Multi-source data lightning approaching early warning method fused with three-dimensional attention mechanism
By constructing a three-dimensional U-Net network model, combining radar and historical lightning data, the problem of reduced accuracy in traditional lightning warning methods is solved, and higher lightning prediction accuracy and effective early warning of small-scale thunderstorms are achieved.
Patent Information
- Application Number
- CN202510584818.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2025-08-19
AI Technical Summary
Traditional lightning early warning methods are difficult to accurately simulate the sudden and nonlinear growth of lightning activities, resulting in a decrease in prediction accuracy with the increase in early warning time, and a high rate of underreport for weak echoes and small-scale thunderstorms.
The multi-source data lightning proximity warning method is adopted with a multi-source data lightning approach, and by constructing a three-dimensional U-Net prediction network model based on the attention mechanism, using radar data and historical lightning data for data splicing and time pairing, training and optimization prediction network model to achieve lightning prediction and early warning.
It improves the accuracy of lightning prediction, can better learn the space-time laws of lightning activities, highlight important spatial positions, and enhances the accuracy of prediction.
Smart Images
Figure CN120508954A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of lightning approach prediction, and in particular to a multi-source data lightning approach warning method integrating a three-dimensional attention mechanism. Background Art
[0002] Lightning disasters pose a serious threat to key sectors such as power systems, aviation, communications networks, and transportation. Therefore, lightning monitoring, forecasting, and mitigation are becoming increasingly important. Lightning, a sudden electrical discharge that occurs during thunderstorms, is characterized by small scale, rapid changes, and complex evolution. However, due to the rapid evolution and localized nature of thunderstorms, traditional monitoring and forecasting methods struggle to accurately simulate the sudden and nonlinear growth of lightning activity, making accurate early warning of approaching lightning a significant challenge.
[0003] Early lightning predictions primarily relied on radar extrapolation algorithms and numerical models. Traditional extrapolation algorithms include TREC (Tracking Radar Echoes by Correlation), TITAN (Thunderstorm Identification, Tracking, Analysis, and Nowcasting), Johnson's SCIT (Storm Cell Identification and Tracking), and optical flow methods. Most extrapolation algorithms rely on fixed physical model assumptions and analyze past observational data to predict lightning. However, due to the rapid and highly nonlinear nature of lightning development, it is difficult to extract nonlinear features from a large number of thunderstorm processes and learn the physical laws governing thunderstorm initiation, development, and dissipation. Consequently, the prediction accuracy of these methods degrades significantly as warning time increases. For example, the cross-correlation-based TREC algorithm tracks the motion trends of radar echoes to achieve storm extrapolation. However, its derivation assumes a stable storm structure, making it difficult to adapt to the rapid development and dissipation of thunderstorms. Cell tracking algorithms such as TITAN and SCIT predetermine a set of reflectivity thresholds and then extract PPI storm cell components at different elevation angles above the threshold. These components are combined to generate three-dimensional thunderstorm cells for auxiliary tracking and early warning. However, these methods have a high underreporting rate for lightning activity in weak echo regions, failing to balance the stability of the results with the accuracy of forecasts. Optical flow methods are similar in principle to the TREC method, but offer greater portability and stability across various tasks. However, they are also limited by the simplification of their physical assumptions and are unable to model the complex charge separation mechanisms within thunderstorm clouds. Numerical models parameterize the electrification and discharge processes within thunderstorm clouds and couple them to meteorological models of corresponding scales, enabling them to simulate the charge structure and electric field characteristics within clouds and forecast lightning activity. However, most numerical models have temporal resolutions greater than one hour and low spatial resolution, which can lead to underreporting of small-scale thunderstorms.
[0004] In summary, traditional methods have strong simplifications in physical assumptions, difficulty in capturing the rapid nonlinear evolution of lightning activity, and insufficient temporal and spatial resolution of numerical models. As a result, the prediction accuracy decreases as the warning time increases, and the underreporting rate of weak echoes and small-scale thunderstorms is high. Summary of the Invention
[0005] The technical problem to be solved by the present invention is how to provide a multi-source data lightning approach warning method that integrates a three-dimensional attention mechanism with high lightning warning accuracy.
[0006] To solve the above technical problems, the present invention adopts a technical solution: a multi-source data lightning approach warning method integrating a three-dimensional attention mechanism, comprising the following steps:
[0007] Data collection and preprocessing: Obtain radar data of the area to be predicted and obtain historical lightning data in the area;
[0008] Dataset construction: Lightning data and radar data are spliced together to obtain multi-source data input samples. Time pairing is performed to ensure that each pair of input data corresponds to the same time period, completing the dataset construction.
[0009] Build a prediction network model: Build a three-dimensional U-Net prediction network model based on the attention mechanism;
[0010] Model training: using the constructed data set to train the prediction network model and optimize the prediction network model;
[0011] Lightning prediction: Use the optimized prediction network model to predict lightning and issue warnings based on the prediction results.
[0012] The beneficial effects of adopting the above technical solution are: the method can extract spatiotemporal features and learn the spatiotemporal laws of lightning activities through the constructed three-dimensional U-Net network; the constructed three-dimensional attention mechanism can realize the adaptive allocation of multi-source data weights in the channel dimension, improving the model's ability to select spatiotemporal features; in the spatial dimension, it can highlight important spatial locations and improve the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0014] Figure 1 is a main flow chart of the method according to an embodiment of the present invention;
[0015] Figure 2 Schematic diagram of the structure of the three-dimensional U-Net prediction network model in the method according to an embodiment of the present invention;
[0016] Figure 3 CBAM framework diagram in the method according to an embodiment of the present invention;
[0017] Figure 4 is a schematic diagram of the structure of the channel attention module in the method according to an embodiment of the present invention;
[0018] Figure 5 Schematic diagram of the structure of the spatial attention module in the method described in an embodiment of the present invention. DETAILED DESCRIPTION
[0019] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts are within the scope of protection of the present invention.
[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0021] Overall, such as Figure 1 As shown, the embodiment of the present invention discloses a multi-source data lightning approach warning method integrating a three-dimensional attention mechanism, comprising the following steps:
[0022] S1, data acquisition and preprocessing: obtain radar data of the area to be predicted and obtain historical lightning data in the area;
[0023] S2, dataset construction: Lightning data and radar data are spliced together to obtain multi-source data input samples. Through time pairing, each pair of input data corresponds to the same time period, completing the dataset construction;
[0024] S3, build prediction network model: build a three-dimensional U-Net prediction network model based on the attention mechanism;
[0025] S4, model training: using the constructed data set to train the prediction network model and optimize the prediction network model;
[0026] S5, Lightning Prediction: Use the optimized prediction network model to predict lightning and issue warnings based on the prediction results.
[0027] The following is a detailed description of the above steps combined with specific methods
[0028] Data collection and preprocessing:
[0029] Acquire radar data, perform data preprocessing on the radar data, extract the reflectivity at each elevation angle of the radar base data and interpolate it to a unified grid, compensate for systematic errors by removing or correcting outliers and noise, then use spatial registration and visualization to obtain a radar reflectivity image containing geographic information, and then crop the radar data to obtain only the study area;
[0030] Lightning data were obtained from cloud-to-ground lightning (CG) observation data provided by the China Meteorological Administration's Lightning Detection Network. This data includes information such as lightning return stroke location, occurrence time, lightning intensity, lightning steepness, positioning method, and positioning error. The lightning data were discretized in time and space. First, lightning return stroke data were accumulated at 10-minute intervals to generate lightning data with a 10-minute time resolution. This data was then two-dimensionally gridded. A 1 km × 1 km two-dimensional grid was created for the study area, and lightning strikes within the area were projected onto the grid according to longitude and latitude.
[0031] Dataset construction:
[0032] For lightning data, observation data with a lightning event frequency of more than 20 times were selected for spatiotemporal matching;
[0033] For radar data, radar data with a time difference of less than 3 minutes from lightning data are selected for time matching with lightning data, and the two are spliced in the channel dimension to obtain multi-source data input samples. Time matching is used to ensure that each pair of input data corresponds to the same time period.
[0034] A three-dimensional U-Net based on the attention mechanism is used as a prediction network model to predict the probability of lightning occurrence and the distribution of lightning strike areas;
[0035] 3D-Unet has an encoder and decoder structure. The encoder part on the left uses a series of 3D convolutional layers to extract spatiotemporal features and reduces the scale of the feature map through a 3D pooling layer. The decoder part on the right uses 3D up-convolution and the same 3D convolutional layer to gradually restore the scale. In the middle, the feature information corresponding to the encoder and decoder layers is fused through a shortcut connection layer, and the feature information fusion capability is improved through an improved attention mechanism. Figure 2 The overall structure of the model.
[0036] As shown in the structure diagram, the input is a (6×2×480×480) image with 2 channels, one for radar combined reflectivity and one for lightning location data. 6 represents the time dimension, with a 10-minute temporal resolution. This indicates six images, so the input consists of six images taken 10 minutes apart, for a total of one hour. 480 represents the image size. Therefore, there are two channels, namely six images of radar combined reflectivity and lightning location data taken 10 minutes apart, as input samples.
[0037] The overall model structure is U-shaped, extracting features through layer-by-layer convolution and restoring them through layer-by-layer upconvolution. 3D-CBAM attention is added to the central shortcut connection. The left half is the encoder, which has five layers, each with two 3D convolutional layers. 3D pooling downsampling is used to reduce the feature map dimension. The right half is the decoder, also five layers, each with two 3D convolutional layers. Three-dimensional upconvolution is used to restore the feature map dimension.
[0038] Such as structure Figure 2 As shown in the figure, a data sample of (6×2×480×480) is input into the model and first passes through a 3D convolutional layer, indicated by the purple arrow. This layer has two convolution operations. In each convolution operation, the convolution kernel size is 3×3×3, the stride is 1, and the padding is 1. After each convolution, the convolution layer passes through the BatchNorm (BN) and Prelu activation functions. The BN operation is used to normalize the batch of samples, and the Prelu activation function is used to prevent gradient vanishing or exploding during training. After convolution, the data size of the (6×2×480×480) data sample becomes (6×64×480×480).
[0039] The data of (6×64×480×480) is divided into two branches, one branch continues convolution through 3D pooling downsampling, and the other branch performs channel-spatial attention allocation through shortcut connection; the data of the first branch (6×64×480×480) is converted to (4×64×240×240) through 3D pooling of (2×2×2) size and padding of (1,0,0), and then continues convolution in the second layer; the second branch is a shortcut connection, where the channel-spatial attention mechanism is used, such as Figure 2, the (6×64×480×480) data are respectively subjected to average pooling and maximum pooling to obtain two pooled feature maps, and two (1×64×1×1) feature maps are obtained by pooling. Then, the channel dimension of the input feature map is compressed to (1×4×1×1) by a three-dimensional convolution of size (1×1×1), and then the channel dimension is restored to (1×64×1×1) by a three-dimensional convolution of size (1×1×1); finally, the two output feature maps are added and then passed through the sigmoid activation function to obtain the channel attention weight Mc(1×64×1×1). At the same time, the (6×64×480×480) data will also obtain the spatial attention weight Ms through the spatial attention module. The (6×64×480×480) data will undergo channel-based three-dimensional global maximum pooling and three-dimensional global average pooling, and the two obtained (6×1×480×480) feature maps will be spliced in the channel dimension to obtain a feature map (6×2×480×480). After the 7×7×7 convolution is used to reduce the dimension to (6×1×480×480), the spatial attention weight Ms is obtained after the sigmoid activation function is used. After obtaining the weight, in the subsequent decoding process, the features of this layer (6×64×480×480) will be multiplied by the two attention weights and spliced with the features of the corresponding decoder layer in the channel dimension.
[0040] Similarly, the second layer features of the encoder (4×64×240×240) are obtained through the three-dimensional convolution layer to obtain a feature map of (4×128×240×240) and then downsampled to obtain a feature map of (2×128×120×120), and two attention weights are obtained through the attention module.
[0041] The third layer feature of the encoder (2×128×120×120) is obtained through the three-dimensional convolution layer to obtain a feature map of (2×256×120×120) and then downsampled to obtain a feature map of (1×256×60×60). Two attention weights are obtained through the attention module.
[0042] The fourth layer feature of the encoder (1×256×60×60) is obtained through the three-dimensional convolution layer to obtain a feature map of (1×512×60×60) and then downsampled to obtain a feature map of (1×512×30×30). Two attention weights are obtained through the attention module.
[0043] The fifth-layer features of the encoder (1×512×30×30) are passed through the three-dimensional convolution layer to obtain the dimensionally unchanged (1×512×30×30) features, which are upsampled to obtain the (1×512×60×60) feature map. At this time, the features in the second branch quick connection of the corresponding layer of the original encoder are multiplied by the two attention weights to obtain the weighted features (1×512×60×60); the two (1×512×60×60) features are spliced in the channel dimension (1×1024×60×60), and then passed through the three-dimensional convolution layer to obtain the (1×256×60×60) feature map, which is the fourth-layer feature of the decoder.
[0044] The fourth-layer features of the decoder (1×256×60×60) are upsampled to obtain (2×256×120×120) features. At this time, the features in the second branch quick connection of the fourth layer of the original encoder are multiplied by the two attention weights to obtain the weighted features (2×256×120×120); the two (2×256×60×60) features are spliced in the channel dimension (2×512×120×120), and then passed through the three-dimensional convolution layer to obtain the feature map of (2×128×120×120), which is the third-layer feature of the decoder.
[0045] The third-layer features of the decoder (2×128×120×120) are upsampled to obtain (4×128×240×240) features. At this time, the features in the second branch quick connection of the third layer of the original encoder are multiplied by the two attention weights to obtain the weighted features (4×128×240×240); the two (4×128×240×240) features are spliced in the channel dimension (4×256×240×240), and then passed through the three-dimensional convolution layer to obtain the feature map of (4×64×240×240), which is the second-layer feature of the decoder.
[0046] The second-layer features of the decoder (4×64×240×240) are upsampled to obtain (6×64×480×480) features. At this time, the features in the second branch quick connection of the second layer of the original encoder are multiplied by the two attention weights to obtain the weighted features (6×64×480×480); the two (6×64×480×480) features are spliced in the channel dimension (6×128×480×480), and then passed through the three-dimensional convolution layer to obtain the feature map of (6×64×480×480), which is the first-layer feature of the decoder.
[0047] The first layer features of the decoder (6×64×480×480) are reduced to 1 in time dimension through three-dimensional pooling, that is, (1×64×480×480), and then a two-channel binary classification prediction result (1×2×480×480) is obtained through three-dimensional convolution.
[0048] The prediction result of the determined probability value is obtained through the softmax classifier. There are two channels: channel 1 represents the probability distribution of the area where lightning occurs, and channel 2 represents the probability distribution of the area where lightning does not occur.
[0049] The improved 3D-CBAM attention mechanism is used in the model. CBAM (Convolutional Block Attention Module) has two modules: channel attention module and spatial attention module. Figure 3 The structure of CBAM is shown.
[0050] The module can be formulated as:
[0051]
[0052] like Figure 4 As shown in the figure, the channel attention module uses the channel relationship between features to generate a channel attention map. The input feature map F is subjected to three-dimensional global maximum pooling and three-dimensional global average pooling based on height and width respectively. Two C×1×1×1 feature maps (C is the number of channels) are obtained through pooling. The spatial dimensions of the input feature map are then compressed through three convolutional layers. The output features are then summed and activated through a sigmoid function to obtain the channel attention weight Mc. The features multiplied by the channel attention weight can highlight the feature channels in lightning data and radar data that are most helpful for prediction, helping the network to better integrate multi-source data.
[0053] like Figure 5 As shown in the figure, the spatial attention module uses the spatial relationship between features to generate a spatial attention map. The input feature F′ undergoes channel-based three-dimensional global maximum pooling and three-dimensional global average pooling respectively. The two feature maps are concatenated in the channel dimension, and the dimensionality is reduced to one channel through a 7×7×7 convolution. After passing the sigmoid activation function, the spatial attention weight Ms is obtained. The feature after multiplying with the spatial attention weight can highlight important spatial positions and suppress background noise.
Claims
1. A multi-source data lightning approach warning method integrating a three-dimensional attention mechanism, characterized by The steps include: Data collection and preprocessing: Obtain radar data of the area to be predicted and obtain historical lightning data in the area; Dataset construction: Lightning data and radar data are spliced together to obtain multi-source data input samples. Through time pairing, each pair of input data corresponds to the same time period to complete the dataset construction. Build a prediction network model: Build a three-dimensional U-Net prediction network model based on the attention mechanism; Model training: using the constructed data set to train the prediction network model and optimize the prediction network model; Lightning prediction: Use the optimized prediction network model to predict lightning and issue warnings based on the prediction results.
2. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 1 is characterized in that: In the data collection and preprocessing steps: The radar data is preprocessed to extract the reflectivity at each elevation angle of the radar base data and interpolate it to a uniform grid. System errors are compensated by removing or correcting outliers and noise. Spatial registration and visualization are then used to obtain a radar reflectivity image including geographic information. The radar data that only includes the area to be studied is then cropped.
3. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 1 is characterized in that: In the data collection and preprocessing steps: Lightning data is obtained from a public database and discretized in time and space. First, lightning return stroke data is accumulated at a set time interval to generate lightning data with a time resolution of that period. Then, the lightning data is processed into a two-dimensional grid. A two-dimensional grid of a set size is created according to the study area, and lightning in the area is projected into the grid according to longitude and latitude.
4. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 1 is characterized in that: The dataset construction includes the following steps: For lightning data, observation data with a lightning event frequency higher than a set threshold are selected for spatiotemporal matching; For radar data, radar data with a time difference less than the set time from lightning data are selected for time matching with the lightning data, and the two are spliced in the channel dimension to obtain multi-source data input samples. Time matching is used to ensure that each pair of input data corresponds to the same time period.
5. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 1 is characterized in that: The three-dimensional U-Net prediction network model includes an encoder, a decoder, and a shortcut connection layer. The encoder part is used to extract spatiotemporal features using a series of three-dimensional convolutional layers and reduce the scale of the feature map through a three-dimensional pooling layer; the decoder part is used to gradually restore the scale using three-dimensional up-convolution and three-dimensional convolution layers; the shortcut connection layer is used to fuse the feature information corresponding to the encoder layer and the decoder layer, and improve the feature information fusion capability through an improved attention mechanism.
6. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 5 is characterized in that: The overall structure of the prediction network model is U-shaped, which extracts features through layer-by-layer convolution and restores features through layer-by-layer up-convolution; a 3D-CBAM attention module is added at the middle shortcut connection layer; the left half is the encoder, which has a total of 5 layers, each layer has two 3D convolution layers, and the feature map dimension is reduced by 3D pooling downsampling; the right half is the decoder, which has a total of 5 layers, each layer has two 3D convolution layers, and the feature map dimension is restored by three-dimensional up-convolution.
7. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 5 is characterized in that: The processing method of the prediction network model comprises the following steps: A data sample with a data dimension of (6×2×480×480) is input into the model. It first passes through a 3D convolutional layer in the first layer of the encoder. This layer includes two convolution operations. In a single convolution operation, the convolution kernel size is 3×3×3, the stride is 1, and the padding is 1. After each convolution, the batch normalization and prelu activation functions are applied. After convolution, the data sample with a data dimension of (6×2×480×480) becomes (6×64×480×480). The data with a dimension of 6×64×480×480 is divided into two branches. The first branch continues the convolution through 3D pooling downsampling, and the second branch is connected to the first layer of the decoder through a shortcut connection layer to perform channel-spatial attention allocation; The second layer features of the encoder (4×64×240×240) are passed through the 3D convolutional layer to obtain a feature map of (4×128×240×240). One path is downsampled to obtain a feature map of (2×128×120×120). The other path passes through the attention module to obtain two attention weights. The encoder's third layer features (2×128×120×120) are passed through a 3D convolutional layer to obtain a feature map of (2×256×120×120). One path is downsampled to obtain a feature map of (1×256×60×60). The other path passes through the attention module to obtain two attention weights. The encoder's fourth layer features (1×256×60×60) pass through a three-dimensional convolutional layer to obtain a feature map of (1×512×60×60). One path is downsampled to obtain a feature map of (1×512×30×30). The other path passes through the attention module to obtain two attention weights. The fifth-layer features of the encoder (1×512×30×30) are passed through the three-dimensional convolution layer to obtain the dimension-invariant (1×512×30×30) features. The fifth-layer features of the decoder are upsampled to obtain the feature map (1×512×60×60). At this time, the features in the second branch quick connection layer of the corresponding layer of the original encoder are multiplied by the two attention weights to obtain the weighted features (1×512×60×60); the two (1×512×60×60) features are concatenated in the channel dimension (1×1024×60×60), and then passed through the three-dimensional convolution layer to obtain the feature map (1×256×60×60), which is the fourth-layer feature of the decoder. The decoder's fourth-layer features (1×256×60×60) are upsampled to obtain (2×256×120×120) features. At this time, the features in the second branch quick connection of the original encoder's fourth layer are multiplied by the two attention weights to obtain the weighted features (2×256×120×120); the two (2×256×60×60) features are concatenated in the channel dimension (2×512×120×120), and then passed through the 3D convolution layer to obtain the feature map of the decoder's third layer (2×128×120×120); The decoder's third-layer features (2×128×120×120) are upsampled to obtain (4×128×240×240) features. At this time, the features in the second branch quick connection layer of the original encoder's third layer are multiplied by the two attention weights to obtain weighted features (4×128×240×240); the two (4×128×240×240) features are concatenated in the channel dimension (4×256×240×240), and then passed through a three-dimensional convolutional layer to obtain the feature map of the decoder's second-layer features (4×64×240×240); The second-layer features of the decoder (4×64×240×240) are upsampled to obtain (6×64×480×480) features. At this time, the features in the second branch quick connection of the second layer of the original encoder are multiplied by the two attention weights to obtain the weighted features (6×64×480×480); the two (6×64×480×480) features are concatenated in the channel dimension (6×128×480×480), and then passed through the 3D convolution layer to obtain the feature map of the first-layer features of the decoder (6×64×480×480); The feature map of the first layer of the decoder (6×64×480×480) is reduced to 1 in time dimension through three-dimensional pooling, that is, (1×64×480×480), and then a two-channel binary classification prediction result (1×2×480×480) is obtained through three-dimensional convolution.
8. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 7, characterized in that: In the two branches of the first layer output of the encoder, the data of the first branch (6×64×480×480) is converted to (4×64×240×240) through three-dimensional pooling of (2×2×2) size and filling (1,0,0), and then continues to be convolved in the second layer of the encoder; the second branch is a shortcut connection, where the channel space attention mechanism is used, and the data of (6×64×480×480) is respectively subjected to average pooling and maximum pooling to obtain two pooled feature maps, and two (1×64×1×1) feature maps are obtained by pooling. After that, the channel dimension of the input feature map is compressed to (1×4×1×1) by three-dimensional convolution of (1×1×1), and then the channel dimension is restored to (1×64×1×1) by three-dimensional convolution of (1×1×1); finally, the two output feature maps are added Then, the channel attention weight Mc(1×64×1×1) is obtained by the sigmoid activation function; at the same time, the data of (6×64×480×480) is obtained through the spatial attention module to obtain the spatial attention weight Ms, and the data of (6×64×480×480) is subjected to channel-based three-dimensional global maximum pooling and three-dimensional global average pooling. The two obtained (6×1×480×480) feature maps are spliced in the channel dimension to obtain a feature map of (6×2×480×480), which is reduced to (6×1×480×480) through a 7×7×7 convolution and then subjected to the sigmoid activation function to obtain the spatial attention weight Ms. After obtaining the weight, in the subsequent decoding process, the features of this layer (6×64×480×480) are multiplied by the two attention weights and then spliced with the features of the corresponding decoder layer in the channel dimension.
9. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 5, characterized in that: The 3D-CBAM attention mechanism module includes a channel attention module and a spatial attention module.
10. The multi-source data lightning approach warning method integrating a three-dimensional attention mechanism as claimed in claim 9, characterized in that: The 3D-CBAM attention mechanism module is formulated as follows: The channel attention module uses the channel relationship between features to generate a channel attention map. The input feature map F is subjected to three-dimensional global maximum pooling and three-dimensional global average pooling based on height and width, respectively. This pooling process generates two C×1×1×1 feature maps, where C is the number of channels. The spatial dimensions of the input feature map are then compressed through three convolutional layers. The output features are then summed and activated with a sigmoid function to obtain the channel attention weight Mc. The features multiplied by the channel attention weight are used to highlight the most predictive feature channels in lightning and radar data, helping the network better integrate multi-source data. The spatial attention module uses the spatial relationship between features to generate a spatial attention map. The input feature F′ undergoes channel-based three-dimensional global maximum pooling and three-dimensional global average pooling respectively. The two feature maps are concatenated in the channel dimension, reduced to one channel through a 7×7×7 convolution, and then activated by a sigmoid function to obtain the spatial attention weight Ms. The feature after multiplying with the spatial attention weight highlights important spatial positions and suppresses background noise.