Lightning activity information analysis method based on image recognition
By using multimodal fusion images and a dual-branch deep learning network, the problems of insufficient data utilization and low prediction accuracy in existing lightning monitoring methods are solved, enabling accurate identification and spatiotemporal prediction of lightning activity and providing efficient early warning decision support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QINZHOU POWER SUPPLY BUREAU OF GUANGXI POWER GRID CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-24
AI Technical Summary
Existing lightning monitoring methods rely on single-channel images and fail to fully utilize the complementary information from multi-source data. Traditional methods are insufficient in characterizing the generation and dissipation of thunderstorms, resulting in low prediction accuracy and short warning time. Deep learning models lack a unified framework for lightning identification and spatiotemporal prediction.
A lightning activity analysis method based on image recognition is constructed. This method combines multimodal fusion image generation and a dual-branch deep learning network with a lightning activity recognition and segmentation network and a lightning spatiotemporal evolution prediction network. The model is trained using the Dice loss function and the combined loss function, and the centroid tracking method is used for deep information mining.
It enables accurate identification and spatiotemporal prediction of lightning activity, generating high-quality lightning probability maps and prediction images, providing direct decision-making basis for disaster prevention and mitigation, and improving prediction accuracy and early warning time.
Smart Images
Figure CN121919793A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of lightning activity information analysis technology, specifically to a lightning activity information analysis method based on image recognition. Background Technology
[0002] Lightning is a severe natural disaster, making accurate monitoring and early warning of its activity crucial. Traditional lightning monitoring relies primarily on ground-based lightning location networks, but their coverage is limited and effective forecasting is difficult. With the development of meteorological satellite technology, visible light, infrared, and water vapor channel cloud images provide abundant data for lightning monitoring. However, existing analytical methods have significant limitations: First, most studies rely solely on single-channel (e.g., infrared) images, failing to fully utilize the complementary information from multi-source data; second, traditional methods are mostly based on simple threshold segmentation or extrapolation, lacking the ability to characterize complex nonlinear processes such as the formation and dissipation of thunderstorms, resulting in low prediction accuracy and short warning times; furthermore, existing deep learning models often treat lightning identification and spatiotemporal prediction in isolation, lacking a unified framework integrating identification, prediction, and risk decision-making. Therefore, developing an analytical method that can deeply integrate multimodal satellite data, accurately identify lightning activity, and effectively predict its spatiotemporal evolution is a pressing technical challenge. Summary of the Invention
[0003] The purpose of this invention is to address the problems that most studies rely solely on single-channel (e.g., infrared) images, failing to fully utilize the complementary information from multi-source data; traditional methods, often based on simple threshold segmentation or extrapolation, are insufficient in characterizing complex nonlinear processes such as the generation and dissipation of thunderstorms, resulting in low prediction accuracy and short warning time; and existing deep learning models are often treated separately in terms of lightning identification and spatiotemporal prediction, lacking a unified framework that integrates identification, prediction, and risk decision-making. Therefore, this invention proposes an image recognition-based method for analyzing lightning activity information.
[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A method for analyzing lightning activity information based on image recognition includes the following steps: S1. Acquire visible light image, infrared image and water vapor channel image of lightning, and fuse the three into a multimodal fusion image after standardization preprocessing; S2. Construct a dual-branch deep learning network structure, which includes a lightning activity recognition and segmentation network and a lightning spatiotemporal evolution prediction network. S3. Using a single multimodal fusion image as input, a probability segmentation map can be obtained through the lightning activity recognition and segmentation network. After multiple training sessions, a trained lightning activity recognition and segmentation network is obtained. Using a short sequence of multimodal images as input, a predicted image can be obtained through the lightning activity recognition and segmentation network. S4. Input the predicted image into the trained lightning activity recognition and segmentation network to obtain a lightning activity probability map based on the future predicted state. S5. Deep information mining is performed on the probability map and prediction image of lightning activity. The lightning activity intensity index is used as the vitality diagnosis of lightning activity. The centroid tracking method is used to obtain the motion trajectory prediction of lightning risk and the decision support map of early warning response.
[0005] Based on the above technical solution, the present invention can be further improved as follows.
[0006] Preferably, the specific generation process of the multimodal fusion image is as follows: First, the original satellite images of the visible light, infrared, and water vapor channels are standardized and preprocessed to unify them into the same spatiotemporal grid and numerical range; then, a dynamic weighted fusion method based on physical mechanisms is adopted, that is, weights are assigned to each channel, and then the information of the three channels is compressed and mapped into a new image space through principal component analysis or color synthesis technology; finally, a multimodal fusion image integrating cloud top texture, development intensity, and atmospheric water vapor conditions is generated.
[0007] Preferably, the lightning activity recognition and segmentation network uses an encoder-decoder architecture as its basic architecture. The encoder part consists of four downsampling stages, each of which includes two 3×3 convolutional layers, a ReLU activation function, and a batch normalization layer, and is finally followed by a 2×2 max pooling layer for downsampling. The decoder part also includes four upsampling stages, each of which doubles the size of the feature map through transposed convolution and makes skip connections with the feature map of the corresponding layer of the encoder. The lightning activity identification and segmentation network is trained using a supervised learning approach with labels based on real future lightning observation data, and the Dice loss function is selected as the loss function.
[0008] Preferably, the lightning spatiotemporal evolution prediction network uses a ConvLSTM network as the core to construct ConvLSTM units. Each ConvLSTM unit includes an input gate, a forget gate, an output gate, and a cell state, and all fully connected operations are replaced with convolutional operations. The lightning spatiotemporal evolution prediction network structure adopts an encoder-prediction framework. The encoder part consists of multiple stacked ConvLSTM layers, which progressively extract spatiotemporal features from the input sequence. The prediction part decodes the learned temporal evolution rules into cloud map predictions for multiple future time steps through deconvolutional layers and skip connections. During the training of the lightning spatiotemporal evolution prediction network, a combined loss function is used. The mean squared error ensures that the predicted cloud map is numerically close to the real cloud map, while the gradient difference loss specifically penalizes the problem of blurred prediction boundaries, making the predicted thunderstorm cloud outline clearer.
[0009] Preferably, the Dice loss function is calculated using the following formula: ; in, It is a prediction probability. These are the actual labels, and N is the total number of pixels. It is a smoothing term to prevent division by zero.
[0010] Preferably, the formula for calculating the combined loss function is as follows: Mean square error: ; Gradient difference loss: ; Combination loss function: ; in, This is the mean squared error value, where N is the total number of pixels. It is a predicted value. It is the actual value. It is the gradient difference loss value. It is the gradient operator of the image in the x-direction. It is the gradient operator of the image in the y-direction. It is the loss value of the combined loss function. yes The weighting coefficients, yes The weighting coefficients.
[0011] Preferably, the formula for the lightning activity intensity index is: ; in, Grid points The probability of lightning occurrence, It represents the actual area that the grid point represents. It is the total area of regions with a probability exceeding 0.5. It is the total area of the grid.
[0012] Preferably, the calculation formula for the centroid tracking method is as follows: Formula for calculating the center of mass: ; ; in, Grid points The probability of lightning occurrence, and These are the row index and column index of the grid point, respectively. It is the x-coordinate of the centroid at time t. It is the ordinate of the centroid at time t; Formula for calculating movement speed: ; in, and These are the x and y coordinates of the centroid at time t. and These are the x and y coordinates of the centroid at time t-1, respectively. It's the grid resolution. It is the time interval between two consecutive time steps. This is the movement speed value; Formula for calculating direction of movement: ; in, It is the displacement component of the center of mass in the x-direction. It is the displacement component of the center of mass in the y-direction. It indicates the direction of movement.
[0013] Preferably, atmospheric physical constraints are introduced as weight adjustment factors during the multimodal fusion image generation process to construct a dynamic attention-weighted fusion network, specifically including: First, the cloud top height is determined by inverting the brightness temperature of the infrared channel. The development index of the middle and upper-level wetland area was calculated by combining the water vapor channel gradient. And utilize visible light texture complexity to extract convection primary signals. Construct the physical sensing gate function: ; in, It is the first Dynamic weights of each channel It is the Sigmoid function. For learnable feature encoders, This represents a combination of empirical formulas driven by meteorological physical models. , These are the trainable parameters; The fused feature map is represented as follows: ; in, The features of each channel image after preprocessing, This represents element-wise multiplication; Ultimately, a Transformer module is used to achieve cross-channel context modeling, enhancing the information expression capabilities of key areas.
[0014] Preferably, the lightning activity intensity index incorporates a local environmental modulation factor, specifically including: Calculate the topographic lift coefficient ,in, The gradient represents the terrain height field; it is a vector. Represents wind speed vector, topographic lift coefficient Used to reflect the forced uplift effect of windward slopes; Using the topographic lift coefficient and the intensity of the urban heat island The lightning activity intensity index has been modified as follows: ; in, and These are the calibration parameters for the region.
[0015] Compared with the prior art, the technical solution of this application has the following beneficial technical effects: 1. This invention generates a multimodal image by physically weighting and fusing visible light, infrared, and water vapor channel images. This multimodal image integrates multidimensional information such as cloud top texture, development intensity, and atmospheric water vapor, providing a richer and more reliable data foundation for lightning identification and overcoming the limitations of a single data source.
[0016] 2. This invention combines the "static" identification and segmentation of lightning with the "dynamic" spatiotemporal evolution prediction by constructing a dual-branch network structure. This design can not only accurately identify the current lightning activity, but also learn the generation and dissipation evolution law of thunderstorms through the ConvLSTM network, thereby achieving the prediction of future lightning probability and realizing the leap from monitoring to forecasting.
[0017] 3. This invention alleviates the class imbalance problem by selectively using the Dice loss function, and uses a combined loss function that combines mean squared error and gradient difference loss to ensure accurate prediction cloud map values and clear boundaries, making model training more stable and generating higher quality lightning probability maps and prediction images.
[0018] 4. This invention uses the lightning activity intensity index and centroid tracking method for in-depth analysis, transforming image data into intuitive intensity assessments, movement paths, and early warning maps, providing direct and scientific decision-making basis for disaster prevention and mitigation, and realizing a closed loop from data to intelligent decision-making. Attached Figure Description
[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] A method for analyzing lightning activity information based on image recognition includes the following steps: S1. Acquire visible light image, infrared image and water vapor channel image of lightning, and fuse the three into a multimodal fusion image after standardization preprocessing; S2. Construct a dual-branch deep learning network structure, which includes a lightning activity recognition and segmentation network and a lightning spatiotemporal evolution prediction network. S3. Using a single multimodal fusion image as input, a probability segmentation map can be obtained through the lightning activity recognition and segmentation network. After multiple training sessions, a trained lightning activity recognition and segmentation network is obtained. Using a short sequence of multimodal images as input, a predicted image can be obtained through the lightning activity recognition and segmentation network. S4. Input the predicted image into the trained lightning activity recognition and segmentation network to obtain a lightning activity probability map based on the future predicted state. S5. Deep information mining is performed on the lightning activity probability map and prediction image. The lightning activity intensity index is used as the vitality diagnosis of lightning activity. The lightning activity probability map and prediction image are combined to obtain the motion trajectory forecast and early warning response decision support map of lightning risk.
[0022] Visible light images: used to identify the texture, outline, and shadow of thunderstorm cloud tops to determine the intensity of cloud system development.
[0023] Infrared images: used to invert cloud top brightness temperature. The lower the brightness temperature, the higher the cloud top, the stronger the convection, and the greater the potential for lightning activity. This is a key indicator for judging the intensity of thunderstorms.
[0024] Water vapor channel images: used to analyze water vapor transport in the upper atmosphere and help determine the energy source and movement path of thunderstorm systems.
[0025] The specific generation process of the multimodal fusion image is as follows: First, the original satellite images of visible light, infrared, and water vapor channels are standardized and preprocessed to unify them into the same spatiotemporal grid and numerical range. Then, a dynamic weighted fusion method based on physical mechanisms is adopted, that is, weights are assigned to each channel. Then, through principal component analysis or color synthesis technology, the information of the three channels is compressed and mapped into a new image space. Finally, a multimodal fusion image integrating cloud top texture, development intensity, and atmospheric water vapor conditions is generated.
[0026] The lightning activity recognition and segmentation network uses an encoder-decoder architecture as its basic structure. The encoder consists of four downsampling stages, each containing two 3×3 convolutional layers, a ReLU activation function, and a batch normalization layer. Finally, a 2×2 max pooling layer is added for downsampling. This hierarchical network structure can gradually extract features from multimodal satellite images: the shallow network captures local features such as cloud texture and edges, while the deep network understands global features such as the overall structure and spatial distribution of cloud clusters. In particular, dilated convolution can be introduced into the encoder to expand the receptive field, enabling the network to better understand large-scale weather system configurations. The decoder also includes four upsampling stages. Each stage doubles the feature map size through transposed convolution and makes skip connections with the feature maps of the corresponding layers in the encoder. This design ensures that the decoder can utilize the high-level semantic features extracted by the encoder while preserving necessary spatial details when reconstructing high-resolution probability maps. The last convolutional layer uses a 1×1 convolution and a sigmoid activation function to compress the number of feature channels to 1, outputting the lightning probability value for each pixel. The lightning activity recognition and segmentation network is trained using supervised learning with labels from real future lightning observation data. The loss function chosen is the Dice loss function, which is particularly suitable for handling imbalanced classification problems like lightning, which account for a small proportion of images.
[0027] The lightning spatiotemporal evolution prediction network uses a ConvLSTM network as its core to construct ConvLSTM units. Each ConvLSTM unit includes an input gate, a forget gate, an output gate, and a cell state. All fully connected operations are replaced with convolutional operations, which allows the network to maintain the spatial structure of the feature map while remembering temporal dynamics. Specifically, the forget gate controls which historical spatial information needs to be retained, the input gate determines which newly observed spatial features need to be updated, and the output gate regulates the degree of influence of the current state on the next time step. The lightning spatiotemporal evolution prediction network structure adopts an encoder-prediction framework. The encoder part consists of multiple stacked ConvLSTM layers, which progressively extract spatiotemporal features from the input sequence. The prediction part decodes the learned temporal evolution rules into cloud map predictions for multiple future time steps through deconvolutional layers and skip connections. During the training of the lightning spatiotemporal evolution prediction network, a combined loss function is used. The mean squared error ensures that the predicted cloud map is numerically close to the real cloud map, while the gradient difference loss specifically penalizes the problem of blurred prediction boundaries, making the predicted thunderstorm cloud outline clearer.
[0028] Trained on a large amount of historical data, the lightning spatiotemporal evolution prediction network can learn the typical movement patterns, intensity variation patterns, and evolutionary characteristics of different types of thunderstorm systems. For example, it can learn to identify the propagation direction of thunderstorms, predict the cooling rate of cloud tops, and even simulate the influence of flow boundaries on the triggering of nascent convection.
[0029] The formula for calculating the Dice loss function is as follows: ; in, It is a prediction probability. These are the actual labels, and N is the total number of pixels. It is a smoothing term to prevent division by zero.
[0030] The formula for calculating the combined loss function is as follows: Mean square error: ; Gradient difference loss: ; Combination loss function: ; in, This is the mean squared error value, where N is the total number of pixels. It is a predicted value. It is the actual value. It is the gradient difference loss value. It is the gradient operator of the image in the x-direction. It is the gradient operator of the image in the y-direction. It is the loss value of the combined loss function. yes The weighting coefficients, yes The weighting coefficients.
[0031] The formula for the lightning activity intensity index is: ; in, Grid points The probability of lightning occurrence, It represents the actual area that the grid point represents. It is the total area of regions with a probability exceeding 0.5. It is the total area of the grid.
[0032] The calculation formula for the centroid tracking method is as follows: Formula for calculating the center of mass: ; ; in, Grid points The probability of lightning occurrence, and These are the row index and column index of the grid point, respectively. It is the x-coordinate of the centroid at time t. It is the ordinate of the centroid at time t; Formula for calculating movement speed: ; in, and These are the x and y coordinates of the centroid at time t. and These are the x and y coordinates of the centroid at time t-1, respectively. It's the grid resolution. It is the time interval between two consecutive time steps. This is the movement speed value; ; in, It is the displacement component of the center of mass in the x-direction. It is the displacement component of the center of mass in the y-direction. It indicates the direction of movement.
[0033] Assuming that the lightning activity area maintains its current movement vector and intensity trend over a short period of time, the position, velocity, and direction of the centroid can be used as the initial state for linear or nonlinear extrapolation. For example, the possible location of the centroid in 30 minutes or 60 minutes can be predicted, thus forming a predicted trajectory line, i.e., "trajectory prediction".
[0034] Decision support mapping is a comprehensive risk map integrating spatial location, time, intensity, and evolution trends. It overlays calculated historical centroid movement trajectories and predicted future trajectories onto a geographic base map or satellite cloud image. Based on the current lightning activity probability map, areas with probabilities exceeding a certain threshold are delineated. Along the predicted trajectory, combining the lightning activity intensity index L (reflecting current activity vitality) and movement speed, the potential impact range of risk areas in the future is estimated. For example, a buffer zone on both sides of the predicted trajectory line represents the "potentially affected area," and key decision information is directly marked on the map: different colors (e.g., red, orange, yellow) are used to mark current high-risk areas and potentially affected areas in the future; the estimated arrival time is marked at key nodes of the predicted trajectory (e.g., passing through a town); and arrows visually indicate the current direction and speed of movement. (Topographic lifting coefficient is also included.) This can be incorporated into decision-making as important environmental field information. For example, if there is [a certain condition] ahead of the predicted trajectory. High values for steep, windward slopes indicate that the terrain may strongly lift air currents, triggering or significantly intensifying thunderstorms. Decision maps can overlay terrain lifting coefficient layers on these areas or add special annotations (such as "high-risk terrain-triggered area") to alert decision-makers that thunderstorm activity in these areas may suddenly develop or intensify, requiring an increase in the warning level.
[0035] The multimodal fusion image generation process incorporates atmospheric physical constraints as weight adjustment factors to construct a dynamic attention-weighted fusion network, specifically including: First, the cloud top height is determined by inverting the brightness temperature of the infrared channel. The development index of the middle and upper-level wetlands was calculated by combining the water vapor channel gradient. And utilize visible light texture complexity to extract convection primary signals. Construct a physical sensing gate function: ; in, It is the first Dynamic weights of each channel It is the Sigmoid function. For learnable feature encoders, This represents a combination of empirical formulas driven by meteorological physical models. , These are the trainable parameters; The fused feature map is represented as follows: ; in, The features of each channel image after preprocessing, This represents element-wise multiplication; Ultimately, a Transformer module is used to achieve cross-channel context modeling, enhancing the information expression capabilities of key areas.
[0036] The lightning activity intensity index incorporates a local environmental modulation factor, specifically including: Calculate the topographic lift coefficient ,in, The gradient represents the terrain height field; it is a vector. Represents wind speed vector, topographic lift coefficient Used to reflect the forced uplift effect of windward slopes; Using the topographic lift coefficient and the intensity of the urban heat island The lightning activity intensity index has been modified as follows: ; in, and These are the calibration parameters for the region.
[0037] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for analyzing lightning activity information based on image recognition, characterized in that, Includes the following steps: S1. Acquire visible light image, infrared image and water vapor channel image of lightning, and fuse the three into a multimodal fusion image after standardization preprocessing; S2. Construct a dual-branch deep learning network structure, which includes a lightning activity recognition and segmentation network and a lightning spatiotemporal evolution prediction network. S3. Using a single multimodal fusion image as input, a probability segmentation map can be obtained through the lightning activity recognition and segmentation network. After multiple training sessions, a trained lightning activity recognition and segmentation network is obtained. Using a short sequence of multimodal images as input, a predicted image can be obtained through the lightning activity recognition and segmentation network. S4. Input the predicted image into the trained lightning activity recognition and segmentation network to obtain a lightning activity probability map based on the future predicted state. S5. Deep information mining is performed on the probability map and prediction image of lightning activity. The lightning activity intensity index is used as the vitality diagnosis of lightning activity. The centroid tracking method is used to obtain the motion trajectory prediction of lightning risk and the decision support map of early warning response.
2. The lightning activity information analysis method based on image recognition according to claim 1, characterized in that: The specific generation process of the multimodal fusion image is as follows: First, the original satellite images of visible light, infrared, and water vapor channels are standardized and preprocessed to unify them into the same spatiotemporal grid and numerical range. Then, a dynamic weighted fusion method based on physical mechanisms is adopted, that is, weights are assigned to each channel. Then, through principal component analysis or color synthesis technology, the information of the three channels is compressed and mapped into a new image space. Finally, a multimodal fusion image integrating cloud top texture, development intensity, and atmospheric water vapor conditions is generated.
3. The lightning activity information analysis method based on image recognition according to claim 1, characterized in that: The lightning activity recognition and segmentation network uses an encoder-decoder architecture as its basic structure. The encoder part consists of four downsampling stages, each of which contains two 3×3 convolutional layers, a ReLU activation function and a batch normalization layer, and finally a 2×2 max pooling layer for downsampling. The decoder part also contains four upsampling stages, each of which doubles the size of the feature map through transposed convolution and makes skip connections with the feature map of the corresponding layer of the encoder. The lightning activity identification and segmentation network is trained using a supervised learning approach with labels based on real future lightning observation data, and the Dice loss function is selected as the loss function.
4. The lightning activity information analysis method based on image recognition according to claim 1, characterized in that: The lightning spatiotemporal evolution prediction network uses a ConvLSTM network as the core to construct ConvLSTM units. Each ConvLSTM unit contains an input gate, a forget gate, an output gate, and a cell state, and all fully connected operations are replaced with convolutional operations. The lightning spatiotemporal evolution prediction network structure adopts an encoder-prediction framework. The encoder part consists of multiple stacked ConvLSTM layers, which progressively extract spatiotemporal features from the input sequence. The prediction part decodes the learned temporal evolution rules into cloud map predictions for multiple future time steps through deconvolutional layers and skip connections. During the training of the lightning spatiotemporal evolution prediction network, a combined loss function is used. The mean squared error ensures that the predicted cloud map is numerically close to the real cloud map, while the gradient difference loss specifically penalizes the problem of blurred prediction boundaries, making the predicted thunderstorm cloud outline clearer.
5. The lightning activity information analysis method based on image recognition according to claim 3, characterized in that: The formula for calculating the Dice loss function is as follows: ； in, It is a prediction probability. These are the actual labels, and N is the total number of pixels. It is a smoothing term to prevent division by zero.
6. The lightning activity information analysis method based on image recognition according to claim 4, characterized in that: The formula for calculating the combined loss function is as follows: Mean square error: ; Gradient difference loss: ; Combination loss function: ; in, This is the mean squared error value, where N is the total number of pixels. It is a predicted value. It is the actual value. It is the gradient difference loss value. It is the gradient operator of the image in the x-direction. It is the gradient operator of the image in the y-direction. It is the loss value of the combined loss function. yes The weighting coefficients, yes The weighting coefficients.
7. The lightning activity information analysis method based on image recognition according to claim 1, characterized in that: The formula for the lightning activity intensity index is: ； in, Grid points The probability of lightning occurrence, It represents the actual area that the grid point represents. It is the total area of regions with a probability exceeding 0.
5. It is the total area of the grid.
8. The lightning activity information analysis method based on image recognition according to claim 1, characterized in that: The calculation formula for the centroid tracking method is as follows: Formula for calculating the center of mass: ; ； in, Grid points The probability of lightning occurrence, and These are the row index and column index of the grid point, respectively. It is the x-coordinate of the centroid at time t. It is the ordinate of the centroid at time t; Formula for calculating movement speed: ; in, and These are the x and y coordinates of the centroid at time t. and These are the x and y coordinates of the centroid at time t-1, respectively. It's the grid resolution. It is the time interval between two consecutive time steps. This is the movement speed value; Formula for calculating direction of movement: ; in, It is the displacement component of the center of mass in the x-direction. It is the displacement component of the center of mass in the y-direction. It indicates the direction of movement.
9. The lightning activity information analysis method based on image recognition according to claim 2, characterized in that: The multimodal fusion image generation process incorporates atmospheric physical constraints as weight adjustment factors to construct a dynamic attention-weighted fusion network, specifically including: First, the cloud top height is determined by inverting the brightness temperature of the infrared channel. The development index of the middle and upper-level wetland area was calculated by combining the water vapor channel gradient. And utilize visible light texture complexity to extract convection primary signals. Construct a physical sensing gate function: ； in, It is the first Dynamic weights of each channel It is the Sigmoid function. For learnable feature encoders, This represents a combination of empirical formulas driven by meteorological physical models. , These are the trainable parameters; The fused feature map is represented as follows: ； in, The features of each channel image after preprocessing, This represents element-wise multiplication; Ultimately, a Transformer module is used to achieve cross-channel context modeling, enhancing the information expression capabilities of key areas.
10. The lightning activity information analysis method based on image recognition according to claim 7, characterized in that: The lightning activity intensity index incorporates a local environmental modulation factor, specifically including: Calculate the topographic lift coefficient ,in, The gradient represents the terrain height field; it is a vector. Represents wind speed vector, topographic lift coefficient Used to reflect the forced uplift effect of windward slopes; Using the topographic lift coefficient and the intensity of the urban heat island The lightning activity intensity index has been modified as follows: ； in, and These are the calibration parameters for the region.