Lightning nowcasting method based on multi-source satellite data

Abstract

The invention relates to a lightning nowcasting method based on multi-source satellite data, and belongs to the technical field of lightning nowcasting. The method comprises the steps of obtaining and preprocessing satellite data of a multi-source channel and corresponding lightning data, constructing a U-Net lightning prediction model, collecting a historical multi-source satellite image sequence, constructing a PredRNN satellite image prediction model and predicting the lightning occurrence probability. According to the method, the PredRNN and the U-Net model are combined, the lightning distribution can be predicted at the future moment only by using the satellite image, the dependence on historical lightning data is broken through, and the method has a wide application prospect in the field of lightning nowcasting.

Description

Technical Field

[0001] The invention belongs to the technical field of lightning nowcasting, and in particular relates to a lightning nowcasting method based on multi-source satellite data. Background Art

[0002] Traditional means of lightning nowcasting are generally based on physical models, radar, and ground lightning network detection technology. Commonly used prediction methods rely on meteorological numerical prediction models, and infer the area where lightning may occur by simulating atmospheric dynamics and thermodynamic processes. The main principle is to predict lightning based on changes in meteorological elements. However, the required calculation complexity is high, the process is time-consuming, and requires a lot of computing resources. In addition, lightning occurs quickly, and sometimes has nonlinearity and non-stationarity. It performs poorly in local and short-term situations. In general, the prediction effect of traditional lightning nowcasting methods is unstable. In recent years, with the development of deep learning, deep learning methods have been gradually adopted for lightning nowcasting. It mainly extracts data time series information and spatial feature technology, and extracts features based on multi-source data such as satellite images, radar reflectivity, and lightning data to give future lightning probability predictions, making lightning forecasts more accurate. These methods rely on radar and historical lightning monitoring data, and fail to make full use of satellite data to explore the relationship between cloud changes and lightning events. It is difficult to predict future lightning through the spatiotemporal extrapolation of satellite images. At the same time, the problem of no lightning detection equipment in remote areas is not taken into account, and the model has limitations.

[0003] In addition, the number of lightning events in a certain area is relatively small, and areas without lightning occupy most of the space. Conventionally used loss functions, such as cross entropy loss, will cause the model to ignore the sparse "lightning" category because most pixels are in the "no lightning" category, resulting in the model easily outputting high-precision "no lightning" predictions and failing to identify "lightning". Therefore, the existing lightning forecasting model based on deep learning lacks consideration of category imbalance in the loss function, resulting in insufficient prediction accuracy of the model for lightning events.

[0004] Based on this, how to overcome the shortcomings of existing technologies is an urgent problem to be solved in the field of lightning nowcasting technology. Summary of the invention

[0005] The purpose of the present invention is to solve the deficiencies of the prior art and to provide a lightning nowcasting method based on multi-source satellite data, which can accurately predict future lightning distribution without relying on historical lightning data.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A lightning nowcasting method based on multi-source satellite data comprises the following steps:

[0008] Step (1): Select satellite data from multiple sources and corresponding lightning data, align their spatio-temporal resolutions, and then preprocess the satellite data and the corresponding lightning data. Use the preprocessed data as the dataset for training and predicting the U-Net model.

[0009] Step (2): Divide the dataset obtained in step (1) into a training set and a test set. Use the preprocessed satellite data in step (1) as the input and the preprocessed lightning data in step (1) as the output. Train the U-Net model using the training set obtained in step 1 and test it using the test set obtained in step (1) to obtain the U-Net lightning prediction model.

[0010] Step (3): Collect historical satellite data from multiple sources, preprocess it to form a dataset, and divide it into a training set and a test set.

[0011] Step (4): Use the satellite image sequence of a certain time period as the input and the satellite image sequence of the next time period as the output. Train the PredRNN model using the training set obtained in step (3) and test it using the test set obtained in step (3) to obtain the PredRNN satellite image prediction model.

[0012] Step (5): Collect the real-time multi-source satellite image sequence of the current time period and output it to the PredRNN satellite image prediction model obtained in step (4) to obtain the satellite image of the next time period. Then, after performing the same preprocessing as in step (1) on the satellite image of the next time period, input it into the U-Net lightning prediction model obtained in step (2) to obtain the lightning occurrence probability at each location in the next time period.

[0013] Furthermore, preferably, the specific method of step (1) is as follows:

[0014] Step A: Select the infrared channels and the calculation channels representing the brightness temperature difference in the satellite data that are closely related to convective weather and lightning formation. Divide the data of each channel according to the longitude and latitude range of the area for lightning nowcasting to generate satellite images of the corresponding area, which are used as the input of the U-Net model.

[0015] Step B: Based on the lightning data collected by the ground-based wide-area lightning location network, adjust the time resolution and spatial resolution of the lightning data to be consistent with those of the satellite data, and grid the lightning data to generate pictures, which are used as the labels of the U-Net model.

[0016] Furthermore, preferably, in step A, the infrared channels are channels 8, 9, 10, and 13, and their λ values are 6.2μm, 6.9μm, 7.3μm, and 10.4μm, respectively.

[0017] Calculate the four channels of BTD0910, BTD1013, BTD1513, and TTD;

[0018] Among them, the brightness temperature difference of the BTD0910 calculation channel is the difference between the brightness temperature of channel 9 and the brightness temperature of channel 10;

[0019] The brightness temperature difference of the BTD1013 calculation channel is the difference between the brightness temperature of channel 10 and the brightness temperature of channel 13;

[0020] The brightness temperature difference of the BTD1513 calculation channel is the difference between the brightness temperature of channel 15 and the brightness temperature of channel 13; among them, λ of channel 15 = 12.4μm;

[0021] The brightness temperature difference of the TTD calculation channel = (brightness temperature of channel 11 - brightness temperature of channel 13) - (brightness temperature of channel 13 - brightness temperature of channel 15), where λ of channel 11 = 8.6μm.

[0022] Furthermore, preferably, in step B, the spatial resolution of the lightning data is 0.05°, and the time resolution is 10 minutes;

[0023] The specific method of data grid processing is as follows:

[0024] Generate a lightning image every 10 minutes, and this image contains all lightning events within this time period; each lightning event is mapped to the grid generated according to the longitude and latitude division according to its longitude and latitude coordinates, and the position of each grid point is represented by coordinates (x, y); for each (x, y) pixel position, judge whether a lightning event occurs within this time period. If there is at least one lightning event, set its value to 1, otherwise 0, to generate a binary matrix L(x, y), which specifically satisfies the following relationship:

[0025]

[0026] Among them, i is the number of lightning events occurring within this time window.

[0027] Furthermore, preferably, in the grid, the size of each grid is 5km × 5km.

[0028] Furthermore, preferably, in step (2), the specific structure of the U-Net model includes an encoder, a decoder, and an output layer;

[0029] The encoder includes 5 convolutional blocks, and each convolutional block consists of two 3×3 convolutions + batch normalization + ReLU + Dropout; a max pooling operation is connected after each convolutional block;

[0030] The input of the encoder is an 8-channel image;

[0031] The described decoder starts from the last layer of the encoder and gradually restores the spatial resolution through upsampling operations; in each upsampling block, the feature map of the skip connection is concatenated with the upsampled feature map. After each upsampling, a convolutional block is used to integrate the channel information;

[0032] The described output layer is a 1×1 convolution, which is used to convert the feature map into 1 channel for output.

[0033] Further, preferably, in step (2), when training the U-Net model, a weighted cross-entropy loss function is used for model training;

[0034]

[0035] where y i is the true label, indicating whether a lightning event occurs or not. When a lightning occurs, y i has a value of 1; when no lightning occurs, y i has a value of 1; is the probability predicted by the U-Net model that a lightning occurs; W i represents the weight. When a lightning event occurs, that is, when the true label is lightning, W i is W lightning , when no lightning event occurs, that is, when the label is no lightning, W i is W no-lightning ; W lightning <W no-linghtning ; W lightaning +W no-lightning =1;

[0036] Calculate the loss function according to the loss function formula in Equation (1) and perform backpropagation. By minimizing the loss function through multiple iterations, the U-Net lightning prediction model is obtained.

[0037] Further, preferably, 0.7 ≤ W lightning ≤ 0.99, 0.01 ≤ W no-lightning ≤ 0.3.

[0038] Further, preferably, in step (1), the sample ratio of the training set to the test set is 9:1; in step (3), the sample ratio of the training set to the test set is 7:3;

[0039] In step (3), the specific method of preprocessing is: extract the brightness temperature data of the target area from the historical multi-source satellite data, map all the data to [0,1] through normalization, crop it according to the required longitude and latitude range, and then convert it into an image format for storage to form a data set.

[0040] Furthermore, preferably, in step (4), the PredRNN model is trained using the satellite image sequence of the past one hour as the input and the satellite image sequence of the next one hour as the output, and is optimized using the MSE loss and the decoupling loss; during training, the MSE loss and the decoupling loss are fused by summation, and the two are added to calculate the total loss, and training is performed with the goal of minimizing the total loss.

[0041] The ground-based wide-area lightning location network described in the present invention is an existing one, and the present invention does not make special limitations on this.

[0042] The U-Net model of the present invention extracts the cloud morphology, temperature distribution, cloud thickness, and cloud top height features in the satellite image through the convolutional layer, and gradually extracts high-level global features through downsampling (convolution + pooling). In the upsampling stage, the transposed convolutional layer is used to gradually restore the spatial resolution of the image, and at the same time, the detailed features retained during the downsampling process are fused with the upsampling features through the skip connection mechanism to ensure that the model can retain both the global features and restore the local detailed information of the input image when outputting. Finally, the lightning occurrence probability map of each pixel is output.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] (1) No longer relying on historical lightning data: The PredRNN model used in the present invention can extrapolate the future cloud state based on past satellite images. The U-Net model makes lightning predictions based on the relationship between satellite images and lightning events on the basis of cloud features. Compared with the prior art, the present invention no longer relies on input historical lightning monitoring data and radar reflectivity data, and directly infers the future lightning distribution from past satellite observation information. It overcomes the technical bottleneck of insufficient coverage of the ground lightning monitoring network in traditional methods and expands the applicable range of lightning prediction;

[0045] (2) Handling the class imbalance problem and improving the prediction accuracy of lightning sparse data: By introducing the weighted cross-entropy loss function and assigning higher weights to lightning events, the model can better learn the features of these key regions, the prediction accuracy of lightning events is improved, and the model is more stable; compared with using the traditional cross-entropy loss function, the MAR (miss rate) of the samples used is reduced from 44.67% to 8.11%.

[0046] (3) Spatiotemporal extrapolation of multi-source satellite data: By extrapolating future multi-source satellite images, the present invention can predict the future lightning distribution in advance based on the changes in cloud layers before lightning occurs. This spatiotemporal extrapolation technology provides a longer prediction window for real-time lightning warning and improves the timeliness of the warning. Compared with the prediction results of traditional lightning nowcasting models, the POD (detection rate) reaches 91.89%, the MAR (miss rate) is 8.11%, and the TS (threat score) is 67.3%, all showing better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic flowchart of the lightning nowcasting method based on multi-source satellite data of the present invention;

[0048] Figure 2 It is an example of the satellite image used in the present invention;

[0049] Figure 3 It is an example of the lightning binarized image used in the present invention; among them, the images at four times of 2:30, 2:40, 2:50, and 3:00 on June 1, 2023 are respectively shown;

[0050] Figure 4 It is the U-Net model structure used in the present invention;

[0051] Figure 5 It is the PredRNN model structure used in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0052] The present invention will be further described in detail below in conjunction with the embodiments.

[0053] Those skilled in the art will understand that the following embodiments are only used to illustrate the present invention and should not be construed as limiting the scope of the present invention. For those not specified in the embodiments regarding specific technologies or conditions, they shall be carried out according to the technologies or conditions described in the literature in this field or according to the product specifications. For those materials or equipment without indicating the manufacturer, they are all conventional products that can be obtained by purchase.

[0054] Embodiment 1

[0055] A lightning nowcasting method based on multi-source satellite data includes the following steps:

[0056] Step (1), select satellite data of multi-source channels and corresponding lightning data, align their spatiotemporal resolutions, and then preprocess the satellite data and the corresponding lightning data, and use the preprocessed data as the dataset for training and predicting the U-Net model;

[0057] Step (2): Divide the dataset obtained in step (1) into a training set and a test set. Using the preprocessed satellite data in step (1) as the input and the preprocessed lightning data in step (1) as the output, train the U-Net model with the training set obtained in step 1 and test it with the test set obtained in step (1) to obtain the U-Net lightning prediction model.

[0058] Step (3): Collect historical multi-source satellite data, preprocess it to form a dataset, and divide it into a training set and a test set.

[0059] Step (4): Using a sequence of satellite images in a certain period as the input and a sequence of satellite images in the next period as the output, train the PredRNN model with the training set obtained in step (3) and test it with the test set obtained in step (3) to obtain the PredRNN satellite image prediction model.

[0060] Step (5): Real-time collect a sequence of multi-source satellite images in the current period, output it to the PredRNN satellite image prediction model obtained in step (4) to get the satellite images in the next period. Then, after performing the same preprocessing as in step (1) on the satellite images in the next period, input them into the U-Net lightning prediction model obtained in step (2) to obtain the lightning occurrence probability at each location in the next period.

[0061] Embodiment 2

[0062] A lightning nowcasting method based on multi-source satellite data, comprising the following steps:

[0063] Step (1): Select multi-source channel satellite data and corresponding lightning data, align their spatio-temporal resolutions, and then preprocess the satellite data and the corresponding lightning data. Use the preprocessed data as the dataset for U-Net model training and prediction.

[0064] Step (2): Divide the dataset obtained in step (1) into a training set and a test set. Using the preprocessed satellite data in step (1) as the input and the preprocessed lightning data in step (1) as the output, train the U-Net model with the training set obtained in step 1 and test it with the test set obtained in step (1) to obtain the U-Net lightning prediction model.

[0065] Step (3): Collect historical multi-source satellite data, preprocess it to form a dataset, and divide it into a training set and a test set.

[0066] Step (4): Use the satellite image sequences of a certain period as the input, and the satellite image sequences of the next period as the output. Train the PredRNN model using the training set obtained in step (3), and test it using the test set obtained in step (3) to obtain the PredRNN satellite image prediction model.

[0067] Step (5): Collect the multi-source satellite image sequences of the current period in real time and output them to the PredRNN satellite image prediction model obtained in step (4) to obtain the satellite images of the next period. Then, after performing the same preprocessing on the satellite images of the next period as in step (1), input them into the U-Net lightning prediction model obtained in step (2) to obtain the lightning occurrence probability at each location in the next period.

[0068] The specific method of step (1) is as follows:

[0069] Step A: Select the infrared channels and the calculation channels representing the brightness temperature difference in the satellite data that are closely related to convective weather and lightning formation. Divide the data of each channel according to the longitude and latitude range of the area where lightning nowcasting is to be performed to generate satellite images of the corresponding area as the input of the U-Net model.

[0070] Step B: Based on the lightning data collected by the ground-based wide-area lightning location network, adjust the time resolution and spatial resolution of the lightning data to be consistent with those of the satellite data, and grid the lightning data to generate pictures as the labels of the U-Net model.

[0071] In step A, select the four channels of channel 8, 9, 10, and 13 for the infrared channels; their λ values are 6.2μm, 6.9μm, 7.3μm, and 10.4μm in sequence.

[0072] Select the four channels of BTD0910, BTD1013, BTD1513, and TTD for the calculation channels;

[0073] Among them, the brightness temperature difference of the BTD0910 calculation channel is the difference between the brightness temperature of channel 9 and the brightness temperature of channel 10;

[0074] The brightness temperature difference of the BTD1013 calculation channel is the difference between the brightness temperature of channel 10 and the brightness temperature of channel 13;

[0075] The brightness temperature difference of the BTD1513 calculation channel is the difference between the brightness temperature of channel 15 and the brightness temperature of channel 13; among them, λ of channel 15 = 12.4μm;

[0076] The brightness temperature difference of the TTD calculation channel = (brightness temperature of channel 11 - brightness temperature of channel 13) - (brightness temperature of channel 13 - brightness temperature of channel 15), where λ of channel 11 = 8.6μm.

[0077] In step B, the spatial resolution of the lightning data is 0.05°, and the time resolution is 10 minutes;

[0078] The specific method for data gridding is as follows:

[0079] A lightning image is generated every 10 minutes, and this image contains all lightning events within this time period; each lightning event is mapped to the grid generated according to longitude and latitude based on its longitude and latitude coordinates, and the position of each grid point is represented by coordinates (x, y); for each (x, y) pixel position, it is judged whether a lightning event occurs within this time period. If there is at least one lightning event, its value is set to 1, otherwise it is 0, generating a binary matrix L(x, y), which specifically satisfies the following relationship:

[0080]

[0081] where i is the number of lightning events occurring within this time window.

[0082] In the grid, the size of each grid is 5km × 5km.

[0083] In step (2), the specific structure of the U-Net model includes an encoder, a decoder, and an output layer;

[0084] The encoder mentioned above includes 5 convolutional blocks, and each convolutional block consists of two 3×3 convolutions + batch normalization + ReLU + Dropout; a max pooling operation is performed after each convolutional block;

[0085] The input of the encoder is an 8-channel image;

[0086] The decoder starts from the last layer of the encoder and gradually restores the spatial resolution through upsampling operations; in each upsampling block, the feature map of the skip connection is concatenated with the upsampled feature map. After each upsampling, a convolutional block is used to integrate the channel information;

[0087] The output layer is a 1×1 convolution, which is used to convert the feature map into 1 channel for output.

[0088] In step (2), when training the U-Net model, a weighted cross-entropy loss function is used for model training;

[0089]

[0090] where y i is the true label, indicating whether a lightning event occurs or not. When a lightning occurs, the value of y i is 1; when no lightning occurs, the value of y i is 1; is the probability predicted by the U-Net model that a lightning occurs; Wi Denotes the weight. When a lightning event occurs, i.e., the true label is lightning, W i is W lightning , when no lightning event occurs, i.e., the label is no lightning, W i is W no-lightning ; W lightning <W no-lightning ; W lightning +W no-lightning = 1;

[0091] Calculate the loss function according to the loss function formula in Equation (1) and perform backpropagation. Minimize the loss function through multiple iterations to obtain the U-Net lightning prediction model.

[0092] 0.7 ≤ W lightning ≤ 0.99, 0.01 ≤ W no-lightning ≤ 0.3.

[0093] In step (1), the sample ratio of the training set to the test set is 9:1; in step (3), the sample ratio of the training set to the test set is 7:3;

[0094] In step (3), the specific method of preprocessing is as follows: Extract the brightness temperature data of the target area from historical multi-source satellite data, map all the data to [0,1] through normalization, crop it according to the required longitude and latitude range, and then convert it into an image format for storage to form a data set.

[0095] In step (4), use the satellite image sequence of the past hour as the input and the satellite image sequence of the next hour as the output to train the PredRNN model, and optimize it using the MSE loss and the decoupled loss; during training, fuse the MSE loss and the decoupled loss by summation, add the two to calculate the total loss, and train with the goal of minimizing the total loss.

[0096] Application Example

[0097] As Figure 1 shown, this example discloses a lightning nowcasting method based on multi-source satellite data, including:

[0098] Step (1), select the satellite data of multi-source channels and the corresponding lightning data, and align their spatio-temporal resolutions as the data set for training and predicting the U-Net model;

[0099] Step (1) specifically includes:

[0100] Step A: Select the infrared channels and the calculation channels representing the brightness temperature difference in the satellite data that are closely related to convective weather and lightning formation. Divide the data of each channel according to the longitude and latitude range of the area where lightning nowcasting is to be carried out (one map for each channel and each time step, and one time step is 10 minutes, that is, one map for the same channel every 10 minutes, or 8 maps for each time step), and generate satellite images of the corresponding area as the input of the U-Net model;

[0101] In Step A, the infrared channels selected are channels 8, 9, 10, and 13; their λ values are 6.2μm, 6.9μm, 7.3μm, and 10.4μm respectively;

[0102] The calculation channels selected are BTD0910, BTD1013, BTD1513, and TTD;

[0103] Among them, the brightness temperature difference of the BTD0910 calculation channel is the difference between the brightness temperature of channel 9 and the brightness temperature of channel 10;

[0104] The brightness temperature difference of the BTD1013 calculation channel is the difference between the brightness temperature of channel 10 and the brightness temperature of channel 13;

[0105] The brightness temperature difference of the BTD1513 calculation channel is the difference between the brightness temperature of channel 15 and the brightness temperature of channel 13; among them, the λ of channel 15 = 12.4μm;

[0106] The brightness temperature difference of the TTD calculation channel = (brightness temperature of channel 11 - brightness temperature of channel 13) - (brightness temperature of channel 13 - brightness temperature of channel 15), where the λ of channel 11 = 8.6μm.

[0107] The specifically selected satellite images are the eight-channel satellite image data of the Pearl River Delta region from June to October 2023, one picture every 10 minutes, and the spatial resolution is 0.05°. As Figure 2As shown in the figure, in the infrared channel, four channels of band8 (λ = 6.2μm), band9 (λ = 6.9μm), band10 (λ = 7.3μm), and band13 (λ = 10.4μm) are selected, which respectively represent the upper-layer water vapor, middle-layer water vapor, lower-layer water vapor, and cloud top height in the troposphere. The calculation channel selects BTD0910 (brightness temperature difference between band9 and band10), BTD1013 (brightness temperature difference between band10 and 13), BTD1513 (brightness temperature difference between band15 (λ = 12.4μm) and band13), and TTD (brightness temperature difference between band11 (λ = 8.6μm) and band13 minus the brightness temperature difference between band13 and band15), which respectively represent the cloud top height relative to the troposphere, the cloud top height relative to the bottom layer of the troposphere, cloud optical thickness, and cloud top phase state. These reflect the key meteorological parameters in the cloud layer, and these parameters are closely related to the formation process of lightning.

[0108] Step B: Based on the data collected by the ground-based wide-area lightning location network, align the spatio-temporal resolution of the lightning data with that of the satellite data, that is, the spatial resolution of the lightning data is 0.05°, and the time resolution is 10 minutes; and grid the data as the basic label data for the training of the U-Net model.

[0109] Step B specifically includes:

[0110] Select the lightning data in the Pearl River Delta region, set a spatial resolution of 0.05°, covering an 80×80 pixel area near the central coordinates [113.5, 23.4], so that its spatio-temporal resolution is aligned with that of the satellite data. Each lightning event is mapped to a grid with a size of 5km×5km according to its longitude and latitude coordinates to form a standardized input format.

[0111] Generate a lightning picture every 10 minutes, which contains all lightning events within this time period; each lightning event is mapped to the grid generated according to the longitude and latitude division, and the position of each grid point is represented by coordinates (x, y); for each (x, y) pixel position, judge whether there is a lightning event within this time window. If there is at least one lightning event, set its value to 1, otherwise to 0, to generate a binary matrix L(x, y), which satisfies the following relationship:

[0112]

[0113] where i is the number of lightning events occurring within this time window.

[0114] As Figure 3 shown, the generated binary lightning picture is used as the corresponding label for the model input.

[0115] Step (2): Divide the dataset obtained in step (1) into a training set and a test set; use the preprocessed satellite data in step (1) as the input and the preprocessed lightning data in step (1) as the output, train the U-Net model using the training set obtained in step 1, and test it using the test set obtained in step (1) to obtain the U-Net lightning prediction model;

[0116] Step (2) specifically includes: Divide the dataset collected and processed in step 1 according to a certain ratio. In this example, the training set and the test set are divided in a ratio of 9:1. Build a U-net model for training to obtain model weights. This model can process satellite images through a multi-layer convolutional and deconvolutional structure to extract features such as cloud morphology, temperature distribution, cloud thickness, and cloud top height;

[0117] The specific structure of U-Net is as follows:

[0118] The entire U-Net model is an encoder-decoder structure, which establishes a connection between feature extraction and spatial information recovery through skip connections, as Figure 4 shown.

[0119] Encoder part: The input is an eight-channel image with a shape of (batch_size, 8, 80, 80). The encoder consists of 5 convolutional blocks, each of which is composed of two 3×3 convolutions + batch normalization + ReLU + Dropout, gradually extracting higher-level features. After each convolutional block, a max pooling operation is performed, halving the resolution and increasing the number of channels. The number of feature channels changes sequentially as: 8 → 64 → 128 → 256 → 512 → 1024.

[0120] Decoder part: Starting from the last layer of the encoder, gradually restore the spatial resolution through upsampling operations. The resolution doubles each time, and the number of feature channels gradually decreases: 1024 → 512 → 256 → 128 → 64. In each upsampling block, the feature map of the skip connection is concatenated with the upsampled feature map. After each upsampling, a convolutional block is used to integrate the channel information.

[0121] The last layer converts the feature map from 64 channels to 1 channel through a 1×1 convolution, and the output shape is (batch_size, 1, 80, 80). The input of the entire model is an eight-channel satellite image of 80×80, and the output result maintains the original size with 1 channel. It represents the predicted lightning occurrence probability distribution.

[0122] In the training process of the U-Net model, a weighted cross-entropy loss function is introduced to make the model pay more attention to the sparse lightning areas during the training process.

[0123] Specifically includes:

[0124] To address the class imbalance issue, the present invention assigns higher weights to the pixels where lightning occurs. Specifically, lightning events are rare but crucial, so higher weights need to be assigned to these sparse "lightning pixels" in the loss calculation to ensure that the model pays more attention to these areas. The weighted version of the loss function can be expressed as:

[0125]

[0126] where y i is the ground truth label, indicating whether a lightning event occurs (value 1) or not (value 0), and is the predicted probability of the model. W i represents the weight. When a lightning event occurs, i.e., the ground truth label is lightning, W i is W lightning . When no lightning event occurs, i.e., the label is no lightning, W i is W no-lightning .

[0127] Among them, the weights are used to balance the losses between lightning occurrence (positive class) and non-occurrence (negative class). Usually, a larger weight W lightning is assigned to lightning events (positive class), while a smaller weight W no-lightning is assigned to non-lightning events (negative class), that is, W lightning < W no-lightning .

[0128] The selection of the weights is based on the class distribution of the lightning data. The value range of W lightning is 0.7 ≤ W lightning ≤ 0.99, and the value range of W no-lightning is 0.01 ≤ W no-lightning ≤ 0.3. By calculating the proportion of pixels where lightning occurs in the dataset, the present invention can dynamically adjust the weight values. However, if the value of W lightning is too large, it may cause W no-lightning to be too low, affecting the model stability. When lightning events are extremely sparse (proportion ≤ 1%), it is recommended to set W lightning to 0.99, and always keep W lightning + W no-lightning = 1.

[0129] In this example, if lightning events account for 1% of the total pixels, the positive class weight can be set to 99 times that of the negative class, then W no-lightning = 0.01, W lightning = 0.99, to balance the attention of the loss function to the positive and negative classes.

[0130] Case 1: When the ground truth label is lightning (y = 1), the U-Net model predicts the probability of it being lightning as 0.2:

[0131] Loss = -0.99(1·log(0.2) + 0·log(1 - 0.2)) = 0.692

[0132] Case 2: When the true label is no lightning (y = 0), the probability that the U-Net model predicts it as having lightning is 0.1:

[0133] Loss = -0.01(0·log(0.1) + 1·log(1 - 0.1)) = 0.00046

[0134] According to the results, the loss value in Case 1 is 0.693, indicating that the model's prediction for the positive class (having lightning) is very poor. In Case 2, the loss value is approximately 0.00046, indicating that the model's prediction for the negative class (no lightning) is good. Therefore, the overall lightning prediction effect of the model is not good, and the model needs to be further learned and optimized.

[0135] During the backpropagation process of model training, the model can adjust the convolutional kernel weights and biases of each layer so that the model's prediction can gradually approach the true value at the pixel points of sparse lightning events.

[0136] Step (3), collect historical multi-source satellite image sequences, perform preprocessing to form a dataset, and divide it into a training set and a test set;

[0137] The specific method of preprocessing is: extract the brightness temperature data of the target area from historical multi-source satellite data, map all the data to [0, 1] through normalization, crop it according to the required longitude and latitude range, and then convert it into an image format for storage to form a dataset.

[0138] The sample ratio of the training set to the test set is 7:3.

[0139] Step (4), use the satellite image sequence of a certain time period as the input and the satellite image sequence of the next time period as the output, train the PredRNN model using the training set obtained in step (3), and test it using the test set obtained in step (3) to obtain the PredRNN satellite image prediction model;

[0140] The PredRNN model simultaneously captures the spatial features and temporal dynamics in the satellite image through a spatio-temporal convolutional unit (referred to as spatio-temporal LSTM, abbreviated as ST-LSTM), learns the evolution trend of clouds over time, and outputs the satellite image at the future moment.

[0141] The ConvLSTM model structure is commonly used for spatio-temporal prediction. The PredRNN model takes into account the shortcoming of the ConvLSTM model that it ignores the features extracted from the top layer of the previous sequence, and improves its structure, making it have better performance in image sequence prediction. The neural unit of PredRNN is specifically as Figure 5 shown; in which the network is set with 4 spatio-temporal LSTM layers, X t is the input image of the network model at time t, is the output image of the network at time t, is the hidden layer state of the i-th neuron of the network at time t, is the output tensor of the i-th layer of the network at time t. The specific formula is as follows:

[0142]

[0143] Among them, g t is the abstract extraction process of the PredRNN neural unit for the input information at the current time and the input information at the previous time. Among them, X t l {l＝1} is the image input to the first layer of the PredRNN network at time t, W xg is the convolution kernel of the abstract information extraction layer for the input image at the current time, W hg is the convolution kernel of the abstract information extraction layer for the output image at the previous time, b g is the bias; i t is the input gate, which controls how much of the current input information is written into the unit. W xi , W hi , W ci are the convolution kernels of the input gate for the input image at the current time, the input image at the previous time, and the hidden layer state of the network at the previous time respectively. b i is the input gate bias; f t is the forget gate, which determines which information in the unit needs to be discarded. W xf , W hf , W cf are the convolution kernels of the forget gate for the input image at the current time, the input image at the previous time, and the hidden layer state of the network at the previous time respectively. b f is the forget gate bias; o t is the output gate, which controls which information is output from the unit to the next time. W xo , W ho , W co are the convolution kernels of the output gate for the input image at the current time, the input image at the previous time, and the hidden layer state of the network at the previous time respectively. b o is the output gate bias. i t , f t , o tIt is the core part of the spatio-temporal LSTM cell. The symbol * represents the convolution operation, and the symbol ⊙ represents the dot product operation.

[0144] Figure 5 The model in can infer a sequence after a period of time by inputting a sequence. For example, input X t-1 、X t 、X t+1 of the true value, and output X t 、X t+1 、X t+2 of the predicted value.

[0145] The model combines a convolutional neural network and a recurrent neural network, adopts a multi-level recursive structure, and processes the input sequence through a spatio-temporal feature fusion branch. The model contains multiple spatio-temporal convolutional units, which gradually transfer the states and memories of the outputs of each layer while extracting spatio-temporal features. Finally, the spatio-temporal features are extracted through the spatio-temporal LSTM layer, and the training is completed by optimizing the MSE loss and the decoupling loss.

[0146] The training is completed by optimizing the MSE loss and the decoupling loss. Specifically, the two losses are fused by summation, and the total loss is calculated by adding the two, and the training is carried out with the goal of minimizing the total loss.

[0147] Step (5), real-time collect the multi-source satellite image sequence of the current period, output it to the PredRNN satellite image prediction model obtained in step (4) to obtain the satellite image of the next period; then preprocess the satellite image of the next period in the same way as in step (1), and input it into the U-Ne t lightning prediction model obtained in step (2), so as to obtain the lightning occurrence probability at each position in the next period.

[0148] The PredRNN and U-Net models are combined to construct a complete spatio-temporal lightning prediction system. Step 5 specifically includes:

[0149] Step a, use the PredRNN model to perform spatio-temporal extrapolation on the satellite images of past multiple moments to generate the satellite images of future moments.

[0150] In this example, the trained model weights are used, and the parameters are set to extrapolate the satellite images of the future hour from the past hour, that is, extrapolate the future six images from the past six images.

[0151] Step b, input the satellite images of future moments generated by PredRNN into the U-Net model to extract the spatial features of the clouds and generate the future lightning distribution map.

[0152] In this example, the six future multi-source satellite images generated in step a are input into the U-Net network model, and the network outputs six corresponding lightning probability prediction pictures.

[0153] Step c, finally predict future lightning based on past satellite images, and the output lightning distribution map contains the lightning occurrence probability at each future location.

[0154] To evaluate the lightning nowcasting effect of the present invention, this example conducted a test experiment based on the data of the Pearl River Delta in July 2024. The evaluation indicators used were POD (detection probability, that is, among the points where lightning actually exists, what proportion is correctly predicted as lightning by the model), MAR (false negative rate, that is, among the points where lightning actually exists, the proportion that the model fails to correctly predict), and TS (threat score, considering the impacts of both false negatives and false positives, measuring the overall ability of the model to predict lightning). The prediction effect of the traditional numerical weather prediction model was: POD was between 70% and 90%, MAR was between 10% and 30%, and TS was between 20% and 40%. The evaluation indicators of the lightning nowcasting provided by the present invention in the sample data used were: POD = 91.89%, MAR = 8.11%, TS = 67.3%. Compared with the traditional numerical model, the prediction performance of the present invention has been significantly improved.

[0155] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A lightning nowcasting method based on multi-source satellite data, characterized in that: The steps include: Step (1), selecting satellite data and corresponding lightning data from multiple source channels, aligning their temporal and spatial resolutions, and then preprocessing the satellite data and the corresponding lightning data, and using the preprocessed data as the data set for U-Net model training and prediction; Step (2), dividing the data set obtained in step (1) into a training set and a test set; Using the satellite data preprocessed in step (1) as input and the lightning data preprocessed in step (1) as output, the U-Net model is trained using the training set obtained in step 1, and the test set obtained in step (1) is used for testing to obtain a U-Net lightning prediction model; Step (3), collecting historical multi-source satellite data, preprocessing to form a data set, and dividing it into a training set and a test set; Step (4), taking the satellite image sequence of a certain period as input and the satellite image sequence of the next period as output, using the training set obtained in step (3) to train the PredRNN model, and using the test set obtained in step (3) to test, to obtain the PredRNN satellite image prediction model; Step (5), real-time acquisition of a multi-source satellite image sequence of the current period, output to the PredRNN satellite image prediction model obtained in step (4), and obtain the satellite image of the next period; The satellite image of the next period is then preprocessed in the same way as in step (1) and input into the U-Net lightning prediction model obtained in step (2) to obtain the probability of lightning occurrence at each location in the next period.

2. The lightning nowcasting method based on multi-source satellite data according to claim 1, characterized in that: The specific method of step (1) is: Step A, selecting the infrared channel closely related to convective weather and lightning formation and the calculation channel representing the brightness temperature difference in the satellite data, dividing the data of each channel according to the latitude and longitude range of the area where lightning nowcasting is to be performed, and generating satellite images of the corresponding area as the input of the U-Net model; Step B, based on the lightning data collected by the ground-based wide-area lightning location network, adjust the time resolution and spatial resolution of the lightning data to be consistent with the satellite data, and grid the lightning data to generate images as labels for the U-Net model.

3. The lightning nowcasting method based on multi-source satellite data according to claim 2, characterized in that: In step A, the infrared channels select channels 8, 9, 10, and 13; their λ are 6.2μm, 6.9μm, 7.3μm, and 10.4μm respectively; The calculation channels include BTD0910, BTD1013, BTD1513 and TTD. Among them, the brightness temperature difference of the BTD0910 calculation channel is the difference between the brightness temperature of channel 9 and the brightness temperature of channel 10; The brightness temperature difference of the channels calculated by BTD1013 is the difference between the brightness temperature of channel 10 and the brightness temperature of channel 13; The brightness temperature difference of the channel calculated by BTD1513 is the difference between the brightness temperature of channel 15 and the brightness temperature of channel 13; where λ of channel 15 = 12.4 μm; The brightness temperature difference of the TTD calculation channel = (brightness temperature of channel 11 - brightness temperature of channel 13) - (brightness temperature of channel 13 - brightness temperature of channel 15), where λ of channel 11 = 8.6 μm.

4. The lightning nowcasting method based on multi-source satellite data according to claim 2, characterized in that: In step B, the spatial resolution of lightning data is 0.05° and the temporal resolution is 10 minutes; The specific method of data gridding processing is: A lightning image is generated every 10 minutes. The image contains all lightning events in this time period. Each lightning event is mapped to a grid generated according to the latitude and longitude coordinates. The position of each grid point is represented by the coordinates (x, y). For each (x, y) pixel position, it is determined whether a lightning event occurred in this time period. If there is at least one lightning event, its value is set to 1, otherwise it is set to 0. A binary matrix L(x, y) is generated, which satisfies the following relationship: Where i is the number of lightning events occurring in the time window.

5. The lightning nowcasting method based on multi-source satellite data according to claim 4, characterized in that: In the grid, each cell is 5km×5km in size.

6. The lightning nowcasting method based on multi-source satellite data according to claim 1, characterized in that: In step (2), the specific structure of the U-Net model includes an encoder, a decoder, and an output layer; The encoder includes 5 convolution blocks, each of which consists of two 3×3 convolutions + batch normalization + ReLU + Dropout; each convolution block is followed by a maximum pooling operation; The input of the encoder is an 8-channel image; The decoder starts from the last layer of the encoder and gradually restores the spatial resolution through upsampling operations; in each upsampling block, the feature map of the jump connection is concatenated with the upsampled feature map. After each upsampling, a convolution block is used to integrate the channel information; The output layer is a 1×1 convolution, which is used to convert the feature map into 1 channel for output.

7. The lightning nowcasting method based on multi-source satellite data according to claim 1, characterized in that: In step (2), when training the U-Net model, a weighted cross entropy loss function is used for model training; Among them, y i is the true label, indicating whether a lightning event occurs or not. When lightning occurs, y i The value is 1; when no lightning occurs, y i The value is 1; is the probability of lightning predicted by the U-Net model; W i represents the weight. When a lightning event occurs, that is, when the true label is lightning, W i W lightning , when no lightning event occurs, that is, when the label is no lightning, W i W no-lightning ; W lightning <W no-lightning ; W lightning +W no-lightning =1; The loss function is calculated according to the loss function formula of formula (1) and back propagated. The loss function is minimized through multiple iterations to obtain the U-Net lightning prediction model.

8. The lightning nowcasting method based on multi-source satellite data according to claim 7, characterized in that: 0.7≤W lightning ≤0.99,0.01≤W no-lightning ≤0.3。 9. The lightning nowcasting method based on multi-source satellite data according to claim 1, characterized in that: In step (1), the sample ratio of the training set to the test set is 9:1; in step (3), the sample ratio of the training set to the test set is 7:3; In step (3), the specific method of preprocessing is: extract the brightness temperature data of the target area from the historical multi-source satellite data, map all the data to [0,1] through normalization, crop it according to the required longitude and latitude range, and then convert it into image format and save it to form a data set.

10. The lightning nowcasting method based on multi-source satellite data according to claim 1, characterized in that: In step (4), the satellite image sequence of the past hour is used as input, and the satellite image sequence of the next hour is used as output. The PredRNN model is trained and optimized using MSE loss and decoupling loss. During training, the MSE loss and decoupling loss are fused by summing, and the total loss is calculated by adding the two together. The training is performed with the goal of minimizing the total loss.

Images