064] in, is the output of the previous encoder, is the convolution module, It is The output of the convolutional module in the layer encoder, represents the output function of the global feature mining module based on self-attention mechanism and convolution, It is The output of the layer encoder;

[0065] The decoder module includes five layers of decoders connected in sequence. Each layer of decoders takes the output of the corresponding layer of encoders and the output of the previous layer of decoders as input. Each layer of decoders contains a convolution module. The features output by the last layer of decoders are sequentially passed through The convolutional layers of size and function to obtain the final output of the thunderstorm identification model.

[0066] As a preferred example, the convolution modules include a first convolution block, a second convolution block and a downsampling block connected in sequence; the first convolution block and the second convolution block each include a depth convolution unit, a first point-by-point convolution channel, a first group of normalization units, a first activation unit, a second convolution channel and a first fusion unit connected in sequence;

[0067] The calculation formula of the first convolution block is:

[0068] ;

[0069] ;

[0070] ;

[0071] The calculation formula of the second convolution block is:

[0072] ;

[0073] ;

[0074] ;

[0075] in, and are the outputs of the depthwise convolutional units of the first and second convolutional blocks, respectively. and is the output of the first convolutional block and the first point-wise convolution channel of the second convolutional block, , , , They represent the group normalization function, Activation function, depthwise convolution function and pointwise convolution function;

[0076] The downsampling block outputs the feature map of the second convolution block Downsample to get the output of the convolution module ; Among them, the downsampling block converts the feature map into The length, width and height are reduced to half of the original ones, and the number of channels is doubled.

[0077] As a preferred example, the global feature mining module includes a second group of normalization units, a multi-head self-attention unit, a first convolution unit, a splicing unit, a second convolution unit, a second activation unit and a second fusion unit;

[0078] The second group normalization unit performs group normalization on the output features of the convolution module, and the group normalization results are respectively input into the multi-head self-attention unit and the first convolution unit to extract local features and global features :

[0079] ;

[0080] ;

[0081] In the formula Represented as a multi-head self-attention mechanism;

[0082] The splicing unit converts the local features and global features After splicing, it is input to the second convolution unit for a 3D convolution operation with a convolution kernel size of 1 to complete the local feature and global features The fusion of , the formula is:

[0083] ;

[0084] The fused feature vector The output feature of the second activation unit is input into the second activation unit, and the second fusion unit is used to fuse the output feature of the second activation unit with the input feature of the global feature mining module to obtain the output feature vector of the global feature mining module. :

[0085] ;

[0086] In the formula, is the activation function.

[0087] The beneficial effects of the present invention are:

[0088] First, the 3D convolutional neural network thunderstorm recognition method and system based on volumetric radar data of the present invention can convert meteorological radar volume scan data into volume data, which can be used for deep learning model training after being combined with ground lightning positioning data. It helps to better utilize the radar data accumulated by meteorological stations for a long time, discover more data features from the interlayer structure of radar data, and improve the accuracy of thunderstorm recognition.

[0089] Second, the 3D convolutional neural network thunderstorm recognition method and system based on volume radar data of the present invention improves the 3D convolutional neural network to address the deficiency that the convolutional network cannot extract the global features of volume radar data, so that the 3D convolutional neural network can be applied to the volume radar data set, thereby improving the accuracy of thunderstorm recognition. In addition, the present invention also adds a global feature mining module to the network, which can enhance the network's ability to extract global features of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 It is a flow chart of the 3D convolutional neural network thunderstorm recognition method based on volume radar data;

[0091] Figure 2 This is a schematic diagram of the thunderstorm identification model structure;

[0092] Figure 3 It is a structural diagram of the convolution module;

[0093] Figure 4 It is a structural diagram of the global feature mining module;

[0094] Figure 5 Schematic diagram for comparing experimental results of different models.

[0095] Figure 6 Schematic diagram for comparing the recognition results of different models. DETAILED DESCRIPTION

[0096] The following examples will enable those skilled in the art to more fully understand the present invention, but are not intended to limit the present invention in any way.

[0097] The present invention discloses a 3D convolutional neural network thunderstorm recognition method based on volume radar data, the method comprising the following steps:

[0098] Step A, obtaining original radar volume scan data, converting the original radar volume scan data into radar reflectivity images of corresponding layers by linear interpolation, graying the radar reflectivity images of all layers, combining the grayed images together in order of layers and saving them as radar volume data files, using the radar volume data files as training data, and constructing a volume radar data set;

[0099] Step B, obtaining ground-to-ground lightning location data corresponding to the time and region of the original radar volume scan data, and marking the volume radar data set according to the latitude and longitude coordinates of the ground-to-ground lightning location data as a label of the volume radar data set;

[0100] Step C, constructing a thunderstorm recognition model based on a 3D convolutional neural network, and using a volumetric radar dataset and its labels to train the network weights of the thunderstorm recognition model;

[0101] Step D: inputting the radar volume scan data to be identified into the trained thunderstorm identification model, identifying the probability of thunderstorm occurrence in different areas, and generating thunderstorm identification results.

[0102] See also Figure 1 The 3D convolutional neural network thunderstorm recognition method based on volume radar data of the present invention comprises the following steps:

[0103] 1. Obtain the original radar volume scan data and transform the original radar volume scan data to obtain the corresponding volume radar data set as training data, including:

[0104] The original radar volume scan data is converted into radar reflectivity images of the corresponding number of layers, the radar reflectivity images of all layers are grayed out, and these images are combined together in the order of the number of layers and saved as radar volume data files.

[0105] 2. Obtain the ground-to-ground lightning location data corresponding to the time and region of the original radar volume scan data, and complete the annotation of the volume radar data set according to the latitude and longitude coordinates of the ground-to-ground lightning as the label of the training data, including:

[0106] The corresponding area is rasterized, and the location of the ground flash is located in the corresponding grid according to the latitude and longitude coordinates recorded in the ground flash data. After completing the positioning of all ground flash data in the corresponding time period, the relatively isolated ground flash locations are cleared, and the relatively densely distributed ground flash areas are marked as high-risk thunderstorm areas. The surrounding areas are marked as medium-risk areas and low-risk areas according to the distance from the high-risk areas. Finally, the labels are generated in combination with the radar volume data.

[0107] 3. Use the volumetric radar dataset and its labels to train the network weights of the thunderstorm recognition model and train the thunderstorm recognition model. Figure 2 The thunderstorm recognition model includes an input module, a large kernel feature extraction module, an encoder module, a decoder module, and an output module. The training process of the thunderstorm recognition model includes:

[0108] 3.1 For the volume radar dataset Group Data , randomly extract a size The offspring As the input of the model, the number of channels of this sub-body .

[0109] 3.2 Calculating model input using the large kernel feature extraction module The eigenvector of As input to the model encoder, include:

[0110] ;

[0111] in, Represents a 3D convolution operation with a kernel size of 7, a step size of 2, and a padding of 3. This convolution operation is used to calculate the feature vector of the model input. ; represents the group normalization operation, the eigenvector The size is , the number of channels is expanded to .

[0112] 3.3 The encoder of the thunderstorm recognition model of the present invention is divided into four layers, and the calculation formula of each layer of encoder is:

[0113] ;

[0114] ;

[0115] in, is the output of the previous encoder, is the convolution module, is the output result of the convolution module in this layer encoder, represents the global feature mining module based on self-attention mechanism and convolution, is the output of the encoder at this layer.

[0116] 3.3.1 Convolutional Module

[0117] See also Figure 3 , the convolution module of each layer of the encoder contains two convolution blocks and one downsampling block. The formula used for the two convolution blocks is:

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] in, and is the output of the depthwise convolution of different convolutional blocks, and is the output of the first point-wise convolution of different convolutional blocks, , , , They represent group normalization, Activation functions, depthwise convolutions, and pointwise convolutions.

[0125] Get the output of two convolutional blocks After that, it is input into the downsampling block to obtain the output of the convolution module The downsampling block converts the feature map into The length, width and height are reduced to half of the original ones, and the number of channels is doubled.

[0126] 3.3.2 Global Feature Mining Module

[0127] See also Figure 4 The global feature mining module first converts the output of the convolution module of the encoder to its own layer They are respectively input into the convolution operation and self-attention mechanism in the feature mining module to obtain local features. and global features , the formula is:

[0128] ;

[0129] ;

[0130] in Represented as a multi-head self-attention mechanism.

[0131] Then the local features and global features After splicing, a 3D convolution operation with a convolution kernel size of 1 is performed to perform local features and global features The fusion of , the formula is:

[0132] .

[0133] Finally, the fused feature vector Input into the activation function to obtain the output feature vector of the global feature mining module , the formula is:

[0134] ;

[0135] In the formula, is an activation function; the output feature vector It will continue to be input into the next layer of encoder.

[0136] 3.4 Eigenvectors After passing through the encoder, the intermediate layer vector is obtained , the middle layer vector Input the decoder to obtain the initial prediction result.

[0137] 3.4.1 Decoder

[0138] The decoder is divided into five layers, and its input comes from the output of the corresponding layer encoder and the output of the previous layer decoder or the intermediate layer vector The structure of the decoder is similar to that of the encoder, except that an upsampling block is used to replace the downsampling block in the convolution module. The upsampling block doubles the length, width and height of the feature vector through a convolution operation, and the number of channels becomes the original .

[0139] The training loss value is calculated according to the initial prediction result and the model input, the parameters of the thunderstorm recognition model are optimized according to the training loss value, the training process of the thunderstorm recognition model is iterated repeatedly until the training loss value converges, and the trained thunderstorm recognition model is output and saved.

[0140] 4. Convert the radar volume scan data required for thunderstorm identification and input it into the pre-trained thunderstorm identification model to obtain the final thunderstorm identification result;

[0141] This embodiment uses a volumetric radar data set and its annotations to train a thunderstorm recognition model, which can directly rely on the original radar volume scan data to identify areas with a high probability of thunderstorms, making up for the deficiency that only two-dimensional combined reflectivity images could be used for recognition in the past, and can effectively utilize the interlayer structure information of the original radar volume scan data.

[0142] Depend on Figure 5It can be seen that this embodiment also modifies some traditional models to enable them to process volume radar data sets for comparative experiments. These traditional models include 3D UX-Net, UNETR, SwinUNTER, TransBTS, and nnFormer. From the comparative experiments, it can be seen that the modifications such as the global feature mining module added to the network by the present invention can effectively improve the accuracy of the final recognition result.

[0143] Figure 6 By visualizing the recognition results of some radar data by the model of the present invention and the above traditional models, it can be seen that the thunderstorm recognition result of the model of the present invention is closer to the actual thunderstorm area, and the detailed features are more detailed and accurate. Compared with other traditional models, the recognition effect is better overall.

[0144] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A 3D convolutional neural network thunderstorm recognition method based on volume radar data, characterized in that: The method comprises the following steps: Step A, obtaining original radar volume scan data, converting the original radar volume scan data into radar reflectivity images of corresponding layers by linear interpolation, graying the radar reflectivity images of all layers, combining the grayed images together in order of layers and saving them as radar volume data files, using the radar volume data files as training data, and constructing a volume radar data set; Step B, obtaining ground-to-ground lightning location data corresponding to the time and region of the original radar volume scan data, and marking the volume radar data set according to the latitude and longitude coordinates of the ground-to-ground lightning location data as a label of the volume radar data set; Step C, constructing a thunderstorm recognition model based on a 3D convolutional neural network, and using a volumetric radar dataset and its labels to train the network weights of the thunderstorm recognition model; Step D, subjecting the radar volume scan data to be identified to the same conversion process as in step A and then inputting it into the trained thunderstorm identification model, and outputting the location of the thunderstorm affected area within the coverage of the radar volume scan data; Step B further comprises: The corresponding area is rasterized, and the location of the ground-to-ground lightning is located in the corresponding grid according to the latitude and longitude coordinates recorded in the ground-to-ground lightning location data, and isolated ground-to-ground lightning locations are removed; The lightning area with a distribution density greater than the preset density threshold and the grids within a certain distance around it are marked as thunderstorm affected areas. The annotations are extended to three-dimensional space based on the volume radar dataset, and labels are set for the data in the thunderstorm affected area to complete the generation of dataset labels. In step C, the thunderstorm recognition model includes an input module, a large kernel feature extraction module, an encoder module, a decoder module and an output module; The input module is for the i-th group of data V in the volume radar data set i , randomly extract a sub-body P of size H×W×D i As the input of the large kernel feature extraction module, the sub-body P i The number of channels C = 1; The large kernel feature extraction module calculates the sub-volume P i The feature vector Z is taken as the input of the encoder module: Z=GN(Conv(P i )); Where Conv(·) represents a 3D convolution operation with a kernel size of 7, a step size of 2, and a padding of 3; GN(·) represents a group normalization operation with a feature vector Z of size The number of channels is expanded to C = 48; The encoder module includes four layers of encoders connected in sequence, and each layer of encoder includes a convolution module and a global feature mining module connected to each other; wherein the calculation formula of the i-th layer encoder is: Z′ i =Conv Block(Z i-1 ); Z i =GFE(Z′ i ); Among them, Z i-1 is the output of the previous encoder, Conv Block (·) is the convolution module, Z′ i is the output result of the convolution module in the i-th layer encoder, GFE(·) represents the output function of the global feature mining module based on the self-attention mechanism and convolution, and Z i is the output of the i-th layer encoder; The decoder module includes five layers of decoders connected in sequence. Each decoder layer takes the output of the encoder of the corresponding layer and the output of the decoder of the previous layer as input. Each decoder layer contains a convolution module. The features output by the last decoder layer are sequentially passed through a 1×1×1 convolution layer and a softmax function to obtain the final output of the thunderstorm recognition model. The convolution modules each include a first convolution block, a second convolution block, and a downsampling block connected in sequence; the first convolution block and the second convolution block each include a depth convolution unit, a first point-by-point convolution channel, a first group of normalization units, a first activation unit, a second convolution channel, and a first fusion unit connected in sequence; The calculation formula of the first convolution block is: The calculation formula of the second convolution block is: in, and are the outputs of the depthwise convolutional units of the first and second convolutional blocks, respectively. and is the output of the first convolutional block and the first point-wise convolution channel of the second convolutional block, GN(·), GELU(·), DWC(·), DCS(·) represent the group normalization function, GELU activation function, depth convolution function and point-wise convolution function, respectively; The downsampling block outputs the feature map z of the second convolution block l+1 Downsample to get the output of the convolution module Among them, the downsampling block converts the feature map z into l+1 The length, width and height are reduced to half of the original ones, and the number of channels is doubled; The global feature mining module includes a second group of normalization units, a multi-head self-attention unit, a first convolution unit, a splicing unit, a second convolution unit, a second activation unit and a second fusion unit; The second group normalization unit performs group normalization on the output features of the convolution module, and the group normalization results are respectively input into the multi-head self-attention unit and the first convolution unit to extract the local features Z1 and the global features Z2: Z1=Conv(GN(Z′ i )); Z2=MSA(GN(Z i i )); Where MSA(·) represents the multi-head self-attention mechanism; After the splicing unit splices the local feature Z1 and the global feature Z2, the splicing unit inputs the splicing unit into the second convolution unit for a 3D convolution operation with a convolution kernel size of 1 to complete the fusion of the local feature Z1 and the global feature Z2, and obtain the fused feature vector Z″ i , the formula is: Z″ i =Conv(Concat(Z1,Z2)); The fused feature vector Z″ i The output feature of the second activation unit is input into the second activation unit, and the second fusion unit is used to fuse the output feature of the second activation unit with the input feature of the global feature mining module to obtain the output feature vector Z of the global feature mining module. i : FROM i =GELU(Z″ i )+Z′ i ; Where GELU(·) is the activation function.

2. A 3D convolutional neural network thunderstorm recognition system based on volumetric radar data, characterized in that: The system comprises: The volume radar data set processing module is used to obtain the original radar volume scan data, convert the original radar volume scan data into radar reflectivity images of the corresponding number of layers through linear interpolation, grayscale the radar reflectivity images of all layers, combine the grayscale images together according to the order of the layers and save them as radar volume data files, use the radar volume data files as training data, and construct a volume radar data set; The label processing module is used to obtain the ground-to-ground lightning location data corresponding to the time and region of the original radar volume scan data, and to mark the volume radar data set according to the latitude and longitude coordinates of the ground-to-ground lightning location data as the label of the volume radar data set; A thunderstorm recognition model building module is used to build a thunderstorm recognition model based on a 3D convolutional neural network and train the network weights of the thunderstorm recognition model using a volumetric radar dataset and its labels; The thunderstorm identification model is used to identify the input radar volume scan data to be identified to obtain the probability of thunderstorm occurrence in different areas and generate thunderstorm identification results; The label processing module includes a grid positioning unit, a positioning data screening unit, a positioning data recording unit and a volume marking unit; The grid positioning unit is used to grid the corresponding area; The positioning data screening unit is used to clear isolated ground-to-ground lightning positioning; The positioning data recording unit locates the location where the ground lightning occurs to the corresponding grid according to the latitude and longitude coordinates recorded in the ground lightning positioning data; The volume labeling unit marks the ground-to-ground lightning area with a distribution density greater than a preset density threshold as a high-risk thunderstorm area, and marks the surrounding areas of the high-risk thunderstorm area as medium-risk thunderstorm areas and low-risk thunderstorm areas according to the distance, and completes the generation of labels in combination with the volume radar data set; The thunderstorm recognition model includes an input module, a large kernel feature extraction module, an encoder module, a decoder module and an output module; The input module is for the i-th group of data V in the volume radar data set i , randomly extract a sub-body P of size H×W×D i As the input of the large kernel feature extraction module, the sub-body P i The number of channels C = 1; The large kernel feature extraction module calculates the sub-volume P i The feature vector Z is taken as the input of the encoder module: Z=GN(Conv(P i )); Where Conv(·) represents a 3D convolution operation with a kernel size of 7, a step size of 2, and a padding of 3; GN(·) represents a group normalization operation with a feature vector Z of size The number of channels is expanded to C = 48; The encoder module includes four layers of encoders connected in sequence, and each layer of encoder includes a convolution module and a global feature mining module connected to each other; wherein the calculation formula of the i-th layer encoder is: Z′ i =Conv Block(Z i-1 ); Z i =GFE(Z′ i ); Among them, Z i-1 is the output of the previous encoder, Conv Block (·) is the convolution module, Z′ i is the output result of the convolution module in the i-th layer encoder, GFE(·) represents the output function of the global feature mining module based on the self-attention mechanism and convolution, and Z i is the output of the i-th layer encoder; The decoder module includes five layers of decoders connected in sequence, each layer of decoders concatenates the output of the encoder of the corresponding layer with the output of the decoder of the previous layer as input, and each layer of decoders includes a convolution module; the features output by the last layer of decoders are sequentially passed through a 1×1×1 convolution layer and a softmax function to obtain the final output of the thunderstorm recognition model; The convolution modules each include a first convolution block, a second convolution block, and a downsampling block connected in sequence; the first convolution block and the second convolution block each include a depth convolution unit, a first point-by-point convolution channel, a first group of normalization units, a first activation unit, a second convolution channel, and a first fusion unit connected in sequence; The calculation formula of the first convolution block is: The calculation formula of the second convolution block is: in, and are the outputs of the depthwise convolutional units of the first and second convolutional blocks, respectively. and is the output of the first convolutional block and the first point-wise convolution channel of the second convolutional block, GN(·), GELU(·), DWC(·), DCS(·) represent the group normalization function, GELU activation function, depth convolution function and point-wise convolution function, respectively; The downsampling block outputs the feature map 2 of the second convolution block l+1 Downsample to get the output of the convolution module Among them, the downsampling block converts the feature map 2 into l+1 The length, width and height are reduced to half of the original ones, and the number of channels is doubled; The global feature mining module includes a second group of normalization units, a multi-head self-attention unit, a first convolution unit, a splicing unit, a second convolution unit, a second activation unit and a second fusion unit; The second group normalization unit performs group normalization on the output features of the convolution module, and the group normalization results are respectively input into the multi-head self-attention unit and the first convolution unit to extract the local features Z1 and the global features Z2: Z1=Conv(GN(Z′ i )); Z2=MSA(GN(Z′ i )); Where MSA(·) represents the multi-head self-attention mechanism; After the splicing unit splices the local feature Z1 and the global feature Z2, the splicing unit inputs the splicing unit into the second convolution unit for a 3D convolution operation with a convolution kernel size of 1 to complete the fusion of the local feature Z1 and the global feature Z2, and obtain the fused feature vector Z″ i , the formula is: Z″ i =Conv(Concat(Z1,Z2)); The fused feature vector Z″ i The output feature of the second activation unit is input into the second activation unit, and the second fusion unit is used to fuse the output feature of the second activation unit with the input feature of the global feature mining module to obtain the output feature vector Z of the global feature mining module. i : FROM i =GELU(Z″ i )+Z′ i ; Where GELU(·) is the activation function.

Images