A hail identification method based on U-Net and multi-modal data fusion
By fusing U-Net convolutional neural network with multimodal data, the problem of effectively fusing multi-source data in hail identification is solved, achieving high-precision hail identification and improving the accuracy and robustness of hail identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHENGDU YUANWANG DETECTION TECH CO LTD
- Filing Date
- 2026-04-15
- Publication Date
- 2026-07-21
Smart Images
Figure CN122045957B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hail recognition technology, and more specifically to a hail recognition method based on U-Net and multimodal data fusion. Background Technology
[0002] Hail is a highly destructive meteorological disaster in severe convective weather, characterized by its strong locality, significant suddenness, and high risk of causing damage. It poses a serious and direct threat to agricultural production, infrastructure construction, and the safety of people's lives and property. Efficient and accurate hail identification and early warning technologies are not only a core component in improving the monitoring and early warning capabilities of severe convective weather and effectively reducing hail damage losses, but also a key and challenging problem that urgently needs to be addressed in the field of meteorological disaster research.
[0003] Currently, hail monitoring primarily relies on three observation methods: weather radar, numerical models, and meteorological satellites. While each type of equipment has its own technological advantages, it also has significant and specific limitations in its application. Weather radar can achieve refined detection of hail clouds, effectively capturing the dynamic changes in the internal dynamics and microphysical structure of the cloud. Radar data, such as echo intensity, is crucial for hail identification, but it is easily affected by ground obstruction, resulting in impulse noise. Numerical models can retrieve key atmospheric environmental background information such as temperature and dynamic fields, reflecting the dynamic conditions for hail formation, but they suffer from inherent biases in numerical forecasting systems. Meteorological satellites can acquire macroscopic weather features such as cloud top brightness and temperature from a wide-ranging upper-air perspective, reflecting the macroscopic evolution of hail clouds, but they have low spatial resolution and blurred cloud boundaries. Furthermore, due to multiple factors such as data dimensionality, spatiotemporal resolution, and differences in observation mechanisms, the heterogeneity of the three types of observation data is significant, and it is difficult to achieve accurate spatial alignment and temporal matching of data from different sources, further preventing their direct and effective fusion for hail identification.
[0004] Currently, the mainstream hail identification technologies in the industry are mainly divided into two categories. One category uses single observation data as model input. This type of method often suffers from inaccurate hail identification rates due to incomplete information on hail cloud development. The other category integrates multi-source data as input, but still uses traditional fuzzy logic methods to mine relevant features of various types of data, failing to fully leverage the complementary advantages of multi-source information. Summary of the Invention
[0005] The purpose of this invention is to provide a hail identification method based on U-Net and multimodal data fusion in order to solve the above-mentioned technical problems.
[0006] To achieve the above objectives, the present invention specifically adopts the following technical solution:
[0007] This invention provides a hail identification method based on U-Net and multimodal data fusion, comprising the following steps:
[0008] S1. Input the raw satellite data, raw radar data, and raw model data into the data dimensionality reduction and resampling module;
[0009] S2. The data dimensionality reduction and resampling module performs dimensionality reduction and resampling on the three-dimensional data in the input data to obtain a two-dimensional planar structure (the original satellite data is a two-dimensional planar structure that can be used directly, so no dimensionality reduction processing is required; the original radar data and the original model data contain three-dimensional data, so both the original radar data and the original model data need to be dimensionality reduced and resampling to obtain a two-dimensional planar structure in order to achieve uniformity of all data dimensions). The principle of dimensionality reduction and resampling is data time alignment and spatial resampling, and extracts the satellite modal data, radar modal data and model modal data that have been dimensionality reduced and resampling.
[0010] S3. Based on satellite modal data, radar modal data, and mode modal data, construct a multi-source modal weighted adaptive hybrid loss function;
[0011] S4. Based on the multi-source modality weighted adaptive hybrid loss function in step S3, construct and train the U-Net convolutional neural network model;
[0012] S5. Identify hail based on the U-Net convolutional neural network model and output the identification results.
[0013] In one implementation, the method for dimensionality reduction extraction of the original radar data in step S2 is as follows:
[0014] The radar data is volume scan data, divided into multiple elevation angles, each containing reflectivity data and dual polarization parameter data. From the perspective of hail observation, hail clouds often appear as strong convective cells. Therefore, the optimal solution input to the model is the radar data with the strongest reflectivity among the strong convective cells. The specific method for dimensionality reduction extraction of radar data is as follows:
[0015] S21. The radar volume scan range is a circular scanning area centered on the radar station. A latitude and longitude network is constructed with a resolution of 0.02 degrees to ensure that the latitude and longitude grid completely covers the radar scanning area.
[0016] S22. For each latitude and longitude grid point, iterate through the reflectivity data of all elevation angle layers, all radial and distance databases to find the strongest reflectivity at all elevation angles, and record the strongest reflectivity. and their corresponding positions ,in, For elevation angle index, Radial index, For distance database indexing, via turn up The strongest reflectivity at the corresponding location ,in Indicates differential reflectivity, Indicates differential phase shift rate, Indicates the position corresponding to the correlation coefficient;
[0017] S23. After all grid points within the latitude and longitude grid have been traversed, data such as the combined reflectivity, differential emissivity, differential phase shift rate, correlation coefficient, and echo top height of the two-dimensional planar structure are obtained.
[0018] In one implementation, the method for dimensionality reduction extraction of the original pattern data in step S2 is as follows:
[0019] The temperature field and dynamic field elements in the original model data are all three-dimensional data. ,in, For different height layers, Longitude The latitude dimension is considered. From the perspective of hail formation mechanism, hail is formed by liquid water in strong convective cells being lifted by updrafts to a height layer above 0°C and condensing and colliding. Therefore, it is not necessary to use all height elements for model training. In addition, the multi-source data used in this scheme are used to train the U-Net convolutional neural network model at the same time. In order to ensure the uniformity of data dimensions, it is also necessary to perform dimensionality reduction extraction on the profile data.
[0020] For temperature field elements, the traversal pattern uses the original data. Collect the height information corresponding to 0℃ / -20℃ to form the height distribution of the 0℃ / -20℃ isotherm; for dynamic field elements, first traverse the original model data. Longitude and latitude and dimensional Then traverse different height dimensions Take the maximum value of the updraft to form the maximum updraft velocity.
[0021] Specifically, after the above data processing, the input data of the U-Net convolutional neural network model includes radar modes such as combined reflectivity, differential reflectivity, differential phase shift rate, correlation coefficient, and echo top height; mode modes such as graupel mass mixing ratio, vertical updraft velocity, 0℃ / -20℃ isotherm height, and vertical liquid water content; and satellite modes such as 6.25um / 6.95um water vapor channel brightness temperature and 8.55um / 12.00um infrared channel brightness temperature. These total 14 data channels, each with a dimension of [missing information]. ,composition Data tensors .
[0022] In one implementation, the construction method for the multi-source modality weighted adaptive hybrid loss function in step S3 is as follows:
[0023] The loss function is the core basis for training deep learning models. It is used to quantify the error between the prediction results of the U-Net convolutional neural network model and the true labels, and directly determines the model's convergence effect, recognition accuracy, and generalization ability. In the U-Net convolutional neural network model, the design of the loss function directly affects the extraction efficiency and synergistic utilization effect of different modal features. Traditional equal-weight loss weakens the contribution of high-value modalities, resulting in insufficient robustness of the model for hail recognition. Therefore, this scheme adopts a multi-source modality weighted adaptive hybrid loss function.
[0024] S31. Calculate the basic sub-loss;
[0025] S32. Calculate the sub-losses of radar data modes, pattern data modes, and satellite data modes;
[0026] S33. Calculate the modal adaptive weights;
[0027] S34. Calculate the weighted fusion total loss function to obtain the multi-source modal weighted adaptive hybrid loss function.
[0028] In one implementation, the basic sub-loss is calculated in step S31 as follows:
[0029] S311. Calculate the cross-entropy loss. The formula for calculating the cross-entropy loss is as follows:
[0030] ;
[0031] In the formula, These represent the height and width of the label image and the prediction image, respectively. This represents the total number of pixels. For the first The real label of each pixel Represents hail pixels, Indicates non-hail pixels; For the first The model predicts the probability for each pixel. ;
[0032] S312, Calculation loss, The loss calculation formula is as follows:
[0033] ;
[0034] In the formula, It is a very small constant used to prevent the denominator from being 0.
[0035] In one implementation, the calculation formulas for calculating the radar data mode, mode data mode, and satellite data mode sub-loss in step S32 are as follows:
[0036] ;
[0037] ;
[0038] ;
[0039] In the formula, , , These correspond to the calculation of three data modes: radar data mode, pattern data mode, and satellite data mode, respectively. , , These are the model prediction outputs for radar data modes, pattern data modes, and satellite data modes, respectively. The actual label value; for Norm, for Norm; The smoothing weighting coefficient is typically between 0.5 and 0.8. This is a height correction factor, preset based on the number of height layers in the model data, generally between 0.3 and 1.0; Mean Absolute Deviation Loss Function Measure the deviation between predicted and actual values in a model mode; It is a very small constant used to prevent the denominator from being 0.
[0040] In one implementation, in step S33, the modal adaptive weights are calculated using the following formula:
[0041] ;
[0042] In the formula, For the first Adaptive weights for each modality Corresponding to radar data modes, pattern data modes, and satellite data modes; for Feature confidence of each modality; It is the sum of the confidence scores of the three modalities; It is a very small constant used to prevent the denominator from being 0.
[0043] In one implementation, in step S34, the weighted fusion total loss function is calculated, and the calculation formula is as follows:
[0044] ;
[0045] In the formula, The total loss is the multi-source modal weighted adaptive hybrid loss function. For cross-entropy loss and The loss balance coefficient, These are the weighting coefficients for the modal loss term.
[0046] In one implementation, the U-Net convolutional neural network model is constructed and trained in step S4 as follows:
[0047] Based on the U-Net convolutional neural network framework, the multi-source modality weighted adaptive hybrid loss function constructed in step S3 is embedded into the model loss calculation module as the core evaluation criterion for model training. The U-Net convolutional neural network model adopts an encoder-decoder symmetric structure. The encoder extracts key features of hail targets from multi-source modality data through convolutional layers and pooling layers. The decoder upsamples through transposed convolution and combines the different features of the three data modalities with a jump connection structure to achieve pixel-level localization and recognition of hail areas.
[0048] During the model training phase, the constructed data tensors The corresponding real-labeled hail samples are synchronously input into the U-Net convolutional neural network model for iterative training. During training, the multi-source modality weighted adaptive hybrid loss function calculates the pixel-level deviation between the model's predicted values and the real labels in real time. Based on the real-time contribution of each modality feature and the degree of noise interference, the training weights of each modality are dynamically adjusted to avoid the influence of single-modality defects on the training results, thereby achieving stable convergence of the U-Net convolutional neural network model and obtaining the trained U-Net convolutional neural network model.
[0049] In one implementation, the specific method for hail identification based on the U-Net convolutional neural network model in step S5 is as follows:
[0050] The U-Net convolutional neural network model performs pixel-by-pixel regression prediction on the input data and outputs a hail probability field with the same spatial size as the input. Each data point corresponds to a hail probability value between 0 and 1, and the magnitude of the probability value represents the likelihood of hail occurring at that data point.
[0051] Set a threshold for the hail probability field output by the U-Net convolutional neural network model. , will be greater than The data points are processed to form the outer boundary of the hailstone area.
[0052] The beneficial effects of this invention are as follows:
[0053] This invention proposes a hail recognition method based on U-Net convolutional neural network training and fusion of multimodal data. Addressing the challenges of balancing multimodal loss functions and the difficulty of manual parameter tuning in multi-source data fusion deep learning models, this invention proposes a multimodal weighted adaptive hybrid loss function. By dynamically weighting and adaptively optimizing different modal features, it effectively solves the problems of unstable model training and poor convergence under multi-source input, providing a practical technical route for achieving high-precision and high-reliability multi-source data fusion hail recognition. Attached Figure Description
[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.
[0055] Figure 1 This is a flowchart of a hail identification method based on U-Net and multimodal data fusion.
[0056] Figure 2 This is a radar composite reflectivity map.
[0057] Figure 3 This is a diagram showing the height distribution of the 0℃ isotherm.
[0058] Figure 4 This is a map showing the height distribution of the -20℃ isotherm.
[0059] Figure 5 It is a diagram of the vertical updraft velocity.
[0060] Figure 6 This is a comparison chart of training loss and validation loss.
[0061] Figure 7 Hailstone recognition effect Figure 1 .
[0062] Figure 8 Hailstone recognition effect Figure 2 . Detailed Implementation
[0063] To make the technical problems, technical solutions, and technical effects of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0064] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0065] Example 1
[0066] like Figures 1 to 8 As shown, this embodiment provides a hail identification method based on U-Net and multimodal data fusion, including the following steps:
[0067] S1. Input the raw satellite data, raw radar data, and raw model data into the data dimensionality reduction and resampling module;
[0068] S2. The data dimensionality reduction and resampling module performs dimensionality reduction and resampling on the three-dimensional data in the input data to obtain a two-dimensional planar structure (the original satellite data is a two-dimensional planar structure that can be used directly, so no dimensionality reduction processing is required; the original radar data and the original model data contain three-dimensional data, so both the original radar data and the original model data need to be dimensionality reduced and resampling to obtain a two-dimensional planar structure in order to achieve uniformity of all data dimensions). The principle of dimensionality reduction and resampling is data time alignment and spatial resampling, and extracts the satellite modal data, radar modal data and model modal data that have been dimensionality reduced and resampling.
[0069] The method for dimensionality reduction extraction of raw radar data is as follows:
[0070] The radar data is volume scan data, divided into multiple elevation angles, each containing reflectivity data and dual polarization parameter data. From the perspective of hail observation, hail clouds often appear in the form of strong convective cells. Therefore, the optimal solution input to the model is the radar data with the strongest reflectivity in the strong convective cells.
[0071] S21. The radar volume scan range is a circular scanning area centered on the radar station. A latitude and longitude network is constructed with a resolution of 0.02 degrees to ensure that the latitude and longitude grid completely covers the radar scanning area.
[0072] S22. For each latitude and longitude grid point, iterate through the reflectivity data of all elevation angle layers, all radial and distance databases to find the strongest reflectivity at all elevation angles, and record the strongest reflectivity. and their corresponding positions ,in, For elevation angle index, Radial index, For distance database indexing, via turn up The strongest reflectivity at the corresponding location ,in Indicates differential reflectivity, Indicates differential phase shift rate, Indicates the position corresponding to the correlation coefficient;
[0073] S23. After all grid points within the latitude and longitude grid have been traversed, data such as the combined reflectivity, differential emissivity, differential phase shift rate, correlation coefficient, and echo top height of the two-dimensional planar structure are obtained.
[0074] The method for dimensionality reduction extraction of the original pattern data is as follows:
[0075] The temperature field and dynamic field elements in the original model data are all three-dimensional data. ,in, Longitude Latitude dimension The data is divided into different height layers. From the perspective of hail formation mechanism, hail is formed by liquid water in strong convective cells being lifted by updrafts to a height layer above 0°C and condensing and colliding. Therefore, it is not necessary to use all height elements for model training. In addition, the multi-source data used in this scheme are used to train the U-Net convolutional neural network model at the same time. In order to ensure the uniformity of data dimensions, it is also necessary to perform dimensionality reduction extraction on the profile data.
[0076] For temperature field elements, the traversal pattern uses the original data. We extract the altitude information corresponding to 0℃ / -20℃ to form the altitude distribution of the 0℃ / -20℃ isotherm; for dynamic field elements, we first traverse the original model data. Longitude and latitude and dimensional Then iterate through different height layers and dimensions. Take the maximum value of the updraft to form the maximum updraft velocity.
[0077] Specifically, after the above data processing, the input data of the U-Net convolutional neural network model includes radar modes such as combined reflectivity, differential reflectivity, differential phase shift rate, correlation coefficient, and echo top height; mode modes such as graupel mass mixing ratio, vertical updraft velocity, 0℃ / -20℃ isotherm height, and vertical liquid water content; and satellite modes such as 6.25um / 6.95um water vapor channel brightness temperature and 8.55um / 12.00um infrared channel brightness temperature. These total 14 data channels, each with a dimension of [missing information]. ,composition Data tensors This is the data carrier prepared for the U-Net model.
[0078] S3. Based on satellite modal data, radar modal data, and mode modal data, construct a multi-source modal weighted adaptive hybrid loss function:
[0079] The loss function is the core basis for training deep learning models. It is used to quantify the error between the prediction results of the U-Net convolutional neural network model and the true labels, and directly determines the model's convergence effect, recognition accuracy, and generalization ability. In the U-Net convolutional neural network model, the design of the loss function directly affects the extraction efficiency and synergistic utilization effect of different modal features. Traditional equal-weight loss weakens the contribution of high-value modalities, resulting in insufficient robustness of the model for hail recognition. Therefore, this scheme adopts a multi-source modality weighted adaptive hybrid loss function.
[0080] S31. Calculate the basic sub-loss:
[0081] S311. Calculate the cross-entropy loss. The formula for calculating the cross-entropy loss is as follows:
[0082] ;
[0083] In the formula, These represent the height and width of the label image and the prediction image, respectively. This represents the total number of pixels. For the first The real label of each pixel Represents hail pixels, Indicates non-hail pixels; For the first The model predicts the probability for each pixel. ;
[0084] S312, Calculation loss, The loss calculation formula is as follows:
[0085] ;
[0086] In the formula, It is a very small constant used to prevent the denominator from being 0.
[0087] S32. The formulas for calculating the sub-losses of radar data modes, pattern data modes, and satellite data modes are as follows:
[0088] ;
[0089] ;
[0090] ;
[0091] In the formula, , , These correspond to the calculation of three data modes: radar data mode, pattern data mode, and satellite data mode, respectively. , , These are the model prediction outputs for radar data modes, pattern data modes, and satellite data modes, respectively. The actual label value; for Norm, for Norm; The smoothing weighting coefficient is typically between 0.5 and 0.8. This is a height correction factor, preset based on the number of height layers in the model data, generally between 0.3 and 1.0; Mean Absolute Deviation Loss Function Measure the deviation between predicted and actual values in a model mode; It is a very small constant used to prevent the denominator from being 0.
[0092] S33. Calculate the modal adaptive weights using the following formula:
[0093] ;
[0094] In the formula, For the first Adaptive weights for each modality Corresponding to radar data modes, pattern data modes, and satellite data modes; for Feature confidence of each modality; It is the sum of the confidence levels of the three modalities; It is a very small constant used to prevent the denominator from being 0.
[0095] S34. Calculate the weighted fusion total loss function. The calculation formula is as follows:
[0096] ;
[0097] In the formula, The total loss is the multi-source modal weighted adaptive hybrid loss function. For cross-entropy loss and The loss balance coefficient, These are the weighting coefficients for the modal loss term.
[0098] S4. Based on the multi-source modality weighted adaptive hybrid loss function in step S3, construct and train the U-Net convolutional neural network model:
[0099] Based on the U-Net convolutional neural network framework, the multi-source modality weighted adaptive hybrid loss function constructed in step S3 is embedded into the model loss calculation module as the core evaluation criterion for model training. The U-Net convolutional neural network model adopts an encoder-decoder symmetric structure. The encoder extracts key features of hail targets from multi-source modality data through convolutional layers and pooling layers. The decoder upsamples through transposed convolution and combines a skip connection structure to fuse different features from 14 data channels of three data modalities, thereby achieving pixel-level localization and recognition of hail areas.
[0100] During the model training phase, the constructed data tensors The corresponding real-world hail samples are synchronously input into the U-Net convolutional neural network model for iterative training. During training, a multi-modal weighted adaptive hybrid loss function is used to calculate the pixel-level deviation between the model's predicted values and the real-world annotations in real time. Based on the real-time contribution of each modality's features and the degree of noise interference, the training weights of each modality are dynamically adjusted to avoid the impact of single-modal defects on the training results. Represents a data tensor.
[0101] from Figure 6 As can be seen: in the early stage of training, both training loss and validation loss decrease rapidly, and the model quickly learns the basic hail features of multi-source modal data; in the middle stage of training, the loss fluctuates slightly, and the model continuously optimizes the modal weight allocation to adapt to the feature differences of multi-source data; in the later stage of training, both training loss and validation loss tend to stabilize, and their values decrease to below 0.08, with the fluctuation range narrowing to a small range, and the validation set loss does not show a continuous upward trend, indicating that the model has not overfitted and has achieved stable convergence.
[0102] S5. Identify hail based on the U-Net convolutional neural network model and output the identification results:
[0103] The U-Net convolutional neural network model performs pixel-by-pixel regression prediction on the input data and outputs a hail probability field with the same spatial size as the input. Each data point corresponds to a hail probability value between 0 and 1, and the magnitude of the probability value represents the likelihood of hail occurring at that data point.
[0104] Set a threshold for the hail probability field output by the U-Net convolutional neural network model. , will be greater than The data points are processed to form the outer boundary of the hailstone area.
Claims
1. A hail identification method based on U-Net and multimodal data fusion, characterized in that, Includes the following steps: S1. Input the raw satellite data, raw radar data, and raw model data into the data dimensionality reduction and resampling module; S2. The data dimensionality reduction and resampling module performs dimensionality reduction and resampling processing on the three-dimensional data in the input data to obtain a two-dimensional planar structure. The principle of dimensionality reduction and resampling processing is data time alignment and spatial resampling. It extracts satellite modal data, radar modal data and mode modal data after dimensionality reduction and resampling processing. S3. Based on satellite modal data, radar modal data, and mode modal data, a multi-source modal weighted adaptive hybrid loss function is constructed as follows: S31. Calculate the basic sub-loss. The method for calculating the basic sub-loss is as follows: S311. Calculate the cross-entropy loss using the following formula: ； In the formula, These are the height and width of the label image and the prediction image, respectively. This represents the total number of pixels. For the first The real label of each pixel Represents hail pixels, Indicates non-hail pixels; For the first The model predicts the probability for each pixel. ; S312, Calculation The loss is calculated using the following formula: ； In the formula, It is a very small constant used to prevent the denominator from being zero; S32. Calculate the sub-losses of radar data modes, mode data modes, and satellite data modes. The calculation formulas are as follows: ； ； ； In the formula, , , These correspond to the calculation of three data modes: radar data mode, pattern data mode, and satellite data mode, respectively. , , These are the model prediction outputs for radar data modes, pattern data modes, and satellite data modes, respectively. The actual label value; for Norm, for Norm; The smoothing weighting coefficient is typically between 0.5 and 0.
8. This is a height correction factor, preset based on the number of height layers in the model data, generally between 0.3 and 1.0; The mean absolute deviation loss function is used to measure the deviation between predicted and actual values in a model mode. It is a very small constant used to prevent the denominator from being zero; S33. Calculate the modal adaptive weights using the following formula: ； In the formula, For the first Adaptive weights for each modality Corresponding to radar data modes, pattern data modes, and satellite data modes; for Feature confidence of each modality; It is the sum of the confidence levels of the three modalities; It is a very small constant used to prevent the denominator from being zero; S34. Calculate the weighted fusion total loss function to obtain the multi-source modal weighted adaptive fusion loss function. The calculation formula is as follows: ； In the formula, The total loss is the multi-source modal weighted adaptive hybrid loss function. For cross-entropy loss and The loss balance coefficient, These are the weighting coefficients for the modal loss term; S4. Based on the multi-source modality weighted adaptive hybrid loss function in step S3, construct and train the U-Net convolutional neural network model; S5. Identify hail based on the U-Net convolutional neural network model and output the identification results.
2. The hail identification method based on U-Net and multimodal data fusion according to claim 1, characterized in that, In step S2, the method for dimensionality reduction extraction of the original radar data is as follows: S21. The radar volume scan range is a circular scanning area centered on the radar station. A latitude and longitude network is constructed with a resolution of 0.02 degrees to ensure that the latitude and longitude grid completely covers the radar scanning area. S22. For each latitude and longitude grid point, iterate through the reflectivity data of all elevation angle layers, all radial and distance databases to find the strongest reflectivity at all elevation angles, and record the strongest reflectivity. and their corresponding positions ,in, For elevation angle index, Radial index, For distance database indexing, via turn up The strongest reflectivity at the corresponding location ,in Indicates differential reflectivity, Indicates differential phase shift rate, Indicates the position corresponding to the correlation coefficient; S23. After all grid points within the latitude and longitude grid have been traversed, the combined reflectivity, differential emissivity, differential phase shift rate, correlation coefficient, and echo top height data of the two-dimensional planar structure are obtained.
3. The hail identification method based on U-Net and multimodal data fusion according to claim 2, characterized in that, In step S2, the method for dimensionality reduction extraction of the original pattern data is as follows: For temperature field elements, the traversal pattern uses the original data. ,in, Longitude Latitude dimension For different height layers, take the height information corresponding to 0℃ / -20℃ to form the height distribution of the 0℃ / -20℃ isotherm; for dynamic field elements, first traverse the original model data. Longitude in and latitude and dimensional Then iterate through different height layers and dimensions. Take the maximum value of the updraft to form the maximum updraft velocity.
4. The hail identification method based on U-Net and multimodal data fusion according to claim 1, characterized in that, In step S4, the U-Net convolutional neural network model is constructed and trained as follows: Based on the U-Net convolutional neural network framework, the multi-source modality weighted adaptive hybrid loss function constructed in step S3 is embedded into the model loss calculation module as the core evaluation criterion for model training. The U-Net convolutional neural network model adopts an encoder-decoder symmetric structure. The encoder extracts key features of hail targets from multi-source modality data through convolutional layers and pooling layers. The decoder upsamples through transposed convolution and combines the different features of the three data modalities with a jump connection structure to achieve pixel-level localization and recognition of hail areas. During the model training phase, the constructed data tensors The corresponding real-world hail samples are simultaneously input into the U-Net convolutional neural network model mentioned above for iterative training. During training, the multi-modal weighted adaptive hybrid loss function calculates the pixel-level deviation between the model's predicted values and the actual annotations in real time. Based on the real-time contribution of each modality's features and the degree of noise interference, the training weights of each modality are dynamically adjusted to avoid the impact of single modality defects on the training results, thereby achieving stable convergence of the U-Net convolutional neural network model and obtaining a well-trained U-Net convolutional neural network model.
5. The hail identification method based on U-Net and multimodal data fusion according to claim 4, characterized in that, In step S5, the specific method for hail identification based on the U-Net convolutional neural network model is as follows: The U-Net convolutional neural network model performs pixel-by-pixel regression prediction on the input data and outputs a hail probability field with the same spatial size as the input. Each data point corresponds to a hail probability value between 0 and 1, and the magnitude of the probability value represents the likelihood of hail occurring at that data point. Set a threshold for the hail probability field output by the U-Net convolutional neural network model. , will be greater than The data points are processed to form the outer boundary of the hailstone area.