Hailstorm warning method under complex terrain action

Abstract

The application discloses a hailstorm early warning method under complex terrain action and relates to the technical field of weather forecasting. The method constructs an integrated probability model comprising a feature screening module, multiple logistic regression sub-models, an integrated layer and a physical gating mechanism layer. Historical hailstorm process samples are acquired and labeled, terrain, dynamic and cloud microphysical basic features are extracted, and a candidate feature set is constructed. The integrated probability model and early warning release parameters are trained and optimized. In the real-time monitoring stage, multi-source observation data are acquired, the tracking state of hail cloud objects is judged, early warning signs are generated based on model output probability, physical gating results and state transition rules. After the process ends, the real-time hailfall results are associated and new labeled samples are generated. The method integrates complex terrain disturbance, updraft evolution and cloud microphysical processes into a unified early warning framework, improving the accuracy, stability and interpretability of short-term and nowcasting early warning of hailstorms in complex terrain areas.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting, and in particular to a method for early warning of hailstorms under complex terrain. Background Technology

[0002] Hailstorms often exhibit stronger heterogeneity and suddenness in complex terrain areas such as mountains and plateaus. Orographic uplift, flow splitting, and valley wind circulation can modulate convection triggering and updrafts, thereby altering the thermal environment and supercooled liquid water budget of the 0-20℃ mixed phase layer, and regulating hail embryo evolution and hail precipitation. For X-band radar observations, improvements such as introducing phase state vertical continuity constraints have significantly enhanced the fine-scale inversion accuracy of graupel and hail particle distribution in the core area of ​​hailstorm cells. Research reveals that complex terrain, through interaction with boundary layer processes, can excite mesoscale vortices or persistent convergence lines under specific circulation backgrounds, providing key dynamic conditions for the generation and development of hailstorms. Case studies in piedmont areas show that topographically triggered mesoscale uplift and inflow channels can reshape storm echo structures and hail core maintenance mechanisms, highlighting the constraining effect of topography on convection organization and hail precipitation areas. Observations in plateau and mountainous areas have revealed the uniqueness of hail particle spectra and particle growth patterns. In summary, complex terrain not only dynamically regulates the overall evolution of hailstorms, but its modification of thermal stratification and water vapor transport is also a crucial factor influencing hail formation and growth. Existing publicly available hailstorm forecasting and early warning technologies generally use hailstorm characteristic parameters as core inputs, failing to fully incorporate the regulatory effects of terrain factors on hailstorms, resulting in limitations in technology adaptability. Results obtained from pure machine learning models lack clear physical mechanism support, making it difficult to explain the rationality of the early warning logic; while early warning methods based on single physical criteria are influenced by subjective experience, lacking sufficient quantification and failing to meet the needs of accurate early warning. All of these factors significantly reduce the scientific validity and credibility of early warning decisions, making it impossible to effectively support the precise prevention and control of hailstorms in complex terrain. Summary of the Invention

[0003] The purpose of this invention is to design a hailstorm warning method for complex terrain conditions to solve the aforementioned problems. This invention achieves the above objective through the following technical solution:

[0004] Hailstorm warning methods under complex terrain include:

[0005] S1. Construct the initial ensemble probability model, which includes a feature selection module, m logistic regression sub-models, an ensemble layer, and a physical gating mechanism layer. The output of the feature selection module serves as the input to each of the m logistic regression sub-models, and the outputs of the m logistic regression sub-models all serve as the input to the ensemble layer. The outputs of the ensemble layer and the feature selection module both serve as the input to the physical gating mechanism layer, where m is a positive integer greater than 2. Each logistic regression sub-model is trained using the liblinear solver and L2 regularization. The parameter settings for the m logistic regression sub-models are different.

[0006] S2. Obtain historical hailstorm process samples and label multiple historical hailstorm process samples with warning tags. Then, divide the labeled historical hailstorm process samples into a training sample set and a validation sample set. Each hailstorm process sample includes multi-source observation data at multiple radar volume scan times within the hailstorm process. The multi-source observation data includes X-band dual-polarization radar base data, 30m resolution ASTER GDEM terrain elevation data, cloud microphysical feature data, and dynamic feature data. The X-band dual-polarization radar base data includes the reflectivity factor Z. H Differential reflectivity Z DR , differential phase K DP Radial velocity Vr and correlation coefficient ρ hv ;

[0007] S3. Extract the basic features corresponding to each radar volume scan time based on multi-source observation data. The basic features include topographic features, dynamic features, and cloud microphysical features. Topographic features include the average elevation change ΔH of the core area with a reflectivity greater than 40 dBZ of hail cloud combination and the topographic slope S on the path of hail cloud movement. Dynamic features include the 90th percentile value UpP90 of the positive vertical updraft between 0℃ and −20℃ temperature layers, the number of radar arrays NV0-0.5 for vertical updrafts between 0℃ and −20℃ temperature layers with a speed of 0 to 0.5 m / s, and the number of radar arrays NV0.5-1 for vertical updrafts between 0℃ and −20℃ temperature layers with a speed of 0.5 to 1 m / s. Cloud microphysical features include the number of radar arrays NSWA for supercooled water, NRH for hail, NCR for ice crystals, NAG for ice crystal polymers, NLDG for low-density graupel, and NHDG for high-density graupel.

[0008] S4. Construct a candidate feature set based on the basic features. When entering S3 from S2 or S10, proceed to S5; when returning to S3 from S7, proceed to S8.

[0009] S5. Use the candidate feature set to train and optimize the current ensemble probability model to obtain the optimized ensemble probability model and the corresponding early warning release parameters.

[0010] S6. Acquire multi-source observation data at multiple radar volume scan times during real-time monitoring;

[0011] S7. Determine whether the current hail cloud object enters monitoring and tracking, maintains monitoring and tracking, or stops tracking and issues an early warning based on the tracking start and stop conditions; if it is determined that the current hail cloud object enters monitoring and tracking or maintains monitoring and tracking, proceed to S3; if it is determined that the current hail cloud object stops tracking and issues an early warning, proceed to S10.

[0012] S8. Analyze the candidate feature set using the optimized ensemble probability model to obtain the early warning judgment parameters;

[0013] S9. Analyze the early warning judgment parameters to obtain the analysis results. The analysis results are either an early warning state or no early warning. If the analysis results are an early warning state and an early warning flag is generated, return to S6. If the analysis results are no early warning, return directly to S6.

[0014] S10. After the current hailstorm process ends, based on the current hailstorm process and the actual ground hail situation, the object tracking results stored in the real-time warning process are associated with the actual hail period to generate new labeled samples as training samples, and then returned to S3.

[0015] The beneficial effects of this invention are as follows: (1) By integrating terrain disturbance, slope disturbance, microphysical activity, graupel ratio and updraft information into the feature system, the hail formation evolution process under complex terrain can be more accurately characterized; (2) The probability results are first output by the integrated probability model, and then the physical gating mechanism layer and state machine are used for time-series constraints and physical consistency verification, which not only retains the recognition ability of machine learning, but also enhances the interpretability and reliability of the results; (3) By re-labeling and retraining the newly added samples after the process is completed, the warning release parameters can be updated with the accumulation of samples, which is suitable for operational deployment and continuous optimization. Attached Figure Description

[0016] Figure 1 This is a flowchart of the hail storm warning method under complex terrain conditions according to the present invention;

[0017] Figure 2 This is a hail storm movement path map provided in an embodiment of the present invention, wherein hail occurred between 10:24 and 10:35, and (a1)-(a4) are superimposed with combined reflectance maps at times greater than 40 dBZ at 10:21, 10:27, 10:35 and 10:43 respectively;

[0018] Figure 3The following are the reflectivity profiles (a1-a5) at 10:21, 10:24, 10:27, 10:35, and 10:43 provided in the embodiment; radial velocity profiles (b1-b5); particle identification profiles (c1-c5); topographic lift height difference and slope value (d) in the movement path of the hail cloud above 40 dBZ; with an average altitude H0 of 1208.6 m.

[0019] Figure 4 This is a time series of the number of hydromorphic particles and updraft parameters retrieved by radar during hailfall provided in this embodiment of the invention; (a) the number of hydromorphic particles and the difference in topographic lift height ΔH; (b) the rate of change of the number of hydromorphic particles and the difference in topographic lift height ΔH; (c) the number of updrafts in the 0~−20℃ layer and the difference in topographic lift height ΔH; the red boxes mark the main periods of enhanced topographic-related updrafts. Detailed Implementation

[0020] like Figure 1 As shown, the hail storm warning method under complex terrain includes:

[0021] S1. Construct the initial ensemble probability model, which includes a feature selection module, m logistic regression sub-models, an ensemble layer, and a physical gating mechanism layer. The output of the feature selection module serves as the input to each of the m logistic regression sub-models, and the outputs of the m logistic regression sub-models all serve as the input to the ensemble layer. The outputs of the ensemble layer and the feature selection module both serve as the input to the physical gating mechanism layer, where m is a positive integer greater than 2. Each logistic regression sub-model is trained using the liblinear solver and L2 regularization. The parameter settings for the m logistic regression sub-models are different.

[0022] S2. Obtain historical hailstorm process samples and label multiple historical hailstorm process samples with warning tags. Then, divide the labeled historical hailstorm process samples into a training sample set and a validation sample set. The validation sample set is an independent test set independent of model training and candidate parameter search, and is only used to evaluate the final effect of the ensemble probability model after training and optimization. The object extrapolation probability is the grouped extrapolation prediction probability generated within the training sample set using leave-one-out validation for each hailstorm process. Each hailstorm process sample includes multi-source observation data at multiple radar volume scan times within the hailstorm process. The multi-source observation data includes X-band dual-polarization radar base data, 30m resolution ASTER GDEM terrain elevation data, cloud microphysical feature data, and dynamic feature data. The X-band dual-polarization radar base data includes reflectivity factor Z. H Differential reflectivity Z DR , differential phase K DP Radial velocity Vr and correlation coefficient ρ hv ;

[0023] S3. Based on multi-source observation data, extract the basic features corresponding to each radar volume scan time. These basic features include topographic features, dynamic features, and cloud microphysical features. Topographic features include the average elevation change ΔH of the core area with a hail cloud reflectivity > 40 dBZ and the topographic slope S along the hail cloud's movement path. Dynamic features include the 90th percentile of the positive vertical updraft between 0℃ and −20℃ temperatures (UpP90), the number of radar arrays for vertical updrafts between 0℃ and −20℃ temperatures (0–0.5 m / s, NV0-0.5), and the number of radar arrays for vertical updrafts between 0℃ and −20℃ temperatures (0.5–1 m / s, NV0.5-1). Cloud microphysical features include the number of radar arrays for supercooled water (NSWA), hail (NRH), ice crystals (NCR), ice crystal polymers (NAG), low-density graupel (NLDG), and high-density graupel (NHDG). The cloud microphysical features are determined by the quality-controlled reflectivity factor Z. H Differential reflectivity Z DR , differential phase K DP Correlation coefficient ρ HV Using ambient temperature T as input, the types of water condensate particles in each radar array are identified through an improved fuzzy logic algorithm, and then statistically obtained. The improved fuzzy logic algorithm includes the following steps:

[0024] S301, Select Z after quality control H Z DR K DP ρ HV Using ambient temperature T as input parameters for phase state recognition of water-soluble particles, a water-soluble particle recognition model is constructed.

[0025] S302, set Z for drizzle DR, rain RA, ice crystal polymer AG, ice crystal CR, low-density graupel LDG, high-density graupel HDG, and hail RH. H Z DR K DP ρ HV The membership function weight coefficients and membership function thresholds corresponding to T are defined. The membership functions of each parameter adopt trapezoidal or semi-trapezoidal membership functions determined by X1, X2, X3, and X4. X1 and X4 are used to limit the effective value range of the parameter for the corresponding particle type, and X2 and X3 are used to limit the high membership value range of the parameter for the corresponding particle type. The weight coefficients and membership function thresholds are optimized locally based on historical hail storm observation data in the study area to improve the applicability of the water condensate particle identification model in hail storm processes in complex terrain.

[0026] S303. Calculate the comprehensive membership score of any radar library under different water-based particle types, as follows: ;Sc is the comprehensive membership score of radar-collected particles belonging to class c hydrogel particles; c is DR, RA, AG, CR, LDG, HDG or RH; , , , and Z H Z DR K DP ρ HV The weighting coefficients of T; , , , and Z H Z DR K DP ρ HV The membership degree between T and T. The overall membership degree score is expressed as: Due to hail , It does not include a temperature membership term; the membership functions are represented using trapezoidal membership functions. ; X is a radar parameter or ambient temperature parameter used for phase state identification of water-soluble particles. k,x,c ( ) represents the threshold of the k-th member function of the c-th type of hydrogel particles under parameter x; X 1,x,c and X 4,x,c The valid range of values ​​for parameter x is limited. 2,x,c and X 3,x,c Limit the range of values ​​for the high membership degree of parameter x. 1,x,c =X 2,x,c or X 3,x,c =X 4,x,c When this happens, the corresponding linear ascending or descending segment is omitted, and the membership function degenerates into a semi-trapezoidal or rectangular membership function;

[0027] S304. For the radar database set corresponding to the hail cloud object at the current volume scan time t. For any radar library q within the range, the type of water-soluble particles is determined based on the comprehensive membership score. And extract cloud microphysical features, , ; The comprehensive membership score of the q-th radar column belonging to the c-th type of water condensate particle at the current volume scan time t. Let T(q,t)<0 be the set of water-condensed particle types, where T(q,t)<0 means that the q-th radar cell is located above the 0℃ isotherm.

[0028] S4. Construct a candidate feature set based on the basic features. When entering S3 from S2 or S10, proceed to S5; when returning to S3 from S7, proceed to S8. Specifically, this includes:

[0029] S4001. Construct core features based on basic features. Core features include the comprehensive quantity of ice phase particles at the t-th radar volume scan time. relative order of magnitude compressibility of ice Comprehensive microphysical activity index The proportion of graupel Terrain disturbance and slope disturbance ; , , , , , j is the sequence number of the radar volume scan time within the time window; The time window consists of the t-th radar volume scan time and all radar volume scan times within the preceding 6 minutes. For time window The number of radar volume scan samples included; The average elevation change of the core area where the reflectivity of the hail cloud combination is >40 dBZ at the j-th radar volume scan time. The radar volume scan times obtained from the start of the current hailstorm process to the current radar volume scan time t are as follows: The average value; ΔS(j) = S(j) - S(j-1) is the slope change value on the hail cloud's moving path at the j-th radar volume scan time; for the first effective radar volume scan time of the current hail cloud object, since there is no previous effective radar volume scan time, let the slope change value at that time be... =0; The data obtained from the start of the current hailstorm process to the current radar volume scan time t are as follows. The average value; To prevent extremely small positive numbers with a denominator of zero;

[0030] S4002. Construct the ice facies enhancement trend based on the actual time interval between adjacent effective radar volume scan times. and the trend of graupel enhancement Let W3(t) represent the time window consisting of the t-th radar volume scan time and all valid radar volume scan times within the preceding 3 minutes; let... Let W3(t) contain the set of time indices corresponding to the previous valid radar volume scan time. Let represent the number of effective adjacent difference terms, then: ; ,in, This represents the actual time interval, in seconds, between the (j-1)th and j-th effective radar volume scan times; when When = 0, it means there are no valid adjacent difference terms within the time window. Let =0 and =0; when When the value is greater than 0, the calculation is performed according to the above effective adjacent difference terms. This process avoids referencing a previous time that does not exist when the first effective radar volume scan time or an intermediate volume scan is missing.

[0031] S4003. Construct coupling characteristics, including characterizing the degree of synergistic enhancement between strong updrafts in the mixed phase layer and ice particle growth. The coupling relationship between the response intensity and the characterization of microphysical activity under terrain disturbance background. , , ;

[0032] S4004. Constructing historical window sorting features for core features and coupled features. Robust anomaly characteristics of historical windows Historical window maximum value characteristics Historical window average characteristics and short-term variation characteristics , Indicates core or coupled features; short-term changing features The feature value at the current time and its value at... The difference between the feature values ​​corresponding to the start time of the time window;

[0033] S4005. Constructing lag features based on the aforementioned core features, coupled features, and historical statistical features: , The preset lag duration is used; the lag feature represents the preset duration of the aforementioned features before the current volume scan. The state at the location; the preset lag time is 0 min, 3 min, 6 min, 9 min or 12 min, preferably 6 min; the short-term change time window is preferably 3 min or 6 min;

[0034] S4006. Construct key features related to hail formation processes under complex terrain conditions as fixed features. These fixed features include: the maximum value of the 90th percentile of the updraft UpP90 within a continuous 12-minute time window prior to the current radar volume scan time t (UpP90_max12(t)); and the ice particle intensity index within a continuous 12-minute time window prior to the current radar volume scan time t. The maximum value of ice_max12(t), the maximum value of micro_log of the comprehensive microphysical activity index micro_log within the 12-minute time window before the current radar volume scan time t, the maximum value of graffrac of the graupel proportion index gra_max12(t) within the 12-minute time window before the current radar volume scan time t, the cooperative characteristic value of updraft and ice phase particles UpIce_L0(t) at the current radar volume scan time t, the cooperative characteristic value of updraft and ice phase particles UpIce_L3(t) at the current radar volume scan time t with a lag of 3 minutes, the cooperative characteristic value of terrain disturbance and microphysical activity DHxMicro_L0(t) at the current radar volume scan time t, and the cooperative characteristic value of terrain disturbance and microphysical activity at the current radar volume scan time t with a lag of 3 minutes. The following parameters are considered: the degree of coordination characteristic value DHxMicro_L3(t), the percentile value of the comprehensive microphysical activity index micro_log(t) at the current radar volume scan time t within the preset historical time window, the percentile value of the graupel proportion index graFrac(t) at the current radar volume scan time t within the preset historical time window, the larger value of the historical window percentile values ​​of the ice phase enhancement trend dIce(t) and the graupel enhancement trend dGra(t) at the current radar volume scan time t, and the percentile value of the updraft index UpP90(t) at the current radar volume scan time t within the preset historical time window; among them, ice_max12(t) is the largest value of the historical window percentile values ​​of the ice phase enhancement trend dIce(t) and the graupel enhancement trend dGra(t) at the current radar volume scan time t. For the calculation of the basic quantities, micro_max12(t) and micro_pr_L0(t) are both calculated based on micro_log(t), gra_max12(t) and gra_pr_L0(t) are both calculated based on graFrac(t), trend_pr_L0(t) is determined by the larger of the historical window sorting percentile values ​​of dIce(t) and dGra(t), and UpP90_pr_L0(t) is calculated based on UpP90(t).

[0035] S4007. Construct dynamic statistical features for the coupling features. These dynamic statistical features include the maximum value of UpIce (UpIce_max6(t)) and the average value of UpIce (UpIce_mean6(t)) within a continuous 6-minute time window preceding the current volumetric scan time t; the first-order linear slope of UpIce relative to time (UpIce_slope3(t)) within a continuous 3-minute time window preceding the current volumetric scan time t; the second-order change of UpIce at the current volumetric scan time t (UpIce_d2(t)); the difference between the UpIce value at the current volumetric scan time t and the UpIce value at the previous 3-minute time window (UpIce_rise3(t)); and the values ​​within the continuous 6-minute time window preceding the current volumetric scan time t. The maximum value of DHxMicro within the min time window is DHxMicro_max6(t), the average value of DHxMicro within the 6 consecutive min time windows before the current volume scan time t is DHxMicro_mean6(t), the slope of the first-order linear change of DHxMicro relative to time within the 3 consecutive min time windows before the current volume scan time t is DHxMicro_slope3(t), the second-order change of DHxMicro at the current volume scan time t is DHxMicro_d2(t), and the difference between the DHxMicro value at the current volume scan time t and the DHxMicro value at the previous 3 min time windows is DHxMicro_rise3(t).

[0036] S4008. Combine fixed features, automatically selected lag class features, and dynamic statistical features of coupled features to form a candidate feature set;

[0037] S4009. When entering S3 from S2 or S10, enter S5; when returning to S3 from S7, enter S8.

[0038] S5. Train and optimize the current ensemble probability model using the candidate feature set to obtain the optimized ensemble probability model and the corresponding early warning release parameters; specifically including:

[0039] S5001. Construct a training sample matrix X, a label vector y, and a process grouping vector g based on the candidate feature set. Set sample weights for the training samples: the process grouping vector g must satisfy g... i Samples with weight = q, belonging to the q-th hailstorm process, are assigned basic weights. The unnormalized sample weights of the i-th training sample in the r-th logistic regression sub-model. Normalized sample weights are obtained by normalizing the unnormalized sample weights. Let be the number of training samples for the q-th hailstorm process. and Let a be the number of positive samples and the number of negative samples in the q-th hailstorm process. rLet be the positive sample augmentation coefficient corresponding to the r-th logistic regression sub-model, where N is the total number of training samples. The training sample index is denoted as r, and the logistic regression sub-model index is denoted as r = 1, 2, ..., m. When both positive and negative samples exist in the q-th hailstorm process, class imbalance enhancement is applied to the positive samples. When neither positive nor negative samples exist, the ratio of positive to negative samples is not calculated, and the basic weights are used directly.

[0040] S5002, the feature selection module uses the ExtraTreesClassifier to rank candidate features by importance, selects optional features based on the importance ranking results, and retains at most one representative feature for each type of dynamic statistical feature to avoid feature redundancy and improve model stability; the ExtraTreesClassifier evaluates the contribution of each candidate feature to the early warning label discrimination by constructing multiple random decision trees.

[0041] S5003. Merge the optional features with the fixed features, and remove duplicates according to the feature name to obtain the final input feature set F;

[0042] S5004. For the r-th logistic regression sub-model, the weighted logistic regression objective function is optimized using the liblinear solver and L2 regularization, and is expressed as: ,in, Let be the final input feature vector corresponding to the i-th training sample; To classify the binary labels y i The values ​​obtained after mapping to {−1,1} , and Let the intercept term and regression coefficient of the r-th logistic regression sub-model be represented respectively. Let represent the L2 regularization strength corresponding to the r-th logistic regression sub-model, and =1 / C r C r Let C1 be the regularization parameter for the r-th logistic regression sub-model; r = 1, 2, ..., m. The three logistic regression sub-models each have different regularization parameters, class weight settings, and positive sample enhancement coefficients. For the first sub-model, the regularization parameter C1 is 0.08, the class weights are balanced, and the positive sample enhancement coefficient a1 is 1.2. For the second sub-model, the regularization parameter C2 is 0.20, the class weights are balanced, and the positive sample enhancement coefficient a2 is 1.0. For the third sub-model, the regularization parameter C3 is 0.50, the class weights are not balanced, and the positive sample enhancement coefficient a3 is 1.0. When m = 3, the first to third sub-models use the above parameter settings.

[0043] S5005. During training, a leave-one-out validation method is adopted, grouped by hailstorm process. In each round of leave-one-out validation, samples from one hailstorm process are used as the leave-out subset, and samples from the remaining hailstorm processes are used as the in-process training subset, so that samples from the same hailstorm process do not enter the in-process training subset and the leave-out subset at the same time. The samples in the leave-out subset of each round are predicted by the ensemble probability model obtained from the corresponding in-process training, and the predicted probabilities are spliced ​​together according to the order of hailstorm process and radar volume scan time to form the object extrapolation probability.

[0044] S5006, Parameter Search and Early Warning Release Parameters: Select the Optimal Candidate Parameter Group And together with the logistic regression ensemble model, it constitutes the optimized ensemble probability model; specifically including:

[0045] (1) The kth candidate parameter group Two types of binary warning sequences are generated by applying the extrapolation probability and the in-training set probability to the object respectively; the extrapolation verification warning sequence is then obtained. and early warning sequences within the training set If the warning state is entered at the corresponding time, the value is 1; otherwise, the value is 0.

[0046] Each group of candidate parameters includes probability smoothing parameters, probability channel parameters, physical gating parameters, early warning status rule parameters, early warning activation constraint parameters, weak alarm segment pruning parameters, short-process rescue parameters, and long-process activation protection parameters.

[0047] The probability smoothing parameters include the probability smoothing time constant smooth_tau_min;

[0048] The probability channel parameters include the absolute probability channel high threshold T, the absolute probability channel high-low threshold interval hysteresis_gap, the probability history background window length prob_bg_window_min, the probability anomaly channel high threshold prob_z_T, the probability anomaly channel high-low threshold interval prob_z_gap, the probability increment channel high threshold prob_delta_T, the probability increment channel high-low threshold interval prob_delta_gap, the probability ranking channel high threshold prob_pr_T, the probability ranking channel high-low threshold interval prob_pr_gap, the probability background window effective sample number threshold prob_ctx_min_pts, and the probability background window effective time span threshold prob_ctx_min_min.

[0049] The physical gating parameters include the physical gating discrimination mode (gate_mode), the effective hit threshold for the micro-physical enhancement channel (gate_micro), the effective hit threshold for the graupel enhancement channel (gate_gra), the effective hit threshold for the trend enhancement channel (gate_trend), the effective hit threshold for the updraft enhancement channel (gate_up), the physical gating hit count threshold (gate_need), the effective sample count threshold for the gating context (gate_ctx_min_pts), the effective time span threshold for the gating context (gate_ctx_min_min), the high threshold for physical consistency score (phys_T), and the high-low threshold interval for physical consistency score (phys_gap).

[0050] The warning status rule parameters include the high threshold comprehensive vote count threshold hi_need, the low threshold comprehensive vote count threshold lo_need, the warning entry confirmation review time confirm_min, the warning entry confirmation hit count threshold confirm_hits, the warning hold time window hold_min, the maximum continuous warning duration max_on_min, the repeated triggering cooldown time after exiting the warning status retrigger_cooldown_min, and the actual hail cooling down time after hail;

[0051] The warning activation constraint parameters include the absolute probability activation constraint switch start_need_abs and the probability increase activation constraint switch start_need_rise, both of which have a value of 0.

[0052] The weak alarm segment pruning parameters include the minimum score threshold for retaining weak alarm segments (prune_min_keep_score), the average score threshold for weak alarm segments (prune_min_mean_score), the minimum duration threshold for weak alarm segments (prune_min_duration_min), and the additional pruning duration for weak alarm segments that exceed the warning time limit (prune_far_lead_extra_min). Among them, prune_far_lead_extra_min is used to determine whether a continuous warning segment belongs to a far-early weak alarm segment during the training optimization or historical process replay stage.

[0053] Short-process rescue parameters include the short-process sample number threshold short_proc_n, the short-process absolute probability threshold short_abs_T, the short-process probability rise threshold short_rise_T, and the short-process physical channel past window sorting threshold short_gate_pr_T;

[0054] The long process start protection parameters include the long process sample number threshold long_proc_n, the long process start protection time window long_start_guard_min, the long process absolute probability threshold long_abs_T, the long process probability anomaly threshold long_z_T, the long process probability ranking threshold long_pr_T, the long process physical consistency score threshold long_phys_T, and the long process start protection hit number threshold long_start_need;

[0055] The initial candidate value ranges are as follows: smooth_tau_min 0.70–1.20; T 0.405–0.455; hysteresis_gap 0.015–0.060; confirm_min 1.20–2.15; confirm_hits fixed at 2; hold_min 0.70–1.20; max_on_min 3.8–5.6; prob_bg_window_min 18.0–30.0; prob_z_T 0.8–2.2; prob_z_gap 0.10–0.70; prob_delta_T −0.008–0.012. `prob_delta_gap` is set to 0.004–0.030; `prob_pr_T` to 0.76–0.84; `prob_pr_gap` to 0.05–0.17; `prob_ctx_min_pts` to 4–7; `prob_ctx_min_min` to 5.0–10.5; `hi_need` is fixed at 1; `lo_need` is fixed at 1; `gate_mode` is fixed at `mix`; `gate_micro` is set to 0.42–0.70; `gate_gra` to 0.06–0.24; `gate_trend` to 0.22–0.48; `gate_up` to 0.05–0.18; `gate_need` is fixed. =1; start_need_abs is fixed at 0; start_need_rise is fixed at 0; gate_ctx_min_pts is fixed at 5; gate_ctx_min_min is 6.0~9.5; phys_T is 0.54~0.66; phys_gap is 0.07~0.16; short_proc_n is 10~15; short_abs_T is 0.34~0.47; short_rise_T is 0.001~0.016; short_gate_pr_T is 0.80~0.90; long_proc_n is 22~30; long_start_g `long_ard_min` is set to 14.0–22.0; `long_abs_T` to 0.48–0.58; `long_z_T` to 1.3–1.95; `long_pr_T` to 0.84–0.91; `long_phys_T` to 0.62–0.74; `long_start_need` is fixed at 1; `prune_min_keep_score` is set to 0.93–0.985; `prune_min_mean_score` to 0.69–0.78; `prune_min_duration_min` to 0.28–0.90; `prune_far_lead_extra_min` to 2.0–6.0; retrigger_cooldown_min is 1.2 to 2.8; hail_cooldown_min is 1.8 to 3.6.

[0056] In the local search phase, the currently retained candidate parameter set or the candidate parameter set with a high score of the comprehensive objective function in the previous round is selected as the basic candidate parameter set , on the basis of which a local candidate parameter set is generated , which is added to the candidate parameter set as the k-th group of candidate parameters to participate in the subsequent early warning sequence generation and comprehensive objective function evaluation.

[0057] For continuous scalar parameters, a zero-mean normal perturbation method is adopted for local search, that is, a random perturbation is superimposed on the basic parameter value, and the perturbed value is clipped to the local value range corresponding to the parameter; for integer scalar parameters, an integer step perturbation method of -1, 0 or 1 is adopted for local search, and the perturbed value is clipped to the integer value range corresponding to the parameter; for fixed parameters, the preset value remains unchanged during the local search process.

[0058] For any continuous scalar parameter x, let its value in the basic candidate parameter set be , the lower limit of local search be , the upper limit of local search be , and the perturbation scale be , then the value of this parameter in the local candidate parameter set is expressed as: , ; for any integer scalar parameter , let its value in the basic candidate parameter set be , the lower limit of local search be , the upper limit of local search be , then the value of this integer parameter in the local candidate parameter set is expressed as: , ; means restricting the variable V to the interval [L, U], taking L when V<L, taking U when V>U, and taking V when L≤V≤U; represents the local random perturbation of the continuous parameter; represents the local random perturbation of the integer parameter; represents a normal distribution with a mean of 0 and a variance of .

[0059] ​​(2) Calculate the grouped extrapolation validation index and the training set index. Compare the grouped extrapolation validation warning sequence generated by the k-th candidate parameter group with the corresponding real warning label, and calculate the grouped extrapolation validation precision, recall and F1 score, which are expressed as follows: , and , , and Let be the number of true positive samples, false positive samples, and false negative samples in the extrapolation validation of the k-th candidate parameter group, respectively; compare the warning sequences generated in the training set from the k-th candidate parameter group with the corresponding true warning labels, and calculate the precision, recall, and F1 score in the training set, respectively denoted as: , and ,in, , and Let k = 1, 2, ..., K, and ε be the number of true positive samples, false positive samples, and false negative samples in the training set, respectively. ε is a very small positive number to prevent the denominator from being zero. The macro-average index, process-level hit rate, and process robustness index are calculated. Specifically, the process-level grouped extrapolation verification index of the kkk-th candidate parameter group in the qqq-th hailstorm process is calculated. The grouped extrapolation verification warning sequence in this hailstorm process is compared with the corresponding real warning label to obtain the number of true positive samples in the process-level grouped extrapolation verification. Number of false positive samples and the number of false negative samples The precision, recall, and F1 score of process-level grouped extrapolation validation are expressed as follows: , , The process-level grouped extrapolation validation precision is used to characterize the proportion of true positive samples among the radar volume scan samples judged as warning states during the q-th hailstorm process; the process-level grouped extrapolation validation recall is used to characterize the proportion of true positive samples in the hailstorm process that are identified by the warning sequence; the process-level grouped extrapolation validation F1 score is used to comprehensively reflect precision and recall. The warning sequences in the training set of the k-th candidate parameter group during the q-th hailstorm process are compared with the corresponding true warning labels to obtain the number of true positive samples in the training set. Number of false positive samples and the number of false negative samples Precision, recall, and F1 score within the process-level training set are expressed as follows: , , The aforementioned process-level metrics within the training set are used to evaluate the fitting results of the candidate parameter groups to the training samples, and are compared with the corresponding group extrapolation validation metrics to evaluate the consistency between performance within and outside the training set.

[0060] (3) Let M represent the total number of hailstorm processes participating in the evaluation. The process-level grouped extrapolation validation indices for each hailstorm process are averaged using equal weights to obtain the grouped extrapolation validation macro-average precision, macro-average recall, and macro-average F1 score for the k-th candidate parameter group: , , Accordingly, the indices within the training set for each hailstorm process are averaged using equal weights to obtain the macro-average precision, macro-average recall, and macro-average F1 score within the training set: , , Macro-averaging ensures that each hailstorm event has equal weight in the evaluation results, preventing individual events with a large number of radar volume scan samples from dominating the overall evaluation. Furthermore, the hit rate of the warning sequence is calculated by extrapolating whether the continuous time periods of real positive samples are grouped. Let... Let represent the total number of consecutive periods of true positive samples during all hailstorm processes included in the evaluation. Let represent the number of consecutive time periods in which at least one radar volume scan time is hit by the grouped extrapolation verification warning sequence generated by the k-th group of candidate parameters. Then, the process-level hit rate is expressed as: The process-level hit rate evaluates the candidate parameter group from the perspective of whether the continuous hail risk period is covered by the warning, which can avoid evaluating the warning effect based solely on the classification results of a single radar volume scan sample. To evaluate the weakest performance and inter-process stability of the candidate parameter group under different hail-prone processes, a process robustness index is further calculated. Let... This indicates a set of hailstorm processes that, in true warning labels, contain at least one positive sample. This represents the number of hailstorm events included in the set. The minimum process recall and minimum process precision of the k-th candidate parameter group in the set are respectively expressed as: , Based on the process-level grouping of each hail-affected event, extrapolate and verify the deviation of the recall rate relative to the average value, and calculate the standard deviation of the process recall rate: , The minimum process recall and minimum process precision are used to reflect the weakest identification performance of the candidate parameter group in a single hailstorm process; the standard deviation of the process recall is used to reflect the degree of recall fluctuation of the candidate parameter group among different hailstorm processes. The smaller the standard deviation, the more stable the warning performance of the candidate parameter group is in different hailstorm processes. q=1,2,…,M, k=1,2,…,K, and ε is a very small positive number set to prevent the denominator from being zero.

[0061] (4) Calculate the negative sample false alarm rate, time-series penalty term, and state transition penalty term; specifically, based on the k-th candidate parameter group The generated grouped extrapolation verification warning sequence is used to calculate the false alarm rate of negative samples. State transition penalty item Penalties for exceeding the warning lead time limit Punishment for traveling too far in advance Early warning and punishment Even after hail, warnings and penalties will still be in effect. False alarm penalties for hailstorms occurring far from the target , is represented as: ; ; ; ; ; ; ; and These represent the set of hailstorm events without any real positive samples and the number of hailstorm events, respectively. The number of radar volume scan times during the q-th hailstorm process; The grouped extrapolation verification warning flag generated for the k-th candidate parameter group during the q-th hailstorm process and at the t-th radar volume scan time;

[0062] Penalty for exceeding the warning lead time limit Punishment for traveling too far in advance Early warning and punishment Even after hail, warnings and penalties will still be in effect. False alarms during hailstorms far from the target and number of state transitions Represented as: ; ; ; ; ; ; This represents the lead time from the t-th radar volume scan time during the q-th hailstorm process to the next actual hailfall time; when there is no correlated actual hailfall time after the current radar volume scan time t. =+∞; H is the upper limit of the warning period; This indicates whether the radar volume scan time at the t-th moment within the q-th hailstorm process falls within the actual hail observation period. If it falls within the actual hail observation period, the value is 1; otherwise, the value is 0. The average number of state transitions during each hailstorm event;

[0063] (5) The negative sample false alarm rate, time series penalty term, and state transition penalty term are taken as deduction terms in the comprehensive objective function, and are expressed as follows: ;

[0064] (6) Calculate the target gap term, generalization difference term, precision-recall balance, target overall distance, and warning ratio penalty term, specifically: based on the target precision... and target recall rate Calculate the difference in accuracy between the grouped extrapolation verifications for the k-th candidate parameter group. Group extrapolation to verify recall gap In-training set accuracy gap The difference between the recall rate and the in-training set , is represented as: ; ; ; The aforementioned target gap term characterizes the degree to which the corresponding precision or recall is lower than the target value; when the corresponding indicator reaches or exceeds the target value, the corresponding target gap term is set to 0. Precision generalization difference is calculated based on the differences between the in-training set indicators and the grouped extrapolation validation indicators. Recall rate generalization difference Difference in generalization with F1 score , is represented as: ; ; The smaller the generalization difference, the closer the in-training set metrics of the candidate parameter groups are to the group extrapolation validation metrics. The balance between precision and recall in group extrapolation validation is calculated. And the balance between precision and recall within the training set. , is represented as: ; The balance value ranges from 0 to 1; the closer the precision and recall are, the closer the balance value is to 1. The combined distances between the extrapolation validation result and the in-training set result of the k-th candidate parameter group relative to the target precision and target recall are calculated and expressed as: ; The weighting coefficient for recall bias is 0.85; the smaller the overall target distance, the closer the precision and recall of the candidate parameter group are to the target value. The prediction and warning ratio of the k-th candidate parameter group during the q-th hailstorm process is calculated. The proportion of true positive samples in the q-th hailstorm process , is represented as: ; ,in ; When the predicted warning ratio deviates by no more than 0.10 from the actual positive sample ratio, or deviates by no more than 0.18 from the actual positive sample ratio, the corresponding penalty term is set to 0. Let... Let be the set of hailstorm processes that contain at least one positive sample in the true warning label. The penalties for excessively high average warning ratios, excessively low average warning ratios, and the maximum warning ratio deviation for the k-th candidate parameter group are calculated and expressed as follows: ; ; ,in, and These respectively characterize the degree to which the average warning ratio of the k-th candidate parameter group is too high and the degree to which the average warning ratio is too low during a hailstorm with positive samples; This characterizes the maximum warning proportion deviation during hailstorm processes with positive samples. The above indicators are input into the subsequent comprehensive objective function; when... When =0, let , , , , ;

[0065] (7) Based on the aforementioned group extrapolation validation index, training set index, macro average index, process-level hit rate, process robustness index, target gap term, generalization difference term, precision and recall balance, target comprehensive distance, negative sample false alarm rate, time series penalty term, state transition penalty term, and warning ratio penalty term, construct the k-th candidate parameter group. , is represented as: Where, k=1,2,…,K, K represents the total number of candidate parameter groups; the positive coefficient terms in the comprehensive objective function are used to evaluate the early warning recognition capability, process-level stability and performance inside and outside the training set of the candidate parameter groups, and the negative coefficient terms are used to deduct the effects caused by false alarms, time series deviations, frequent state transitions, insufficient target indicators, differences inside and outside the training set and early warning ratio deviations.

[0066] (8) For candidate parameter sets Calculate the integrated objective function separately The candidate parameter sets are then compared from highest to lowest based on their comprehensive evaluation scores, and the set that maximizes the comprehensive objective function is selected as the optimal candidate parameter set. , is represented as: Optimal candidate parameter group index k * Among them, k * This represents the index of the candidate parameter group corresponding to the optimal candidate parameter group;

[0067] S6. Acquire multi-source observation data at multiple radar volume scan times during real-time monitoring;

[0068] S7. Determine whether the current hail cloud object should enter monitoring and tracking, maintain monitoring and tracking, or stop tracking and issue a warning based on the tracking start / stop conditions; if the current hail cloud object meets the conditions for entering monitoring and tracking or is in monitoring and tracking status but does not meet the conditions for stopping tracking and issuing a warning, proceed to S3; if the current hail cloud object meets the conditions for stopping tracking and issuing a warning, proceed to S10; the specific tracking start / stop conditions are: when the same hail cloud object in the combined reflectivity of the same radar, at two consecutive adjacent volume scan times and When all parameters meet the requirement of "maximum combined reflectivity echo intensity ≥ 40 dBZ", the system is determined to enter monitoring and tracking; when the hail cloud object meets the requirement of "maximum combined reflectivity echo intensity ≥ 40 dBZ" in three consecutive adjacent volume scan times of the same radar, the system is considered to enter monitoring and tracking. , and When both conditions are met that "maximum combined reflectivity echo intensity < 40 dBZ", or when the target matching results based on two adjacent volume scan times determine that the hail cloud object has dissipated or cannot be continuously tracked, a tracking stop warning is issued. The target matching results are specifically: based on the spatial position continuity, area overlap, echo intensity continuity, and motion direction continuity of the hail cloud object in two adjacent radar volume scan times, a matching judgment is made to determine whether the hail cloud objects in two adjacent radar volume scan times belong to the same tracking object; if the hail cloud objects in two adjacent radar volume scan times meet the preset matching conditions, they are determined to be the same tracking hail cloud object; if the preset matching conditions are not met, the hail cloud object is determined to have dissipated or cannot be continuously tracked; the preset matching conditions are that the combined reflectivity core areas of the hail cloud object in the previous radar volume scan time and the hail cloud object in the current radar volume scan time have spatial overlap, or the centroid distance is not greater than a preset distance threshold; whereby the combined reflectivity core area is a set of radars with a combined reflectivity ≥ 40 dBZ.

[0069] S8. Analyze the candidate feature set using the optimized ensemble probability model to obtain early warning judgment parameters; specifically including:

[0070] S8001. Based on the final input feature set F determined in the training and optimization phase, extract the input features corresponding to the final input feature set F from the candidate feature set corresponding to the current radar volume scan time t, and form the final input feature vector of the current radar volume scan time t.

[0071] S8002 and m logistic regression sub-models analyze the probability of the first warning based on the final input set F. , is represented as: ; The current radar volume scan time t is the first... One input feature; and , respectively, are the intercept term and regression coefficient of the r-th logistic regression sub-model; d is the final input feature dimension; r=1,2,…,m;

[0072] S8003, weighted fusion of m first early warning probabilities at the integrated layer Probability of obtaining a second warning , , Let be the ensemble weights corresponding to the r-th logistic regression sub-model, and satisfy . When m=3, ;

[0073] S8004, physical gating mechanism layer based on second early warning probability A probability channel is constructed, and a physical gating channel is constructed based on feature derivations related to the hail formation physical process. Warning judgment parameters are output based on the probability-gated channel and the physical gating channel; specifically, this includes:

[0074] ① Exponential time smoothing second warning probability Obtain the smoothed early warning probability at the current radar volume scan time t. , For the remaining effective radar volume scan times, let: The dynamic smoothing coefficient is expressed as: ;in, The second early warning probability at the first radar volume scan moment; Let be the smoothed early warning probability at the (t-1)th radar volume scan time. This represents the actual time interval between the current radar volume scan time t and the (t-1)th radar volume scan time.

[0075] ② Based on the smoothed early warning probability at the current radar volume scan time t Construct absolute probabilities, and before calculating statistics dependent on historical context, determine whether the historical context of the probabilities is valid. The absolute probability is expressed as: ;make This indicates the length of the preset probability historical background window before the current radar volume scan time t. The set of historical radar volume scan times that contain effective smoothing of early warning probabilities: The number of effective radar volume scan samples within the probability history background window is: ,when When the effective time span of the probabilistic historical background window is the time difference between the current radar volume scan time and the earliest effective radar volume scan time within the background window; when When the effective time span is set to 0, it indicates that there is no available historical background at the current moment. The valid flag for probabilistic historical context is represented as: ,when At that time, the probability historical background window simultaneously satisfies the requirements for both the number of effective samples and the effective time span. Let The earliest valid radar volume scan time within the probabilistic historical background window: The median absolute deviation of the probability of a smoothed early warning within the historical background window is expressed as: Based on this, the probability outlier, probability increment, and probability ranking factor are calculated: , , ;when If the number of valid samples or the valid time span of the probability history background window does not meet the requirements, the calculation of the median, median absolute deviation, background window start time, probability outlier, probability increment, and probability ranking factor is not performed. Instead, the following is directly set: =0, 0 =0.5, where the probability anomaly quantity and probability increment are set to 0 to avoid false probability anomalies or probability increase signals when historical background is insufficient; the probability ranking quantity is set to 0.5 to represent the neutral ranking position. Because The summation formula is only used when Execute at the time, and Guaranteed Therefore, it will not appear. This results in a division-by-zero problem. Here, u and v represent historical radar volume scan times within the probabilistic historical background window; The threshold for the number of valid samples in the probability background window; The threshold for the effective time span of the probability background window; This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. To prevent extremely small positive numbers with zero denominators when the absolute deviation of the median is zero;

[0076] ③ Based on the absolute probability, probability anomaly, probability increment, and probability ranking obtained in step ②, construct the absolute probability channel, probability anomaly channel, probability increment channel, and probability ranking channel respectively. The corresponding channel values ​​are as follows: , , and Simultaneously, based on the derived features related to hail formation processes, such as comprehensive microphysical activity, graupel enhancement characteristics, microphysical change trends, and updraft intensity, microphysical enhancement channels, graupel enhancement channels, trend enhancement channels, and updraft enhancement channels are constructed. For any physical gated channel r, where r∈{micro,gra,trend,up}, let... This represents the order of the current channel value within the past window of the physical history background window. This represents the robust outlier of the current channel value relative to the historical median and robust scale. Since the past window ranking value ranges from 0 to 1, while the original robust outlier's range is not limited by this interval, to enable weighted fusion, the robust outlier is first truncated and normalized to obtain the normalized robust outlier. , is represented as: ;in, The expression (,0,1) represents truncating the input variable ξ to a closed interval between 0 and 1; ξ < 0 is taken as 0, ξ > 1 is taken as 1, and the original value is retained when 0 ≤ ξ ≤ 1. A mixed value is constructed based on the past window sort quantity and the normalized robust outlier. , is represented as: The physical gating mode `gate_mode` determines which form of channel value is used for the physical gating channel in the current calculation. When `gate_mode` is set to `pr`, the past window sorting value is used. As the actual channel value, this method mainly reflects the relative position of the current physical feature within its own historical context; when gate_mode is z, the original robust anomaly value is used. As the actual channel value, this method primarily reflects the degree of deviation of the current physical characteristics from the historical median state; when gate_mode is set to mix, the mixed value is used. As the actual channel value, it utilizes both relative sorting information and robust anomaly information. The actual channel value of the r-th physically gating channel, determined according to the above physical gating discrimination method, at the current radar volume scan time t is denoted as... By setting selectable physical gating discrimination methods, a more suitable physical feature representation method can be selected during the training and optimization phase based on the statistical characteristics of different research areas and historical samples, without changing the overall structure of subsequent physical gating and state machines;

[0077] ④ Compare the absolute probability channel, probability anomaly channel, probability increment channel, and probability ranking channel with their respective high thresholds to obtain the high threshold vote count at the current radar volume scan time t. , represented as Among them, the absolute probability channel does not depend on historical context, so it participates in high-threshold voting regardless of whether the probability historical context is valid; the probability anomaly channel, probability increment channel, and probability ranking channel are all multiplied by the probability historical context validity flag. Voting is only conducted when the probabilistic historical context is valid. When the historical background is insufficient, the historical statistics channel does not increase high-threshold voting, thereby avoiding erroneous initiation caused by unstable outliers or ranking quantities.

[0078] ⑤ Compare the absolute probability channel, probability anomaly channel, probability increment channel, and probability ranking channel with their respective low thresholds to obtain the low threshold vote count at the current radar volume scan time t. , represented as The low threshold is obtained by subtracting the preset high-low threshold interval from the corresponding high threshold. The high threshold is mainly used to control the activation of new warnings, while the low threshold is mainly used to determine whether an already activated warning should be maintained. By setting the high-low threshold difference, frequent activation and deactivation caused by small fluctuations in probability or physical quantities around a single threshold can be reduced. Similar to high threshold voting, the three probability channels that rely on historical background only... When the threshold is 1, participate in low-threshold voting;

[0079] ⑥ Before performing routine physical gating judgments, first determine whether the current radar volume scan time t has sufficient physical history background. Let This represents the number of valid radar volume scan samples within the physical gated history background window. Let N represent the effective time span from the earliest effective radar volume scan time to the current radar volume scan time t within the physical gated history background window. obj (t) represents the cumulative number of valid radar volume scan samples obtained from the current hail cloud object since it entered the monitoring and tracking phase up to the current radar volume scan time t. Physical gating context valid flag. Represented as: ;when When =1, it indicates that the physical history background meets the requirements in terms of both the number of effective samples and the time coverage, and the process can proceed to step ⑦ to perform normal physical gating judgment. When t = 0, directly determining the physical gating condition as invalid might miss hailstorm processes with short development times, lacking a complete historical background, but whose probability or hail formation physical characteristics are rapidly increasing. Therefore, this embodiment sets up a short-process rescue mechanism. Let t − This represents the most recent effective radar volume scan time t before the current radar volume scan time. The increase in the current absolute probability relative to the previous effective radar volume scan time is expressed as: When there is no previous valid radar volume scan time at the current radar volume scan time, it means that there is no comparable previous probability benchmark at the current time. In this case, let... =0. Short-term rescue sign Represented as: Short-process rescue requires the following conditions to be met simultaneously: the cumulative number of valid radar volume scan samples of the current hail cloud object has not exceeded the short-process sample number threshold; the current absolute probability reaches the short-process absolute probability threshold; and the increase in absolute probability reaches the short-process probability increase threshold, or the past window ranking of at least one physical gating channel reaches the short-process physical channel past window ranking threshold. Therefore, short-process rescue is not simply about relaxing physical gating, but requires that the current object already has a high absolute risk, and simultaneously exhibit a rapid increase in probability or a significant enhancement in physical characteristics. When G(t) = 1, the physical gating condition at the current moment is temporarily considered to be true, so let G(t) = 1. However, since the physical history context has not yet met the normal gating requirements, the probability votes are not increased by using the physical consistency score. Therefore, let gate_hits(t) = 0, phys_score(t) = 0, VH∗(t) = VH(t) and VL∗(t) = VL(t), and then proceed to step ⑩ to perform the long process start protection judgment. When Rescue(t) = 0, it means that there is neither a valid physical history background nor enough short process fast enhancement evidence at the current moment. Let G(t) = 0, B(t) = 0, gate_hits(t) = 0, phys_score(t) = 0, VH∗(t) = VH(t) and VL∗(t) = VL(t), and then proceed to step ⑪.

[0080] ⑦ When the physics-gated context is valid, the actual channel values ​​of the microphysics enhancement channel, hail enhancement channel, trend enhancement channel, and updraft enhancement channel are compared with their corresponding valid hit thresholds. The physics-gated hit count gate_hits(t) is represented as: The physical gating hit count ranges from 0 to 4, representing the number of channels among the four key hail-forming physical channels that reach the effective hit threshold. The physical gating condition fulfillment flag G(t) is represented as: When G(t) = 1, it indicates that the number of physical gating channels that reach the effective hit threshold is not less than the preset physical gating hit threshold, and the current probability signal has the corresponding hail formation physical process support; when G(t) = 0, it indicates that the common support of physical channels is insufficient. After completing the judgment, proceed to step ⑧;

[0081] ⑧ To further reflect the extent to which each physical gated channel exceeds or approaches the effective hit threshold, and instead of simply using binary results of hit or miss, normalized gating scores relative to the effective hit threshold are calculated for the microphysics enhancement channel, graupel enhancement channel, trend enhancement channel, and updraft enhancement channel; for any physical gated channel r, r∈{micro,gra,trend,up}; when using past window sorting... When used as a channel value, the corresponding normalized gating score is... Represented as: When the robust outlier Z_r(t) is used as the channel value, the corresponding normalized gating score is... Represented as: When using mixed values When used as a channel value, the corresponding normalized gating score is... Represented as: When r takes the values ​​micro, gra, trend, and up respectively, These correspond to gate_micro, gate_gra, gate_trend, and gate_up, respectively. The closer the normalized gating score is to 1, the stronger the support of the corresponding physical channel for the current hail risk. The closer the normalized gating score is to 0, the lower the corresponding channel value is to the effective hit threshold. clip(ξ,0,1) is used to truncate the input variable ξ, where ξ is the intermediate calculation amount within the parentheses of the corresponding function. When ξ<0, clip(ξ,0,1)=0; when ξ>1, clip(ξ,0,1)=1; when 0≤ξ≤1, clip(ξ,0,1)=ξ.

[0082] ⑨ Calculate the physical consistency score phys_score(t) based on the normalized gating scores of the microphysics enhancement channel, graupel enhancement channel, trend enhancement channel, and updraft enhancement channel. The physical consistency score is used to characterize the overall support of four key hail-forming physical processes—microphysical enhancement, hail enhancement, changing trends, and updrafts—for the current probabilistic warning signal. The physical consistency score does not directly replace the output of the probabilistic model, but is used to make limited corrections to the probability votes obtained in steps ④ and ⑤. When phys_score(t) is not lower than the high threshold phys_T of the physical consistency score, it indicates that the current physical process strongly supports the high-risk probability signal, and one vote is added to the high threshold vote count; otherwise, no addition is made. High threshold comprehensive vote count. The low threshold retention criterion employs a relatively lenient physical consistency condition. When If the total number of votes is not less than max{0.25, phys_T−phys_gap}, add 1 vote to the low threshold vote count; otherwise, do not add any votes. Low threshold combined vote count Represented as: Since step 9 is only executed after the physical gating context is valid and normal physical gating calculation is completed, the above physical consistency score has valid historical background support. When the physical gating context is insufficient and the process proceeds from the short-process rescue branch to subsequent steps, the physical consistency score has already been set to 0 in step 6, and the overall vote count remains the original probability vote count, without using insufficient physical history information to increase the vote;

[0083] ⑩ To prevent premature warning activation due to local probability fluctuations when hail clouds just enter the long-process phase, this embodiment sets up long-process activation protection for objects that have reached the long-process sample number threshold. N represents the cumulative number of valid radar volume scan samples of the current hail cloud object. obj(t) The radar volume scan time at which the long-process sample count threshold (long_proc_n) is first reached. If the current hail cloud object has not yet reached the long-process sample count threshold, it indicates that the object has not yet entered the long-process initiation protection phase, so the long-process initiation protection period flag (Guard(t)) is set to 0. After reaching this threshold, the elapsed time between the current radar volume scan time and the time when the long-process sample count threshold is first reached is calculated: The long-process start-up protection period indicator is represented as follows: During the long-process start-up protection period, the long-process start-up protection hit count (LongStartVotes(t)) is calculated based on the absolute probability, the probability anomaly and probability ranking in the historical context of the effective probability, and the physical consistency score. Where, when the probabilistic history context is invalid, V prob When (t) = 0, probability anomalies and probability rankings are not included in the long-process start-up protection hit count, avoiding insufficient historical statistical information to release protection. The long-process start-up protection suppression flag B(t) is expressed as: When the current moment is still within the long-process start-up protection period and the number of long-process start-up protection hits is lower than the preset hit threshold, B(t) = 1 is set to suppress the activation of new warnings at the current moment; when the number of hits reaches the threshold, or when the current moment has left the long-process start-up protection period, B(t) = 0 is set. This suppression flag is only used in the subsequent high-threshold triggering stage and does not directly participate in the low-threshold maintenance judgment of already activated warnings, thereby avoiding the long-process start-up protection from erroneously interrupting the ongoing effective warnings.

[0084] ⑪ After completing the probability channel, physical gating, short-process rescue, physical consistency score correction, and long-process start-up protection, set the physical gating condition establishment flag G(t), the long-process start-up protection suppression flag B(t), and the high-threshold comprehensive voting number. and low threshold combined vote count The probability channel value, physical gating channel value, and early warning state timing parameters obtained in the aforementioned steps are used together as the early warning judgment parameters for the current radar volume scan time t and input into step S9. The early warning state timing parameters include at least the previous early warning state of the current hail cloud object, the start time of this continuous early warning, the most recent effective holding time, and the most recent exit time of the early warning state. Even if the physical gating condition is not met at the current time, the above-mentioned state timing parameters are still input into step S9 so that the state machine can determine whether the existing early warning should continue to be maintained or exit smoothly based on the early warning holding time window and the maximum continuous early warning duration, rather than immediately terminating the early warning due to a single failure of physical gating.

[0085] S9. Based on the warning judgment parameters output in step S8, and combined with the previous warning status of the current hail cloud object, the warning entry confirmation rule, the repeated trigger cooling rule, the actual hailfall cooling rule, the warning maintenance rule, the maximum continuous warning duration constraint, and the hourly weak alarm judgment rule, determine the final warning flag and analysis result for the current radar volume scan time. The analysis result includes a warning status and a no-warning status; when the final warning flag is 1, the analysis result is a warning status and a warning flag is generated; when the final warning flag is 0, the analysis result is a no-warning status. After completing the status update for the current radar volume scan time, return to step S6 to continue acquiring real-time monitoring data for the next radar volume scan time:

[0086] S9001. Determine the previous warning state at the current radar volume scan time. When the current radar volume scan time is the first valid radar volume scan time of the current hail cloud object, since there is no previous valid radar volume scan state to inherit, the previous warning state is initialized to a no-warning state; in other cases, the previous warning state is determined according to the final warning flag of the previous valid radar volume scan time. Therefore, subsequent steps can uniformly use the "previous warning state" without needing to set a non-existent previous time for the first valid radar volume scan time. When G(t)=1, it indicates that the current time meets the physical gating condition or the short-process rescue condition, and proceeds to S9002; when G(t)=0, the current time does not have new physical gating support, and the low threshold hold flag L(t)=0. If the previous warning state is a no-warning state, a new warning cannot be initiated at the current time, the initial warning state Y0(t)=0, and proceeds to S9006. If the previous warning state was a warning state, the system will not immediately exit due to the failure of the physical gating at the current moment. Instead, it will determine whether the time interval between the current radar volume scan time and the most recent effective holding time is greater than the warning holding time window hold_min, or whether the duration of this continuous warning exceeds the maximum continuous warning duration max_on_min. The duration of the continuous warning is determined by the actual time difference between the current radar volume scan time and the start time of this continuous warning. If any exit condition is met, Y0(t) = 0; if neither exit condition is met, Y0(t) = 1. After the judgment is completed, the system proceeds to S9006.

[0087] S9002. Based on the high-threshold integrated voting number, low-threshold integrated voting number, physical gating condition establishment flag, and long-process start-up protection suppression flag at the current radar volume scan time, determine the high-threshold trigger flag H(t) and the low-threshold hold flag L(t) respectively. This formula requires that the high-threshold comprehensive vote count reaches the high-threshold comprehensive vote count threshold, the physical gating condition is met, and the current time is not subject to long-process start-up protection suppression. High-threshold triggering only occurs when all three conditions are met simultaneously. The low-threshold hold flag is represented as: The low-threshold hold flag does not include the long-process start protection suppression flag B(t) because long-process start protection is only used to prevent new warnings from starting too early and should not be used to interrupt hail cloud objects that have already entered the warning state. When the current warning state is no warning state, it is necessary to confirm whether to enter the warning state through the high-threshold triggering results of continuous time intervals, so it enters S9003. When the current warning state is the warning state, it is first determined whether the duration of the current continuous warning exceeds the maximum duration of the continuous warning max_on_min. If it has exceeded it, regardless of whether the current low-threshold hold flag is valid, Y0(t)=0 is set to ensure that the warning does not continue indefinitely. If the duration of the continuous warning has not exceeded max_on_min and L(t)=1, it means that the current probability voting and physical gating still meet the hold conditions, so Y0(t)=1 is set to 1. If the duration of continuous early warning has not exceeded max_on_min, but L(t)=0, then determine whether the time interval between the current radar volume scan time and the most recent effective hold time is greater than hold_min; if it is greater, set Y0(t)=0, otherwise set Y0(t)=1. After completing the judgment, proceed to S9006;

[0088] S9003. When the current warning state is no warning state, with the current radar volume scan time t as the endpoint, construct a set of warning entry confirmation times based on the warning entry confirmation review duration confirm_min. , is represented as: The set includes only valid radar volume scan times of the current hail cloud object within the confirmed time range. (Statistical set) The effective hit count C(t) of the high threshold trigger flag: ;make This indicates the number of valid radar volume scan times within the set of times when an early warning is confirmed: Because radar volume scans may have uneven time intervals or missing data for some scans, the actual number of valid radar volume scans obtained within the confirmation time window may be less than the preset confirmation hit threshold. If a fixed `confirm_hits` is always used, there may be situations where all valid moments within the confirmation window meet the high threshold trigger, but the warning cannot be triggered due to insufficient valid volume scans. Therefore, the effective confirmation hit threshold is determined based on the actual number of valid volume scans within the confirmation time window: ;when If the current high threshold trigger result has sufficient continuity support, proceed to S9004 to execute the cooling rule judgment; otherwise, set Y0(t)=0 and proceed to S9006. This setting does not lower the confirmation threshold under normal data integrity conditions, but only matches the confirmation requirement with the amount of available data when the actual number of effective volume scans is insufficient.

[0089] S9004. After the high-threshold triggering result meets the early warning entry confirmation requirements, further determine whether the current radar volume scan time meets the repeated triggering cooling rule and the actual hailfall post-cooling rule. Repeated triggering cooling is used to prevent the same hail cloud object from repeatedly triggering early warnings in a short period due to fluctuations in probability or physical channel values ​​near the threshold after exiting the early warning period. Actual hailfall post-cooling is used to prevent repeated early warnings from being issued in a short period due to residual strong echoes or probability smoothing inertia after actual hailfall has been recorded. The valid flag for repeated triggering cooling is denoted as... If the current hail cloud object does not have a radar volume scan time of its most recent exit from the warning state, it indicates that the current object does not have the problem of repeated triggering after exiting the warning state. =1. When there is a radar volume scan time that most recently exited the early warning state. When the cooling is repeatedly triggered, the valid flag is indicated as follows: Therefore, when the time interval between the current moment and the last exit from the warning state has reached the repeated trigger cooling period, the repeated trigger cooling is effective; when the cooling period has not yet been reached, but the current hail cloud object meets the strong restart conditions, the repeated trigger cooling restriction can be lifted in advance. The effective cooling flag after actual hailfall is recorded as follows: The most recent actual hail observation time The determination is based on the actual hail observation information already obtained and recorded during the current warning process; when no actual hail observation information has been obtained, it is considered that there is no most recent actual hail observation time. =1; when there is a most recent actual hail observation time t hail At that time, the effective indicator of cooling after actual hail is: When the time interval after actual hailfall has reached the cooling duration, the cooling restriction is lifted. If the cooling duration has not yet been reached, but the physical consistency score reaches a high threshold and the probability ranking channel also reaches a high threshold, it indicates that the hail cloud object has shown strong probabilistic and physical enhancement evidence again after actual hailfall, and the cooling restriction can also be lifted early. The strong restart condition flag R(t) is used to handle the situation where the hail cloud object rapidly enhances again shortly after the warning is lifted, and is represented as: The strong restart condition requires that the number of valid hits within the confirmation time window reaches the greater of the valid confirmation threshold and the high-threshold comprehensive vote threshold, and that at least one of the physical consistency score, probability ranking quantity, or probability anomaly quantity reaches the strong signal condition. Therefore, strong restart does not unconditionally skip the cooling-off period, but rather restores the early warning activation capability in advance when the current object regains sufficient temporal confirmation and strong probability or physical evidence. After completing the above analysis, proceed to S9005;

[0090] S9005. Since S9004 is only executed when the previous warning state was a no-warning state and the current high-threshold trigger has met the confirmation requirements, this step is used to determine whether entering the warning state is allowed at the current moment based on the two cooling validity flags. When =1 and When =1, it indicates that the current moment simultaneously satisfies both the repeated triggering cooling rule and the actual hailfall cooling rule, allowing the current hail cloud object to enter the warning state, and setting the initial warning state Y0(t)=1. When =0 or When Y0(t) = 0, it indicates that at least one cooling limit has not been lifted, and restarting the warning is not allowed at the current moment. Therefore, set Y0(t) = 0. After completing the judgment, proceed to S9006.

[0091] S9006. To suppress initial warnings that are too short in duration and have insufficient overall alarm strength, a time-by-time weak alarm judgment is performed on the continuous initial warning segments up to the current radar volume scan time. When Y0(t)=0, there are no initial warnings that need to be retained at the current time, so the final warning flag Y(t) is directly set to 0, and proceed to S9007. When Y0(t)=1, let... This indicates the start time of the radar volume scan for the current continuous preliminary warning phase, i.e., from... The initial warning status up to the current time t is continuously set to 1. Construct a set of continuous initial warning segments up to the current time: For sets At any radar volume scan time u, let This represents the probability history background window and the maximum absolute probability at the current moment: The four physical gating channels share the valid physical gating context flag, and the number of valid physical gating channels is represented as follows: The overall alarm score of u at the current radar volume scan time. It is composed of absolute probability, probability ranking, physical consistency score, maximum absolute probability within the historical context, and weighted average of physical gating hit rate, and is expressed as: Among them, in the denominator This is used to prevent division by zero when the physical gating context is invalid; in short process branches with invalid physical contexts, the physical gating hit count is already set to 0, therefore this item will not increase the overall alarm score. (Based on the set) Calculate the highest, average, and duration of the comprehensive alarm score obtained up to the current radar volume scan time, based on the comprehensive alarm scores obtained so far. ; ; Two types of conditions are set for retaining consecutive preliminary warning segments. The first type of retention condition is used to retain warning segments that have already reached a high peak alarm score: The second type of retention condition is used to retain warning segments where the average alarm score meets the requirement and the duration meets the minimum duration requirement: The continuous preliminary warning segment retention sign is indicated as follows: If at least one of the two retention conditions is met, the current preliminary warning is retained; if neither condition is met, the current preliminary warning is classified as a weak warning. The final warning indicator is as follows: Since the highest score, average score, and duration are updated progressively with the current radar volume scan time, a short and insufficient initial warning will not be immediately output as a final warning; when the subsequently acquired data makes the peak intensity or average intensity and duration meet the retention conditions, the final warning can be output from the current time. After the judgment is completed, proceed to S9007;

[0092] S9007. Based on the previous warning state determined in S9001 and the final warning flag Y(t) at the current radar volume scan time, determine whether a warning state transition has occurred at the current time, and update the warning start time, the most recent valid hold time, and the most recent exit time from the warning state accordingly. If the previous warning state is no warning state and the current final warning flag Y(t) = 1, the warning state changes from 0 to 1, the analysis result is a warning state, a new warning flag is generated, the current radar volume scan time t is recorded as the start time of this continuous warning, and the most recent valid hold time is initialized to the current radar volume scan time t. The warning start time is only recorded when the state changes from 0 to 1, and is not repeatedly updated in each subsequent volume scan time of maintaining the warning, thus enabling the correct calculation of the continuous warning duration. If the previous warning state is a warning state and the current final warning flag Y(t) = 1, the final warning flag remains at 1, the analysis result continues to be a warning state, and the start time of this continuous warning is not re-recorded. Only when the current low-threshold hold flag L(t) = 1 is the most recent valid hold time updated to the current radar volume scan time t. When L(t) = 0 but the warning continues because the hold time window has not expired, the most recent valid hold time is not updated, so that subsequent times can correctly determine whether the hold time window has expired based on that time. When the current warning state is a warning state and the current final warning flag Y(t) = 0, the warning state changes from 1 to 0, the analysis result is a no-warning state, and the current radar volume scan time t is recorded as the most recent time t of exiting the warning state. off This is used for subsequent repeated triggering of cooldown checks.

[0093] When the current warning state is no warning state and the current final warning flag Y(t) = 0, the final warning flag remains at 0, and the warning start time, the most recent valid holding time, and the most recent exit time from the warning state are not updated. After completing the current state recording, return to step S6;

[0094] S10. After the current hailstorm process ends, based on the current hailstorm process and the actual ground hail situation, the object tracking results stored during the real-time warning process are correlated with the actual hail period to generate newly labeled samples as training samples, and the process is returned to S3. After the newly labeled samples are merged into the original training sample set, S3 to S5 are executed again for training and optimization to obtain the updated ensemble probability model and the updated warning release parameters. The updated warning release parameters are not a simple superposition or fixed correction of the previous round of warning release parameters, but rather the optimal parameter set is re-determined based on the updated training sample set formed by merging the original training sample set and the newly labeled sample set, using the same parameter search and comprehensive objective function evaluation method as in S5. Therefore, the relationship between the warning release parameters obtained in the current round of training and the warning release parameters obtained in the previous round of training is a replacement update relationship, that is, the parameters of the later round are used to replace the parameters of the previous round and applied to the subsequent real-time warning process. The parameter differences reflect the correction results of the new samples on the model output characteristics, physical gating conditions, and warning release rules. After training, the updated ensemble probability model is evaluated. If the updated ensemble probability model has a Precision, Recall, and F1 score of at least 0.7 on the preset evaluation dataset, the updated ensemble probability model and the corresponding warning release parameters are used for subsequent real-time warnings. Otherwise, the ensemble probability model and warning release parameters obtained in the previous training round are kept unchanged.

[0095] Example: The experimental hailstorm process in a certain area on a certain day of a certain month of a certain year is an isolated case of hailstorm that produced continuous hail. There were two hailstorm processes. The analysis period is the second hailstorm process from 10:18 to 10:43 UTC.

[0096] like Figure 2 As shown, the hail cloud traversed complex and undulating mountainous areas along its path. Affected by the mechanical lifting effect of the mountains, from 10:21 (a1) to 10:35 (a3), the intensity of the hail cloud continued to increase, and the area with radar echo intensity greater than 40 dBZ continued to expand, with the area of ​​strong echo intensity greater than 50 dBZ showing a particularly significant increase. Hail fell between 10:24 and 10:35. After the hail fell, until 10:43 (a4), the intensity of the hail cloud gradually weakened, and the area of ​​echo intensity greater than 40 dBZ shrank significantly, with the area of ​​strong echo intensity greater than 50 dBZ shrinking considerably.

[0097] like Figure 3 As shown, under terrain conditions favorable for mechanical uplift, the storm exhibited two updraft pulses: one being a net uplift below the core of >40 dBZ (during 10:18–10:21 UTC). The second is that the storm enters the windward slope and the slope along the path turns positive (during 10:27–10:29 UTC). rad), corresponding to two rising pulses respectively ( Figure 3 (d, b2–b3). The supercooled water SWA exhibits a two-stage replenishment-rapid consumption evolution: the reservoir increased by 246 units at 10:24 UTC, and reached its peak at 10:29 UTC. Library, 10:32 UTC reduction The total SWA volume remained around 2000 CU between 10:29 and 10:35 UTC, indicating the presence of a relatively persistent supercooled liquid water reservoir within the mixed phase layer. The phased evolution of hail embryos facilitated a prolonged effective hail RH growth period, resulting in the maximum hail diameter during this process. The approximately 11-minute duration of ground hailfall was due to the temporal continuity and partial overlap of RH deposition generated by the two rising pulses.

[0098] like Figure 4 As shown, during the period dominated by topographic effects from 10:21 to 10:27, the stock count of updrafts with velocities of 0–0.5 m / s (NV0–0.5) peaked twice consecutively. Topographic effects further enhanced the dynamic conditions between the mixed phase layers, maintaining high levels of supercooled water glacial activity (SWA), which in turn promoted the continuous growth of ice crystals (CR) and ice crystal polymers (AG). Subsequently, CR and AG, through the adsorption of supercooled water, formed low-density graupel (LDG) and high-density graupel (HDG) through a wet growth process. Based on this, the two types of graupel particles, through mutual collision and coalescence and a wet growth process, ultimately formed hail (RH), and the stock count of RH also showed a rapid growth trend.

[0099] As shown in Table 1, in this embodiment, the model successfully triggered the warning before the hail occurred. The "Whether a warning was issued" in Table 1 indicates whether the system is in the warning issuance state at the time of the volume sweep, rather than whether the warning was triggered for the first time at that time; when the warning has reached the preset lead time requirement before the hail occurred, a new warning sign will not be triggered again at subsequent hail occurrence times.

[0100] As shown in Table 2, the evaluation was conducted using a 6-sample training set and a 2-sample test set. The TRAIN_OOF set, reflecting cross-process generalization ability, achieved Precision, Recall, and F1 scores of 0.8958, 0.7167, and 0.7963, respectively, indicating that the model maintains good early warning performance even under the strict evaluation condition of leave-one extrapolation across processes. The TEST set, with Precision, Recall, and F1 scores of 0.8571, 0.7500, and 0.8000, respectively, is close to the TRAIN_OOF results, indicating that the method performs stably on the test samples overall, without significant overfitting, and has good potential for practical application.

[0101] Table 1 Comparison of ground hail conditions and early warning results

[0102]

[0103] Table 2. Results of grouped extrapolation validation and independent test set.

[0104]

[0105] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A method for early warning of hailstorms under complex terrain, characterized in that, include: S1. Construct the initial ensemble probability model, which includes a feature selection module, m logistic regression sub-models, an ensemble layer, and a physical gating mechanism layer. The output of the feature selection module serves as the input to each of the m logistic regression sub-models, and the outputs of the m logistic regression sub-models all serve as the input to the ensemble layer. The outputs of the ensemble layer and the feature selection module both serve as the input to the physical gating mechanism layer, where m is a positive integer greater than 2. Each logistic regression sub-model is trained using the liblinear solver and L2 regularization. The parameter settings for each of the m logistic regression sub-models are different; S2. Obtain historical hailstorm process samples and label multiple historical hailstorm process samples with warning tags. Then, divide the labeled historical hailstorm process samples into a training sample set and a validation sample set. Each hailstorm process sample includes multi-source observation data at multiple radar volume scan times within the hailstorm process. The multi-source observation data includes X-band dual-polarization radar base data, 30m resolution ASTER GDEM terrain elevation data, cloud microphysical feature data, and dynamic feature data. The X-band dual-polarization radar base data includes the reflectivity factor Z. H Differential reflectivity Z DR , differential phase K DP Radial velocity Vr and correlation coefficient ρ hv ; S3. Extract the basic features corresponding to each radar volume scan time based on multi-source observation data. The basic features include terrain features, dynamic features and cloud microphysical features. Topographic features include the average elevation change ΔH of the core area with a reflectivity greater than 40 dBZ and the topographic slope S along the hail cloud's path. Dynamic features include the 90th percentile of the positive vertical updraft between 0℃ and −20℃, UpP90, the radar counts NV0-0.5 for vertical updrafts between 0℃ and −20℃ with a speed of 0 to 0.5 m / s, and NV0.5-1 for vertical updrafts between 0℃ and −20℃ with a speed of 0.5 to 1 m / s. Cloud microphysical features include the radar counts NSWA for supercooled water, NRH for hail, NCR for ice crystals, NAG for ice crystal polymers, NLDG for low-density graupel, and NHDG for high-density graupel. S4. Construct a candidate feature set based on the basic features. When entering S3 from S2 or S10, proceed to S5; when returning to S3 from S7, proceed to S8. S5. Use the candidate feature set to train and optimize the current ensemble probability model to obtain the optimized ensemble probability model and the corresponding early warning release parameters. S6. Acquire multi-source observation data at multiple radar volume scan times during real-time monitoring; S7. Determine whether the current hail cloud object should be entered into monitoring and tracking, maintained in monitoring and tracking, or stopped in tracking and warning based on the tracking start and stop conditions; If it is determined that the current hail cloud object has entered the monitoring and tracking stage or will continue to be monitored and tracked, proceed to S3; If it is determined that the current hail cloud object will stop tracking and warning, proceed to S10; S8. Analyze the candidate feature set using the optimized ensemble probability model to obtain the early warning judgment parameters; S9. Analyze the early warning judgment parameters to obtain the analysis results. The analysis results are either an early warning state or no early warning. If the analysis results are an early warning state and an early warning flag is generated, return to S6. If the analysis results are no early warning, return directly to S6. S10. After the current hailstorm process ends, based on the current hailstorm process and the actual ground hail situation, the object tracking results stored in the real-time warning process are associated with the actual hail period to generate new labeled samples as training samples, and then returned to S3.

2. The hailstorm warning method under complex terrain conditions according to claim 1, characterized in that, S4 includes: S4001. Construct core features based on basic features. Core features include the comprehensive quantity of ice phase particles at the t-th radar volume scan time. ice relative order of magnitude compressibility Comprehensive microphysical activity index , proportion of granite Terrain disturbance and slope disturbance ; indicates as: , , , , , , where j is the sequence number of the radar volume scan time within the time window; The time window consists of the t-th radar volume scan time and all radar volume scan times within the previous 6 minutes. Indicates time window The number of radar volume scan samples included; This indicates the average elevation change of the core area where the reflectivity of the hail cloud combination is greater than 40 dBZ at the j-th radar volume scan time. The average elevation change values ​​obtained at each radar volume scan time from the start of the current hailstorm to the current radar volume scan time t. The average value; ΔS(j) = S(j) - S(j-1) is the slope change value on the hail cloud's moving path at the j-th radar volume scan time; for the first effective radar volume scan time of the current hail cloud object, since there is no previous effective radar volume scan time, let the slope change value at that time be... =0; The data obtained from the start of the current hailstorm process to the current radar volume scan time t are as follows. The average value; To prevent extremely small positive numbers with a denominator of zero; S4002, Tectonic Ice Phase Enhancement Trend and the trend of graupel enhancement , represented as: ; ; in, Let J3(t) represent the time window consisting of the t-th radar volume scan time and all radar volume scan times within the preceding 3 minutes; let J3(t) be the set of time indices in W3(t) that contain the previous valid radar volume scan time. =|J3(t)| represents the number of effective adjacent difference terms; dt s (j) represents the time interval between adjacent volume scans in seconds; when When = 0, let dIce(t) = 0 and dGra(t) = 0; when When the value is greater than 0, it is calculated based on the effective adjacent difference terms; S4003. Construct coupling features, including the synergistic enhancement effect of complex terrain. and coupling relationship , represented as: , ; S4004. Constructing historical window sorting features for core features and coupled features. Robust anomaly characteristics of historical windows Historical window maximum value characteristics Historical window average characteristics and short-term variation characteristics , Indicates core features or coupled features. Indicates a preset historical time window; S4005. Constructing hysteresis features based on the aforementioned features. ,in, Indicates the preset lag time; S4006. Construct key features related to hail formation processes under complex terrain conditions as fixed features. These fixed features include: the maximum value of the 90th percentile of the updraft UpP90 within a continuous 12-minute time window prior to the current radar volume scan time t (UpP90_max12(t)); and the ice particle intensity index within a continuous 12-minute time window prior to the current radar volume scan time t. The maximum value of ice_max12(t), the maximum value of micro_log of the comprehensive microphysical activity index micro_log within the 12-minute time window before the current radar volume scan time t, the maximum value of graffrac of the graupel proportion index gra_max12(t) within the 12-minute time window before the current radar volume scan time t, the cooperative characteristic value of updraft and ice phase particles UpIce_L0(t) at the current radar volume scan time t, the cooperative characteristic value of updraft and ice phase particles UpIce_L3(t) at the current radar volume scan time t with a lag of 3 minutes, the cooperative characteristic value of terrain disturbance and microphysical activity DHxMicro_L0(t) at the current radar volume scan time t, and the cooperative characteristic value of terrain disturbance and microphysical activity at the current radar volume scan time t with a lag of 3 minutes. The following parameters are considered: the degree of coordination characteristic value DHxMicro_L3(t), the percentile value of the comprehensive microphysical activity index micro_log(t) at the current radar volume scan time t within the preset historical time window, the percentile value of the graupel proportion index graFrac(t) at the current radar volume scan time t within the preset historical time window, the larger value of the historical window percentile values ​​of the ice phase enhancement trend dIce(t) and the graupel enhancement trend dGra(t) at the current radar volume scan time t, and the percentile value of the updraft index UpP90(t) at the current radar volume scan time t within the preset historical time window; among them, ice_max12(t) is the largest value of the historical window percentile values ​​of the ice phase enhancement trend dIce(t) and the graupel enhancement trend dGra(t) at the current radar volume scan time t. For the calculation of the basic quantities, micro_max12(t) and micro_pr_L0(t) are both calculated based on micro_log(t), gra_max12(t) and gra_pr_L0(t) are both calculated based on graFrac(t), trend_pr_L0(t) is determined by the larger of the historical window sorting percentile values ​​of dIce(t) and dGra(t), and UpP90_pr_L0(t) is calculated based on UpP90(t). S4007. Construct dynamic statistical features of the coupled features. These dynamic statistical features include: the maximum value of UpIce (UpIce_max6(t)) within a consecutive 6-minute time window before the current radar volume scan time t; the average value of UpIce (UpIce_mean6(t)) within a consecutive 6-minute time window before the current radar volume scan time t; the first-order linear slope of UpIce relative to time (UpIce_slope3(t)) within a consecutive 3-minute time window before the current radar volume scan time t; the second-order change of UpIce at the current radar volume scan time t (UpIce_d2(t)); and the difference between the UpIce value at the current radar volume scan time t and the UpIce value at the previous 3-minute time window (UpIce_rise3(t)). The maximum value of DHxMicro within a continuous 6-minute time window before radar volume scan time t, the average value of DHxMicro within a continuous 6-minute time window before the current radar volume scan time t, the slope of the first-order linear change of DHxMicro relative to time within a continuous 3-minute time window before the current radar volume scan time t, the second-order change of DHxMicro at the current radar volume scan time t, DHxMicro_d2(t), and the difference between the DHxMicro value at the current radar volume scan time t and the DHxMicro value at the previous 3-minute time window, DHxMicro_rise3(t). S4008. Combine fixed features, automatically selected lag class features, and dynamic statistical features of coupled features to form a candidate feature set; S4009. When entering S3 from S2 or S10, enter S5; when returning to S3 from S7, enter S8.

3. The hailstorm warning method under complex terrain conditions according to claim 2, characterized in that, S5 includes: S5001. Construct a training sample matrix X, a label vector y, and a process grouping vector g based on the candidate feature set. Set sample weights for the training samples: the process grouping vector g must satisfy g... i Samples with weight = q, belonging to the q-th hailstorm process, are assigned basic weights. The unnormalized sample weights of the i-th training sample in the r-th logistic regression sub-model. Normalized sample weights are obtained by normalizing the unnormalized sample weights. ; ; Let be the number of training samples for the q-th hailstorm process. and Let a be the number of positive samples and the number of negative samples in the q-th hailstorm process. r Let be the positive sample enhancement coefficient corresponding to the r-th logistic regression sub-model, where r = 1, 2, ..., m; S5002, the feature selection module uses the ExtraTreesClassifier to rank the importance of candidate features, and selects optional features from the candidate features other than the fixed features based on the importance ranking results; among them, fixed features are not considered as deletable objects and are not included in the output of optional features; for dynamic statistical features of the same type of coupled features, at most one representative feature is retained as optional features. S5003. Merge the optional features with the fixed features, and remove duplicates according to the feature name to obtain the final input feature set F; S5004, the r-th logistic regression sub-model, uses the liblinear solver and L2 regularization to optimize the weighted logistic regression objective function. , Let be the final input feature vector corresponding to the i-th training sample; To classify the binary labels y i The values ​​mapped to {−1,1} are represented as follows: , and Let be the intercept term and regression coefficient of the r-th logistic regression sub-model, respectively. Let L2 regularization strength be the value corresponding to the r-th logistic regression sub-model. S5005. During the training process, a leave-one-out verification method is adopted, which groups samples according to the hail storm process, so that samples of the same hail storm process do not enter the training subset and the leave-out subset at the same time; the predicted probabilities of each round of leave-out subsets are spliced ​​together according to the order of hail storm process and radar volume scan time to form the object extrapolation probability. S5006. Based on object extrapolation warning probability and training set in-set warning probability, in K candidate parameter groups... In the process, parameter search is performed to determine the optimal candidate parameter set, which is then combined with the logistic regression ensemble model to form the optimized ensemble probability model. Each group of candidate parameters includes probability smoothing parameters, probability channel parameters, physical gating parameters, early warning status rule parameters, early warning activation constraint parameters, weak alarm segment pruning parameters, short-process rescue parameters, and long-process activation protection parameters. The probability smoothing parameters include the probability smoothing time constant smooth_tau_min; The probability channel parameters include the absolute probability channel high threshold T, the absolute probability channel high-low threshold interval hysteresis_gap, the probability history background window length prob_bg_window_min, the probability anomaly channel high threshold prob_z_T, the probability anomaly channel high-low threshold interval prob_z_gap, the probability increment channel high threshold prob_delta_T, the probability increment channel high-low threshold interval prob_delta_gap, the probability ranking channel high threshold prob_pr_T, the probability ranking channel high-low threshold interval prob_pr_gap, the probability background window effective sample number threshold prob_ctx_min_pts, and the probability background window effective time span threshold prob_ctx_min_min. The physical gating parameters include the physical gating discrimination mode (gate_mode), the effective hit threshold for the micro-physical enhancement channel (gate_micro), the effective hit threshold for the graupel enhancement channel (gate_gra), the effective hit threshold for the trend enhancement channel (gate_trend), the effective hit threshold for the updraft enhancement channel (gate_up), the physical gating hit count threshold (gate_need), the effective sample count threshold for the gating context (gate_ctx_min_pts), the effective time span threshold for the gating context (gate_ctx_min_min), the high threshold for physical consistency score (phys_T), and the high-low threshold interval for physical consistency score (phys_gap). The warning status rule parameters include the high threshold comprehensive vote count threshold hi_need, the low threshold comprehensive vote count threshold lo_need, the warning entry confirmation review time confirm_min, the warning entry confirmation hit count threshold confirm_hits, the warning hold time window hold_min, the maximum continuous warning duration max_on_min, the repeated triggering cooldown time after exiting the warning status retrigger_cooldown_min, and the actual hail cooling down time after hail; The warning activation constraint parameters include the absolute probability activation constraint switch start_need_abs and the probability increase activation constraint switch start_need_rise, both of which have a value of 0. The weak alarm segment pruning parameters include the minimum score threshold for retaining weak alarm segments (prune_min_keep_score), the average score threshold for weak alarm segments (prune_min_mean_score), the minimum duration threshold for weak alarm segments (prune_min_duration_min), and the additional pruning duration for weak alarm segments that exceed the warning time limit (prune_far_lead_extra_min). Among them, prune_far_lead_extra_min is used to determine whether a continuous warning segment belongs to a far-early weak alarm segment during the training optimization or historical process replay stage. Short-process rescue parameters include the short-process sample number threshold short_proc_n, the short-process absolute probability threshold short_abs_T, the short-process probability rise threshold short_rise_T, and the short-process physical channel past window sorting threshold short_gate_pr_T; The long process start-up protection parameters include the long process sample number threshold long_proc_n, the long process start-up protection time window long_start_guard_min, the long process absolute probability threshold long_abs_T, the long process probability anomaly threshold long_z_T, the long process probability ranking threshold long_pr_T, the long process physical consistency score threshold long_phys_T, and the long process start-up protection hit number threshold long_start_need.

4. The hailstorm warning method under complex terrain conditions according to claim 3, characterized in that, S5006 includes: (1) The k-th candidate parameter group The extrapolation probability and the in-training set probability are applied respectively to generate a grouped extrapolation verification warning sequence. and early warning sequences within the training set Among them, the warning flag at the t-th radar volume scan time generated based on the object extrapolation probability of the k-th candidate parameter group is: The warning flag generated based on the probability within the training set for the k-th candidate parameter group at the t-th radar volume scan time is: If the warning state is entered at the corresponding time, then set it to 1; otherwise, set it to 0. (2) Calculate the grouped extrapolation validation metric and the training set indices, specifically: the grouped extrapolation validation metric includes the accuracy of the grouped extrapolation validation. Recall rate and F1 value , , and These represent the number of true positive samples, false positive samples, and false negative samples in the extrapolation verification of the k-th candidate parameter group, respectively. In-training set metrics include accuracy within the training set. Recall rate and F1 value , , and These represent the number of true positive samples, false positive samples, and false negative samples in the training set of the k-th candidate parameter group, respectively. To prevent extremely small positive numbers with a denominator of zero; (3) Calculate the macro average index, process-level hit rate and process robustness index, specifically: calculate the process-level grouping extrapolation verification accuracy of the k-th candidate parameter group in the q-th hailstorm process. Process-level grouping extrapolation validation recall rate Verification of F1 value by process-level grouping extrapolation , represented as: , , ; Calculate the accuracy of the k-th candidate parameter group within the process-level training set during the q-th hailstorm process. In-set recall at the process level F1 score within the process-level training set , represented as: , , Based on the process-level indices of each hailstorm process, the macro-average precision, macro-average recall, and macro-average F1 score of the grouped extrapolation validation of the k-th candidate parameter group are calculated and expressed as follows: , , Calculate the macro-average precision, macro-average recall, and macro-average F1 score within the training set of the k-th candidate parameter group, expressed as: , , The process-level hit rate of the k-th candidate parameter group. Represented as: Process robustness metrics include minimum process recall. Minimum process accuracy and process recall standard deviation ,in, , and These represent the number of true positive samples, false positive samples, and false negative samples, respectively, for the group extrapolation verification of the k-th candidate parameter group during the q-th hailstorm process. , and These represent the number of true positive samples, false positive samples, and false negative samples in the training set for the k-th candidate parameter group during the q-th hailstorm process, respectively; M is the total number of hailstorm processes participating in the evaluation. The total number of consecutive periods of real positive samples; The number of consecutive time periods in which the grouped extrapolation verification warning sequence generated by the k-th candidate parameter group hits the true positive samples; The set of hailstorm processes that contain at least one positive sample in the actual warning labels; ;q=1,2,…,M;k=1,2,…,K; (4) Based on the k-th candidate parameter group The generated grouped extrapolation verification warning sequence is used to calculate the false alarm rate of negative samples. Timing penalty term and state transition penalty term The time-series penalty items include penalties for exceeding the warning lead time limit. Punishment for traveling too far in advance Early warning and punishment Even after hail, warnings and penalties will still be in effect. False alarm penalties for hailstorms occurring far from the target ; This is a set of hailstorm processes that do not contain positive samples in the actual warning label sequence. , for The number of hailstorm events in the region, , The grouped extrapolation verification warning flag generated for the k-th candidate parameter group during the q-th hailstorm process and at the t-th radar volume scan time is I(.), which is an indicator function that takes the value 1 when the condition is met and 0 otherwise; the penalty for the warning lead time of the q-th hailstorm process exceeding the warning time limit is... Punishment for traveling too far in advance Early warning and punishment Even after hail, warnings and penalties will still be in effect. False alarms during hailstorms far from the target and number of state transitions , This represents the lead time from the t-th radar volume scan time during the q-th hailstorm process to the next actual hailfall time; when there is no correlated actual hailfall time after the current radar volume scan time t. =+∞; H represents the upper limit of the warning time limit; when the t-th radar volume scan time within the q-th hailstorm process falls within the actual hail observation period, then =1, otherwise =0; (5) The negative sample false alarm rate, time-series penalty term, and state transition penalty term are used as deduction terms in the comprehensive objective function to calculate the deduction value, which is expressed as: ; (6) Calculate the target gap term, generalization difference term, precision-recall balance, target overall distance, and warning ratio penalty term, specifically: based on the target precision... and target recall rate Calculate the difference in accuracy between the grouped extrapolation verifications for the k-th candidate parameter group. recall rate gap In-training set accuracy gap The gap between recall rate and , represented as: ; ; ; Based on the differences between the in-training set metrics and the group extrapolation validation metrics, calculate the precision generalization difference. Recall rate generalization difference Difference in generalization with F1 score , represented as: ; ; ;Calculate the balance between precision and recall in group extrapolation validation. And the balance between precision and recall within the training set. , represented as: , ; Calculate the combined distances between the grouped extrapolation validation results and the training set results of the k-th candidate parameter group relative to the target precision and target recall, respectively, expressed as: ; ; Calculate the prediction and warning ratio of the k-th candidate parameter group during the q-th hailstorm process. The proportion of true positive samples in the q-th hailstorm process , represented as: ; Based on the predicted warning ratio and the actual positive sample ratio, calculate the penalty for an excessively high warning ratio in the q-th hailstorm process. ; Based on the penalties for excessively high and low warning ratios for each hailstorm process that contains at least one positive sample in the actual warning labels, calculate the average penalty for excessively high warning ratios for the k-th candidate parameter group. Penalties for excessively low average warning rate Deviation from maximum warning ratio , represented as: ; ; ;in, and These are target precision and target recall, respectively. The set of hailstorm processes that contain at least one positive sample in the actual warning labels; , where n is the number of hailstorm events contained in the set; q The number of radar volume scan times during the q-th hailstorm process; The grouped extrapolation verification warning flag generated for the k-th candidate parameter group at the t-th radar volume scan time of the q-th hailstorm process; The actual warning label corresponding to the t-th radar volume scan time of the q-th hailstorm process; and The values ​​are all 0 or 1; (7) Based on the aforementioned evaluation indicators and penalty terms, construct the k-th candidate parameter group. The comprehensive objective function , represented as: ; (8) For candidate parameter sets Calculate the integrated objective function separately The candidate parameter sets are then compared from highest to lowest based on their comprehensive evaluation scores, and the set that maximizes the comprehensive objective function is selected as the optimal candidate parameter set. , represented as: Optimal candidate parameter group index k * ; where k * This represents the index of the candidate parameter group corresponding to the optimal candidate parameter group.

5. The hailstorm warning method under complex terrain conditions according to claim 4, characterized in that, In S7, the specific conditions for tracking start / stop are: when the same hail cloud object is detected in two consecutive adjacent radar volume scans within the combined reflectivity of the same radar. and When all conditions are met, the maximum combined reflectivity echo intensity is greater than or equal to 40 dBZ, and the hail cloud object is determined to be included in the monitoring and tracking; when the hail cloud object is within three consecutive adjacent radar volume scan times in the combined reflectivity product of the same radar, the hail cloud object is considered to be included in the monitoring and tracking. , and When both conditions are met that "the maximum combined reflectivity echo intensity is less than 40 dBZ", a warning is issued to stop tracking the hail cloud object; or when the target matching results based on two adjacent radar volume scan times indicate that the hail cloud object has dissipated or cannot be continuously tracked, a warning is issued to stop tracking the hail cloud object.

6. The hailstorm warning method under complex terrain conditions according to claim 5, characterized in that, The target matching results are specifically analyzed as follows: based on the spatial position continuity, area overlap, echo intensity continuity, and motion direction continuity of hail cloud objects in two adjacent radar volume scan times, the matching judgment results are made on whether hail cloud objects in two adjacent radar volume scan times belong to the same tracking object. If the hail cloud objects in two adjacent radar volume scan times meet the preset matching conditions, they are determined to be the same tracked hail cloud object; If the preset matching conditions are not met, it is determined that the hail cloud object has dissipated or cannot be continuously tracked. The preset matching condition is that the combined reflectivity core area of ​​the hail cloud object at the previous radar volume scan time and the hail cloud object at the current radar volume scan time has spatial overlap, or the centroid distance is not greater than a preset distance threshold; wherein, the combined reflectivity core area is a set of radar libraries with a combined reflectivity of not less than 40 dBZ.

7. The hailstorm warning method under complex terrain conditions according to claim 6, characterized in that, S8 includes: S8001. Based on the final input feature set F determined in the training and optimization phase, extract the input features corresponding to the final input feature set F from the candidate feature set corresponding to the current radar volume scan time t, and form the final input feature vector of the current radar volume scan time t. S8002 and m logistic regression sub-models analyze the probability of the first warning based on the final input set F. , represented as: ; The current radar volume scan time t is the first... One input feature; and , respectively, are the intercept term and regression coefficient of the r-th logistic regression sub-model; d is the final input feature dimension; S8003, weighted fusion of m first early warning probabilities at the integrated layer Probability of obtaining a second warning , , Let be the ensemble weights corresponding to the r-th logistic regression sub-model, and satisfy . ; S8004, physical gating mechanism layer based on second early warning probability A probability channel is constructed, and a physical gating channel is constructed based on the feature derivation quantities related to the hail formation physical process. Then, the warning judgment parameters are output based on the probability gating channel and the physical gating channel.

8. The hailstorm warning method under complex terrain conditions according to claim 7, characterized in that, S8004 includes: ①Based on the second early warning probability Perform exponential time smoothing to obtain the smoothed early warning probability at the current radar volume scan time t. ; , Dynamic smoothing coefficient , The second early warning probability at the first radar volume scan moment; Let be the smoothed early warning probability at the (t-1)th radar volume scan time. This represents the actual time interval between the current radar volume scan time t and the (t-1)th radar volume scan time. ② Smooth the early warning probability As the absolute probability of the current radar volume scan time t Based on the number of effective radar volume scan samples within the probability historical background window and effective time span Calculate the probability of valid historical context flags ; indicates as: ;when When =1, let the earliest effective radar volume scan time within the probability history background window be... ; and calculate the median absolute deviation of the smoothed early warning probability within the historical background window. u and v are both historical radar volume scan times within the probability history background window; , ; The length of the probability history background window before the current radar volume scan time t. A set of historical radar volume scan times that effectively smooths out early warning probabilities; ③ Constructing absolute probability channels Probability Anomaly Channel Probability Increment Channel Probability sorting channel The channel values ​​for the microphysics enhancement channel, graupel enhancement channel, trend enhancement channel, and updraft enhancement channel, as well as the absolute probability channel, probability anomaly channel, probability increment channel, and probability sorting channel, are respectively... , , and ;when season , , ;when When =1, the channel values ​​of the microphysics enhancement channel, graupel enhancement channel, trend enhancement channel, and updraft enhancement channel are... for , or mixed value ; indicates as: , , To normalize robust outliers; Indicates input variables Truncate to a closed interval between 0 and 1; ④ Compare the absolute probability channel, probability anomaly channel, probability increment channel, and probability ranking channel with their respective high thresholds to obtain the high threshold vote count VH(t) at the current radar volume scan time t, expressed as: ; ⑤ Compare the absolute probability channel, probability anomaly channel, probability increment channel, and probability ranking channel with their respective low thresholds to obtain the low threshold vote count VL(t) at the current radar volume scan time t; expressed as: ; ⑥ Calculate the valid flags of the physical gating context , represented as: ; and These represent the number of valid radar volume scan samples and the valid time span within the physically gated historical background window, respectively. This represents the cumulative number of valid radar volume scan samples obtained from the current hail cloud object since it entered the monitoring and tracking phase, up to the current radar volume scan time t; when When gate_valid(t) = 0, proceed to step ⑦; when gate_valid(t) = 0, calculate the increase in absolute probability. and short-process rescue signs , , ;when At that time, the physical gating condition is established. , , , and , directly enter ⑩; when season , , , , and Proceed directly to step 11; ⑦ The microphysics enhancement channel, graupel enhancement channel, trend enhancement channel, and updraft enhancement channel are compared with their corresponding effective hit thresholds. The number of physics-gated hits is expressed as: ; ; ⑧ Calculate the normalized gating scores of the micro-physics enhancement channel, gra-like enhancement channel, trend enhancement channel, and updraft enhancement channel relative to the effective hit threshold; Channel value for hour, Channel value for hour, Channel value For mixed values hour, ;r represents micro, gra, trend, or up; ⑨ Based on the normalized gating scores of the microphysics enhancement channel, graupel enhancement channel, trend enhancement channel, and updraft enhancement channel, calculate the physics consistency score phys_score(t), and then adjust the high-threshold vote count and low-threshold vote count according to the physics consistency score to obtain the high-threshold composite vote count. and low threshold combined vote count , represented as: ; ; ; ⑩ When If the long process sample count threshold long_proc_n has not yet been reached, set the long process start protection period flag. =0, in After reaching long_proc_n, let , for The first time the long-process sample number threshold has been reached The radar volume scan time was calculated. Calculate the number of long-process start-up protection hits. And calculate the long-process start-up protection inhibition flag. ; ⑪ , , , The probability channel value, physical gating channel value, and early warning status timing parameters are used as early warning judgment parameters and then processed in S9.

9. The hailstorm warning method under complex terrain conditions according to claim 8, characterized in that, Specifically, S9 includes: S9001, Determine the previous warning state at the current radar volume scan time; if the current radar volume scan time is the first valid radar volume scan time for the current hail cloud object, initialize the previous warning state to no warning state; otherwise, determine the previous warning state based on the final warning flag of the previous valid radar volume scan time; when G(t)=1, proceed to S9002; when season It then determines whether the previous warning state was a warning state; if not, it sets the initial warning state to... If so, then determine whether the time interval between the current radar volume scan time t and the most recent effective hold time is greater than the early warning hold time window. Or whether the duration of consecutive warnings exceeds the maximum duration of consecutive warnings. If so, then let Otherwise And directly enter S9006; S9002. Analyze the high-threshold trigger flag at the current radar volume scan time t. and low threshold retention marker , represented as: ; When the current warning state is no warning state, proceed to S9003; when the current warning state is a warning state, first determine whether the duration of continuous warning exceeds the maximum duration of continuous warning max_on_min; if yes, set Y0(t)=0; if no, and L(t)=1, set Y0(t)=1; if no, and L(t)=0, then determine whether the time interval between the current radar volume scan time t and the most recent effective hold time is greater than hold_min; if yes, set Y0(t)=0, otherwise set Y0(t)=1; and directly proceed to S9006. S9003. Construct a set of early warning entry confirmation times with the current radar volume scan time t as the endpoint. And statistically confirm the time window The effective hit count C(t) of the high-threshold trigger flag is expressed as: , ;judge If the condition is not met, set Y0(t) = 0 and proceed directly to S9006; if it is met, proceed to S9004; effectively confirm the hit count threshold. The warning enters the confirmation time set. Number of effective radar volume scan times ; S9004. Analyze the effective flag of repeated triggering cooling at the current radar volume scan time t. And the effective indicator of cooling after actual hail. h (t); when there is no radar volume scan time at which the most recent exit from the warning state occurred, let Cool r (t)=1; when there is a radar volume scan time t that most recently exited the early warning state. off hour, When there is no most recent actual hail observation time, let Cool h (t)=1; when there is a most recent actual hail observation time t hail hour, ;in, This refers to the radar volume scan time when the radar most recently exited the early warning status. The time of the most recent actual hail observation is indicated by ∧, which represents the logical AND. Represents the logical "OR", and R(t) is a strong restart condition flag. ; S9005, The current warning status is no warning status, and , season ;when or season ; S9006. Perform time-by-time weak alarm assessment on the continuous preliminary warning segments up to the current radar volume scan time t; when... At that time, the final warning sign was issued. And enter S9007; when At that time, build And calculate the comprehensive alarm score.

10. , This is the starting radar volume scan time for the current continuous preliminary warning phase; , Based on the comprehensive alarm score ,calculate ; Retain signs for continuous preliminary warning sections Final warning sign ; S9007. Record the warning state transition time based on the previous warning state and the final warning flag Y(t) of the current radar volume scan time; when the final warning flag changes from 0 to 1, the analysis result is a warning state, a new warning flag is generated, the current radar volume scan time t is recorded as the start time of this continuous warning, and the most recent valid hold time is initialized to the current radar volume scan time t; when the final warning flag remains at 1, the analysis result remains a warning state, the start time is not re-recorded, and only when... The most recent valid hold time is updated to the current radar volume scan time t; when the final warning flag changes from 1 to 0, the analysis result is no warning, and the current radar volume scan time t is recorded as the most recent time of exiting the warning state. When the final warning indicator remains at 0, it will not be updated. Return to S6.

Images