Thunderstorm and gale identification method and system based on attention mechanism convolutional neural network

Abstract

The invention discloses a thunderstorm and gale identification method and system based on an attention mechanism convolutional neural network. The thunderstorm and gale identification method comprises the following steps: 1, generating a sample set according to historical weather radar single station base data and thunderstorm and gale event occurrence actual measurement records, and splitting the sample set into a training set and a verification set; 2, training the recognition model by adopting a training set, verifying the generalization ability of the recognition model by adopting a verification set, and selecting an optimal model parameter according to the model performance; the input of the identification model is a radar reflectivity graph and a radar radial velocity graph of a to-be-identified area at the same time, and the output is a thunderstorm and gale estimation label; and 3, generating a radar reflectivity chart and a radar radial velocity chart according to radar data acquired by a real-time weather radar single station, and inputting the radar reflectivity chart and the radar radial velocity chart into the recognition model to obtain a thunderstorm and gale recognition result. According to the method, for nonlinearity, randomness and burstiness of thunderstorm and gale event space distribution, multi-scale information and position information can be effectively utilized, and the thunderstorm and gale recognition accuracy is improved.

Description

Technical Field

[0001] The present invention relates to a thunderstorm gale recognition method and system based on an attention mechanism convolutional neural network, and belongs to the technical field of weather radar data processing. Background Art

[0002] Thunderstorms are severe storm phenomena caused by thunderstorms, often accompanied by heavy rainfall, hail and lightning, which have a significant impact on human society, infrastructure and the environment. The formation mechanism of thunderstorms is complex, involving atmospheric dynamics, thermodynamics and microphysical processes, so its identification has always been one of the difficult problems in meteorological research. With the development of meteorological observation technology and numerical weather models, the ability of thunderstorms has been significantly improved. However, high-resolution numerical models still have many limitations, and it is difficult to accurately capture the occurrence and intensity of thunderstorms. Therefore, the development of accurate thunderstorm identification technology is of great practical significance.

[0003] Traditional thunderstorm and gale identification methods mainly rely on numerical weather models and statistical methods. These methods have been used in meteorology for many years, and have accumulated rich experience and data, playing an important role in understanding and identifying severe weather phenomena such as thunderstorms and gale. However, these methods also have some limitations, especially when dealing with complex nonlinear weather phenomena, which are not enough to fully capture the details of nonlinear and complex weather phenomena such as thunderstorms and gale. Summary of the invention

[0004] Purpose of the invention: The technical problem to be solved by the present invention is to provide a thunderstorm and gale identification method based on an attention mechanism convolutional neural network in view of the shortcomings of the prior art. The method is highly nonlinear, random and sudden in the spatial distribution of thunderstorm and gale events, can effectively utilize multi-scale information and location information, is suitable for identifying thunderstorm and gale occurrence areas, and can improve the accuracy of thunderstorm and gale identification.

[0005] In order to solve the above technical problems, on the one hand, the present invention discloses a thunderstorm gale recognition method based on an attention mechanism convolutional neural network, comprising:

[0006] Step 1: Generate a sample set based on historical weather radar single-station base data and actual record data of thunderstorm gale events, wherein a sample in the sample set includes a radar reflectivity map, a radar radial velocity map and a real label of the thunderstorm gale in the area to be identified at the same time; split the sample set into a training set and a validation set;

[0007] Step 2: Use the training set to train the recognition model, use the validation set to verify the generalization ability of the recognition model, and select the optimal model parameters according to the model performance; the input of the recognition model is the radar reflectivity map and the radar radial velocity map of the area to be identified at the same time, and the output is the thunderstorm gale estimation label of the area to be identified;

[0008] Step 3: Generate a radar reflectivity map and a radar radial velocity map for the area to be identified based on the radar reflectivity data and radial velocity data of the area to be identified obtained by the real-time weather radar single station, and input them into the recognition model to obtain the thunderstorm gale recognition result of the area to be identified.

[0009] In step 1, the steps of constructing the sample set are as follows:

[0010] S11, gridding the area to be identified according to longitude and latitude to obtain a two-dimensional space grid map;

[0011] S12, obtaining a radar reflectivity map and a radar radial velocity map corresponding to the two-dimensional space grid map at the same time;

[0012] S13. Detect the storm location in radar data, specifically:

[0013] The radar reflectivity image is binarized, the pixel values ​​exceeding the set threshold are set to 255, and the remaining pixel values ​​are set to 0. A 3×3 morphological kernel is created, and the noise is removed by morphological opening operation.

[0014] The processed image is marked with connected areas, and the connected areas with an area smaller than the set number of grid points are filtered out. The remaining connected areas are storm drop areas.

[0015] For each storm location, the longitude and latitude coordinates of the contour points are obtained to form a storm location polygon, thus completing the detection of the storm location based on radar data.

[0016] S14. Obtain the real label of thunderstorm gale according to the measured record of thunderstorm gale event:

[0017] Select the records with wind speed greater than the preset wind speed threshold from the measured records of thunderstorm and gale events, and select the radar data whose time difference with the selected thunderstorm and gale events is within the radar time resolution;

[0018] If the longitude and latitude coordinates of a thunderstorm event are within the storm area outline polygon corresponding to the filtered radar data, the storm area is marked as "with thunderstorms and strong winds"; the storm area that does not contain the longitude and latitude points of the thunderstorm event is marked as "without thunderstorms and strong winds", and the remaining areas are marked as "others"; the three types of areas are distinguished by numerical values ​​to create thunderstorm labels.

[0019] Specifically, the radar reflectivity map and radar radial velocity map in a sample are constructed based on the radar data points obtained in a volume scanning cycle; the radar combined reflectivity is mapped to a two-dimensional space grid according to longitude and latitude to obtain a radar reflectivity map; the effective radial velocity is mapped to a two-dimensional space grid to obtain a radar radial velocity map; the radar combined reflectivity is the maximum reflectivity among the 10 altitude layers corresponding to the same longitude and latitude, that is, the combined reflectivity REF at the longitude and latitude coordinates (lon, lat) C (lon,lat) is:

[0020] REF C (lon,lat)=max(REF n (lon,lat)),n=0,1,2…,9; (1)

[0021] Where REF n (lon,lat) is the reflectivity of the nth altitude layer at the latitude and longitude coordinates (lon,lat);

[0022] The effective radial velocity is composed of the radial velocities of the lowest three layers of the 10 altitude layers corresponding to the same longitude and latitude. The effective radial velocity DOP at the longitude and latitude coordinates (lon, lat) S (lon,lat) is:

[0023] DOP s (lon,lat)=(DOP 0 (lon,lat),DOP 1 (lon,lat),DOP 2 (lon,lat)) (2)

[0024] Among them, DOP m (lon, lat) is the radial velocity of the mth altitude layer at the latitude and longitude coordinates (lon, lat), m = 0, 1, 2, that is, the effective radial velocity DOP S (lon,lat) is a 3D vector;

[0025] Given the longitude and latitude information, the reflectivity and radial velocity at each altitude layer are obtained by interpolation.

[0026] Specifically, the reflectivity and radial velocity at each altitude layer are obtained by trilinear interpolation method, and the specific steps are as follows:

[0027] a. Calculate the coordinates of the two-dimensional space grid map in the spherical coordinate system with the radar station as the origin at each altitude layer based on the latitude, longitude and altitude information; the longitude and latitude coordinates of the i-th small grid in the two-dimensional space grid map are (i lat ,i lon), the azimuth, elevation and distance to the radar at the nth altitude layer are calculated as follows:

[0028] Convert radar station latitude and longitude from degrees to radians:

[0029]

[0030] The spherical distance from the radar station to the target point is calculated using the spherical trigonometry formula. The longitude and latitude coordinates of the target point are (i lat ,i lon ), height hgt n :

[0031] s=Re×arccos(sin(lat r ))×sin(i lat )+cos(lat r )×cos(i χlat )×cos(i lon -lon r )

[0032] According to the relative position of the target point and the radar station, calculate the azimuth a of the target point:

[0033]

[0034] a=arcsin(C)

[0035] Adjust the target point's azimuth numerically:

[0036]

[0037] Calculate the elevation angle e and distance r of the target point:

[0038]

[0039] where lon radar ,lat radar and hgt r They represent the longitude, latitude and altitude of the radar station respectively; Re represents the radius of the earth; Rm represents the equivalent radius of the earth; ∈>0, which means to avoid small zero values.

[0040] b. Get the spherical coordinate index of the target point:

[0041] Azi=Az_index(i lat ,i lon ,hgt n )

[0042] Eli=EL_index(i lat ,i lon ,hgt n)

[0043] Dist = dist_index(i lat ,i lon ,hgt n )

[0044] The spherical coordinate index is a mapping between longitude, latitude and altitude and azimuth, elevation and distance to the radar; Az_index, EL_index and dist_index respectively represent obtaining the azimuth index Azi, the elevation index Eli and the distance index Disti according to the longitude, latitude and altitude values.

[0045] c. Get the spherical coordinates of the radar data points adjacent to the target point:

[0046] e1=El(Eli-1)

[0047] e2=El(Eli+1)

[0048] a1=Az(Azi-1)

[0049] a2=Az(Azi+1)

[0050] r1=Dist(Disti-1)

[0051] r2=Dist(Dist+1)

[0052] El, Az, Dist represent the values ​​of elevation, azimuth and distance obtained according to the elevation index, azimuth index and distance index respectively; the radar data point with e1 as elevation, a1 as azimuth and r1 as distance and the radar data point with e2 as elevation, a2 as azimuth and r2 as distance are two adjacent points of the target point;

[0053] d. Calculate the interpolation weight based on the azimuth, elevation and distance of the target point and adjacent data points:

[0054]

[0055] e. Get the radar variable values ​​corresponding to adjacent data points:

[0056] f1=Z all (Disti,Azi,Eli)

[0057] f2=Z all (Disti+1,Azi,Eli)

[0058] f3=Z all (Disti,Azi+1,Eli)

[0059] f4=Z all(Disti+1,Azi+1,Eli)

[0060] f5=Z all (Disti,Azi,Eli+1)

[0061] f6=Z all (Disti+1,Azi,Eli+1)

[0062] f7=Z all (Disti,Azi+1,Eli+1)

[0063] f8=Z all (Disti+1,Azi+1,Eli+1)

[0064] Z all () is to obtain radar reflectivity or radial velocity according to the spherical coordinate index;

[0065] Use the trilinear interpolation formula to calculate the radar corresponding variable value of the target point:

[0066] ff=we1×(wa1×(wr1×f1+wr2×f2)+wa2×(wr1×f3+wr2×f4))

[0067] +we2×(wa1×(wr1×f5+wr2×f6)+wa2×(wr1×f7

[0068] +wr2×f8))

[0069] The radar corresponding variable is radar reflectivity or radial velocity, which is selected according to the variable to be calculated currently.

[0070] Specifically, the recognition model adopts an attention mechanism and a convolutional neural network, including an encoder 100, a decoder 200 and a compression excitation module 300;

[0071] The encoder 100 includes three encoding layers and three downsampling layers. The input of the first encoding layer En-conv1 is the combined data obtained by splicing the radar reflectivity map and the radial velocity map. The combined data is 4-channel data obtained by splicing the radar reflectivity map and the radar radial velocity map. The output feature map of En-conv1 is The first channel weight s is obtained through the compression excitation module 300 1 , the first channel weight s 1 and After the residual connection, the first downsampling layer performs downsampling, and the feature map after downsampling and After the residual connection, the second encoding layer En-conv2 is input for encoding. The output feature map of En-conv2 The second channel weight s is obtained through the compression excitation module 300 2 , the second channel weight s 2 and After the residual connection, the second downsampling layer performs downsampling, and the feature map after downsampling and After the residual connection, the third encoding layer En-conv3 is input for encoding. The output feature map of En-conv3 After passing through the third downsampling layer, the data is input to the decoder 200 for decoding;

[0072] Each encoding layer performs a 3×3 convolution operation on the input feature map. The first downsampling layer, the second downsampling layer, and the third downsampling layer are all downsampled by a 2×2 maximum pooling operation.

[0073] The decoder 200 includes three decoding layers and three upsampling layers; the input of the third decoding layer De-conv3 is the output of the encoder 100, the output of De-conv3 passes through the third upsampling layer and then enters the second decoding layer De-conv2, the output of De-conv2 and the output of the third upsampling layer are connected through residual connection and then enter the second upsampling layer, the output of the second upsampling layer is input into the first decoding layer De-conv1, and the output of De-conv1 and the output of the second upsampling layer pass through the first upsampling layer to obtain a thunderstorm gale prediction label;

[0074] Each decoding layer performs a 3×3 convolution operation on the input feature map, and the first upsampling layer, the second upsampling layer, and the third upsampling layer are all upsampled by 2×2 deconvolution operations;

[0075] The compression excitation module 300 acts after each encoding layer of the encoder 100, performs global average pooling on each channel, and the output is the channel weight. The channel weight is applied to each channel data of the feature map in the form of residual connection to realize the effect of the attention mechanism; the compression excitation module 300 includes a global average pooling layer, a first fully connected layer, a ReLU activation function, a second fully connected layer and a Sigmoid function cascaded in sequence; wherein the processing of the feature map X input to the compression excitation module (300) by the global average pooling layer is expressed as:

[0076]

[0077] where x ijis the element value of the feature map X at (i, j) of the input compression excitation module 300, H and W are the height and width of the feature map X, i = 1, 2, ..., H; j = 1, 2, ..., W; z is the feature map after global average pooling;

[0078] z learns the correlation between channels through the first fully connected layer, ReLU activation function, the second fully connected layer and Sigmoid function, which is expressed as:

[0079] s=σ(W 2 ReLU(W 1 ·z))

[0080] Where σ represents the sigmoid function, W 1 and W 2 are the weights of the first fully connected layer and the second fully connected layer, respectively. ReLU represents the ReLU activation function. s is the input feature map X channel weight, which represents the correlation between channels.

[0081] The residual connection uses channel weights to recalibrate the original input:

[0082] Y=X·s

[0083] So the input of the first downsampling layer That is, the first channel weight s 1 right Recalibrate to get:

[0084]

[0085] right After downsampling, the feature map is obtained

[0086] The training set is used to train the recognition model, and the training loss function L HL It is composed of the weighted sum of the multi-classification cross entropy loss function and the mean absolute error loss function, which is:

[0087] L HL =αL MCL +L MAE

[0088] Where L MCL is the multi-classification cross entropy loss function; L MAE is the mean absolute error loss function, α is the loss distribution coefficient, where:

[0089]

[0090] Where: N is the number of samples; C is the number of categories; y i,c is the true label of the i-th sample belonging to category c; pi,c is the probability that the i-th sample is predicted to be class c.

[0091]

[0092] Among them, y i is the thunderstorm wind label of the i-th sample, It is the thunderstorm wind forecast label of the i-th sample output by the recognition model.

[0093] On the other hand, the present invention also discloses a recognition system for realizing the thunderstorm gale recognition method based on the attention mechanism convolutional neural network, comprising:

[0094] The sample set generation module 1 is used to generate a sample set based on historical weather radar single-station base data and actual record data of thunderstorm gale events, wherein a sample in the sample set includes a radar reflectivity map, a radar radial velocity map and a real label of the thunderstorm gale in the area to be identified at the same time; and split the sample set into a training set and a validation set;

[0095] The recognition model training module 2 is used to train the recognition model using the training set, verify the generalization ability of the recognition model using the verification set, and select the optimal model parameters according to the model performance; the input of the recognition model is the radar reflectivity map and the radar radial velocity map of the area to be recognized at the same time, and the output is the thunderstorm gale estimation label of the area to be recognized;

[0096] The recognition result acquisition module 3 is used to generate a radar reflectivity map and a radar radial velocity map to be identified based on the radar reflectivity data and radial velocity data of the area to be identified obtained by the real-time weather radar single station, and input them into the recognition model to obtain the thunderstorm gale recognition result of the area to be identified.

[0097] On the other hand, the present invention further discloses a computer-readable storage medium on which computer instructions are stored, and when the computer instructions are executed, the above-mentioned thunderstorm and gale identification method is executed.

[0098] On the other hand, the present invention also discloses an electronic device, including a processor and a storage medium, wherein the storage medium is a computer-readable storage medium; the processor loads and executes instructions and data in the storage medium to implement the above-mentioned thunderstorm and gale identification method.

[0099] Beneficial effects: Compared with the prior art, the thunderstorm and gale recognition and system based on the attention mechanism convolutional neural network provided by the present invention takes into account the mapping relationship between thunderstorm and gale events and radar reflectivity and radial velocity, and introduces the convolution block attention mechanism and compression excitation mechanism in the recognition model, integrates multi-scale information, and learns the correlation between radar reflectivity and radial velocity channels to screen useful feature channels and extract rich and effective features, thereby realizing accurate recognition of the location of thunderstorm and gale events. In addition, the present invention takes into account the uneven distribution of different thunderstorm and gale events in the sample set, and realizes the model weight distribution during the training process by combining the multi-classification cross entropy loss function and the mean absolute error loss function, thereby improving the model's recognition ability for positive samples of thunderstorm and gale events. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0101] Figure 1 It is a flow chart of a thunderstorm gale recognition method based on an attention mechanism convolutional neural network disclosed in the present invention;

[0102] Figure 2 The radar reflectivity map and the radar radial velocity map in the embodiment;

[0103] Figure 3 Schematic diagram of a real label of a thunderstorm gale in the embodiment;

[0104] Figure 4 This is a schematic diagram of the composition of the recognition model;

[0105] Figure 5 It is a schematic diagram of the test results in the embodiment;

[0106] Figure 6 The present invention discloses a schematic diagram of the composition of a thunderstorm and gale recognition system based on an attention mechanism convolutional neural network. DETAILED DESCRIPTION

[0107] This embodiment takes Guizhou Province as an example to illustrate the thunderstorm and gale identification method disclosed in the present invention.

[0108] A thunderstorm and gale recognition method based on attention mechanism convolutional neural network, such as Figure 1 As shown, including:

[0109] Step 1: Generate a sample set based on historical weather radar single-station base data and actual record data of thunderstorm gale events, wherein a sample in the sample set includes a radar reflectivity map, a radar radial velocity map and a real label of the thunderstorm gale in the area to be identified at the same time; split the sample set into a training set and a validation set;

[0110] In this embodiment, radar-based data and measured records of thunderstorm gale events from March to September in Guizhou Province from 2021 to 2023 are obtained, and radar-based data and measured records of thunderstorm gale events with time differences within the radar time resolution range are selected as raw data of a sample to construct a sample set. In this embodiment, the radar time resolution is 6 minutes, and the steps for constructing the sample set are as follows:

[0111] S11, gridding the area to be identified according to longitude and latitude to obtain a two-dimensional space grid map;

[0112] The longitude of Guizhou Province is (103, 110) and the latitude is (24, 30). The range is two-dimensionally gridded. In this embodiment, the resolution of the two-dimensional space grid map is 0.01°×0.01°, that is, the longitude and latitude corresponding to a small grid in the two-dimensional space grid map is 0.01°×0.01°, and the longitude and latitude range of Guizhou Province is divided into 701×601 small grids.

[0113] S12, obtaining a radar reflectivity map and a radar radial velocity map corresponding to the two-dimensional space grid map at the same time;

[0114] Most of the weather radars currently used adopt the volume scanning mode. In a volume scanning cycle, the radar antenna first points to the minimum elevation angle, transmits a pulse and then receives the echo signal; maintains the elevation angle, changes the azimuth angle to transmit and receive again; when the radar completes a 360° scan, the antenna is raised to the next elevation angle to continue the above process until the elevation angle reaches the maximum value. The number of elevation angles in a volume scanning cycle determines the number of antenna lifting heights. The time resolution of the radar single-station base data used in this embodiment is 6 minutes, that is, the volume scanning cycle is 6 minutes; the number of elevation angles is 10, respectively [0, 0.5, 1.45, 2.4, 3.35, 4.3, 6.0, 9.9, 14.6, 19.5], that is, 10 scanning cones can be obtained in one volume scanning cycle. The attributes of the data point p obtained during the scanning process are: azimuth a, elevation e, distance r to the radar station, reflectivity ref and radial velocity dop, expressed as p(t, a, e, r, ref, dop). That is, the coordinates (a, e, r) of the data points acquired by the radar are positions in the spherical coordinate system.

[0115] The radar reflectivity map and radar radial velocity map in one sample are constructed based on the radar data points acquired within one volume scanning cycle.

[0116] The radar combined reflectivity is mapped to a two-dimensional space grid according to longitude and latitude to obtain a radar reflectivity map; the effective radial velocity is mapped to a two-dimensional space grid to obtain a radar radial velocity map.

[0117] The radar combined reflectivity is the maximum reflectivity among the 10 altitude layers corresponding to the same longitude and latitude, that is, the combined reflectivity REF at the longitude and latitude coordinates (lon, lat) C (lon,lat) is:

[0118] REF C (lon,lat)=max(REF n (lon,lat)),n=0,1,2…,9; (1)

[0119] Where REF n (lon, lat) is the reflectivity of the nth altitude layer at the latitude and longitude coordinates (lon, lat). The 10 altitude layers start from 1000 meters above the ground, with a layer spacing of 1000 meters, that is, 1-10 kilometers in height. From bottom to top, they are 0-9 layers, and the height of the nth layer is hgt n It is 1000+1000n, n=0,1,2,……,9. The altitude layer data is obtained by interpolating the data scanned at different elevation angles to the corresponding altitude layer.

[0120] The effective radial velocity is composed of the radial velocities of the lowest three layers of the 10 altitude layers corresponding to the same longitude and latitude. The effective radial velocity DOP at the longitude and latitude coordinates (lon, lat) S (lon,lat) is:

[0121] DOP S (lon,lat)=(DOP 0 (lon,lat),DOP 1 (lon,lat),DOP 2 (lon,lat)) (2)

[0122] Among them, DOP m (lon, lat) is the radial velocity of the mth altitude layer at the latitude and longitude coordinates (lon, lat), m = 0, 1, 2, that is, the effective radial velocity DOP s (lon,lat) is a 3D vector;

[0123] Given the longitude and latitude information, the reflectivity and radial velocity at each altitude layer are obtained by interpolation. The present invention adopts trilinear interpolation method, and the specific steps are as follows:

[0124] a. Calculate the coordinates of the two-dimensional space grid map in the spherical coordinate system with the radar station as the origin at each altitude layer based on the latitude, longitude and altitude information;

[0125] The present invention takes into account the effect of earth curvature, and the longitude and latitude coordinates of the i-th small grid in the two-dimensional space grid map are (i lat ,i lon), its azimuth, elevation and distance to the radar at the nth altitude layer are calculated as follows:

[0126] Convert radar station latitude and longitude from degrees to radians:

[0127]

[0128] The spherical distance from the radar station to the target point is calculated using the spherical trigonometry formula. The longitude and latitude coordinates of the target point are (i lat ,i lon ), height hgt n :

[0129] s=Re×arccos(sin(lat r ))×sin(i lat )+cos(lat r )×cos(i lat )×cos(i lon -lon r )

[0130] According to the relative position of the target point and the radar station, calculate the azimuth a of the target point:

[0131]

[0132] a=arcsin(C)

[0133] Adjust the target point's azimuth numerically:

[0134]

[0135] Calculate the elevation angle e and distance r of the target point:

[0136]

[0137] where lon radar ,lat radar and hgt r They represent the longitude, latitude and altitude of the radar station respectively; Re represents the radius of the earth; Rm represents the equivalent radius of the earth; ∈>0, which means to avoid small zero values.

[0138] b. Get the spherical coordinate index of the target point:

[0139] Azi=Az_index(i lat ,i lon ,hgt n )

[0140] Eli=EL_index(i lat ,i lon ,hgtn )

[0141] Dist = dist_index(i lat ,i lon ,hgt n )

[0142] The spherical coordinate index is a mapping between longitude, latitude and altitude and azimuth, elevation and distance to the radar; Az_index, EL_index and dist_index respectively represent obtaining the azimuth index Azi, the elevation index Eli and the distance index Disti according to the longitude, latitude and altitude values.

[0143] c. Get the spherical coordinates of the radar data points adjacent to the target point:

[0144] e1=El(Eli-1)

[0145] e2=El(Eli+1)

[0146] a1=Az(Azi-1)

[0147] a2=Az(Azi+1)

[0148] r1=Dist(Disti-1)

[0149] r2=Dist(Dist+1)

[0150] El, Az, Dist represent the values ​​of elevation, azimuth and distance obtained according to the elevation index, azimuth index and distance index respectively; the radar data point with e1 as elevation, a1 as azimuth and r1 as distance and the radar data point with e2 as elevation, a2 as azimuth and r2 as distance are two adjacent points of the target point;

[0151] d. Calculate the interpolation weight based on the azimuth, elevation and distance of the target point and adjacent data points:

[0152]

[0153] e. Get the radar variable values ​​corresponding to adjacent data points:

[0154] f1=Z all (Disti,Azi,Eli)

[0155] f2=Z all (Disti+1,Azi,Eli)

[0156] f3=Z all (Disti,Azi+1,Eli)

[0157] f4=Zall (Disti+1,Azi+1,Eli)

[0158] f5=Z all (Disti,Azi,Eli+1)

[0159] f6=Z all (Disti+1,Azi,Eli+1)

[0160] f7=Z all (Disti,Azi+1,Eli+1)

[0161] f8=Z all (Disti+1,Azi+1,Eli+1)

[0162] Z all () is to obtain radar reflectivity or radial velocity according to the spherical coordinate index.

[0163] Use the trilinear interpolation formula to calculate the radar corresponding variable value of the target point:

[0164] ff=we1×(wa1×(wr1×f1+wr2×f2)+wa2×(wr1×f3+wr2×f4))

[0165] +we2×(wa1×(wr1×f5+wr2×f6)+wa2×(wr1×f7

[0166] +wr2×f8))

[0167] The above radar corresponding variable, namely radar reflectivity or radial velocity, is selected according to the variable currently to be calculated.

[0168] Using the interpolated radar reflectivity and radial velocity, the radar combined reflectivity and radar effective radial velocity are obtained according to equations (1) and (2), and then a radar reflectivity map and a radar radial velocity map are generated. For the convenience of display, the pixel values ​​in the radar reflectivity map and the radar radial velocity map are converted into the color values ​​of the image pixels. In this embodiment, the converted radar reflectivity map and radar radial velocity map are respectively as follows: Figure 2 (a) and (b). Each image actually contains 701×601 small grids, each of which is a pixel. In order to make the image clear, the grid lines in the two-dimensional space grid image are not displayed in the figure. Figure 2 The combined reflectivity range in (a) is 0-65dbz, divided into 15 levels at 5dbz intervals, and represented by 15 colors. The radar effective radial velocity is a 3D vector. Figure 2(b) Select the first dimension of the three-dimensional vector, i.e., the radial velocity at the 0th altitude layer, to generate an image. The radial velocity range of the 0th altitude layer is -20-20m / s, divided into 11 levels at every 5m / s, and represented by 11 colors. The positive and negative values ​​of the radial velocity represent the wind direction. A positive value indicates that the wind direction is along the radar transmission direction, and a negative value indicates that the wind direction is opposite to the radar transmission direction.

[0169] S13. Detect the storm location in radar data, specifically:

[0170] The radar reflectivity map is binarized, the pixel values ​​exceeding the set threshold are set to 255, and the remaining pixel values ​​are set to 0, a 3×3 morphological kernel is created, and noise is removed by morphological opening operation; in this embodiment, the pixel value of the radar reflectivity map with a combined reflectivity exceeding 20dbz is set to 255, and the remaining pixel values ​​are set to 0.

[0171] The processed image is marked with connected areas, and the connected areas with an area smaller than the set number of grid points are filtered out. The remaining connected areas are storm drop areas.

[0172] For each storm location, the longitude and latitude coordinates of the contour points are obtained to form a storm location polygon, thus completing the detection of the storm location in the radar data.

[0173] S14, obtaining the real label of the thunderstorm gale according to the measured record of the thunderstorm gale event;

[0174] In the measured records of thunderstorm events, the records with wind speed greater than 17.2m / s are selected, and the radar data with a time difference of less than 6 minutes from the selected thunderstorm events are selected. If the longitude and latitude coordinates of the thunderstorm event are within the storm area contour polygon corresponding to the selected radar data, the storm area is marked as "with thunderstorm" (value marked as 2); the storm area that does not contain the longitude and latitude points of the thunderstorm event is marked as "without thunderstorm" (value marked as 1), and the rest of the area is marked as 0, and the thunderstorm label is made as described above.

[0175] like Figure 3 As shown, Figure 2 (a) and Figure 2 (b) The corresponding thunderstorm and gale real label. In order to show the image clearly, not all grid lines in the two-dimensional space grid map are shown in the figure; the figure actually includes 701×601 small grids. The light gray area in the figure, that is, the area marked with a value of 2, is the area with thunderstorm responses; the medium gray area, that is, the area marked with a value of 1, is the area without thunderstorm and gale.

[0176] In order to verify the effect of the recognition model, this embodiment divides the sample set into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0177] Step 2: Use the training set to train the recognition model, use the validation set to verify the generalization ability of the recognition model, and select the optimal model parameters according to the model performance; the input of the recognition model is the radar reflectivity map and the radar radial velocity map of the area to be identified at the same time, and the output is the thunderstorm gale estimation label of the area to be identified;

[0178] In the present invention, the recognition model adopts the attention mechanism and convolutional neural network, such as Figure 4 As shown, it includes an encoder 100, a decoder 200 and a compression excitation module 300;

[0179] The radar reflectivity map and the radar radial velocity map are spliced ​​to obtain 4-channel 701×601 combined data, that is, the dimension of the combined data after splicing is (4, 701, 601).

[0180] The encoder consists of three encoding layers and three downsampling layers. The input of the first encoding layer En-conv1 is the combined data obtained by splicing the radar reflectivity map and the radial velocity map. The output feature map of En-conv1 is The first channel weight s is obtained through the compression excitation module 300 1 , the first channel weight s 1 and After the residual connection, the first downsampling layer down1() is used for downsampling. The feature map after downsampling and After the residual connection, the second encoding layer En-conv2 is input for encoding. The output feature map of En-conv2 The second channel weight s is obtained through the compression excitation module 300 2 , the second channel weight s 2 and After the residual connection, the second downsampling layer down2() performs downsampling, and the feature map after downsampling and After the residual connection, the third encoding layer En-conv3 is input for encoding. The output feature map of En-conv3 After the third downsampling layer down3(), it is input into the decoder for decoding; in the figure represents residual connection;

[0181] Each encoding layer performs a 3×3 convolution operation on the input feature map, and the three downsampling layers down1(), down2(), and down3() are all implemented by 2×2 maximum pooling operations;

[0182] The decoder includes three decoding layers and three upsampling layers; the input of the third decoding layer De-conv3 is the output of the encoder, the output of De-conv3 passes through the third upsampling layer Up3() and then enters the second decoding layer De-conv2, the output of De-conv2 and the output of the third upsampling layer Up3() are connected through residual connection and then input into the second upsampling layer Up2(), the output of the second upsampling layer Up2() is input into the first decoding layer De-conv1, the output of De-conv1 and the output of the second upsampling layer Up2() pass through the first upsampling layer Up1() to obtain the thunderstorm wind estimation label;

[0183] Each decoding layer performs a 3×3 convolution operation on the input feature map, and the three upsampling layers Up1(), Up2(), and Up3() are all implemented by 2×2 deconvolution operations;

[0184] Residual connections are used between the encoding layers in the encoder and between the decoding layers in the decoder. Residual connections are a means of interaction and fusion of information from feature maps of different scales. The core idea of ​​residual connections is to retain the original input information while learning the residual function between input and output, that is, the difference between input and output, through the nonlinear transformation of the network. This not only helps to retain the original information, so that the deep network will not completely lose the basic information of the input data, but also enhances the stability and robustness of the model; at the same time, residual connections also promote feature enhancement, by combining the original information with the new information, so that the model can simultaneously utilize low-level and high-level features. The compression excitation module 300 is a module for the inter-channel attention mechanism, which acts after each encoding layer of the encoder and performs global average pooling on each channel; its output is the channel weight, and the channel weight is applied to each channel data of the feature map in the form of a residual connection to realize the effect of the attention mechanism; the compression excitation module 300 includes a global average pooling layer, a first fully connected layer, a ReLU activation function, a second fully connected layer and a Sigmoid function that are cascaded in sequence; wherein

[0185] The processing of the feature map X of the input compression excitation module 300 by the global average pooling layer is expressed as:

[0186]

[0187] where x ij is the element value of the feature map X at (i, j) of the input compression excitation module 300, H and W are the height and width of the feature map X, i = 1, 2, ..., H; j = 1, 2, ..., W; z is the feature map after global average pooling;

[0188] z learns the correlation between channels through the first fully connected layer, ReLU activation function, the second fully connected layer and Sigmoid function, which is expressed as:

[0189] s=σ(W 2 ReLU(W 1 ·z))

[0190] Where σ represents the sigmoid function, W 1 and W 2 are the weights of the first fully connected layer and the second fully connected layer, respectively. ReLU represents the ReLU activation function. s is the input feature map X channel weight, which represents the correlation between channels.

[0191] The residual connection uses channel weights to recalibrate the original input:

[0192] Y=X·s

[0193] So the input of the first downsampling layer That is, the first channel weight s 1 right Recalibrate to get:

[0194]

[0195] right After downsampling, the feature map is obtained

[0196] The loss function L of the training in the present invention is HL It is composed of the weighted sum of the multi-classification cross entropy loss function and the mean absolute error loss function, which is:

[0197] L HL =αL MCL +L MAE

[0198] Where L MCL is the multi-classification cross entropy loss function; L MAE is the mean absolute error loss function, α is the loss distribution coefficient, where:

[0199]

[0200] Where: N is the number of samples; C is the number of categories; y i,c is the true label of the i-th sample belonging to category c; p i,c is the probability that the i-th sample is predicted to be class c.

[0201]

[0202] Among them, y i is the thunderstorm wind label of the i-th sample, It is the thunderstorm wind forecast label of the i-th sample output by the recognition model.

[0203] Step 3: Generate a radar reflectivity map and a radar radial velocity map for the area to be identified based on the radar reflectivity data and radial velocity data of the area to be identified obtained by the real-time weather radar single station, and input them into the recognition model to obtain the thunderstorm gale recognition result of the area to be identified.

[0204] In this embodiment, the optimal recognition model obtained after training is used to identify thunderstorm and gale on the radar data in the test set, and compared with the observed data to evaluate the performance of the model. Figure 5 (a)-(c) are schematic diagrams of three test results, where the left image is the real label of the thunderstorm in the test sample, the middle image is the thunderstorm recognition result output by the recognition model, and the right image is the difference between the real label of the thunderstorm and the thunderstorm recognition result. In this difference, the black area represents the area with incorrect recognition, and the gray area represents the area with correct recognition. Figure 5 The three examples in the figure show that the recognition model can make a good judgment on whether a thunderstorm event occurs in the real-time radar data, and can also better identify the area where the thunderstorm occurs.

[0205] In this example, IoU is used to score the thunderstorm and gale recognition model. The scoring results are shown in Table 1. IoU (Intersection over Union) is used to measure the overlap between the predicted bounding box and the true bounding box. It is obtained by calculating the intersection area of ​​the two bounding boxes divided by their union area. Mean IoU measures the overlap between the recognition result predicted by the model for each category of pixels in the image and the true label. It can be expressed as:

[0206]

[0207] Among them, M is the number of categories, TP is the true positive (the number of samples correctly predicted as positive by the model), FP is the false positive (the number of samples incorrectly predicted as positive by the model), and FN is the false negative (the number of samples incorrectly predicted as negative by the model); Mean IoU ranges from 0 to 1, and the higher the value, the better the performance of the model.

[0208] Table 1

[0209]

[0210] The present invention also discloses a thunderstorm and gale recognition system based on an attention mechanism convolutional neural network for implementing the above method, such as Figure 6 As shown, including:

[0211] The sample set generation module 1 is used to generate a sample set based on historical weather radar single-station base data and actual record data of thunderstorm gale events, wherein a sample in the sample set includes a radar reflectivity map, a radar radial velocity map and a real label of the thunderstorm gale in the area to be identified at the same time; and split the sample set into a training set and a validation set;

[0212] The recognition model training module 2 is used to train the recognition model using the training set, verify the generalization ability of the recognition model using the verification set, and select the optimal model parameters according to the model performance; the input of the recognition model is the radar reflectivity map and the radar radial velocity map of the area to be recognized at the same time, and the output is the thunderstorm gale estimation label of the area to be recognized;

[0213] The recognition result acquisition module 3 is used to generate a radar reflectivity map and a radar radial velocity map to be identified based on the radar reflectivity data and radial velocity data of the area to be identified obtained by the real-time weather radar single station, and input them into the recognition model to obtain the thunderstorm gale recognition result of the area to be identified.

[0214] The present invention also discloses a computer-readable storage medium on which computer instructions are stored. When the computer instructions are run, the above-mentioned thunderstorm and gale identification method based on the attention mechanism convolutional neural network is executed.

[0215] The present invention also discloses an electronic device, including a processor and a storage medium, wherein the storage medium is a computer-readable storage medium; the processor loads and executes instructions and data in the storage medium to implement the above-mentioned thunderstorm and gale identification method based on the attention mechanism convolutional neural network.

[0216] The present invention provides a thought and method for thunderstorm gale recognition based on attention mechanism convolutional neural network. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A thunderstorm and gale recognition method based on an attention mechanism convolutional neural network, characterized in that: include: Step 1: Generate a sample set based on historical weather radar single-station base data and actual record data of thunderstorm gale events, wherein a sample in the sample set includes a radar reflectivity map, a radar radial velocity map and a real label of the thunderstorm gale in the area to be identified at the same time; split the sample set into a training set and a validation set; Step 2: Use the training set to train the recognition model, use the validation set to verify the generalization ability of the recognition model, and select the optimal model parameters according to the model performance; the input of the recognition model is the radar reflectivity map and the radar radial velocity map of the area to be identified at the same time, and the output is the thunderstorm gale estimation label of the area to be identified; Step 3: Generate a radar reflectivity map and a radar radial velocity map for the area to be identified based on the radar reflectivity data and radial velocity data of the area to be identified obtained by the real-time weather radar single station, and input them into the recognition model to obtain the thunderstorm gale recognition result of the area to be identified.

2. The thunderstorm gale identification method according to claim 1, characterized in that: In step 1, the steps of constructing the sample set are as follows: S11, gridding the area to be identified according to longitude and latitude to obtain a two-dimensional space grid map; S12, obtaining a radar reflectivity map and a radar radial velocity map corresponding to the two-dimensional space grid map at the same time; S13. Detect the storm location in radar data, specifically: The radar reflectivity image is binarized, the pixel values ​​exceeding the set threshold are set to 255, and the remaining pixel values ​​are set to 0. A 3×3 morphological kernel is created, and the noise is removed by morphological opening operation. The processed image is marked with connected areas, and the connected areas with an area smaller than the set number of grid points are filtered out. The remaining connected areas are storm drop areas. For each storm location, the longitude and latitude coordinates of the contour points are obtained to form a storm location polygon, thus completing the detection of the storm location based on radar data. S14. Obtain the real label of thunderstorm gale according to the measured record of thunderstorm gale event: Select the records with wind speed greater than the preset wind speed threshold from the measured records of thunderstorm and gale events, and select the radar data whose time difference with the selected thunderstorm and gale events is within the radar time resolution; If the longitude and latitude coordinates of the thunderstorm event are within the storm area outline polygon corresponding to the filtered radar data, the storm area is marked as "with thunderstorm"; the storm area that does not contain the longitude and latitude points of the thunderstorm event is marked as "without thunderstorm", and the remaining areas are marked as "others"; the three types of areas are distinguished by numerical values ​​to create thunderstorm labels.

3. The thunderstorm gale identification method according to claim 1, characterized in that: The radar reflectivity map and radar radial velocity map in a sample are constructed based on the radar data points obtained in a volume scanning cycle; the radar combined reflectivity is mapped to a two-dimensional space grid according to longitude and latitude to obtain a radar reflectivity map; the effective radial velocity is mapped to a two-dimensional space grid to obtain a radar radial velocity map; the radar combined reflectivity is the maximum reflectivity among the 10 altitude layers corresponding to the same longitude and latitude, that is, the combined reflectivity REF at the longitude and latitude coordinates (lon, lat). C (lon, lat) is: REF C (lon,lat)=max(REF n (lon,lat)),n=0,1,2...,9; (1) Where REF n (lon, lat) is the reflectivity of the nth altitude layer at the latitude and longitude coordinates (lon, lat); The effective radial velocity is composed of the radial velocities of the lowest three layers of the 10 altitude layers corresponding to the same longitude and latitude. The effective radial velocity DOP at the longitude and latitude coordinates (lon, lat) S (lon, lat) is: DOP S (lon,lat)=(DOP 0 (lon,lat),DOP 1 (lon,lat),DOP 2 (lon,lat)) (2) Among them, DOP m (lon, lat) is the radial velocity of the mth altitude layer at the latitude and longitude coordinates (lon, lat), m = 0, 1, 2, that is, the effective radial velocity DOP S (l℃n, lat) is a 3D vector; Given the longitude and latitude information, the reflectivity and radial velocity at each altitude layer are obtained by interpolation.

4. The thunderstorm gale identification method according to claim 2, characterized in that: In step S13, the radar reflectivity map is binarized, and the threshold is set to 20 dbz, that is, the pixel value of the radar reflectivity map with a combined reflectivity exceeding 20 dbz is set to 255, and the remaining pixel values ​​are set to 0; In step S14, the wind speed threshold is preset to 17.2 m / s.

5. The thunderstorm gale identification method according to claim 3, characterized in that: The reflectivity and radial velocity at each altitude layer are obtained by trilinear interpolation. The specific steps are as follows: a. Calculate the coordinates of the two-dimensional space grid map in the spherical coordinate system with the radar station as the origin at each altitude layer based on the latitude, longitude and altitude information; the longitude and latitude coordinates of the i-th small grid in the two-dimensional space grid map are (i lat ,i lon ), the azimuth, elevation and distance to the radar at the nth altitude layer are calculated as follows: Convert radar station latitude and longitude from degrees to radians: The spherical distance from the radar station to the target point is calculated using the spherical trigonometry formula. The longitude and latitude coordinates of the target point are (i lat ,i lon ), height hgt n : s=Re×arccos(sin(lat r ))×sin(i lat )+cos(lat r )×cos(i iat )×cos(i lon -lon r ) According to the relative position of the target point and the radar station, calculate the azimuth a of the target point: a=arcsin(C) Adjust the target point's azimuth numerically: Calculate the elevation angle e and distance r of the target point: where lon radar ,lat radar and hgt r They represent the longitude, latitude and altitude of the radar station respectively; Re represents the radius of the earth; Rm represents the equivalent radius of the earth; ∈>0, which means to avoid small zero values. b. Get the spherical coordinate index of the target point: Azi=Azi_index(i lat ,I lon ,hgt n ) Eli=EL_index(i lat ,i lon ,hgt n ) Disti=dist_index(i lat ,i lon ,hgt n ) The spherical coordinate index is a mapping between longitude, latitude and altitude and azimuth, elevation and distance to the radar; Az_index, EL_index and dist_index respectively represent obtaining the azimuth index Azi, the elevation index Eli and the distance index Disti according to the longitude, latitude and altitude values. c. Get the spherical coordinates of the radar data points adjacent to the target point: e1=El(Eli-1) e2=El(Eli+1) a1=Az(Azi-1) a2=Az(Azi+1) r1=Dist(Disti-1) r2=Dist(Dist+1) El, Az, Dist represent the values ​​of elevation, azimuth and distance obtained according to the elevation index, azimuth index and distance index respectively; the radar data point with e1 as elevation, a1 as azimuth and r1 as distance and the radar data point with e2 as elevation, a2 as azimuth and r2 as distance are two adjacent points of the target point; d. Calculate the interpolation weight based on the azimuth, elevation and distance of the target point and adjacent data points: e. Get the radar variable values ​​corresponding to adjacent data points: f1=Z all (Disti, Azi, Eli) <h2 style=";text-align:left;direction:ltr">f2 = Z<h2 style=";text-align:left;direction:ltr"> all <h2 style=";text-align:left;direction:ltr"> (Disti+1,Azi,Eli) f3=Z all (Disti,Azi+1,Eli) f4=Z all (Disti+1, Azi+1, Eli) f5=Z all (Disti,Azi,Eli+1) f6=Z all (Disti+1, Azi, Eli+1) f7=Z all (Disti, Azi+1, Eli+1) f8=Z all (Disti+1, Azi+1, Eli+1) Z all () is to obtain radar reflectivity or radial velocity according to the spherical coordinate index; Use the trilinear interpolation formula to calculate the radar corresponding variable value of the target point: ff=we1×(wa1×(wr1×f1+wr2×f2)+wa2×(wr1×f3+wr2×f4))+we2×(wa1×(wr1×f5+wr2×f6)+wa2×(wr1×f7+wr2×f8)) The radar corresponding variable is radar reflectivity or radial velocity, which is selected according to the variable to be calculated currently.

6. The thunderstorm gale identification method according to claim 1, characterized in that: The recognition model adopts an attention mechanism and a convolutional neural network, and includes an encoder (100), a decoder (200) and a compression excitation module (300); The encoder (100) comprises three encoding layers and three downsampling layers, wherein the input of the first encoding layer En-conv1 is the combined data obtained by splicing the radar reflectivity map and the radial velocity map, and the combined data is the 4-channel data obtained by splicing the radar reflectivity map and the radar radial velocity map; the output feature map of En-conv1 is The first channel weight s1 is obtained through the compression excitation module (300), and the first channel weight s1 is After the residual connection, the first downsampling layer performs downsampling, and the feature map after downsampling and After the residual connection, the second encoding layer En-conv2 is input for encoding. The output feature map of En-conv2 The second channel weight s2 is obtained through the compression excitation module (300), and the second channel weight s2 is After the residual connection, the second downsampling layer performs downsampling, and the feature map after downsampling and After the residual connection, the third encoding layer En-conv3 is input for encoding. The output feature map of En-conv3 After passing through the third downsampling layer, the data is input into a decoder (200) for decoding; Each encoding layer performs a 3×3 convolution operation on the input feature map. The first downsampling layer, the second downsampling layer, and the third downsampling layer are all downsampled by a 2×2 maximum pooling operation. The decoder (200) comprises three decoding layers and three upsampling layers; the input of the third decoding layer De-conv3 is the output of the encoder (100); the output of De-conv3 is input into the second decoding layer De-conv2 after passing through the third upsampling layer; the output of De-conv2 and the output of the third upsampling layer are input into the second upsampling layer after being residually connected; the output of the second upsampling layer is input into the first decoding layer De-conv1; the output of De-conv1 and the output of the second upsampling layer are input into the thunderstorm gale estimation label after passing through the first upsampling layer; Each decoding layer performs a 3×3 convolution operation on the input feature map, and the first upsampling layer, the second upsampling layer, and the third upsampling layer are all upsampled by 2×2 deconvolution operations; The compression excitation module (300) acts after each encoding layer of the encoder (100), performs global average pooling on each channel, and the output is the channel weight. The channel weight is applied to each channel data of the feature map in the form of residual connection to realize the effect of the attention mechanism; The compression excitation module (300) comprises a global average pooling layer, a first fully connected layer, a ReLU activation function, a second fully connected layer and a Sigmoid function which are cascaded in sequence; wherein the processing of the feature map X input to the compression excitation module (300) by the global average pooling layer is expressed as follows: where x ij is the element value of the feature map X at (i, j) of the input compression excitation module (300), H and W are the height and width of the feature map X, i = 1, 2, ..., H; j = 1, 2, ..., W; z is the feature map after global average pooling; z learns the correlation between channels through the first fully connected layer, ReLU activation function, the second fully connected layer and Sigmoid function, which is expressed as: s=σ(W2·ReLU(W1·z)) Where σ represents the sigmoid function, W1 and W2 are the weights of the first fully connected layer and the second fully connected layer respectively, ReLU represents the ReLU activation function; s is the channel weight of the input feature map X, which represents the correlation between channels; The residual connection uses channel weights to recalibrate the original input: Y=X·s So the input of the first downsampling layer That is, the first channel weight s1 is Recalibrate to get: right After downsampling, the feature map is obtained 7. The thunderstorm gale identification method according to claim 1, characterized in that: The training set is used to train the recognition model, and the training loss function L HL It is composed of the weighted sum of the multi-classification cross entropy loss function and the mean absolute error loss function, which is: L HL =αL MCL +L MAE Where L MCL is the multi-classification cross entropy loss function; L MAE is the mean absolute error loss function, α is the loss distribution coefficient, where: Where: N is the number of samples; C is the number of categories; y i,c is the true label of the i-th sample belonging to category c; p i,c is the probability that the i-th sample is predicted to be class c. Among them, y i is the thunderstorm wind label of the i-th sample, It is the thunderstorm wind forecast label of the i-th sample output by the recognition model.

8. A thunderstorm and gale recognition system based on an attention mechanism convolutional neural network, characterized in that: include: A sample set generation module (1) is used to generate a sample set based on historical weather radar single-station base data and actual recorded data of thunderstorm gale events, wherein a sample in the sample set includes a radar reflectivity map, a radar radial velocity map and a real label of the thunderstorm gale in the area to be identified at the same time; and split the sample set into a training set and a validation set; The recognition model training module (2) is used to train the recognition model using a training set, verify the generalization ability of the recognition model using a verification set, and select optimal model parameters according to the model performance; the input of the recognition model is a radar reflectivity map and a radar radial velocity map of the area to be recognized at the same time, and the output is a thunderstorm gale prediction label of the area to be recognized; The recognition result acquisition module (3) is used to generate a radar reflectivity map and a radar radial velocity map to be identified based on the radar reflectivity data and radial velocity data of the area to be identified obtained by a single real-time weather radar station, and input them into a recognition model to obtain a thunderstorm gale recognition result of the area to be identified.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed, the thunderstorm and gale identification method according to any one of claims 1 to 7 is executed.

10. An electronic device, characterized in that: It includes a processor and a storage medium, wherein the storage medium is the computer-readable storage medium as described in claim 9; the processor loads and executes the instructions and data in the storage medium to implement the thunderstorm and gale identification method as described in any one of claims 1-7.

Images