Hail recognition and prediction method based on multi-source meteorological data fusion and attention mechanism

Abstract

The embodiment of the invention provides a hail identification and prediction method based on multi-source meteorological data fusion and an attention mechanism. The method is applied to the technical field of meteorology and artificial intelligence processing, and comprises the following steps: collecting dual-polarization radar data, satellite remote sensing data and ground meteorological observation data, and carrying out time sequence alignment, data standardization and feature splicing processing; extracting physical mechanism features and statistical texture features of hail clouds from the dual-polarization radar data, the satellite remote sensing data and the ground meteorological observation data; extracting multiple spatio-temporal features of the hail cloud by using a spatio-temporal feature extraction network; and inputting real-time observation data into the trained spatial-temporal feature extraction network, and outputting a hail occurrence probability, a nuclear region position and intensity grade distribution. In this way, the technical problems that in the prior art, multi-modal data noise, inconsistency and insufficient time-space feature capture are caused, and the precision and real-time performance of hail recognition are limited can be solved.

Description

Technical Field

[0001] The present disclosure relates to the fields of meteorology and artificial intelligence processing, and in particular to a hail recognition and prediction method based on multi-source meteorological data fusion and attention mechanism. Background Art

[0002] Currently, hail, a small-scale severe convective weather phenomenon, is difficult to monitor and predict due to its rapid occurrence and highly destructive nature. Traditional hail identification methods rely primarily on radar reflectivity and empirical formulas, but suffer from high false alarm rates and delayed recognition. In recent years, with the abundance of radar, satellite, and ground-based observation data and the application of deep learning techniques, intelligent hail identification has become a research hotspot. However, issues such as multimodal data noise, inconsistencies, and insufficient capture of spatiotemporal features continue to limit the accuracy and real-time performance of hail identification systems. Summary of the Invention

[0003] The present disclosure provides a hail recognition and prediction method based on multi-source meteorological data fusion and attention mechanism, which solves the problems of multimodal data noise, inconsistency and insufficient capture of spatiotemporal features in the existing technology, and the technical problems that limit the accuracy and real-time performance of hail recognition.

[0004] According to a first aspect of the present disclosure, a method for hail recognition and prediction based on multi-source meteorological data fusion and attention mechanism is provided, comprising:

[0005] Collect dual-polarization radar data, satellite remote sensing data, and ground-based meteorological observation data, and perform time series alignment, spatial resampling, and missing value processing;

[0006] The physical mechanism characteristics and statistical texture characteristics of hail clouds are extracted from dual-polarization radar data, satellite remote sensing data, and ground-based meteorological observation data. The physical mechanism characteristics include high echo ratio, effective core thickness, and liquid ratio, while the statistical texture characteristics include gradient change, statistical moment characteristics, and co-occurrence matrix characteristics.

[0007] The multi-temporal features of hail clouds are extracted using a spatiotemporal feature extraction network. The spatiotemporal feature extraction network structure includes a multimodal input module, a dynamic attention module, a 3D convolution module, and a time series modeling module.

[0008] The real-time observation data is input into the trained spatiotemporal feature extraction network to output the probability of hail occurrence, core area location and intensity level distribution.

[0009] Compared with the prior art, the advantages and positive effects achieved by the present disclosure are:

[0010] The present invention achieves comprehensive modeling of the three-dimensional structure, evolution dynamics, and probability distribution of hail clouds through multimodal fusion of dual-polarization radar, satellite data, and ground meteorological station observation data, combined with a spatiotemporal feature extraction network (ST-FEN). It uses an innovative spatiotemporal attention module and multi-scale fusion technology to address the problems of data noise, missing high-altitude signals, and capturing the rapid development of small-scale severe convection in hail recognition. It significantly improves the accuracy and timeliness of intelligent hail recognition, and provides real-time warning and regional risk assessment capabilities. By constructing a spatiotemporal feature extraction network (ST-FEN), the present invention introduces multimodal data fusion, a dynamic attention mechanism, and multi-scale feature extraction technology, significantly improving the ability to intelligently identify and predict hail.

[0011] It should be understood that the contents described in the Summary of the Invention section are not intended to limit the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0013] Figure 1 A flowchart of a hail recognition and prediction method based on multi-source meteorological data fusion and attention mechanism according to an embodiment of the present disclosure is shown;

[0014] Figure 2 A block diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0015] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0016] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0017] Figure 1 FIG. 1 shows a flow chart of a hail recognition and prediction method 100 based on multi-source meteorological data fusion and attention mechanism in an embodiment of the present disclosure, as shown in FIG. Figure 1 As shown, the method 100 includes:

[0018] S110: Collects dual-polarization radar data, satellite remote sensing data, and ground-based meteorological observation data, and performs time series alignment, data normalization, and feature splicing. Dual-polarization radar data includes reflectivity, differential reflectivity, and common partial correlation coefficients; satellite remote sensing data includes cloud top brightness temperature and infrared brightness temperature; and ground-based meteorological observation data includes temperature, humidity, and precipitation. Time series alignment aligns multi-source data using timestamps and spatially resamples them to a unified grid. Data normalization normalizes radar, satellite, and ground-based data. Feature splicing groups radar signals, satellite brightness temperature, and ground-based observation data according to their physical properties and constructs a feature tensor.

[0019] Optionally, in some embodiments, the process of performing time series alignment, data normalization, and feature splicing processing specifically includes the following steps:

[0020] By aligning timestamps, the time series of dual-polarization radar data, satellite remote sensing data, and ground-based meteorological observation data are unified; multi-source data are resampled onto a unified grid;

[0021] Normalize radar, satellite, and ground data to convert data of different dimensions to the same scale; use extrapolation to fill in missing data;

[0022] Radar signals, satellite brightness temperature and ground observation data are grouped according to physical properties to construct feature tensors; the grouped data are spliced according to physical properties to form a multidimensional feature tensor.

[0023] It should be noted that, in the embodiment, the timestamp alignment of the time series alignment is:

[0024]

[0025] Where, T align represents the aligned time series; t i Indicates the timestamp of the i-th data source; t ref The timestamp representing the reference time series; w irepresents the weight of the i-th data source, which is used to adjust the alignment accuracy; σ t represents the smoothing parameter of time alignment, which controls the locality of alignment; N represents the total number of data sources.

[0026] Spatial resampling formula:

[0027]

[0028] Where G resample Represents the resampled grid data; S j represents the value of the jth original data point; p j represents the spatial position of the jth original data point; p grid represents the spatial position of the target grid point; σ p Represents the spatial smoothing parameter, which controls the smoothness of resampling; W j represents the spatial weight matrix of the jth data point; M represents the total number of original data points.

[0029] The normalization processing formula for data standardization is:

[0030]

[0031] Where, X norm represents the normalized data; X raw represents the original data; μ global Represents the global mean, which is the average value of all data sources; σ global represents the global standard deviation, which calculates the degree of dispersion of all data sources; α″ represents the scaling factor, which is used to adjust the normalization range; β″ represents the offset factor, which is used to adjust the normalization benchmark; μ local Represents the local mean, which calculates the average value of local data; σ local Represents the local standard deviation, which calculates the degree of dispersion of local data.

[0032] The missing value filling (extrapolation) formula is:

[0033]

[0034] V l Where, X fill represents the missing value after filling; X neighbor，l Represents the data value of the lth adjacent time point: t neighbor，l Indicates the timestamp of the lth adjacent time point; t missing Indicates the timestamp corresponding to the missing value; τ represents the time decay parameter, which controls the smoothness of the extrapolation; V l represents the weight vector of the lth neighboring time point; L represents the total number of neighboring time points.

[0035] Physical property grouping formula for feature splicing:

[0036]

[0037] Where, F group represents the feature vector after grouping; f n represents the nth original eigenvector; w n represents the weight vector of the nth feature; f center represents the center vector of the feature group; γ represents the radius of the feature group, which controls the group range; U n Represents the group mapping matrix of the nth feature.

[0038] The feature tensor construction formula is expressed as:

[0039]

[0040] Where, T feature Represents the constructed multi-dimensional feature tensor; F group，k represents the k-th group feature matrix; M k represents the kth feature mapping matrix; Represents tensor product operation, used to splice multi-dimensional features; B bias represents the bias matrix, used to adjust the baseline of the feature tensor; K represents the total number of feature groups. This design ensures scientific and creative time series alignment, data standardization, and feature concatenation. Each formula enhances the flexibility and adaptability of the calculation process by introducing elements such as weights, decay parameters, and mapping matrices.

[0041] In the embodiments of this application, the hierarchical processing described above ensures the temporal, spatial, and physical consistency of multi-source meteorological data, providing a high-quality data foundation for subsequent hail identification and prediction. This process is not only technologically innovative but also demonstrates cross-disciplinary thinking, meeting the patent law's definition of inventive step.

[0042] S120: Extract the physical mechanism characteristics and statistical texture characteristics of hail clouds from dual-polarization radar data, satellite remote sensing data, and ground-based meteorological observation data. The physical mechanism characteristics include high echo ratio, effective core thickness, and liquid ratio, while the statistical texture characteristics include gradient changes, statistical moment characteristics, and co-occurrence matrix characteristics.

[0043] Optionally, in some embodiments, the process of extracting the physical mechanism characteristics and statistical texture characteristics of the hail cloud specifically includes the following steps:

[0044] During the initial feature screening stage, core physical quantities such as reflectivity, differential reflectivity and common partial correlation coefficient are extracted from dual-polarization radar data; thermodynamic characteristics such as cloud top brightness temperature and infrared brightness temperature are extracted from satellite remote sensing data; environmental parameters such as temperature, humidity and precipitation are extracted from ground meteorological observation data; and key variables related to hail formation are preliminarily screened out.

[0045] The high echo ratio of the hail cloud is calculated through the reflectivity and differential reflectivity in the dual-polarization radar data. The high echo ratio area corresponds to the strong convective core in the hail cloud and is the key area for hail formation. By analyzing the temporal and spatial distribution of the high echo ratio, the potential generation location of the hail cloud is identified.

[0046] The effective core thickness of the hail cloud is calculated from the common partial correlation coefficient and reflectivity in the dual-polarization radar data, combined with the atmospheric temperature profile data; the liquid ratio of the hail cloud is calculated through the differential reflectivity and common partial correlation coefficient in the dual-polarization radar data, combined with the temperature and humidity in the ground meteorological observation data. Areas with lower liquid ratios correspond to areas with a high probability of hail formation.

[0047] Calculate the spatial gradient changes of physical quantities in hail clouds (such as reflectivity, cloud top brightness temperature, etc.) to capture the rapid spatial change characteristics of hail clouds; calculate the statistical moment characteristics such as mean, variance, skewness and kurtosis of physical quantities in hail clouds to reveal the dynamic evolution law of hail clouds; construct the symbiosis matrix of physical quantities in hail clouds, calculate its statistical quantities such as contrast, correlation, energy and entropy, and reveal the inherent laws of hail formation.

[0048] According to the different stages of hail formation, the physical mechanism characteristics and statistical texture features are weightedly fused; based on the feature dimensionality reduction method of physical laws, the thermodynamic conditions and kinetic conditions of hail formation are utilized to screen out features closely related to hail formation, such as liquid ratio, gradient change and statistical moment characteristics.

[0049] Among them, high echo ratio is used to describe the intensity of the strong convective core in the hail cloud. The calculation formula is as follows:

[0050]

[0051] Where Z h Indicates the reflectivity of the dual-polarization radar (unit: dBZ); Z h，min Indicates the minimum value of reflectivity; Z h，max Indicates the maximum value of reflectivity; Z dr Indicates differential reflectivity (unit: dB); Z dr，min Indicates the minimum value of differential reflectivity; Z dr，max Indicates the maximum value of the differential reflectivity.

[0052] The effective core thickness is used to describe the distribution of ice crystals and liquid water in hail clouds. The calculation formula is as follows:

[0053]

[0054] Where h1 and h2 represent the vertical height range of the hail cloud (in km); ρ ice (h) represents the ice crystal density at height h (unit: kg / m 3 );Z h (h) represents the reflectivity at height h (unit: dBZ); Z h，ref Represents the reference reflectivity (unit: dBZ); ρ cc (h) represents the partial correlation coefficient at height h; ρ cc，ref represents the reference partial correlation coefficient.

[0055] The liquid ratio is used to describe the ratio of liquid water to solid ice in a hail cloud. The calculation formula is as follows:

[0056]

[0057] Where, ρ liq (h) represents the density of liquid water at height h (unit: kg / m 3 );Z dr (h) represents the differential reflectivity at height h (unit: dB); Z dr，ref represents the reference differential reflectivity (in dB); T(h) represents the temperature at height h (in K); T ref Indicates the reference air temperature (unit: K).

[0058] Spatial gradient changes are used to capture the rapid changes in physical quantities in hail clouds. The calculation formula is as follows:

[0059]

[0060] Where, Represents the gradient of reflectivity in the horizontal direction x; Represents the gradient of reflectivity in the horizontal direction y; Represents the gradient of reflectivity in the vertical direction z.

[0061] The statistical moment characteristics are used to describe the dynamic evolution of hail clouds. The calculation formula is as follows:

[0062]

[0063] Where N represents the number of samples; Z h (i) represents the reflectivity of the i-th sample (unit is dBZ); represents the mean of the reflectivity (unit is dBZ); k represents the order of the statistical moment (usually 2, 3, and 4 corresponding to variance, skewness, and kurtosis, respectively); ρ cc (i) represents the partial correlation coefficient of the i-th sample; ρ cc，ref represents the reference partial correlation coefficient.

[0064] The co-occurrence matrix characteristics are used to analyze the correlation between different physical quantities in hail clouds. The calculation formula is as follows:

[0065]

[0066] Where p(i, j) represents the probability of co-occurrence of physical quantities i and j; p ref represents the reference symbiosis probability; Z h (i) represents the reflectivity of the i-th sample (unit: dBZ); Z dr (j) represents the differential reflectivity of the jth sample (in dB); Z h，ref Indicates the reference reflectivity (unit: dBZ); Z dr，ref Indicates the reference differential reflectivity (in dB).

[0067] Feature weighted fusion is used to combine physical mechanism features and statistical texture features. The calculation formula is as follows:

[0068]

[0069] In the formula, α, β, γ, and δ represent weight coefficients; R hb，ref Indicates the reference high echo ratio; T eff，ref Indicates the effective thickness of the reference core; L r，ref Indicates the reference liquid ratio; G sp，ref The above formula can meet the requirements of extracting the physical mechanism characteristics and statistical texture characteristics of hail clouds.

[0070] In the embodiment of the present application, through the above-mentioned hierarchical technical process, not only the physical mechanism characteristics and statistical texture characteristics of hail clouds are extracted, but also through cross-domain technical integration, a new hail recognition and prediction method is formed.

[0071] S130: Extracting multi-spatiotemporal features of hail clouds using a spatiotemporal feature extraction network. The spatiotemporal feature extraction network structure includes a multimodal input module, a dynamic attention module, a 3D convolution module, and a time series modeling module.

[0072] Optionally, in some embodiments, the process of constructing the spatiotemporal feature extraction network specifically includes the following steps:

[0073] Through the cross-modal attention mechanism, the weights of data of different modalities are dynamically assigned; for example, in the early stages of hail cloud development, the cloud top brightness temperature characteristics of satellite remote sensing data may be more critical; while in the mature stage of hail cloud, the reflectivity characteristics of dual-polarization radar are more discriminative; through the attention mechanism, the spatiotemporal feature extraction network can adaptively focus on the most valuable features at the current stage.

[0074] The spatiotemporal attention mechanism adopts a long short-term memory variant structure in the time dimension to capture the dynamic characteristics of hail clouds evolving over time; in the spatial dimension, a local-global attention mechanism is designed, with local attention focusing on tiny changes in the hail core area, and global attention capturing the evolution trend of the overall structure of the hail cloud.

[0075] Multi-scale convolution kernels are used to extract the features of hail clouds at different spatial scales. For example, small-scale convolution kernels capture the high-frequency changes in the hail core area, while large-scale convolution kernels extract the overall morphological characteristics of the hail cloud. Combined with the dynamic attention mechanism, the optimal scale features can be adaptively selected.

[0076] A three-dimensional convolution kernel is designed to simultaneously capture the changes of hail clouds in time and space dimensions; through nonlinear activation functions and residual connections; a feature pyramid structure is constructed to fuse features at different levels; low-level features contain detailed information about hail clouds, while high-level features capture the overall evolution of hail clouds; in the time series modeling module, temporal feature enhancement is adopted, and a sliding window mechanism is used to capture the mutation characteristics of hail clouds in a short period of time, while long-term memory units are used to extract the long-term evolution of hail clouds.

[0077] Among them, in the process of building the spatiotemporal feature extraction network, the calculation formulas for designing the three-dimensional convolution kernel, building the feature pyramid structure, and the time series modeling module are as follows:

[0078] The three-dimensional convolution kernel is used to capture the changes of hail clouds in time and space dimensions at the same time. Assume that the input data is Among them, T represents the time dimension (number of time steps); H and W represent the height and width of the spatial dimension respectively; C represents the number of channels of the input data.

[0079] 3D convolution kernel The calculation formula is expressed as:

[0080]

[0081] Where Y t，h，w，c′ is the output feature map; k t , k h , k w are the convolution kernel sizes in time, height and width directions respectively; b c′is the bias term; t, h, w are the time, height, and width dimensions of the output feature map, respectively.

[0082] After the three-dimensional convolution, the feature expression ability is enhanced through nonlinear activation functions (such as ReLU) and residual connections. Let Y be the convolution output, and the calculation formula of the residual connection is expressed as:

[0083] Z t，h，w，c′ =σ(Y t，h，w，c′ )+X t，h，w，c′

[0084] Where σ(.) represents a nonlinear activation function (such as ReLU); Z is the output after the residual connection.

[0085] The feature pyramid structure extracts features at different levels through multi-scale convolution kernels and fuses these features. l is the feature map of the lth layer, and the calculation formula of the feature pyramid is expressed as:

[0086] F l =Upsample(Z l )+Conv 1×1 (Z l+1 )

[0087] Where Upsample(·) represents the upsampling operation; Conv 1×1 (·) represents a 1×1 convolution operation; F l is the fusion feature of the lth layer.

[0088] The time series modeling module captures the short-term mutation and long-term evolution of hail clouds through a sliding window mechanism and long-term memory units. t is the feature map of time step t, and the calculation formula of the sliding window mechanism is expressed as:

[0089]

[0090] Where w is the size of the sliding window; W i′ is the weight matrix of time step i′; S t is the output of the sliding window.

[0091] The calculation formula of long-term memory unit is expressed as:

[0092] L t =LSTM(F t , L t-1 )

[0093] Where LSTM(·) represents the computational process of the long short-term memory network; L t is the long-term memory output at time step t.

[0094] Temporal feature enhancement captures the dynamic characteristics of hail clouds by combining the output of sliding windows and long-term memory units. The calculation formula is expressed as:

[0095] E t =S t +L t

[0096] Where, E t It is the output after temporal feature enhancement.

[0097] Finally, the features at different levels are integrated to generate the spatiotemporal features of hail clouds. The calculation formula is:

[0098]

[0099] Where 0 is the final spatiotemporal feature output; L is the number of layers of the feature pyramid.

[0100] The spatiotemporal feature extraction network specifically includes:

[0101] The multimodal input module processes dual-polarization radar, satellite data, and ground observation data, extracts modality-specific features from each, and then fuses them into the core network.

[0102] Input data:

[0103] Radar data: three-dimensional data such as reflectivity, differential reflectivity, and common partial correlation coefficient; satellite data: two-dimensional data such as infrared brightness temperature (IR) and cloud top height; ground observation data: point data such as temperature, humidity, and precipitation, interpolated into a two-dimensional grid.

[0104] Module structure:

[0105] Each modality is individually subjected to preliminary feature extraction through a 2D or 3D convolutional layer with a convolution kernel size of 3×3 or 3×3×3.

[0106] The output feature map size is unified to a spatial resolution of 128×128 (for satellite and ground data) or 64×64×20 (for radar 3D data).

[0107] The channel fusion mechanism is used to concatenate the feature maps of all modalities into a multi-channel tensor.

[0108] Dynamic Attention Module:

[0109] Function: Give higher weights to key hail target areas (such as high echo cores and overhanging areas) and time nodes (such as the beginning of severe convection).

[0110] Core Technology

[0111] Spatial Attention: A learnable weight matrix is introduced to automatically identify salient features of high-echo cores (such as areas above 50dBz) and overhanging areas based on input features.

[0112] Temporal attention: Utilizes the temporal weights of GRU output to identify time series patterns that are critical to hail occurrence.

[0113] Implementation

[0114] Calculate the attention weights for the input feature map:

[0115] 3D convolution module:

[0116] Function: Extract the spatial characteristics of hail targets from 3D radar data, including core morphology, liquid ratio and vertical structure.

[0117] Module structure:

[0118] A three-layer cascaded 3D convolutional network is used, with each layer containing two convolutional blocks.

[0119] The convolution kernel size decreases layer by layer: 5×5×5, 3×3×3, and 1×1×1.

[0120] Each layer is followed by a ReLU activation function and Batch Normalization to prevent overfitting.

[0121] Output:

[0122] The three-dimensional position of the hail core area and its structural characteristics (such as overhang and core thickness).

[0123] Time Series Modeling Module

[0124] Purpose: To study the dynamic evolution of hail targets and predict the location and intensity of future hail.

[0125] Core components:

[0126] Use GRU (Gated Recurrent Unit) instead of traditional LSTM to reduce the number of parameters.

[0127] The input is a sequence of feature maps over multiple time steps.

[0128] Implementation:

[0129] The feature map of each time step is updated through the GRU unit;

[0130] The final output state represents the comprehensive spatiotemporal characteristics of the hail target.

[0131] Fusion and prediction module

[0132] Function: Integrate spatial and temporal features to output hail probability maps and other prediction results.

[0133] Fully connected layer: maps the spatiotemporal features of GRU to the prediction task (hail probability, location, intensity classification);

[0134] Output layer.

[0135] It should be noted that, in the embodiment, the training of the spatiotemporal feature extraction network specifically includes the following steps:

[0136] The prediction of hail occurrence probability, core area location and intensity level distribution prediction are taken as multi-task objectives, and a joint loss function is designed; through multi-task joint training, multiple objectives are optimized simultaneously.

[0137] An adaptive learning rate adjustment strategy is adopted to dynamically adjust the learning rate according to the change of loss during training; in the early stage of training, a larger learning rate is used for rapid convergence; in the later stage of training, the learning rate is gradually reduced to avoid overfitting.

[0138] Through feature visualization technology, we analyze whether the key features extracted by the network are consistent with the physical mechanism of hail clouds; verify whether the high echo ratio features are consistent with the actual location of the hail core area; introduce adversarial sample generation technology to generate interfering meteorological data samples, and through adversarial training, improve the network's prediction accuracy under complex meteorological conditions.

[0139] Among them, the formula of the joint loss function is expressed as:

[0140]

[0141] Where, represents the total loss function, which indicates the overall loss of multi-task joint training; α′, β′, and γ′ represent weight coefficients, which are used to balance the losses of hail occurrence probability prediction, core area location positioning, and intensity level distribution prediction, respectively; The loss function for predicting the probability of hail occurrence is expressed using the cross entropy loss function:

[0142]

[0143] Where N′ represents the number of samples; y i″ represents the true label (0 or 1) of the i″th sample; represents the predicted probability of the i″th sample.

[0144] The adaptive learning rate adjustment strategy is expressed as:

[0145] During training, the learning rate η is dynamically adjusted according to the loss change:

[0146]

[0147] Where ηt′ represents the learning rate of the t′th iteration; η0 represents the initial learning rate; η min represents the minimum learning rate; T represents the total number of iterations.

[0148] In the embodiment of the present application, through the above-mentioned hierarchical construction and training process, the spatiotemporal feature extraction network can efficiently integrate multi-source meteorological data, accurately identify and predict the occurrence and development of hail clouds, which is significantly different from the existing technology and has high creativity and practicality.

[0149] S140: Input the real-time observation data into the trained spatiotemporal feature extraction network to output the probability of hail occurrence, core area location and intensity level distribution.

[0150] Optionally, in some embodiments, the output is the probability of hail occurrence, core area location and intensity level distribution, and data visualization technology, such as the heat map generation algorithm in the geographic information system (GIS), is used to present the prediction results to the user in an intuitive form.

[0151] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present disclosure is not limited by the order of the actions described, because according to the present disclosure, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present disclosure.

[0152] The above is an introduction to the method embodiment. The following is a further explanation of the solution disclosed in the present disclosure through an apparatus embodiment.

[0153] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device and a readable storage medium.

[0154] Figure 2 A schematic block diagram of an electronic device 300 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0155] The electronic device 300 includes a computing unit 301, which can perform various appropriate actions and processes according to a computer program stored in a ROM 302 or a computer program loaded from a storage unit 308 into a RAM 303. The RAM 303 may also store various programs and data required for the operation of the electronic device 300. The computing unit 301, the ROM 302, and the RAM 303 are connected to each other via a bus 304. An I / O interface 305 is also connected to the bus 304.

[0156] Multiple components in the electronic device 300 are connected to the I / O interface 305, including an input unit 306, such as a keyboard, a mouse, etc.; an output unit 307, such as various types of displays, speakers, etc.; a storage unit 308, such as a magnetic disk, an optical disk, etc.; and a communication unit 309, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 309 allows the electronic device 300 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0157] The computing unit 301 can be any general-purpose and / or specialized processing component with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as the method for risk zoning for aircraft flights. For example, in some embodiments, the method for risk zoning for aircraft flights can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 308. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 300 via the ROM 302 and / or the communication unit 309. When the computer program is loaded into the RAM 303 and executed by the computing unit 301, one or more steps of the method for risk zoning for aircraft flights described above can be performed. Alternatively, in other embodiments, the computing unit 301 may be configured in any other appropriate manner (for example, by means of firmware) to execute the method for risk zoning for aircraft flight.

[0158] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0159] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0160] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0161] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0162] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0163] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0164] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0165] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A hail recognition and prediction method based on multi-source meteorological data fusion and attention mechanism, characterized by: include: Collect dual-polarization radar data, satellite remote sensing data, and ground-based meteorological observation data, and perform time series alignment, spatial resampling, and missing value processing; The physical mechanism characteristics and statistical texture characteristics of hail clouds are extracted from dual-polarization radar data, satellite remote sensing data, and ground-based meteorological observation data. The physical mechanism characteristics include high echo ratio, effective core thickness, and liquid ratio, while the statistical texture characteristics include gradient change, statistical moment characteristics, and co-occurrence matrix characteristics. The multi-temporal features of hail clouds are extracted using a spatiotemporal feature extraction network. The spatiotemporal feature extraction network structure includes a multimodal input module, a dynamic attention module, a 3D convolution module, and a time series modeling module. The real-time observation data is input into the trained spatiotemporal feature extraction network to output the probability of hail occurrence, core area location and intensity level distribution.

2. The method according to claim 1, characterized in that The dual-polarization radar data includes reflectivity, differential reflectivity and common partial correlation coefficient, the satellite remote sensing data includes cloud top brightness temperature and infrared brightness temperature, and the ground meteorological observation data includes temperature, humidity and precipitation.

3. The method according to claim 1, characterized in that The process of time series alignment, data standardization, and feature splicing includes the following steps: By aligning timestamps, the time series of dual-polarization radar data, satellite remote sensing data, and ground-based meteorological observation data are unified; multi-source data are resampled onto a unified grid; Normalize radar, satellite, and ground data to convert data of different dimensions to the same scale; fill in missing data; Radar signals, satellite brightness temperature and ground observation data are grouped according to physical properties to construct feature tensors; the grouped data are spliced according to physical properties to form a multidimensional feature tensor.

4. The method according to claim 3, characterized in that The time series are aligned with the timestamps: Where, T align represents the aligned time series; t i Indicates the timestamp of the i-th data source; t ref The timestamp representing the reference time series; w i represents the weight of the i-th data source, which is used to adjust the alignment accuracy; σ t represents the smoothing parameter of time alignment, controlling the locality of alignment; N represents the total number of data sources; Spatial resampling formula: Where G resample Represents the resampled grid data; S j represents the value of the jth original data point; p j represents the spatial position of the jth original data point; p grid represents the spatial position of the target grid point; σ p Represents the spatial smoothing parameter, which controls the smoothness of resampling; W j represents the spatial weight matrix of the jth data point; M represents the total number of original data points; The normalization processing formula for data standardization is: Where, X norm represents the normalized data; X raw represents the original data; μ global Represents the global mean, which is the average value of all data sources; σ global Represents the global standard deviation, which calculates the degree of dispersion of all data sources; α″ represents the scaling factor, which is used to adjust the normalization range; β″ is the offset factor used to adjust the normalized reference; μ local Represents the local mean, which calculates the average value of local data; σ local Represents the local standard deviation and calculates the degree of dispersion of local data; The missing value filling formula is: V l Where, X fill represents the missing value after filling; X neighbor，l Represents the data value of the lth adjacent time point; t neighbor，l Indicates the timestamp of the lth adjacent time point; t missing Indicates the timestamp corresponding to the missing value; τ represents the time decay parameter, which controls the smoothness of the extrapolation; V l represents the weight vector of the lth neighboring time point; L represents the total number of neighboring time points.

5. The method according to claim 3, characterized in that The physical property grouping formula of the feature splicing is: Where, F group represents the feature vector after grouping; f n represents the nth original eigenvector; w n represents the weight vector of the nth feature; f center represents the center vector of the feature group; γ represents the radius of the feature group, which controls the group range; U n Represents the group mapping matrix of the nth feature; The feature tensor construction formula is expressed as: Where, T feature Represents the constructed multi-dimensional feature tensor; F group，k represents the k-th group feature matrix; M k represents the kth feature mapping matrix; Represents the tensor product operation, which is used to splice multi-dimensional features; B bias Represents the bias matrix, which is used to adjust the basis of the feature tensor; K represents the total number of feature groups.

6. The method according to claim 1, characterized in that The process of extracting the physical mechanism characteristics and statistical texture characteristics of hail clouds includes the following steps: During the initial feature screening phase, core physical quantities such as reflectivity, differential reflectivity, and common partial correlation coefficient were extracted from dual-polarization radar data; cloud top brightness temperature and infrared brightness temperature thermodynamic characteristics were extracted from satellite remote sensing data; and environmental parameters such as temperature, humidity, and precipitation were extracted from ground-based meteorological observation data. Key variables related to hail formation were preliminarily screened. The high echo ratio of hail clouds is calculated using the reflectivity and differential reflectivity in dual-polarization radar data. High echo ratio areas correspond to strong convective cores within hail clouds and are key areas for hail formation. The potential locations for hail formation are identified by analyzing the spatiotemporal distribution of high echo ratios. The effective core thickness of hail clouds is calculated from the partial correlation coefficient and reflectivity in dual-polarization radar data, combined with atmospheric temperature profile data. The liquid fraction of hail clouds is calculated from the differential reflectivity and partial correlation coefficient in dual-polarization radar data, combined with temperature and humidity from ground meteorological observation data. Calculate the spatial gradient changes of physical quantities in hail clouds to capture the rapid spatial changes of hail clouds; calculate the statistical moment characteristics of the mean, variance, skewness and kurtosis of physical quantities in hail clouds to reveal the dynamic evolution of hail clouds; construct the symbiosis matrix of physical quantities in hail clouds and calculate its contrast, correlation, energy and entropy statistics to reveal the inherent laws of hail formation; According to the different stages of hail formation, the physical mechanism characteristics and statistical texture features are weightedly fused; based on the feature dimensionality reduction method of physical laws, the thermodynamic conditions and kinetic conditions of hail formation are utilized to screen out the features closely related to hail formation, such as liquid ratio, gradient change and statistical moment characteristics.

7. The method according to claim 6, characterized in that High echo ratio is used to describe the intensity of the strong convective core in hail clouds. The calculation formula is as follows: Where Z h Represents the reflectivity of dual-polarization radar; Z h，min Indicates the minimum value of reflectivity; Z h，max Indicates the maximum value of reflectivity; Z dr represents the differential reflectivity; Z dr，min Indicates the minimum value of differential reflectivity; Z dr，max Indicates the maximum value of differential reflectivity; The effective core thickness is used to describe the distribution of ice crystals and liquid water in hail clouds. The calculation formula is as follows: Where h1 and h2 represent the vertical height range of the hail cloud; ρ ice (h) represents the ice crystal density at height h; Z h (h) represents the reflectivity at height h; Z h，ref represents the reference reflectivity; ρ cc (h) represents the partial correlation coefficient at height h; ρ cc，ref represents the reference common partial correlation coefficient; The liquid ratio is used to describe the ratio of liquid water to solid ice in a hail cloud. The calculation formula is as follows: Where, ρ liq (h) represents the density of liquid water at height h; Z dr (h) represents the differential reflectivity at height h; Z dr，ref represents the reference differential reflectivity; T(h) is the temperature at height h (unit is K); T ref represents the reference temperature; Spatial gradient changes are used to capture the rapid changes in physical quantities in hail clouds. The calculation formula is as follows: Where, Represents the gradient of reflectivity in the horizontal direction x; Represents the gradient of reflectivity in the horizontal direction y; Represents the gradient of reflectivity in the vertical direction z; The statistical moment characteristics are used to describe the dynamic evolution of hail clouds. The calculation formula is as follows: Where N represents the number of samples; Z h (i) represents the reflectivity of the i-th sample; represents the mean of reflectivity; k represents the order of statistical moment; ρ cc (i) represents the partial correlation coefficient of the i-th sample; ρ cc，ref represents the reference common partial correlation coefficient; The co-occurrence matrix characteristics are used to analyze the correlation between different physical quantities in hail clouds. The calculation formula is as follows: Where p(i, j) represents the probability of co-occurrence of physical quantities i and j; p ref represents the reference symbiosis probability; Z h (i) represents the reflectivity of the i-th sample; Z dr (j) represents the differential reflectivity of the jth sample; Z h，ref Indicates the reference reflectivity; Z dr，ref Indicates the reference differential reflectivity.

8. The method according to claim 6, characterized in that Feature weighted fusion is used to combine physical mechanism features and statistical texture features. The calculation formula is as follows: In the formula, α, β, γ, and δ represent weight coefficients; R hb，ref Indicates the reference high echo ratio; T eff，ref Indicates the effective thickness of the reference core; L r，ref Indicates the reference liquid ratio; G sp，ref Represents the reference space gradient change.

9. The method according to claim 1, characterized in that The construction process of the spatiotemporal feature extraction network specifically includes the following steps: Through the cross-modal attention mechanism, the weights of different modal data are dynamically assigned; through the attention mechanism, the spatiotemporal feature extraction network adaptively focuses on the most valuable features at the current stage; The spatiotemporal attention mechanism uses a variant of the long short-term memory structure in the temporal dimension to capture the dynamic characteristics of hail clouds evolving over time. In the spatial dimension, a local-global attention mechanism is designed, with local attention focusing on subtle changes in the hail core area and global attention capturing the evolution trend of the overall structure of the hail cloud. The features of hail clouds at different spatial scales are extracted through multi-scale convolution kernels; A three-dimensional convolution kernel is designed to simultaneously capture the changes of hail clouds in time and space dimensions; through nonlinear activation functions and residual connections; a feature pyramid structure is constructed to fuse features at different levels; low-level features contain detailed information about hail clouds, while high-level features capture the overall evolution of hail clouds; in the time series modeling module, temporal feature enhancement is adopted, and a sliding window mechanism is used to capture the mutation characteristics of hail clouds in a short period of time, while long-term memory units are used to extract the long-term evolution of hail clouds.

10. The method according to claim 9, characterized in that The spatiotemporal feature extraction network specifically includes: The multimodal input module processes dual-polarization radar, satellite data, and ground observation data, extracting modality-specific features from each modality and fusing them into the core network. Each modality is individually processed through 2D or 3D convolutional layers for preliminary feature extraction. The feature maps of all modalities are then concatenated into a multi-channel tensor using a channel fusion mechanism. The dynamic attention module assigns higher weights to key hail target areas and time nodes; it uses the time weights output by the GRU to identify time series patterns that are critical to hail occurrence; The 3D convolution module extracts the spatial characteristics of hail targets from 3D radar data, including core morphology, liquid ratio, and vertical structure; and outputs the 3D position of the hail core and its structural characteristics. The time series modeling module learns the dynamic evolution of hail targets and predicts the future location and intensity of hail. The feature map of each time step is updated through the GRU unit; the output state represents the comprehensive spatiotemporal characteristics of the hail target. The fusion and prediction module integrates spatial and temporal features and outputs a hail occurrence probability map; The fully connected layer maps the spatiotemporal features of GRU to the prediction task.

Images