Convection Occurrence Probability Forecast Method Based on Fengyun Geostationary Satellite Spectral Imager

Abstract

The present invention discloses a method for predicting the probability of convection based on a Fengyun geostationary satellite spectral imager, comprising the following steps: step one, constructing a convection probability prediction model; step two, outputting a preliminary result of the convection probability prediction of a target area: based on the constructed convection probability prediction model, inputting the satellite real-time observation data of the target grid point #imgabs0#, the preliminary result of the convection probability prediction #imgabs1# can be obtained; step three, correcting the preliminary result of the convection probability prediction of the target grid point: by further judging whether the cloud cluster of the target grid point #imgabs2# is a convective cloud cluster, it is determined whether the preliminary result of the convection probability prediction needs to be corrected, so as to output the final result of the convection probability prediction. It can be seen that the present invention can realize real-time early warning of the probability of convection based on satellite observation data, provide technical support for convection monitoring, especially for areas with insufficient ground-based observation data, such as plateaus and over the ocean, and can well improve the disaster prevention and mitigation capabilities for severe convective weather.

Description

Technical Field

[0001] The invention relates to a method for predicting the occurrence probability of convection based on a wind-cloud geostationary satellite spectral imager, and belongs to the technical field of weather forecasting. Background Art

[0002] Convective weather systems, especially severe convective weather, are the main manifestations of sudden disasters, such as thunderstorms, gale force winds, and short-term heavy rainfall. This type of weather system is characterized by strong suddenness, great destructive power, and short duration, which poses a huge threat to people's lives, agriculture, transportation, etc., and is often accompanied by derivative disasters such as floods and mudslides, which further increase the degree of harm. Therefore, achieving accurate and detailed observation and early warning of convective weather has always been the focus and difficulty of meteorological research and forecasting.

[0003] Traditional ground-based radars and observation sites play an important role in convection monitoring. However, due to the sparse and uneven distribution of observation sites and the fact that ground-based radars are greatly affected by terrain obstruction, the observation data have problems such as low temporal and spatial resolution and limited coverage. In contrast, satellite remote sensing technology has the advantages of wide coverage, high temporal and spatial resolution, and high observation frequency, and can provide more comprehensive and continuous convection monitoring data.

[0004] The hyperspectral detector and spectral imager carried by the Fengyun-4 geostationary satellite can provide high-precision, high-temporal and high-resolution infrared observation data. Using these data, the formation and development of severe convective weather can be discovered earlier, tens of minutes to hours earlier than traditional ground-based radar observations, and early identification and warning of convective weather can be achieved. By combining advanced machine learning algorithms, satellite remote sensing data can significantly improve the recognition accuracy and warning timeliness of convective weather, providing strong technical support for disaster prevention and mitigation work. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting the probability of convection based on the Fengyun geostationary satellite spectral imager to solve the problem of insufficient accuracy of convective weather forecasts in the prior art. It significantly improves the forecast accuracy and timeliness of convective weather by providing a convective probability forecast model that can learn and evolve, combining satellite observation data (obtained by the spectral imager carried on the Fengyun geostationary satellite) and radar combined reflectivity data (obtained by ground-based weather radar observations). The aforementioned convective probability forecast model can not only adapt to different meteorological conditions, but also continuously improve the forecast performance through continuous learning and optimization. In addition, the present invention also takes into account that the difference between satellite observation data and radar combined reflectivity data may cause deviations in the prediction results. Therefore, the present invention reduces such deviations by correcting the preliminary results of the convective probability prediction output by the convective probability forecast model, thereby further improving the accuracy of the convective probability prediction results.

[0006] In order to achieve the above technical objectives, the present invention will adopt the following technical solutions:

[0007] A method for predicting the probability of convection occurrence based on a Fengyun geostationary satellite spectral imager comprises the following steps:

[0008] Step 1: Construct a convection probability prediction model;

[0009] The convection probability prediction model is built based on a deep convolutional neural network and trained using a training dataset;

[0010] The training data set includes historical satellite observation data and historical convection occurrence area label data that match time and space;

[0011] The historical satellite observation data are observation data of key spectral channels of the spectral imager carried on the Fengyun geostationary satellite in a specified historical period; the key spectral channels of the spectral imager include infrared spectral channels, water vapor spectral channels and visible light spectral channels;

[0012] The historical convection zone label data is generated based on the historical radar combined reflectivity data of the same period as the historical satellite observation data;

[0013] The historical radar combined reflectivity data is the observation data of the laser radar in a specified historical period;

[0014] Step 2: Output the preliminary results of the convection probability prediction in the target area;

[0015] Focus the key spectral channels of the spectral imager on the target grid The real-time satellite observation data is preprocessed and then input into the convection probability prediction model to predict the target grid point. The convection probability of the cloud cluster is calculated and the target grid is output. Preliminary results of convective probability forecast ;

[0016] Step 3: Correct the preliminary results of the convection probability prediction of the target grid:

[0017] Based on the target grid Preliminary results of convective probability forecast , combined with the target grid The cloud top temperature, cloud layer thickness, convective region thickness and / or microstructure of the cloud cluster are used to further determine the target grid point. Is the cloud cluster a convective cloud cluster? When the judgment result shows that the target grid If the cloud cluster is a convective cloud cluster, then the target grid Preliminary results of convective probability forecast As the final result of the convection probability prediction Output; otherwise, the target grid is output The final result of the convection probability prediction =0.

[0018] Preferably, in step 3, the target grid point The cloud top temperature, cloud layer thickness, convective region thickness and / or microstructure of a cloud cluster are determined based on the brightness temperature output by a typical spectral channel of a spectral imager;

[0019] The typical spectral channels of the spectral imager include three infrared spectral channels and one water vapor spectral channel; wherein: the three infrared spectral channels correspond to the infrared 10.8 micron channel, the infrared 12 micron channel and the infrared 8.55 micron channel; the one water vapor spectral channel is specifically the water vapor 6.25 micron channel;

[0020] The cloud top temperature of the cloud cluster is output through the brightness temperature BT of the infrared 10.8 micron channel 10.8 To observe and when BT 10.8 <263K, indicating that the cloud top temperature is less than 0 degrees Celsius;

[0021] The cloud thickness of the cloud cluster is calculated by calculating the brightness temperature difference between the water vapor 6.25 μm channel and the infrared 10.8 μm channel. 6.25-10.8 to observe and when BTD 6.25-10.8 When >-2K, it indicates that the cloud thickness of the cloud cluster has reached the preset thickness;

[0022] The thickness of the convective region of the cloud cluster is calculated by calculating the brightness temperature difference between the infrared 12 μm channel and the infrared 10.8 μm channel. 12-10.8 To observe, and when -2K <BTD 10.8-12 When <2K, it indicates that there is a convection area of ​​preset thickness in the cloud;

[0023] The cloud microstructure is determined by calculating the brightness temperature difference (BTD) between the infrared 8.55 μm channel and the infrared 10.8 μm channel. 8.55-10.8 to observe and when BTD 8.55-10.8 When >0, it indicates that the microstructure of the cloud is a typical ice cloud.

[0024] Preferably, in step 3, based on the target grid Preliminary results of convective probability forecast The size of the target grid The cloud top temperature, cloud layer thickness, convective region thickness and / or cloud microstructure of the cloud cluster are used to calculate the target grid point Preliminary results of convective probability forecast Perform segmented correction, including:

[0025] When the target grid Preliminary results of convective probability forecast satisfy When Whether the cloud cluster meets the following requirements at the same time: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud; when the judgment result shows that the target grid point The target grid point is determined if the cloud cluster satisfies the following conditions at the same time: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud. The cloud cluster is a convective cloud cluster, and the target grid point Preliminary results of convective probability forecast As the final result of the convection probability prediction Output;

[0026] When the target grid Preliminary results of convective probability forecast satisfy When The cloud cluster meets any two of the following four conditions: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud; when the judgment result shows that the target grid point If the cloud cluster does not meet any two of the following four conditions: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness of the cloud cluster meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the shape of the cloud cluster is a typical ice cloud, then the target grid point is judged The cloud cluster is a non-convective cloud cluster, and the target grid is output The final result of the convection probability prediction =0;

[0027] When the target grid Preliminary results of convective probability forecast satisfy When Whether the cloud cluster satisfies the following conditions at the same time: the cloud top temperature is less than 0 degrees Celsius and the cloud microstructure is a typical ice cloud; if the judgment result shows that the target grid point The cloud cluster satisfies: the cloud top temperature is less than 0 degrees Celsius and the cloud microstructure is a typical ice cloud, then the target grid point is judged The cloud cluster is a convective cloud cluster, and the target grid point Preliminary results of convective probability forecast As the final result of the convection probability prediction Output; otherwise, the target grid is determined The cloud cluster is a non-convective cloud cluster, and the target grid is output The final result of the convection probability prediction =0;

[0028] in: is a preset first convection probability threshold; is a preset second convection probability threshold;

[0029] The first and second convection probability thresholds meet: and , .

[0030] Preferably, the first convection probability threshold ; Second convection probability threshold .

[0031] Preferably, in step 1, the training data set is constructed by the following steps:

[0032] Step 1.1, obtaining the satellite historical observation data and historical radar combined reflectivity data of the specified historical period, and generating the convection history occurrence area label data of the specified historical period based on the historical radar combined reflectivity data;

[0033] Satellite historical observation data comes from the observation data of key spectral channels of the Fengyun geostationary satellite imager in a specified historical period;

[0034] The historical radar combined reflectivity data comes from the observation data of the lidar in a specified historical period;

[0035] Step 1.2, time matching is performed on the acquired satellite historical observation data and the convection history occurrence area label data to ensure that the satellite historical observation data and the convection history occurrence area label data are at the same observation time;

[0036] Step 1.3: spatially match the satellite historical observation data and the convection history occurrence area label data that have been time-matched to ensure that the satellite historical observation data and the convection history occurrence area label data are on the same spatial grid;

[0037] Step 1.4, statistically calculate the mean and standard deviation of each key spectral channel observation data in the satellite historical observation data, and normalize the observation data of each key spectral channel;

[0038] Through the above steps, the construction of the training data set can be completed.

[0039] Preferably, in step 1.1, the generation of the flow history drop zone label data specifically includes the following steps:

[0040] Step 111, obtaining historical radar combined reflectivity data for a specified historical period;

[0041] Step 112: Identify the convection occurrence area:

[0042] Based on the historical radar combined reflectivity data obtained in step 111, the corresponding radar echo intensity information is extracted, and the convective cloud cluster areas with radar echo intensity information greater than 35dBZ and an area greater than 25 square kilometers are screened out and marked, so as to complete the identification of the convective occurrence area;

[0043] The radar combined reflectivity data in the convection occurrence area are processed by nearest neighbor spatial fusion interpolation to make it consistent with the spatial resolution of the historical satellite observation data in the same period;

[0044] Step 113: Construct label data of the convection occurrence area:

[0045] The convection occurrence area label is the convection occurrence area label within the next 1 hour at any time t in the specified historical period, which is specifically constructed in the following way: first, the convection occurrence area within the next 1 hour at any time t is collected; then, the radar combined reflectivity data in the convection occurrence area within the next 1 hour at any time t is spatially merged by using the convective cloud cluster tracking algorithm, so as to generate the convection occurrence area label data within the next 1 hour at any time t;

[0046] Step 114: Filter the historical satellite observation data of the same period:

[0047] Based on the time t recorded in step 113, the satellite historical observation data two hours before the time t is screened, and the screened satellite historical observation data is interpolated by the nearest neighbor to be consistent with the spatial resolution of the radar combined reflectivity data.

[0048] Preferably, in step 112, the nearest neighbor spatial fusion interpolation processing is specifically as follows: for each target area grid point, first collect all radar combined reflectivity data in an area with a radius of 0.02 degrees around the target area grid point, and then adopt the distance inverse weighted proportional method, that is, the weight of the target area grid point is 1, and the boundary weight of the radius 0.02 degrees is 0 for weighting, so that the radar combined reflectivity data of the 0.04 degree equal longitude and latitude grid can be fused and generated.

[0049] Preferably, in step 2, the key spectral channel of the spectral imager is directed to the target grid point The satellite real-time observation data is preprocessed through the following steps:

[0050] Step 2.1, the satellite real-time observation data is processed by nearest neighbor spatial fusion interpolation to make the spatial resolution of the satellite real-time observation data consistent with the convection occurrence area label data;

[0051] Step 2.2: Use the mean and standard deviation of the observation data of each key spectral channel in the satellite historical observation data counted in step 1.4 to normalize the satellite real-time observation data to ensure data consistency.

[0052] Preferably, the convection probability prediction model includes an encoder, a decoder and a loss function, wherein:

[0053] The input data of the encoder is the hourly satellite observation data of the target grid point 2 hours before any observation time T, with a total of 2 time steps: T-2h and T-1h; the encoder includes Layer, each layer of the encoder includes a first convolution layer, a first pooling layer and a first connection layer. In the first layer of the encoder, the first convolution layer is used to convolve the input data, and the first pooling layer is used to downsample the feature map output by the first convolution layer at the corresponding layer of the encoder and output it to the first connection layer; the first connection layer uses feature splicing to fuse the features of different time steps output by the first pooling layer at the corresponding layer of the encoder, and outputs the fused features to the next layer of the encoder, and so on, until it is completed in the encoder. Sub-convolution downsampling feature fusion, extracting high-level features and outputting them to the decoder;

[0054] The decoder includes an upsampling layer, a second connection layer, a second convolutional layer and a Softmax activation function; the upsampling layer is used to upsample the high-level features output by the encoder and output them to the second connection layer; the second connection layer, through a jump connection, splices the feature map input to the second connection layer with the feature map of the corresponding scale in the encoder to obtain high- and low-level fusion features, and outputs them to the second convolutional layer; the second convolutional layer convolves the input high- and low-level fusion features and outputs them to the Softmax activation function; the Softmax activation function can calculate the preliminary result of the convection probability prediction of the target grid point according to the output of the second convolutional layer;

[0055] The loss function is used to measure the difference between the preliminary convection probability prediction results output by the convection probability prediction model and the historical convection zone label data during the training process of the convection probability prediction model.

[0056] Preferably, the output of the encoder is:

[0057] ;

[0058] ;

[0059] ;

[0060] Where: Indicates that the encoder has passed The high-level features output after the sub-convolution downsampling features are fused; Indicates the satellite observation data for time step T-2h, the encoder's The layer passes through the output of the first convolutional layer; Represents the satellite observation data for time step T-1h. The encoder The output of the layer; Indicates the satellite observation data for time step T-2h, the encoder's The layer passes through the output of the first pooling layer; Indicates the satellite observation data for time step T-1h, the encoder The layer passes through the output of the first pooling layer; For encoder Bias parameters for layer outputs; For encoder The convolution kernel of the layer output; BN means batch normalization; ReLU is the activation function; * indicates the convolution operation; hour, Indicates that the data input to the first convolutional layer is the satellite observation data of time step T-2h, Indicates that the data input to the first convolutional layer is the satellite observation data of time step T-1h;

[0061] In the decoder, the output of the upsampling layer, the second connection layer, and the Softmax activation function is:

[0062] ;

[0063] ;

[0064] ;

[0065] Where: represents the output of the upsampling layer, ; represents the output of the second connection layer; Represents the output of the encoder and upsampling layer Feature map of corresponding scale; It represents the preliminary result of the convection probability prediction of the target grid point; The convolution kernel representing the decoder output; Represents the bias parameter of the decoder output; when hour, represents the output of the encoder;

[0066] The loss function used in the training process of the convection probability prediction model is the weighted cross entropy loss function, expressed as follows:

[0067] ;

[0068] Where: is the total number of grid points in the region; For the The convection occurrence domain label of each grid point; The output of the convection probability prediction model is The probability of convection occurring at each grid point; is the positive sample weight coefficient.

[0069] Based on the above technical objectives, the present invention has the following advantages over the prior art:

[0070] The present invention combines the observation data of the Fengyun geostationary satellite spectral imager (i.e., historical satellite observation data) and the ground-based weather radar observation data (i.e., historical radar combined reflectivity data), and adopts a deep learning model capable of learning evolution (i.e., a convective probability prediction model), thereby realizing the probabilistic forecast of convective weather. Compared with the prior art, the present invention has significant advantages in many aspects:

[0071] First, the convective probability prediction model described in the present invention significantly improves the forecast accuracy through multi-source data fusion and convolutional neural network model. At the same time, the convective probability prediction model combines historical satellite observation data and historical radar combined reflectivity data to enhance the ability to identify complex meteorological phenomena. Through the construction and training of the convective probability prediction model, the present invention can predict the occurrence of convective clouds 1 hour in advance, improve the timeliness of early warning, and gain valuable time for disaster prevention and mitigation.

[0072] Secondly, automated processing and rapid response mechanisms have greatly improved forecast efficiency. Model recognition and calculation time is less than 30 seconds, and early warning information can be provided in a timely manner. Continuous online learning mechanisms ensure that the model can be dynamically adjusted and optimized to adapt to different meteorological conditions and improve forecast performance. In addition, forecast products have high temporal and spatial resolution and multiple output formats, providing high-quality forecast information. Efficient deep learning algorithms reduce computing resource consumption, and data preprocessing and quality control ensure the validity of the data.

[0073] Finally, the present invention takes into account that the differences between historical satellite observation data and historical radar combination reflectivity data may cause deviations in the prediction results output by the convective probability prediction model. Therefore, the present invention corrects the preliminary convective probability prediction results output by the convective probability forecast model to reduce such deviations, thereby further improving the accuracy of the convective probability prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a flow chart of the method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager described in the present invention. DETAILED DESCRIPTION

[0075] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means any limitation to the present invention and its application or use. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Unless otherwise specified, the relative arrangement, expressions and numerical values ​​of the components and steps described in these embodiments do not limit the scope of the present invention. The techniques, methods and equipment known to ordinary technicians in the relevant fields may not be discussed in detail, but in appropriate cases, the techniques, methods and equipment should be regarded as part of the specification. In all examples shown and discussed here, any specific value should be interpreted as merely exemplary, rather than as a limitation. Therefore, other examples of the exemplary embodiments may have different values.

[0076] like Figure 1 As shown, the method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager of the present invention comprises the following steps:

[0077] Step 1: Construct a convection probability prediction model;

[0078] The convection probability prediction model described in the present invention is constructed based on a deep convolutional neural network and is trained using a training data set.

[0079] The training data set includes historical satellite observation data with time and space matching and historical convection occurrence area label data.

[0080] The historical satellite observation data are observation data of key spectral channels of the spectral imager carried on the Fengyun geostationary satellite in a specified historical period; the key spectral channels of the spectral imager include infrared spectral channels, water vapor spectral channels and visible light spectral channels;

[0081] The historical convection zone label data is generated based on the historical radar combined reflectivity data of the same period as the historical satellite observation data;

[0082] The historical radar combined reflectivity data is the observation data of the ground-based weather radar in a specified historical period.

[0083] The convection probability prediction model described in the present invention is constructed based on a deep convolutional neural network (CNN) model. Therefore, in order to improve the generalization ability of the convection probability prediction model, it is necessary to construct a large-scale, high-quality training data set. The training data set is constructed through the following steps:

[0084] Step 1.1, obtain the satellite historical observation data and historical radar combined reflectivity data of the specified historical period, and generate the convection history occurrence area label data of the specified historical period based on the historical radar combined reflectivity data.

[0085] Satellite historical observation data comes from the observation data of key spectral channels of Fengyun geostationary satellite imagers in a specified historical period;

[0086] The historical radar composite reflectivity data comes from the observation data of ground-based weather radar in a specified historical period.

[0087] The specified historical period can be June to August 2019 and June to August 2020, which includes two summers in total, to construct the training data set.

[0088] The generation of label data of the historical occurrence area of ​​convection specifically includes the following steps:

[0089] Step 111, obtaining historical radar combined reflectivity data for a specified historical period;

[0090] The temporal resolution of historical radar combined reflectivity data can be 6 minutes and the spatial resolution can be 0.01 degrees.

[0091] Step 112: Identify the convection occurrence area:

[0092] Based on the historical radar combined reflectivity data obtained in step 111, the corresponding radar echo intensity information is extracted, and the convective cloud cluster areas with radar echo intensity information greater than 35dBZ and an area greater than 25 square kilometers are screened out and marked, so as to complete the identification of the convective occurrence area. In the present invention, the convective cloud cluster area is marked as 1, and the area is considered to be the convective occurrence area, and other areas are marked as 0.

[0093] The radar combined reflectivity data in the convection occurrence area are processed by nearest neighbor spatial fusion interpolation to make it consistent with the spatial resolution of the historical satellite observation data in the same period.

[0094] The nearest neighbor spatial fusion interpolation processing is specifically as follows: for each target area grid point, first collect all radar combined reflectivity data in the area with a radius of 0.02 degrees around the target area grid point, and then use the distance inverse weight ratio method, that is, the weight of the target area grid point is 1, and the boundary weight of the radius 0.02 degrees is 0 for weighting, so that the radar combined reflectivity data of 0.04 degrees of equal longitude and latitude grids can be fused and generated. The advantage of this method is that it fully considers the proportion and representativeness of the sampling distance. The closer the distance to the target, the better its representativeness, and the more it can reflect the characteristics of the target area. Conversely, the farther the distance from the target, the worse the representativeness.

[0095] This step aims to obtain accurate ground radar convective cloud distribution information and provide accurate data support for the model to construct convective drop zone label data.

[0096] Step 113: Construct label data of the convection occurrence area:

[0097] The convection occurrence area label is the convection occurrence area label within the next 1 hour at any time t in the specified historical period. It is constructed in the following way: first, the convection occurrence area within the next 1 hour at any time t is collected, and the convective cloud cluster data of the next 10 moments (a total of 1 hour) at any time t are continuously obtained at a time interval of every 6 minutes; then, the convective cloud cluster tracking algorithm is used to spatially merge the radar combined reflectivity data in the convection occurrence area within the next 1 hour at any time t, so as to generate the convection occurrence area label data within the next 1 hour at any time t.

[0098] In this process, the hourly observation time corresponding to each data is recorded so as to time match with the satellite observation data in the subsequent step 114. At the same time, for the boundaries of different observation times, if there is a spatial gap, two observations in adjacent spaces and adjacent times are used for splicing and fusion, that is, for the target point x, if the point is closer to the area at time T1 and farther from the area at time T2, the inverse distance weighted method similar to that in step 112 is also used for fusion.

[0099] The purpose of this step is to construct the target labels (label data of the convective drop zone) required for model training, reflecting the spatial distribution of convective clouds in the next hour.

[0100] Step 114: Filter the historical satellite observation data of the same period:

[0101] Based on the time t recorded in step 113, the hourly observation data of the key spectral channels of the Fengyun geostationary satellite imager in the two hours before the time t are screened, and the hourly observation data of the screened key channels are interpolated by the nearest neighbor to be constructed into 0.04 degree equal longitude and latitude grid data to be consistent with the spatial resolution of the radar combined reflectivity data.

[0102] Select key spectral channels, which mainly involves screening the spectral channels of the Fengyun geostationary satellite imager through information volume and sensitivity analysis, that is, performing spectral analysis on target features, such as slight changes in cloud thickness , compare the spectral changes of each spectral channel before and after the change , observe with caused by changes in The proportion of change : .

[0103] That is, for each satellite spectral channel observation, when the ratio When it is very small, it means that the spectral channel is less sensitive to the target characteristics and the amount of target characteristic information obtained is small. On the contrary, the spectral channel is more sensitive to the target characteristics and can obtain more target characteristic information in actual satellite observations, which can be effectively used to monitor the target characteristics. At the same time, considering the vertical changes in temperature and water vapor in the atmosphere, the spectral channel analysis also considers the sensitivity of each spectral channel to the changes in temperature and water vapor at different altitudes, and the sensitivity to the changes in cloud vertical height, cloud thickness and particle size. The final results include: infrared channel (reflecting cloud top temperature and atmospheric temperature), water vapor channel (reflecting the water vapor content in the atmosphere), and visible light channel (reflecting cloud morphology, thickness, and particle size).

[0104] In other words, the key channels screened out by the present invention include infrared channel, water vapor channel and visible light channel, among which the hourly observation data in the infrared channel can reflect the cloud top temperature and atmospheric temperature, the hourly observation data in the water vapor channel can reflect the water vapor content in the atmosphere, and the hourly observation data in the visible light channel can reflect the cloud morphology, thickness and particle size.

[0105] This step aims to obtain satellite observation information before the occurrence of convective clouds and provide rich input features for the model.

[0106] Step 1.2, time matching is performed on the acquired satellite historical observation data and the convection history occurrence area label data to ensure that the satellite historical observation data and the convection history occurrence area label data are at the same observation time;

[0107] Step 1.3: spatially match the satellite historical observation data and the convection history occurrence area label data that have been time-matched to ensure that the satellite historical observation data and the convection history occurrence area label data are on the same spatial grid;

[0108] Step 1.4: Calculate the mean and standard deviation of each key spectral channel observation data in the satellite historical observation data, and perform Z-Score normalization on each key spectral channel observation data. The formula is:

[0109] ;

[0110] in, is the original data, is the mean, is the standard deviation. To increase the diversity of training data, data augmentation techniques are used, including rotation (0-360 degree random rotation) operations. This step aims to build a large-scale, high-quality training dataset to improve the generalization ability of the model.

[0111] Through the above steps 1.1-1.4, the construction of the training data set can be completed.

[0112] In the present invention, the convective probability prediction model is specifically a multi-input, multi-task deep convolutional neural network (CNN) model, which is used to predict the probability of convective cloud clusters. The main innovation of the convective probability prediction model is to integrate multi-time and multi-channel satellite observation data, and adopt an improved U-Net architecture, combining spatial and temporal features to achieve accurate prediction of convective clouds.

[0113] The convection probability prediction model includes an encoder, a decoder, and a loss function, where:

[0114] The input data of the encoder is the hourly satellite observation data of the target grid point 2 hours before any observation time T, totaling 2 time steps: T-2h and T-1h; the satellite observation data of each time step has a spatial size of H W, the number of channels is C (including infrared channel, water vapor channel, visible light channel, a total of 26).

[0115] The encoder comprises Layer, each layer of the encoder includes the first convolution layer, the first pooling layer and the first connection layer.

[0116] In the first layer of the encoder, the first convolutional layer is used to perform 3 3 convolutions, the number of output channels is 32, the step size is 1, the padding is 1, the activation function is ReLU, and the The output of the layer is , then:

[0117] ;

[0118] in, and Respectively The convolution kernel and bias of the layer, * indicates the convolution operation, BN (abbreviation for BatchNormalization) indicates batch normalization, is the activation function.

[0119] It can be seen that when t=T-2h, the Output of the layer Recorded as , when t=T-1h, then the Output of the layer Recorded as . hour, Indicates that the data input to the first convolutional layer is the satellite observation data of time step T-2h, It means that the data input to the first convolutional layer is the satellite observation data of time step T-1h.

[0120] The first pooling layer is used to downsample the feature map output by the first convolutional layer at the corresponding layer of the encoder and output it to the first connection layer; the first pooling layer uses the maximum pooling layer for downsampling, and the pooling window size is 2 2. Reduce the feature map size by half. It can be expressed as:

[0121] ;

[0122] Where: Represents the first convolutional layer Output of the layer , the output of the first pooling layer.

[0123] The first connection layer uses feature concatenation to fuse the features of different time steps output by the first pooling layer in the corresponding layer of the encoder, and outputs the fused features to the next layer of the encoder. This cycle is repeated until the encoder is completed. The convolution downsamples features and fuses them to extract high-level features, which are then output to the decoder.

[0124] The present invention performs 4 convolution downsampling feature fusions to extract high-level features, and the number of feature channels is gradually doubled (64, 128, 256, 512).

[0125] The output of the encoder can be expressed as:

[0126] ;

[0127] Where: Indicates that the encoder has passed The high-level features output after the sub-convolution downsampling features are fused; Indicates the satellite observation data for time step T-2h, the encoder's The layer passes through the output of the first pooling layer; Indicates the satellite observation data for time step T-1h, the encoder The layer passes the output of the first pooling layer.

[0128] The decoder includes an upsampling layer, a second connection layer, a second convolutional layer and a Softmax activation function.

[0129] The upsampling layer is used to upsample the high-level features output by the encoder and output them to the second connection layer; specifically, the upsampling layer upsamples the features fused by the encoder and restores them to the original spatial size. Bilinear interpolation is used to gradually restore the feature map size. The upsampling layer can be expressed as:

[0130] ;

[0131] In the above formula: represents the output of the upsampling layer, ; represents the output of the second connection layer; when hour, Represents the output of the encoder.

[0132] The second connection layer, through a jump connection, concatenates the feature map input to the second connection layer with the feature map of the corresponding scale in the encoder to obtain high- and low-level fusion features, and outputs them to the second convolutional layer; the second connection layer can be expressed as:

[0133] ;

[0134] In the above formula: represents the output of the second connection layer; Represents the output of the encoder and upsampling layer Feature maps of corresponding scales.

[0135] At the end of the decoder, after a 1 The second convolution layer of 1 performs convolution on the input high- and low-level fusion features, maps the number of feature channels to the number of output channels, and outputs it to the Softmax activation function; the Softmax activation function can calculate the preliminary result of the convection probability prediction of the target grid point based on the output of the second convolution layer. The Softmax activation function can be expressed as:

[0136] ;

[0137] In the above formula: It represents the preliminary result of the convection probability prediction of the target grid point; The convolution kernel representing the decoder output; Represents the bias parameter of the decoder output.

[0138] The final output of the convection probability prediction model is a probability map with the same size as the input. Each pixel point (regional grid point) represents the probability of the location being a convection drop zone, ranging from 0 to 1.

[0139] The loss function is used to measure the difference between the preliminary convection probability prediction results output by the convection probability prediction model and the historical convection zone label data during the training process of the convection probability prediction model.

[0140] Weighted Cross Entropy Loss Function It can be expressed as:

[0141] ;

[0142] Where: is the total number of regional grid points; For the The convection occurrence label of each regional grid point (1 means convection occurs, 0 means no convection occurs); The output of the convection probability prediction model is The probability of convection occurring at each grid point in the region; is the positive sample weight coefficient. Usually set > 0.5 to increase the focus on positive samples.

[0143] The adaptive moment estimation (Adam) optimization algorithm is used, with the initial learning rate set to 0.001 and the batch size to 32. The training process includes the following steps: input the preprocessed training data set into the convection probability prediction model, and calculate the output probability map through the U-Net model; use the weighted cross entropy loss function to calculate the error between the prediction result and the true label; calculate the gradient of the loss function to the model parameters and update the model parameters. Evaluate the model performance on the validation set, monitor the weighted cross entropy loss evaluation index, and prevent overfitting. According to the validation results, adjust the hyperparameters such as learning rate and batch size to optimize the model performance. Repeat the above process until the loss function converges or reaches the preset training rounds. Finally, the model parameters that perform best on the validation set are saved for subsequent real-time prediction.

[0144] Step 2: Output the preliminary results of the convection probability prediction in the target area;

[0145] Focus the key spectral channels of the spectral imager on the target grid The real-time satellite observation data is preprocessed and then input into the convection probability prediction model to predict the target grid point. The convection probability of the cloud cluster is calculated and the target grid is output. Preliminary results of convective probability forecast .

[0146] In this step, the key spectral channel of the spectral imager is aimed at the target grid point The satellite real-time observation data is preprocessed through the following steps:

[0147] Step 2.1, the satellite real-time observation data is processed by nearest neighbor spatial fusion interpolation to make the spatial resolution of the satellite real-time observation data consistent with the convection occurrence area label data;

[0148] Step 2.2: Use the mean and standard deviation of the observation data of each key spectral channel in the satellite historical observation data counted in step 1.4 to normalize the satellite real-time observation data to ensure data consistency.

[0149] Step 3: Correct the preliminary results of the convection probability prediction of the target grid:

[0150] Considering the target grid output by the convection probability prediction model Preliminary results of convective probability forecast , there may be some deviations due to the differences between satellite observation data and weather radar observation methods, so in order to reduce this deviation to obtain more accurate convective probability prediction results. Based on step 2, the present invention corrects the preliminary results of convective probability prediction, specifically: based on the target grid Preliminary results of convective probability forecast , combined with the target grid The cloud top temperature, cloud layer thickness, convective region thickness and / or microstructure of the cloud cluster are used to further determine the target grid point. Is the cloud cluster a convective cloud cluster? When the judgment result shows that the target grid If the cloud cluster is a convective cloud cluster, the target grid Preliminary results of convective probability forecast As the final result of the convection probability prediction Output; otherwise, the target grid is output The final result of the convection probability prediction =0.

[0151] In order to achieve the correction of the preliminary results of the above-mentioned convective probability prediction, the present invention further leverages the advantages of the multi-channel observation of the Fengyun geostationary satellite. Specifically, according to the sensitivity differences of different spectral channels to cloud and atmospheric characteristics, the brightness temperature difference between different spectral channels is used to carry out identification and further analysis and threshold construction of target grid points. The cloud top temperature, cloud layer thickness, convective region thickness and / or microstructure of the cloud cluster are determined. The typical spectral channels of the spectral imager used here include three infrared spectral channels and one water vapor spectral channel; wherein: the three infrared spectral channels correspond to the infrared 10.8 micron channel, the infrared 12 micron channel and the infrared 8.55 micron channel; the one water vapor spectral channel is specifically the water vapor 6.25 micron channel.

[0152] The basic theory is:

[0153] A. The infrared channel 10.8 microns is the atmospheric window channel. Under clear sky conditions, the radiation information of the surface and atmosphere can easily reach the top of the atmosphere. However, when there are clouds, due to the absorption of clouds, the brightness temperature (BT) of the infrared channel 10.8 microns is 10.8 ) mainly reflects the characteristics of the cloud top. When the brightness temperature is less than 263K, the cloud top temperature is less than 0 degrees Celsius. It can be seen that the present invention can output the brightness temperature BT through the infrared 10.8 micron channel. 10.8 To observe the cloud top temperature of the cloud cluster, and when BT 10.8 When <263K, it indicates that the cloud top temperature is less than 0 degrees Celsius.

[0154] B. The absorption of water vapor in the atmosphere in the 12-micron infrared channel is stronger than that in the 10.8-micron infrared channel. Therefore, the energy of the 10.8-micron infrared channel observed by the satellite is greater than that of the 12-micron infrared channel. Therefore, the brightness temperature difference BTD between the 12-micron infrared channel and the 10.8-micron infrared channel is 10.8-12 Usually greater than 0K. With the presence of clouds, especially when the clouds are thick, the infrared channel emission radiation capacity of the clouds gradually increases, which will significantly weaken the influence of water vapor absorption in the atmosphere below the clouds. The energy at the top of the atmosphere observed by satellites is mainly contributed by the emission energy of clouds. Therefore, for the development area of ​​deep convective clouds, the brightness temperature difference (BTD) between the infrared 12 micron channel and the infrared 10.8 micron channel is 10.8-12 ) is smaller or equal to 0. Therefore, the present invention calculates the brightness temperature difference BTD between the infrared 12 micron channel and the infrared 10.8 micron channel 12-10.8 To observe the thickness of the convective region of the cloud, and when -2K <BTD 10.8-12 When it is <2K, it indicates that there is a convective area of ​​preset thickness in the cloud.

[0155] C. For the infrared 8.55 micron channel compared with the infrared 10.8 micron channel, although the absorption capacity of ice clouds and water clouds in the infrared 8.55 micron channel is relatively close, the absorption capacity of ice cloud particles in the infrared 10.8 micron channel is significantly greater than that of water cloud particles. Therefore, for typical ice cloud areas, the energy of the infrared 10.8 micron channel observed by the satellite is significantly smaller than the channel energy of the infrared 8.55 micron channel, that is, for ice cloud areas, the brightness temperature difference BTD between the infrared 8.55 micron channel and the infrared 10.8 micron channel is 8.55-10.8 >0, while for the water cloud region, the opposite is true. The brightness temperature difference BTD between the infrared 8.55 μm channel and the infrared 10.8 μm channel 8.55-10.8 <0. Therefore, the present invention can calculate the brightness temperature difference BTD between the infrared 8.55 micron channel and the infrared 10.8 micron channel 8.55-10.8 To observe the microstructure of clouds and to serve as a BTD 8.55-10.8 When >0, it indicates that the microstructure of the cloud is a typical ice cloud.

[0156] D. The 6.25-micron water vapor channel is a strong absorption channel for water vapor. Therefore, the energy of this channel observed by the satellite under clear sky conditions is relatively small. The brightness temperature difference BTD between the 6.25-micron water vapor channel and the 10.8-micron infrared channel is 6.25-10.8 It is far less than 0, or even less than -50. However, when clouds exist, the higher the cloud height and the thicker the cloud thickness, the brightness temperature difference BTD between the water vapor 6.25 micron channel and the infrared 10.8 micron channel 6.25-10.8 The brightness temperature difference BTD between the water vapor 6.25 μm channel and the infrared 10.8 μm channel gradually decreases and approaches 0, especially for thicker clouds with higher cloud tops. 6.25-10.8 will be greater than 0. Therefore, the present invention calculates the brightness temperature difference BTD between the water vapor 6.25 micron channel and the infrared 10.8 micron channel 6.25-10.8 To observe the cloud thickness of the cloud cluster and to serve as a BTD 6.25-10.8 When >-2K, it indicates that the cloud thickness of the cloud cluster has reached the preset thickness.

[0157] In addition, when correcting the preliminary results of the convection probability prediction, the present invention uses the target grid Preliminary results of convective probability forecast The size of the target grid The cloud top temperature, cloud layer thickness, convective region thickness and / or cloud microstructure of the cloud cluster are used to calculate the target grid point Preliminary results of convective probability forecast Perform segmented correction, including:

[0158] When the target grid Preliminary results of convective probability forecast satisfy When Whether the cloud cluster meets the following requirements at the same time: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud; when the judgment result shows that the target grid point The target grid point is determined if the cloud cluster satisfies the following conditions at the same time: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud. The cloud cluster is a convective cloud cluster, and the target grid point Preliminary results of convective probability forecast As the final result of the convection probability prediction Output;

[0159] When the target grid Preliminary results of convective probability forecast satisfy When The cloud cluster meets any two of the following four conditions: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud; when the judgment result shows that the target grid point If the cloud cluster does not meet any two of the following four conditions: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness of the cloud cluster meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the shape of the cloud cluster is a typical ice cloud, then the target grid point is judged The cloud cluster is a non-convective cloud cluster, and the target grid is output The final result of the convection probability prediction =0;

[0160] When the target grid Preliminary results of convective probability forecast satisfy When Whether the cloud cluster satisfies the following conditions at the same time: the cloud top temperature is less than 0 degrees Celsius and the cloud microstructure is a typical ice cloud; if the judgment result shows that the target grid point The cloud cluster satisfies: the cloud top temperature is less than 0 degrees Celsius and the cloud microstructure is a typical ice cloud, then the target grid point is judged The cloud cluster is a convective cloud cluster, and the target grid point Preliminary results of convective probability forecast As the final result of the convection probability prediction Output; otherwise, the target grid is determined The cloud cluster is a non-convective cloud cluster, and the target grid is output The final result of the convection probability prediction =0;

[0161] in: is a preset first convection probability threshold; is the preset second convection probability threshold; the first and second convection probability thresholds meet: and , In the specific implementation of the present invention, the first convection probability threshold ; Second convection probability threshold .

[0162] The convection occurrence probability prediction method based on the Fengyun geostationary satellite spectral imager described in the present invention is deployed on a server or a cloud computing platform to ensure the stability and reliability of computing resources.

[0163] Access the real-time observation data stream of the Fengyun geostationary satellite imager, including satellite observation data of infrared channel, water vapor channel and visible light channel.

[0164] After access, the real-time satellite observation data is processed, and the nearest neighbor interpolation is performed to generate 0.04 degree latitude and longitude grid data. The mean and standard deviation statistically calculated in the training data set are used to perform Z-Score normalization on the real-time satellite observation data to ensure data consistency. The pre-processed real-time satellite observation data is input into the convective probability prediction model to generate the predicted probability of convective clouds.

[0165] Each pixel point (regional grid point) output by the convective probability prediction model represents the probability of severe convection occurring within the next hour. The prediction results are presented in the form of images, contour lines or numerical values ​​for analysis and interpretation by meteorological researchers or users. Combined with other meteorological data, severe convective weather warning information is issued to guide the public and relevant departments to take disaster prevention and mitigation measures.

Claims

1. A method for predicting the probability of convection based on the Fengyun geostationary satellite spectral imager, characterized in that: The steps include: Step 1: Construct a convection probability prediction model; The convection probability prediction model is built based on a deep convolutional neural network and trained using a training dataset; The training data set includes historical satellite observation data and historical convection occurrence area label data that match time and space; The satellite historical observation data is the observation data of the key spectral channels of the spectral imager carried on the Fengyun geostationary satellite in a specified historical period; the key spectral channels of the spectral imager include infrared spectral channels, water vapor spectral channels and visible light spectral channels; The historical convection zone label data is generated based on the historical radar combined reflectivity data of the same period as the historical satellite observation data; The historical radar combined reflectivity data is the observation data of the laser radar in a specified historical period; Step 2: Output the preliminary results of the convection probability prediction in the target area; Focus the key spectral channels of the spectral imager on the target grid The real-time satellite observation data is preprocessed and then input into the convection probability prediction model to predict the target grid point. The convection probability of the cloud cluster is calculated and the target grid is output. Preliminary results of convective probability forecast ; Step 3: Correct the preliminary results of the convection probability prediction of the target grid: Based on the target grid Preliminary results of convective probability forecast , combined with the target grid The cloud top temperature, cloud layer thickness, convective region thickness and / or microstructure of the cloud cluster are used to further determine the target grid point. Is the cloud cluster a convective cloud cluster? When the judgment result shows that the target grid If the cloud cluster is a convective cloud cluster, the target grid Preliminary results of convective probability forecast As the final result of the convection probability prediction Output; otherwise, the target grid is output The final result of the convection probability prediction =0; In step 3, based on the target grid Preliminary results of convective probability forecast The size of the target grid The cloud top temperature, cloud layer thickness, convective region thickness and / or cloud microstructure of the cloud cluster are used to calculate the target grid point Preliminary results of convective probability forecast Perform segmented correction, including: When the target grid Preliminary results of convective probability forecast satisfy When Whether the cloud cluster meets the following requirements at the same time: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud; when the judgment result shows that the target grid point The target grid point is determined if the cloud cluster satisfies the following conditions at the same time: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud. The cloud cluster is a convective cloud cluster, and the target grid point Preliminary results of convective probability forecast As the final result of the convection probability prediction Output; When the target grid Preliminary results of convective probability forecast satisfy When The cloud cluster meets any two of the following four conditions: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the microstructure of the cloud cluster is a typical ice cloud; when the judgment result shows that the target grid point If the cloud cluster does not meet any two of the following four conditions: the cloud top temperature is less than 0 degrees Celsius, the cloud layer thickness of the cloud cluster meets the preset requirements, there is a convection area of ​​preset thickness in the cloud cluster, and the shape of the cloud cluster is a typical ice cloud, then the target grid point is judged The cloud cluster is a non-convective cloud cluster, and the target grid is output The final result of the convection probability prediction =0; When the target grid Preliminary results of convective probability forecast satisfy When Whether the cloud cluster satisfies the following conditions at the same time: the cloud top temperature is less than 0 degrees Celsius and the cloud microstructure is a typical ice cloud; if the judgment result shows that the target grid point The cloud cluster satisfies: the cloud top temperature is less than 0 degrees Celsius and the cloud microstructure is a typical ice cloud, then the target grid point is judged The cloud cluster is a convective cloud cluster, and the target grid point Preliminary results of convective probability forecast As the final result of the convection probability prediction Output; otherwise, the target grid is determined The cloud cluster is a non-convective cloud cluster, and the target grid is output The final result of the convection probability prediction =0; in: is a preset first convection probability threshold; is a preset second convection probability threshold; The first and second convection probability thresholds meet: and , .

2. The method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager according to claim 1 is characterized in that: In step 3, the target grid The cloud top temperature, cloud layer thickness, convective region thickness and / or microstructure of a cloud cluster are determined based on the brightness temperature output by a typical spectral channel of a spectral imager; The typical spectral channels of the spectral imager include three infrared spectral channels and one water vapor spectral channel; wherein: the three infrared spectral channels correspond to the infrared 10.8 micron channel, the infrared 12 micron channel and the infrared 8.55 micron channel; the one water vapor spectral channel is specifically the water vapor 6.25 micron channel; The cloud top temperature of the cloud cluster is output through the brightness temperature BT of the infrared 10.8 micron channel 10.8 To observe and when BT 10.8 <263K, indicating that the cloud top temperature is less than 0 degrees Celsius; The cloud thickness of the cloud cluster is calculated by calculating the brightness temperature difference between the water vapor 6.25 μm channel and the infrared 10.8 μm channel. 6.25-10.8 to observe and when BTD 6.25-10.8 When >-2K, it indicates that the cloud thickness of the cloud cluster has reached the preset thickness; The thickness of the convective region of the cloud cluster is calculated by calculating the brightness temperature difference between the infrared 12 μm channel and the infrared 10.8 μm channel. 12-10.8 To observe, and when -2K <BTD 10.8-12 When <2K, it indicates that there is a convection area of ​​preset thickness in the cloud; The cloud microstructure is determined by calculating the brightness temperature difference (BTD) between the infrared 8.55 μm channel and the infrared 10.8 μm channel. 8.55-10.8 to observe and when BTD 8.55-10.8 When >0, it indicates that the microstructure of the cloud is a typical ice cloud.

3. The method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager according to claim 1 is characterized in that: First convection probability threshold ; Second convection probability threshold .

4. The method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager according to claim 1 is characterized in that: In step 1, the training data set is constructed through the following steps: Step 1.1, obtaining the satellite historical observation data and historical radar combined reflectivity data of the specified historical period, and generating the convection history occurrence area label data of the specified historical period based on the historical radar combined reflectivity data; Satellite historical observation data comes from the observation data of key spectral channels of the Fengyun geostationary satellite imager in a specified historical period; The historical radar combined reflectivity data comes from the observation data of the lidar in a specified historical period; Step 1.2, time matching is performed on the acquired satellite historical observation data and the convection history occurrence area label data to ensure that the satellite historical observation data and the convection history occurrence area label data are at the same observation time; Step 1.3: spatially match the satellite historical observation data and the convection history occurrence area label data that have been time-matched to ensure that the satellite historical observation data and the convection history occurrence area label data are on the same spatial grid; Step 1.4, statistically calculate the mean and standard deviation of each key spectral channel observation data in the satellite historical observation data, and normalize the observation data of each key spectral channel; Through the above steps, the construction of the training data set can be completed.

5. The method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager according to claim 4 is characterized in that: In step 1.1, the generation of label data of the flow history falling area includes the following steps: Step 111, obtaining historical radar combined reflectivity data for a specified historical period; Step 112: Identify the convection occurrence area: Based on the historical radar combined reflectivity data obtained in step 111, the corresponding radar echo intensity information is extracted, and the convective cloud cluster areas with radar echo intensity information greater than 35dBZ and an area greater than 25 square kilometers are screened out and marked, so as to complete the identification of the convective occurrence area; The radar combined reflectivity data in the convection occurrence area are processed by nearest neighbor spatial fusion interpolation to make it consistent with the spatial resolution of the historical satellite observation data in the same period; Step 113: Construct label data of the convection occurrence area: The convection occurrence area label is the convection occurrence area label within the next 1 hour at any time t in the specified historical period, which is specifically constructed in the following way: first, the convection occurrence area within the next 1 hour at any time t is collected; then, the radar combined reflectivity data in the convection occurrence area within the next 1 hour at any time t is spatially merged by using the convective cloud cluster tracking algorithm, so as to generate the convection occurrence area label data within the next 1 hour at any time t; Step 114: Filter the historical satellite observation data of the same period: Based on the time t recorded in step 113, the satellite historical observation data two hours before the time t is screened, and the screened satellite historical observation data is interpolated by the nearest neighbor to be consistent with the spatial resolution of the radar combined reflectivity data.

6. The method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager according to claim 5 is characterized in that: In step 112, the nearest neighbor spatial fusion interpolation processing is specifically as follows: for each target area grid point, first collect all radar combined reflectivity data in an area with a radius of 0.02 degrees around the target area grid point, and then use the distance inverse weighted proportional method, that is, the weight of the target area grid point is 1, and the boundary weight of the radius 0.02 degrees is 0 for weighting, so that the radar combined reflectivity data of the 0.04 degree equal longitude and latitude grid can be fused and generated.

7. The method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager according to claim 5 is characterized in that: In step 2, the key spectral channel of the spectral imager is aimed at the target grid point The satellite real-time observation data is preprocessed through the following steps: Step 2.1, the satellite real-time observation data is processed by nearest neighbor spatial fusion interpolation to make the spatial resolution of the satellite real-time observation data consistent with the convection occurrence area label data; Step 2.2: Use the mean and standard deviation of the observation data of each key spectral channel in the satellite historical observation data counted in step 1.4 to normalize the satellite real-time observation data to ensure data consistency.

8. The method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager according to claim 1 is characterized in that: The convection probability prediction model includes an encoder, a decoder and a loss function, wherein: The input data of the encoder is the hourly satellite observation data of the target grid point 2 hours before any observation time T, with a total of 2 time steps: T-2h and T-1h; the encoder includes Layer, each layer of the encoder includes a first convolution layer, a first pooling layer and a first connection layer. In the first layer of the encoder, the first convolution layer is used to convolve the input data, and the first pooling layer is used to downsample the feature map output by the first convolution layer at the corresponding layer of the encoder and output it to the first connection layer; the first connection layer uses feature splicing to fuse the features of different time steps output by the first pooling layer at the corresponding layer of the encoder, and outputs the fused features to the next layer of the encoder, and so on, until it is completed in the encoder. Sub-convolution downsampling feature fusion, extracting high-level features and outputting them to the decoder; The decoder includes an upsampling layer, a second connection layer, a second convolutional layer and a Softmax activation function; the upsampling layer is used to upsample the high-level features output by the encoder and output them to the second connection layer; the second connection layer, through a jump connection, splices the feature map input to the second connection layer with the feature map of the corresponding scale in the encoder to obtain high- and low-level fusion features, and outputs them to the second convolutional layer; the second convolutional layer convolves the input high- and low-level fusion features and outputs them to the Softmax activation function; the Softmax activation function can calculate the preliminary result of the convection probability prediction of the target grid point according to the output of the second convolutional layer; The loss function is used to measure the difference between the preliminary convection probability prediction results output by the convection probability prediction model and the historical convection zone label data during the training process of the convection probability prediction model.

9. The method for predicting the probability of convection occurrence based on the Fengyun geostationary satellite spectral imager according to claim 8 is characterized in that: The output of the encoder is: ； ； ； Where: Indicates that the encoder has passed The high-level features output after the sub-convolution downsampling features are fused; Indicates the satellite observation data for time step T-2h, the encoder's The layer passes through the output of the first convolutional layer; Represents the satellite observation data for time step T-1h. The encoder The output of the layer; Indicates the satellite observation data for time step T-2h, the encoder's The layer passes through the output of the first pooling layer; Indicates the satellite observation data for time step T-1h, the encoder The layer passes through the output of the first pooling layer; BN means batch normalization; ReLU is the activation function; * indicates the convolution operation; For encoder Bias parameters for layer outputs; For encoder The convolution kernel of the layer output; hour, Indicates that the data input to the first convolutional layer is the satellite observation data of time step T-2h, Indicates that the data input to the first convolutional layer is the satellite observation data of time step T-1h; In the decoder, the output of the upsampling layer, the second connection layer, and the Softmax activation function is: ； ； ； Where: represents the output of the upsampling layer, ; represents the output of the second connection layer; Represents the output of the encoder and upsampling layer Feature map of corresponding scale; It represents the preliminary result of the convection probability prediction of the target grid point; The convolution kernel representing the decoder output; Represents the bias parameter of the decoder output; when hour, represents the output of the encoder; The loss function used in the training process of the convection probability prediction model is the weighted cross entropy loss function, expressed as follows: ； Where: is the total number of grid points in the region; For the The convection occurrence domain label of each grid point; The output of the convection probability prediction model is The probability of convection occurring at each grid point; is the positive sample weight coefficient.

Images