Hail cloud feature analysis and identification method, system and device based on multi-source data and storage medium

Abstract

This invention discloses a method, system, device, and storage medium for hail cloud feature analysis and identification based on multi-source data. The method includes: obtaining satellite feature products, derived feature products, and radar feature products strongly correlated with the hail cloud formation and dissipation process; performing spatial registration processing on the satellite feature products, derived feature products, and radar feature products to obtain spatiotemporally aligned satellite feature sets and radar feature sets; constructing a dataset based on the spatiotemporally aligned satellite feature sets and radar feature sets; constructing a GSCF deep learning recognition model; and training and testing the model. This application effectively solves the problem of heterogeneous fusion of multi-source data, fully leverages the advantages of macroscopic meteorological information from satellite data and fine structural information from radar data, improves the quality and representational ability of fused features, provides reliable data support for accurate hail cloud identification, and is of great significance for improving disaster early warning efficiency.

Description

Technical Field

[0001] This invention belongs to the field of meteorological technology, and in particular relates to a method, system, device and storage medium for hail cloud feature analysis and identification based on multi-source data. Background Technology

[0002] Hail is a severe weather phenomenon caused by strong convection. my country experiences frequent hailstorms annually, making it one of the four most hail-prone regions in the world. As a serious meteorological disaster, hail significantly impacts agriculture, construction, and human activities. Due to the suddenness, localized nature, and short lifespan of hailstorms, they are difficult to predict in advance. Therefore, accurate and timely identification of hail clouds is crucial for mitigating the losses caused by hail disasters.

[0003] With the development of Doppler weather radar detection technology, research on convective weather nowcasting technology based on radar data has progressed rapidly in China. Radar data has the advantage of high spatiotemporal resolution in hail identification, especially in real-time monitoring and refined observation stages. Although it has achieved certain results in identifying hail clouds, there are still shortcomings, such as the detection range of a single radar, distance attenuation, and blind spots. Therefore, exploring new monitoring methods, such as satellite remote sensing technology, is of great significance in making up for the shortcomings of radar. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method, system, device and storage medium for hail cloud feature analysis and identification based on multi-source data, so as to solve the technical problem in the prior art that it is difficult to efficiently integrate satellite and radar multi-source data, thereby affecting the accuracy of hail cloud identification.

[0005] The technical solution adopted in this application to solve the above-mentioned technical problems is as follows: providing a method for hail cloud feature analysis and identification based on multi-source data, including: S1. Obtain satellite feature products, satellite product derivative feature products, and radar feature products that are strongly correlated with the hail cloud formation and dissipation process; S2. Perform spatial registration processing on satellite feature products, satellite product-derived feature products, and radar feature products to obtain spatiotemporally aligned satellite feature sets and radar feature sets; S3. Construct a dataset based on spatiotemporally aligned satellite and radar feature sets; S4. Construct the GSCF deep learning recognition model; S5. Model training and testing.

[0006] Optionally, in some embodiments of this application, the method for obtaining satellite signature products and radar signature products strongly correlated with the formation and dissipation process of hail clouds includes: S11. Obtain satellite products and radar data. S12. Based on satellite products, obtain derivative products of satellite products; S13. Obtain radar products based on radar data; S14. Based on satellite products, satellite product derivatives, and radar products, select satellite characteristic products, satellite product derivatives, and radar characteristic products that are strongly correlated with the hail cloud formation and dissipation process.

[0007] Optionally, in some embodiments of this application, satellite products and radar data are collected within a 30-minute window before and after the time T0 when the target hail event occurs, i.e., the time window of [T0-30min, T0+30min]; and / or The radar data is inverted using radar meteorological equations to calculate radar products, which include combined reflectivity (CR), echo top height (ET), and vertical cumulative liquid water (VIL); and / or Map satellite feature products and derived feature products of satellite products to the grid coordinates of radar feature products using interpolation methods; and / or The method for constructing the dataset includes: based on satellite and radar feature sets, combined with real-time ground-based hail observations, binary classification labeling is performed on the satellite and radar feature sets. Specifically: locations where hail actually occurred are labeled as "hail clouds," i.e., positive samples; locations where no hail occurred are labeled as "non-hail clouds," i.e., negative samples. The labeled dataset is then divided into training, validation, and test sets using a stratified sampling strategy, with a ratio of 8:1:1; and / or The GSCF deep learning recognition model includes a multi-channel input layer, a depthwise separable convolutional layer, a feature fusion layer, a convolutional feature extraction layer, and an output layer; and / or After the satellite product and radar data are processed through S1 and S2 as described above, they are used as input to the trained GSCF deep learning recognition model, and the recognition results are output.

[0008] Optionally, in some embodiments of this application, the satellite products include cloud type CLT, blackbody brightness temperature TBB, 0.65μm visible light band, 1.61μm shortwave infrared band; and / or The formula for calculating the combined reflectance CR is as follows: ; ; In the formula, It is the reflectivity factor. This is a category number for precipitation particles. It is the diameter of precipitation particles. Particle number concentration, For the first The reflectivity factor values ​​for each elevation angle layer, where N is the number of elevation angle layers. For the combined reflectivity values ​​within the radar detection grid; and / or The echo peak height ET is calculated using the following formula: ; ; ; ; ; In the formula, , They are respectively , The corresponding echo intensity value, where r is the radial distance. For the equivalent Earth radius, The altitude of the radar is [altitude]. This is the echo height of the radar's first elevation angle layer after correction for Earth curvature. This is the echo height of the radar's second elevation angle layer after correction for Earth curvature. The final calculated radar echo top height, The corrected height for the second elevation angle layer The weighting coefficients, The corrected height for the first elevation angle layer The weighting coefficients, The threshold for determining the echo top height; and / or The vertically accumulated liquid water volume (VIL) is calculated using the following formula: ; In the formula, Let i be the reflectivity factor value of the i-th elevation angle layer. The vertical height difference between the elevation angles of the (i+1)th and i-th layers is given. D The density of liquid water is typically taken as 1.0; and / or Methods for screening satellite signature products, satellite signature products derived from satellite products, and radar signature products that are strongly correlated with the formation and dissipation of hail clouds include: S141. Using the hailfall time T0 as a reference, the time window is... Feature values ​​corresponding to internal satellite products, satellite product derivatives and radar products are extracted and time series sequences are constructed. Based on the constructed time series sequences, the nearest neighbor time matching method is used to unify satellite products, satellite product derivatives and radar products to the same time scale, resulting in time-aligned time series sequences. S142. For continuous numerical and ratio-based categories, if a significant peak and trend appear near the hailfall time T0, retain the category; if there is no significant peak and the trend is stable before and after T0, remove the category; for cloud type categories, observe whether ice cloud type appears, and retain the cloud type category if ice cloud type exists; and / or The interpolation process is as follows: S21. Let the pixel coordinates of the satellite feature product and the derived feature product of the satellite product be... The corresponding pixel value The pixel coordinates of the radar feature product are The corresponding pixel value ; S22. Calculate the width of satellite feature products, derived feature products of satellite products, and radar feature products. The scaling factor for the direction is calculated using the following formula: ; in , It is the width of the radar feature product. It refers to the breadth of satellite-specific products and derivative products of satellite products; S23. Based on the scaling ratio, the pixel coordinates of the radar feature product are... The pixel coordinates mapped to satellite feature products and derived feature products of satellite products are: The formula for obtaining non-integer coordinates is as follows: ; in and The coordinates are the non-integer coordinates of the satellite feature product and its derived feature products. These non-integer coordinates are then rounded to obtain the nearest integer coordinates among the satellite feature product and its derived feature products. ,Right now: ; S24. Place satellite feature products and derivative feature products of satellite products on integer coordinates. pixel value at Directly assign the coordinates of the radar feature product Corresponding pixel value ,Right now ; S25. Perform the above interpolation operations on the satellite feature products and their derived feature products, and combine them with the radar feature products to obtain the spatiotemporally aligned satellite feature set and radar feature set; and / or The multi-channel input layer includes 7 channels of input data, corresponding to 3 channels for radar and 4 channels for satellite; and / or Depthwise separable convolutional layers include channel-wise convolution and pointwise convolution; and / or The feature fusion layer includes an adaptive channel weighting module and a gated global weighting fusion module; and / or The convolutional feature extraction layer comprises alternating stacks of convolutional blocks and max-pooling layers, wherein each convolutional block sequentially contains a convolutional layer, a batch normalization layer, and a ReLU activation function; and / or The output layer uses the Sigmoid activation function to output the probability value of each sample belonging to a hail cloud. The final output is a binary map of the spatial distribution of the hail cloud region.

[0009] Optionally, in some embodiments of this application, derivatives of the satellite product include the TBB spatial gradient and the 0.65μm / 1.61μm band ratio; and / or Continuous numerical categories include combined reflectance (CR), echo top height (ET), vertical cumulative liquid water (VIL), TBB spatial gradient, and 0.65 μm visible light reflectance; and / or Ratio-type categories include 0.65μm / 1.61μm band ratios; and / or Cloud type categories include CLT; and / or The radar has three channels, including CR, ET, and VIL; and / or The satellite's four channels include cloud type CLT, TBB spatial gradient, 0.65μm visible light band, and 0.65μm / 1.61μm band ratio; and / or After passing through the adaptive channel weighting module, features are extracted and dimensionality reduced by convolution and max pooling respectively. Then, the data is sent to the gated global weighted fusion module. This module generates a gated mask through global average pooling and fully connected layers, and outputs a weight vector with a value range of [0,1].

[0010] Optionally, in some embodiments of this application, the TBB spatial gradient is obtained by: extracting the spatial gradient of the blackbody brightness temperature TBB using the Sobel operator. The Sobel operator performs convolution operations with the blackbody brightness temperature TBB data using two 3×3 convolution kernels to extract the TBB spatial gradient information; assuming the blackbody brightness temperature TBB data is represented in two-dimensional matrix form as follows. For two-dimensional matrices For each pixel, calculate the gradient in the x-direction, the gradient in the y-direction, and the total gradient using the following formulas: ; ; ; ; ; In the formula, This represents the cloud top brightness temperature value in the i-th row and j-th column. The row index represents the y-direction. For column index, i.e., in the x-direction, where express The top left and top right neighboring pixel values, express The pixel values ​​of the lower left and lower right neighboring pixels. express The left and right neighboring pixel values, , express The upper and lower neighbor pixel values, This represents the gradient value in the x-direction. This represents the gradient value in the y-direction. This represents the total gradient value; The calculated The TBB spatial gradient is converted into a standardized raster format using a spatial reference system consistent with the blackbody brightness temperature (TBB) data.

[0011] Optionally, in some embodiments of this application, satellite feature products, derived feature products of satellite products, and radar feature products strongly correlated with the hail cloud formation and dissipation process include combined reflectivity CR, echo top height ET, vertical cumulative liquid water VIL, cloud type CLT, TBB spatial gradient, 0.65μm visible light band, and 0.65μm / 1.61μm band ratio.

[0012] Accordingly, embodiments of this application also provide a hail cloud feature analysis and identification system based on multi-source data, including: The strong correlation module is used to obtain satellite feature products, satellite product derivative feature products, and radar feature products that are strongly correlated with the hail cloud formation and dissipation process. The spatial registration module is used to perform spatial registration processing on satellite feature products, satellite product-derived feature products and radar feature products to obtain spatiotemporally aligned satellite feature sets and radar feature sets; The dataset acquisition module constructs a dataset based on spatiotemporally aligned satellite and radar feature sets; The model building module is used to build GSCF deep learning recognition models; The training and testing modules are used for training and testing the model.

[0013] Accordingly, embodiments of this application also provide a computer device, including a storage device and a processor, wherein the storage device stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0014] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of this application are as follows: its core contribution lies in overcoming the technical bottlenecks of difficulty in multi-source data fusion and insufficient recognition accuracy in existing hail cloud identification technologies. Through innovative fusion strategies and model improvements, it achieves greater accuracy and efficiency in hail cloud identification. The specific beneficial effects are as follows: Precise selection of core features lays a solid data foundation for hail cloud identification, significantly improving the timeliness and reliability of identification and early warning. This application uses satellite and radar data before and after actual hail events to calculate and invert satellite products, derivative products, and radar products. Based on this, it analyzes their dynamic changes within the hail time window, selecting satellite feature products, derivative feature products, and radar feature products strongly correlated with the hail cloud formation and dissipation process. This provides strong support for subsequent threshold setting and model training.

[0016] An adaptive-gated dual-fusion strategy is proposed, significantly improving the quality and efficiency of multi-source data fusion and laying a core foundation for enhancing recognition accuracy. This strategy differs from traditional single-fusion methods and simple feature splicing, highlighting its innovation: Through an adaptive fusion module, feature weights are dynamically allocated based on different modalities of satellite and radar data and the differences in information entropy across channels, prioritizing the retention of key information; a gated fusion module performs confidence screening and nonlinear enhancement on the initial fused features, filtering redundancy, suppressing noise, and improving data utilization. Compared to traditional single-fusion methods, this strategy offers more targeted and reliable fusion results, maximizing the collaborative value of multi-source data and effectively solving the problem of poor heterogeneous fusion performance of multi-source data.

[0017] This application optimizes deep learning recognition models, balancing recognition accuracy and computational efficiency, to meet the needs of high-dimensional feature processing through multi-source fusion. The optimization replaces traditional convolution with depthwise separable convolution, reducing the number of parameters and computational cost while maintaining feature extraction capabilities, thus improving training and inference efficiency and adapting to high-dimensional fusion feature processing.

[0018] Through the above innovations and optimizations, the recognition performance of this application is significantly better than that of the prior art: the GSCF model has an accuracy of 0.919 and an F1 score of 0.928, while the traditional CNN model has an accuracy of only 0.781 and an F1 score of 0.765. The GSCF model has better overall recognition performance. Attached Figure Description

[0019] Figure 1This is a flowchart of the hail cloud feature analysis and identification method based on multi-source data according to the present invention; Figure 2 This is a radar time-series sequence analysis diagram of the present invention; Figure 3 This is an analysis diagram of the cloud type CLT of this invention; Figure 4 This is a satellite time series analysis diagram of the present invention; Figure 5 This is a flowchart illustrating the feature fusion process of the present invention; Figure 6 This is a diagram of the depth-separable convolutional architecture of the present invention; Figure 7 This is a diagram of the convolutional feature extraction architecture of the present invention; Figure 8 This is a diagram showing the characteristic indicators and identification results of hailfall time according to the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] The technical solution of this application is as follows: Firstly, please refer to Figure 1 This application provides a method for hail cloud feature analysis and identification based on multi-source data, including: S1. Obtain satellite feature products, satellite product derivative feature products, and radar feature products that are strongly correlated with the hail cloud formation and dissipation process; S2. Perform spatial registration processing on satellite feature products, satellite product-derived feature products, and radar feature products to obtain spatiotemporally aligned satellite feature sets and radar feature sets; S3. Construct a dataset based on spatiotemporally aligned satellite and radar feature sets; S4. Construct the GSCF deep learning recognition model; S5. Model training and testing.

[0022] Satellite data and radar data have significant heterogeneity differences in spatiotemporal resolution and physical observation dimensions. Directly combining them can easily lead to data feature conflicts or information redundancy, making it impossible for the two types of data to complement each other and difficult to efficiently input into the recognition model to complete high-precision hail cloud recognition.

[0023] To overcome the aforementioned technical bottlenecks, this application proposes a method, system, device, and storage medium for hail cloud feature analysis and identification based on multi-source data. The core of this approach lies in constructing an adaptive-gated dual-fusion strategy: First, considering the heterogeneous characteristics of multi-source data, this application designs a fusion module embedded within the model architecture, abandoning the static fusion method of the traditional preprocessing stage. Second, an adaptive weighted fusion mechanism dynamically calculates the information entropy weights of features from each channel, achieving differentiated weighting of features from different satellite and radar products. Simultaneously, a gated fusion mechanism is used to perform confidence screening and nonlinear enhancement on the initially fused features, ultimately achieving end-to-end intelligent fusion of satellite-radar multi-channel data.

[0024] This application effectively solves the problem of heterogeneous fusion of multi-source data, fully leverages the advantages of macro-meteorological information from satellite data and fine-structure information from radar data, improves the quality and characterization capabilities of fusion features, provides reliable data support for accurate identification of hail clouds, and is of great significance for improving the efficiency of disaster early warning.

[0025] In S1: In some embodiments, a method for obtaining satellite signature products and radar signature products strongly correlated with hail cloud formation and dissipation processes includes: S11. Obtain satellite products and radar data. S12. Based on satellite products, obtain derivative products of satellite products; S13. Obtain radar products based on radar data; S14. Based on satellite products, satellite product derivatives, and radar products, select satellite characteristic products, satellite product derivatives, and radar characteristic products that are strongly correlated with the hail cloud formation and dissipation process.

[0026] In S11: In some embodiments, satellite product and radar data are collected 30 minutes before and after the time T0 when the target hail event occurs, i.e., the time window of [T0-30min, T0+30min].

[0027] In S12: Furthermore, the satellite products include cloud type CLT, blackbody brightness temperature TBB, 0.65μm visible light band, and 1.61μm shortwave infrared band.

[0028] Furthermore, satellite product derivatives include TBB spatial gradient and 0.65μm / 1.61μm band ratio.

[0029] Furthermore, the spatial gradient of the blackbody brightness temperature (TBB) is obtained, specifically by extracting the spatial gradient of the TBB using the Sobel operator. The Sobel operator performs convolution operations with the blackbody brightness temperature TBB data using two 3×3 convolution kernels to extract the TBB spatial gradient information. Let the blackbody brightness temperature TBB data be represented in two-dimensional matrix form as follows: For two-dimensional matrices For each pixel, calculate the gradient in the x-direction, the gradient in the y-direction, and the total gradient using the following formulas: ; ; ; ; ; In the formula, This represents the cloud top brightness temperature value in the i-th row and j-th column. The row index represents the y-direction. For column index, i.e., in the x-direction, where express The top left and top right neighboring pixel values, express The pixel values ​​of the lower left and lower right neighboring pixels. express The left and right neighboring pixel values, , express The upper and lower neighbor pixel values, This represents the gradient value in the x-direction. This represents the gradient value in the y-direction. This represents the total gradient value; The calculated The TBB spatial gradient is converted into a standardized raster format using a spatial reference system consistent with the blackbody brightness temperature (TBB) data.

[0030] It is understandable that the calculated result A standardized raster format was adopted by using a spatial reference system consistent with the blackbody brightness temperature (TBB) data to ensure that the gradient data and the blackbody brightness temperature (TBB) data correspond one-to-one in spatial location.

[0031] It is understandable that the two 3×3 convolution kernels correspond to the x and y directions respectively. By analyzing the spatial gradient of TBB, the active areas of strong convective clouds can be identified, thereby determining the potential hail area.

[0032] Furthermore, the 0.65μm / 1.61μm band ratio was obtained, specifically by comparing the 0.65μm visible light band with the 1.61μm shortwave infrared band.

[0033] It is understandable that this ratio is sensitive to the effective radius of particles at the cloud top and can be used to identify the evolution of particle phases in hail clouds.

[0034] In S13: Furthermore, the radar data is inverted through the radar meteorological equation to calculate radar products, which include combined reflectivity CR, echo top height ET, and vertical cumulative liquid water VIL.

[0035] It is understandable that the radar data is the raw echo data obtained using weather radar detection.

[0036] It is understandable that the combined reflectivity (CR) is the strongest reflectivity combination selected by the radar in all elevation angle scans, used to identify the strongest echo area of ​​hail clouds.

[0037] Furthermore, the combined reflectance CR is calculated using the following formula: ; ; In the formula, It is the reflectivity factor. This is a category number for precipitation particles. It is the diameter of precipitation particles. Particle number concentration, For the first The reflectivity factor values ​​for each elevation angle layer, where N is the number of elevation angle layers. This represents the combined reflectivity value within the radar detection grid.

[0038] It is understandable that the echo top height ET reflects the strongest vertical height of the radar echo signal and is positively correlated with the intensity of the updraft.

[0039] Furthermore, the echo peak height ET is calculated using the following formula: ; ; ; ; ; In the formula, , They are respectively , The corresponding echo intensity value, where r is the radial distance. For the equivalent Earth radius, The altitude of the radar is [altitude]. This is the echo height of the radar's first elevation angle layer after correction for Earth curvature. This is the echo height of the radar's second elevation angle layer after correction for Earth curvature. The final calculated radar echo top height, The corrected height for the second elevation angle layer The weighting coefficients, The corrected height for the first elevation angle layer The weighting coefficients, The threshold for determining the echo top height.

[0040] It is understandable that the vertically accumulated liquid water (VIL) converts the reflectivity factor value into an equivalent liquid water value, and a sudden increase in VIL usually indicates the start of hail.

[0041] Furthermore, the vertically accumulated liquid water volume (VIL) is calculated using the following formula: ; In the formula, Let i be the reflectivity factor value of the i-th elevation angle layer. The vertical height difference between the elevation angles of the (i+1)th and i-th layers is given. D Let be the density of liquid water, which is usually taken as 1.0.

[0042] In S14: In some embodiments, satellite feature products, satellite feature products derived from satellite products, and radar feature products strongly correlated with the hail cloud formation and dissipation process include combined reflectivity (CR), echo top height (ET), vertical cumulative liquid water (VIL), cloud type (CLT), TBB spatial gradient, 0.65 μm visible light band, and 0.65 μm / 1.61 μm band ratio.

[0043] It is understandable that, based on obtaining satellite products, satellite product derivatives, and radar products, further analysis of their dynamic changes within the hailfall time window is needed to screen out satellite characteristic products, satellite product derivative characteristic products, and radar characteristic products that are strongly correlated with the hail cloud formation and dissipation process.

[0044] In some embodiments, the method for screening satellite feature products, derived feature products of satellite products, and radar feature products that are strongly correlated with the hail cloud formation and dissipation process includes: S141. Using the hailfall time T0 as a reference, the time window is... Feature values ​​corresponding to internal satellite products, satellite product derivatives and radar products are extracted and time series sequences are constructed. Based on the constructed time series sequences, the nearest neighbor time matching method is used to unify satellite products, satellite product derivatives and radar products to the same time scale, resulting in time-aligned time series sequences. S142. For continuous numerical categories and ratio categories, if a significant peak appears near the hailfall time T0 and the trend of change is significant, then retain it; if there is no significant peak before and after T0 and the trend of change is stable, then remove it; for cloud type categories, observe whether ice cloud type appears, and if ice cloud type exists, then retain the cloud type category.

[0045] For example, to address the issue of inconsistent radar and satellite observation times (satellite observations every 15 minutes, radar observations every 6 minutes), the nearest neighbor time matching method is used to unify satellite products, satellite product derivatives, and radar products to the same time scale, resulting in a time-aligned time series.

[0046] It is understandable that, based on each satellite observation time, the time axis is divided into consecutive 15-minute intervals. If the radar observation time is within 6 minutes after the current satellite observation time, the radar product is matched with the satellite product and its derivative products at the current satellite observation time, thereby forming a time-aligned sequence.

[0047] For example, if the satellite observation time is between 0 minutes and 15 minutes, then a radar product with a radar observation time of 0 minutes to 6 minutes should be selected; if the satellite observation time is between 15 minutes and 30 minutes, then a radar product with a radar observation time of 12 minutes to 18 minutes should be selected.

[0048] In S142: In some embodiments, the continuous numerical categories include combined reflectance CR, echo top height ET, vertical cumulative liquid water VIL, TBB spatial gradient, and 0.65 μm visible light reflectance.

[0049] In some embodiments, the ratio type includes a 0.65μm / 1.61μm band ratio.

[0050] In some embodiments, cloud type categories include CLT.

[0051] In S2: In some embodiments, satellite feature products and derived feature products of satellite products are mapped to the grid coordinates of radar feature products through an interpolation method.

[0052] It is understandable that after interpolation, all feature products have the same spatial resolution and grid range.

[0053] Furthermore, the interpolation process is as follows: S21. Let the pixel coordinates of the satellite feature product and the derived feature product of the satellite product be... The corresponding pixel value The pixel coordinates of the radar feature product are The corresponding pixel value ; S22. Calculate the width of satellite feature products, derived feature products of satellite products, and radar feature products. The scaling factor for the direction is calculated using the following formula: ; in , It is the width of the radar feature product. It refers to the breadth of satellite-specific products and derivative products of satellite products; S23. Based on the scaling ratio, the pixel coordinates of the radar feature product are... The pixel coordinates mapped to satellite feature products and derived feature products of satellite products are: The formula for obtaining non-integer coordinates is as follows: ; in and The coordinates are the non-integer coordinates of the satellite feature product and its derived feature products. These non-integer coordinates are then rounded to obtain the nearest integer coordinates among the satellite feature product and its derived feature products. ,Right now: ; S24. Place satellite feature products and derivative feature products of satellite products on integer coordinates. pixel value at Directly assign the coordinates of the radar feature product Corresponding pixel value ,Right now ; S25. Perform the above interpolation operation on the satellite feature products and the derived feature products of the satellite products, and combine them with the radar feature products to obtain the spatiotemporally aligned satellite feature set and radar feature set.

[0054] In S3: In some embodiments, the method for constructing the dataset includes: based on satellite feature sets and radar feature sets, and combined with ground-based hail observation data, performing binary classification labeling on the satellite feature sets and radar feature sets. Specifically, the locations where hail actually occurred are labeled as "hail clouds," i.e., positive samples; the locations where no hail occurred are labeled as "non-hail clouds," i.e., negative samples. The labeled dataset is then divided into a training set, a validation set, and a test set according to a stratified sampling strategy, with a ratio of 8:1:1.

[0055] It is understandable that the training set is used for learning model parameters, the validation set is used for hyperparameter tuning and early stopping judgment, and the test set is used for independent evaluation of the final model performance.

[0056] In S4: Understandably, this application improves upon the CNN model by employing depthwise separable convolution and a dual fusion strategy, proposing a hail cloud recognition model based on a gated and self-adaptive convolution fusion network (GSCF).

[0057] Please see Figure 1 In some embodiments, the GSCF deep learning recognition model includes a multi-channel input layer, a depthwise separable convolutional layer, a feature fusion layer, a convolutional feature extraction layer, and an output layer.

[0058] Furthermore, the multi-channel input layer includes 7 channels of input data, corresponding to 3 radar channels and 4 satellite channels respectively.

[0059] Furthermore, the radar has three channels, including CR, ET, and VIL.

[0060] Furthermore, the satellite's four channels include cloud type CLT, TBB spatial gradient, 0.65μm visible light band, and 0.65μm / 1.61μm band ratio.

[0061] It is understandable that the traditional standard convolution operation is broken down into two independent steps: channel-wise convolution and point-wise convolution, which are performed sequentially.

[0062] Please see Figure 6 Furthermore, depth-separable convolutional layers include channel-wise convolution and point-wise convolution.

[0063] As can be understood, firstly, each channel of the input is convolved separately through channel-wise convolution, with each channel corresponding to a dedicated convolution kernel responsible only for extracting the feature information within that channel; then, pointwise convolution is performed using a 1×1 convolution kernel to fuse all channel feature maps output by channel-wise convolution, integrating the single features extracted from each channel to obtain the final feature output.

[0064] Compared to standard convolution, this decomposition design significantly reduces the number of model parameters and computational cost, lowers model training complexity, and enhances the model's ability to capture local details of hail clouds (such as the strong echo core region of hail clouds and cloud top texture gradient).

[0065] Please see Figure 5 Furthermore, the feature fusion layer includes an adaptive channel weighting module and a gated global weighting fusion module.

[0066] It is understandable that the adaptive channel weighting module adopts a channel-level adaptive weighting mechanism, using learnable weight parameters. (c=1,…,7, which is the number of channels for the corresponding modality), the contribution of each channel feature is dynamically adjusted; the weights are automatically optimized by the model during training based on the statistical characteristics of the input data, thereby strengthening the role of channels strongly correlated with hail clouds and suppressing the interference of noise channels.

[0067] Furthermore, after passing through the adaptive channel weighting module, features are extracted and dimensionality reduced through convolution and max pooling respectively, and then sent to the gated global weighted fusion module. This module generates a gated mask through global average pooling and fully connected layers, and outputs a weight vector with a value range of [0,1].

[0068] Please see Figure 7 Furthermore, the convolutional feature extraction layer includes multiple sets of alternating stacked convolutional blocks and max pooling layers, where each convolutional block sequentially contains a convolutional layer, a batch normalization layer, and a ReLU activation function.

[0069] It is understandable that convolutional feature extraction layers enhance the nonlinear representation of features while improving the stability and convergence speed of model training.

[0070] Furthermore, the output layer employs the Sigmoid activation function to output the probability value of each sample belonging to a hail cloud. The final output is a binary map of the spatial distribution of the hail cloud region.

[0071] Understandably, the final output is a binary map of the spatial distribution of the hail cloud region, indicating whether each location is identified as a hail cloud.

[0072] In S5: In some embodiments, the training set is input into the GSCF deep learning recognition model, the binary cross-entropy is used as the loss function, and the Adam optimizer is used for end-to-end training. During the training process, the model performance is tested using the validation set, and training is stopped when the validation set loss does not decrease for several consecutive epochs.

[0073] It is understandable that training should be stopped when the validation set loss does not decrease for several consecutive epochs to prevent overfitting. This application uses precision, hit rate, and F1 score as core performance evaluation metrics for dynamically optimizing model hyperparameters.

[0074] After training is complete, the model is evaluated using an independent test set.

[0075] The test results of this application show that the GSCF deep learning recognition model significantly outperforms the traditional CNN model in all core metrics of hail cloud recognition, demonstrating higher recognition accuracy.

[0076] In some embodiments, satellite products and radar data, after being processed by S1 and S2 as described above, are used as input to the trained GSCF deep learning recognition model, and the recognition result is output.

[0077] Secondly, embodiments of this application provide a hail cloud feature analysis and identification system based on multi-source data, including: The strong correlation module is used to obtain satellite feature products, satellite product derivative feature products, and radar feature products that are strongly correlated with the hail cloud formation and dissipation process. The spatial registration module is used to perform spatial registration processing on satellite feature products, satellite product-derived feature products and radar feature products to obtain spatiotemporally aligned satellite feature sets and radar feature sets; The dataset acquisition module constructs a dataset based on spatiotemporally aligned satellite and radar feature sets; The model building module is used to build GSCF deep learning recognition models; The training and testing modules are used for training and testing the model.

[0078] In the strongly correlated module In some embodiments, a method for obtaining satellite signature products and radar signature products strongly correlated with hail cloud formation and dissipation processes includes: S11. Obtain satellite products and radar data. S12. Based on satellite products, obtain derivative products of satellite products; S13. Obtain radar products based on radar data; S14. Based on satellite products, satellite product derivatives, and radar products, select satellite characteristic products, satellite product derivatives, and radar characteristic products that are strongly correlated with the hail cloud formation and dissipation process.

[0079] In S11: In some embodiments, satellite product and radar data are collected 30 minutes before and after the time T0 when the target hail event occurs, i.e., the time window of [T0-30min, T0+30min].

[0080] In S12: Furthermore, the satellite products include cloud type CLT, blackbody brightness temperature TBB, 0.65μm visible light band, and 1.61μm shortwave infrared band.

[0081] Furthermore, satellite product derivatives include TBB spatial gradient and 0.65μm / 1.61μm band ratio.

[0082] Furthermore, the spatial gradient of the blackbody brightness temperature (TBB) is obtained, specifically by extracting the spatial gradient of the TBB using the Sobel operator. The Sobel operator performs convolution operations with the blackbody brightness temperature TBB data using two 3×3 convolution kernels to extract the TBB spatial gradient information. Let the blackbody brightness temperature TBB data be represented in two-dimensional matrix form as follows: For two-dimensional matrices For each pixel, calculate the gradient in the x-direction, the gradient in the y-direction, and the total gradient using the following formulas: ; ; ; ; ; In the formula, This represents the cloud top brightness temperature value in the i-th row and j-th column. The row index represents the y-direction. For column index, i.e., in the x-direction, where express The top left and top right neighboring pixel values, express The pixel values ​​of the lower left and lower right neighboring pixels. express The left and right neighboring pixel values, , express The upper and lower neighbor pixel values, This represents the gradient value in the x-direction. This represents the gradient value in the y-direction. This represents the total gradient value; The calculated The TBB spatial gradient is converted into a standardized raster format using a spatial reference system consistent with the blackbody brightness temperature (TBB) data.

[0083] It is understandable that the calculated result A standardized raster format was adopted by using a spatial reference system consistent with the blackbody brightness temperature (TBB) data to ensure that the gradient data and the blackbody brightness temperature (TBB) data correspond one-to-one in spatial location.

[0084] It is understandable that the two 3×3 convolution kernels correspond to the x and y directions respectively. By analyzing the spatial gradient of TBB, the active areas of strong convective clouds can be identified, thereby determining the potential hail area.

[0085] Furthermore, the 0.65μm / 1.61μm band ratio was obtained, specifically by comparing the 0.65μm visible light band with the 1.61μm shortwave infrared band.

[0086] It is understandable that this ratio is sensitive to the effective radius of particles at the cloud top and can be used to identify the evolution of particle phases in hail clouds.

[0087] In S13: Furthermore, the radar data is inverted through the radar meteorological equation to calculate radar products, which include combined reflectivity CR, echo top height ET, and vertical cumulative liquid water VIL.

[0088] It is understandable that the radar data is the raw echo data obtained using weather radar detection.

[0089] It is understandable that the combined reflectivity (CR) is the strongest reflectivity combination selected by the radar in all elevation angle scans, used to identify the strongest echo area of ​​hail clouds.

[0090] Furthermore, the combined reflectance CR is calculated using the following formula: ; ; In the formula, It is the reflectivity factor. This is a category number for precipitation particles. It is the diameter of precipitation particles. Particle number concentration, For the first The reflectivity factor values ​​for each elevation angle layer, where N is the number of elevation angle layers. This represents the combined reflectivity value within the radar detection grid.

[0091] It is understandable that the echo top height ET reflects the strongest vertical height of the radar echo signal and is positively correlated with the intensity of the updraft.

[0092] Furthermore, the echo peak height ET is calculated using the following formula: ; ; ; ; ; In the formula, , They are respectively , The corresponding echo intensity value, where r is the radial distance. For the equivalent Earth radius, The altitude of the radar is [altitude]. This is the echo height of the radar's first elevation angle layer after correction for Earth curvature. This is the echo height of the radar's second elevation angle layer after correction for Earth curvature. The final calculated radar echo top height, The corrected height for the second elevation angle layer The weighting coefficients, The corrected height for the first elevation angle layer The weighting coefficients, The threshold for determining the echo top height.

[0093] It is understandable that the vertically accumulated liquid water (VIL) converts the reflectivity factor value into an equivalent liquid water value, and a sudden increase in VIL usually indicates the start of hail.

[0094] Furthermore, the vertically accumulated liquid water volume (VIL) is calculated using the following formula: ; In the formula, Let i be the reflectivity factor value of the i-th elevation angle layer. The vertical height difference between the elevation angles of the (i+1)th and i-th layers is given. D Let be the density of liquid water, which is usually taken as 1.0.

[0095] In S14: In some embodiments, satellite feature products, satellite feature products derived from satellite products, and radar feature products strongly correlated with the hail cloud formation and dissipation process include combined reflectivity (CR), echo top height (ET), vertical cumulative liquid water (VIL), cloud type (CLT), TBB spatial gradient, 0.65 μm visible light band, and 0.65 μm / 1.61 μm band ratio.

[0096] It is understandable that, based on obtaining satellite products, satellite product derivatives, and radar products, further analysis of their dynamic changes within the hailfall time window is needed to screen out satellite characteristic products, satellite product derivative characteristic products, and radar characteristic products that are strongly correlated with the hail cloud formation and dissipation process.

[0097] In some embodiments, the method for screening satellite feature products, derived feature products of satellite products, and radar feature products that are strongly correlated with the hail cloud formation and dissipation process includes: S141. Using the hailfall time T0 as a reference, the time window is... Feature values ​​corresponding to internal satellite products, satellite product derivatives and radar products are extracted and time series sequences are constructed. Based on the constructed time series sequences, the nearest neighbor time matching method is used to unify satellite products, satellite product derivatives and radar products to the same time scale, resulting in time-aligned time series sequences. S142. For continuous numerical categories and ratio categories, if a significant peak appears near the hailfall time T0 and the trend of change is significant, then retain it; if there is no significant peak before and after T0 and the trend of change is stable, then remove it; for cloud type categories, observe whether ice cloud type appears, and if ice cloud type exists, then retain the cloud type category.

[0098] For example, to address the issue of inconsistent radar and satellite observation times (satellite observations every 15 minutes, radar observations every 6 minutes), the nearest neighbor time matching method is used to unify satellite products, satellite product derivatives, and radar products to the same time scale, resulting in a time-aligned time series.

[0099] It is understandable that, based on each satellite observation time, the time axis is divided into consecutive 15-minute intervals. If the radar observation time is within 6 minutes after the current satellite observation time, the radar product is matched with the satellite product and its derivative products at the current satellite observation time, thereby forming a time-aligned sequence.

[0100] For example, if the satellite observation time is between 0 minutes and 15 minutes, then a radar product with a radar observation time of 0 minutes to 6 minutes should be selected; if the satellite observation time is between 15 minutes and 30 minutes, then a radar product with a radar observation time of 12 minutes to 18 minutes should be selected.

[0101] In S142: In some embodiments, the continuous numerical categories include combined reflectance CR, echo top height ET, vertical cumulative liquid water VIL, TBB spatial gradient, and 0.65 μm visible light reflectance.

[0102] In some embodiments, the ratio type includes a 0.65μm / 1.61μm band ratio.

[0103] In some embodiments, cloud type categories include CLT.

[0104] In the spatial registration module In some embodiments, satellite feature products and derived feature products of satellite products are mapped to the grid coordinates of radar feature products through an interpolation method.

[0105] It is understandable that after interpolation, all feature products have the same spatial resolution and grid range.

[0106] Furthermore, the interpolation process is as follows: S21. Let the pixel coordinates of the satellite feature product and the derived feature product of the satellite product be... The corresponding pixel value The pixel coordinates of the radar feature product are The corresponding pixel value ; S22. Calculate the width of satellite feature products, derived feature products of satellite products, and radar feature products. The scaling factor for the direction is calculated using the following formula: ; in , It is the width of the radar feature product. It refers to the breadth of satellite-specific products and derivative products of satellite products; S23. Based on the scaling ratio, the pixel coordinates of the radar feature product are... The pixel coordinates mapped to satellite feature products and derived feature products of satellite products are: The formula for obtaining non-integer coordinates is as follows: ; in and The coordinates are the non-integer coordinates of the satellite feature product and its derived feature products. These non-integer coordinates are then rounded to obtain the nearest integer coordinates among the satellite feature product and its derived feature products. ,Right now: ; S24. Place satellite feature products and derivative feature products of satellite products on integer coordinates. pixel value at Directly assign the coordinates of the radar feature product Corresponding pixel value ,Right now ; S25. Perform the above interpolation operation on the satellite feature products and the derived feature products of the satellite products, and combine them with the radar feature products to obtain the spatiotemporally aligned satellite feature set and radar feature set.

[0107] In the dataset acquisition module In some embodiments, the method for constructing the dataset includes: based on satellite feature sets and radar feature sets, and combined with ground-based hail observation data, performing binary classification labeling on the satellite feature sets and radar feature sets. Specifically, the locations where hail actually occurred are labeled as "hail clouds," i.e., positive samples; the locations where no hail occurred are labeled as "non-hail clouds," i.e., negative samples. The labeled dataset is then divided into a training set, a validation set, and a test set according to a stratified sampling strategy, with a ratio of 8:1:1.

[0108] It is understandable that the training set is used for learning model parameters, the validation set is used for hyperparameter tuning and early stopping judgment, and the test set is used for independent evaluation of the final model performance.

[0109] In the model construction module Understandably, this application improves upon the CNN model by employing depthwise separable convolution and a dual fusion strategy, proposing a hail cloud recognition model based on a gated and self-adaptive convolution fusion network (GSCF).

[0110] Please see Figure 1 In some embodiments, the GSCF deep learning recognition model includes a multi-channel input layer, a depthwise separable convolutional layer, a feature fusion layer, a convolutional feature extraction layer, and an output layer.

[0111] Furthermore, the multi-channel input layer includes 7 channels of input data, corresponding to 3 radar channels and 4 satellite channels respectively.

[0112] Furthermore, the radar has three channels, including CR, ET, and VIL.

[0113] Furthermore, the satellite's four channels include cloud type CLT, TBB spatial gradient, 0.65μm visible light band, and 0.65μm / 1.61μm band ratio.

[0114] It is understandable that the traditional standard convolution operation is broken down into two independent steps: channel-wise convolution and point-wise convolution, which are performed sequentially.

[0115] Please see Figure 6 Furthermore, depth-separable convolutional layers include channel-wise convolution and point-wise convolution.

[0116] As can be understood, firstly, each channel of the input is convolved separately through channel-wise convolution, with each channel corresponding to a dedicated convolution kernel responsible only for extracting the feature information within that channel; then, pointwise convolution is performed using a 1×1 convolution kernel to fuse all channel feature maps output by channel-wise convolution, integrating the single features extracted from each channel to obtain the final feature output.

[0117] Compared to standard convolution, this decomposition design significantly reduces the number of model parameters and computational cost, lowers model training complexity, and enhances the model's ability to capture local details of hail clouds (such as the strong echo core region of hail clouds and cloud top texture gradient).

[0118] Please see Figure 5 Furthermore, the feature fusion layer includes an adaptive channel weighting module and a gated global weighting fusion module.

[0119] It is understandable that the adaptive channel weighting module adopts a channel-level adaptive weighting mechanism, using learnable weight parameters. (c=1,…,7, which is the number of channels for the corresponding modality), the contribution of each channel feature is dynamically adjusted; the weights are automatically optimized by the model during training based on the statistical characteristics of the input data, thereby strengthening the role of channels strongly correlated with hail clouds and suppressing the interference of noise channels.

[0120] Furthermore, after passing through the adaptive channel weighting module, features are extracted and dimensionality reduced through convolution and max pooling respectively, and then sent to the gated global weighted fusion module. This module generates a gated mask through global average pooling and fully connected layers, and outputs a weight vector with a value range of [0,1].

[0121] Please see Figure 7 Furthermore, the convolutional feature extraction layer includes multiple sets of alternating stacked convolutional blocks and max pooling layers, where each convolutional block sequentially contains a convolutional layer, a batch normalization layer, and a ReLU activation function.

[0122] It is understandable that convolutional feature extraction layers enhance the nonlinear representation of features while improving the stability and convergence speed of model training.

[0123] Furthermore, the output layer employs the Sigmoid activation function to output the probability value of each sample belonging to a hail cloud. The final output is a binary map of the spatial distribution of the hail cloud region.

[0124] Understandably, the final output is a binary map of the spatial distribution of the hail cloud region, indicating whether each location is identified as a hail cloud.

[0125] In the training and testing modules, In some embodiments, the training set is input into the GSCF deep learning recognition model, the binary cross-entropy is used as the loss function, and the Adam optimizer is used for end-to-end training. During the training process, the model performance is tested using the validation set, and training is stopped when the validation set loss does not decrease for several consecutive epochs.

[0126] It is understandable that training should be stopped when the validation set loss does not decrease for several consecutive epochs to prevent overfitting. This application uses precision, hit rate, and F1 score as core performance evaluation metrics for dynamically optimizing model hyperparameters.

[0127] After training is complete, the model is evaluated using an independent test set.

[0128] The test results of this application show that the GSCF deep learning recognition model significantly outperforms the traditional CNN model in all core metrics of hail cloud recognition, demonstrating higher recognition accuracy.

[0129] In some embodiments, satellite products and radar data, after being processed by S1 and S2 as described above, are used as input to the trained GSCF deep learning recognition model, and the recognition result is output.

[0130] Thirdly, this application provides a computer device including a storage device and a processor. The storage device stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the hail cloud feature analysis and identification method based on multi-source data as described above.

[0131] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0132] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the computer device. Of course, the memory may include both the internal storage unit and the external storage device of the computer device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the hail cloud feature analysis and identification method based on multi-source data. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0133] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is used to run program code stored in the memory or process data, for example, to run program code for a hail cloud feature analysis and identification method based on multi-source data.

[0134] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the hail cloud feature analysis and identification method based on multi-source data as described above.

[0135] The computer-readable storage medium stores an interface display program that can be executed by at least one processor to cause the at least one processor to perform the steps of the hail cloud feature analysis and identification method based on multi-source data as described above.

[0136] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the hail cloud feature analysis and identification method based on multi-source data described in the embodiments of this application.

[0137] Application examples This method is applied in the fields of weather radar, FY4-B satellite data processing, and hail cloud identification. (See also...) Figure 1 , Figure 1 This is a schematic diagram of the overall process of a hail cloud identification method according to an application example of this application. The identification method completes hail cloud identification through three core steps: data analysis of multi-source data, spatial registration, and model identification. The specific implementation methods of each step are as follows: This application example conducts time-series analysis on satellite products, satellite product derivatives, and radar products, identifying seven core characteristic indicators strongly correlated with hail clouds, laying the foundation for subsequent identification.

[0138] 1. Selection of data sources Satellite products: L1-level radiance data (0.65μm visible light band, 1.61μm shortwave infrared band) from FY-4B satellite (Table 1), L2-level products (cloud type CLT, cloud top temperature, blackbody brightness temperature TBB), and derivative products of satellite products (TBB spatial gradient and 0.65μm / 1.61μm band ratio) were selected to cover the cloud top microphysical, optical reflection, and spatial distribution characteristics required for hail cloud identification; Radar data: Radar data from a C-band single-polarization Doppler weather radar were selected. The inverted products include combined reflectivity (CR), echo top height (ET), and vertical integrated liquid water content (VIL), which accurately characterize key features of hail clouds such as vertical dynamics and liquid water accumulation. Table 1 Performance parameters of some channels of FY-4BAGRI

[0139] 2. Radar time series analysis (see...) Figure 2 ) Box plot statistical analysis was conducted on three indicators—CR, ET, and VIL—30 minutes before and after hailfall to explore their temporal evolution patterns during the hail process. The characteristics of each indicator are as follows: Combined reflectance CR ( Figure 2 a) 30 minutes before hail, the box continuously moved upward, and the median and mean rose synchronously, indicating that CR showed a significant upward trend overall; the box position reached its peak at the moment of hail, and the median and mean were the highest in the entire period, indicating that the strong reflection characteristics of the hail cloud were most significant at this time; 30 minutes after hail, the box rapidly moved downward, and the value showed a downward trend, which is consistent with the physical process of hail particles falling and convection intensity weakening; Echo top height ET ( Figure 2 b): 30 minutes before hail falls, the box gradually moves upward and its width increases, indicating that the updraft intensity is increasing, the overall echo height is rising and the dispersion is increasing; at the moment of hail falls, the box is at its highest position and widest, reflecting that the strong updraft lifts the hailstones to grow fully, providing the core driving force for hail formation; 30 minutes after hail falls, the box slowly moves downward and gradually narrows, representing that the updraft is weakening, the overall echo height is decreasing and the characteristics tend to stabilize. Vertical integral liquid water content (VIL) Figure 2 c): The box height increases continuously 30 minutes before hail, indicating that the vertical liquid water in the cloud accumulates rapidly, providing a material basis for hail formation; the box height reaches its peak at the moment of hail, and the liquid water content is the highest during the entire period; the box height decreases rapidly 30 minutes after hail, and the liquid water content drops sharply, reflecting the rapid consumption of liquid water by hail particles during formation and descent, and the "jump phenomenon" of VIL can serve as a key early warning signal for the start of hail.

[0140] 3. Satellite time series analysis (see...) Figure 3 , Figure 4 ) Time-series analysis was conducted on four indicators—cloud type CLT, TBB spatial gradient, 0.65 μm visible light band, and 0.65 / 1.61 μm band ratio—30 minutes before and after hailfall. The results for each indicator are as follows: Cloud type CLT ( Figure 3 30 minutes before hail: The hail-prone area is mainly composed of supercooled water clouds. Figure 3 a) 15 minutes before hail, it gradually develops into a mixed cloud (a mixture of supercooled water droplets and ice crystals). Figure 3 b); Obvious ice clouds appeared at the time of hail ( Figure 3 c); Full sample frequency statistics show ( Figure 3 d) The number of occurrences of mixed clouds and ice clouds increases continuously as hail approaches. Among them, the number of occurrences of ice clouds at the moment of hail is more than twice that in the 30 minutes before hail, which is the core indicator of hail cloud development. TBB spatial gradient ( Figure 4a): 30 minutes before the hailstorm, the container showed no significant changes, and the spatial distribution of strong convective activity was relatively dispersed; at the moment of hailstorm, the container moved significantly upward, indicating that the spatial change of TBB was greatly enhanced, and the activity of strong convective clouds was more concentrated, which could accurately locate the core area of ​​hailstorm; after the hailstorm, the container moved rapidly downward, and the spatial concentration of strong convective activity weakened. 0.65μm visible light band ( Figure 4 b): 30 minutes before the hail, the box continued to move upward and the reflectivity gradually increased; the box reached its peak at the moment of hail, with the most ice crystals and supercooled water droplets on the cloud top, and the most significant reflectivity; after the hail, although the box moved downward, it was still higher than the level before the hail, indicating that although the cloud top convection characteristics weakened after the hail, it still maintained strong microphysical characteristics. 0.65 / 1.61μm band ratio ( Figure 4 c): 30 minutes before the hailstorm, the box steadily moved upward, and the difference in reflection characteristics across different bands gradually increased; the box reached its peak at the moment of hailstorm, and the difference in reflection characteristics between the cloud top and the two bands was the most significant, with a good correspondence between the high ratio area and the hailstorm area; after the hailstorm, the box moved slightly downward, but its overall position remained high, indicating that the optical characteristics of the cloud top microphysical structure remained at a high level after the hailstorm ended.

[0141] 4. Determination of strong correlation characteristics Using the peak value of the feature value at time T0 as the core screening condition, three radar feature products (CR, ET, VIL) and four satellite feature products, as well as the derivative feature products of the satellite products (cloud type CLT, TBB spatial gradient, 0.65μm visible light band, 0.65 / 1.61μm band ratio) were finally determined as strongly correlated feature indicators for hail cloud identification. The indicator values ​​before and after hail are shown in Table 2.

[0142] Table 2. Data characteristics of different products at different time points

[0143] Due to the inherent difference in detection resolution between the FY-4B satellite and C-band radar, spatial registration processing is required for the selected satellite feature products, derived feature products from the satellite products, and radar feature products. This ensures a one-to-one spatial correspondence between satellite and radar features for the same hail event, providing a unified dataset foundation for multi-source data fusion and identification. The specific implementation is as follows: Spatial registration: Interpolation is used to unify the resolution. The spatial resolution of satellite feature products and their derived feature products is 4 kilometers, while the spatial resolution of radar feature products is 1 kilometer. In this embodiment, the 1-kilometer high resolution of radar feature products is used as the target benchmark. Spatial resampling is performed on satellite feature products and their derived feature products using interpolation: the low-resolution raster data of satellite feature products and their derived feature products are mapped to the high-resolution grid of radar feature products. This is achieved through calculation... Scaling ratio of direction ( The target grid coordinates are reverse-mapped to the satellite feature product. The data coordinates of the derived feature product data of the satellite product are rounded to find the nearest neighbor satellite pixel and assigned a value, thus realizing the satellite feature product. The spatial resolution of the satellite product's derived feature product and the radar feature product is unified to ensure that their spatial positions correspond one-to-one. Finally, the spatiotemporally aligned satellite feature set and radar feature set are obtained.

[0144] Based on satellite and radar feature sets, and combined with ground-based hail observations, binary classification labels were applied to the satellite and radar feature sets: locations where hail actually occurred were labeled as "hail clouds" (positive samples), and locations where no hail occurred were labeled as "non-hail clouds" (negative samples). The labeled dataset was then divided into training, validation, and test sets using a stratified sampling strategy, with a ratio of 8:1:1. The dataset was input into the GSCF deep learning recognition model to perform intelligent hail cloud recognition, and the performance was compared with that of a traditional CNN model (Table 3).

[0145] Table 3 Results of different model indicators

[0146] 1. Dataset preprocessing and model training Based on measured hail data, the satellite feature set and radar feature set are labeled with binary classification (hail cloud sample / non-hail cloud sample), and are divided into training set, validation set and independent test set in a ratio of 8:1:1 to ensure that the ratio of hail cloud / non-hail cloud sample is consistent in each set. The training set is input into the GSCF model for training. The model hyperparameters (such as the number of convolutional layers, learning rate, fusion weights, etc.) are optimized through the validation set. The model performance is verified by metrics such as precision, recall, and F1 score, and finally the optimal training model is obtained. The model achieves independent extraction of 7 strongly correlated feature products through a multi-channel parallel input layer, and utilizes depthwise separable convolution (…). Figure 5 This reduces the number of model parameters and computational cost, lowers model training complexity, and enhances the model's ability to capture local details of hail clouds through a feature fusion layer. Figure 6 This method achieves deep complementary fusion of radar and satellite, enhances the nonlinear expression capability of features through convolutional feature extraction layers, improves the stability and convergence speed of model training, and finally outputs the probability value of each sample as a hail cloud and a spatial distribution binary map, thereby realizing spatial identification of hail cloud regions.

[0147] 2. Verification of recognition performance (see...) Figure 8 ) Hailfall timing recognition effect ( Figure 8 ): Figure 8The image shows the characteristic indicators of hailfall (a~g) and the identification results (h) (the identified hailfall areas are marked in red, and black dots represent the actual hailfall areas). The identified areas in the image include the actual hailfall areas, indicating that the model correctly identified the hail cloud areas. Compared with the CNN model, as shown in Table 3, the GSCF model outperforms the CNN model in terms of accuracy, hit rate, and F1 score.

[0148] This embodiment is based on a multi-source data fusion identification method using FY-4B satellite and C-band radar. It overcomes the technical obstacles of heterogeneous data fusion between satellite and radar, fully leverages the advantages of both data sources, improves the accuracy of hail cloud identification, and ultimately provides efficient and reliable technical support for hail disaster prevention and mitigation.

[0149] In this embodiment, the terms used are explained as follows: Visible light band: This is a specific signal channel used in satellites or remote sensing equipment to receive and process electromagnetic radiation in the visible light band with a wavelength of 0.65μm. When convective clouds develop vigorously, the number of particles such as ice crystals and supercooled water droplets on the cloud top increases. These particles have a strong reflective effect on light in the 0.65μm band, resulting in an increase in the reflectivity of this channel.

[0150] Shortwave infrared band: This wavelength of radiation is extremely sensitive to the absorption and reflection characteristics of water molecules, vegetation moisture content, and some minerals, making it a key distinguishing feature from the visible light band. In this band, the scattering and absorption characteristics of cloud particles are closely related to the microphysical structure of clouds.

[0151] Cloud type: Based on the microphysical structure and thermodynamic properties of clouds, and utilizing the different effective absorption optical thickness ratios of different types and phases of clouds in the four infrared channels, a full-disk cloud top type is generated. Nominal cloud type products include: warm (liquid) water clouds, supercooled water clouds, mixed clouds, opaque ice clouds, cirrus clouds (i.e., semi-transparent ice clouds), and multilayered clouds (semi-transparent upper layer, opaque lower layer), as well as undetermined types. Hailfall is usually closely related to strong convective cloud systems. Clouds like cumulonimbus, which are considered vigorously developing convective clouds in cloud type classification, often have the potential to produce hail.

[0152] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for hail cloud feature analysis and identification based on multi-source data, characterized in that, include: S1. Obtain satellite feature products, satellite product derivative feature products, and radar feature products that are strongly correlated with the hail cloud formation and dissipation process; S2. Perform spatial registration processing on satellite feature products, satellite product-derived feature products, and radar feature products to obtain spatiotemporally aligned satellite feature sets and radar feature sets; S3. Construct a dataset based on spatiotemporally aligned satellite and radar feature sets; S4. Construct the GSCF deep learning recognition model; S5. Model training and testing.

2. The hail cloud feature analysis and identification method based on multi-source data according to claim 1, characterized in that, Methods for obtaining satellite and radar signature products strongly correlated with hail cloud formation and dissipation processes include: S11. Obtain satellite product and radar data S12. Based on satellite products, obtain derivative products of satellite products; S13. Obtain radar products based on radar data; S14. Based on satellite products, satellite product derivatives, and radar products, select satellite characteristic products, satellite product derivatives, and radar characteristic products that are strongly correlated with the hail cloud formation and dissipation process.

3. The hail cloud feature analysis and identification method based on multi-source data according to claim 2, characterized in that, Collect satellite products and radar data for 30 minutes before and after the target hail event occurs (T0), i.e., within the time window of [T0-30min, T0+30min]; and / or The radar data is inverted using radar meteorological equations to calculate radar products, which include combined reflectivity (CR), echo top height (ET), and vertical cumulative liquid water (VIL); and / or The satellite feature products and their derived feature products are mapped to the grid coordinates of the radar feature products using an interpolation method. and / or The method for constructing the dataset includes: based on satellite and radar feature sets, combined with real-time ground-based hail observations, binary classification labeling is performed on the satellite and radar feature sets. Specifically: locations where hail actually occurred are labeled as "hail clouds," i.e., positive samples; locations where no hail occurred are labeled as "non-hail clouds," i.e., negative samples. The labeled dataset is then divided into training, validation, and test sets using a stratified sampling strategy, with a ratio of 8:1:1; and / or The GSCF deep learning recognition model includes a multi-channel input layer, a depthwise separable convolutional layer, a feature fusion layer, a convolutional feature extraction layer, and an output layer; and / or After the satellite product and radar data are processed through S1 and S2 as described above, they are used as input to the trained GSCF deep learning recognition model, and the recognition results are output.

4. The hail cloud feature analysis and identification method based on multi-source data according to claim 3, characterized in that, Satellite products include cloud type CLT, blackbody brightness temperature TBB, 0.65μm visible light band, 1.61μm shortwave infrared band; and / or The formula for calculating the combined reflectance CR is as follows: ； ； In the formula, It is the reflectivity factor. This is a category number for precipitation particles. It is the diameter of precipitation particles. Particle number concentration, For the first The reflectivity factor values ​​for each elevation angle layer, where N is the number of elevation angle layers. For the combined reflectivity values ​​within the radar detection grid; and / or The echo peak height ET is calculated using the following formula: ； ； ； ； ； In the formula, , They are respectively , The corresponding echo intensity value, where r is the radial distance. For the equivalent Earth radius, The altitude of the radar is [altitude]. This is the echo height of the radar's first elevation angle layer after correction for Earth curvature. This is the echo height of the radar's second elevation angle layer after correction for Earth curvature. The final calculated radar echo top height, The corrected height for the second elevation angle layer The weighting coefficients, The corrected height for the first elevation angle layer The weighting coefficients, The threshold for determining the echo top height; and / or The vertically accumulated liquid water volume (VIL) is calculated using the following formula: ； In the formula, Let i be the reflectivity factor value of the i-th elevation angle layer. The vertical height difference between the elevation angles of the (i+1)th and i-th layers is given. D The density of liquid water is typically taken as 1.0; and / or Methods for screening satellite signature products, satellite signature products derived from satellite products, and radar signature products that are strongly correlated with the formation and dissipation of hail clouds include: S141. Using the hailfall time T0 as a reference, the time window is... Feature values ​​corresponding to internal satellite products, satellite product derivatives and radar products are extracted and time series sequences are constructed. Based on the constructed time series sequences, the nearest neighbor time matching method is used to unify satellite products, satellite product derivatives and radar products to the same time scale, resulting in time-aligned time series sequences. S142. For continuous numerical categories and ratio categories, if a significant peak and a marked trend of change occur near the hailfall time T0, the category is retained; if there is no significant peak and a stable trend of change before and after T0, the category is removed; for cloud type categories, observe whether ice cloud type appears, and if ice cloud type exists, the cloud type category is retained; and / or The interpolation process is as follows: S21. Let the pixel coordinates of the satellite feature product and the derived feature product of the satellite product be... The corresponding pixel value The pixel coordinates of the radar feature product are The corresponding pixel value ; S22. Calculate the width of satellite feature products, derived feature products of satellite products, and radar feature products. The scaling factor for the direction is calculated using the following formula: ； in , It is the width of the radar feature product. It refers to the breadth of satellite-specific products and derivative products of satellite products; S23. Based on the scaling ratio, the pixel coordinates of the radar feature product The pixel coordinates mapped to satellite feature products and derived feature products of satellite products are: The formula for obtaining non-integer coordinates is as follows: ; in and The coordinates are the non-integer coordinates of the satellite feature product and its derived feature products. These non-integer coordinates are then rounded to obtain the nearest integer coordinates among the satellite feature product and its derived feature products. ,Right now: ； S24. Place satellite feature products and derived feature products of satellite products on integer coordinates. pixel value at Directly assign the coordinates of the radar feature product Corresponding pixel value ,Right now ; S25. Perform the above interpolation operations on satellite feature products and derived feature products of satellite products, and combine them with radar feature products to obtain spatiotemporally aligned satellite feature sets and radar feature sets; and / or The multi-channel input layer includes 7 channels of input data, corresponding to 3 channels for radar and 4 channels for satellite; and / or Depthwise separable convolutional layers include channel-wise convolution and pointwise convolution; and / or The feature fusion layer includes an adaptive channel weighting module and a gated global weighting fusion module; and / or The convolutional feature extraction layer comprises alternating stacks of convolutional blocks and max-pooling layers, wherein each convolutional block sequentially contains a convolutional layer, a batch normalization layer, and a ReLU activation function; and / or The output layer uses the Sigmoid activation function to output the probability value of each sample belonging to a hail cloud. The final output is a binary map of the spatial distribution of the hail cloud region.

5. The hail cloud feature analysis and identification method based on multi-source data according to claim 4, characterized in that, Derivatives of satellite products include TBB spatial gradients and 0.65μm / 1.61μm band ratios; and / or Continuous numerical categories include combined reflectance (CR), echo top height (ET), vertical cumulative liquid water (VIL), TBB spatial gradient, and 0.65 μm visible light reflectance; and / or Ratio-type categories include 0.65μm / 1.61μm band ratios; and / or Cloud type categories include CLT; and / or The radar has three channels, including CR, ET, and VIL; and / or The satellite's four channels include cloud type CLT, TBB spatial gradient, 0.65μm visible light band, and 0.65μm / 1.61μm band ratio; and / or After passing through the adaptive channel weighting module, features are extracted and dimensionality reduced by convolution and max pooling respectively. Then, the data is sent to the gated global weighted fusion module. This module generates a gated mask through global average pooling and fully connected layers, and outputs a weight vector with a value range of [0,1].

6. The method for hail cloud feature analysis and identification based on multi-source data according to claim 5, characterized in that, The TBB spatial gradient is obtained specifically by extracting the spatial gradient of the blackbody brightness temperature (TBB) using the Sobel operator. The Sobel operator performs convolution operations with the blackbody brightness temperature TBB data using two 3×3 convolution kernels to extract the TBB spatial gradient information. Let the blackbody brightness temperature TBB data be represented in two-dimensional matrix form. For two-dimensional matrices For each pixel, calculate the gradient in the x-direction, the gradient in the y-direction, and the total gradient using the following formulas: ； ； ； ； ； In the formula, This represents the cloud top brightness temperature value in the i-th row and j-th column. The row index represents the y-direction. For column index, i.e., in the x-direction, where express The top left and top right neighboring pixel values, express The pixel values ​​of the lower left and lower right neighboring pixels. express The left and right neighboring pixel values, , express The upper and lower neighbor pixel values, This represents the gradient value in the x-direction. This represents the gradient value in the y-direction. This represents the total gradient value; The calculated The TBB spatial gradient is converted into a standardized raster format using a spatial reference system consistent with the blackbody brightness temperature (TBB) data.

7. The method for hail cloud feature analysis and identification based on multi-source data according to claim 1, characterized in that, Satellite characteristic products, satellite product derivative characteristic products, and radar characteristic products strongly correlated with hail cloud formation and dissipation processes include combined reflectivity (CR), echo top height (ET), vertical cumulative liquid water (VIL), cloud type (CLT), TBB spatial gradient, 0.65μm visible light band, and 0.65μm / 1.61μm band ratio.

8. A hail cloud feature analysis and identification system based on multi-source data, characterized in that, include: The strong correlation module is used to obtain satellite feature products, satellite product derivative feature products, and radar feature products that are strongly correlated with the hail cloud formation and dissipation process. The spatial registration module is used to perform spatial registration processing on satellite feature products, derived feature products of satellite products, and radar feature products to obtain spatiotemporally aligned satellite feature sets and radar feature sets; The dataset acquisition module constructs a dataset based on spatiotemporally aligned satellite and radar feature sets; The model building module is used to build GSCF deep learning recognition models; The training and testing modules are used for training and testing the model.

9. A computer device, characterized in that, It includes a storage device and a processor, the storage device storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1-7.

Images