Ground hail identification method and system based on hydrogel classification result

Abstract

The invention relates to the technical field of meteorological observation, and provides a ground hail identification method and system based on a hydrogel classification result, and the method comprises the steps: carrying out the time-space correlation of multi-source hail data through a time-space matching algorithm, and obtaining a hail event data set of time-space matching; through a dynamic membership function optimization algorithm, self-adaptive phase state identification is carried out on the dual-polarization radar data, and multi-elevation hail phase state characteristic parameters containing rain-ice mixture categories are obtained; based on the multi-elevation hail phase state characteristic parameters, performing integrated preprocessing on the multi-source meteorological data to obtain standardized multi-dimensional meteorological characteristic data fused with phase state characteristics; performing unsupervised pre-training and supervised fine-tuning training on the DCNN-DBN hybrid neural network through the standardized multi-dimensional meteorological feature data to obtain a ground hail recognition model; and outputting a hail falling area identification result through the ground hail identification model. According to the invention, the distinguishing capability of easily-confused phase states is improved, and the false alarm rate and the missing report rate of hail identification are reduced.

Description

Technical Field

[0001] The present invention relates to the field of meteorological observation technology, and in particular to a method and system for identifying ground hail based on hydrometeor classification results. Background Art

[0002] Hail is a significant and potentially devastating weather phenomenon associated with severe convective systems. Currently, hail identification technology primarily relies on dual-polarization Doppler radar detection, which analyzes the polarization characteristics of radar echoes to identify different hydrometeor phases. Traditional hail identification methods primarily employ a dual-polarization radar hydrometeor classification algorithm. This algorithm uses fixed membership function thresholds based on parameters such as reflectivity factor, differential reflectivity, and differential phase shift to classify precipitation particles into 10 categories: graupel, ice crystals, wet snow, dry snow, ground objects, clear-air echoes, large raindrops, hail, heavy rain, and light rain. Furthermore, traditional machine learning methods, such as BP neural networks, are also widely used in hail identification. These methods train radar feature parameters through the construction of multilayer perceptron networks for classification.

[0003] However, existing technologies have significant shortcomings: First, traditional algorithms use fixed membership function thresholds and cannot adaptively adjust 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 hydrometeor classification algorithms only consider pure phases and ignore the rain-ice mixture state that is prevalent near the melt layer, resulting in missed and false positives in hail identification; third, traditional BP neural networks are prone to overfitting under small sample data and lack effective feature extraction capabilities, making it difficult to fully exploit the high-order spatial feature information contained in radar data, resulting in insufficient accuracy in hail identification. Summary of the Invention

[0004] The present invention provides a ground hail identification method and system based on hydrometeor classification results, so as to solve the defects of the prior art.

[0005] The present invention provides a method for identifying ground hail based on hydrometeor classification results, comprising: S1: Using the space-time matching algorithm, the multi-source hail data is subjected to space-time correlation processing to obtain a space-time matching hail event dataset; S2: Adaptively identifying the phase state of 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 including the rain-ice mixture category; S3: Based on the multi-elevation-angle hail phase characteristic parameters, integrating and preprocessing the multi-source meteorological data through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data of the fused phase characteristics; S4: performing unsupervised pre-training and supervised fine-tuning training on the DCNN-DBN hybrid neural network using the standardized multi-dimensional meteorological feature data to obtain a ground hail recognition model; S5: Outputting a hailfall area recognition result through the ground hail recognition model.

[0006] According to a method for identifying ground hail based on hydrometeor classification results provided by the present invention, step S1 further includes: S11: spatially screen the radar observation area according to the preset weather station location to obtain the effective detection area; S12: Using a dynamic time window strategy, time matching is performed on hail events within the effective detection area to obtain a time matching window based on the volume scanning time; S13: performing spatial correlation processing on the actual position of the hailstone within the time matching window to obtain a radar echo matching result within a preset distance search radius; S14: Perform multi-source fusion on the radar echo matching results and public data to obtain a time-space matched hail event dataset.

[0007] According to a method for identifying ground hail based on hydrometeor classification results provided by the present invention, step S2 further includes: S21: adjusting the membership function threshold of the polarization parameter in the Park dual-polarization phase state classification algorithm in real time according to the A test statistic of the hail event dataset to obtain a dynamically optimized phase state discrimination threshold; wherein the polarization parameter includes a horizontal reflectivity factor and a differential reflectivity; S22: performing adaptive correction on the weight matrix of the polarization parameter error to obtain an optimized dual polarization parameter weight configuration; S23: performing rain-ice mixture identification on the region above the melting layer under the dual polarization parameter weight configuration according to the mutual correlation coefficient, to obtain an extended classification result including a mixed phase state; S24: Performing a multi-elevation-angle comprehensive evaluation on the extended classification results to obtain multi-elevation-angle hail phase characteristic parameters including a rain-ice mixture category.

[0008] According to a method for identifying ground hail based on hydrometeor classification results provided by the present invention, step S21 further includes: S211: Calculate the A test statistic of the hail event data set to obtain a statistical test index of the radar data distribution; S212: dynamically adjusting the membership function thresholds of the reflectivity factor and the differential reflectivity in the Park algorithm according to the statistical test index to obtain an adaptive phase state discrimination parameter; S213: reclassifying the easily confused phases under the phase discrimination parameters by real-time parameter updating to obtain a phase recognition result with enhanced discrimination; the easily confused phases include ice crystals and dry snow; S214: Apply the phase state recognition result to the entire observation period, perform batch processing, and obtain a dynamically optimized phase state discrimination threshold.

[0009] According to a method for identifying ground hail based on hydrometeor classification results provided by the present invention, step S23 further includes: S231: Using the cross-correlation coefficient threshold to determine the mixed phase region of radar observation data under the optimized dual-polarization parameter weight configuration, the potential distribution data of the rain-ice mixed region is obtained; S232: Determine the position of the melting layer in the rain-ice mixed zone corresponding to the rain-ice mixed zone distribution data based on the height information, and obtain the range of the ice phase area above the 0°C layer height; S233: Performing spatial superposition analysis on the ice phase region and the mixed phase region to obtain position information of the rain-ice mixture; S234: Performing phase classification expansion on the rain-ice mixture position information through mixture category labeling to obtain an expanded classification result including the mixture phase.

[0010] According to a method for identifying ground hail based on hydrometeor classification results provided by the present invention, step S3 further includes: S31: Gridding the ERA5 data and ground observation data to obtain unified grid data; S32: extracting the hailfall area in the unified grid data to obtain a multidimensional feature window containing phase characteristics; S33: Based on the positive and negative samples, labels are generated for the multidimensional feature window to obtain a labeled training sample set with phase identification; S34: Standardizing the labeled training sample set to obtain standardized multi-dimensional meteorological feature data fused with phase characteristics.

[0011] According to a method for identifying ground hail based on hydrometeor classification results provided by the present invention, step S4 further includes: S41: performing multi-layer convolution on the standardized multi-dimensional meteorological feature data to pre-train a deep belief network structure to obtain high-order feature representation weights; S42: Optimizing the high-order feature representation weights layer by layer through an inter-layer unsupervised learning algorithm to obtain a pre-trained DCNN-DBN hybrid network structure; S43: Using the BP algorithm, the DCNN-DBN hybrid network structure is fine-tuned to obtain a ground hail recognition model.

[0012] According to a method for identifying ground hail based on hydrometeor classification results provided by the present invention, step S41 further includes: S411: reorganizing the standardized multi-dimensional meteorological characteristic data into tensors according to a preset window structure to obtain a data tensor suitable for a convolution operation; S412: Inputting the data tensor into a multi-layer convolution kernel to perform feature extraction on the data tensor to obtain a high-order feature map; S413: Initializing the weights of each convolutional layer corresponding to the high-order feature map layer by layer according to an unsupervised learning criterion to obtain an initial weight configuration; S414: Integrate the pre-trained convolutional layer with the initial weight configuration and the fully connected layer to obtain high-order feature representation weights.

[0013] According to a method for identifying ground hail based on hydrometeor classification results provided by the present invention, step S42 further includes: S421: performing unsupervised training on the first hidden layer weights in the high-order feature representation weights according to the restricted Boltzmann machine principle to obtain first-layer optimized weight parameters; S422: Optimizing the subsequent hidden layers of the first layer's optimized weight parameters in sequence through a greedy layer-by-layer training strategy to obtain a multi-layer weight configuration in which each layer is independently trained; S423: Globally adjust the multi-layer weight configuration according to a reconstruction error minimization criterion to obtain a global weight distribution; S424: Performing network integration on the global weight distribution to obtain a pre-trained DCNN-DBN hybrid network structure.

[0014] The present invention also provides a ground hail identification system based on hydrometeor classification results, comprising: Correlation module: used to perform spatiotemporal correlation processing on multi-source hail data through spatiotemporal matching algorithm to obtain a spatiotemporal matching hail event dataset; Identification module: used to perform adaptive phase identification 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 including rain-ice mixture categories; Preprocessing module: used for integrating and preprocessing multi-source meteorological data based on the multi-elevation angle hail phase characteristic parameters through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data of fused phase characteristics; Training module: used to perform unsupervised pre-training and supervised fine-tuning training on the DCNN-DBN hybrid neural network using the standardized multi-dimensional meteorological feature data to obtain a ground hail recognition model; The recognition module is configured as the ground hail recognition model obtained through training by the training module, and is used to output a hailfall area recognition result.

[0015] The present invention provides a method and system for ground hail identification based on hydrometeor 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 the present invention uses the A test statistic to adjust the polarization parameter membership function threshold in the Park algorithm in real time, so that the present invention can adaptively optimize the recognition parameters according to different meteorological conditions and geographical environments, effectively solving the problem of reduced recognition accuracy of the traditional fixed threshold method in complex meteorological environments, especially improving the ability to distinguish easily confused phases such as ice crystals and dry snow, and greatly reducing the false alarm rate and missed alarm rate of hail recognition; secondly, the rain-ice mixture recognition technology based on the mutual correlation coefficient of the present invention fills the technical gap of the traditional hydrometeor classification algorithm that only considers pure phases. By identifying the mixed phase near the melting layer, the traditional 10 phases are expanded to an 11-category classification system including rain-ice mixtures, making the hail recognition coverage more comprehensive, the recognition accuracy is improved, and the recognition blind spots caused by ignoring the mixed phase are effectively reduced; in addition, the introduction of the DCNN-DBN hybrid neural network architecture of the present invention realizes the algorithm The features improve the 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 contained in the radar data through multi-layer convolution operations, effectively solving the overfitting problem of the traditional BP neural network in small sample scenarios. At the same time, the inter-layer unsupervised learning algorithm enables the network to learn richer feature representations under unlabeled data conditions through the restricted Boltzmann machine principle and greedy layer-by-layer training strategy, significantly reducing the dependence on a large number of labeled samples, and improving the recognition accuracy with the same amount of training data; in addition, the multi-source data fusion mechanism organically integrates dual-polarization radar data, ERA5 reanalysis data and ground observation data through unified grid processing, and constructs a high-dimensional feature space containing 43 meteorological elements, enabling 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 the recognition accuracy under complex meteorological conditions compared with a single data source.

[0016] Overall, this invention, through the organic combination of algorithm innovation and network architecture optimization, not only achieves a leap in the accuracy of hail recognition technology, but also provides more reliable and efficient technical support for practical application scenarios such as insurance claims and agricultural disaster prevention. It has important scientific value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic flow chart of a method for identifying ground hail based on hydrometeor classification results provided by the present invention; Figure 2 A schematic structural diagram of a ground hail identification system based on hydrometeor classification results provided by the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0020] The following describes embodiments of the present invention with reference to the accompanying drawings.

[0021] like Figure 1 As shown, the present invention provides a ground hail identification method based on hydrometeor classification results, comprising: S1: Through the space-time matching algorithm, the multi-source hail data is processed in space-time correlation to obtain a space-time matching hail event dataset.

[0022] Wherein, step S1 further includes: S11: Spatially screen the radar observation area according to the preset weather station location to obtain an effective detection area.

[0023] Furthermore, step S11 of the present invention first performs spatial screening. In a specific embodiment, the latitude and longitude coordinate information of all weather stations in area A is first read, including the coordinates of six stations equipped with S-band dual-polarization radars. Subsequently, the Euclidean distance calculation formula is used to calculate the spherical distance between each weather 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 due to the radar's lowest elevation angle of 0.5°, which results in a detection altitude that is too low. Areas exceeding 200 kilometers are excluded due to severe signal attenuation, which results in insufficient detection accuracy. Then, screening is performed based on the distance threshold and the calculated spherical distance. The screening process traverses all weather station coordinates, calculates the distance to each radar station one by one, marks weather stations that meet the distance range conditions as valid sites, and constructs a geographical boundary polygon of the effective detection area. The final effective detection area is centered on the radar station and has a ring range of 50-200 kilometers, forming a sector-shaped or ring-shaped spatial geometry. Each effective detection area includes attribute information such as its geographical boundary coordinates, coverage area, and the number of weather stations included.

[0024] S12: Using a dynamic time window strategy, time matching is performed on hail events within the effective detection area to obtain a time matching window based on the volume scanning time.

[0025] Furthermore, the dynamic time window strategy of the present invention is designed based on the radar volume scan interval. In this embodiment of the present invention, the standard volume scan interval for the dual-polarization radar in area A is 6 minutes. Specifically, in step S12, the start time record of the hail event is first obtained. Then, the volume scan moment closest to the hail start time is searched in the radar volume scan time series as the reference time point. After determining the reference time, the algorithm extrapolates three volume scan times forward as the starting boundary of the time window (i.e., the reference time minus 18 minutes). It also extrapolates two volume scan times backward as the ending boundary of the time window (i.e., the reference time plus 12 minutes), forming a time matching window with a total length of 30 minutes.

[0026] When a hail event has a clear end time record, this method uses a dynamic adjustment mechanism to use the actual end time as the cutoff boundary of the time window, replacing the default two-scan time setting. The time matching window is stored in timestamp format and contains time attribute information such as the window start time, end time, reference scan time, and total window length.

[0027] S13: performing spatial correlation processing on the actual position of the hail within the time matching window to obtain a radar echo matching result within a preset distance search radius.

[0028] Furthermore, spatial correlation processing uses a point-to-plane neighborhood test 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. The search radius is determined based on the typical spatial scale of the hail system and the spatial resolution of radar detection. First, the radar base data at 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. The radar polar coordinates are then converted to a geographic coordinate system. After the conversion is completed, a circular buffer zone with a radius of 20 kilometers is established with the actual hail location as the center. Spatial geometric calculations are used to determine whether the radar pixel falls within the buffer zone. Radar pixels that meet the spatial conditions are extracted, forming a radar echo data set spatially associated with the actual hail location. Each pixel contains complete information such as its geographic coordinates, dual-polarization parameter values, and observation time.

[0029] S14: Perform multi-source fusion on the radar echo matching results and public data to obtain a time-space matched hail event dataset.

[0030] Furthermore, the publicly available data includes weather station observation data, disaster survey data, and insurance claims data. Weather station observation data includes hourly observations of common 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, location, and severity of the disaster. Insurance claims data records crop losses caused by hail in agricultural insurance, including detailed information such as the claim time, location, and extent of the loss. After acquiring the data, a spatiotemporal joint matching strategy is used for data fusion. First, temporal matching is performed to link the records closest to the radar observation time from each data source. Then, spatial matching is performed to fuse observations within a reasonable spatial distance, using the actual hail location as a reference. For disaster survey and insurance claims data, the coordinates of the geometric center point of the lowest-precision area are extracted as the representative location. During the fusion process, when the combined reflectivity exceeds 45 dBZ for more than three radar observation moments, and the proportion of hail pixels within a 20-kilometer radius reaches a maximum value using the HCL hydrometeor classification product, that moment is marked as the hail occurrence moment. The final spatiotemporal matching hail event dataset contains multi-dimensional information such as the time information, spatial location, radar observation parameters, ground observation data, disaster records, etc. of each hail event, constituting a complete description of the hail event characteristics.

[0031] S2: Adaptively identify the phase state of 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 including the rain-ice mixture category.

[0032] Wherein, step S2 further includes: S21: According to the A test statistic of the hail event data set, the membership function threshold of the polarization parameter in the Park dual-polarization phase state classification algorithm is adjusted in real time to obtain a dynamically optimized phase state discrimination threshold; wherein the polarization parameter includes a horizontal reflectivity factor and a differential reflectivity.

[0033] Wherein, step S21 further includes: S211: Calculate the A test statistic of the hail event data set to obtain a statistical test index of the radar data distribution.

[0034] Furthermore, 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, the present invention extracts radar data corresponding to each hail event according to elevation angle layers, 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 layer, a numerical sequence is constructed, and the sequence length is equal to the number of all valid pixels in the elevation layer. Subsequently, the Zh and ZDR data are arranged in ascending order, and then the cumulative probability value of each data point under the assumed normal distribution is calculated. The cumulative probability calculation adopts the standard normal distribution function, and finally the A test statistic of Zh and ZDR is calculated respectively. The calculation formula is: in, To calculate the obtained A test statistic, is the sample size, is the observation sample index value, For the sorted observations, is the cumulative distribution function, and finally two independent statistical test indicators are obtained, which quantify the degree of deviation of the actual radar observation data from the theoretical normal distribution.

[0035] S212: Dynamically adjust the membership function thresholds of the reflectivity factor and the differential reflectivity in the Park algorithm according to the statistical test index to obtain an adaptive phase state discrimination parameter.

[0036] Furthermore, the membership function of the Park algorithm adopts a trapezoidal function form, with each phase state category corresponding to a set of trapezoidal parameters. The dynamic adjustment process in step S212 performs parameter correction based on the numerical value of the A test statistic. When the Az value exceeds the critical threshold, indicating that the Zh data distribution significantly deviates from the normal distribution, the parameter adjustment mechanism is activated, and the adjustment mechanism is a linear correction through the adjustment coefficient. In addition, the present invention also simultaneously adjusts the membership function parameters of other phase state categories, including large raindrops, graupel, ice crystals, dry snow, etc. The adjustment amplitude of each phase state is weighted and calculated based on its similarity to the hail phase state. After the adjustment is completed, an adaptive phase state discrimination parameter matrix containing the correction parameters of all phase state categories is finally generated. The matrix dimension is 10×8, corresponding to the membership function boundary values ​​of 10 phase state categories and 8 dual-polarization parameters.

[0037] S213: reclassifying the easily confused phases under the phase discrimination parameters by real-time parameter updating to obtain a phase recognition result with enhanced discrimination.

[0038] Furthermore, the easily confused phases mentioned primarily refer to ice crystals and dry snow, which overlap with hail in the dual-polarization parameter space. Specifically, during the reclassification process, the membership of each radar pixel to each phase category is calculated under the modified membership function. Specifically, the membership of each pixel to hail, ice crystals, and dry snow is calculated separately under parameters such as Zh, ZDR, ρHV, and KDP to form a 3×4 membership matrix. The weighted geometric mean method is then used to calculate the overall membership. For pixels with similar membership values, the present invention introduces a secondary discrimination mechanism to calculate the relative membership difference between phases. When the membership difference between hail and ice crystals is less than 0.1, the present invention also adds temperature profile information for auxiliary discrimination. Areas with temperatures below -20°C tend to be discriminated as ice crystals, and areas above -10°C tend to be discriminated as hail. After reclassification, the final phase category identifier and corresponding confidence value of each pixel are output to form a phase recognition result matrix with enhanced discrimination.

[0039] S214: Apply the phase state recognition result to the entire observation period, perform batch processing, and obtain a dynamically optimized phase state discrimination threshold.

[0040] Specifically, the batch processing process is intended to expand the phase recognition results of a single moment to continuous data of the entire observation period. During the data processing process, the radar data of all body scan moments in the time matching window are first read, which usually contains 5-8 consecutive observation moments; then the sliding window technology is used, the window size is set to 3 consecutive moments, and the sliding step is 1 moment. For each sliding window, the time consistency index of the phase recognition result in the window is calculated. The consistency calculation method is: for each spatial pixel point, the frequency of being identified as the same phase state within 3 moments is counted. The pixel points with a frequency greater than or equal to 2 are marked as stable pixels, and the pixel points with a frequency of 1 are marked as unstable pixels; after calculating the consistency index, a phase transition probability matrix is ​​established to record the transition frequency between each phase state. , the matrix dimension is 10×10, corresponding to the mutual conversion relationship between 10 phase state categories; based on the conversion probability matrix, the present invention subsequently calculates the stability weight coefficient of each phase state, the phase state with high stability (such as hail, heavy rain) has a larger weight coefficient, and the phase state with low stability (such as ice crystals, graupel) has a smaller weight coefficient; then the stability weight coefficient is fed back to the membership function parameter for adjustment, the adjustment amplitude is increased for the phase state with low stability, and the adjustment amplitude is reduced for the phase state with high stability. After 3-5 iterations, it finally converges to the dynamically optimized phase state discrimination threshold. The obtained threshold not only maintains the adaptability to the current meteorological conditions, but also takes into account the stability requirements of the time series.

[0041] S22: Perform adaptive weight matrix correction on the polarization parameter error to obtain an optimized dual polarization parameter weight configuration.

[0042] Furthermore, in the error analysis process in 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 in the matrix are the error estimates of the parameters at the observation points. The specific expression is: in, is the Zh parameter error, is the calibration error, is the attenuation error (0.1-0.5dB), is the ground clutter error; in, is the ZDR parameter error, is the differential reflectivity calibration error, is the differential attenuation error, is the differential phase shift error.

[0043] The subsequent adaptive correction of the weight matrix adopts an error sensitivity test method. That is, the influence of each parameter error on the phase state identification result is calculated. A 4×10 weight matrix is ​​established, with rows corresponding to the four dual-polarization parameters (Zh, ZDR, ρHV, KDP) and columns corresponding to the 10 phase state categories. The weight correction formula is then used based on the sensitivity coefficient. The corrected weight configuration reduces the weight of high-error parameters and increases the weight of low-error parameters, forming an optimized dual-polarization parameter weight configuration matrix.

[0044] S23: According to the mutual correlation coefficient, the rain-ice mixture is identified in the region above the melting layer under the dual polarization parameter weight configuration to obtain an extended classification result including the mixed phase.

[0045] Wherein, step S23 further includes: S231: The radar observation data under the optimized dual-polarization parameter weight configuration are used to locate the mixed phase area through the cross-correlation coefficient threshold judgment to obtain the potential distribution data of the rain-ice mixed area.

[0046] Specifically, the cross-correlation coefficient ρHV is an important parameter for measuring the correlation between horizontal and vertical polarization echoes. Its value range is 0-1. The ρHV value of pure phase hydrometeor is usually close to 1, while the ρHV value of mixed phase region is significantly reduced. Therefore, in step S231, the present invention locates the mixed phase region by ρHV.

[0047] Furthermore, the ρHV threshold discrimination standard is first set, and the discrimination threshold is determined according to 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 area is between 0.7-0.9. The present invention adopts 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 area; when 0.7 ≤ ρHV < 0.9, the pixel is marked as a potential mixing area; when ρHV ≥ 0.9, the pixel is marked as a pure phase area. Then, the 8-connected domain algorithm is used to identify continuous mixed phase areas. Spatially adjacent pixels that meet the mixed area conditions are merged 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 mixed area attribute table is established to record the spatial position and intensity characteristics of each connected domain, forming potential rain-ice mixed area distribution data.

[0048] S232: Determine the melting layer position in the rain-ice mixed zone corresponding to the rain-ice mixed zone distribution data based on the height information, and obtain the range of the ice phase area above the 0°C layer height.

[0049] Furthermore, the determination of the melting layer position requires the combination of temperature profile information in the ERA5 reanalysis data. Therefore, in step S232, the present invention first reads the ERA5 hourly data closest to the radar observation time, extracts the temperature data of five isobaric surfaces, namely 1000hPa, 975hPa, 950hPa, 925hPa, and 900hPa, and then calculates the height position of the 0°C isotherm by linear interpolation. The interpolation process is performed between two adjacent isobaric surfaces. When the temperature of one isobaric surface is greater than 0°C and the temperature of the next isobaric surface is less than 0°C, the 0°C layer height is calculated as follows: in, is the height of the 0℃ layer, is the height of the first isobaric surface, is the height of the second isobaric surface, is the first isobaric surface temperature, The temperature of the second isobaric surface is then calculated. The observation height corresponding to the radar elevation angle is then compared to the 0°C layer for each radar pixel. Pixels above the 0°C layer are marked as ice-phase regions, while pixels below the 0°C layer are marked as liquid-phase regions. The resulting ice-phase region is stored in a highly layered manner, including information such as the spatial boundaries, height range, and number of pixels included in each elevation layer.

[0050] S233: Performing spatial overlay analysis on the ice phase region and the mixed phase region to obtain the position information of the rain-ice mixture.

[0051] Furthermore, the spatial overlay analysis in step S233 adopts the GIS spatial analysis method, and the ice phase area range and the mixed phase area are treated as two independent spatial layers for overlay operation. The basic unit of the overlay operation is the radar pixel point, and each pixel point contains its geographic coordinates, height information, phase properties and other spatial attributes.

[0052] Specifically, all radar pixels are first traversed to check whether each pixel satisfies both the ice phase region and the mixed phase region conditions. Pixels that meet the conditions are marked as rain-ice mixture pixels. The criteria for this are: (1) the pixel height > 0°C layer height; (2) the pixel ρHV value is within the range of 0.7-0.9; and (3) the pixel belongs to a connected mixed phase region. An attribute table of rain-ice mixture pixels is then established, recording detailed information such as the 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 2 pixels, and the minimum number of cluster points is set to 5 pixels. Finally, 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.

[0053] S234: Performing phase classification expansion on the rain-ice mixture position information through mixture category labeling to obtain an expanded classification result including the mixture phase.

[0054] Furthermore, the phase classification expansion in step S234 aims to expand the traditional 10-category hydrometeor classification system to an 11-category classification system that includes rain-ice mixtures. Specifically, during this expansion, the present invention first establishes a new category identifier for rain-ice mixtures, setting it to category number 11. The phase classification code table is then reconstructed, retaining the original 10 categories: 1-graupel, 2-ice crystals, 3-wet snow, 4-dry snow, 5-ground features, 6-clear air echo, 7-large raindrops, 8-hail, 9-heavy rain, 10-light rain, with the addition of 11-rain-ice mixtures. After this, the phase recognition results for all radar pixels are traversed, and the category identifiers for pixels meeting the rain-ice mixture condition are modified from the original category to category 11. This modification process utilizes a priority rule. When a pixel simultaneously meets multiple phase conditions, it is classified according to the following priority: rain-ice mixture > hail > graupel > large raindrops > other phases. In addition, the present invention also establishes a phase conversion mapping table to record the conversion relationship of each pixel point from the original classification to the expanded classification. The final expanded classification results are stored in the form of a three-dimensional array with the array dimensions of [number of elevation angles × number of distance bins × number of azimuth angles]. The array element values ​​are phase category numbers from 1 to 11. The classification results of each elevation layer are stored separately, forming seven two-dimensional classification matrices corresponding to the observation data of seven standard elevation angles. The expanded classification results also include statistical information for each phase category, such as quantitative indicators such as the number of pixels, proportion, and spatial distribution characteristics.

[0055] S24: Performing a multi-elevation-angle comprehensive evaluation on the extended classification results to obtain multi-elevation-angle hail phase characteristic parameters including a rain-ice mixture category.

[0056] In step S24, the present invention uses a vertical profile analysis method to perform a comprehensive multi-elevation assessment, aiming to perform a three-dimensional reconstruction and statistical analysis of the expanded classification results of the seven elevation layers. During the reconstruction process, a three-dimensional coordinate system is first established, with the radar station as the origin, to establish a range-altitude-azimuth cylindrical coordinate system. Then, an interpolation algorithm is used to spatially interpolate the discrete observation points in the range-altitude plane to generate a continuous phase distribution profile. The interpolation method uses the inverse distance weighted algorithm. Specifically, the statistical characteristic parameters of the hail and rain-ice mixture in each elevation layer are calculated, including six basic parameters: the number of pixels N, the total area S, the average reflectivity Zm, the maximum reflectivity Zmax, the average differential reflectivity ZDRm, and the average cross-correlation coefficient ρHVm. Then, a multi-elevation hail phase characteristic parameter vector is established. The vector dimension is 42, containing a combination of 7 elevation layers × 6 basic parameters. In addition, the characteristic parameter vector also includes three geometric parameters of the vertical structure and two mixed phase statistical parameters (the proportion of rain-ice mixture and the number of mixed areas). Finally, a complete description of the multi-elevation hail phase characteristic parameters including the rain-ice mixture category is formed.

[0057] S3: Based on the multi-elevation-angle hail phase characteristic parameters, the multi-source meteorological data are integrated and preprocessed by a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data of the fused phase characteristics.

[0058] Wherein, step S3 further includes: S31: Grid the ERA5 data and ground observation data to obtain unified grid data.

[0059] Furthermore, in step S31, ERA5 reanalysis data is first read. The original resolution of ERA5 data is a 0.25° × 0.25° latitude and longitude grid. It contains 25 atmospheric parameters, including relative humidity, U wind component, V wind component, temperature, and vertical velocity, at five isobaric surfaces: 1000 hPa, 975 hPa, 950 hPa, 925 hPa, and 900 hPa. A target grid is then established with a resolution of 0.001° × 0.001°. The grid covers Area 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. The ERA5 data is then interpolated using a bilinear interpolation algorithm, while the ground station data is interpolated using the nearest neighbor method. The system calculates the spherical distance between each target grid point and all observation stations and selects the value at the closest station as the interpolation result for that grid point. After interpolation, the 25 parameters of ERA5 and the 18 parameters of ground observations were merged to form a final meteorological element containing 43 elements. The 25 parameters of ERA5 include 5 isobaric surfaces × 5 parameters. The parameters of each layer include relative humidity, U wind component (east-west wind speed), V wind component (north-south wind speed), temperature, and vertical velocity. The ground observation data include 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, station pressure, sea level pressure, 3-hour pressure change, visibility, and precipitation.

[0060] S32: extracting the hailfall area in the unified grid data to obtain a multi-dimensional feature window containing phase characteristics.

[0061] Furthermore, in step S32, the present invention first reads the multi-elevation hail phase characteristic parameters, including the rain-ice mixture category, and extracts the latitude and longitude coordinates of the hail and rain-ice mixture pixels. The hailfall zone is defined as the geographic area where hail or rain-ice mixture phases exist. Specifically, during the extraction process, the present invention utilizes a buffer analysis method, establishing a circular buffer zone with a radius of 5 km centered around each hail pixel. Multiple overlapping buffer zones are then merged into continuous hailfall zone polygons using a Boolean union operation. Then, the corresponding data blocks are extracted from the unified grid data, using the centroid coordinates of each hailfall zone as the window center.

[0062] The final feature window data structure contains attributes such as spatial coordinate information, 43-dimensional meteorological parameter values, and phase characteristic identifiers. The phase characteristic identifier records the phase category distribution of each pixel point 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.

[0063] S33: Based on the positive and negative samples, labels are generated for the multidimensional feature window to obtain a labeled training sample set with phase identification.

[0064] Furthermore, the positive sample is defined as a feature window containing hail or rain-ice mixture phase, and the label value is set to 1; the negative sample is defined as a feature window not containing hail and rain-ice mixture phase, and the label value is set to 0.

[0065] Positive sample extraction is directly based on the characteristic window of the hailfall area. The total number of hail pixels and rain-ice mixture pixels in each window is counted. When the total number is greater than a 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 within 0.1° longitude and latitude in front, behind, left and right of the hail record location is randomly selected as the negative sample center point.

[0066] The present invention further processes the samples. First, the validity of negative samples is verified by checking whether hail or rain-ice mixture pixels exist within the window. When the number of pixels is less than a threshold of two pixels, the window is confirmed as a valid negative sample. Secondly, the samples are balanced by using undersampling technology to calculate the ratio of positive to negative samples. When the number of negative samples exceeds three times the number of positive samples, excess negative samples are randomly deleted to maintain a positive-to-negative sample ratio between 1:2 and 1:3.

[0067] The label generation process is to establish a sample label mapping table to record the complete attributes of each feature window, such as sample number, center coordinates, label value, and phase statistical information contained. The phase identification extension records the distribution of the number of pixels of 11 phase categories in the window to form an 11-dimensional phase feature vector. The vector element value represents the pixel proportion of the corresponding phase in the window. The labeled training sample set with phase identification contains the meteorological data of the feature window, 1-dimensional label value, and 11-dimensional phase feature vector, ultimately forming a complete sample description.

[0068] S34: Standardizing the labeled training sample set to obtain standardized multi-dimensional meteorological feature data fused with phase characteristics.

[0069] Furthermore, the standardization processing in step S34 adopts the Z-score standardization method to eliminate the influence of the differences in dimensions and numerical ranges between different meteorological elements. The numerical range of the standardized phase characteristic vector is limited to between 0 and 1. The phase characteristics are finally integrated into the standardized multidimensional meteorological feature data. The data fusion processing is to splice the standardized 43-dimensional meteorological characteristics with the 11-dimensional phase characteristics to form a 54-dimensional fusion feature vector. The standardized multidimensional meteorological feature data that finally integrates the phase characteristics is a four-dimensional tensor structure. Each sample contains spatially distributed multidimensional feature information and corresponding label values, forming a complete machine learning training data set.

[0070] S4: Using the standardized multi-dimensional 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.

[0071] Wherein, 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.

[0072] Wherein, step S41 further includes: S411: reorganizing the standardized multi-dimensional meteorological characteristic data into tensors according to a preset window structure to obtain a data tensor suitable for a convolution operation.

[0073] Furthermore, the tensor reorganization process converts the standardized multi-dimensional meteorological feature data of the fused phase characteristics in step S3 from the original four-dimensional tensor structure into the input format required by the convolutional neural network. 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). The 79 feature dimensions of each sample are then grouped and arranged according to their physical meaning. The first 43 meteorological parameters are sorted vertically, with the five parameters of the 1000hPa layer first, followed by the 975hPa, 950hPa, 925hPa, and 900hPa layer parameters. Each layer contains five elements: relative humidity, U wind, V wind, temperature, and vertical velocity. The 18 ground-based observation parameters are sorted by element type, with temperature-related parameters first, humidity parameters in the middle, and wind speed and direction parameters at the end. The last 11 phase features are arranged in order of phase numbers 1-11, corresponding to the pixel proportions of graupel, ice crystals, wet snow, dry snow, ground objects, clear sky echoes, large raindrops, hail, heavy rain, light rain, and rain-ice mixtures. The reorganized data tensor meets the data format requirements of the deep learning framework that the number of channels comes first.

[0074] S412: Input the data tensor into a multi-layer convolution kernel to perform feature extraction on the data tensor to obtain a high-order feature map.

[0075] Furthermore, the multi-layer convolution kernel design of the present invention adopts a hierarchical feature extraction architecture, which includes 4 convolution layers, and each layer uses convolution kernels of different sizes and numbers for feature extraction. Specifically: the first convolution layer uses 64 7×7 convolution kernels, the convolution step size is set to 2, the padding method is same padding, the tensor obtained in step S411 is input, and the feature map is output after the convolution operation; the second convolution layer uses 128 5×5 convolution kernels with a step size of 1; the third convolution layer uses 256 3×3 convolution kernels with a step size of 1; the fourth convolution layer uses 512 3×3 convolution kernels with a step size of 2, and each convolution layer is connected to a batch normalization layer and a ReLU activation function. During the feature extraction process, the shallow convolution kernel extracts low-level features such as spatial edges and textures, and the deep convolution kernel extracts high-level semantic features such as the vertical distribution pattern of hail particles and the spatial structure of the meteorological field. Finally, a high-order feature map is output, which contains the feature representation output by each convolution layer, forming a multi-scale, multi-semantic feature pyramid structure.

[0076] S413: Initializing the weights of each convolutional layer corresponding to the high-order feature map layer by layer according to an unsupervised learning criterion to obtain an initial weight configuration.

[0077] Furthermore, in step S413, the present invention uses the principle of autoencoder reconstruction error minimization based on unsupervised learning principles to initialize weights. The specific initialization process is performed layer by layer. First, the first convolutional layer is unsupervised pre-trained, with the input data tensor as the target output. The intrinsic representation of the data is learned through an encoder-decoder structure. The encoder uses 64 7×7 convolution kernels to encode the input into a feature representation, and the decoder uses transposed convolution to reconstruct the feature representation back to the original input size. The second convolutional layer is initialized using the output feature map of the first layer as input. The encoding-decoding process is repeated to learn higher-level feature representations. The encoder uses 128 5×5 convolution kernels, and the decoder uses the corresponding transposed convolution kernels. The initialization process for the third and fourth layers is similar, learning 256-dimensional and 512-dimensional feature representations, respectively. After each layer is trained, the weight parameters of that layer are saved, including the convolution kernel weight matrix, bias vector, and batch normalization parameters. The initial weight configuration contains the complete parameter set of four convolutional layers, with a total number of approximately 3 million parameters.

[0078] S414: Integrate the pre-trained convolutional layer with the initial weight configuration and the fully connected layer to obtain high-order feature representation weights.

[0079] Furthermore, in step S414, the present invention connects the four pre-trained convolutional layers and three fully connected layers in series, aiming to build a complete deep neural network architecture. Specifically, a global average pooling layer is first 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 a 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.

[0080] The final high-order feature representation weights obtained include the spatial feature extraction weights of the convolutional layer and the semantic mapping weights of the fully connected layer, totaling approximately 3.5 million trainable parameters. The integrated network structure is saved in a computational graph format, which includes a complete network description such as the forward propagation path, backpropagation gradient flow, and parameter update rules.

[0081] S42: Through an inter-layer unsupervised learning algorithm, the high-order feature representation weights are optimized layer by layer to obtain a pre-trained DCNN-DBN hybrid network structure.

[0082] Wherein, step S42 further includes: S421: Perform unsupervised training on the first hidden layer weights in the high-order feature representation weights according to the restricted Boltzmann machine principle to obtain first-layer optimized weight parameters.

[0083] Furthermore, the restricted Boltzmann machine (RBM) is a probabilistic graphical model for unsupervised learning, consisting of a visible layer and a hidden layer. Neurons within a layer are unconnected, while neurons between layers are fully connected. The weight optimization of the first hidden layer uses the 512-dimensional feature vector output by global average pooling as the visible layer, with 256 hidden units forming the hidden layer. Unsupervised training of the RBM utilizes the contrastive divergence (CD) algorithm. The training process consists of two phases: forward propagation, which calculates hidden layer activation probabilities based on the visible layer state, and backward propagation, which reconstructs the visible layer based on the hidden layer state. Training proceeds through 100 epochs, with each epoch comprising a complete pass through all training samples. Convergence is determined based on the reconstruction error. Training is terminated when the error remains less than 0.001 over 10 consecutive epochs. The resulting optimized weight parameters for the first layer include a 512×256 weight matrix, a 512-dimensional visible layer bias, and a 256-dimensional hidden layer bias, totaling approximately 130,000 parameters.

[0084] S422: Optimizing the subsequent hidden layers of the first layer's optimized weight parameters in sequence through a greedy layer-by-layer training strategy to obtain a multi-layer weight configuration in which each layer is trained independently.

[0085] Furthermore, in step S422, the present invention trains each hidden layer in sequence in a bottom-up order according to a greedy layer-by-layer training strategy. After each layer of training is completed, the weight of the layer is fixed, and the training of the next layer is continued. The second hidden layer training uses the 256-dimensional output of the first hidden layer as input, and sets 128 hidden units. The training process repeats the contrast divergence algorithm of RBM, but the energy function is adjusted to the first hidden layer state, the second hidden layer state, and the second layer bias. The forward propagation calculates the activation probability of the second layer, and the negative propagation reconstructs the first layer state. The weight update formula remains unchanged, and the learning rate is adjusted to 0.005 to adapt to deeper training. After the second layer training is completed, the 256×128 weight matrix, the 256-dimensional input layer bias, and the 128-dimensional hidden layer bias are saved. The third hidden layer training uses the 128-dimensional output of the second layer as input, and sets 2 output units corresponding to the two-classification task. The training process of the third layer is similar, but the guidance of supervisory information is added, and the classification error term is added to the RBM objective function: in, is the loss of RBM, The balance coefficient is 0.1, The multi-layer weight configuration for each layer trained independently contains the weight parameters of the three-layer RBM, the first layer [512×256], the second layer [256×128], the third layer [128×2], and the corresponding bias parameters, totaling about 180,000 parameters.

[0086] S423: According to a reconstruction error minimization criterion, globally adjust the multi-layer weight configuration to obtain a global weight distribution.

[0087] Furthermore, in step S423, the reconstruction error minimization criterion performs end-to-end global optimization of the entire network using 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 symmetrical encoder-decoder network structure. The encoder path performs feature compression in a dimensionality reduction order of 512 → 256 → 128 → 2, while the decoder path performs feature reconstruction in a dimensionality increase order of 2 → 128 → 256 → 512. The Adam optimizer is used for global optimization, with the learning rate set to 0.0001, the first momentum parameter set to 0.9, and the second momentum parameter set to 0.999. Batch gradient descent is used for the optimization process, and the batch size is set to 32 samples. The gradient calculation is performed through the back propagation algorithm. The reconstruction error on the validation set is monitored during the training process. When the error does not decrease after 20 consecutive epochs, the early stopping strategy is used to terminate the training. Then, a global adjustment is made according to the set parameters, that is, all weight parameters are jointly optimized to eliminate the accumulated error in layer-by-layer training. The global weight distribution maintains the original three-layer structure, but the weight values ​​are adjusted through global optimization, and the numerical distribution of the weight matrix of each layer is more uniform, avoiding the problem of gradient vanishing or exploding.

[0088] S424: Performing network integration on the global weight distribution to obtain a pre-trained DCNN-DBN hybrid network structure.

[0089] 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 series 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 into probabilistic outputs for hail recognition.

[0090] During the integration process, the interface compatibility of the two networks is first verified. The global average pooling layer of the DCNN outputs a 512-dimensional feature vector, matching the input dimension of the first layer of the DBN. Using a network integration algorithm, the present invention sequentially connects the four convolutional layers and one pooling layer of the DCNN with the three fully connected layers of the DBN to form an end-to-end deep network. The weight parameters of each layer remain unchanged during the connection process, and only the data flow and gradient propagation paths 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.

[0091] The resulting pre-trained DCNN-DBN hybrid network structure possesses powerful feature representation capabilities. The DCNN component automatically extracts the spatial distribution pattern of the hail system, while the DBN component establishes a nonlinear mapping relationship between features and hail occurrence probability. The resulting 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 foundation for subsequent supervised fine-tuning training.

[0092] S43: Using the BP algorithm, the DCNN-DBN hybrid network structure is fine-tuned to obtain a ground hail recognition model.

[0093] Furthermore, in step S43, the present invention uses a back propagation (BP) algorithm to perform end-to-end supervised learning on the pre-trained DCNN-DBN hybrid network through supervised fine-tuning training, aiming to ultimately use a labeled training sample set with phase identification to convert the hail recognition task into a binary classification problem and output the calculated probability distribution of hail occurrence.

[0094] The BP algorithm calculates the gradient of the loss function with respect to the weights of each layer through the chain rule. The gradient calculation is propagated layer by layer from the output layer to the input layer. The stochastic gradient descent optimizer is used for fine-tuning training. The learning rate is set to 0.0001 and the momentum coefficient is 0.9. Data enhancement technology is used in the training process, including random rotation, flipping, scaling and other operations to increase sample diversity. The batch size is set to 16 samples and the number of training rounds is set to 100 epochs. The verification strategy adopts 5-fold cross-validation, and the training data is divided into training set and verification set in a ratio of 8:2. In addition, the present invention also introduces an early stopping mechanism to monitor the accuracy on the verification set. Training is stopped when the accuracy does not improve for 15 consecutive epochs. The model performance evaluation adopts indicators such as accuracy, precision, recall rate, F1 score, and the prediction performance of the model is analyzed through the confusion matrix.

[0095] Finally, after fine-tuning is completed, 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 falling areas.

[0096] S5: Outputting a hailfall area recognition result through the ground hail recognition model.

[0097] Furthermore, the final recognition result of step S5 is the spatial location information of the hailfall area, specifically including: spatial location data: The hailfall area is identified using a 50×50 pixel window, and the output is whether hail exists within a specific latitude and longitude coordinate range, equivalent to making a point-by-point judgment on a fine 0.001°×0.001° grid; binary classification results: For each spatial window, a label of 0 or 1 is output, where 1 indicates hail has occurred in the area and 0 indicates no hail has occurred; probability information: Because the recognition model obtained in step S4 of the present invention uses a softmax output layer based on a BP neural network, the final model outputs a probability value (between 0 and 1) for hail occurrence in each area, facilitating the setting of different confidence thresholds. Furthermore, based on the aforementioned hailfall area recognition results, the present invention also provides a planar distribution map of the hailfall area, showing which areas are likely to produce hail and which are not, forming a continuous spatial distribution pattern.

[0098] For example, if severe convective weather occurs in a certain area and the radar detects an echo with a reflectivity exceeding 50dBZ, the algorithm of the present invention first extracts valid observation data within 200 kilometers of the radar station, and establishes a time window from 18 minutes before to 12 minutes after the reference time. The dual-polarization parameters show Zh=52dBZ, ZDR=0.3dB, and ρHV=0.87, and are identified as hail phase through a dynamically optimized membership function. Combined with ERA5 reanalysis data, the 0°C layer height is determined to be 3200 meters, and the echo is located at an altitude of 4000 meters, belonging to the ice phase area. 43 meteorological elements such as ground temperature of 28°C, relative humidity of 65%, and wind speed of 12m / s are integrated to form a 79-dimensional feature vector, which is input into the DCNN-DBN model. The model outputs a probability of hail occurrence in the area of ​​0.85, which exceeds the 0.5 threshold and is ultimately identified as a hailfall area. A binary classification result indicating the presence of hail within the latitude and longitude of the area is output.

[0099] like Figure 2 As shown, the present invention also provides a ground hail identification system based on hydrometeor classification results, comprising: Correlation module 100: used to perform spatiotemporal correlation processing on multi-source hail data using a spatiotemporal matching algorithm to obtain a spatiotemporal matching hail event dataset; Identification module 200: configured to perform adaptive phase identification 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 including a rain-ice mixture category; Preprocessing module 300: for performing integrated preprocessing on multi-source meteorological data through a data fusion algorithm based on the multi-elevation-angle hail phase characteristic parameters to obtain standardized multi-dimensional meteorological characteristic data of fused phase characteristics; Training module 400: used to perform unsupervised pre-training and supervised fine-tuning training on the DCNN-DBN hybrid neural network using the standardized multi-dimensional meteorological feature data to obtain a ground hail recognition model; The recognition module 500 is configured as the ground hail recognition model obtained through training by the training module 400, and is used to output a hailfall area recognition result.

[0100] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0101] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software and a necessary general-purpose hardware platform, or alternatively, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes instructions for enabling a computer device (such as a personal computer, server, or network device) to execute the ground hail identification method based on hydrometeor classification results, as described in various embodiments or portions thereof.

[0102] The present invention realizes the precise identification and positioning of hail on the ground, significantly improves the accuracy and timeliness of hail warning, makes phase identification more refined, effectively solves the problem that hail is easily confused with other hydrometeors, and significantly improves the accuracy of phase discrimination, especially in complex weather conditions. It has stronger recognition ability, laying a solid foundation for the accurate identification of hail events. The overall method provides important technical support for monitoring, early warning, prevention and mitigation of hail disasters.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for identifying ground hail based on hydrometeor classification results, characterized in that: include: S1: Using the space-time matching algorithm, the multi-source hail data is subjected to space-time correlation processing to obtain a space-time matching hail event dataset; S2: Adaptively identifying the phase state of 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 including the rain-ice mixture category; S3: Based on the multi-elevation-angle hail phase characteristic parameters, integrating and preprocessing the multi-source meteorological data through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data of the fused phase characteristics; S4: performing unsupervised pre-training and supervised fine-tuning training on the DCNN-DBN hybrid neural network using the standardized multi-dimensional meteorological feature data to obtain a ground hail recognition model; S5: Outputting a hailfall area recognition result through the ground hail recognition model.

2. The method for identifying ground hail based on hydrometeor classification results according to claim 1, characterized in that: Step S1 further comprises: S11: spatially screen the radar observation area according to the preset weather station location to obtain the effective detection area; S12: Using a dynamic time window strategy, time matching is performed on hail events within the effective detection area to obtain a time matching window based on the volume scanning time; S13: performing spatial correlation processing on the actual position of the hailstone within the time matching window to obtain a radar echo matching result within a preset distance search radius; S14: Perform multi-source fusion on the radar echo matching results and public data to obtain a time-space matched hail event dataset.

3. The method for identifying ground hail based on hydrometeor classification results according to claim 1, characterized in that: Step S2 further comprises: S21: adjusting the membership function threshold of the polarization parameter in the Park dual-polarization phase state classification algorithm in real time according to the A test statistic of the hail event dataset to obtain a dynamically optimized phase state discrimination threshold; wherein the polarization parameter includes a horizontal reflectivity factor and a differential reflectivity; S22: performing adaptive correction on the weight matrix of the polarization parameter error to obtain an optimized dual polarization parameter weight configuration; S23: performing rain-ice mixture identification on the region above the melting layer under the dual polarization parameter weight configuration according to the mutual correlation coefficient, to obtain an extended classification result including a mixed phase state; S24: Performing a multi-elevation-angle comprehensive evaluation on the extended classification results to obtain multi-elevation-angle hail phase characteristic parameters including a rain-ice mixture category.

4. The method for identifying ground hail based on hydrometeor classification results according to claim 3, characterized in that: Step S21 further includes: S211: Calculate the A test statistic of the hail event data set to obtain a statistical test index of the radar data distribution; S212: dynamically adjusting the membership function thresholds of the reflectivity factor and the differential reflectivity in the Park algorithm according to the statistical test index to obtain an adaptive phase state discrimination parameter; S213: reclassifying the easily confused phases under the phase discrimination parameters by real-time parameter updating to obtain a phase recognition result with enhanced discrimination; the easily confused phases include ice crystals and dry snow; S214: Apply the phase state recognition result to the entire observation period, perform batch processing, and obtain a dynamically optimized phase state discrimination threshold.

5. The method for identifying ground hail based on hydrometeor classification results according to claim 3, characterized in that: Step S23 further includes: S231: Using the cross-correlation coefficient threshold to determine the mixed phase region of radar observation data under the optimized dual-polarization parameter weight configuration, the potential distribution data of the rain-ice mixed region is obtained; S232: Determine the position of the melting layer in the rain-ice mixed zone corresponding to the rain-ice mixed zone distribution data based on the height information, and obtain the range of the ice phase area above the 0°C layer height; S233: Performing spatial superposition analysis on the ice phase region and the mixed phase region to obtain position information of the rain-ice mixture; S234: Performing phase classification expansion on the rain-ice mixture position information through mixture category labeling to obtain an expanded classification result including the mixture phase.

6. The method for identifying ground hail based on hydrometeor classification results according to claim 1, characterized in that: Step S3 further comprises: S31: Gridding the ERA5 data and ground observation data to obtain unified grid data; S32: extracting the hailfall area in the unified grid data to obtain a multidimensional feature window containing phase characteristics; S33: Based on the positive and negative samples, labels are generated for the multidimensional feature window to obtain a labeled training sample set with phase identification; S34: Standardizing the labeled training sample set to obtain standardized multi-dimensional meteorological feature data fused with phase characteristics.

7. The method for identifying ground hail based on hydrometeor classification results according to claim 1, characterized in that: Step S4 further comprises: S41: performing multi-layer convolution on the standardized multi-dimensional meteorological feature data to pre-train a deep belief network structure to obtain high-order feature representation weights; S42: Optimizing the high-order feature representation weights layer by layer through an inter-layer unsupervised learning algorithm to obtain a pre-trained DCNN-DBN hybrid network structure; S43: Using the BP algorithm, the DCNN-DBN hybrid network structure is fine-tuned to obtain a ground hail recognition model.

8. The method for identifying ground hail based on hydrometeor classification results according to claim 7, characterized in that: Step S41 further includes: S411: reorganizing the standardized multi-dimensional meteorological characteristic data into tensors according to a preset window structure to obtain a data tensor suitable for a convolution operation; S412: Inputting the data tensor into a multi-layer convolution kernel to perform feature extraction on the data tensor to obtain a high-order feature map; S413: Initializing the weights of each convolutional layer corresponding to the high-order feature map layer by layer according to an unsupervised learning criterion to obtain an initial weight configuration; S414: Integrate the pre-trained convolutional layer with the initial weight configuration and the fully connected layer to obtain high-order feature representation weights.

9. The method for identifying ground hail based on hydrometeor classification results according to claim 7, characterized in that: Step S42 further includes: S421: performing unsupervised training on the first hidden layer weights in the high-order feature representation weights according to the restricted Boltzmann machine principle to obtain first-layer optimized weight parameters; S422: Optimizing the subsequent hidden layers of the first layer's optimized weight parameters in sequence through a greedy layer-by-layer training strategy to obtain a multi-layer weight configuration in which each layer is independently trained; S423: Globally adjust the multi-layer weight configuration according to a reconstruction error minimization criterion to obtain a global weight distribution; S424: Performing network integration on the global weight distribution to obtain a pre-trained DCNN-DBN hybrid network structure.

10. A ground hail identification system based on hydrometeor classification results, characterized in that: include: Correlation module: used to perform spatiotemporal correlation processing on multi-source hail data through spatiotemporal matching algorithm to obtain a spatiotemporal matching hail event dataset; Identification module: used to perform adaptive phase identification 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 including rain-ice mixture categories; Preprocessing module: used for integrating and preprocessing multi-source meteorological data based on the multi-elevation angle hail phase characteristic parameters through a data fusion algorithm to obtain standardized multi-dimensional meteorological characteristic data of fused phase characteristics; Training module: used to perform unsupervised pre-training and supervised fine-tuning training on the DCNN-DBN hybrid neural network using the standardized multi-dimensional meteorological feature data to obtain a ground hail recognition model; The recognition module is configured as the ground hail recognition model obtained through training by the training module, and is used to output a hailfall area recognition result.

Images