A hail event quality control calibration method and system

Abstract

The application discloses a hail event quality control calibration method and system, first acquires hail related multi-source data and original saves management, gets a suspected hail clue set through clue correlation degree preliminary screening, extracts hail event elements of the suspected hail clue, retains extraction confidence and traceability evidence, encapsulates into network evidence objects after completing space-time normalization, constructs an evidence propagation diagram, divides original source groups, completes candidate event clustering, generates and matches a meteorological support area, executes hard conflict rule gating to identify non-supplementable hard conflicts, calculates comprehensive confidence again, combines gating processing label to output quality control level, finally completes quality control calibration under rule constraints, and outputs credible hail event records. The application fuses network multi-source data and meteorological data, distinguishes soft scores and non-supplementable hard conflicts, reduces misjudgment and missed reports, and realizes objective and credible automatic quality control calibration of hail events.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology for meteorological disaster information, specifically a quality control calibration method and system for hail events. Background Technology

[0002] Hail is characterized by its sudden occurrence, small impact area, short duration, and discontinuous spatial distribution. Weather radar can detect strong convective cells and atmospheric water vapor, but it is limited by factors such as radar beam height, hail melting process, and sparse ground observation stations. Relying solely on professional meteorological observation data makes it difficult to accurately determine the actual location of hail on the ground. Social media, short videos, news websites, government platforms, and manual reporting channels can provide a large amount of on-site text and video data, which can supplement information such as the fact of hail on the ground, local damage, and hail particle size. However, this type of multi-source online data is not directly usable truth value of the event and has many problems, such as the release time being later than the time of the disaster, inconsistencies between the account's location and the actual shooting location, repeated reposting and dissemination of materials, secondary diffusion of historical disaster materials in new convective processes, and strong subjectivity in textual descriptions.

[0003] Existing processing solutions mostly use keyword retrieval or simple weighted scoring to process multi-source data, which has many limitations: (1) The data structures of online posts, media materials and meteorological grid points are different, and it is impossible to directly establish a calculable and traceable relationship; (2) Multiple reposts of the same material will cause the number of evidence to be inflated, and the credibility will be calculated by accumulating the number of posts; (3) The shooting, posting and actual occurrence time of online materials are mixed, and the location description is vague, aliased, cross-administrative region and account location misleading; (4) A single weighted formula is difficult to handle the irreparable hard conflict such as old news from other places and complete time conflict; (5) The radar support area and the network candidate location have different scales, and simple point-to-point distance cannot reflect the movement path and positioning uncertainty of the individual; (6) There is no calibration mechanism to retain uncertainty between the hail particle size, damage degree and meteorological intensity index described in the network; (7) The final event result lacks a complete chain of evidence, and it is impossible to trace the basis for the formation of the event confirmation, denial or calibration results. Furthermore, strong radar echoes, dual-polarization hail identification, or lightning surges can only characterize the degree of support for hail or strong convection in the air, and cannot prove that actual hail fell on the ground in all cases. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a hail event quality control calibration method and system, which converts multi-source data on the network into computable network evidence objects, and constructs a multi-layered evidence chain by utilizing the same source relationship, source independence, professional meteorological support and authoritative ground records; performs hard conflict rule gating before credibility calculation, and completes hail event calibration under the premise of retaining time, space and intensity uncertainties, and finally forms a credible hail event record that can be used for meteorological review, disaster assessment, model verification and business services.

[0005] The technical solution of this invention is as follows: A quality control calibration method for hail events specifically includes the following steps: (1) Obtain multi-source data related to hail events on the Internet and complete the original storage and management of multi-source data; (2) Calculate the correlation of clues from multiple sources, and identify a set of suspected hail clues after initial screening; (3) Extract hail event elements from each suspected hail clue in the suspected hail clue set, retain the extraction confidence of hail event elements and the corresponding multi-source tracing evidence, then perform spatiotemporal normalization processing on hail event elements, and finally encapsulate the normalized hail event elements of each suspected hail clue, their extraction confidence and the corresponding multi-source tracing evidence into a network evidence object. (4) Based on all network evidence objects corresponding to the suspected hail clue set, calculate the multi-dimensional evidence similarity pairwise. When the multi-dimensional evidence similarity meets the same source threshold requirement, establish a propagation edge to complete the construction of the evidence propagation graph. Based on the propagation edge, divide the data into multiple original source groups. Then, for all network evidence objects corresponding to the suspected hail clue set, cluster candidate events based on spatiotemporal, semantic, evidence consistency, and same source or conflict penalty to generate candidate hail events. (5) Generate and match meteorological support areas for candidate hail events, and at the same time perform hard conflict rule gating on candidate hail events, identify non-replaceable hard conflicts and output corresponding processing labels; (6) After calculating the credibility of network evidence, meteorological support, multi-source cross-validation gain and conflict penalty of candidate hail events, the comprehensive credibility is calculated by weighting and the quality control level is output by combining the processing label output by the hard conflict rule gating. Among them, when a hard conflict that cannot be supplemented is hit, the corresponding processing result takes precedence over the comprehensive credibility and cannot be offset by the credibility of network evidence, meteorological support or multi-source cross-validation gain. (7) Perform quality control calibration on the time, location and intensity of candidate hail events under the constraints of the rules, retain the original information of candidate hail events, and generate a credible hail event record containing the chain of evidence, quality control level, uncertainty range and calibration instructions.

[0006] The multi-source data related to hail events on the network includes text, titles, tags, images, videos, comments, and metadata related to hail events obtained from authorized or compliant data interfaces, public web pages, government announcements, news media, manual reports, and business systems. The original preservation and management of the multi-source data specifically involves assigning a unique original record identifier to each piece of collected data, completely preserving the original information, and only adding to the original storage layer without modifying it. Subsequent clue identification and event calibration results are associated with the original record identifier through the version number, preserving complete traceability evidence.

[0007] The process of calculating the relevance of clues from multi-source data, and identifying a set of suspected hail clues after initial screening, specifically involves: S21. Construct a keyword dictionary and calculate keyword scores for text data from multiple sources using the following formula (1). ; (1); In equation (1), The first in the representative keyword dictionary The weight of each term is determined by the fact that the keyword dictionary includes terms of different categories, and different weights are assigned to terms of different categories, including core terms of hail, phenomenon terms, particle size terms, and disaster damage terms. This represents the weight of different text fields extracted from the text data, specifically the title, body text, tags, or comments. Represents a binary indicator variable. Representing the Each term appears in the text field. Representing the The term did not appear in the text field; This represents iterating through all entries and all text fields, and accumulating keyword scores; This represents a truncation function, with a maximum keyword score of 1. S22. Calculate the relevance of the clues according to the following formula (2). : (2); In equation (2), Represents semantic confidence, specifically the probability that the content output by the semantic classification model describes a real hail event; This represents the media confidence score, specifically the confidence score given by the media recognition model when it identifies image targets such as hail particles, hail accumulation, broken vehicle windows, and damaged crops. When no image target is identified, the confidence score is lower. ; This represents a reverse rule penalty; when low-relevance content such as science popularization, historical reviews, film and television content, reposting from other locations, and weather forecasts is identified, Take the penalty term that is greater than 0; , and These are non-negative weighting coefficients. The contribution ratios of keywords, semantic classification models, and media evidence in the calculation of clue relevance are controlled respectively. This represents the penalty coefficient, used to control the severity of the penalty for the reverse rule; The pruning function forces the calculation result to be restricted to... The interval; S23, Clue Relevance Not lower than the relevance threshold Multi-source data is incorporated into the suspected hail clue set; relevance threshold. The F1 score, recall lower limit, or business cost function on historical labeled samples are used to determine the optimal value.

[0008] The specific steps for extracting the hail event elements from each suspected hail clue in the suspected hail clue set and encapsulating them into a network evidence object are as follows: S31. Extracting Hail Event Elements: When the suspected hail clue is text, a rule base and named entity recognition model are used to extract entities with time, location, particle size, damage, organization, and negative relation; when the suspected hail clue is an image or video, an image classification model, target detection model, or video keyframe model are used to identify hail and damage targets; the extracted entities and the identified hail and damage targets together constitute the hail event elements; For each hail event element, save source metadata: for the extracted entities, save the original text fragment, the character position of the entity, the version number of the rule base and named entity recognition model, and the extraction confidence; for the identified hail and damage targets, save the media identifier, keyframe number, detection box coordinates, the version number of the image classification model, target detection model or video keyframe model, and the extraction confidence. S32. Time normalization processing of hail event elements: Convert the time entities in the hail event elements into a candidate time window or a set of candidate time windows. For each candidate time window, output the event start time, event end time, time precision level, time confidence level, and time basis. For each candidate time window, the overall credibility score is calculated using the following formula (3): (3); In equation (3), Representing the The overall credibility score of each candidate time window, Representing the The contextual evidence score for each candidate time window is calculated based on the temporal consistency between the candidate time window and the original text context. Representing the The media metadata time evidence score for each candidate time window is calculated based on the overlap between the candidate time window and the timestamps in the image and video data. Representing the The cross-evidence score for each candidate time window is calculated based on the degree of overlap between the candidate time window and the time windows of other resolved hail cues. Representing the The meteorological matching support score of each candidate time window is calculated based on the degree of matching between the candidate time window and the actual strong convection records of the meteorological support area. , , and Represents the weight coefficient of each item; each weight coefficient is a non-negative number, and the sum of all valid item weights is 1; if there are unusable items, the corresponding items are removed and the remaining valid weights are renormalized; Based on the comprehensive credibility score of the candidate time windows, multiple candidate time windows in a set of candidate time windows are sorted. If the comprehensive scores of multiple candidate time windows are close to each other, they are not eliminated and all candidate time windows are retained. Before the meteorological support area is generated, when calculating the comprehensive credibility score of the candidate time window, the meteorological matching support score is not calculated, and the weights of the remaining valid sub-items are renormalized; after the meteorological support area is generated, the meteorological matching support score is introduced, the comprehensive credibility score of the retained candidate time windows is recalculated, and the ranking is updated a second time. S33. Spatial normalization of hail event elements: First, the location entities in the hail event elements are analyzed and recalled according to the administrative division database, place name dictionary and POI database to generate a candidate location or a set of candidate locations; for each candidate location, the positioning information is specified, and the latitude and longitude points and error radius are output. For fuzzy location entities without corresponding latitude and longitude points, the geographic boundary vector polygon and spatial accuracy level are output. For each candidate position, the overall credibility score is calculated using the following formula (4): (4); In equation (4), Representing the The overall credibility score of each candidate position. Representing the The text name matching score of each candidate location is obtained by calculating the text similarity between the candidate location and the standard place name of the candidate location in the administrative division database and place name dictionary. Representing the The context score of each candidate location is calculated by matching the candidate location with the context location description in the original text. Representing the The POI matching score for each candidate position is calculated by matching the candidate position with the matching objects in the POI library. Representing the The cross-clue evidence score of each candidate location is calculated by spatially matching the candidate location with the other hail clue locations. Representing the The meteorological region overlap score of each candidate location is obtained by spatially matching the candidate location with the meteorological support area. , , , and Represents the weight coefficient of each item; each weight coefficient is a non-negative number, and the sum of all valid item weights is 1; if there are unusable items, the corresponding items are removed and the remaining valid weights are renormalized; The candidate positions in a set of candidate positions are ranked according to their comprehensive credibility scores. Before the meteorological support area is generated, when calculating the comprehensive confidence score of the candidate locations, the meteorological region overlap score is not calculated, and the weights of the remaining valid sub-items are renormalized; after the meteorological support area is generated, the meteorological region overlap score is introduced, the comprehensive confidence score of the retained candidate locations is recalculated, and the ranking is updated a second time. S34. For suspected hail clues that have been extracted from hail event elements and have undergone time and space normalization, if there are independent images, videos or comments on the same webpage corresponding to them, the suspected hail clues are split into multiple sub-objects, each of which is a network evidence object; otherwise, a corresponding network evidence object is directly generated for the suspected hail clues. The suspected hail clue before splitting is used as the parent original clue. The parent original clue and its corresponding multiple child objects are associated. The parent source identifier of the child object is taken as the unique evidence identifier of the parent original clue. The network evidence objects are assembled using fixed fields, which include a unique evidence identifier, parent original clue record number, source type, content carrier type, publication time, parent source identifier, candidate time window, candidate spatial region, hail size range, disaster impact label, media fingerprint, original source link, hail event element extraction confidence level, original source group number, conflict marker, and processing version.

[0009] The specific steps for generating candidate hail events are as follows: First, based on all network evidence objects corresponding to the suspected hail clue set, multi-dimensional evidence similarity is calculated pairwise. When the multi-dimensional evidence similarity meets the same-origin threshold requirement, a propagation edge is established to complete the construction of the evidence propagation graph. Multiple original source groups are then obtained based on the propagation edges. Next, for all network evidence objects corresponding to the suspected hail clue set, candidate events are clustered based on spatiotemporal, semantic, evidence consistency, and same-origin or conflict penalties. S41. Construct an evidence propagation graph G=(V,E), where the nodes in the node set V are network evidence objects, and the edge set E consists of propagation edges; For any two network evidence objects, the multi-dimensional evidence similarity is calculated using the following formula (5). : (5); In equation (5), Representative node and nodes Multidimensional evidence similarity between them Representative node and nodes Text similarity between them Representative node and nodes Perceptual hash similarity between images Representative node and nodes Similarity of keyframes between videos Representative node and nodes Audio fingerprint similarity between them Represents the similarity of link references. , , , and The weight coefficients represent the weights of each dimension. When two nodes belong to the same content modality, the similarity corresponding to that content modality and the link reference similarity are calculated, and the weighted sum of the similarity items actually involved in the calculation is performed. When two nodes belong to different content modalities, the link reference similarity is used as the multi-dimensional evidence similarity. The weight coefficients corresponding to the similarity items not involved in the calculation are not included in this calculation. The effective weight coefficients corresponding to the similarity items actually involved in the calculation are renormalized so that the sum of the effective weight coefficients is 1. when Not lower than the set homogeneity threshold When the two nodes are connected, a propagation edge is established between them and added to the edge set E. After completing the multi-dimensional evidence similarity calculation and edge construction for all network evidence objects, perform connected component partitioning or community discovery on the evidence propagation graph to obtain multiple original source groups; each original source group is generated based on graph connectivity, which includes evidence pairs that meet the multi-dimensional evidence similarity criteria, as well as capturing the indirect propagation homologous evidence set formed by chain retransmission; The credibility of all network evidence objects within each original source group is aggregated to obtain the aggregate credibility of that original source group, and an equivalent evidence object is generated. During credibility aggregation, the aggregate credibility of the original source group is equal to the maximum base credibility of all network evidence objects within the original source group, or the aggregate credibility obtained by attenuation and superposition according to the propagation distance. The equivalent evidence object is a virtual evidence record, and its fixed fields are selected from the network evidence object corresponding to the maximum base credibility or the fields obtained by fusion operation of the fields of all network evidence objects within the original source group. If there is only one network evidence object in the original source group, and this network evidence object is different from other network evidence objects describing the same hail event in terms of account, shooting angle, media fingerprint, and spatiotemporal information describing the event, then it is regarded as an independent evidence object, and its basic credibility is retained as the aggregate credibility of the original source group. The basic credibility of the network evidence object is calculated using the following formula (6): (6); In equation (6), Representing the The credibility of the source of each piece of online evidence is derived from the source type; Representing the The credibility of the evidentiary form of an individual network evidence object is derived from the type of content carrier; Representing the The reliability of the temporal precision of each network evidence object is derived from the candidate time window; The reliability of the location accuracy is derived from the candidate spatial region; Representing the The originality and credibility of each piece of online evidence is obtained by comparing and retrieving media fingerprints and original source links; Representing the The conflict penalty for each network evidence object is derived from the conflict marker; Representing the The punishment for reposting old news by an individual online evidence subject is derived from media fingerprints and historical material databases retrieved based on publication time. , , , , , and Represents the weighting coefficients of each item; The pruning function forces the calculation result to be restricted to... The range; among which, source credibility, evidence form credibility, time accuracy credibility, location accuracy credibility and originality credibility are positive sub-items, each sub-item weight is a non-negative number, and the sum of the sub-item weights is 1; the weights of conflict penalty and old news reposting penalty are set separately. S42. Calculate the event association score using the following formula (7). : (7); In equation (7), Representative object and object The event correlation score is based on whether the object is an equivalent evidence object or an independent evidence object; It represents the temporal similarity between two objects, that is, the degree of overlap between the candidate time windows of the two objects; It represents the spatial similarity between two objects, that is, it calculates the degree of overlap or distance decay between the candidate spatial regions of the two objects; This represents the semantic similarity between two objects, specifically calculating the matching similarity of the semantic tags of the hail event elements of the two objects. It represents the consistency of evidence between two objects, that is, the degree of agreement between the values ​​of structured elements in the hail event elements of the two objects; The same-origin or conflict penalty between two objects is calculated from the original source group number and conflict flag of the two objects; , , , and The weight coefficients represent the weights of each item; among them, temporal similarity, spatial similarity, semantic similarity, and evidence consistency are positive items, each item's weight is a non-negative number, and the sum of the item weights is 1; the weights for homology or conflict penalties are set separately. When the event is associated with the score Not less than the clustering threshold If the hard conflict conditions of being geographically distant and having completely disjoint dates are not met, then an edge with the same event is established between the two objects, and all objects are grouped and clustered according to all edges with the same event by connected component partitioning, hierarchical clustering or DBSCAN algorithm to obtain multiple candidate event clusters. The aggregated confidence scores of all objects within each candidate event cluster are read and fused to obtain the event confidence score of that candidate event cluster; the structured information of all objects within each candidate event cluster is aggregated to obtain the candidate hail events.

[0010] After obtaining multiple candidate event clusters, post-verification processing is performed: the maximum time span and maximum spatial span among all objects within the candidate event cluster are calculated, and the consistency of meteorological units corresponding to the candidate event cluster is verified; when the maximum time span and maximum spatial span exceed preset thresholds, or when meteorological units are inconsistent, the candidate event cluster is split a second time to obtain split candidate event clusters; the calculation of event confidence and the aggregation of candidate hail events are performed after post-verification processing.

[0011] The specific steps for generating and matching meteorological support zones for candidate hail events, applying hard conflict rule gating to candidate hail events, identifying irreparable hard conflicts, and outputting corresponding processing labels are as follows: S51. Obtain meteorological data through the meteorological operational data interface, and convert the meteorological data into various meteorological support areas, which are divided into radar support areas, dual-polarization support areas, path support areas, and surface support areas; calculate the time interval overlap between the candidate time windows of candidate hail events and the corresponding time slices of the meteorological support areas; calculate the area intersection-union ratio, nearest distance, and path relationship between the candidate spatial regions of candidate hail events and the meteorological support areas; combine the time interval overlap calculation results and the area intersection-union ratio, nearest distance, and path relationship calculation results into the spatiotemporal matching result between the candidate hail events and the meteorological support areas. S52. Execute hard conflict rule gating: Perform hard conflict rule matching for each candidate hail event, determine whether each rule meets the triggering conditions, and execute the corresponding handling action when the triggering conditions are met. Finally, summarize the handling results of all rules and output the handling flags. The handling flags include exclusion flags, downgrade flags, manual review flags, and conflict tags. Conflict tags include supplementary conflict tags and non-supplementary hard conflict tags.

[0012] After calculating the credibility of network evidence, meteorological support, multi-source cross-validation gain, and conflict penalty for candidate hail events, a weighted average is calculated to obtain the comprehensive credibility. This is then combined with a quality control level output using a processing label gating system based on hard conflict rules. The specific steps are as follows: S61, Original source group Aggregate credibility Substituting into equation (8), the baseline value of the candidate hail event is calculated. ; (8); In equation (8), Representing the original source group Weighting coefficients; The strength of multi-source mutual verification is calculated using the following formula (9). : (9); In equation (9), This represents the total number of original source groups corresponding to the current candidate hail event. It is an integer not less than 1; Represents the saturation coefficient. Greater than 0 controls the rate at which saturation occurs; when When =1, the strength of multi-source mutual verification is 0; as As the value increases, the strength of multi-source mutual verification monotonically increases and gradually approaches 1; the strength of multi-source mutual verification... Represents the strength of mutual corroboration among the original source groups, with a value range of [value range missing]. ; Finally, the credibility of the network evidence for the candidate hail event is calculated using the following formula (10). : (10); In equation (10), The weighting coefficients representing the strength of multi-source mutual verification. Weighting coefficients representing the baseline values ​​of candidate hail events; using The function constrains the fusion result to The interval; S62. Calculate the meteorological support of candidate hail events according to the following formula (11). : (11); In equation (11), Used to measure the degree of support from professional meteorological data for the existence of hail or severe convective conditions in the candidate spatial region of a candidate hail event. The time matching degree is represented by calculating the time distance between the meteorological observation time and the time window of the candidate hail event. The spatial matching degree is obtained by calculating the spatial distance or the percentage of spatial overlap between the meteorological observation location and the spatial region of the candidate hail event. The radar hail support is represented by the radar-identified hail potential calculated in the spatial region of the candidate hail event. This represents the dual polarization support, obtained by calculating the significance of hail features; The score represents the support for lightning, which is obtained by statistically analyzing the frequency and distance of lightning within the spatial region of the candidate hail event. The higher the frequency of lightning and the closer the observation location is to the spatial region of the candidate hail event, the higher the score. The satellite's convective support is obtained by extracting convective features and identifying strong convective potential. The weather support level of the automatic weather station is calculated by extracting environmental meteorological features that are conducive to the occurrence of hail events and giving a comprehensive score. Representing ground record support, this is determined by statistically analyzing whether there are actual records of hail in the spatial area of ​​the candidate hail event; the more records, the higher the score. , , , , , , and The weight coefficients represent the weights of each item. These weight coefficients are global hyperparameters and are optimally determined from the historical hail validation set. When all data sources are valid, the sum of all weight coefficients equals 1. When a meteorological data source is missing, the weight coefficients corresponding to the missing items are removed, and the remaining valid weight coefficients are renormalized. At the same time, a data availability identifier is output to mark the availability status of each sub-item data source. S63. Calculate the multi-source cross-validation gain using the following formula (12). : (12); In equation (12), The sub-item consistency score represents the pairwise comparison results between two original source groups for the same candidate hail event. The comparison types include cross-platform comparison, cross-user comparison, cross-media comparison, and cross-category source comparisons conducted by the network multi-source group with radar, radar dual polarization, and ground real-time records, respectively. Each comparison outputs a corresponding sub-item consistency score. The gain coefficient representing the consistency score of each sub-item is a hyperparameter; subscript An index representing the consistency score of each item; each set of comparisons corresponds to one. ; The gain effect of cross-validation of information from multiple sources is characterized; only valid comparisons between different original source groups are included in the product operation; when there are no valid comparisons, the gain of cross-validation of multiple sources is 0, and reprints within the same original source group are not counted repeatedly. S64. Calculate the conflict penalty using the following formula (13). : (13); In equation (13), The conflict severity represents the severity of the conflict, which is not supported by old news, time, location, intensity, media or weather. The conflict label is a supplementary conflict label, which is obtained by comparing the conflict situation of multiple sources of information under the same candidate hail event and assigning a value according to the intensity of the conflict. Represents the conflict weight, which is a hyperparameter; Characterizes the comprehensive conflict degree of multi-source information of candidate hail events; non-supplementable hard conflicts are not included in the calculation of equation (13) and are directly processed by hard conflict rule gating; S65. The overall confidence level of candidate hail events is calculated using the following formula (14): (14); In equation (14), , , The weights are non-negative and their sum is 1. This is the conflict penalty coefficient; After calculating the overall credibility of the candidate hail event, the candidate hail event is classified into the corresponding preliminary quality control level according to the credibility level classification principle. Various processing labels are received from the hard conflict rule gating output. The processing labels are used to exclude, downgrade, add manual review mark and limit the highest level of the preliminary quality control level to obtain the final quality control level of the candidate hail event.

[0013] The specific steps for performing quality control calibration on the time, location, and intensity of candidate hail events under rule constraints, retaining the original information of the candidate hail events, and generating a credible hail event record containing a chain of evidence, quality control level, uncertainty range, and calibration instructions are as follows: S71. Time Calibration: When the time range of a candidate hail event intersects with the effective time interval corresponding to the meteorological support area, if there is only one overlapping intersection time period, it is used as the recommended time window after meteorological constraints. If there are multiple overlapping intersection sub-time periods, the meteorological support corresponding to each intersection sub-time period is calculated, and the intersection sub-time period with the highest meteorological support is selected as the recommended time window after meteorological constraints. At the same time, the original time range of the original candidate hail event is retained, and time conflicts are marked to retain the uncertainty of time. When there is no overlap but the time interval is within the set buffer threshold, Bayesian or weighted updates are performed by combining the radar single-unit transit time, the dual-polarization hail identification time, the lightning jump time, and the independent network clue sequence to obtain the recommended time window after meteorological constraints; at the same time, the original time range of the original candidate hail events is retained, and time conflicts are marked to retain the uncertainty of time. When there is no overlap but the time interval exceeds the set buffer threshold, the original time range of the original candidate hail event is retained, the time conflict is marked, the time of the event is not modified, and the uncertainty of the time is retained. S72. Location Calibration: When the candidate spatial region of a candidate hail event overlaps with the meteorological support region, if there is only a single overlapping region, the overlapping region is used as the candidate spatial region after meteorological constraints. If there are multiple discontinuous overlapping sub-regions, the meteorological support corresponding to each overlapping sub-region is calculated, and the overlapping sub-region with the highest meteorological support is selected as the candidate spatial region after meteorological constraints. At the same time, the original candidate spatial region of the original candidate hail event is retained, and location conflicts are marked to retain spatial uncertainty. When the candidate spatial region description of a candidate hail event is ambiguous, but the meteorological cell path information in the meteorological support area is clear, the candidate spatial region of the candidate hail event is narrowed to the intersection of the meteorological cell path, the dual-polarized hail region, and the multiple ground clues corresponding to the candidate hail event, to obtain the meteorologically constrained candidate spatial region; at the same time, the original candidate spatial region of the original candidate hail event is retained, and positional conflicts are marked to preserve spatial uncertainty. When the candidate spatial region of a candidate hail event does not overlap with the meteorological support area, the quality control level of the candidate hail event is reduced, the original candidate spatial region of the original candidate hail event is retained, and the location conflict is marked to retain spatial uncertainty. S73, Intensity Calibration: The hail size descriptions obtained from suspected hail cues for candidate hail events are mapped to initial size intervals with upper and lower boundaries. Based on image scale, video target references, manual observations, and official disaster damage records, the intensity confidence of the suspected hail cues is adjusted. The rationality of the initial size interval is then verified using radar and dual polarization features of the meteorological support area. Finally, the upper and lower limits of hail size, hail intensity level, impact level, and intensity confidence are output. After calibration, the intensity descriptions of the original candidate hail events are retained, and intensity conflicts are marked to preserve the uncertainty of intensity. S74. Based on the constructed evidence propagation graph, the nodes in the evidence propagation graph include original records, evidence objects, original source groups, candidate hail events, time windows, spatial regions, meteorological support areas, ground records, rule actions, and calibration results; the edges in the evidence propagation graph include origin, same source, support, correction, conflict, time matching, spatial matching, and calibrations generated therefrom. Each edge stores the relationship score, rule or model version, generation time, and original data index; thereby tracing back the formation basis of any quality control level, time calibration, location calibration, or intensity calibration, a credible hail event record containing the evidence chain, quality control level, uncertainty range, and calibration description is generated; Among them, the evidence propagation graph is used for the same source identification of network evidence objects; based on the evidence propagation graph, the nodes and corresponding relationship edges corresponding to the rule actions and calibration results are improved and expanded to form a full-process evidence chain graph.

[0014] A hail event quality control calibration system includes a data acquisition and storage module, a clue identification module, an event element extraction module, a spatiotemporal normalization module, an evidence object generation module, a homology identification module, a candidate event clustering module, a meteorological support area matching module, a quality control rule engine module, a credibility assessment module, a quality control calibration module, and an event recording module. The data acquisition and retention module is used to acquire multi-source data related to hail events and to complete the raw storage and management of multi-source data; The clue identification module is used to calculate the relevance of clues from multiple sources, and after initial screening, it identifies a set of suspected hail clues. The event element extraction module is used to extract hail event elements from each suspected hail clue in the suspected hail clue set, and retain the extraction confidence of hail event elements and corresponding multi-source tracing evidence; The spatiotemporal normalization module is used to perform spatiotemporal normalization processing on hail event elements; The evidence object generation module is used to encapsulate the normalized hail event elements of each suspected hail clue, their extracted confidence level, and the corresponding multi-source tracing evidence into a network evidence object. The same source identification module is used to calculate the multi-dimensional evidence similarity pairwise based on all network evidence objects corresponding to the suspected hail clue set. When the multi-dimensional evidence similarity meets the same source threshold requirement, a propagation edge is established to complete the construction of the evidence propagation graph, and multiple original source groups are obtained based on the propagation edge. The candidate event clustering module clusters candidate events based on spatiotemporal, semantic, evidence consistency, and homology or conflict penalty for all network evidence objects corresponding to the suspected hail clue set, and generates candidate hail events. The meteorological support area matching module generates and matches meteorological support areas for candidate hail events. The quality control rule engine module is used to perform hard conflict rule gating on candidate hail events, identify irreparable hard conflicts, and output the corresponding processing labels. The credibility assessment module is used to calculate the credibility of network evidence, meteorological support, multi-source cross-validation gain and conflict penalty of candidate hail events, and then calculate the comprehensive credibility by weighting, and combine the processing label output by hard conflict rule gating to output the quality control level. The quality control calibration module is used to perform quality control calibration on the time, location, and intensity range of candidate hail events under rule constraints; The event logging module is used to generate credible hail event logs that include the chain of evidence, quality control level, uncertainty range, and calibration instructions.

[0015] Advantages of this invention: (1) This invention encapsulates multi-source data related to hail events on the Internet into standardized network evidence objects, standardizes the storage of hail event elements, extracts confidence levels and corresponding multi-source tracing evidence, eliminates the structural barriers of multi-source data, and enables various clues to complete similarity comparison, clustering and confidence aggregation under the same computing framework, so as to realize that heterogeneous data is computable, indexable and traceable.

[0016] (2) The present invention constructs an evidence propagation graph and divides the original source group. When aggregating network evidence objects, it applies the same source or conflict penalty, identifies the reprinted material of the same source propagation, and suppresses the reprinted duplication gain, so that the network evidence credibility calculation of candidate hail events is not affected by the propagation heat, and improves the objectivity of event assessment.

[0017] (3) This invention generates and matches meteorological support areas for candidate hail events. During matching, the candidate spatial area of ​​the candidate hail event is compared with the meteorological support area by area intersection ratio, shortest distance and path relationship calculation, avoiding the problem that simple point-to-point distance judgment cannot reflect the movement path of convective cells and the positioning uncertainty.

[0018] (4) This invention decouples hard conflict rule gating from comprehensive credibility calculation. Hard conflict is given priority as a pre-constraint. Even if the keyword scores and media confidence of multi-source data are very high, once irreplaceable hard conflicts such as old news from other places and complete spatiotemporal contradictions are identified, such conflicts will not be offset by high score weighting, effectively filtering contradictory evidence and avoiding the logical defects of simple weighting models.

[0019] (5) This invention does not output a unique definite value, but retains the candidate time window, candidate spatial region and hail particle size range, and records the uncertainty range simultaneously to prevent the generation of overly accurate false definite results under insufficient evidence.

[0020] (6) The present invention first calculates the credibility of network evidence and meteorological support for candidate hail events separately, and then calculates the overall credibility by weighting. Network credibility represents the direct support of ground-based multi-source clues for hail events, while meteorological support represents that meteorological observations such as radar can only prove that there are environmental conditions for hail to occur in the air. The two are calculated separately to avoid directly equating strong convective signals in the air with the fact of hail falling on the ground.

[0021] (7) The present invention constructs a complete evidence chain based on the node-edge relationship of the evidence propagation graph. Each credible hail event record is associated with the corresponding evidence chain, quality control level, uncertainty range and calibration instructions; the quality control calibration process can be audited, reviewed and iterated. Attached Figure Description

[0022] Figure 1 This is a flowchart of the hail event quality control calibration method of the present invention.

[0023] Figure 2 This is a schematic diagram of the hail event quality control calibration system of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See Figure 1 A quality control calibration method for hail events, specifically including the following steps: (1) Obtain multi-source data related to hail events on the Internet. Multi-source data related to hail events on the Internet includes text, titles, tags, pictures, videos, comments and metadata related to hail events obtained from authorized or compliant data interfaces, public web pages, government releases, news media, manual reports and business systems; and then complete the original preservation and management of multi-source data. Specifically, assign a unique original record identifier to each piece of collected data, retain the original information completely, and only add to the original storage layer without changing it. Subsequent clue identification and event calibration results are associated with the original record identifier through the version number to retain complete traceability evidence. (2) The correlation of clues was calculated from the multi-source data. After initial screening, a set of suspected hail clues was identified, specifically: S21. Construct a keyword dictionary and calculate keyword scores for text data from multiple sources using the following formula (1). ; (1); In equation (1), The first in the representative keyword dictionary The weight of each term is determined by the fact that the keyword dictionary includes terms of different categories, and different weights are assigned to terms of different categories, including core terms of hail, phenomenon terms, particle size terms, and disaster damage terms. This represents the weight of different text fields extracted from the text data, specifically the title, body text, tags, or comments. Represents a binary indicator variable. Representing the Each term appears in the text field. Representing the The term did not appear in the text field; This represents iterating through all entries and all text fields, and accumulating keyword scores; This represents a truncation function, with a maximum keyword score of 1. S22. Calculate the relevance of the clues according to the following formula (2). : (2); In equation (2), Represents semantic confidence, specifically the probability that the content output by the semantic classification model describes a real hail event; This represents the media confidence score, specifically the confidence score given by the media recognition model when it identifies image targets such as hail particles, hail accumulation, broken vehicle windows, and damaged crops. When no image target is identified, the confidence score is lower. ; This represents a reverse rule penalty; when low-relevance content such as science popularization, historical reviews, film and television content, reposting from other locations, and weather forecasts is identified, Take the penalty term that is greater than 0; , and These are non-negative weighting coefficients. The contribution ratios of keywords, semantic classification models, and media evidence in the calculation of clue relevance are controlled respectively. This represents the penalty coefficient, used to control the severity of the penalty for the reverse rule; The pruning function forces the calculation result to be restricted to... The interval; S23, Clue Relevance Not lower than the relevance threshold Multi-source data is incorporated into the suspected hail clue set; relevance threshold. Determined by optimizing the F1 score, recall lower limit, or business cost function on historical labeled samples; (3) Extract hail event elements from each suspected hail clue in the suspected hail clue set, and retain the extraction confidence of the hail event elements and the corresponding multi-source tracing evidence. Then, perform spatiotemporal normalization processing on the hail event elements. Finally, encapsulate the normalized hail event elements of each suspected hail clue, their extraction confidence, and the corresponding multi-source tracing evidence into a network evidence object. The specific steps are as follows: S31. Extracting Hail Event Elements: When the suspected hail clue is text, entities with time, location, particle size, damage, organization, and negation relation are extracted using a rule base and named entity recognition model. The named entity recognition model uses a conditional random field, a bidirectional long short-term memory network-conditional random field, or a Transformer-based sequence labeling model. When the suspected hail clue is an image or video, an image classification model, a target detection model, or a video keyframe model is used to identify hail and damage targets. The extracted entities and the identified hail and damage targets together constitute the hail event elements. For each hail event element, save source metadata: for the extracted entities, save the original text fragment, the character position of the entity, the version number of the rule base and named entity recognition model, and the extraction confidence; for the identified hail and damage targets, save the media identifier, keyframe number, detection box coordinates, the version number of the image classification model, target detection model or video keyframe model, and the extraction confidence. The rule base stores pre-set keywords, matching patterns, and entity normalization mapping relationships for deterministic entity extraction, which compensates for the missed detection and false detection problems of the named entity recognition model. It also performs standardized mapping and legality verification on the extracted entities and outputs the extraction confidence corresponding to the rule matching. This is then fused with the output of the named entity recognition model to obtain the hail event elements and their corresponding extraction confidence. S32. Time normalization processing of hail event elements: Convert the time entities in the hail event elements into a candidate time window or a set of candidate time windows. For each candidate time window, output the event start time, event end time, time precision level, time confidence level, and time basis. For each candidate time window, the overall credibility score is calculated using the following formula (3): (3); In equation (3), Representing the The overall credibility score of each candidate time window, Representing the The contextual evidence score for each candidate time window is calculated based on the temporal consistency between the candidate time window and the original text context. Representing the The media metadata time evidence score for each candidate time window is calculated based on the overlap between the candidate time window and the timestamps in the image and video data. Representing the The cross-evidence score for each candidate time window is calculated based on the degree of overlap between the candidate time window and the time windows of other resolved hail cues. Representing the The meteorological matching support score of each candidate time window is calculated based on the degree of matching between the candidate time window and the actual strong convection records of the meteorological support area. , , and Represents the weight coefficient of each item; each weight coefficient is a non-negative number, and the sum of all valid item weights is 1; if there are unusable items, the corresponding items are removed and the remaining valid weights are renormalized; Multiple candidate time windows in a set of candidate time windows are sorted according to their comprehensive confidence scores. If the comprehensive scores of multiple candidate time windows are close to each other, they are not eliminated and all candidate time windows are retained. The time confidence is obtained by mapping the comprehensive confidence scores of the candidate time windows. Before the meteorological support area is generated, when calculating the comprehensive credibility score of the candidate time window, the meteorological matching support score is not calculated, and the weights of the remaining valid sub-items are renormalized; after the meteorological support area is generated, the meteorological matching support score is introduced, the comprehensive credibility score of the retained candidate time windows is recalculated, and the ranking is updated a second time. S33. Spatial normalization of hail event elements: First, the location entities in the hail event elements are analyzed and recalled according to the administrative division database, place name dictionary, and POI database to generate a candidate location or a set of candidate locations; for each candidate location, the positioning information is specified, and the latitude and longitude points and error radius are output; for fuzzy location entities without corresponding latitude and longitude points (POI, township, county, etc.), the geographic boundary vector polygon and spatial accuracy level are output. For each candidate position, the overall credibility score is calculated using the following formula (4): (4); In equation (4), Representing the The overall credibility score of each candidate position. Representing the The text name matching score of each candidate location is obtained by calculating the text similarity between the candidate location and the standard place name of the candidate location in the administrative division database and place name dictionary. Representing the The context score of each candidate location is calculated by matching the candidate location with the context location description in the original text. Representing the The POI matching score for each candidate position is calculated by matching the candidate position with the matching objects in the POI library. Representing the The cross-clue evidence score of each candidate location is calculated by spatially matching the candidate location with the other hail clue locations. Representing the The meteorological region overlap score of each candidate location is obtained by spatially matching the candidate location with the meteorological support area. , , , and Represents the weight coefficient of each item; each weight coefficient is a non-negative number, and the sum of all valid item weights is 1; if there are unusable items, the corresponding items are removed and the remaining valid weights are renormalized; The candidate positions in a set of candidate positions are ranked according to their comprehensive credibility scores. Before the meteorological support area is generated, when calculating the comprehensive confidence score of the candidate locations, the meteorological region overlap score is not calculated, and the weights of the remaining valid sub-items are renormalized; after the meteorological support area is generated, the meteorological region overlap score is introduced, the comprehensive confidence score of the retained candidate locations is recalculated, and the ranking is updated a second time. S34. For suspected hail clues that have been extracted from hail event elements and have undergone time and space normalization, if there are independent images, videos or comments on the same webpage corresponding to them, the suspected hail clues are split into multiple sub-objects, each of which is a network evidence object; otherwise, a corresponding network evidence object is directly generated for the suspected hail clues. The suspected hail clue before splitting is used as the parent original clue. The parent original clue and its corresponding multiple child objects are associated. The parent source identifier of the child object is taken as the unique evidence identifier of the parent original clue. The network evidence objects are assembled using fixed fields, which include: unique evidence identifier, parent original clue record number, source type, content carrier type, publication time, parent source identifier, candidate time window, candidate spatial region, hail size range, disaster impact label, media fingerprint, original source link, hail event element extraction confidence level, original source group number, conflict marker, and processing version; the candidate time window is a single candidate time window or a sorted set of candidate time windows; the candidate spatial region is a single candidate location or a sorted set of candidate locations; the hail size range and disaster impact label are derived from the extracted hail event elements; (4) Based on all network evidence objects corresponding to the suspected hail clue set, calculate the multi-dimensional evidence similarity pairwise. When the multi-dimensional evidence similarity meets the same source threshold requirement, establish a propagation edge to complete the construction of the evidence propagation graph, and divide it into multiple original source groups according to the propagation edge. Then, for all network evidence objects corresponding to the suspected hail clue set, cluster candidate events based on spatiotemporal, semantic, evidence consistency, and same source or conflict penalty to generate candidate hail events. The specific steps are as follows: S41. Construct an evidence propagation graph G=(V,E), where the nodes in the node set V are network evidence objects, and the edge set E consists of propagation edges; For any two network evidence objects, the multi-dimensional evidence similarity is calculated using the following formula (5). : (5); In equation (5), Representative node and nodes Multidimensional evidence similarity between them Representative node and nodes Text similarity between them Representative node and nodes Perceptual hash similarity between images Representative node and nodes Similarity of keyframes between videos Representative node and nodes Audio fingerprint similarity between them Represents the similarity of link references. , , , and The weight coefficients represent the weights of each dimension. When two nodes belong to the same content modality, the similarity corresponding to that content modality and the link reference similarity are calculated, and the weighted sum of the similarity items actually involved in the calculation is performed. When two nodes belong to different content modalities, the link reference similarity is used as the multi-dimensional evidence similarity. The weight coefficients corresponding to the similarity items not involved in the calculation are not included in this calculation. The effective weight coefficients corresponding to the similarity items actually involved in the calculation are renormalized so that the sum of the effective weight coefficients is 1. when Not lower than the set homogeneity threshold When the two nodes are connected, a propagation edge is established between them and added to the edge set E. After completing the multi-dimensional evidence similarity calculation and edge construction for all network evidence objects, perform connected component partitioning or community detection on the evidence propagation graph to obtain multiple original source groups. Each original source group is generated based on the graph connectivity relationship and includes evidence pairs that meet the multi-dimensional evidence similarity standard, as well as capturing the indirect propagation homologous evidence set formed by chain relay. Optionally, to avoid misgrouping caused by chain propagation, after obtaining the original source group, an intra-group consistency check can also be performed to remove objects with too low similarity within the group. The credibility of all network evidence objects within each original source group is aggregated to obtain the aggregate credibility of that original source group, and an equivalent evidence object is generated. During credibility aggregation, the aggregate credibility of the original source group is equal to the maximum base credibility of all network evidence objects within the original source group, or the aggregate credibility obtained by attenuation and superposition according to the propagation distance. The equivalent evidence object is a virtual evidence record, and its fixed fields are selected from the network evidence object corresponding to the maximum base credibility or the fields obtained by fusion operation of the fields of all network evidence objects within the original source group. If there is only one network evidence object in the original source group, and this network evidence object is different from other network evidence objects describing the same hail event in terms of account, shooting angle, media fingerprint, and spatiotemporal information describing the event, then it is regarded as an independent evidence object, and its basic credibility is retained as the aggregate credibility of the original source group. The basic credibility of the network evidence object is calculated using the following formula (6): (6); In equation (6), Representing the The credibility of the source of each piece of online evidence is derived from the source type; Representing the The credibility of the evidentiary form of an individual network evidence object is derived from the type of content carrier; Representing the The reliability of the temporal precision of each network evidence object is derived from the candidate time window; The reliability of the location accuracy is derived from the candidate spatial region; Representing the The originality and credibility of each piece of online evidence is obtained by comparing and retrieving media fingerprints and original source links; Representing the The conflict penalty for each network evidence object is derived from the conflict marker; Representing the The punishment for reposting old news by an individual online evidence subject is derived from media fingerprints and historical material databases retrieved based on publication time. , , , , , and Represents the weighting coefficients of each item; The pruning function forces the calculation result to be restricted to... The range; among which, source credibility, evidence form credibility, time accuracy credibility, location accuracy credibility and originality credibility are positive sub-items, each sub-item weight is a non-negative number, and the sum of the sub-item weights is 1; the weights of conflict penalty and old news reposting penalty are set separately. S42. Calculate the event association score using the following formula (7). : (7); In equation (7), Representative object and object The event correlation score is based on whether the object is an equivalent evidence object or an independent evidence object; It represents the temporal similarity between two objects, that is, the degree of overlap between the candidate time windows of the two objects; It represents the spatial similarity between two objects, that is, it calculates the degree of overlap or distance decay between the candidate spatial regions of the two objects; This represents the semantic similarity between two objects, specifically calculating the matching similarity of the semantic tags of the hail event elements of the two objects. It represents the consistency of evidence between two objects, that is, the degree of agreement between the values ​​of structured elements in the hail event elements of the two objects; The same-origin or conflict penalty between two objects is calculated from the original source group number and conflict flag of the two objects; , , , and The weight coefficients represent the weights of each item; among them, temporal similarity, spatial similarity, semantic similarity, and evidence consistency are positive items, each item's weight is a non-negative number, and the sum of the item weights is 1; the weights for homology or conflict penalties are set separately. When the event is associated with the score Not less than the clustering threshold If the hard conflict conditions of being geographically distant and having completely disjoint dates are not met, then an edge with the same event is established between the two objects, and all objects are grouped and clustered according to all edges with the same event by connected component partitioning, hierarchical clustering or DBSCAN algorithm to obtain multiple candidate event clusters. The aggregated confidence scores of all objects within each candidate event cluster are read and fused to obtain the event confidence score of that candidate event cluster; the structured information of all objects within each candidate event cluster is aggregated to obtain candidate hail events; After obtaining multiple candidate event clusters, post-verification processing is performed: the maximum time span and maximum spatial span among all objects within the candidate event cluster are calculated, and the consistency of meteorological units corresponding to the candidate event cluster is verified; when the maximum time span and maximum spatial span exceed preset thresholds, or when meteorological units are inconsistent, the candidate event cluster is split a second time to obtain split candidate event clusters; the calculation of event confidence and the aggregation of candidate hail events are performed after post-verification processing. (5) Generate and match meteorological support areas for candidate hail events, and perform hard conflict rule gating on the candidate hail events to identify irreplaceable hard conflicts and output the corresponding processing labels. The specific steps are as follows: S51. Meteorological data is acquired through the meteorological operational data interface and converted into various meteorological support zones. The meteorological support zones are divided into radar support zones, dual-polarization support zones, path support zones, and ground support zones. The radar support zone, also known as the radar strong echo support zone, is generated by reflectivity, echo top height, vertical cumulative liquid water content, maximum reflectivity height, and strong echo vertical structure. The dual-polarization support zone, also known as the dual-polarization hail identification zone, is generated by correlation coefficient reduction, differential reflectivity anomaly, differential phase shift rate, and hail category in water condensate classification. The path support zone includes radar meteorological cell path buffer zones, lightning activity zones, satellite convection zones, and automatic weather station accompanying weather zones. By tracking the convective cell centers of adjacent volume scans, a time-varying buffer zone is generated based on cell scale, movement speed, and positioning error. The ground support zone is generated by the location and error range recorded by manual observation, official disaster reports, agricultural insurance claims, or trusted volunteers. The candidate time windows of the candidate hail events are overlapped with the corresponding time slices of the meteorological support area. The candidate spatial regions of the candidate hail events are compared with the meteorological support area in terms of area intersection-union ratio, nearest distance, and path relationship. The results of the time interval overlap calculation, as well as the results of the area intersection-union ratio, nearest distance, and path relationship calculation, are combined to form the spatiotemporal matching result between the candidate hail events and the meteorological support area. Specifically, for coarse-grained locations such as county level, the spatial buffer range increases with location uncertainty; for minute-level time, the time buffer range adjusts with the radar volume scan cycle. Interval overlap calculation is used to assess the degree of temporal overlap between the time window of the candidate hail event and the time slice of the meteorological support area, providing temporal evidence. Area intersection-union ratio is used to assess the degree of spatial overlap between the candidate spatial region of the hail event and the meteorological support area. Nearest distance is used to characterize the spatial distance between the two regions when there is almost no overlap, adapting to coarse-grained disaster locations with positioning errors. Path relationship is used to determine whether the movement path of the convective cell that produced hail passes through the spatial range of the candidate hail event. Through the above multiple calculations, the spatiotemporal matching degree of each meteorological support area corresponding to the candidate hail event is quantified, and the matching results are output to provide meteorological evidence for subsequent comprehensive credibility calculation. S52. Execute hard conflict rule gating: Perform hard conflict rule matching for each candidate hail event, determine whether each rule meets the triggering conditions, and execute the corresponding handling action when the triggering conditions are met. Finally, summarize the handling results of all rules and output the handling flags. The handling flags include exclusion flags, downgrade flags, manual review flags, and conflict tags. Conflict tags include supplementary conflict tags and non-supplementary hard conflict tags. The hard conflict rules, triggering conditions, and handling actions are shown in Table 1 below: Table 1

[0026] When the rules in Table 1, such as confirming old news, clearly indicating a different location, hard time conflict, or hard location conflict, are triggered, their conflict labels are marked as non-supplementable hard conflict labels; when the weather does not support the rules or the strong weather-weak network rule is triggered, their conflict labels are marked as supplementable conflict labels. The re-clustering in Table 1 refers to removing the network evidence object that triggered the tag from the current candidate event cluster and attempting to merge it into other matching candidate event clusters; if merging is not possible, an independent candidate event is generated. (6) After calculating the credibility of network evidence, meteorological support, multi-source cross-validation gain, and conflict penalty of candidate hail events, the overall credibility is calculated by weighting, and the quality control level is output by combining the processing label of the output with the hard conflict rule gating. The specific steps are as follows: S61, Original source group Aggregate credibility Substituting into equation (8), the baseline value of the candidate hail event is calculated. ; (8); In equation (8), Representing the original source group Weighting coefficients; The strength of multi-source mutual verification is calculated using the following formula (9). : (9); In equation (9), This represents the total number of original source groups corresponding to the current candidate hail event. It is an integer not less than 1; Represents the saturation coefficient. Greater than 0 controls the rate at which saturation occurs; when When =1, the strength of multi-source mutual verification is 0; as As the value increases, the strength of multi-source mutual verification monotonically increases and gradually approaches 1; the strength of multi-source mutual verification... Represents the strength of mutual corroboration among the original source groups, with a value range of [value range missing]. ; Finally, the credibility of the network evidence for the candidate hail event is calculated using the following formula (10). : (10); In equation (10), The weighting coefficients representing the strength of multi-source mutual verification. Weighting coefficients representing the baseline values ​​of candidate hail events; using The function constrains the fusion result to The interval; S62. Calculate the meteorological support of candidate hail events according to the following formula (11). : (11); In equation (11), Used to measure the degree of support from professional meteorological data for the existence of hail or severe convective conditions in the candidate spatial region of a candidate hail event. The time matching degree is represented by the calculation of the time distance between the meteorological observation time and the time window of the candidate hail event, and normalized to 0 to 1. The spatial matching degree is represented by calculating the spatial distance or spatial overlap ratio between the meteorological observation location and the spatial region of the candidate hail event, and normalized to 0 to 1. The radar hail support is represented by the radar-identified hail potential calculated in the spatial region of the candidate hail event. Represents the dual polarization support, which is obtained by calculating the significance of hail features and normalized to 0 to 1; The score represents the support for lightning, which is obtained by statistically analyzing the frequency and distance of lightning within the spatial region of the candidate hail event. The higher the lightning frequency and the closer the observation location is to the spatial region of the candidate hail event, the higher the score, which is normalized to 0 to 1. The satellite's convective support is obtained by extracting convective features and identifying strong convective potential. The weather support level of the automatic weather station is calculated by extracting environmental meteorological features that are conducive to the occurrence of hail events and scoring them comprehensively, and then normalized to 0 to 1. Representing ground record support, this is determined by statistically analyzing whether there are actual records of hail in the spatial area of ​​the candidate hail event. Records with more records are given higher scores, and the scores are normalized to 0 to 1. , , , , , , and The weight coefficients represent the weights of each item. These weight coefficients are global hyperparameters and are optimally determined from the historical hail validation set. When all data sources are valid, the sum of all weight coefficients equals 1. When a meteorological data source is missing, the weight coefficients corresponding to the missing items are removed, and the remaining valid weight coefficients are renormalized. At the same time, a data availability identifier is output to mark the availability status of each sub-item data source. S63. Calculate the multi-source cross-validation gain using the following formula (12). : (12); In equation (12), The sub-item consistency score represents the pairwise comparison results between two original source groups for the same candidate hail event. The comparison types include cross-platform comparison, cross-user comparison, cross-media comparison, and cross-category source comparisons conducted by the network multi-source group with radar, radar dual polarization, and ground real-time records, respectively. Each comparison outputs a corresponding sub-item consistency score. The gain coefficient representing the consistency score of each sub-item is a hyperparameter; subscript An index representing the consistency score of each item; each set of comparisons corresponds to one. ; The gain effect of cross-validation of information from multiple sources is characterized; only valid comparisons between different original source groups are included in the product operation; when there are no valid comparisons, the gain of cross-validation of multiple sources is 0, and reprints within the same original source group are not counted repeatedly. S64. Calculate the conflict penalty using the following formula (13). : (13); In equation (13), The conflict severity represents the severity of the conflict, which is not supported by old news, time, location, intensity, media or weather. The conflict label is a supplementary conflict label, which is obtained by comparing the conflict situation of multiple sources of information under the same candidate hail event and assigning a value according to the intensity of the conflict. Represents the conflict weight, which is a hyperparameter; Characterizes the comprehensive conflict degree of multi-source information of candidate hail events; non-supplementable hard conflicts are not included in the calculation of equation (13) and are directly processed by hard conflict rule gating; S65. The overall confidence level of candidate hail events is calculated using the following formula (14): (14); In equation (14), , , The weights are non-negative and their sum is 1. This is the conflict penalty coefficient; When a hard conflict that cannot be supplemented is hit, the corresponding processing result takes precedence over the overall credibility and cannot be offset by the credibility of network evidence, meteorological support or multi-source cross-validation gain. After calculating the overall credibility of the candidate hail event, the candidate hail event is classified into the corresponding preliminary quality control level according to the credibility level classification principle; various processing labels are received from the hard conflict rule gating output, and the preliminary quality control level is excluded, downgraded, marked with manual review, and restricted to the highest level using the processing labels to obtain the final quality control level of the candidate hail event. Examples of quality control levels are shown in Table 2 below: Table 2

[0027] The thresholds (0.2, 0.4, 0.6, 0.8) used for classifying levels in Table 2 were adjusted based on historical labeled samples and business risks; (7) Under the constraints of the rules, perform quality control calibration on the time, location and intensity of the candidate hail events, retain the original information of the candidate hail events, and generate a credible hail event record containing the chain of evidence, quality control level, uncertainty range and calibration instructions. The specific steps are as follows: S71. Time Calibration: When the time range of a candidate hail event intersects with the effective time interval corresponding to the meteorological support area, if there is only one overlapping intersection time period, it is used as the recommended time window after meteorological constraints. If there are multiple overlapping intersection sub-time periods, the meteorological support corresponding to each intersection sub-time period is calculated, and the intersection sub-time period with the highest meteorological support is selected as the recommended time window after meteorological constraints. At the same time, the original time range of the original candidate hail event is retained, and time conflicts are marked to retain the uncertainty of time. When there is no overlap but the time interval is within the set buffer threshold, Bayesian or weighted updates are performed by combining the radar single-unit transit time, the dual-polarization hail identification time, the lightning jump time, and the independent network clue sequence to obtain the recommended time window after meteorological constraints; at the same time, the original time range of the original candidate hail events is retained, and time conflicts are marked to retain the uncertainty of time. When there is no overlap but the time interval exceeds the set buffer threshold, the original time range of the original candidate hail event is retained, the time conflict is marked, the time of the event is not modified, and the uncertainty of the time is retained. Weighted update involves obtaining the temporal probability distributions corresponding to candidate hail events and the temporal probability distributions corresponding to meteorological support areas, with weights derived from the temporal confidence of network evidence and the temporal matching degree of meteorological support, respectively. The two sets of temporal probability distributions are then weighted and fused, and the recommended time window with meteorological constraints is determined based on the peak value of the fused probability distribution. Bayesian update uses the temporal probability distribution corresponding to candidate hail events as the prior distribution and the meteorological time-series observation information of the effective time interval corresponding to the meteorological support area as the likelihood to calculate the posterior temporal probability distribution. The recommended time window with meteorological constraints is then generated based on the peak value of the posterior distribution. S72. Location Calibration: When the candidate spatial region of a candidate hail event overlaps with the meteorological support region, if there is only a single overlapping region, the overlapping region is used as the candidate spatial region after meteorological constraints. If there are multiple discontinuous overlapping sub-regions, the meteorological support corresponding to each overlapping sub-region is calculated, and the overlapping sub-region with the highest meteorological support is selected as the candidate spatial region after meteorological constraints. At the same time, the original candidate spatial region of the original candidate hail event is retained, and location conflicts are marked to retain spatial uncertainty. When the candidate spatial region description of a candidate hail event is ambiguous, but the meteorological cell path information in the meteorological support area is clear, the candidate spatial region of the candidate hail event is narrowed to the intersection of the meteorological cell path, the dual-polarized hail region, and the multiple ground clues corresponding to the candidate hail event, to obtain the meteorologically constrained candidate spatial region; at the same time, the original candidate spatial region of the original candidate hail event is retained, and positional conflicts are marked to preserve spatial uncertainty. When the candidate spatial region of a candidate hail event does not overlap with the meteorological support area, the quality control level of the candidate hail event is reduced, the original candidate spatial region of the original candidate hail event is retained, and the location conflict is marked to retain spatial uncertainty. The aforementioned time and location calibrations do not rewrite the original time and location of candidate hail events; rather, they are based on meteorological constraints to select and generate recommended time windows and candidate spatial regions, while preserving the original event information completely and marking conflicts and uncertainties. S73, Intensity Calibration: The hail size descriptions obtained from suspected hail cues for candidate hail events are mapped to initial size intervals with upper and lower boundaries. Based on image scale, video target references, manual observations, and official disaster damage records, the intensity confidence of the suspected hail cues is adjusted. The rationality of the initial size interval is then verified using radar and dual-polarization features of the meteorological support area. Finally, the upper and lower limits of hail size, hail intensity level, impact level, and intensity confidence are output. After calibration, the intensity descriptions of the original candidate hail events are retained, and intensity conflicts are marked. Meteorological data is only used to constrain the intensity interval and does not directly replace the ground hail size obtained from suspected hail cues, thus retaining the uncertainty of intensity. S74. Based on the constructed evidence propagation graph, the nodes in the evidence propagation graph include original records, evidence objects, original source groups, candidate hail events, time windows, spatial regions, meteorological support areas, ground records, rule actions, and calibration results; the edges in the evidence propagation graph include origin, same source, support, correction, conflict, time matching, spatial matching, and calibrations generated therefrom. Each edge stores the relationship score, rule or model version, generation time, and original data index; thereby tracing back the formation basis of any quality control level, time calibration, location calibration, or intensity calibration, a credible hail event record containing the evidence chain, quality control level, uncertainty range, and calibration description is generated; Credible hail event records specifically include: unique event identifier, quality control level, overall credibility, credibility of network evidence, meteorological support, event start time, event end time, candidate time window set, candidate spatial region set, spatial uncertainty, minimum hail size, maximum hail size, intensity credibility, disaster impact label, unique evidence identifier set, original source group identifier set, evidence chain, quality control mark set, conflict mark set, rule-based handling action set, calibration instructions, model version set, and creation time. Among them, the evidence propagation graph is used for the same source identification of network evidence objects; based on the evidence propagation graph, the nodes and corresponding relationship edges corresponding to the rule actions and calibration results are improved and expanded to form a full-process evidence chain graph.

[0028] Records Q0 and Q1 are added to the candidate audit database; records Q2 through Q4 can be added to the trusted dataset, with Q2 records retaining the mark that requires manual review. Business systems can be configured to release only Q3 and Q4.

[0029] See Figure 2 A hail event quality control calibration system includes a data acquisition and storage module 1, a clue identification module 2, an event element extraction module 3, a spatiotemporal normalization module 4, an evidence object generation module 5, a homology identification module 6, a candidate event clustering module 7, a meteorological support area matching module 8, a quality control rule engine module 9, a credibility assessment module 10, a quality control calibration module 11, and an event recording module 12. The data acquisition and retention module 1 is used to acquire multi-source data related to hail events and to complete the original storage and management of multi-source data; Clue identification module 2 is used to calculate the correlation of clues from multiple sources, and after initial screening, a set of suspected hail clues is identified. The event element extraction module 3 is used to extract hail event elements from each suspected hail clue in the suspected hail clue set, and retain the extraction confidence of hail event elements and corresponding multi-source tracing evidence. Spatiotemporal normalization module 4 is used to perform spatiotemporal normalization processing on hail event elements; The evidence object generation module 5 is used to encapsulate the normalized hail event elements of each suspected hail clue, their extracted confidence level, and the corresponding multi-source tracing evidence into a network evidence object. The same source identification module 6 is used to calculate the multi-dimensional evidence similarity pairwise based on all network evidence objects corresponding to the suspected hail clue set. When the multi-dimensional evidence similarity meets the same source threshold requirement, a propagation edge is established to complete the construction of the evidence propagation graph and multiple original source groups are obtained based on the propagation edge. Candidate event clustering module 7 clusters candidate events based on spatiotemporal, semantic, evidence consistency, and homology or conflict penalty for all network evidence objects corresponding to the suspected hail clue set, and generates candidate hail events. The meteorological support area matching module 8 generates and matches meteorological support areas for candidate hail events; The quality control rule engine module 9 is used to perform hard conflict rule gating on candidate hail events, identify irreparable hard conflicts and output the corresponding processing labels; The credibility assessment module 10 is used to calculate the credibility of network evidence, meteorological support, multi-source cross-validation gain and conflict penalty of candidate hail events, and then calculate the comprehensive credibility by weighting, and output the quality control level by combining the processing label output by the hard conflict rule gating output. The quality control calibration module 11 is used to perform quality control calibration on the time, location, and intensity range of candidate hail events under rule constraints; The event logging module 12 is used to generate a credible hail event log that includes the chain of evidence, quality control level, uncertainty range, and calibration instructions.

[0030] Example 1: Hailstorm event quality control calibration Regarding the localized severe convection and brief hailstorm in a certain city in a certain province, public reports indicate that around 4:35 PM, a district and some townships in the city experienced a brief hailstorm. Additionally, netizens provided photos and text clues showing "heavy rain, hail, and strong winds 10 kilometers east of a certain location." Simultaneously, a yellow warning for severe convection was in effect for the city, and severe convection warnings were issued for many surrounding areas. The above publicly available information is used to illustrate the event background of this embodiment; the final event time, hail area, and intensity are still based on the actual calculation results from the original network data, radar base data, and operational observation records.

[0031] The hail event quality control calibration system first collects online posts, news pages, images, short videos, and comments related to terms such as "city, district, hail, hail, strong winds" on the day of the incident, and writes the publication time, original link, media hash, and webpage snapshot into the original record layer. The clue identification module 2 calculates the clue relevance based on the title "Just now, hail fell in a certain city," the on-site description, and related media content. Weather warning information, because it describes possible weather rather than actual ground events, is marked as meteorological background clues and is not considered standalone evidence of ground hail.

[0032] Event element extraction module 3 obtains the date "May 14, 2026", the relative time "just now", the location "a certain city", "some townships", "10 kilometers east of a certain location", and the phenomena "small hail, heavy rain, strong wind" from the publicly available content. Using the report's publication time of 16:35 to 16:38 as a reference, "just now" is normalized into a backward-looking candidate time window; the specific backward length is set based on the platform's historical statistics of publication time delays, for example, initially forming an initial window from 15:50 to 16:38. Spatiotemporal normalization module 4 generates candidate spatial regions for the city, relevant townships, and the area east of the location, retaining administrative boundaries, POI candidates, and spatial uncertainty.

[0033] The same source identification module 6 compares images, video keyframes, titles, and text on different pages. If multiple media pages reproduce the same on-site photo or the same video, they are classified into the same original source group; the number of independent evidence objects is only increased when the subject of the photograph, shooting angle, media fingerprint, or location information can prove their independence. This avoids miscounting multiple media reprints of the same material as multiple pieces of on-site evidence. Subsequently, based on conditions such as overlapping time windows, spatial proximity, semantic consistency of hail, and no clear conflict between different locations, two candidate events are formed respectively; if the radar meteorological cell path and evidence timing indicate that the two belong to different convective cells or the spatial span exceeds the upper limit, they remain as two candidate events and are not forcibly merged because they belong to the same administrative region.

[0034] The meteorological support area matching module 8 retrieves radar volume scans, dual-polarization products, lightning location data, automatic weather station records, and early warning records for the surrounding area on May 14, 2026. Based on the actual operational data, it generates a radar strong echo support area, a dual-polarization hail identification area, and a radar meteorological cell path buffer zone. In this operational data backfilling, the two candidate events are recorded as a candidate event for a certain region and a candidate event for a certain location, respectively. The time slice corresponding to the meteorological support area is selected from 16:08 to 16:24 Beijing time on May 14, 2026. The measured results are as follows: In one area, the maximum reflectance of the candidate event was 40.50 dBZ, the maximum vertical cumulative liquid water content (VIL) was 2.20, no strong echo top was identified at 45 dBZ, the -20°C layer height (H2O) was 7.452 km, the wet-bulb zero-degree layer height (WBZ) was 3.403 km, and the maximum area of ​​the strong 45 dBZ echo was 0 km². In another location, the maximum reflectance of the candidate event was 63.50 dBZ, the maximum VIL was 87.90, the 45 dBZ echo top height (ET45) was 12.40 km, the H2O was 7.440 km, the WBZ was 3.482 km, the maximum area of ​​the strong 45 dBZ echo was 132.703 km², and continuous strong convective cells were identified. The system does not presuppose that the parameters of the above operational data backfill necessarily support hail, but calculates the meteorological support based on the measured values.

[0035] The hard conflict rule gating first checks whether the on-site media is historical material or relocated, and verifies whether the network location conflicts with the meteorological unit path. If the media passes the same source and historical database checks, and the region of the first or second candidate event matches the corresponding meteorological support area in time and space, then the credibility of the network evidence, meteorological support, multi-source cross-validation gain, and overall credibility are calculated. If there is only a warning without independent on-site evidence, ground hail cannot be directly confirmed. The final time is taken as the intersection of the network candidate time window and the meteorological support area or the sub-window with the highest support; the location is taken as the intersection of the network candidate spatial area, the meteorological unit path, and independent on-site evidence; the intensity is based on the publicly described "small hail" as the initial intensity level, and is adjusted in combination with the image scale, ground records, and dual polarization results. Precise millimeter values ​​are not output when a reliable particle size reference is not obtained. In this reproduction, the time matching degree of the two candidate events... All are 1.00; the candidate events in a certain district are approximated by the administrative boundaries of that district, resulting in a coarser spatial granularity and a lower spatial matching degree. A value of 0.70 is used; a candidate event in a certain location directly matches an administrative region in that location. Set the value to 1.00. Calculate the radar hail support for candidate events in a certain area based on currently available radar and environmental field data. The meteorological support is 0.0894. The value is 0.3937, which indicates low meteorological support; [This is related to] candidate events in a certain area. The meteorological support is 0.9679. The score is 0.9807, which falls under the category of high meteorological support. Currently, dual-polarization data and ground observation data are unavailable. The status of lightning, satellite, and automatic weather station data has not yet been confirmed. Missing or unconfirmed data will not be scored as unsupported, but will be treated as data availability.

[0036] This allows us to trace back the basis for any quality control level, time calibration, location calibration, or intensity calibration, generating a credible hail event record that includes a chain of evidence, quality control level, uncertainty range, and calibration instructions.

[0037] This embodiment ultimately outputs at least two types of records: one type consists of credible hail events that meet the database entry threshold, storing the calibration time, calibration location, intensity range, quality control level, and complete evidence chain; the other type consists of candidate records with only warnings, reposts, or unclear locations, stored in the audit database and marked "requires manual review." Based on the actual meteorological data backfilling results, candidate events in a certain area showed weak reflectivity and VIL during key time periods, and no 45dBZ strong echo core was identified, resulting in low meteorological support. The meteorological support score is 0.3937, indicating low meteorological support. Downgrading or manual review is recommended. A candidate event in a certain area exhibits strong reflectivity, high VIL, deep 45 dBZ echo, and a large area of ​​strong echoes. A value of 0.9807 is marked as high meteorological support. It should be noted that high meteorological support only indicates strong support from professional meteorological data for the presence of hail or severe convective conditions in the candidate area. In the absence of highly reliable ground records or other independent field evidence, actual hail cannot be directly confirmed based solely on meteorological support. This demonstrates that this invention does not directly equate "severe convective weather in a certain area" with "hail occurred in that area," but rather confirms local hail events by object, region, and evidence. Public background information source: Hail and severe convective weather warnings for the region published on May 14, 2026, by local news channels, public online platforms, and related media.

[0038] Example 2: Eliminating the hard conflict of old news or materials from other locations being passed off as local weather. On May 5, 2024, a netizen posted a video of hail from another area, which they had seen online, on their personal social media account with the caption "Hail in a certain county," leading viewers to mistakenly believe that the video was filmed in their own county and reflected their real-time weather. Official verification revealed that the video was not from the county at all, and the relevant announcement stated that the netizen posted it casually out of momentary curiosity.

[0039] When the system collects content posted by a netizen, it first generates a suspected hail clue, extracting the posting time, posting platform, poster identifier, text "Hail in a certain county," declared location "a certain county," and media type "Hail video." If the original media work is still accessible, it saves the work link, work identifier, video file hash, keyframe fingerprint, and collection time; if the original post has been deleted or cannot be retrieved, it marks the material availability as unavailable, marks the media fingerprint as missing, and saves the official announcement page, announcement title, announcement date, verification entity, and verification conclusion. It is prohibited to replace the original material with other "Hail in a certain county" videos currently found in the search.

[0040] During the clue identification stage, this content, due to its inclusion of terms such as "a certain county" and "hail," along with images of hail, may receive high keyword scores, semantic confidence, and media confidence. If a soft scoring method were used directly, the system might incorrectly identify it as a local hail clue. Therefore, the quality control rule engine module 9 performs a hard conflict check before calculating overall credibility: when an authoritative verification record clearly confirms that the media actually originated from another province and states the location as a certain county, a "confirmed out-of-town media" flag, a "location conflict" flag, and a "misleading local text" flag are set to 1; this conclusion constitutes a hard conflict that cannot be offset by keywords, video clarity, or concurrent strong convection signals.

[0041] If the original video is still available, the system can also compare the keyframe fingerprints of the video with the historical material library and the cross-regional event library, and establish an evidence edge for "events originating from other provinces" based on the earliest release time, original release location and landmark identification results. If the original video is no longer available, the system must not fabricate the above comparison results, but instead use the official verification conclusion as a highly credible verification node to establish a conflict edge for "official verification negating local events", while recording the absence of the original media as an "evidence missing marker".

[0042] The hard conflict rule gating output excludes the clue by setting it to True, Q0 (quality control level), and False (requires verification). However, it remains in the candidate audit database and the negative sample database for online rumors. Reasons for exclusion include at least "video from another location impersonating local weather," "conflict between declared location and verification location," and "denial by authoritative verification." Even if strong radar echoes existed in or around May 5, 2024, and meteorological support was high, it does not change the fact that the video was not taken in this county. Meteorological support only indicates the presence of strong convective conditions in the area; it does not prove the location or authenticity of the online video.

[0043] This embodiment illustrates that the hard conflict exclusion of the present invention is not only applicable to the republication of old news that can be identified through media fingerprinting, but also to the manipulation of materials from other locations that has been confirmed by authoritative verification. When the original post, original video, or work identifier is missing, the system explicitly indicates the evidence boundary through material availability, evidence missing markers, and authoritative verification nodes. This not only excludes false local events but also prevents unverified online videos from being incorrectly spliced ​​into the evidence chain.

[0044] Example 3: Strong meteorological support but insufficient online exposure Radar and dual-polarization data from remote areas show strong support for hail, but only one network clue with an ambiguous location. The system calculates high meteorological support but low network evidence confidence, outputting a flag indicating missing network evidence, and sending it to manual review or an external ground survey queue, instead of directly generating a high-confidence ground hail event. This process avoids equating incorrect aerial hail support with actual ground hail and can also indicate potentially missed areas.

[0045] During system implementation, a labeled set is established, including samples of real events, reposted old news, remotely sourced content, time conflicts, and weather-incompatible data. The training set is used for model learning, and the validation set is used to confirm the relevance threshold. Homology threshold Clustering threshold saturation coefficient The system incorporates fusion weights, and an independent test set is used to evaluate event-level accuracy, recall, location error, timing error, intensity interval coverage, and false alarm rate. System training or threshold evaluation utilizes independent human observations, official disaster reports, or field surveys as references to avoid repeatedly generating true values ​​from the meteorological data to be evaluated.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A quality control calibration method for hail events, characterized in that: Specifically, it includes the following steps: (1) Obtain multi-source data related to hail events on the Internet and complete the original storage and management of multi-source data; (2) Calculate the correlation of clues from multiple sources, and identify a set of suspected hail clues after initial screening; (3) Extract hail event elements from each suspected hail clue in the suspected hail clue set, retain the extraction confidence of hail event elements and the corresponding multi-source tracing evidence, then perform spatiotemporal normalization processing on hail event elements, and finally encapsulate the normalized hail event elements of each suspected hail clue, their extraction confidence and the corresponding multi-source tracing evidence into a network evidence object. (4) Based on all network evidence objects corresponding to the suspected hail clue set, calculate the multi-dimensional evidence similarity pairwise. When the multi-dimensional evidence similarity meets the same source threshold requirement, establish a propagation edge to complete the construction of the evidence propagation graph. Based on the propagation edge, divide the data into multiple original source groups. Then, for all network evidence objects corresponding to the suspected hail clue set, cluster candidate events based on spatiotemporal, semantic, evidence consistency, and same source or conflict penalty to generate candidate hail events. (5) Generate and match meteorological support areas for candidate hail events, and at the same time perform hard conflict rule gating on candidate hail events, identify non-replaceable hard conflicts and output corresponding processing labels; (6) After calculating the credibility of network evidence, meteorological support, multi-source cross-validation gain and conflict penalty of candidate hail events, the comprehensive credibility is calculated by weighting and the quality control level is output by combining the processing label output by the hard conflict rule gating. Among them, when a hard conflict that cannot be supplemented is hit, the corresponding processing result takes precedence over the comprehensive credibility and cannot be offset by the credibility of network evidence, meteorological support or multi-source cross-validation gain. (7) Perform quality control calibration on the time, location and intensity of candidate hail events under the constraints of the rules, retain the original information of candidate hail events, and generate a credible hail event record containing the chain of evidence, quality control level, uncertainty range and calibration instructions.

2. The hail event quality control calibration method according to claim 1, characterized in that: The multi-source data related to hail events on the network includes text, titles, tags, images, videos, comments, and metadata related to hail events obtained from authorized or compliant data interfaces, public web pages, government announcements, news media, manual reports, and business systems. The original preservation and management of the multi-source data specifically involves assigning a unique original record identifier to each piece of collected data, completely preserving the original information, and only adding to the original storage layer without modifying it. Subsequent clue identification and event calibration results are associated with the original record identifier through the version number, preserving complete traceability evidence.

3. The hail event quality control calibration method according to claim 1, characterized in that: The process of calculating the relevance of clues from multi-source data, and identifying a set of suspected hail clues after initial screening, specifically involves: S21. Construct a keyword dictionary and calculate keyword scores for text data from multiple sources using the following formula (1). ; （1）； In equation (1), The first in the representative keyword dictionary The weight of each term is determined by the fact that the keyword dictionary includes terms of different categories, and different weights are assigned to terms of different categories, including core terms of hail, phenomenon terms, particle size terms, and disaster damage terms. This represents the weight of different text fields extracted from the text data, specifically the title, body text, tags, or comments. Represents a binary indicator variable. Representing the Each term appears in the text field. Representing the The term did not appear in the text field; This represents iterating through all entries and all text fields, and accumulating keyword scores; This represents a truncation function, with a maximum keyword score of 1. S22. Calculate the relevance of the clues according to the following formula (2). : （2）； In equation (2), Represents semantic confidence, specifically the probability that the content output by the semantic classification model describes a real hail event; This represents the media confidence score, specifically the confidence score given by the media recognition model when it identifies image targets such as hail particles, hail accumulation, broken vehicle windows, and damaged crops. When no image target is identified, the confidence score is lower. ; This represents a reverse rule penalty; when low-relevance content such as science popularization, historical reviews, film and television content, reposting from other locations, and weather forecasts is identified, Take the penalty term that is greater than 0; , and These are non-negative weighting coefficients. The contribution ratios of keywords, semantic classification models, and media evidence in the calculation of clue relevance are controlled respectively. This represents the penalty coefficient, used to control the severity of the penalty for the reverse rule; The pruning function forces the calculation result to be restricted to... The interval; S23, Clue Relevance Not lower than the relevance threshold Multi-source data is incorporated into the suspected hail clue set; relevance threshold. The F1 score, recall lower limit, or business cost function on historical labeled samples are used to determine the optimal value.

4. The hail event quality control calibration method according to claim 1, characterized in that: The specific steps for extracting the hail event elements from each suspected hail clue in the suspected hail clue set and encapsulating them into a network evidence object are as follows: S31. Extracting Hail Event Elements: When the suspected hail clue is text, a rule base and named entity recognition model are used to extract entities with time, location, particle size, damage, organization, and negative relation; when the suspected hail clue is an image or video, an image classification model, target detection model, or video keyframe model are used to identify hail and damage targets; the extracted entities and the identified hail and damage targets together constitute the hail event elements; For each hail event element, save source metadata: for the extracted entities, save the original text fragment, the character position of the entity, the version number of the rule base and named entity recognition model, and the extraction confidence; for the identified hail and damage targets, save the media identifier, keyframe number, detection box coordinates, the version number of the image classification model, target detection model or video keyframe model, and the extraction confidence. S32. Time normalization processing of hail event elements: Convert the time entities in the hail event elements into a candidate time window or a set of candidate time windows. For each candidate time window, output the event start time, event end time, time precision level, time confidence level, and time basis. For each candidate time window, the overall credibility score is calculated using the following formula (3): （3）； In equation (3), Representing the The overall credibility score of each candidate time window, Representing the The contextual evidence score for each candidate time window is calculated based on the temporal consistency between the candidate time window and the original text context. Representing the The media metadata time evidence score for each candidate time window is calculated based on the overlap between the candidate time window and the timestamps in the image and video data. Representing the The cross-evidence score for each candidate time window is calculated based on the degree of overlap between the candidate time window and the time windows of other resolved hail cues. Representing the The meteorological matching support score of each candidate time window is calculated based on the degree of matching between the candidate time window and the actual strong convection records of the meteorological support area. , , and This represents the weight coefficient of each item; all weight coefficients are non-negative, and the sum of all valid item weights is 1. If there are unusable items, remove the corresponding items and renormalize the remaining valid weights; Based on the comprehensive credibility score of the candidate time windows, multiple candidate time windows in a set of candidate time windows are sorted. If the comprehensive scores of multiple candidate time windows are close to each other, they are not eliminated and all candidate time windows are retained. When the meteorological support zone has not yet been generated, the meteorological matching support score is not calculated when calculating the comprehensive confidence score of the candidate time window, and the weights of the remaining valid sub-items are renormalized. After the meteorological support zone is generated, the meteorological matching support score is introduced, the comprehensive confidence score of the retained candidate time windows is recalculated, and the sorting is updated a second time. S33. Spatial normalization of hail event elements: First, the location entities in the hail event elements are analyzed and recalled according to the administrative division database, place name dictionary and POI database to generate a candidate location or a set of candidate locations; for each candidate location, the positioning information is specified, and the latitude and longitude points and error radius are output. For fuzzy location entities without corresponding latitude and longitude points, the geographic boundary vector polygon and spatial accuracy level are output. For each candidate position, the overall credibility score is calculated using the following formula (4): （4）； In equation (4), Representing the The overall credibility score of each candidate position. Representing the The text name matching score of each candidate location is obtained by calculating the text similarity between the candidate location and the standard place name of the candidate location in the administrative division database and place name dictionary. Representing the The context score of each candidate location is calculated by matching the candidate location with the context location description in the original text. Representing the The POI matching score for each candidate position is calculated by matching the candidate position with the matching objects in the POI library. Representing the The cross-clue evidence score of each candidate location is calculated by spatially matching the candidate location with the other hail clue locations. Representing the The meteorological region overlap score of each candidate location is obtained by spatially matching the candidate location with the meteorological support area. , , , and This represents the weight coefficient of each item; all weight coefficients are non-negative, and the sum of all valid item weights is 1. If there are unusable items, remove the corresponding items and renormalize the remaining valid weights; The candidate positions in a set of candidate positions are ranked according to their comprehensive credibility scores. When the meteorological support area has not yet been generated, the comprehensive confidence score of the candidate location is calculated without calculating the meteorological area overlap score, and the weights of the remaining valid sub-items are renormalized. After the meteorological support area is generated, the meteorological region overlap score is introduced, the comprehensive credibility score of the retained candidate locations is recalculated, and the ranking is updated a second time. S34. For suspected hail clues that have been extracted from hail event elements and have undergone time and space normalization, if there are independent images, videos or comments on the same webpage corresponding to them, the suspected hail clues are split into multiple sub-objects, each of which is a network evidence object; otherwise, a corresponding network evidence object is directly generated for the suspected hail clues. The suspected hail clue before splitting is used as the parent original clue. The parent original clue and its corresponding multiple child objects are associated. The parent source identifier of the child object is taken as the unique evidence identifier of the parent original clue. The network evidence objects are assembled using fixed fields, which include a unique evidence identifier, parent original clue record number, source type, content carrier type, publication time, parent source identifier, candidate time window, candidate spatial region, hail size range, disaster impact label, media fingerprint, original source link, hail event element extraction confidence level, original source group number, conflict marker, and processing version.

5. The hail event quality control calibration method according to claim 4, characterized in that: The specific steps for generating candidate hail events are as follows: First, based on all network evidence objects corresponding to the suspected hail clue set, multi-dimensional evidence similarity is calculated pairwise. When the multi-dimensional evidence similarity meets the same-origin threshold requirement, a propagation edge is established to complete the construction of the evidence propagation graph. Multiple original source groups are then obtained based on the propagation edges. Next, for all network evidence objects corresponding to the suspected hail clue set, candidate events are clustered based on spatiotemporal, semantic, evidence consistency, and same-origin or conflict penalties. S41. Construct an evidence propagation graph G=(V,E), where the nodes in the node set V are network evidence objects, and the edge set E consists of propagation edges; For any two network evidence objects, the multi-dimensional evidence similarity is calculated using the following formula (5). : （5）； In equation (5), Representative node and nodes Multidimensional evidence similarity between them Representative node and nodes Text similarity between them Representative node and nodes Perceptual hash similarity between images Representative node and nodes Similarity of keyframes between videos Representative node and nodes Audio fingerprint similarity between them Represents the similarity of link references. , , , and The weight coefficients represent the weights of each dimension. When two nodes belong to the same content modality, the similarity corresponding to that content modality and the link reference similarity are calculated, and the similarity components actually involved in the calculation are weighted and summed. When two nodes belong to different content modalities, the link reference similarity is used as the multi-dimensional evidence similarity. The weight coefficients corresponding to similarity items that were not included in the calculation are not included in this calculation. The effective weight coefficients corresponding to the similarity items that were actually included in the calculation are renormalized so that the sum of the effective weight coefficients is 1. when Not lower than the set homogeneity threshold When the two nodes are connected, a propagation edge is established between them and added to the edge set E. After completing the multi-dimensional evidence similarity calculation and edge construction for all network evidence objects, perform connected component partitioning or community discovery on the evidence propagation graph to obtain multiple original source groups; each original source group is generated based on graph connectivity, which includes evidence pairs that meet the multi-dimensional evidence similarity criteria, as well as capturing the indirect propagation homologous evidence set formed by chain retransmission; The credibility of all network evidence objects within each original source group is aggregated to obtain the aggregate credibility of that original source group, and an equivalent evidence object is generated. When the credibility is aggregated, the aggregate credibility of the original source group is equal to the maximum base credibility of all network evidence objects within the original source group, or the aggregate credibility obtained by attenuation according to the propagation distance. The equivalent evidence object is a virtual evidence record, whose fixed fields are selected from the network evidence object corresponding to the maximum basic credibility or are obtained by merging the fields of all network evidence objects in the original source group. If there is only one network evidence object in the original source group, and this network evidence object is different from other network evidence objects describing the same hail event in terms of account, shooting angle, media fingerprint, and spatiotemporal information describing the event, then it is regarded as an independent evidence object, and its basic credibility is retained as the aggregate credibility of the original source group. The basic credibility of the network evidence object is calculated using the following formula (6): （6）； In equation (6), Representing the The credibility of the source of each piece of online evidence is derived from the source type; Representing the The credibility of the evidentiary form of an individual network evidence object is derived from the type of content carrier; Representing the The reliability of the temporal precision of each network evidence object is derived from the candidate time window; The reliability of the location accuracy is derived from the candidate spatial region; Representing the The originality and credibility of each piece of online evidence is obtained by comparing and retrieving media fingerprints and original source links; Representing the The conflict penalty for each network evidence object is derived from the conflict marker; Representing the The punishment for reposting old news by an individual online evidence subject is derived from media fingerprints and historical material databases retrieved based on publication time. , , , , , and Represents the weighting coefficients of each item; The pruning function forces the calculation result to be restricted to... The range; among which, source credibility, evidence form credibility, time accuracy credibility, location accuracy credibility and originality credibility are positive sub-items, each sub-item weight is a non-negative number, and the sum of the sub-item weights is 1; the weights of conflict penalty and old news reposting penalty are set separately. S42. Calculate the event association score using the following formula (7). : （7）； In equation (7), Representative object and object The event correlation score is based on whether the object is an equivalent evidence object or an independent evidence object; It represents the temporal similarity between two objects, that is, the degree of overlap between the candidate time windows of the two objects; It represents the spatial similarity between two objects, that is, it calculates the degree of overlap or distance decay between the candidate spatial regions of the two objects; It represents the semantic similarity between two objects, that is, calculating the matching similarity of the semantic tags of hail event elements of the two objects; It represents the consistency of evidence between two objects, that is, the degree of agreement between the values ​​of structured elements in the hail event elements of the two objects; The same-origin or conflict penalty between two objects is calculated from the original source group number and conflict flag of the two objects; , , , and The weight coefficients represent the weights of each item; among them, temporal similarity, spatial similarity, semantic similarity, and evidence consistency are positive items, each item's weight is a non-negative number, and the sum of the item weights is 1; the weights for homology or conflict penalties are set separately. When the event is associated with the score Not less than the clustering threshold If the hard conflict conditions of being geographically distant and having completely disjoint dates are not met, then an edge with the same event is established between the two objects, and all objects are grouped and clustered according to all edges with the same event by connected component partitioning, hierarchical clustering or DBSCAN algorithm to obtain multiple candidate event clusters. The aggregated confidence scores of all objects within each candidate event cluster are read and fused to obtain the event confidence score of that candidate event cluster; the structured information of all objects within each candidate event cluster is aggregated to obtain the candidate hail events.

6. The hail event quality control calibration method according to claim 5, characterized in that: After obtaining multiple candidate event clusters, post-verification processing is performed: the maximum time span and maximum spatial span among all objects within the candidate event cluster are calculated, and the consistency of meteorological units corresponding to the candidate event clusters is verified. When the maximum time span or the maximum spatial span exceeds the preset threshold, or when meteorological units are inconsistent, the candidate event cluster is split into two sub-clusters to obtain the split candidate event clusters; the calculation of the event confidence and the aggregation of candidate hail events are performed after post-verification processing.

7. The hail event quality control calibration method according to claim 5, characterized in that: The specific steps for generating and matching meteorological support zones for candidate hail events, applying hard conflict rule gating to candidate hail events, identifying irreparable hard conflicts, and outputting corresponding processing labels are as follows: S51. Obtain meteorological data through the meteorological operational data interface, and convert the meteorological data into various meteorological support areas, which are divided into radar support areas, dual-polarization support areas, path support areas, and surface support areas; calculate the time interval overlap between the candidate time windows of candidate hail events and the corresponding time slices of the meteorological support areas; calculate the area intersection-union ratio, nearest distance, and path relationship between the candidate spatial regions of candidate hail events and the meteorological support areas; combine the time interval overlap calculation results and the area intersection-union ratio, nearest distance, and path relationship calculation results into the spatiotemporal matching result between the candidate hail events and the meteorological support areas. S52. Execute hard conflict rule gating: Perform hard conflict rule matching for each candidate hail event, determine whether each rule meets the triggering conditions, and execute the corresponding handling action when the triggering conditions are met. Finally, summarize the handling results of all rules and output the handling flags. The handling flags include exclusion flags, downgrade flags, manual review flags, and conflict tags. Conflict tags include supplementary conflict tags and non-supplementary hard conflict tags.

8. The hail event quality control calibration method according to claim 7, characterized in that: After calculating the credibility of network evidence, meteorological support, multi-source cross-validation gain, and conflict penalty for candidate hail events, a weighted average is calculated to obtain the comprehensive credibility. This is then combined with a quality control level output using a hard conflict rule-gated output label. The specific steps are as follows: S61, Original source group Aggregate credibility Substituting into equation (8), the baseline value of the candidate hail event is calculated. ; （8）； In equation (8), Representing the original source group Weighting coefficients; The strength of multi-source mutual verification is calculated using the following formula (9). : （9）； In equation (9), This represents the total number of original source groups corresponding to the current candidate hail event. It is an integer not less than 1; Represents the saturation coefficient. Greater than 0 controls the rate at which saturation occurs; when When =1, the strength of multi-source mutual authentication is 0; as As the value increases, the strength of multi-source mutual verification monotonically increases and gradually approaches 1; the strength of multi-source mutual verification... Represents the strength of mutual corroboration among the original source groups, with a value range of [value range missing]. ; Finally, the credibility of the network evidence for the candidate hail event is calculated using the following formula (10). : （10）； In equation (10), The weighting coefficients representing the strength of multi-source mutual verification. Weighting coefficients representing the baseline values ​​of candidate hail events; using The function constrains the fusion result to The interval; S62. Calculate the meteorological support of candidate hail events according to the following formula (11). : （11）； In equation (11), Used to measure the degree of support from professional meteorological data for the existence of hail or severe convective conditions in the candidate spatial region of a candidate hail event. The time matching degree is represented by calculating the time distance between the meteorological observation time and the time window of the candidate hail event. The spatial matching degree is obtained by calculating the spatial distance or the percentage of spatial overlap between the meteorological observation location and the spatial region of the candidate hail event. The radar hail support is obtained by calculating the radar-identified hail potential in the spatial region of the candidate hail event. This represents the dual polarization support, obtained by calculating the significance of hail features; The score represents the support for lightning, which is obtained by statistically analyzing the frequency and distance of lightning within the spatial region of the candidate hail event. The higher the frequency of lightning and the closer the observation location is to the spatial region of the candidate hail event, the higher the score. The satellite's convective support is obtained by extracting convective features and identifying strong convective potential. The weather support level of the automatic weather station is calculated by extracting environmental meteorological features that are conducive to the occurrence of hail events and giving a comprehensive score. Representing ground record support, this is determined by statistically analyzing whether there are actual records of hail in the spatial area of ​​the candidate hail event; the more records, the higher the score. , , , , , , and The weight coefficients represent the weights of each item. These weight coefficients are global hyperparameters and are optimally determined from the historical hail validation set. When all data sources are valid, the sum of all weight coefficients equals 1. When a meteorological data source is missing, the weight coefficients corresponding to the missing items are removed, and the remaining valid weight coefficients are renormalized. At the same time, a data availability identifier is output to mark the availability status of each sub-item data source. S63. Calculate the multi-source cross-validation gain using the following formula (12). : （12）； In equation (12), The sub-item consistency score represents the pairwise comparison results between two original source groups for the same candidate hail event. The comparison types include cross-platform comparison, cross-user comparison, cross-media comparison, and cross-category source comparisons conducted by the network multi-source group with radar, radar dual polarization, and ground real-time records, respectively. Each comparison outputs a corresponding sub-item consistency score. The gain coefficient representing the consistency score of each sub-item is a hyperparameter; subscript An index representing the consistency score of each item; each set of comparisons corresponds to one. ; The gain effect of cross-validation of information from multiple sources is characterized; only valid comparisons between different original source groups are included in the product operation; when there are no valid comparisons, the gain of cross-validation of multiple sources is 0, and reprints within the same original source group are not counted repeatedly. S64. Calculate the conflict penalty using the following formula (13). : （13）； In equation (13), The conflict severity represents the severity of the conflict, which is not supported by old news, time, location, intensity, media or weather. The conflict label is a supplementary conflict label, which is obtained by comparing the conflict situation of multiple sources of information under the same candidate hail event and assigning a value according to the intensity of the conflict. Represents the conflict weight, which is a hyperparameter; Characterizes the comprehensive conflict degree of multi-source information of candidate hail events; non-supplementable hard conflicts are not included in the calculation of equation (13) and are directly processed by hard conflict rule gating; S65. The overall confidence level of candidate hail events is calculated using the following formula (14): （14）； In equation (14), , , The weights are non-negative and their sum is 1. This is the conflict penalty coefficient; After calculating the overall credibility of the candidate hail event, the candidate hail event is classified into the corresponding preliminary quality control level according to the credibility level classification principle. Various processing labels are received from the hard conflict rule gating output. The processing labels are used to exclude, downgrade, add manual review mark and limit the highest level of the preliminary quality control level to obtain the final quality control level of the candidate hail event.

9. A hail event quality control calibration method according to claim 8, characterized in that: The specific steps for performing quality control calibration on the time, location, and intensity of candidate hail events under rule constraints, retaining the original information of the candidate hail events, and generating a credible hail event record containing a chain of evidence, quality control level, uncertainty range, and calibration instructions are as follows: S71. Time Calibration: When the time range of a candidate hail event intersects with the effective time interval corresponding to the meteorological support area, if there is only one overlapping intersection time period, it is used as the recommended time window after meteorological constraints. If there are multiple overlapping intersection sub-time periods, the meteorological support corresponding to each intersection sub-time period is calculated, and the intersection sub-time period with the highest meteorological support is selected as the recommended time window after meteorological constraints. At the same time, the original time range of the original candidate hail event is retained, and time conflicts are marked to retain the uncertainty of time. When there is no overlap but the time interval is within the set buffer threshold, Bayesian or weighted updates are performed by combining the radar single-unit transit time, the dual-polarization hail identification time, the lightning jump time, and the independent network clue sequence to obtain the recommended time window after meteorological constraints; at the same time, the original time range of the original candidate hail events is retained, and time conflicts are marked to retain the uncertainty of time. When there is no overlap but the time interval exceeds the set buffer threshold, the original time range of the original candidate hail event is retained, the time conflict is marked, the time of the event is not modified, and the uncertainty of the time is retained. S72. Location calibration: When the candidate spatial region of a candidate hail event overlaps with the meteorological support region, if there is only a single overlapping region, the overlapping region is taken as the candidate spatial region after meteorological constraints. If there are multiple discontinuous overlapping sub-regions, the meteorological support degree corresponding to each overlapping sub-region is calculated, and the overlapping sub-region with the highest meteorological support degree is selected as the candidate spatial region after meteorological constraints. At the same time, the original candidate spatial regions of the original candidate hail events are preserved, and location conflicts are marked to preserve spatial uncertainty; When the candidate spatial region description of a candidate hail event is vague, but the meteorological cell path information in the meteorological support area is clear, the candidate spatial region of the candidate hail event is narrowed down to the intersection of the meteorological cell path, the dual-polarized hail region, and the multiple ground clues corresponding to the candidate hail event, to obtain the meteorologically constrained candidate spatial region. At the same time, the original candidate spatial regions of the original candidate hail events are preserved, and location conflicts are marked to preserve spatial uncertainty; When the candidate spatial region of a candidate hail event does not overlap with the meteorological support area, the quality control level of the candidate hail event is reduced, the original candidate spatial region of the original candidate hail event is retained, and the location conflict is marked to retain spatial uncertainty. S73, Intensity Calibration: The hail size descriptions obtained from suspected hail cues for candidate hail events are mapped to initial size intervals with upper and lower boundaries. Based on image scale, video target references, manual observations, and official disaster damage records, the intensity confidence of the suspected hail cues is adjusted. The rationality of the initial size interval is then verified using radar and dual polarization features of the meteorological support area. Finally, the upper and lower limits of hail size, hail intensity level, impact level, and intensity confidence are output. After calibration, the intensity descriptions of the original candidate hail events are retained, and intensity conflicts are marked to preserve the uncertainty of intensity. S74. Based on the constructed evidence propagation graph, the nodes in the evidence propagation graph include original records, evidence objects, original source groups, candidate hail events, time windows, spatial regions, meteorological support areas, ground records, rule actions, and calibration results; the edges in the evidence propagation graph include origin, same source, support, correction, conflict, time matching, spatial matching, and calibrations generated therefrom. Each edge stores the relationship score, rule or model version, generation time, and original data index; thereby tracing back the formation basis of any quality control level, time calibration, location calibration, or intensity calibration, a credible hail event record containing the evidence chain, quality control level, uncertainty range, and calibration description is generated; Among them, the evidence propagation graph is used for the same source identification of network evidence objects; based on the evidence propagation graph, the nodes and corresponding relationship edges corresponding to the rule actions and calibration results are improved and expanded to form a full-process evidence chain graph.

10. A calibration system for implementing the hail event quality control calibration method of claim 1, characterized in that: It includes modules for data collection and storage, clue identification, event element extraction, spatiotemporal normalization, evidence object generation, source identification, candidate event clustering, meteorological support area matching, quality control rule engine, credibility assessment, quality control calibration, and event recording. The data acquisition and retention module is used to acquire multi-source data related to hail events and to complete the raw storage and management of multi-source data; The clue identification module is used to calculate the relevance of clues from multiple sources, and after initial screening, it identifies a set of suspected hail clues. The event element extraction module is used to extract hail event elements from each suspected hail clue in the suspected hail clue set, and retain the extraction confidence of hail event elements and corresponding multi-source tracing evidence; The spatiotemporal normalization module is used to perform spatiotemporal normalization processing on hail event elements; The evidence object generation module is used to encapsulate the normalized hail event elements of each suspected hail clue, their extracted confidence level, and the corresponding multi-source tracing evidence into a network evidence object. The same source identification module is used to calculate the multi-dimensional evidence similarity pairwise based on all network evidence objects corresponding to the suspected hail clue set. When the multi-dimensional evidence similarity meets the same source threshold requirement, a propagation edge is established to complete the construction of the evidence propagation graph, and multiple original source groups are obtained based on the propagation edge. The candidate event clustering module clusters candidate events based on spatiotemporal, semantic, evidence consistency, and homology or conflict penalty for all network evidence objects corresponding to the suspected hail clue set, and generates candidate hail events. The meteorological support area matching module generates and matches meteorological support areas for candidate hail events. The quality control rule engine module is used to perform hard conflict rule gating on candidate hail events, identify irreparable hard conflicts, and output the corresponding processing labels. The credibility assessment module is used to calculate the credibility of network evidence, meteorological support, multi-source cross-validation gain and conflict penalty of candidate hail events, and then calculate the comprehensive credibility by weighting, and combine the processing label output by hard conflict rule gating to output the quality control level. The quality control calibration module is used to perform quality control calibration on the time, location, and intensity range of candidate hail events under rule constraints; The event logging module is used to generate credible hail event logs that include the chain of evidence, quality control level, uncertainty range, and calibration instructions.

Images