Geostationary satellite thunderstorm cloud cluster dynamic monitoring and identification method based on long and short memory network

Abstract

The invention discloses a static satellite thunderstorm cloud cluster dynamic monitoring and identification method based on a long short term memory network, which comprises the following steps: constructing a thunderstorm cloud cluster dynamic monitoring and identification model based on a long short term memory network model, and finishing training by adopting a dynamic monitoring machine learning data set; the dynamic monitoring machine learning data set comprises a plurality of dynamic monitoring machine learning arrays, and each dynamic monitoring machine learning array comprises a time-space matched historical mesoscale convective cloud cluster characteristic parameter group and lightning activity generation label data; the input of the trained thunderstorm cloud cluster dynamic monitoring identification model is physical characteristic parameters corresponding to all the mesoscale convective cloud clusters identified in the target area at m continuous moments before the current moment, and the output of the trained thunderstorm cloud cluster dynamic monitoring identification model is a thunderstorm cloud cluster identification result at the current moment. Therefore, by dynamically monitoring the characteristic parameters of the mesoscale convective cloud cluster, high-precision identification of offshore and open-sea thunderstorm cloud clusters can be realized.

Description

Technical Field

[0001] The invention belongs to the field of real-time monitoring and identification of cloud characteristics, and in particular to a method for dynamic monitoring and identification of thunderstorm clouds by a geostationary satellite based on a long short-term memory network. Background Art

[0002] Mesoscale convective systems are often accompanied by disastrous weather such as strong winds, hail, and showers, which often cause huge direct and indirect losses to my country's transportation, power grid, agriculture, forestry, and fisheries. Improving the ability to monitor and predict mesoscale convective systems is of great significance to comprehensively improve my country's natural disaster prevention level and reduce economic and property losses caused by extreme weather disasters.

[0003] The spatial resolution of geostationary meteorological satellites reaches 0.5~4 km, and the time resolution of full-disk observations reaches 15 minutes. They have the ability to monitor all-weather and all-day. Combined with ground-based lightning observations, they can provide strong support for the real-time monitoring and dynamic prediction of mesoscale convective systems in my country and its offshore areas. However, in the past, the satellite-based dynamic monitoring method of mesoscale convective systems mainly adopted the traditional averaging method to obtain the average brightness temperature information of the infrared channel of the mesoscale convective system. However, due to the inhomogeneity of the pixel information inside the mesoscale convective cloud cluster itself, it is extremely easy to be affected by the maximum or minimum value, which makes it difficult to reflect the development degree and disaster-causing nature of the mesoscale convective system. If a correction factor reflecting the height correlation with the development within the cloud is introduced, the role of cloud system characteristics with higher development in the mesoscale convective system will be highlighted, and the role of cloud system characteristics with lower development in the mesoscale convective system will be weakened, which will help to further improve the ability to monitor and predict the state of mesoscale convective clouds in real time. In addition, in the dynamic monitoring of mesoscale convective clouds, the macroscopic and microscopic parameters of clouds such as morphology, area and degree of development of convective cores have not been considered in the past during the occurrence and development of the mesoscale convective system itself. From a physical point of view, these parameters are of great indicative significance for the occurrence and development of the mesoscale convective system. Summary of the invention

[0004] Based on the above reasons, the present invention proposes a method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks. The method constructs a dynamic monitoring and identification model of thunderstorm clouds based on the long short-term memory network model, and uses a dynamic monitoring machine learning data set to train the constructed dynamic monitoring and identification model of thunderstorm clouds, so that the trained dynamic monitoring and identification model of thunderstorm clouds can perform good dynamic monitoring of the evolution of characteristic parameters of mesoscale convective clouds, thereby realizing more accurate dynamic monitoring of mesoscale convective systems in my country's near and far seas.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks comprises the following steps:

[0007] A thunderstorm cloud dynamic monitoring and recognition model was built based on the long short-term memory network model, and the dynamic monitoring machine learning dataset was used to complete the training;

[0008] The dynamic monitoring machine learning dataset includes several groups of dynamic monitoring machine learning arrays, each of which includes a group of historical mesoscale convective cloud cluster characteristic parameters matched in time and space and lightning activity occurrence label data;

[0009] The historical mesoscale convective cloud cluster characteristic parameter group is formed by statistically analyzing the physical characteristic parameters of the mesoscale convective cloud cluster at m consecutive moments before any historical forecast time Tm+1.

[0010] The lightning activity occurrence label data is obtained through the lightning activity observed by the lightning location system at the location of the corresponding mesoscale convective cloud cluster at the historical forecast time Tm+1;

[0011] The input of the trained thunderstorm cloud dynamic monitoring and recognition model is all the mesoscale convective clouds identified in the target area at the current time t m+1 The physical characteristic parameters corresponding to the first m consecutive moments are output as the current moment t m+1 Thunderstorm cloud identification results.

[0012] Preferably, the physical characteristic parameters of the mesoscale convective cloud at any time include cloud macroscopic parameters, cloud microscopic parameters A and improved cloud microscopic parameters B; wherein:

[0013] The cloud cluster macroscopic parameters include the central longitude and latitude of the mesoscale convective cloud cluster, the major axis length, minor axis length, and rotation angle of the circumscribed ellipse;

[0014] The cloud micro-parameter A and improved cloud micro-parameter B are calculated based on the brightness temperature data detected by the infrared imager carried on the geostationary satellite in the target area; among them:

[0015] For any mesoscale convective cloud cluster j, the elements of the cloud cluster microscopic parameter A correspond to: the mean value of the brightness temperature of the 10.4 μm channel , minimum value and the highest value , 6.9 micron and 7.3 micron channel brightness temperature difference Mean of TB1 , minimum value Mi and the maximum value Max , 7.3 μm and 10.4 μm channel brightness temperature difference Mean of TB2 , minimum value Mi and the maximum value Max , 12.4 micron and 10.4 micron channel brightness temperature difference Mean of TB3 , minimum value Mi and the maximum value Max ;

[0016] Any element in the improved cloud micro-parameter B is obtained by optimizing the mean element in the cloud micro-parameter A using the infrared channel brightness temperature correction factor;

[0017] For any mesoscale convective cloud cluster j, the elements of the cloud cluster microscopic parameter B correspond to: the mean value of the brightness temperature of the improved 10.4 μm channel ; Improved average of the brightness temperature difference between the 6.9 μm and 7.3 μm channels ; Improved average of the brightness temperature difference between the 7.3 μm and 10.4 μm channels ; Improved average of the brightness temperature difference between the 12.4 micron and 10.4 micron channels .

[0018] Preferably, the infrared channel brightness temperature correction factor is calculated by the following formula:

[0019] ;

[0020] Where: represents the brightness temperature of the 10.4 μm channel at pixel i, Represents the infrared channel brightness temperature correction factor at pixel i;

[0021] Among the improved cloud microscopic parameters B, any element Calculated by the following formula:

[0022] ;

[0023] Where: represents the first of the cloud microscopic parameters B elements, ; Indicates the average brightness temperature of the 10.4 micron channel ; Represents the average value of the brightness temperature difference between the 6.9 μm and 7.3 μm channels ; Represents the average value of the brightness temperature difference between the 7.3 μm and 10.4 μm channels ; Represents the average value of the brightness temperature difference between the 12.4 μm and 10.4 μm channels ; N represents the number of all pixels in any mesoscale convective cloud cluster j.

[0024] Preferably, for any mesoscale convective cloud cluster j, according to the contour points of the mesoscale convective cloud cluster j, based on the least squares method, the distance between the circumscribed ellipse of the mesoscale convective cloud cluster j and the contour points is minimized, so that the major axis length, minor axis length, and rotation angle of the circumscribed ellipse of any mesoscale convective cloud cluster j can be obtained;

[0025] For any mesoscale convective cloud cluster j, the mean brightness temperature of the 10.4 μm channel in the mesoscale convective cloud cluster j can be calculated based on the brightness temperature of the 10.4 μm channel of all pixels in the mesoscale convective cloud cluster j: , minimum value and the highest value ;

[0026] For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 6.9 μm and 7.3 μm channels of all pixels in the mesoscale convective cloud cluster j. TB1, and then statistically obtain the brightness temperature difference between the 6.9 μm and 7.3 μm channels in the mesoscale convective cloud cluster j Mean of TB1 ,lowest and the highest value ;

[0027] For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 7.3 μm and 10.4 μm channels of all pixels in the mesoscale convective cloud cluster j. TB2, and then statistically obtain the brightness temperature difference between the 7.3 μm and 10.4 μm channels in the mesoscale convective cloud cluster j Mean of TB2 ,lowest and the highest value ;

[0028] For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 12.4 μm and 10.4 μm channels of all pixels in the mesoscale convective cloud cluster j. TB3, and then statistically obtain the brightness temperature difference between the 12.4 μm and 10.4 μm channels in the mesoscale convective cloud cluster j TB3 ,lowest and the highest value .

[0029] Preferably, the dynamic monitoring machine learning dataset is constructed by the following steps:

[0030] Step 1: Identification of mesoscale convective clouds:

[0031] Obtain the brightness temperature data of the target area detected by the infrared imager on board the geostationary satellite during the historical period, and identify the continuous low infrared brightness temperature area detected at the initial time of the historical period as a mesoscale convective cloud cluster;

[0032] Step 2: Calculate the characteristic parameters of mesoscale convective clouds:

[0033] According to the physical connotation of each mesoscale convective cloud cluster identified in step 1, the physical characteristic parameters of the corresponding mesoscale convective cloud cluster are counted;

[0034] Step 3: Tracking of mesoscale convective cloud clusters:

[0035] Using a mesoscale convective system tracking algorithm, the motion trajectory of each mesoscale convective cloud group identified in step 1 is determined, so as to dynamically monitor each mesoscale convective cloud group identified in step 1;

[0036] Step 4: Dynamically monitor machine learning data set construction:

[0037] By combining the physical characteristic parameters of any mesoscale convective cloud cluster counted in step two and the motion trajectory of the corresponding mesoscale convective cloud cluster tracked in step three, and using the lightning activity information observed by the lightning location system, a dynamic monitoring machine learning dataset can be constructed.

[0038] Preferably, the mesoscale convective cloud cluster is identified based on the 10.4 micron channel brightness temperature data detected by an infrared imager in the target area, specifically comprising the following steps:

[0039] Step 1.1: In the target area, obtain the brightness temperature of the 10.4 μm channel observed by the infrared imager at any pixel i. ;

[0040] Step 1.2: Compare the brightness temperature of the 10.4 μm channel and the preset brightness temperature threshold Tb0 to determine whether the corresponding pixel i belongs to a mesoscale convective cloud cluster pixel: When ≥Tb0, pixel i is identified as a mesoscale convective cloud pixel, otherwise it is a non-mesoscale convective cloud pixel; i is a positive integer;

[0041] Step 1.3, traverse the pixels surrounding the mesoscale convective cloud cluster pixel i. If there are mesoscale convective cloud cluster pixels around the mesoscale convective cloud cluster pixel i, continue traversing until the surrounding adjacent pixels of the mesoscale convective cloud cluster pixel i are all non-mesoscale convective cloud cluster pixels. In this way, a complete mesoscale convective cloud cluster identification is completed, and the identified mesoscale convective cloud clusters are numbered and recorded as mesoscale convective cloud cluster j, where j is a positive integer.

[0042] Preferably, the mesoscale convective cloud cluster trajectory tracking in step 3 specifically includes:

[0043] Step 3.1: Based on the mesoscale convective cloud cluster MCC at time Tk Tk The center position information is set, and the search radius R is preset. At time Tk+1, the mesoscale convective cloud cluster search is performed within the search radius R. The number of mesoscale convective cloud clusters whose center positions fall within the search radius R is recorded as N, and each mesoscale convective cloud cluster at time Tk+1 is recorded as MCC. 1 Tk+1 , MCC 2 Tk+1 , ..., MCC N Tk+1 , N is a positive integer;

[0044] Step 3.2: If N = 1, then the only mesoscale convective cloud cluster MCC 1 Tk+1 The center position is recorded as the mesoscale convective cloud cluster MCC Tk At the time Tk+1; if N ≥ 2, then the mesoscale convective cloud cluster MCC at the time Tk is the closest to the mesoscale convective cloud cluster MCC at the time Tk. Tk The center position of the nearest mesoscale convective cloud cluster is recorded as the mesoscale convective cloud cluster MCC at time Tk. Tk The predicted position at time Tk+1;

[0045] Step 3.3: For the subsequent moments Tk+2, Tk+3, …, Tk+n, the mesoscale convective cloud cluster trajectory search is performed according to steps 3.1 and 3.2 until no mesoscale convective cloud cluster is found within the search radius R. Then the mesoscale convective cloud cluster MCC is Tk The trajectory search is finished.

[0046] Preferably, the construction of the dynamic monitoring machine learning data set in step 4 specifically includes:

[0047] Step 4.1, using all the mesoscale convective cloud cluster trajectories tracked in step 3, for any mesoscale convective cloud cluster trajectory, according to the lightning location and lightning occurrence time observed by the lightning location system, count the lightning activities of the mesoscale convective cloud cluster at different times, and mark the occurrence of lightning as "1" and the absence of lightning as "0";

[0048] Step 4.2: Using the physical characteristic parameters of the mesoscale convective cloud clusters obtained in step 2, for any mesoscale convective cloud cluster trajectory that lasts for more than m+1 time periods, the physical characteristic parameters of the mesoscale convective cloud clusters from T1 to Tm consecutive moments and the lightning activity information at Tm+1 time period are used as a set of machine learning data;

[0049] Step 4.3: Take the physical characteristic parameters of the mesoscale convective cloud cluster from T2 to Tm+1 and the lightning activity information at Tm+2 as a set of machine learning data, and so on, until the trajectory of the mesoscale convective cloud cluster terminates, and finally construct a dynamic monitoring machine learning data set.

[0050] Preferably, the training of the thunderstorm cloud dynamic monitoring and identification model specifically includes the following steps:

[0051] Step 5.1: Divide the dynamic monitoring machine learning dataset constructed in step 4 into three parts: training set, test set and validation set according to the ratio of 6:2:2;

[0052] Step 5.2: Using the training set and the test set, adopting the long short-term memory network model, designing different mesoscale convective cloud cluster feature parameter inputs, and using the binary classification BCEwithLogitsLoss loss function, test the optimal input feature parameter combination of the thunderstorm cloud cluster dynamic monitoring and recognition model;

[0053] Step 5.3: Combine the optimal input feature parameters obtained in step 5.2, use the training set and the test set, adopt the long short-term memory network model, design multiple sets of network hyperparameters, and use the binary classification BCEwithLogitsLoss loss function to optimize the network hyperparameters of the thunderstorm cloud dynamic monitoring and identification model; for the validation set, use the detection rate and false alarm rate indicators to evaluate the accuracy of the output of the thunderstorm cloud dynamic monitoring and identification model.

[0054] Preferably, the BCEwithLogitsLoss loss function formula in step 5.2 and step 5.3 is as follows:

[0055]

[0056]

[0057] in, is the input historical mesoscale convective cloud cluster characteristic parameter group, is the sample label, is the predicted value; Represents the Sigmoid activation function;

[0058] In steps 5.2 and 5.3, the calculation formula for the output of the thunderstorm cloud dynamic monitoring and identification model is as follows:

[0059]

[0060] Among them, out is the model output, net represents the thunderstorm cloud dynamic monitoring and identification model, is the Sigmoid activation function.

[0061] According to the above technical solution, compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks provided by the present invention, on the basis of constructing a dynamic monitoring and identification model of thunderstorm clouds based on long short-term memory networks, uses characteristic parameters of mesoscale convective clouds as model inputs, which is more physically interpretable than the previous method of directly using multi-channel observations of geostationary satellites as model inputs.

[0063] 2. By testing multiple groups of mesoscale convective cloud characteristic parameter inputs, the present invention proves that simultaneously introducing cloud macro-parameters, micro-parameters A, and improved micro-parameters B as model inputs can achieve higher-precision thunderstorm cloud identification results, which has a beneficial effect on the dynamic monitoring of thunderstorms in nearshore and offshore areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic flow chart of a method for dynamic monitoring and identification of thunderstorm clouds using a geostationary satellite based on a long short-term memory network provided in an embodiment of the present invention.

[0065] Figure 2 A schematic diagram of the trajectory of a mesoscale convective cloud cluster (white area) identified in the example of the present invention, filled with the brightness temperature of the 10.4 μm channel, the black star represents the center of the mesoscale convective cloud cluster in the example, and the white number represents the number of the mesoscale convective cloud cluster. In the figure: (a) indicates the position of the mesoscale convective cloud cluster at 10:50 (UTC) on July 3, 2018, (b) indicates the position of the mesoscale convective cloud cluster at 11:10 on July 3, 2018, (c) indicates the position of the mesoscale convective cloud cluster at 11:30 on July 3, 2018, (d) indicates the position of the mesoscale convective cloud cluster at 11:50 on July 3, 2018, (e) indicates the position of the mesoscale convective cloud cluster at 12:10 on July 3, 2018, (f) indicates the position of the mesoscale convective cloud cluster at 12:30 on July 3, 2018, (g) indicates the position of the mesoscale convective cloud cluster at 12:50 on July 3, 2018.

[0066] Figure 3The cost function convergence curve (a) of the training set of the thunderstorm cloud dynamic monitoring and identification model of the long short-term memory network, and the detection rate (b) and false alarm rate (c) of the validation set when five groups of historical mesoscale convective cloud characteristic parameter groups are input as examples of the present invention.

[0067] Figure 4 The cost function convergence curves of the training set and test set of the thunderstorm cloud dynamic monitoring and identification model of the long short-term memory network cited as an example of the present invention, and the detection rate and false alarm rate of the verification set. DETAILED DESCRIPTION

[0068] The present invention will be further described in detail below in conjunction with the accompanying drawings in the embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0069] The present invention utilizes the infrared imaging instrument of the Sunflower 8 geostationary satellite to construct a historical mesoscale convective cloud cluster characteristic parameter group. On this basis, a thunderstorm cloud cluster dynamic monitoring and identification model is constructed based on a long short-term memory network, and combined with the time-space matching lightning activity occurrence label data, the characteristic history and future of the geostationary satellite mesoscale convective cloud cluster is dynamically monitored to achieve the purpose of identifying thunderstorm clouds, thereby making a more accurate prediction of the occurrence of thunderstorm events. The embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0070] The present invention discloses a method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellite based on long short-term memory network. Figure 1 As shown, a thunderstorm cloud dynamic monitoring and recognition model is constructed based on the long short-term memory network model, and the dynamic monitoring machine learning data set is used to complete the training;

[0071] The dynamic monitoring machine learning dataset includes several groups of dynamic monitoring machine learning arrays. Each group of dynamic monitoring machine learning arrays includes a temporally and spatially matched historical mesoscale convective cloud cluster characteristic parameter group and lightning activity occurrence label data.

[0072] The historical mesoscale convective cloud cluster characteristic parameter group is formed by counting the physical characteristic parameters of the mesoscale convective cloud cluster at m consecutive moments before any historical forecast time Tm+1; m is a constant and can be 2. The lightning activity occurrence label data is obtained by the lightning activity observed by the lightning location system at the location of the corresponding mesoscale convective cloud cluster at the historical forecast time Tm+1. It can be seen that the mesoscale convective cloud cluster used to construct the dynamic monitoring machine learning dataset will last at least m+1 times before it disappears.

[0073] The input of the trained thunderstorm cloud dynamic monitoring and recognition model is all the mesoscale convective clouds identified in the target area at the current time t m+1The physical characteristic parameters corresponding to the first m consecutive moments are output as the current moment t m+1 Thunderstorm cloud identification results.

[0074] like Figure 1 As shown, the dynamic monitoring machine learning data set described in the present invention has the following specific steps:

[0075] Step 1: Identification of mesoscale convective clouds:

[0076] The continuous low infrared brightness temperature area detected within the scanning range of the geostationary satellite infrared imager in the specified historical period is identified as a mesoscale convective cloud cluster. In this step, the identification of the low infrared brightness temperature area is based on the brightness temperature detected by the 10.4 micron channel of the sunflower geostationary satellite infrared imager. The specific method is as follows:

[0077] Step 1.1: Within the scanning range of the Sunflower satellite infrared imager, the infrared imager observes the corresponding 10.4 micron channel brightness temperature at any pixel i within the specified historical period. ;

[0078] Step 1.2: Compare the brightness temperature of the 10.4 μm channel and the preset brightness temperature threshold Tb0 (in this embodiment, the preset brightness temperature threshold can be set to Tb0 = 241K) to determine whether the corresponding pixel i belongs to a mesoscale convective cloud cluster pixel: When ≥Tb0, pixel i is identified as a mesoscale convective cloud pixel, otherwise it is a non-mesoscale convective cloud pixel; i is a positive integer;

[0079] Step 1.3, traverse the pixels around the mesoscale convective cloud cluster pixel i. If there are mesoscale convective cloud cluster pixels around the mesoscale convective cloud cluster pixel i, continue traversing until the surrounding adjacent pixels of the mesoscale convective cloud cluster pixel i are all non-mesoscale convective cloud cluster pixels. Then, a complete mesoscale convective cloud cluster identification is completed, and the identified mesoscale convective cloud cluster is numbered and recorded as mesoscale convective cloud cluster j, where j is a positive integer. It can be seen that the obtained mesoscale convective cloud cluster j is composed of continuous mesoscale convective cloud cluster pixels i.

[0080] Step 2: Statistical physical characteristic parameters of mesoscale convective clouds:

[0081] According to the physical connotation of each mesoscale convective cloud cluster identified in step 1, the physical characteristic parameters of the corresponding mesoscale convective cloud cluster are counted; and the infrared channel brightness temperature correction factor is introduced to improve the physical characteristic parameters of the mesoscale convective cloud cluster;

[0082] In the present invention, the physical characteristic parameters of the mesoscale convective cloud cluster include cloud cluster microscopic parameters, improved cloud cluster microscopic parameters and cloud cluster macroscopic parameters, wherein: the cloud cluster macroscopic parameters include the central longitude and latitude of the mesoscale convective cloud cluster, the major axis length, the minor axis length and the rotation angle of the circumscribed ellipse;

[0083] For any mesoscale convective cloud cluster j, the elements of the cloud cluster microscopic parameter A correspond to: the mean value of the brightness temperature of the 10.4 μm channel , minimum value and the highest value , 6.9 micron and 7.3 micron channel brightness temperature difference Mean of TB1 , minimum value Mi and the maximum value Max , 7.3 μm and 10.4 μm channel brightness temperature difference Mean of TB2 , minimum value Mi and the maximum value Max , 12.4 micron and 10.4 micron channel brightness temperature difference Mean of TB3 , minimum value Mi and the maximum value Max ;

[0084] For any mesoscale convective cloud cluster j, the elements of the cloud cluster microscopic parameter B correspond to: the mean value of the brightness temperature of the improved 10.4 μm channel ; Improved average of the brightness temperature difference between the 6.9 μm and 7.3 μm channels ; Improved average of the brightness temperature difference between the 7.3 μm and 10.4 μm channels ; Improved average of the brightness temperature difference between the 12.4 micron and 10.4 micron channels .

[0085] In the specific implementation, for any mesoscale convective cloud cluster j, the average values ​​of the longitude and latitude values ​​of all pixels in the mesoscale convective cloud cluster j are calculated to obtain the central longitude and latitude information of the mesoscale convective cloud cluster;

[0086] For any mesoscale convective cloud cluster j, according to the contour points of the mesoscale convective cloud cluster j, based on the least squares method, the distance between the circumscribed ellipse of the mesoscale convective cloud cluster j and the contour points is minimized, and the major axis length, minor axis length, and rotation angle of the circumscribed ellipse of any mesoscale convective cloud cluster j can be obtained;

[0087] For any mesoscale convective cloud cluster j, the mean brightness temperature of the 10.4 μm channel in the mesoscale convective cloud cluster j can be calculated based on the brightness temperature of the 10.4 μm channel of all pixels in the mesoscale convective cloud cluster j: , minimum value and the highest value , which is calculated by the following formula:

[0088] ;

[0089] ;

[0090] ;

[0091] in, represents the brightness temperature of the 10.4 μm channel at pixel i, Represents the number of all pixels within any mesoscale convective cloud cluster j.

[0092] For any mesoscale convective cloud cluster j, the improved mean brightness temperature of the 10.4 μm channel is , calculated according to the following formula:

[0093] ;

[0094] in, represents the brightness temperature of the 10.4 μm channel at pixel i, represents the number of all pixels in any mesoscale convective cloud cluster j, represents the correction coefficient at pixel i, which is calculated by the following formula:

[0095] .

[0096] In addition, in practical applications, if the brightness temperature When it is higher, such as When the correction coefficient is calculated, it is not appropriate to use the above formula , then the correction coefficient The value of is directly 0.

[0097] Based on the above correction factor The calculation formula is: if the brightness temperature =210K, =1, if the brightness temperature Below 210K, Can be greater than 1.

[0098] For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 6.9 μm and 7.3 μm channels of all pixels in the mesoscale convective cloud cluster j. TB1, and then statistically obtain the brightness temperature difference between the 6.9 μm and 7.3 μm channels in the mesoscale convective cloud cluster j Mean of TB1 ,lowest and the highest value ; Calculated by the following formula:

[0099] ;

[0100] Mi ;

[0101] ;

[0102] Where: represents the brightness temperature difference between the 6.9 μm and 7.3 μm channels at pixel i, Represents the number of all pixels within any mesoscale convective cloud cluster j.

[0103] For any mesoscale convective cloud cluster j, the improved mean of the brightness temperature difference between the 6.9 μm and 7.3 μm channels is , calculated according to the following formula:

[0104] = ;

[0105] ;

[0106] in, represents the brightness temperature difference between the 6.9 μm and 7.3 μm channels at pixel i, represents the correction coefficient at position i, represents the brightness temperature of the 10.4 μm channel at pixel i, Represents the number of all pixels within any mesoscale convective cloud cluster j.

[0107] For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 7.3 μm and 10.4 μm channels of all pixels in the mesoscale convective cloud cluster j. TB2, and then statistically obtain the brightness temperature difference between the 7.3 μm and 10.4 μm channels in the mesoscale convective cloud cluster j Mean of TB2 ,lowest and the highest value ; Calculated by the following formula:

[0108] ;

[0109] ;

[0110] ;

[0111] For any mesoscale convective cloud cluster j, the improved mean of the brightness temperature difference between the 7.3 μm and 10.4 μm channels is , calculated according to the following formula:

[0112] = ;

[0113] ;

[0114] in, represents the brightness temperature difference between the 7.3 μm and 10.4 μm channels at pixel i, represents the correction coefficient at pixel i, represents the brightness temperature of the 10.4 μm channel at pixel i, Represents the number of all pixels within any mesoscale convective cloud cluster j.

[0115] For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 12.4 μm and 10.4 μm channels of all pixels in the mesoscale convective cloud cluster j. TB3, and then statistically obtain the brightness temperature difference between the 12.4 μm and 10.4 μm channels in the mesoscale convective cloud cluster j Mean of TB3 ,lowest and the highest value ; Calculated by the following formula:

[0116] ;

[0117] ;

[0118] ;

[0119] For any mesoscale convective cloud cluster j, the improved mean of the brightness temperature difference between the 12.4 μm and 10.4 μm channels is , calculated according to the following formula:

[0120] = ;

[0121] ;

[0122] in, represents the brightness temperature difference between the 12.4 μm and 10.4 μm channels at pixel i, represents the correction coefficient at pixel i, represents the brightness temperature of the 10.4 μm channel at pixel i, Represents the number of all pixels within any mesoscale convective cloud cluster j.

[0123] Step 3: Tracking of mesoscale convective cloud clusters:

[0124] The mesoscale convective system tracking algorithm is used to determine the motion trajectory of the mesoscale convective cloud cluster identified in step 1. The specific method is as follows:

[0125] Step 3.1: Based on the mesoscale convective cloud cluster MCC at time Tk Ti The center position information is set, and the search radius R is preset. At time Tk+1, the mesoscale convective cloud cluster search is performed within the search radius R. The number of mesoscale convective cloud clusters whose center positions fall within the search radius R is recorded as N, and each mesoscale convective cloud cluster at time Tk+1 is recorded as MCC. 1 Tk+1 , MCC 2 Tk+1 , ..., MCC N Tk+1 , N is a positive integer;

[0126] Step 3.2: If N = 1, then the only mesoscale convective cloud cluster MCC 1 Tk+1 The center position is recorded as the mesoscale convective cloud cluster MCC Tk At the time Tk+1; if N ≥ 2, then the mesoscale convective cloud cluster MCC at the time Tk+1 is the closest to the mesoscale convective cloud cluster MCC at the time Tk. Tk The center position of the nearest mesoscale convective cloud cluster is recorded as the mesoscale convective cloud cluster MCC at time Tk. Tk The predicted position at time Tk+1;

[0127] Step 3.3: For the subsequent moments Tk+2, Tk+3, …, Tk+n, the mesoscale convective cloud cluster trajectory search is performed according to steps 3.1 and 3.2 until no mesoscale convective cloud cluster is found within the search radius R. Then the mesoscale convective cloud cluster MCC is Tk The trajectory search is finished.

[0128] Figure 2 Schematic diagram of the application example of the mesoscale convective cloud tracking method provided by the present invention. (White area) Track diagram, filled with 10.4 micron channel brightness temperature, black star represents the center of the mesoscale convective cloud in the example, and white numbers represent the number of the mesoscale convective cloud. Figure 2 It can well reflect the birth, development and extinction process of mesoscale convective clouds.

[0129] Step 4: Dynamically monitor machine learning data set construction:

[0130] Combine the physical characteristic parameters of any mesoscale convective cloud cluster counted in step 2 and the motion trajectory of the corresponding mesoscale convective cloud cluster tracked in step 3, and use the lightning activity information observed by the lightning positioning system to construct a dynamic monitoring machine learning dataset. Specifically, use the trajectories of all mesoscale cloud clusters tracked in step 3 to count the number of lightning occurrences in the mesoscale convective cloud clusters at any time, and mark thunderstorms and non-thunderstorm events; then combine the physical characteristic parameters of the mesoscale convective cloud clusters obtained in step 2, take the physical characteristic parameters of the mesoscale convective cloud clusters at continuous moments from T1 to Tm, and the lightning activity information of the mesoscale convective system at moment Tm+1 as a set of input and output data; and so on, to construct a dynamic monitoring machine learning dataset. The specific implementation method includes the following steps:

[0131] Step 4.1: Using all the mesoscale convective cloud cluster trajectories tracked in step 3, for any mesoscale convective cloud cluster trajectory (MCC Tk , MCC Tk+1 , MCC Tk+2 , ...MCC Tk+n ), according to the lightning location observed by the lightning location system, the number of lightning strikes within the pixel of the mesoscale convective cloud cluster within 5 minutes before and after is counted, that is, the number of lightning strikes in the mesoscale convective cloud cluster at the corresponding time (L Tk , L Tk+1 , ..., L Tk+n ), if the number of lightning is greater than 0, it is marked as "1", which means it is a thunderstorm event; if the number of lightning is equal to 0, it is marked as "0", which means it is not a thunderstorm event;

[0132] Step 4.2: Using the physical characteristic parameters of the mesoscale convective cloud cluster obtained in step 2, for any mesoscale convective system that lasts for more than m+1 (in the present invention, the value of m can be set to m=2) time periods (MCC Tk , MCC Tk+1 , MCC Tk+2 , ...MCC Tk+n ), taking the physical characteristic parameters of the mesoscale convective cloud clusters from T1 to Tm consecutive moments and the lightning activity information at Tm+1 as a set of machine learning data;

[0133] Step 4.3: Take the physical characteristic parameters of the mesoscale convective cloud cluster from T2 to Tm+1 and the lightning activity information at Tm+2 as a set of machine learning data, and so on, until the trajectory of the mesoscale convective cloud cluster terminates, and finally construct a dynamic monitoring machine learning data set.

[0134] It can be seen that the constructed dynamic monitoring machine learning dataset of the characteristic parameters of mesoscale convective clouds includes the characteristic parameter combination of mesoscale convective clouds and the lightning activity information matching the characteristic parameter combination of mesoscale convective clouds.

[0135] The training of the thunderstorm cloud dynamic monitoring and identification model includes the following steps:

[0136] The dynamic monitoring machine learning dataset of the characteristic parameters of mesoscale convective clouds constructed in step four is used to train the dynamic monitoring and identification model of thunderstorm clouds constructed based on the long short-term memory network. In this embodiment, the dynamic monitoring machine learning dataset is divided into a training set, a test set, and a validation set according to a certain ratio; further, based on the training set and the test set, different combinations of characteristic parameters of mesoscale convective clouds and multiple groups of hyperparameter combinations are designed, and the training dataset and the test dataset are trained based on the dynamic monitoring and identification model of thunderstorm clouds. After multiple iterations, the optimal model for thunderstorm cloud identification is constructed; finally, the validation set is input to verify the accuracy of thunderstorm cloud identification. The specific implementation method includes the following steps:

[0137] Step 5.1, data set division:

[0138] The dynamic monitoring machine learning dataset constructed in step 4 is divided into three parts: training set, test set and validation set according to 6:2:2;

[0139] Step 5.2: Model input feature parameter design:

[0140] A set of long short-term memory network model hyperparameters are designed: the learning rate is 0.001, the time series length is 3; the hidden layer in the network hyperparameters is set to 50 layers, the batch size is 100, the number of network layers is set to 3 layers, and the number of iterations is 300 times.

[0141] Design the characteristic parameter combination of mesoscale convective cloud clusters as model input:

[0142] The first category is the combination of cloud microscopic parameters (Input 1): the average, minimum and maximum brightness temperature of the 10.4 μm channel, the average, minimum and maximum values ​​of the brightness temperature difference between the 6.9 μm and 7.3 μm channels, the average, minimum and maximum values ​​of the brightness temperature difference between the 7.3 and 10.4 μm channels, and the average, minimum and maximum values ​​of the brightness temperature difference between the 12.4 and 10.4 μm channels.

[0143] The second category is the improved cloud cluster micro-parameter combinations (Input2): improved average value of 10.4 micron channel, minimum and maximum brightness temperatures of 10.4 micron channel, improved average value of brightness temperature difference between 6.9 micron and 7.3 micron channels, minimum and maximum values ​​of brightness temperature difference between 6.9 micron and 7.3 micron channels, improved average value of brightness temperature difference between 7.3 and 10.4 micron channels, minimum and maximum values ​​of brightness temperature difference between 7.3 and 10.4 micron channels, improved average value of brightness temperature difference between 12.4 and 10.4 micron channels, minimum and maximum values ​​of brightness temperature difference between 12.4 and 10.4 micron channels.

[0144] The third category is the combination of cloud macroscopic parameters (Input 3): the major axis length, minor axis length, and rotation angle of the mesoscale convective cloud cluster.

[0145] The fourth category is the combination of cloud micro-parameters and cloud macro-parameters (Input4): the average, minimum and maximum brightness temperature of the 10.4 μm channel, the average, minimum and maximum values ​​of the brightness temperature difference between the 6.9 μm and 7.3 μm channels, the average, minimum and maximum values ​​of the brightness temperature difference between the 7.3 and 10.4 μm channels, the average, minimum and maximum values ​​of the brightness temperature difference between the 12.4 and 10.4 μm channels; the major axis length, minor axis length and rotation angle of mesoscale convective clouds.

[0146] The fifth category is the improved combination of cloud micro-parameters and cloud macro-parameters (Input5): improved average value of 10.4 μm channel, minimum and maximum brightness temperatures of 10.4 μm channel, improved average value of brightness temperature difference between 6.9 μm and 7.3 μm channels, minimum and maximum values ​​of brightness temperature difference between 6.9 μm and 7.3 μm channels, improved average value of brightness temperature difference between 7.3 and 10.4 μm channels, minimum and maximum values ​​of brightness temperature difference between 7.3 and 10.4 μm channels, improved average value of brightness temperature difference between 12.4 and 10.4 μm channels, minimum and maximum values ​​of brightness temperature difference between 12.4 and 10.4 μm channels; major axis length, minor axis length and rotation angle of mesoscale convective clouds.

[0147] Inupt1, Input2, Input3, Input4 and Input5 in the training set and test set in 5.1 are used as model inputs respectively. The long short-term memory network model is adopted, and the binary classification BCEwithLogitsLoss loss function is used to minimize the error between the predicted and observed values. After 300 iterations, the model with the smallest loss function value of the test set is taken as the optimal model, which are denoted as Model1, Model2, Model3, Model4 and Model5 respectively.

[0148] The BCEwithLogitsLoss loss function formula is as follows:

[0149]

[0150]

[0151] in, is the input historical mesoscale convective cloud cluster characteristic parameter group, is the true value (i.e., the lightning activity information mentioned above, which is the sample label), is the predicted value. Represents the Sigmoid activation function. Finally, use Inupt1, Input2, Input3, Input4, and Input5 in the validation set as model inputs, and use Model1, Model2, Model3, Model4, and Model5 to calculate the detection rate and false alarm rate indicators respectively.

[0152] The calculation formula for the output result of the thunderstorm cloud dynamic monitoring and identification model is as follows:

[0153]

[0154] Among them, out is the model output, net represents the optimal thunderstorm cloud dynamic monitoring and identification model, is the Sigmoid activation function.

[0155] Figure 3 The cost function convergence curve of the thunderstorm cloud dynamic monitoring and recognition model training set under the input of two sets of historical mesoscale convective cloud characteristic parameter sets in the example of the present invention [refer to Figure 3 (a)], and the detection rate of the validation set [refer to Figure 3 (b)] and the false alarm rate [refer to Figure 3 (c)]. It can be seen that: when using the improved cloud micro-parameters (Input2), compared with the original (Input1), the loss value of the model cost function decreases, the detection rate increases by about 3%, and the false alarm rate is about 20.6%; in addition, when the improved cloud micro-parameters and macro-parameters are introduced at the same time (Input5), the detection rate increases by about 7%, and the false alarm rate is about 17% compared with the unimproved cloud micro-parameters and macro-parameters (Input4). Therefore, considering the above quantitative evaluation indicators, the improved combination of cloud micro-parameters and macro-parameters is used as the input of the thunderstorm cloud dynamic monitoring and identification model, which can improve the accuracy of thunderstorm cloud identification to a certain extent.

[0156] Step 5.3 Model hyperparameter design and optimal model accuracy verification:

[0157] The training set and the test set, and the improved mesoscale convective cloud microphysical characteristic parameters and macroscopic parameter combination (Input5) obtained in step 5.2 are used as model input.

[0158] Design the hyperparameters of the long short-term memory network model: the learning rate is 0.001, the time series length is 3; design multiple sets of hyperparameters, hidden layers: 40, 50, 60, 80, 100 layers, batch size: 100, 300, 500, 700, 800, network layers: 2, 3, 4, 6, 8, 10 layers. Use the binary BCEwithLogitsLoss loss function to minimize the error between the predicted and observed values, and the number of iterations is 600; through 600 iterations, the LOSS distribution curves under different hyperparameter combinations are obtained through training, and the hyperparameter combination with the smallest loss function value of the test set is taken as the optimal hyperparameter combination.

[0159] The BCEwithLogitsLoss loss function formula is as follows:

[0160]

[0161]

[0162] in, is the input historical mesoscale convective cloud cluster characteristic parameter group, is the true value, is the predicted value. Represents the Sigmoid activation function.

[0163] The optimal model, the long short-term memory network model, was used to conduct quantitative evaluation on the validation set using the detection rate and false alarm rate indicators.

[0164] The calculation formula for the output result of the thunderstorm cloud dynamic monitoring and identification model is as follows:

[0165]

[0166] Among them, out is the model output, net represents the optimal thunderstorm cloud dynamic monitoring and identification model, is the Sigmoid activation function.

[0167] Figure 4 The cost function convergence curves of the training set and test set of the thunderstorm cloud group dynamic monitoring and identification model of the example of the present invention, as well as the detection rate and false alarm rate of the validation set. In this example, the optimal hyperparameter combination of the thunderstorm cloud group dynamic monitoring and identification model is 4 network layers, 50 hidden layers, and 100 batch size. The detection rate of thunderstorms in the validation set is 81.1%, and the false alarm rate is 17.3%.

[0168] 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 scope of protection of the present invention.

Claims

1. A method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellite based on long short-term memory network, characterized in that: The steps include: A thunderstorm cloud dynamic monitoring and recognition model was built based on the long short-term memory network model, and the dynamic monitoring machine learning dataset was used to complete the training; The dynamic monitoring machine learning dataset includes several groups of dynamic monitoring machine learning arrays, each of which includes a group of historical mesoscale convective cloud cluster characteristic parameters matched in time and space and lightning activity occurrence label data; The historical mesoscale convective cloud cluster characteristic parameter group is formed by statistically analyzing the physical characteristic parameters of the mesoscale convective cloud cluster at m consecutive moments before any historical forecast time Tm+1. The lightning activity occurrence label data is obtained through the lightning activity observed by the lightning location system at the location of the corresponding mesoscale convective cloud cluster at the historical forecast time Tm+1; The input of the trained thunderstorm cloud dynamic monitoring and recognition model is all the mesoscale convective clouds identified in the target area at the current time t m+1 The physical characteristic parameters corresponding to the first m consecutive moments are output as the current moment t m+1 Thunderstorm cloud identification results.

2. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks according to claim 1, characterized in that: The physical characteristic parameters of mesoscale convective clouds at any time include cloud macroscopic parameters, cloud microscopic parameters A and improved cloud microscopic parameters B; among which: The cloud cluster macroscopic parameters include the central longitude and latitude of the mesoscale convective cloud cluster, the major axis length, minor axis length, and rotation angle of the circumscribed ellipse; The cloud micro-parameter A and improved cloud micro-parameter B are calculated based on the brightness temperature data detected by the infrared imager carried on the geostationary satellite in the target area; among them: For any mesoscale convective cloud cluster j, the elements of the cloud cluster microscopic parameter A correspond to: the mean value of the brightness temperature of the 10.4 μm channel , minimum value and the highest value , 6.9 micron and 7.3 micron channel brightness temperature difference Mean of TB1 , minimum value Mi and the maximum value Max , 7.3 μm and 10.4 μm channel brightness temperature difference Mean of TB2 , minimum value Mi and the maximum value Max , 12.4 micron and 10.4 micron channel brightness temperature difference Mean of TB3 , minimum value Mi and the maximum value Max ; Any element in the improved cloud micro-parameter B is obtained by optimizing the mean element in the cloud micro-parameter A using the infrared channel brightness temperature correction factor; For any mesoscale convective cloud cluster j, the elements of the cloud cluster microscopic parameter B correspond to: the mean value of the brightness temperature of the improved 10.4 μm channel ; Improved average of the brightness temperature difference between the 6.9 micron and 7.3 micron channels ; Improved average of the brightness temperature difference between the 7.3 μm and 10.4 μm channels ; Improved average of the brightness temperature difference between the 12.4 micron and 10.4 micron channels .

3. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks according to claim 2 is characterized in that: The infrared channel brightness temperature correction factor is calculated by the following formula: ； Where: represents the brightness temperature of the 10.4 μm channel at pixel i, Represents the infrared channel brightness temperature correction factor at pixel i; Among the improved cloud microscopic parameters B, any element Calculated by the following formula: ； Where: represents the first of the cloud microscopic parameters B elements, ; Indicates the average brightness temperature of the 10.4 micron channel ; Represents the average value of the brightness temperature difference between the 6.9 μm and 7.3 μm channels ; Represents the average value of the brightness temperature difference between the 7.3 μm and 10.4 μm channels ; Represents the average value of the brightness temperature difference between the 12.4 μm and 10.4 μm channels ; N represents the number of all pixels in any mesoscale convective cloud cluster j.

4. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks as claimed in claim 3 is characterized in that: For any mesoscale convective cloud cluster j, according to the contour points of the mesoscale convective cloud cluster j, based on the least squares method, the distance between the circumscribed ellipse of the mesoscale convective cloud cluster j and the contour points is minimized, and the major axis length, minor axis length, and rotation angle of the circumscribed ellipse of any mesoscale convective cloud cluster j can be obtained; For any mesoscale convective cloud cluster j, the mean brightness temperature of the 10.4 μm channel in the mesoscale convective cloud cluster j can be calculated based on the brightness temperature of the 10.4 μm channel of all pixels in the mesoscale convective cloud cluster j: , minimum value and the highest value ; For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 6.9 μm and 7.3 μm channels of all pixels in the mesoscale convective cloud cluster j. TB1, and then statistically obtain the brightness temperature difference between the 6.9 μm and 7.3 μm channels in the mesoscale convective cloud cluster j Mean of TB1 ,lowest and the highest value ; For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 7.3 μm and 10.4 μm channels of all pixels in the mesoscale convective cloud cluster j. TB2, and then statistically obtain the brightness temperature difference between the 7.3 μm and 10.4 μm channels in the mesoscale convective cloud cluster j Mean of TB2 ,lowest and the highest value ; For any mesoscale convective cloud cluster j, first calculate the brightness temperature difference between the 12.4 μm and 10.4 μm channels of all pixels in the mesoscale convective cloud cluster j. TB3, and then statistically obtain the brightness temperature difference between the 12.4 μm and 10.4 μm channels in the mesoscale convective cloud cluster j Mean of TB3 ,lowest and the highest value .

5. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks according to claim 1, characterized in that: The dynamic monitoring machine learning dataset is constructed by the following steps: Step 1: Identification of mesoscale convective clouds: Obtain the brightness temperature data of the target area detected by the infrared imager on board the geostationary satellite during the historical period, and identify the continuous low infrared brightness temperature area detected at the initial time of the historical period as a mesoscale convective cloud cluster; Step 2: Calculate the characteristic parameters of mesoscale convective clouds: according to The physical connotation of each mesoscale convective cloud cluster identified in step 1 is used to count the physical characteristic parameters of the corresponding mesoscale convective cloud cluster; Step 3: Tracking of mesoscale convective cloud clusters: Using a mesoscale convective system tracking algorithm, the motion trajectory of each mesoscale convective cloud group identified in step 1 is determined, so as to dynamically monitor each mesoscale convective cloud group identified in step 1; Step 4: Dynamically monitor machine learning data set construction: By combining the physical characteristic parameters of any mesoscale convective cloud cluster counted in step two and the motion trajectory of the corresponding mesoscale convective cloud cluster tracked in step three, and using the lightning activity information observed by the lightning location system, a dynamic monitoring machine learning dataset can be constructed.

6. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks according to claim 5, characterized in that: The mesoscale convective cloud cluster is identified based on the 10.4 micron channel brightness temperature data detected by the infrared imager in the target area, and specifically includes the following steps: Step 1.1: In the target area, obtain the brightness temperature of the 10.4 μm channel observed by the infrared imager at any pixel i. ; Step 1.2: Compare the brightness temperature of the 10.4 μm channel and the preset brightness temperature threshold Tb0 to determine whether the corresponding pixel i belongs to a mesoscale convective cloud cluster pixel: When ≥Tb0, pixel i is identified as a mesoscale convective cloud cluster pixel, otherwise it is a non-mesoscale convective cloud cluster pixel; i is a positive integer; Step 1.3, traverse the pixels surrounding the mesoscale convective cloud cluster pixel i. If there are mesoscale convective cloud cluster pixels around the mesoscale convective cloud cluster pixel i, continue traversing until the surrounding adjacent pixels of the mesoscale convective cloud cluster pixel i are all non-mesoscale convective cloud cluster pixels. In this way, a complete mesoscale convective cloud cluster identification is completed, and the identified mesoscale convective cloud clusters are numbered and recorded as mesoscale convective cloud cluster j, where j is a positive integer.

7. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks according to claim 5 or 6, characterized in that: The mesoscale convective cloud cluster trajectory tracking in step 3 specifically includes: Step 3.1: Based on the mesoscale convective cloud cluster MCC at time Tk Tk The center position information is set, and the search radius R is preset. At time Tk+1, the mesoscale convective cloud cluster search is performed within the search radius R. The number of mesoscale convective cloud clusters whose center positions fall within the search radius R is recorded as N, and each mesoscale convective cloud cluster at time Tk+1 is recorded as MCC. 1 Tk+1 , MCC 2 Tk+1 , ..., MCC N Tk+1 , N is a positive integer; Step 3.2: If N = 1, then the only mesoscale convective cloud cluster MCC 1 Tk+1 The center position is recorded as the mesoscale convective cloud cluster MCC Tk At the time Tk+1; if N ≥ 2, then the mesoscale convective cloud cluster MCC at the time Tk+1 is the closest to the mesoscale convective cloud cluster MCC at the time Tk. Tk The center position of the nearest mesoscale convective cloud cluster is recorded as the mesoscale convective cloud cluster MCC at time Tk. Tk The predicted position at time Tk+1; Step 3.3: For the subsequent moments Tk+2, Tk+3, …, Tk+n, the mesoscale convective cloud cluster trajectory search is performed according to steps 3.1 and 3.2 until no mesoscale convective cloud cluster is found within the search radius R. Then the mesoscale convective cloud cluster MCC is Tk The trajectory search is finished.

8. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks as claimed in claim 5, characterized in that: The construction of the dynamic monitoring machine learning data set in step 4 specifically includes: Step 4.1, using all the mesoscale convective cloud cluster trajectories tracked in step 3, for any mesoscale convective cloud cluster trajectory, according to the lightning location and lightning occurrence time observed by the lightning location system, count the lightning activities of the mesoscale convective cloud cluster at different times, and mark the occurrence of lightning as "1" and the absence of lightning as "0"; Step 4.2: Using the physical characteristic parameters of the mesoscale convective cloud clusters obtained in step 2, for any mesoscale convective cloud cluster trajectory that lasts for more than m+1 time periods, the physical characteristic parameters of the mesoscale convective cloud clusters from T1 to Tm consecutive moments and the lightning activity information at Tm+1 time period are used as a set of machine learning data; Step 4.3: Take the physical characteristic parameters of the mesoscale convective cloud cluster from T2 to Tm+1 and the lightning activity information at Tm+2 as a set of machine learning data, and so on, until the trajectory of the mesoscale convective cloud cluster terminates, and finally construct a dynamic monitoring machine learning data set.

9. The method for dynamic monitoring and identification of thunderstorm clouds by geostationary satellites based on long short-term memory networks according to claim 6, characterized in that: The training of the thunderstorm cloud dynamic monitoring and identification model specifically includes the following steps: Step 5.1: Divide the dynamic monitoring machine learning dataset constructed in step 4 into three parts: training set, test set and validation set according to the ratio of 6:2:2; Step 5.2: Using the training set and the test set, adopting the long short-term memory network model, designing different mesoscale convective cloud cluster feature parameter inputs, and using the binary classification BCEwithLogitsLoss loss function, test the optimal input feature parameter combination of the thunderstorm cloud cluster dynamic monitoring and recognition model; Step 5.3: Combine the optimal input feature parameters obtained in step 5.2, use the training set and the test set, adopt the long short-term memory network model, design multiple sets of network hyperparameters, and use the binary classification BCEwithLogitsLoss loss function to optimize the network hyperparameters of the thunderstorm cloud dynamic monitoring and identification model; for the validation set, use the detection rate and false alarm rate indicators to evaluate the accuracy of the output of the thunderstorm cloud dynamic monitoring and identification model.

10. The method for dynamic monitoring and identification of geostationary satellite thunderstorm clouds based on long short-term memory network according to claim 9, characterized in that: The BCEwithLogitsLoss loss function formula in step 5.2 and step 5.3 is as follows: ； ； in, is the input historical mesoscale convective cloud cluster characteristic parameter group, is the sample label, is the predicted value; Represents the Sigmoid activation function; In steps 5.2 and 5.3, the calculation formula for the output of the thunderstorm cloud dynamic monitoring and identification model is as follows: ； Among them, out is the model output, net represents the thunderstorm cloud dynamic monitoring and identification model, is the Sigmoid activation function.

Images