A ground hailstone identification method and system based on a water set classification result

Abstract

This invention relates to the field of meteorological observation technology, and provides a method and system for ground hail identification based on water condensate classification results. The method includes: spatiotemporally correlating multi-source hail data using a spatiotemporal matching algorithm to obtain a spatiotemporally matched hail event dataset; adaptively identifying the phase state of dual-polarization radar data using a dynamic membership function optimization algorithm to obtain multi-elevation angle hail phase state feature parameters containing rain-ice mixture categories; integrating and preprocessing multi-source meteorological data based on the multi-elevation angle hail phase state feature parameters to obtain standardized multi-dimensional meteorological feature data with fused phase state features; performing unsupervised pre-training and supervised fine-tuning training on a DCNN-DBN hybrid neural network using the standardized multi-dimensional meteorological feature data to obtain a ground hail identification model; and outputting hail landing area identification results through the ground hail identification model. This invention improves the ability to distinguish easily confused phase states and reduces the false alarm rate and false negative rate of hail identification.

Description

Technical Field

[0001] This invention relates to the field of meteorological observation technology, and in particular to a method and system for identifying ground hail based on water condensate classification results. Background Technology

[0002] Hail is a significant hazardous weather phenomenon accompanying severe convective systems. Currently, hail identification technology primarily relies on dual-polarization Doppler radar for detection, identifying different water condensate phases by analyzing the polarization characteristics of radar echoes. Traditional hail identification methods mainly employ dual-polarization radar water condensate classification algorithms. This algorithm, based on parameters such as reflectivity factor, differential reflectivity, and differential phase shift, uses fixed membership function thresholds to classify precipitation particles into 10 categories: graupel, ice crystals, wet snow, dry snow, ground objects, clear-sky echoes, large raindrops, hail, heavy rain, and light rain. Simultaneously, traditional machine learning methods, such as backpropagation (BP) neural networks, are also widely applied in hail identification, using multilayer perceptron networks to classify and train radar feature parameters.

[0003] However, existing technologies have significant shortcomings: First, traditional algorithms use fixed membership function thresholds, which cannot be adaptively adjusted according to different meteorological conditions and geographical environments, resulting in low recognition accuracy for easily confused phases such as ice crystals and dry snow; second, existing hydrophobic classification algorithms only consider pure phases, ignoring the rain-ice mixture state that is commonly present near the melting layer, causing missed and false alarms in hail identification; third, traditional BP neural networks are prone to overfitting under small sample data, lack effective feature extraction capabilities, and are unable to fully mine the high-order spatial feature information contained in radar data, resulting in insufficient accuracy in hail identification. Summary of the Invention

[0004] This invention provides a ground hail identification method and system based on hydrogel classification results to overcome the shortcomings of existing technologies.

[0005] This invention provides a method for identifying ground hail based on hydrogel classification results, comprising:

[0006] S1: Using a spatiotemporal matching algorithm, multi-source hail data are spatiotemporally correlated to obtain a spatiotemporally matched hail event dataset;

[0007] S2: Adaptive phase identification is performed on the dual-polarization radar data in the hail event dataset using a dynamic membership function optimization algorithm to obtain multi-elevation angle hail phase characteristic parameters that include rain-ice mixture categories;

[0008] S3: Based on the multi-elevation angle hail phase characteristic parameters, multi-source meteorological data are integrated and preprocessed through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data with fused phase characteristics;

[0009] S4: Using the standardized multidimensional meteorological feature data, the DCNN-DBN hybrid neural network is subjected to unsupervised pre-training and supervised fine-tuning training to obtain a ground hail recognition model;

[0010] S5: Output the hail landing area identification result through the ground hail identification model.

[0011] According to the ground hail identification method based on hydrogel classification results provided by the present invention, step S1 further includes:

[0012] S11: Spatial filtering of the radar observation area is performed based on the preset weather station locations to obtain the effective detection area;

[0013] S12: By using a dynamic time window strategy, hail events within the effective detection area are time-matched to obtain a time matching window based on the volume scan time.

[0014] S13: Perform spatial correlation processing on the actual hail locations within the time matching window to obtain radar echo matching results within a preset distance search radius;

[0015] S14: The radar echo matching results are fused with publicly available data to obtain a spatiotemporally matched hail event dataset.

[0016] According to the ground hail identification method based on hydrogel classification results provided by the present invention, step S2 further includes:

[0017] S21: Based on the A-test statistic of the hail event dataset, the membership function threshold of the polarization parameter in the Park dual-polarization phase classification algorithm is adjusted in real time to obtain a dynamically optimized phase discrimination threshold; wherein the polarization parameter includes the horizontal reflectivity factor and differential reflectivity;

[0018] S22: Adaptively correct the polarization parameter error using the weight matrix to obtain an optimized dual polarization parameter weight configuration;

[0019] S23: Based on the cross-correlation coefficient, identify the rain-ice mixture in the region above the melting layer under the dual polarization parameter weight configuration to obtain an extended classification result containing the mixed phase state;

[0020] S24: Perform a multi-angle comprehensive evaluation on the extended classification results to obtain multi-angle hail phase characteristic parameters that include rain-ice mixture categories.

[0021] According to the ground hail identification method based on hydrocondensate classification results provided by the present invention, step S21 further includes:

[0022] S211: Calculate the A-test statistic of the hail event dataset to obtain the statistical test index of the radar data distribution;

[0023] S212: Based on statistical test indicators, the membership function thresholds of reflectivity factor and differential reflectivity in the Park algorithm are dynamically adjusted to obtain adaptive phase discrimination parameters;

[0024] S213: The easily confused phases under the phase discrimination parameters are reclassified by real-time parameter updates to obtain phase recognition results with enhanced discriminative power; the easily confused phases include ice crystals and dry snow;

[0025] S214: Apply the phase recognition results to the entire observation period for batch processing to obtain a dynamically optimized phase discrimination threshold.

[0026] According to the ground hail identification method based on hydrogel classification results provided by the present invention, step S23 further includes:

[0027] S231: By using cross-correlation coefficient threshold discrimination, the radar observation data under the optimized dual polarization parameter weight configuration is used to locate the mixed phase region and obtain the distribution data of potential rain-ice mixing zone;

[0028] S232: Based on the height information, determine the position of the melting layer in the rain-ice mixing zone corresponding to the rain-ice mixing zone distribution data, and obtain the range of the ice phase region above the 0°C layer height;

[0029] S233: Spatial superposition analysis of the ice phase region and the mixed phase region is performed to obtain the location information of the rain-ice mixture;

[0030] S234: By using the mixture category label, the phase classification of the location information of the rain-ice mixture is extended to obtain an extended classification result that includes the mixed phase.

[0031] According to the ground hail identification method based on hydrogel classification results provided by the present invention, step S3 further includes:

[0032] S31: Grid the ERA5 data and ground observation data to obtain uniform grid data;

[0033] S32: Extract hailfall areas from the unified grid data to obtain a multi-dimensional feature window containing phase characteristics;

[0034] S33: Based on positive and negative samples, generate labels for the multidimensional feature window to obtain a labeled training sample set with phase identifiers;

[0035] S34: Standardize the labeled training sample set to obtain standardized multidimensional meteorological feature data that incorporates phase features.

[0036] According to the ground hail identification method based on hydrocondensate classification results provided by the present invention, step S4 further includes:

[0037] S41: Perform multi-layer convolution on the standardized multi-dimensional meteorological feature data to pre-train the deep belief network structure and obtain high-order feature representation weights;

[0038] S42: The weights of the higher-order feature representations are optimized layer by layer through an inter-layer unsupervised learning algorithm to obtain a pre-trained DCNN-DBN hybrid network structure;

[0039] S43: The DCNN-DBN hybrid network structure is supervised and fine-tuned using the BP algorithm to obtain a ground hail recognition model.

[0040] According to the ground hail identification method based on hydrocondensate classification results provided by the present invention, step S41 further includes:

[0041] S411: The standardized multidimensional meteorological feature data is reorganized into tensors according to a preset window structure to obtain a data tensor suitable for convolution operations;

[0042] S412: Input the data tensor into a multi-layer convolutional kernel to extract features from the data tensor and obtain a high-order feature map;

[0043] S413: According to the unsupervised learning principle, the weights of each convolutional layer corresponding to the higher-order feature map are initialized layer by layer to obtain the initial weight configuration;

[0044] S414: Integrate the pre-trained convolutional layer with the initial weight configuration with the fully connected layer to obtain higher-order feature representation weights.

[0045] According to the ground hail identification method based on hydrogel classification results provided by the present invention, step S42 further includes:

[0046] S421: Based on the principle of restricted Boltzmann machine, the weights of the first hidden layer in the weights of the higher-order feature representation are trained in an unsupervised manner to obtain the optimized weight parameters of the first layer.

[0047] S422: The subsequent hidden layers with optimized weight parameters of the first layer are sequentially optimized using a greedy layer-by-layer training strategy to obtain multi-layer weight configurations for independent training of each layer;

[0048] S423: Based on the criterion of minimizing reconstruction error, the multi-layer weight configuration is globally adjusted to obtain the global weight distribution;

[0049] S424: Integrate the global weight distribution into a network to obtain a pre-trained DCNN-DBN hybrid network structure.

[0050] This invention also provides a ground hail identification system based on hydrogel classification results, comprising:

[0051] The association module is used to perform spatiotemporal association processing on multi-source hail data using a spatiotemporal matching algorithm to obtain a spatiotemporally matched hail event dataset.

[0052] Identification module: used to adaptively identify the phase state of dual-polarization radar data in the hail event dataset through a dynamic membership function optimization algorithm, and obtain multi-elevation angle hail phase state characteristic parameters including rain-ice mixture categories;

[0053] Preprocessing module: Based on the multi-elevation angle hail phase characteristic parameters, it integrates and preprocesses multi-source meteorological data through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data with fused phase characteristics;

[0054] Training module: used to perform unsupervised pre-training and supervised fine-tuning training of the DCNN-DBN hybrid neural network using the standardized multidimensional meteorological feature data to obtain a ground hail recognition model;

[0055] The identification module is configured to use the ground hail identification model trained by the training module to output hail landing area identification results.

[0056] This invention provides a ground hail identification method and system based on hydrogel classification results. By constructing an adaptive phase recognition technology based on a dynamic membership function optimization algorithm and combining it with an unsupervised pre-training mechanism of a DCNN-DBN hybrid neural network, the accuracy of ground hail identification is significantly improved. First, the dynamic membership function optimization algorithm of this invention adjusts the polarization parameter membership function threshold in the Park algorithm in real time using the A-test statistic. This enables the invention to adaptively optimize the identification parameters according to different meteorological conditions and geographical environments, effectively solving the problem of decreased identification accuracy of traditional fixed threshold methods under complex meteorological environments. In particular, it improves the ability to distinguish easily confused phases such as ice crystals and dry snow, significantly reducing the false alarm rate and false negative rate of hail identification. Second, the rain-ice mixture identification technology based on cross-correlation coefficients of this invention fills the technical gap of traditional water condensate classification algorithms that only consider pure phases. By identifying mixed phases near the melting layer, the traditional 10 phase categories are expanded to an 11-category classification system that includes rain-ice mixtures, making the hail identification coverage more comprehensive and improving the identification accuracy. This effectively reduces the identification blind spots caused by ignoring mixed phases. In addition, the introduction of the DCNN-DBN hybrid neural network architecture of this invention realizes the algorithm The features enhance recognition performance. The unsupervised pre-training mechanism of the deep belief network automatically extracts high-order spatial features such as the vertical distribution pattern of hail particles from radar data through multi-layer convolution operations, effectively solving the overfitting problem of traditional BP neural networks in small-sample scenarios. Simultaneously, the inter-layer unsupervised learning algorithm, through the restricted Boltzmann machine principle and greedy layer-by-layer training strategy, enables the network to learn richer feature representations under unlabeled data conditions, significantly reducing dependence on a large number of labeled samples and improving recognition accuracy with the same amount of training data. Furthermore, the multi-source data fusion mechanism organically integrates dual-polarization radar data, ERA5 reanalysis data, and ground observation data through unified gridding processing, constructing a high-dimensional feature space containing 43 meteorological elements. This allows the model to comprehensively judge the meteorological conditions for hail occurrence from multiple dimensions, enhancing the stability and reliability of the recognition results and improving recognition accuracy under complex meteorological conditions compared to a single data source.

[0057] Overall, this invention, through the organic combination of algorithm innovation and network architecture optimization, not only achieves a leap in the accuracy of hail identification technology, but also provides more reliable and efficient technical support for practical application scenarios such as insurance claims and agricultural disaster prevention, and has important scientific value and broad application prospects. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0059] Figure 1 A schematic flowchart of a ground hail identification method based on hydrogel classification results provided by the present invention;

[0060] Figure 2 This is a schematic diagram of a ground hail identification system based on hydrogel classification results provided by the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0062] The embodiments of the present invention are described below with reference to the figures.

[0063] like Figure 1 As shown, this invention provides a ground hail identification method based on hydrogel classification results, comprising:

[0064] S1: Using a spatiotemporal matching algorithm, multi-source hail data is spatiotemporally correlated to obtain a spatiotemporally matched hail event dataset.

[0065] Step S1 further includes:

[0066] S11: Based on the preset weather station locations, the radar observation area is spatially filtered to obtain the effective detection area.

[0067] Furthermore, step S11 of the present invention first performs spatial filtering. In a specific embodiment, the latitude and longitude coordinates of all meteorological stations in region A are first read, including the coordinates of 6 stations equipped with S-band dual-polarization radars. Subsequently, the Euclidean distance calculation formula is used to calculate the spherical distance between each meteorological station and the radar station. After calculating the distance, the present invention sets a distance threshold range of 50-200 kilometers. Areas below 50 kilometers are excluded because the radar's minimum elevation angle of 0.5° results in an excessively low detection altitude, and areas above 200 kilometers are excluded because severe signal attenuation leads to insufficient detection accuracy. Then, filtering is performed based on the distance threshold and the calculated spherical distance. The filtering process involves traversing all meteorological station coordinates and calculating the distance to each radar station one by one. Meteorological stations that meet the distance range conditions are marked as valid stations, and the geographical boundary polygon of the valid detection area is constructed. The final obtained valid detection area is a fan-shaped or annular spatial geometry with the radar station as the center and a 50-200 kilometer radius as the ring range. Each valid detection area includes its geographical boundary coordinates, coverage area, number of meteorological stations included, and other attribute information.

[0068] S12: Using a dynamic time window strategy, hail events within the effective detection area are time-matched to obtain a time matching window based on the volume scan time.

[0069] Furthermore, the dynamic time window strategy of this invention is designed based on the radar volume scan time interval. In this embodiment, the standard volume scan interval of the dual-polarization radar in region A is 6 minutes. Specifically, in step S12, the start time record of the hail event is first obtained, and then the volume scan time closest to the start time of the hail is found in the radar volume scan time series as the reference time point. After determining the reference time, the algorithm calculates 3 volume scan times forward as the starting boundary of the time window, that is, the reference time minus 18 minutes, and calculates 2 volume scan times backward as the ending boundary of the time window, that is, the reference time plus 12 minutes, forming a time matching window with a total length of 30 minutes.

[0070] When a hail event has a clearly recorded end time, this invention employs a dynamic adjustment mechanism, using the actual end time as the cutoff boundary of the time window, replacing the default setting of pushing back two body scan times. The time matching window is stored in timestamp format, containing time attribute information such as window start time, end time, reference body scan time, and total window length.

[0071] S13: Perform spatial correlation processing on the actual hail locations within the time matching window to obtain radar echo matching results within a preset distance search radius.

[0072] Furthermore, the spatial correlation processing employs a point-to-area neighborhood verification method. Specifically, the actual hail observation point is first used as the spatial center, and a circular search area is established around it for spatial matching of radar echo data. The search radius is set to 20 kilometers, determined based on the typical spatial scale of the hail system and the spatial resolution of radar detection. First, the radar base data for each moment within the time matching window is read, including dual-polarization parameters such as basic reflectivity, co-correlation coefficient, differential reflectivity, and differential propagation phase shift rate. Then, the radar polar coordinates are converted to the geographic coordinate system. After the conversion, a circular buffer zone is established with the hail location as the center and a radius of 20 kilometers. Spatial geometric calculations are used to determine whether radar pixels fall within the buffer zone. Radar pixels that meet the spatial conditions are extracted, forming a set of radar echo data spatially correlated with the hail location. Each pixel contains complete information such as its geographic coordinates, dual-polarization parameter values, and observation time.

[0073] S14: The radar echo matching results are fused with publicly available data to obtain a spatiotemporally matched hail event dataset.

[0074] Furthermore, the publicly available data includes meteorological station observation data, disaster survey data, and insurance claim data. Meteorological station observation data includes hourly records of conventional meteorological elements such as temperature, humidity, wind speed, and wind direction. Disaster survey data stores detailed hail disaster records down to the township level, including information such as the time of disaster, location of disaster, and disaster intensity. Insurance claim data records crop losses caused by hail in agricultural insurance, including detailed information such as claim time, claim location, and extent of loss. After acquiring the data, a spatiotemporal joint matching strategy is used for data fusion. First, temporal matching is performed, associating records from each data source closest to the radar observation time. Then, spatial matching is performed, fusing observation records within a reasonable spatial distance, using the actual hail location as a benchmark. For disaster survey and insurance claim data, the geometric center coordinates of the area with the lowest precision are extracted as representative locations. During the fusion process, when the combined reflectivity exceeds 45 dBZ and the duration exceeds three radar observation times, and the proportion of hail pixels in the HCL hydrogel classification product reaches its highest value within a 20 km range, that moment is marked as the hail occurrence time. The resulting spatiotemporal matching hail event dataset contains multi-dimensional information such as time information, spatial location, radar observation parameters, ground observation data, and disaster records for each hail event, forming a complete feature description of the hail event.

[0075] S2: Using a dynamic membership function optimization algorithm, adaptive phase identification is performed on the dual-polarization radar data in the hail event dataset to obtain multi-elevation angle hail phase characteristic parameters that include rain-ice mixture categories.

[0076] Step S2 further includes:

[0077] S21: Based on the A-test statistic of the hail event dataset, the membership function threshold of the polarization parameter in the Park dual-polarization phase classification algorithm is adjusted in real time to obtain a dynamically optimized phase discrimination threshold; wherein the polarization parameter includes the horizontal reflectivity factor and differential reflectivity.

[0078] Step S21 further includes:

[0079] S211: Calculate the A-test statistic for the hail event dataset to obtain the statistical test index for the radar data distribution.

[0080] Further, in step S211, the basic reflectivity factor Zh and differential reflectivity ZDR data of the dual-polarization radar are first extracted from the spatiotemporally matched hail event dataset. Specifically, this invention extracts the radar data corresponding to each hail event in layers according to elevation angle, including observation data at seven standard elevation angles: 0.5°, 1.5°, 2.4°, 3.3°, 4.3°, 6.0°, and 9.9°. For the Zh and ZDR data of each elevation angle layer, a numerical sequence is constructed, with the sequence length equal to the number of all valid pixels within that elevation angle layer. Subsequently, the Zh and ZDR data are sorted in ascending order, and then the cumulative probability value of each data point under the assumed normal distribution is calculated. The cumulative probability calculation uses the standard normal distribution function. Finally, the A test statistics for Zh and ZDR are calculated respectively, and the calculation formula is as follows:

[0081]

[0082] in, To calculate the obtained A-test statistic, For the sample size, For the sample index value of the observation, For the first The sorted observations Using the cumulative distribution function, two independent statistical test indicators are obtained, which quantify the degree of deviation between actual radar observation data and the theoretical normal distribution.

[0083] S212: Based on statistical test indicators, the membership function thresholds of reflectivity factor and differential reflectivity in the Park algorithm are dynamically adjusted to obtain adaptive phase discrimination parameters.

[0084] Furthermore, the membership function of the Park algorithm adopts a trapezoidal function form, with each phase category corresponding to a set of trapezoidal parameters. The dynamic adjustment process in step S212 is based on the magnitude of the A test statistic to correct the parameters. When the Az value exceeds the critical threshold, it indicates that the Zh data distribution significantly deviates from the normal distribution, and the parameter adjustment mechanism is activated. The adjustment mechanism is a linear correction through adjustment coefficients. In addition, this invention also adjusts the membership function parameters of other phase categories, including large raindrops, graupel, ice crystals, dry snow, etc. The adjustment range of each phase is calculated by weighting according to its similarity to the hail phase. After the adjustment is completed, an adaptive phase discrimination parameter matrix containing the correction parameters of all phase categories is finally generated. The matrix dimension is 10×8, corresponding to the boundary values ​​of the membership functions of 10 phase categories and 8 dual polarization parameters.

[0085] S213: By updating the parameters in real time, the easily confused phases under the phase discrimination parameters are reclassified to obtain phase recognition results with enhanced discriminative power.

[0086] Furthermore, the easily confused phases mainly refer to two types of hydrogels: ice crystals and dry snow, which overlap with hail in the dual polarization parameter space. Specifically, in the reclassification process, the membership degree of each radar pixel to each phase category under the corrected membership function is first calculated. That is, the membership degree of each pixel to the three phases of hail, ice crystals, and dry snow is calculated separately under parameters such as Zh, ZDR, ρHV, and KDP, forming a 3×4 membership degree matrix. Then, the weighted geometric mean method is used to calculate the comprehensive membership degree. For pixels with similar membership values, this invention introduces a secondary discrimination mechanism to calculate the relative membership difference between phases. When the membership difference between hail and ice crystal is less than 0.1, this invention also adds temperature profile information for auxiliary discrimination. Areas with temperatures below -20℃ tend to be classified as ice crystals, and areas with temperatures above -10℃ tend to be classified as hail. After reclassification, the final phase category identifier and corresponding confidence value of each pixel are output, forming a phase recognition result matrix with enhanced discriminative power.

[0087] S214: Apply the phase recognition results to the entire observation period for batch processing to obtain a dynamically optimized phase discrimination threshold.

[0088] Specifically, the batch processing aims to extend the phase identification results from a single moment to continuous data throughout the entire observation period. During data processing, radar data from all volume scan moments within the time-matching window are first read, typically containing 5-8 consecutive observation moments. Then, a sliding window technique is used, with the window size set to 3 consecutive moments and the sliding step size to 1 moment. For each sliding window, the temporal consistency index of the phase identification results within the window is calculated. The consistency calculation method is as follows: for each spatial pixel, the frequency of being identified as the same phase within 3 moments is counted; pixels with a frequency greater than or equal to 2 are marked as stable pixels, and pixels with a frequency of 1 are marked as unstable pixels. After calculating the consistency index, a phase transition probability matrix is ​​established, recording the transition frequencies between each phase. The matrix has a dimension of 10×10, corresponding to the mutual transformation relationships between 10 phase categories. Based on the transformation probability matrix, this invention subsequently calculates the stability weight coefficient for each phase. Phases with high stability (such as hail and heavy rain) have larger weight coefficients, while phases with low stability (such as ice crystals and graupel) have smaller weight coefficients. Then, the stability weight coefficients are fed back to the membership function parameters for adjustment. The adjustment range is increased for phases with low stability and decreased for phases with high stability. After 3-5 iterations, the results finally converge to the dynamically optimized phase discrimination threshold. The obtained threshold maintains adaptability to the current meteorological conditions while also taking into account the stability requirements of the time series.

[0089] S22: Adaptively correct the polarization parameter error using the weight matrix to obtain an optimized dual polarization parameter weight configuration.

[0090] Furthermore, in the error analysis process of step S22, the standard deviation and relative error of each dual polarization parameter are first calculated. Specifically, an error estimation matrix is ​​first established, where the elements of the matrix are the error estimates of the parameters at the observation points. The specific expression is as follows:

[0091]

[0092] in, For the Zh parameter error, For calibration error, The attenuation error is 0.1-0.5dB. For ground clutter error;

[0093]

[0094] in, For ZDR parameter error, This is for differential reflectivity calibration error. For differential attenuation error, This represents the differential phase shift error.

[0095] The subsequent adaptive correction of the weight matrix adopts the error sensitivity test method, that is, to calculate the influence of each parameter error on the phase identification result, and establish a 4×10 weight matrix, with rows corresponding to the four dual polarization parameters (Zh, ZDR, ρHV, KDP) and columns corresponding to the ten phase categories. Then, the weight correction formula is performed through the sensitivity coefficient. The corrected weight configuration reduces the weight of high error parameters and enhances the weight of low error parameters, forming an optimized dual polarization parameter weight configuration matrix.

[0096] S23: Based on the cross-correlation coefficient, identify the rain-ice mixture in the region above the melting layer under the dual polarization parameter weight configuration to obtain an extended classification result containing the mixed phase state.

[0097] Step S23 further includes:

[0098] S231: By using cross-correlation coefficient threshold discrimination, radar observation data under optimized dual polarization parameter weight configuration is used to locate mixed phase regions and obtain potential rain-ice mixed zone distribution data.

[0099] Specifically, the cross-correlation coefficient ρHV is an important parameter for measuring the correlation between horizontal and vertical polarized echoes. Its value ranges from 0 to 1. The ρHV value of pure phase hydrogels is usually close to 1, while the ρHV value of mixed phase regions is significantly reduced. Therefore, in step S231, the present invention uses ρHV to locate mixed phase regions.

[0100] Furthermore, a ρHV threshold discrimination standard is first set, and the discrimination threshold is determined based on the ρHV characteristic values ​​of different phases. The typical ρHV value of pure hail is 0.85-0.95, the ρHV value of pure raindrops is 0.98-1.0, and the ρHV value of the rain-ice mixed region is between 0.7-0.9. This invention employs a dual-threshold discrimination method, setting an upper threshold of 0.9 and a lower threshold of 0.7. During the threshold discrimination process, all radar pixels under the optimized weight configuration are traversed, and the ρHV value of each pixel is extracted. When ρHV < 0.7, the pixel is marked as a strong mixing region; when 0.7 ≤ ρHV < 0.9, the pixel is marked as a potential mixing region; and when ρHV ≥ 0.9, the pixel is marked as a pure phase region. Then, an 8-connected-domain algorithm is used to identify continuous mixing phase regions. Spatially adjacent pixels that meet the mixing region conditions are grouped into the same connected domain. Each connected domain contains attribute information such as its boundary coordinates, area size, average ρHV value, and number of pixels. Finally, a mixing region attribute table is established to record the spatial location and intensity characteristics of each connected domain, forming potential rain-ice mixing zone distribution data.

[0101] S232: Based on the height information, determine the location of the melting layer in the rain-ice mixing zone corresponding to the rain-ice mixing zone distribution data, and obtain the range of the ice phase region above the 0°C layer height.

[0102] Furthermore, determining the location of the melt layer requires combining the temperature profile information from the ERA5 reanalysis data. Therefore, in step S232, this invention first reads the ERA5 hourly data closest to the radar observation time, extracts the temperature data of five isobaric surfaces (1000 hPa, 975 hPa, 950 hPa, 925 hPa, and 900 hPa), and then calculates the height of the 0℃ isotherm using a linear interpolation method. The interpolation process is performed between two adjacent isobaric surfaces. When the temperature of one isobaric surface is > 0℃ and the temperature of the next isobaric surface is < 0℃, the formula for calculating the height of the 0℃ layer is:

[0103]

[0104] in, The height of the 0℃ layer. The height of the first isobaric surface. The height of the second isobaric surface. The temperature of the first isobaric surface. The temperature of the second isobaric surface is then calculated. The observation height corresponding to the radar elevation angle is then calculated. Subsequently, the observation height of each radar pixel is compared with the height of the 0°C layer. Pixels above the 0°C layer are marked as ice phase regions, and pixels below the 0°C layer are marked as liquid phase regions. Finally, the range of the ice phase region is obtained and stored in a height-layered manner, including information such as the spatial boundary, height range, and number of pixels contained in each elevation angle layer.

[0105] S233: Spatial superposition analysis of the ice phase region and the mixed phase region is performed to obtain the location information of the rain-ice mixture.

[0106] Furthermore, the spatial overlay analysis in step S233 adopts the GIS spatial analysis method, treating the ice facies region and the mixed facies region as two independent spatial layers for overlay operation. The basic unit of the overlay operation is the radar pixel, and each pixel contains spatial attributes such as its geographic coordinates, height information, and facies attributes.

[0107] Specifically, firstly, all radar pixels are traversed to check whether each pixel simultaneously meets the conditions of both ice phase region and mixed phase region. Pixels that meet the conditions are marked as rain-ice mixture pixels. The discrimination conditions are: (1) pixel height > 0℃ layer height; (2) pixel ρHV value is in the range of 0.7-0.9; (3) pixel belongs to a connected mixed phase region. Then, an attribute table of rain-ice mixture pixels is established to record detailed information such as latitude and longitude coordinates, height value, ρHV value, and connected domain number of each pixel. Subsequently, the DBSCAN algorithm is used to aggregate spatially adjacent rain-ice mixture pixels into mixture regions. The clustering radius of the DBSCAN algorithm is set to a distance of 2 pixels, and the minimum number of cluster points is set to 5 pixels. Finally, the clustering results are obtained, forming multiple independent rain-ice mixture regions. Each region contains geometric and physical attribute information such as its spatial boundary, centroid coordinates, area size, and average intensity.

[0108] S234: By using the mixture category label, the phase classification of the location information of the rain-ice mixture is extended to obtain an extended classification result that includes the mixed phase.

[0109] Furthermore, the phase classification extension in step S234 aims to expand the traditional 10-category hydrogel classification system into an 11-category system that includes rain-ice mixtures. Specifically, during the extension process, the present invention first establishes a new category identifier for rain-ice mixtures, set as category number 11, and then reconstructs the phase classification coding table. The original 10 categories retain their original numbers: 1-graupel, 2-ice crystals, 3-wet snow, 4-dry snow, 5-ground features, 6-clear sky echo, 7-large raindrops, 8-hail, 9-heavy rain, 10-light rain, and adds 11-rain-ice mixture. After setting, the phase identification results of all radar pixels are traversed, and the category identifier of pixels that meet the rain-ice mixture condition is modified from the original category to 11. The modification process adopts a priority rule. When a pixel meets multiple phase conditions simultaneously, it is classified according to the following priority: rain-ice mixture > hail > graupel > large raindrops > other phases. In addition, this invention establishes a phase transformation mapping table to record the transformation relationship of each pixel from the original classification to the extended classification. The final extended classification results are stored in the form of a three-dimensional array with the array dimensions being [elevation angle × distance library number × azimuth angle]. The array element values ​​are phase category numbers from 1 to 11. The classification results for each elevation angle layer are stored separately, forming 7 two-dimensional classification matrices, corresponding to the observation data of 7 standard elevation angles. The extended classification results also include statistical information for each phase category, such as quantitative indicators like the number of pixels, proportion, and spatial distribution characteristics.

[0110] S24: Perform a multi-angle comprehensive evaluation on the extended classification results to obtain multi-angle hail phase characteristic parameters that include rain-ice mixture categories.

[0111] In step S24, the present invention employs a vertical profile analysis method for multi-elevation angle comprehensive evaluation, aiming to perform three-dimensional reconstruction and statistical analysis of the extended classification results of the seven elevation angle layers. During the reconstruction process, a three-dimensional coordinate system is first established, with the radar station as the origin, creating a cylindrical coordinate system of range-altitude-azimuth. Then, an interpolation algorithm is used to spatially interpolate discrete observation points in the range-altitude plane, generating a continuous phase distribution profile. The interpolation method employs an inverse distance weighting algorithm, specifically calculating the statistical feature parameters of hail and rain-ice mixtures within each elevation angle layer. These parameters include six basic parameters: number of pixels N, total area S, average reflectivity Zm, maximum reflectivity Zmax, average differential reflectivity ZDRm, and average cross-correlation coefficient ρHVm. A multi-elevation angle hail phase feature parameter vector is then established, with a dimension of 42, containing a combination of 7 elevation angle layers × 6 basic parameters. Furthermore, the feature parameter vector also includes 3 geometric parameters of the vertical structure and 2 statistical parameters of the mixed phase (proportion of rain-ice mixture and number of mixed regions). This results in a complete description of the multi-elevation angle hail phase feature parameters, encompassing the categories of rain-ice mixtures.

[0112] S3: Based on the multi-elevation angle hail phase characteristic parameters, the multi-source meteorological data are integrated and preprocessed through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data with fused phase characteristics.

[0113] Step S3 further includes:

[0114] S31: Grid the ERA5 data and ground observation data to obtain uniform grid data.

[0115] Further, in step S31, the ERA5 reanalysis data is first read. The original ERA5 data has a latitude and longitude grid with a resolution of 0.25°×0.25°, containing 25 atmospheric parameters including relative humidity, U-wind component, V-wind component, temperature, and vertical velocity at five isobaric surfaces: 1000hPa, 975hPa, 950hPa, 925hPa, and 900hPa. Then, a target grid is established with a resolution of 0.001°×0.001°, covering region A and its surrounding areas, with a longitude range of 113°E-120°E and a latitude range of 36°N-43°N, forming a 7000×7000 two-dimensional grid matrix. Subsequently, bilinear interpolation is used to interpolate the ERA5 data, while the nearest neighbor interpolation method is used for ground station data. The system calculates the spherical distance between each target grid point and all observation stations, selecting the value from the nearest station as the interpolation result for that grid point. After interpolation, the 25 parameters from ERA5 and the 18 parameters from ground observations are combined to form a final meteorological element consisting of 43 elements. The 25 parameters from ERA5 include 5 isobaric surfaces × 5 parameters. Each layer of parameters includes relative humidity, U-wind component (east-west wind speed), V-wind component (north-south wind speed), temperature, and vertical velocity. The ground observation data includes 18 parameters: temperature, maximum temperature, minimum temperature, ground temperature, relative humidity, minimum relative humidity, dew point temperature, 2-minute average wind speed, 2-minute average wind direction, maximum wind speed, maximum wind speed and direction, 10-minute average wind speed, 10-minute average wind direction, local air pressure, sea level air pressure, 3-hour pressure variation, visibility, and precipitation.

[0116] S32: Extract the hailfall area from the unified grid data to obtain a multi-dimensional feature window containing phase characteristics.

[0117] Furthermore, in step S32, the present invention first reads the multi-elevation angle hail phase feature parameters containing rain-ice mixture categories, extracts the latitude and longitude coordinates of hail and rain-ice mixture pixels, and defines the hail landing area as the geographical area where hail or rain-ice mixture phase exists. Specifically, in the extraction process, the present invention uses a buffer analysis method, establishing a circular buffer with a radius of 5km centered on each hail pixel. Multiple overlapping buffers are merged into continuous hail landing area polygons through Boolean joint operations. Then, using the centroid coordinates of each hail landing area as the window center, the corresponding data block is extracted from the unified grid data.

[0118] The final feature window data structure includes spatial coordinate information, 43-dimensional meteorological parameter values, phase feature identifiers and other attributes. The phase feature identifiers record the phase category distribution of each pixel in the window, including statistical information such as the number of hail pixels, the number of rain-ice mixture pixels, and the number of other phase pixels.

[0119] S33: Based on positive and negative samples, generate labels for the multidimensional feature window to obtain a labeled training sample set with phase identifiers.

[0120] Furthermore, the positive sample is defined as a feature window containing hail or a mixture of rain and ice phases, with a label value set to 1, and the negative sample is defined as a feature window not containing hail or a mixture of rain and ice phases, with a label value set to 0.

[0121] Positive sample extraction is directly based on the feature window of the hail landing area. The total number of hail pixels and rain-ice mixture pixels in each window is counted. When the total number is greater than the threshold of 10 pixels, the window is marked as a positive sample. Negative sample extraction adopts a spatial sampling strategy. At the time corresponding to the hail event, a non-hail area is randomly selected as the negative sample center point within a 0.1° latitude and longitude range in front, behind, left, and right of the hail record location.

[0122] The present invention further processes the samples. First, the validity of negative samples is verified by checking whether there are pixels of hail or rain-ice mixture within the window. When the number of pixels is less than a threshold of 2 pixels, the window is confirmed as a valid negative sample. Second, the samples are balanced by using undersampling technology to count the ratio of positive to negative samples. When the number of negative samples exceeds 3 times the number of positive samples, the excess negative samples are randomly deleted to maintain the ratio of positive to negative samples between 1:2 and 1:3.

[0123] The label generation process involves establishing a sample label mapping table, recording complete attributes such as sample number, center coordinates, label value, and included phase statistics for each feature window. The phase identifier extends the recording of the pixel distribution of 11 phase categories within the window, forming an 11-dimensional phase feature vector. The vector element value represents the pixel proportion of the corresponding phase within the window. The labeled training sample set with phase identifiers contains meteorological data of the feature window, 1-dimensional label value, and 11-dimensional phase feature vector, ultimately forming a complete sample description.

[0124] S34: Standardize the labeled training sample set to obtain standardized multidimensional meteorological feature data that incorporates phase features.

[0125] Furthermore, the standardization process in step S34 adopts the Z-score standardization method to eliminate the influence of differences in the dimensions and numerical ranges of different meteorological elements. The numerical range of the standardized phase feature vector is limited to 0-1. The final standardized multidimensional meteorological feature data incorporates phase features. The data fusion process involves concatenating the standardized 43-dimensional meteorological features with the 11-dimensional phase features to form a 54-dimensional fused feature vector. The final standardized multidimensional meteorological feature data with fused phase features is a four-dimensional tensor structure. Each sample contains spatially distributed multidimensional feature information and corresponding label values, constituting a complete machine learning training dataset.

[0126] S4: Using the standardized multidimensional meteorological feature data, the DCNN-DBN hybrid neural network is subjected to unsupervised pre-training and supervised fine-tuning training to obtain a ground hail recognition model.

[0127] Step S4 further includes:

[0128] S41: Perform multi-layer convolution on the standardized multi-dimensional meteorological feature data to pre-train the deep belief network structure and obtain high-order feature representation weights.

[0129] Step S41 further includes:

[0130] S411: The standardized multidimensional meteorological feature data is reorganized into tensors according to a preset window structure to obtain a data tensor suitable for convolution operations.

[0131] Furthermore, the tensor recombination process transforms the standardized multidimensional meteorological feature data with fused phase features from the original four-dimensional tensor structure into the input format required by the convolutional neural network in step S3. The system first reads the standardized data of all training samples. Each sample contains a 50×50 pixel spatial grid and 54 feature dimensions (43 meteorological parameters + 11 phase features). Subsequently, the 79 feature dimensions of each sample are grouped and arranged according to their physical meaning. The first 43 meteorological parameters are sorted according to vertical hierarchy, with the 5 parameters of the 1000hPa layer at the beginning, followed by the parameters of the 975hPa, 950hPa, 925hPa, and 900hPa layers. Each layer contains 5 elements: relative humidity, U-wind, V-wind, temperature, and vertical velocity. The 18 parameters of ground observation are sorted according to element type, with temperature-related parameters at the beginning, humidity parameters in the middle, and wind speed and direction parameters at the end. The last 11 phase features are arranged in the order of phase number 1-11, corresponding to the pixel proportions of graupel, ice crystals, wet snow, dry snow, ground features, clear sky echo, large raindrops, hail, heavy rain, light rain, and rain-ice mixture. The recombined data tensor conforms to the data format requirement of the deep learning framework, with the number of channels first.

[0132] S412: Input the data tensor into a multi-layer convolutional kernel to extract features from the data tensor and obtain a high-order feature map.

[0133] Furthermore, the multi-layer convolutional kernel design of this invention adopts a hierarchical feature extraction architecture, comprising four convolutional layers, each using convolutional kernels of different sizes and numbers for feature extraction. Specifically: the first convolutional layer uses 64 7×7 convolutional kernels with a stride of 2 and same padding, inputting the tensor obtained in step S411, and outputting a feature map after convolution; the second convolutional layer uses 128 5×5 convolutional kernels with a stride of 1; the third convolutional layer uses 256 3×3 convolutional kernels with a stride of 1; and the fourth convolutional layer uses 512 3×3 convolutional kernels with a stride of 2. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. During feature extraction, shallow convolutional kernels extract low-level features such as spatial edges and textures, while deep convolutional kernels extract high-level semantic features such as the vertical distribution pattern of hail particles and the spatial structure of the weather field. Finally, a high-order feature map is output, containing the feature representations output by each convolutional layer, forming a multi-scale, multi-semantic hierarchical feature pyramid structure.

[0134] S413: According to the unsupervised learning criterion, the weights of each convolutional layer corresponding to the higher-order feature map are initialized layer by layer to obtain the initial weight configuration.

[0135] Furthermore, in step S413, this invention is based on the unsupervised learning principle and uses the principle of minimizing reconstruction error of the autoencoder for weight initialization. The specific initialization process is performed layer by layer. First, the first convolutional layer is pre-trained unsupervised, using the input data tensor as the target output. The encoder-decoder structure learns the intrinsic representation of the data. The encoder uses 64 7×7 convolutional kernels to encode the input into feature representations, and the decoder uses transposed convolutions to reconstruct the feature representations back to the original input size. The second convolutional layer is initialized using the output feature map of the first layer as input, repeating the encoding-decoding process to learn higher-level feature representations. The encoder uses 128 5×5 convolutional kernels, and the decoder uses corresponding transposed convolutional kernels. The initialization process for the third and fourth layers is similar, learning 256-dimensional and 512-dimensional feature representations respectively. After training each layer, the weight parameters of that layer are saved, including the convolutional kernel weight matrix, bias vector, and batch normalization parameters. The initial weight configuration includes the complete parameter set of the four convolutional layers, with a total of approximately 3 million parameters.

[0136] S414: Integrate the pre-trained convolutional layer with the initial weight configuration with the fully connected layer to obtain higher-order feature representation weights.

[0137] Furthermore, in step S414, the present invention concatenates the four pre-trained convolutional layers with three fully connected layers to construct a complete deep neural network architecture. Specifically, firstly, a global average pooling layer is added after the fourth convolutional layer to convert the feature map output by the fourth convolutional layer into a one-dimensional feature vector. Then, a first fully connected layer is established to receive the 512-dimensional pooled feature vector and output a 256-dimensional feature representation. A second fully connected layer is established to map the 256-dimensional feature to 128 dimensions. A third fully connected layer is established to output a 2-dimensional vector, corresponding to the binary classification task of hail recognition. A Dropout layer is added after each fully connected layer to prevent overfitting, and the Dropout rate is set to 0.5.

[0138] The final high-order feature representation weights include spatial feature extraction weights from convolutional layers and semantic mapping weights from fully connected layers, totaling approximately 3.5 million trainable parameters. The integrated network structure is saved in computation graph format, containing a complete network description including forward propagation paths, backpropagation gradient flows, parameter update rules, and more.

[0139] S42: The weights of the higher-order feature representations are optimized layer by layer through an inter-layer unsupervised learning algorithm to obtain a pre-trained DCNN-DBN hybrid network structure.

[0140] Step S42 further includes:

[0141] S421: Based on the principle of restricted Boltzmann machine, the weights of the first hidden layer in the higher-order feature representation weights are trained in an unsupervised manner to obtain the optimized weight parameters of the first layer.

[0142] Furthermore, the Restricted Boltzmann Machine (RBM) is an unsupervised probabilistic graphical model consisting of a visible layer and hidden layers. Neurons within a layer are unconnected, while neurons between layers are fully connected. The first hidden layer weight optimization uses the 512-dimensional feature vector output by global average pooling as the visible layer, and sets 256 hidden units to form the hidden layer. Unsupervised training of the RBM employs the contrastive divergence (CD) algorithm, with two phases: forward propagation and negative propagation. In the forward propagation phase, the activation probability of the hidden layer is calculated based on the visible layer state; in the negative propagation phase, the visible layer is reconstructed based on the hidden layer state. The training process iterates for 100 epochs, with each epoch containing a complete traversal of all training samples. The final convergence is determined based on the reconstruction error; training stops when the error change is less than 0.001 over 10 consecutive epochs. The obtained first-layer optimized weight parameters include a 512×256 weight matrix, a 512-dimensional visible layer bias, and a 256-dimensional hidden layer bias, totaling approximately 130,000 parameters.

[0143] S422: The subsequent hidden layers with optimized weight parameters of the first layer are sequentially optimized using a greedy layer-by-layer training strategy to obtain multi-layer weight configurations for independent training of each layer.

[0144] Further, in step S422, this invention trains each hidden layer sequentially from bottom to top according to a greedy layer-by-layer training strategy. After each layer is trained, the weights of that layer are fixed, and the next layer is trained. The training of the second hidden layer uses the 256-dimensional output of the first hidden layer as input and sets 128 hidden units. The training process repeats the contrastive divergence algorithm of RBM, but the energy function is adjusted to the state of the first hidden layer, the state of the second hidden layer, and the bias of the second layer. Forward propagation calculates the activation probability of the second layer, and negative propagation reconstructs the state of the first layer. The weight update formula remains unchanged, and the learning rate is adjusted to 0.005 to adapt to deeper training. After the training of the second layer is completed, the 256×128 weight matrix, the 256-dimensional input layer bias, and the 128-dimensional hidden layer bias are saved. The training of the third hidden layer uses the 128-dimensional output of the second layer as input and sets 2 output units corresponding to the binary classification task. The training process of the third layer is similar, but with the addition of supervision information. A classification error term is added to the RBM objective function.

[0145]

[0146] in, For RBM loss, The balance coefficient is 0.1. For cross-entropy loss, the multi-layer weight configuration trained independently for each layer includes the weight parameters of three layers of RBM: the first layer [512×256], the second layer [256×128], and the third layer [128×2], as well as the corresponding bias parameters, totaling approximately 180,000 parameters.

[0147] S423: Based on the criterion of minimizing reconstruction error, the multi-layer weight configuration is globally adjusted to obtain the global weight distribution.

[0148] Furthermore, in step S423, the reconstruction error minimization criterion performs end-to-end global optimization of the entire network through a deep autoencoder structure. During this global adjustment process, the weights obtained from layer-by-layer pre-training are first used as initial values ​​to construct a symmetric encoder-decoder network structure. The encoder path performs feature compression in decreasing dimensions from 512 to 256 to 128 to 2, while the decoder path performs feature reconstruction in increasing dimensions from 2 to 128 to 256 to 512. Global optimization employs the Adam optimizer with a learning rate of 0.0001, a first momentum parameter of 0.9, and a second momentum parameter of 0.999. The optimization process utilizes batch gradient descent with a batch size of 32 samples. Gradient calculation is performed using the backpropagation algorithm. During training, the reconstruction error on the validation set is monitored. If the error fails to decrease for 20 consecutive epochs, early stopping is employed to terminate training. Then, global adjustments are made based on the set parameters, i.e., all weight parameters are jointly optimized to eliminate accumulated errors during layer-by-layer training. The global weight distribution maintains the original three-layer structure, but the weight values ​​are adjusted through global optimization, resulting in a more uniform distribution of weight matrices across layers, thus avoiding gradient vanishing or exploding problems.

[0149] S424: Integrate the global weight distribution into a network to obtain a pre-trained DCNN-DBN hybrid network structure.

[0150] Furthermore, in step S424, the present invention organically integrates the feature extraction part of the deep convolutional neural network (DCNN) and the classification part of the deep belief network (DBN) through a network integration process. The integration adopts a serial connection method. The DCNN part is responsible for extracting spatial features from the original meteorological data, and the DBN part is responsible for mapping the spatial features to the probability output of hail recognition.

[0151] During the integration process, the interface compatibility of the two network parts was first verified. The global average pooling layer of the DCNN part outputs a 512-dimensional feature vector, which matches the input dimension of the first layer of the DBN part. This invention uses a network ensemble algorithm to sequentially connect the four convolutional layers and one pooling layer of the DCNN with the three fully connected layers of the DBN, forming an end-to-end deep network. During the connection process, the weight parameters of each layer remain unchanged; only the data flow direction and gradient propagation path between layers are adjusted. The total depth of the integrated network is 8 layers (4 convolutional layers + 1 pooling layer + 3 fully connected layers), with a total of approximately 3.68 million parameters.

[0152] The final pre-trained DCNN-DBN hybrid network structure possesses powerful feature representation capabilities. The DCNN part can automatically extract the spatial distribution patterns of the hail system, while the DBN part can establish a non-linear mapping relationship between features and the probability of hail occurrence. The obtained hybrid network structure is saved in a standard deep learning model format, containing complete information such as network topology definition, weight parameter values, and hyperparameter configuration, forming the basis model for subsequent supervised fine-tuning training.

[0153] S43: The DCNN-DBN hybrid network structure is supervised and fine-tuned using the BP algorithm to obtain a ground hail recognition model.

[0154] Furthermore, in step S43, the present invention uses supervised fine-tuning training and backpropagation (BP) algorithm to perform end-to-end supervised learning on the pre-trained DCNN-DBN hybrid network. The aim is to finally use the labeled training sample set with phase labels to transform the hail recognition task into a binary classification problem and output the probability distribution of hail occurrence.

[0155] The BP algorithm calculates the gradient of the loss function with respect to the weights of each layer using the chain rule. The gradient calculation propagates layer by layer from the output layer to the input layer. Fine-tuning training uses a stochastic gradient descent optimizer with a learning rate of 0.0001 and a momentum coefficient of 0.9. Data augmentation techniques are used during training, including random rotation, flipping, and scaling to increase sample diversity. The batch size is set to 16 samples, and the number of training epochs is set to 100. The validation strategy uses 5-fold cross-validation, dividing the training data into training and validation sets in an 8:2 ratio. In addition, this invention introduces an early stopping mechanism to monitor the accuracy on the validation set. Training stops when the accuracy does not improve for 15 consecutive epochs. Model performance is evaluated using metrics such as accuracy, precision, recall, and F1 score. The model's predictive performance is analyzed using a confusion matrix.

[0156] Finally, after fine-tuning, the obtained model saves information including the complete network structure, optimized weight parameters, preprocessing parameters, etc., forming a hail recognition system that can be directly deployed and applied. The obtained ground hail recognition model can output the probability of hail occurrence between 0 and 1, and areas with a probability greater than 0.5 are identified as hail landing areas.

[0157] S5: Output the hail landing area identification result through the ground hail identification model.

[0158] Furthermore, the final identification result of step S5 is the spatial location information of the hail landing area, specifically including: Spatial location data: the hail landing area is identified through a 50×50 pixel window, and the output shows whether hail exists within the specific latitude and longitude coordinate range, which is equivalent to making a point-by-point judgment on a fine grid of 0.001°×0.001°; Binary classification result: each spatial window is output with a label of 0 or 1, where 1 indicates that hail occurred in the area and 0 indicates that hail did not occur; Probability information: since the identification model obtained in step S4 of this invention uses a Softmax output layer based on a BP neural network, the final model will output the probability value (between 0 and 1) of hail occurring in each area, which facilitates the setting of different confidence thresholds. In addition, based on the aforementioned hail landing area identification results, this invention also provides a planar distribution map of the hail landing area, showing which areas may produce hail and which areas do not produce hail, forming a continuous spatial distribution pattern.

[0159] For example, if severe convective weather occurs in a certain area and radar detects an echo with a reflectivity exceeding 50 dBZ, based on the algorithm of this invention, firstly, effective observation data within a 200 km radius of the radar station is extracted, and a time window of 18 minutes before and 12 minutes after the reference time is established. The dual polarization parameters show Zh=52 dBZ, ZDR=0.3 dB, and ρHV=0.87. The membership function is dynamically optimized to identify it as a hail phase. Combining ERA5 reanalysis data, the 0°C layer height is determined to be 3200 meters. The echo is located at an altitude of 4000 meters, belonging to the ice phase region. By integrating 43 meteorological elements such as surface temperature 28°C, relative humidity 65%, and wind speed 12 m / s, a 79-dimensional feature vector is formed and input into the DCNN-DBN model. The model outputs that the probability of hail occurrence in this area is 0.85, exceeding the 0.5 threshold, and is finally identified as a hailfall area. The binary classification result of the presence of hail within the latitude and longitude range of this area is output.

[0160] like Figure 2 As shown, the present invention also provides a ground hail identification system based on hydrogel classification results, comprising:

[0161] Association module 100: Used to perform spatiotemporal association processing on multi-source hail data through a spatiotemporal matching algorithm to obtain a spatiotemporally matched hail event dataset;

[0162] Identification module 200: Used to perform adaptive phase identification on dual-polarization radar data in the hail event dataset through a dynamic membership function optimization algorithm, and obtain multi-elevation angle hail phase characteristic parameters including rain-ice mixture categories;

[0163] Preprocessing module 300: Based on the multi-elevation angle hail phase characteristic parameters, it integrates and preprocesses multi-source meteorological data through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data with fused phase characteristics;

[0164] Training module 400: Used to perform unsupervised pre-training and supervised fine-tuning training of the DCNN-DBN hybrid neural network using the standardized multidimensional meteorological feature data to obtain a ground hail recognition model;

[0165] The identification module 500 is configured to use the ground hail identification model trained by the training module 400 to output hail landing area identification results.

[0166] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0167] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the ground hail identification method based on hydrogel classification results described in the various embodiments or some parts of the embodiments.

[0168] This invention enables precise identification and location of ground hail, significantly improving the accuracy and timeliness of hail warnings, refining phase identification, effectively solving the problem of hail being easily confused with other water-borne substances, and significantly improving the accuracy of phase discrimination, especially in complex weather conditions. It lays a solid foundation for the accurate identification of hail events, and the overall method provides important technical support for the monitoring, early warning, prevention and mitigation of hail disasters.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying ground hail based on hydrogel classification results, characterized in that, include: S1: Using a spatiotemporal matching algorithm, multi-source hail data are spatiotemporally correlated to obtain a spatiotemporally matched hail event dataset; S2: Adaptive phase identification is performed on the dual-polarization radar data in the hail event dataset using a dynamic membership function optimization algorithm to obtain multi-elevation angle hail phase characteristic parameters that include rain-ice mixture categories; Step S2 further includes: S21: Based on the A-test statistic of the hail event dataset, the membership function threshold of the polarization parameter in the Park dual-polarization phase classification algorithm is adjusted in real time to obtain a dynamically optimized phase discrimination threshold; wherein the polarization parameter includes the horizontal reflectivity factor and differential reflectivity; Step S21 further includes: S211: Calculate the A-test statistic of the hail event dataset to obtain the statistical test index of the radar data distribution; S212: Based on statistical test indicators, the membership function thresholds of reflectivity factor and differential reflectivity in the Park algorithm are dynamically adjusted to obtain adaptive phase discrimination parameters; S213: The easily confused phases under the phase discrimination parameters are reclassified by real-time parameter updates to obtain phase recognition results with enhanced discriminative power; the easily confused phases include ice crystals and dry snow; S214: Apply the phase recognition results to the entire observation period for batch processing to obtain a dynamically optimized phase discrimination threshold; S22: Adaptively correct the polarization parameter error using the weight matrix to obtain an optimized dual polarization parameter weight configuration; S23: Based on the cross-correlation coefficient, identify the rain-ice mixture in the region above the melting layer under the dual polarization parameter weight configuration to obtain an extended classification result containing the mixed phase state; Step S23 further includes: S231: By using cross-correlation coefficient threshold discrimination, the radar observation data under the optimized dual polarization parameter weight configuration is used to locate the mixed phase region and obtain the distribution data of potential rain-ice mixing zone; S232: Based on the height information, determine the position of the melting layer in the rain-ice mixing zone corresponding to the rain-ice mixing zone distribution data, and obtain the range of the ice phase region above the 0°C layer height; S233: Spatial superposition analysis of the ice phase region and the mixed phase region is performed to obtain the location information of the rain-ice mixture; S234: By using the mixture category label, the phase classification of the location information of the rain-ice mixture is extended to obtain an extended classification result that includes the mixed phase. S24: Perform a multi-angle comprehensive evaluation on the extended classification results to obtain multi-angle hail phase characteristic parameters that include rain-ice mixture categories; S3: Based on the multi-elevation angle hail phase characteristic parameters, multi-source meteorological data are integrated and preprocessed through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data with fused phase characteristics; S4: Using the standardized multidimensional meteorological feature data, the DCNN-DBN hybrid neural network is subjected to unsupervised pre-training and supervised fine-tuning training to obtain a ground hail recognition model; wherein, the DCNN-DBN hybrid neural network is a fusion network in which the feature extraction part of the deep convolutional neural network and the classification part of the deep belief network are connected in series. S5: Output the hail landing area identification result through the ground hail identification model.

2. The ground hail identification method based on hydrogel classification results according to claim 1, characterized in that, Step S1 further includes: S11: Spatial filtering of the radar observation area is performed based on the preset weather station locations to obtain the effective detection area; S12: By using a dynamic time window strategy, hail events within the effective detection area are time-matched to obtain a time matching window based on the volume scan time. S13: Perform spatial correlation processing on the actual hail locations within the time matching window to obtain radar echo matching results within a preset distance search radius; S14: The radar echo matching results are fused with publicly available data to obtain a spatiotemporally matched hail event dataset.

3. The ground hail identification method based on hydrocondensate classification results according to claim 1, characterized in that, Step S3 further includes: S31: Grid the ERA5 data and ground observation data to obtain uniform grid data; S32: Extract hailfall areas from the unified grid data to obtain a multi-dimensional feature window containing phase characteristics; S33: Based on positive and negative samples, generate labels for the multidimensional feature window to obtain a labeled training sample set with phase identifiers; S34: Standardize the labeled training sample set to obtain standardized multidimensional meteorological feature data that incorporates phase features.

4. The ground hail identification method based on hydrocondensate classification results according to claim 1, characterized in that, Step S4 further includes: S41: Perform multi-layer convolution on the standardized multi-dimensional meteorological feature data to pre-train the deep belief network structure and obtain high-order feature representation weights; S42: The weights of the higher-order feature representations are optimized layer by layer through an inter-layer unsupervised learning algorithm to obtain a pre-trained DCNN-DBN hybrid network structure; S43: The DCNN-DBN hybrid network structure is supervised and fine-tuned using the BP algorithm to obtain a ground hail recognition model.

5. The ground hail identification method based on hydrogel classification results according to claim 4, characterized in that, Step S41 further includes: S411: The standardized multidimensional meteorological feature data is reorganized into tensors according to a preset window structure to obtain a data tensor suitable for convolution operations; S412: Input the data tensor into a multi-layer convolutional kernel to extract features from the data tensor and obtain a high-order feature map; S413: According to the unsupervised learning principle, the weights of each convolutional layer corresponding to the higher-order feature map are initialized layer by layer to obtain the initial weight configuration; S414: Integrate the pre-trained convolutional layer with the initial weight configuration with the fully connected layer to obtain higher-order feature representation weights.

6. The ground hail identification method based on hydrocondensate classification results according to claim 4, characterized in that, Step S42 further includes: S421: Based on the principle of restricted Boltzmann machine, the weights of the first hidden layer in the weights of the higher-order feature representation are trained in an unsupervised manner to obtain the optimized weight parameters of the first layer. S422: The subsequent hidden layers with optimized weight parameters of the first layer are sequentially optimized using a greedy layer-by-layer training strategy to obtain multi-layer weight configurations for independent training of each layer; S423: Based on the criterion of minimizing reconstruction error, the multi-layer weight configuration is globally adjusted to obtain the global weight distribution; S424: Integrate the global weight distribution into a network to obtain a pre-trained DCNN-DBN hybrid network structure.

7. A ground hail identification system based on hydrocondensate classification results, characterized in that, include: The association module is used to perform spatiotemporal association processing on multi-source hail data using a spatiotemporal matching algorithm to obtain a spatiotemporally matched hail event dataset. Identification module: used to adaptively identify the phase state of dual-polarization radar data in the hail event dataset through a dynamic membership function optimization algorithm, and obtain multi-elevation angle hail phase state characteristic parameters including rain-ice mixture categories; The identification module is further configured to: adjust the membership function threshold of the polarization parameters in the Park dual-polarization phase classification algorithm in real time based on the A-test statistic of the hail event dataset to obtain a dynamically optimized phase discrimination threshold; wherein the polarization parameters include the horizontal reflectivity factor and differential reflectivity; adaptively correct the polarization parameter error using the weight matrix to obtain an optimized dual-polarization parameter weight configuration; identify rain-ice mixtures in the region above the melting layer under the dual-polarization parameter weight configuration based on the cross-correlation coefficient to obtain an extended classification result containing the mixed phase; and perform multi-elevation angle comprehensive evaluation on the extended classification result to obtain multi-elevation angle hail phase feature parameters containing the rain-ice mixture category. The identification module is further used to: calculate the A-test statistic of the hail event dataset to obtain the statistical test index of the radar data distribution; and dynamically adjust the membership function thresholds of the reflectivity factor and differential reflectivity in the Park algorithm according to the statistical test index to obtain adaptive phase discrimination parameters. By updating parameters in real time, easily confused phases under the phase discrimination parameters are reclassified to obtain phase recognition results with enhanced discriminative power; the easily confused phases include ice crystals and dry snow; the phase recognition results are applied to the entire observation period for batch processing to obtain dynamically optimized phase discrimination thresholds; The identification module is further used to: locate the mixed-phase region of radar observation data under optimized dual-polarization parameter weight configuration by using cross-correlation coefficient threshold discrimination to obtain potential rain-ice mixing zone distribution data; determine the melting layer position in the rain-ice mixing zone corresponding to the rain-ice mixing zone distribution data based on height information to obtain the range of ice phase region above 0°C layer height; and perform spatial overlay analysis of the ice phase region range and the mixed-phase region to obtain rain-ice mixture location information. By labeling the mixture category, the location information of the rain-ice mixture is extended to include the phase classification, resulting in an extended classification result that includes the mixed phases. Preprocessing module: Based on the multi-elevation angle hail phase characteristic parameters, it integrates and preprocesses multi-source meteorological data through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data with fused phase characteristics; Training module: used to perform unsupervised pre-training and supervised fine-tuning training of the DCNN-DBN hybrid neural network using the standardized multidimensional meteorological feature data to obtain a ground hail recognition model; wherein, the DCNN-DBN hybrid neural network is a fusion network in which the feature extraction part of a deep convolutional neural network and the classification part of a deep belief network are connected in series. The identification module is configured to use the ground hail identification model trained by the training module to output hail landing area identification results.

Images