Networking weather radar combined reflectivity false alarm identification and quality control method

Abstract

The present application relates to the technical field of weather radar data processing, and particularly relates to a networked weather radar combined reflectivity false alarm identification and quality control method, which comprises the following steps: acquiring historical continuous multi-time combined reflectivity data of networked weather radar, performing object-level connected region segmentation processing on the historical continuous multi-time combined reflectivity data to generate a historical echo object set; generating high-confidence reserved pseudo-labels and high-confidence false alarm pseudo-labels based on time sequence persistence scoring, and training a probability classification model; acquiring current-time combined reflectivity data, constructing a current object feature vector, outputting object reserved confidence, and generating a quality-controlled combined reflectivity product. The present scheme realizes improving the combined reflectivity false alarm identification accuracy and real-time quality control stability in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of weather radar data processing, specifically a method for identifying and controlling false alarms based on the reflectivity of networked weather radar systems. Background Technology

[0002] Weather radar is an important remote sensing device used to detect precipitation clouds, severe convective systems, and hazardous weather processes. It can continuously monitor precipitation intensity, spatial distribution, and echo evolution. With multiple weather radars networked together, the combined reflectivity products of these networked weather radars can integrate observations from different radar stations to form spatially continuous echo distribution information over a large area. This technology has been widely applied in operational scenarios such as short-term nowcasting, severe convection monitoring, automatic alarms, quantitative precipitation estimation, and hazardous weather risk warnings.

[0003] Currently, quality control for non-meteorological false alarm echoes in networked weather radar reflectivity products typically employs rule-based thresholding, fuzzy logic, temporal continuity analysis, or machine learning-based identification methods. Rule-based thresholding usually sets screening conditions based on reflectivity intensity, area, morphology, or spatial location to delete or weaken suspected false alarm areas. Fuzzy logic typically combines multiple echo features to construct comprehensive discrimination rules and score the confidence level of target areas. Temporal continuity analysis typically utilizes the persistence of echoes in adjacent time intervals, trajectory correlations, or echo evolution patterns to identify short-term isolated echoes. Machine learning-based methods typically extract area, texture, morphological, or statistical features of echo objects and use supervised classification models to classify and identify meteorological and non-meteorological echoes.

[0004] The inventors of this application have discovered the following technical problems in the actual network weather radar combined reflectivity service processing:

[0005] Existing networked weather radar reflectivity products contain non-meteorological false alarm echoes such as ground clutter, anomalous propagation, electromagnetic interference, and sidelobe interference. Traditional rule-based thresholding methods typically rely on fixed empirical parameters, which have limited adaptability to different regions, radar layouts, and complex interference scenarios, making it difficult to reliably identify isolated, fragmented, or strip-shaped non-meteorological echoes. Some supervised machine learning methods rely on a large number of manually labeled samples, but the large number and rapid changes of objects in networked radar echoes result in high manual labeling costs and make it difficult to sustain in the long term.

[0006] Furthermore, existing false alarm identification methods based on temporal continuity typically require the use of future time-series data, multi-time-series trajectory correlation, or echo evolution information. While these methods can leverage echo persistence to improve identification capabilities, they introduce latency in online processing, making it difficult to meet the real-time processing requirements of current single-frame combined reflectivity products. Conversely, while existing single-frame identification methods offer real-time performance, the lack of temporal persistence constraints makes it difficult to effectively distinguish between short-term isolated false alarm echoes and genuine weak-life-history meteorological echoes. Moreover, current technologies lack a solution that can enable models to utilize historical temporal persistence patterns without relying on future time-series data and continuous trajectory information, thus achieving real-time false alarm identification of current single-frame networked weather radar combined reflectivity objects. In simpler terms, existing methods often face the contradiction of "real-time operation is difficult to achieve with temporal continuity, while utilizing temporal continuity is difficult to achieve in real-time processing." Summary of the Invention

[0007] This application provides a method for false alarm identification and quality control of combined reflectivity of networked weather radar, which solves the problems of existing false alarm identification methods relying on future time data, insufficient real-time performance, and low stability of false alarm identification in complex scenarios.

[0008] To achieve the above objectives, the embodiments of this application disclose the following technical solutions:

[0009] This solution discloses a method for identifying and controlling false alarms based on the reflectivity of networked weather radar systems, including:

[0010] Acquire historical continuous multi-time combined reflectivity data of networked weather radar; historical continuous multi-time combined reflectivity data of networked weather radar is used to characterize the spatial evolution process of historical weather systems in the continuous time dimension, so that the subsequent constructed time series persistence score can be based on the continuous propagation law of real weather echoes, rather than static judgment based on discrete instantaneous echo states.

[0011] Object-level connected region segmentation is performed on historical continuous multi-time combined reflectivity data to generate a set of historical echo objects. Object-level connected region segmentation maps the continuously spatially distributed echo grid into a set of historical echo objects with complete boundary structures, so that the subsequent time series correlation process no longer depends on single-pixel-level random changes, thereby reducing the interference of isolated noise points on time series stability analysis.

[0012] A temporal persistence score is constructed based on the spatial overlap, morphological continuity, and reflectivity persistence relationships between historical echo objects from adjacent time periods. High-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels are then generated based on the temporal persistence score. The temporal persistence score does not simply count the existence time of objects, but utilizes the continuous and stable characteristics of real meteorological echoes in the spatial structure and reflectivity propagation process to automatically map historical echo objects with stable evolutionary relationships as high-confidence retained pseudo-labels, and to map short-term abrupt changes or spatially discrete objects as high-confidence false alarm pseudo-labels. This solves the problems of high cost of traditional manual annotation and difficulty in building large-scale training samples in complex weather scenarios.

[0013] Spatial structural features, reflectivity statistical features, spatial connectivity features, and radar geometric features of historical echo objects are extracted to construct a feature vector for historical objects. Spatial structural features are used to characterize the boundary stability and overall morphological characteristics of the echo object; reflectivity statistical features are used to characterize the energy distribution pattern; spatial connectivity features are used to characterize the degree of regional clustering; and radar geometric features are used to characterize the spatial relationship between the echo object and the radar observation direction. This results in a multi-dimensional feature vector for historical objects that can simultaneously reflect the differences between real meteorological echoes and anomalous radial structures.

[0014] A probabilistic classification model is trained based on historical object feature vectors, high-confidence retained pseudo-labels, and high-confidence false alarm pseudo-labels, generating a trained probabilistic classification model. This model establishes an implicit mapping between stable evolutionary patterns across consecutive historical timeframes and single-frame object features, enabling the model to identify false alarms online using only the combined reflectivity data of the current single frame, without waiting for data from future timeframes.

[0015] Obtain the combined reflectance data of the current single frame; this data is used to construct the real-time inference input for the online business phase, thereby decoupling the training and online inference phases.

[0016] The combined reflectance data of the current time frame is processed by object-level connected component segmentation to generate the current time echo object set. The current time echo object set is used to maintain the consistency of object structure representation between the online and offline training phases, thereby avoiding instability of the probabilistic classification model inference due to differences in input structure.

[0017] The spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features of the current echo object are extracted to construct the current object feature vector. The current object feature vector is used to map the current echo object to a unified feature space formed by historical training, so as to inherit the historical time series stability knowledge.

[0018] The current object's feature vector is input into the trained probabilistic classification model, which outputs the object retention confidence score. The object retention confidence score represents the probability level that the current echo object belongs to a real meteorological echo, thus avoiding the boundary deletion problem caused by traditional fixed rule discrimination methods in complex weather scenarios.

[0019] Based on the object retention confidence level, the corresponding echo object is subjected to retention, attenuation, or deletion processing, and the processing results are backfilled into the current time-series combined reflectivity grid to generate the quality-controlled combined reflectivity product. Through a probabilistic hierarchical control method, weak true echoes can be retained while suppressing false alarm echoes, thereby improving the structural continuity and operational reliability of the quality-controlled combined reflectivity product under complex weather conditions.

[0020] Furthermore, to avoid the problem that traditional single-time existence judgment methods cannot accurately distinguish between real meteorological echoes and random false alarm echoes, this application does not only rely on the number of times an object appears for persistence judgment, but also simultaneously introduces the object area change rate, principal axis direction change rate, average reflectivity change rate, and boundary overlap ratio to establish a consistent relationship of the object's structural evolution in the continuous time dimension. Real meteorological echoes usually have stable spatial expansion patterns and continuous reflectivity change patterns during propagation, while false alarm echoes are more likely to exhibit structural abrupt changes, boundary discrepancies, and abnormal directional changes. Therefore, continuous clustering can automatically form stable and unstable object sets. This method is not a simple time-series tracking, but rather transforms the temporal continuity features into pseudo-label constraints that can be learned by the subsequent probabilistic classification model, thereby providing a temporal knowledge transfer foundation for single-frame inference in the online stage.

[0021] Furthermore, to avoid the problem of difficulty in distinguishing between true weak echoes and anomalous propagation echoes due to relying solely on reflectivity intensity features, this application does not employ a single-dimensional feature discrimination method. Instead, it constructs a unified object feature vector through spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features. True convective echoes typically exhibit high spatial clustering and stable boundary structures, while radial interference echoes show anomalous consistency between the principal axis direction angle and the radar radial azimuth angle. Therefore, multi-dimensional normalization fusion processing can form a more stable classification boundary and improve the generalization ability under complex weather scenarios.

[0022] Furthermore, to avoid the probabilistic classification model learning only single reflectivity intensity differences and ignoring structural continuity features, this application does not employ a single-objective loss constraint approach. Instead, it simultaneously introduces temporal consistency constraints, spatial morphological consistency constraints, and radial structural consistency constraints. Temporal consistency constraints are used to maintain the continuous evolutionary relationship of historically stable objects; spatial morphological consistency constraints are used to maintain the structural stability of the true echo boundary; and radial structural consistency constraints are used to suppress the influence of radial stripe interference and sidelobe echoes on the classification boundary. Through the joint constraint of the loss function, the probabilistic classification model can still inherit the stable evolutionary patterns from continuous historical time intervals under single-frame input conditions, thus enabling the probabilistic classification model to utilize these stable evolutionary patterns even under single-frame input conditions.

[0023] Furthermore, to avoid the problem of falsely deleting weak true echoes entirely due to traditional hard threshold deletion methods, this application adopts a hierarchical probability control mechanism instead of a single binary discrimination mechanism. When the object retention confidence is in the middle range, the corresponding echo object is not directly deleted, but reflectivity attenuation processing is performed to preserve its potential true structural information. This method can achieve a dynamic balance between false alarm suppression capability and true echo retention capability, and improve the continuous stability of the combined reflectivity product after quality control in the boundary region.

[0024] Furthermore, to avoid the problem that single-scale structural analysis cannot identify weak convective boundaries and complex nested structures, this application establishes a multi-scale stability analysis mechanism through local-scale, meso-scale, and global-scale regions. The local-scale region is used to perceive changes in boundary details, the meso-scale region is used to identify object expansion trends, and the global-scale region is used to maintain overall spatial continuity. By performing enhancement processing on the object feature vector through multi-scale stability scoring, the probabilistic classification model can simultaneously obtain local texture features and overall spatial evolution features, thereby improving the recognition stability in complex weather scenarios.

[0025] Furthermore, to maintain temporal continuity discrimination capability even when future time-series data is unavailable during the online phase, this application does not directly utilize the current single-frame state for static judgment. Instead, it simulates object diffusion trends and reflectivity changes through a pseudo-evolutionary sliding window to generate pseudo-temporal evolution data. Furthermore, it dynamically adjusts the object retention threshold by combining local clutter density and regional echo spread rate, allowing the object retention conditions to dynamically change with the current spatial environment. This approach can improve false alarm suppression strength in high-clutter regions while reducing the probability of false deletion in truly weak-echo regions, thereby enhancing scene adaptability during the online inference phase.

[0026] Furthermore, to suppress anomalous objects with obvious radar radial structure characteristics, such as radial stripe interference, sidelobe echoes, and anomalous propagation echoes, this application does not rely solely on reflectivity intensity for anomaly identification. Instead, it establishes a radial consistency analysis mechanism between the principal axis direction angle and the radar radial azimuth angle. Real meteorological echoes typically exhibit strong spatial spread randomness, while anomalous radial structure objects are more likely to align with the radar radial direction. Therefore, by using radial conflict scoring and object suppression weight correction mechanisms, the interference of anomalous radial structures on the probability classification results can be effectively reduced, improving the spatial structure stability of the combined reflectivity product.

[0027] Furthermore, to avoid the interference of discrete noise pixels and local void structures on the subsequent feature extraction results, this application does not directly perform feature analysis based on the original combined reflectivity grid. Instead, it first forms standardized echo objects through neighborhood connectivity analysis, small region culling, and void filling. Further, boundary smoothing and morphological uniformity processing are used to eliminate local boundary spikes and random deformations, enabling the object-level echo candidate region set to possess a more stable spatial structure representation capability, thereby improving the structural consistency in the subsequent object feature vector construction process.

[0028] Furthermore, to prevent the classification boundary of the probabilistic classification model from gradually becoming unstable due to seasonal changes, terrain differences, and changes in radar operating status during long-term operation, this application establishes an incremental feedback training mechanism rather than a static operation mechanism with fixed parameters. By dynamically updating the weight parameters in the joint loss function through object false alarm rate, region residual rate, and structural deviation rate, the probabilistic classification model can continuously adapt to the changing patterns of false alarm structures in different business environments, thereby improving the model's stability and scenario generalization ability under long-term operating conditions.

[0029] This invention acquires historical, continuous, multi-time-time combined reflectivity data from a networked weather radar, constructs a temporal persistence score, and generates high-confidence retention pseudo-labels and high-confidence false alarm pseudo-labels. This enables the probabilistic classification model to learn the differences between real meteorological echoes and false alarm echoes in terms of temporal continuity, spatial structure stability, and radial structure consistency. This solves the problems of existing technologies that rely solely on single-frame threshold judgments, which easily lead to the accidental deletion of real weak echoes, and that relying on multi-time-time trajectory correlations is insufficient to meet real-time operational requirements. In practical applications, this solution does not simply look at the reflectivity magnitude at a single moment. Instead, it maps the stable evolution patterns from continuous historical time periods to the object retention confidence judgment process of the current single-frame echo object, much like "learning historical patterns first and then quickly judging the current situation." Therefore, even when only the current single-frame combined reflectivity data is acquired during the online phase, high false alarm identification stability can still be maintained. Meanwhile, through the coordinated processing of object-level connected region segmentation, multi-scale stability scoring, pseudo-temporal evolution data construction, and radial consistency correction, the impact of radial stripe interference, sidelobe echoes, and anomalous propagation echoes on the combined reflectivity product can be effectively suppressed, and the risk of misjudgment caused by random isolated noise can be reduced, thereby improving the spatial structural integrity, business availability, and long-term operational stability of the combined reflectivity product after quality control. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the overall method of the present invention according to an embodiment of the present invention;

[0031] Figure 2 This is a flowchart of the time-series persistence scoring and pseudo-label generation process according to an embodiment of the present invention;

[0032] Figure 3 This is a flowchart of the probabilistic classification model training process according to an embodiment of the present invention;

[0033] Figure 4 This is a flowchart illustrating the echo object retention, attenuation, and deletion processes according to an embodiment of the present invention.

[0034] Figure 5 This is a flowchart of the multi-scale enhancement processing according to an embodiment of the present invention. Detailed Implementation

[0035] 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 a part of the embodiments of the present invention, and not all of them. 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. In the following description, numerous specific details are set forth to provide a comprehensive understanding of the present invention. The present invention may be practiced without some or all of these specific details. In other instances, well-known processes have not been described in detail to avoid unnecessarily obscuring the present invention.

[0036] When used in conjunction with the terms "comprising," "method comprising," or similar language in this specification and appended claims, the singular forms "a," "some," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0037] Terminology definition:

[0038] In this embodiment, "networked weather radar combined reflectivity data" refers to two-dimensional or three-dimensional reflectivity raster data formed by spatial stitching, time synchronization and coordinate unification of multiple weather radar stations. Its data units contain reflectivity intensity values, spatial coordinate information and time index information, which are used to characterize the distribution of meteorological echoes at different locations within the region.

[0039] "Object-level connected region segmentation processing" refers to performing pixel connectivity analysis, region clustering, and boundary consistency processing on high reflectivity regions in combined reflectivity data, extracting echo objects with independent spatial structures from continuous grids, and transforming subsequent processing from traditional pixel-level analysis to object-level analysis.

[0040] "Temporal persistence score" refers to a stability evaluation index obtained by quantitatively calculating the degree of continuous stability of historical echo objects in terms of spatial location, morphological structure and reflectivity evolution. It is used to reflect the persistence of the echo object's existence in the time dimension.

[0041] "High-confidence retained pseudo-labels" and "high-confidence false alarm pseudo-labels" are weakly supervised labels automatically generated based on the time-series stability characteristics in continuous historical time periods. The high-confidence retained pseudo-labels are used to represent real meteorological echo objects, while the high-confidence false alarm pseudo-labels are used to represent short-term isolated false alarm objects.

[0042] The "object feature vector" is a multi-dimensional feature set formed by uniformly encoding the spatial structural features, reflectivity statistical features, spatial connectivity features, and radar geometric features of the echo object. It is used to describe the comprehensive attributes of the current echo object.

[0043] The "probabilistic classification model" is a probabilistic mapping model trained on historical object feature vectors and pseudo-labels. It is used to retain confidence in the output object of the current single-frame echo during the online business phase.

[0044] The “pseudo-evolutionary sliding window” is a virtual temporal extension structure constructed based on the current single-frame combined reflectivity data. By performing diffusion, attenuation and spatial propagation simulations on the current echo object, a pseudo-temporal data structure is formed to estimate future evolution trends.

[0045] Example 1

[0046] This embodiment provides a method for false alarm identification and quality control based on reflectivity of a networked weather radar system. It transforms the traditional time-continuity-based false alarm identification logic, which relies on future time data, into a decoupled structure of "offline time-series knowledge learning and online single-frame real-time inference." In existing technologies, identification methods based on time-series continuity typically require waiting for future time data or rely on multi-frame trajectory correlation processing, thus failing to meet real-time business requirements. This solution automatically generates high-confidence pseudo-labels using historical multi-time data and uses these pseudo-labels to train a probabilistic classification model that relies solely on single-frame object features. This allows for false alarm identification without acquiring future time data during the online phase, thereby ensuring real-time performance while inheriting the ability to discriminate time-series continuity. This solution establishes a clear technical boundary with traditional rule-based thresholding methods, multi-frame trajectory tracking methods, and purely single-frame static classification methods. Its core lies not in simple classification but in achieving time-series knowledge transfer through an implicit mapping of historical time-series stability to the single-frame probability space; including:

[0047] Acquire historical continuous multi-timetime combined reflectivity data from the networked weather radar; perform object-level connected region segmentation on the historical continuous multi-timetime combined reflectivity data to generate a set of historical echo objects; the object-level connected region segmentation process includes: performing preset reflectivity threshold filtering on the current timetime single-frame combined reflectivity data to generate candidate echo regions; performing neighborhood connectivity analysis on the candidate echo regions to generate initial connected objects; performing small region culling and hole filling on the initial connected objects to generate standardized echo objects; performing boundary smoothing and morphology uniformity processing on the standardized echo objects to generate a set of object-level echo candidate regions; and using the set of object-level echo candidate regions as input data for subsequent object feature extraction processing.

[0048] In this embodiment, the historical continuous multi-time combined reflectivity data of the networked weather radar can be obtained from a national or regional weather radar operational platform. The data from each time period are stored at a uniform time interval, and a combined reflectivity grid is constructed with a fixed spatial resolution. First, a preset reflectivity threshold is applied to the combined reflectivity data to remove background noise regions and generate candidate echo regions. Then, based on eight-neighborhood or multi-neighborhood connectivity analysis, connectivity clustering is performed on the continuous grids in the candidate echo regions to generate initial connected objects. Small regions with areas below the preset region threshold are eliminated, and objects with internal void structures are filled to improve object integrity. Finally, morphological smoothing and principal axis unification are performed on the object boundaries to form a standardized set of echo objects. Preferably, boundary smoothing and morphological unification can be achieved through morphological closing operations and region principal direction fitting.

[0049] In this embodiment, to avoid the object breakage problem in weak echo boundary regions caused by traditional pixel adjacency determination methods, a spatial connectivity mapping model is first constructed based on the spatial distance relationship, reflectivity difference relationship, and local orientation consistency relationship between candidate pixels. Since real meteorological echoes typically exhibit high spatial continuity and gradual reflectivity changes in local regions, the spatial adjacency and reflectivity similarity between pixels can be jointly constrained to obtain the object region connectivity determination function:

[0050] ;

[0051] Furthermore, based on the object-level region connectivity determination function, an object-level region label propagation equation is constructed:

[0052] ;

[0053] in, This represents the object region connectivity determination result between candidate pixel i and candidate pixel j. The candidate pixels originate from the candidate echo regions in the combined reflectivity data of the current single frame. When C(i,j)=1, it means that candidate pixel i and candidate pixel j satisfy the spatial connectivity condition and belong to the same object-level echo region; when C(i,j)=0, it means that the two do not satisfy the object connectivity condition.

[0054] This represents the spatial distance between candidate pixel i and candidate pixel j, which can be calculated based on two-dimensional coordinates in the combined reflectivity grid.

[0055] This represents a preset spatial distance threshold, used to limit the maximum local connectivity span of an object region. The preset spatial distance threshold is obtained from the neighborhood range preset in the object-level connected region segmentation process.

[0056] This represents the reflectance difference between candidate pixel i and candidate pixel j, which is obtained by reading the reflectance of the corresponding grid in the combined reflectance data;

[0057] This indicates a preset reflectivity difference threshold, used to limit erroneous connections between regions of different intensities.

[0058] This represents the change in the boundary direction of the candidate pixel, which is calculated from the boundary gradient direction.

[0059] This represents a preset boundary direction threshold, used to constrain the continuity of the object's boundary structure. This represents the label of the object region corresponding to candidate pixel i. This represents the label of the object region corresponding to candidate pixel j.

[0060] By calculating the object region connectivity determination function, stable connectivity analysis between candidate pixels can be achieved, and erroneous object connections caused by abrupt changes in reflectivity, isolated noise points, or abnormal boundary direction changes can be avoided. Furthermore, by using the object-level region label propagation equation, candidate pixels that meet the spatial connectivity conditions can be merged under a unified object region label, thereby generating a complete set of object-level echo candidate regions, which can then be used as input data for subsequent object feature extraction processing.

[0061] Specifically, when a candidate pixel satisfies the spatial connectivity constraint, the corresponding region label is propagated to the target pixel to form an object-level connected region.

[0062] By segmenting connected components at the object level, traditional radar echo analysis based on discrete pixels can be transformed into an analysis method based on complete spatial structures, thereby reducing the impact of random isolated pixel noise on subsequent identification processes. Simultaneously, boundary smoothing and hole filling processes improve the stability of the object's spatial structure, providing a unified structural foundation for subsequent temporal persistence analysis.

[0063] A temporal persistence score is constructed based on the spatial overlap, morphological continuity, and reflectivity persistence relationships among historical echo objects from adjacent time periods. High-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels are generated based on the temporal persistence score. The calculation of the temporal persistence score includes: performing centroid position matching processing on echo objects from adjacent time periods to generate a candidate associated object set; calculating the object area change rate, principal axis direction change rate, average reflectivity change rate, and boundary overlap ratio based on the candidate associated object set; generating an object temporal stability index based on the area change rate, principal axis direction change rate, average reflectivity change rate, and boundary overlap ratio; calculating the temporal persistence score based on the object temporal stability index; performing persistent clustering processing on historical consecutive multi-time period echo objects based on the temporal persistence score to generate a stable object set and an unstable object set; marking the stable object set as a high-confidence retained pseudo-label and the unstable object set as a high-confidence false alarm pseudo-label. The boundary overlap ratio is used to characterize spatial overlap relationships, the area change rate and the principal axis direction change rate are used to characterize morphological continuity relationships, and the average reflectance change rate is used to characterize reflectance continuity relationships.

[0064] In this embodiment, adjacent time-series correlation analysis is performed on historical echo objects from multiple consecutive time periods. Specifically, for echo objects in adjacent time periods, the centroid distance is first calculated, and a candidate associated object set is established based on a preset distance threshold. Then, the object area change rate, principal axis direction change rate, average reflectivity change rate, and boundary overlap ratio are calculated to characterize the object's evolutionary stability over time. Subsequently, an object temporal stability index is constructed based on the above indicators, and persistent clustering is performed on objects from multiple consecutive time periods. Objects with continuous and stable structures are classified into a stable object set, while objects that appear in short-term isolation or undergo structural abrupt changes are classified into an unstable object set. Finally, the stable object set is marked with high-confidence retained pseudo-labels, and the unstable object set is marked with high-confidence false alarm pseudo-labels.

[0065] In this embodiment, to improve the ability to characterize the persistence of real meteorological echoes, spatial overlap, morphological change, and reflectivity persistence are jointly modeled. Since real echoes typically exhibit high boundary overlap, low morphological abrupt changes, and relatively stable reflectivity changes between adjacent time intervals, an object temporal persistence mapping relationship is constructed based on the aforementioned multidimensional temporal relationships, thereby obtaining the temporal persistence scoring equation:

[0066] ;

[0067] in: 1;

[0068] Furthermore, to enhance the stability discrimination capability of objects across multiple time periods, a temporal stability enhancement function is constructed based on the persistent cumulative relationship in historical consecutive time periods:

[0069] ;

[0070] in, This represents the temporal persistence score between adjacent echo objects. The temporal persistence score is calculated jointly by spatial overlap, morphological change, and reflectivity persistence.

[0071] This indicates the spatial overlap ratio between adjacent echo objects. The spatial overlap ratio is obtained by calculating the ratio of the number of overlapping grids between two object regions to the number of grids in the joint region. The relevant object regions are derived from the set of object-level echo candidate regions in historical continuous multi-time combined reflectivity data.

[0072] This represents the rate of morphological change between adjacent echo objects. The rate of morphological change is calculated based on the differences in area parameters, principal axis direction parameters, and boundary structure parameters, and is used to characterize the continuity of the object's morphology.

[0073] This represents the average rate of change of reflectivity between adjacent echo objects. The average rate of change of reflectivity is obtained by reading the average reflectivity of objects in the combined reflectivity grid.

[0074] These represent the weights for spatial overlap, morphological continuity, and reflectivity continuity, respectively. All are preset weight parameters, which can be obtained by presetting based on historical business statistics or model training results, and must meet the weight normalization constraint.

[0075] The temporal stability enhancement score represents the object's overall stability over a long time series, representing the object's overall stability.

[0076] N represents the number of consecutive historical time periods involved in the continuous cumulative calculation. The number of consecutive historical time periods is determined by the time window length of the combined reflectance data from multiple consecutive historical time periods.

[0077] This represents the time series persistence score corresponding to the kth historical time period;

[0078] This represents the time decay weight corresponding to the kth historical time. The time decay weight is calculated based on the time interval between the historical time and the current analysis time, and is used to enhance the influence of the nearest time objects on the stability judgment result.

[0079] By calculating the time-series persistence score, a quantitative analysis of the short-term structural stability of echo objects can be achieved. Furthermore, through the time-series stability enhancement function, the stability of objects across multiple consecutive time periods can be cumulatively enhanced, thereby improving the ability to distinguish between true persistent meteorological echoes and short-term random false alarm echoes. The final generated time-series stability enhancement score can be used to perform persistent clustering processing and generate high-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels.

[0080] Among them, the persistence scores at different times are subject to weight decay processing based on the time distance.

[0081] Real meteorological echoes typically exhibit strong temporal continuity, while non-meteorological false alarm echoes usually display short-term isolation, structural instability, or rapid dissipation characteristics. Therefore, by constructing a temporal persistence score, stable patterns in historical time series can be automatically discovered. Compared to manual annotation methods, this scheme automatically generates pseudo-labels using continuous historical time series, which can significantly reduce annotation costs while improving adaptability to complex weather scenarios.

[0082] The spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features of historical echo objects are extracted to construct a feature vector for historical objects. The spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features of the current echo object are extracted, including: calculating the area parameters, perimeter parameters, morphological compactness parameters, and boundary fragmentation parameters of the current echo object to generate spatial structure features; calculating the maximum reflectivity, average reflectivity, reflectivity gradient, and local variance of the current echo object to generate reflectivity statistical features; calculating the number of connected regions, neighborhood connectivity strength, and local clustering density of the current echo object to generate spatial connectivity features; obtaining the centroid coordinates, radar radial distance, radial azimuth angle, and principal axis direction angle of the current echo object to generate radar geometric features; and performing normalization and fusion processing on the spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features to generate the feature vector for the current object.

[0083] In this embodiment, spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features are extracted for the current echo object. Spatial structure features include area parameters, perimeter parameters, morphological compactness parameters, and boundary fragmentation parameters; reflectivity statistical features include maximum reflectivity, average reflectivity, reflectivity gradient, and local variance; spatial connectivity features include the number of connected regions, neighborhood connectivity strength, and local clustering density; radar geometric features include centroid coordinates, radar radial distance, radial azimuth angle, and principal axis direction angle. These features are then normalized and fused to generate a unified object feature vector.

[0084] In this embodiment, due to the dimensional differences among spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features, a unified scale mapping process is first performed on the different features. Based on the statistical distribution relationship of each feature in historical samples, a feature normalization mapping function is constructed, thereby obtaining the normalized feature expression:

[0085] ;

[0086] Furthermore, to enhance the joint expressive power among different types of features, an object feature fusion equation is constructed based on the feature weight mapping relationship:

[0087] ;

[0088] in, Represents the characteristics of the i-th class of normalized objects;

[0089] This represents the original object feature of the i-th type. The original object feature is derived from the spatial structure feature, reflectivity statistical feature, spatial connectivity feature, or radar geometric feature corresponding to the echo object at the current time.

[0090] This represents the minimum eigenvalue of the i-th type of object in the historical multi-time combination reflectance data. The minimum eigenvalue is obtained by performing statistical analysis on the historical object feature set.

[0091] This represents the maximum eigenvalue of the i-th type of object in the historical multi-time combination reflectance data. The maximum eigenvalue is obtained by performing statistical analysis on the historical object feature set.

[0092] ε represents the normalization stabilization factor, used to avoid calculation anomalies caused by a denominator of zero. The normalization stabilization factor is a preset minimum parameter.

[0093] This represents the object feature vector, which is used to characterize the multidimensional structural properties of the echo object at the current time.

[0094] This represents a set of spatial structural features, which includes area parameters, perimeter parameters, morphological compactness parameters, and boundary fragmentation parameters.

[0095] This represents the set of reflectance statistical features, which includes maximum reflectance, average reflectance, reflectance gradient, and local variance.

[0096] This represents the set of spatial connectivity features, which includes the number of connected regions, neighborhood connectivity strength, and local clustering density.

[0097] This represents the set of radar geometric features, which includes the centroid coordinates, radar radial range, radial azimuth angle, and principal axis direction angle. , , as well as These represent the weights of spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features, respectively. The weight parameters are preset based on historical operational statistics, model training results, or feature importance analysis results.

[0098] By performing object feature normalization, object features with different dimensions and numerical ranges can be mapped to a unified scale space, thereby improving the training stability of the probabilistic classification model. Furthermore, through object feature fusion equations, a joint expression of spatial structure attributes, reflectivity distribution attributes, spatial connectivity attributes, and radar observation geometric attributes can be achieved, thus enhancing the ability of object feature vectors to distinguish complex false alarm structures. The final generated object feature vectors can be used as input data for the probabilistic classification model to output object retention confidence.

[0099] Among them, the weights of various features are dynamically assigned based on their historical training contributions.

[0100] By fusing multidimensional object features, the spatial structural properties, reflectivity distribution properties, and radar observation geometric properties of echo objects can be simultaneously characterized, thereby improving the model's ability to distinguish complex false alarm structures. For example, real convective echoes typically exhibit high spatial clustering and stable morphology, while radial interference echoes show high consistency between the principal axis direction and the radar radial direction. Therefore, multidimensional feature fusion helps improve the stability of the classification boundary.

[0101] A probabilistic classification model is trained based on historical object feature vectors, high-confidence retained pseudo-labels, and high-confidence false alarm pseudo-labels, generating a trained probabilistic classification model. The training of the probabilistic classification model includes: inputting the current object feature vector into a feature mapping network to generate object mapping features for probabilistic classification; performing classification boundary fitting on the object mapping features using a supervised training sample set; constructing a supervised training sample set based on high-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels; and constructing a supervised training sample set based on high-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels.

[0102] The historical object feature vectors are input into the feature mapping network to generate object mapping features, which are the latent representation vectors of objects output from the intermediate layers of the feature mapping network.

[0103] The confidence level of the output object is preserved through the probability mapping layer of the probability classification model;

[0104] Perform classification boundary fitting on object mapping features using a supervised training sample set;

[0105] Temporal consistency constraints are generated based on the temporal variation relationship between adjacent echo objects, spatial morphology consistency constraints are generated based on the spatial structural characteristics of the echo objects, and radial structure consistency constraints are generated based on the relationship between the principal axis direction of the echo objects and the radial azimuth of the radar. A joint loss function is constructed based on the temporal consistency constraints, spatial morphology consistency constraints and radial structure consistency constraints.

[0106] Based on the joint loss function, the feature mapping network and the object-preserving confidence output layer are iteratively updated to generate a trained probabilistic classification model.

[0107] In this embodiment, the current object feature vector is input into a feature mapping network to generate object mapping features for probabilistic classification. Then, a supervised training sample set is constructed based on high-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels, and the object mapping features are fitted with classification boundaries using this supervised training sample set. Furthermore, a joint loss function is constructed based on temporal consistency constraints, spatial morphological consistency constraints, and radial structural consistency constraints, and the feature mapping network is iteratively updated according to the joint loss function to generate a trained probabilistic classification model. Preferably, the feature mapping network can be implemented using a shallow fully connected network, a tree model, or a lightweight convolutional network.

[0108] In this embodiment, to avoid the model relying solely on a single reflectivity intensity feature for classification while ignoring spatial structural continuity and radial structural regularity, temporal consistency constraints, spatial morphological consistency constraints, and radial structural consistency constraints are jointly modeled. Based on the multi-constraint collaborative mapping relationship, a joint loss function is constructed:

[0109] ;

[0110] in:

[0111] ;

[0112] The rate of change of the object area in adjacent echoes is calculated from the area difference of the object region.

[0113] The rate of change in the principal axis direction is calculated from the angular difference in the principal axis direction of the object.

[0114] The rate of change of average reflectance is obtained from the difference in average reflectance of the object region.

[0115] The spatial overlap ratio of adjacent time-series objects is calculated from the intersection-union ratio of object regions.

[0116] These are preset weight parameters used to adjust the degree of influence of each time-series variation item;

[0117] ;

[0118] The deviation from the ideal structure is calculated as the difference between the compactness index, which is constructed from the object's area and perimeter, and the compactness index.

[0119] The boundary fragmentation is obtained by normalizing the number of discontinuous segments at the object boundary.

[0120] This is the spatial distribution offset, calculated from the variance of the internal reflectivity distribution of the object;

[0121] Preset weight parameters;

[0122] ;

[0123] The angle deviation between the principal axis of the object and the radial azimuth of the radar is calculated from geometric coordinates.

[0124] is the radial sensitivity coefficient, a preset parameter used to control the steepness of the weighting curve;

[0125] The number of objects involved in the calculation;

[0126] ;

[0127] Furthermore, based on the mapping relationship between the object retention confidence output and the pseudo-label, a probabilistic classification error function is constructed:

[0128] ;

[0129] in, This represents the joint loss function value, which is used to characterize the overall training error of the current probabilistic classification model under multidimensional constraints.

[0130] The temporal consistency error is calculated based on the temporal persistence score among historical consecutive echo objects and is used to constrain the model's ability to maintain a continuous and stable echo structure.

[0131] The spatial structure error is represented by the spatial morphological consistency constraint, which corresponds to the spatial structure error term in the computational implementation. The spatial structure error is calculated based on the deviation between the area parameter, morphological compactness parameter and boundary fragmentation parameter in the spatial structure features, and is used to constrain the spatial morphological consistency of the object.

[0132] The radial structural error is calculated based on the structural deviation between the radar radial range, radial azimuth angle, and principal axis direction angle, and is used to suppress radial stripe interference and abnormal propagation echoes.

[0133] This represents the probabilistic classification error, which is calculated based on the classification difference between the object retention confidence output and the high-confidence retained false labels and high-confidence false alarm labels.

[0134] These represent the temporal consistency error weight, spatial structure error weight, radial structure error weight, and probabilistic classification error weight, respectively, all of which are preset weight parameters. The weight parameters can be preset based on historical training error statistics, feature importance analysis results, or incremental feedback training results, and must satisfy normalization constraints.

[0135] N represents the number of object samples in the supervised training sample set, and the object samples are derived from the historical echo object set in the historical continuous multi-time combined reflectivity data.

[0136] This represents the pseudo-label value corresponding to the i-th object sample. The value is 1 when the object sample belongs to the high-confidence retained pseudo-label, and 0 when the object sample belongs to the high-confidence false alarm pseudo-label.

[0137] This represents the object retention confidence output for the i-th object sample. The object retention confidence output is obtained by the probabilistic classification model based on object feature vector inference.

[0138] By calculating the joint loss function, joint constraints can be applied to temporal continuous structure, spatial structure features, and radial structure features, thereby avoiding the probabilistic classification model's over-reliance on a single reflectivity feature. Furthermore, the probabilistic classification error function quantifies the matching degree between the object retention confidence output and the pseudo-label, and is used to guide the feature mapping network and the object retention confidence output layer in performing iterative update processing. The final generated joint loss function value can be used to complete the training of the probabilistic classification model and improve the stability and robustness of single-frame false alarm identification in the online stage.

[0139] By using joint loss constraints, a stable transfer of temporal knowledge to the single-frame probability space is achieved.

[0140] The core of this step lies in compressing and mapping the stability patterns from continuous historical time intervals into the single-frame object feature space, enabling the model to complete false alarm identification during the online phase without acquiring data from future time intervals. By using joint loss function constraints, the model avoids learning only a single reflectivity intensity feature while ignoring spatial and radial structure features, thereby improving robustness in complex scenarios.

[0141] Obtain the combined reflectance data of the current time frame; perform object-level connected component segmentation on the combined reflectance data of the current time frame to generate the echo object set of the current time frame.

[0142] In this embodiment, the online service phase only acquires the combined reflectivity data of the current single frame and repeats the object-level connected component segmentation process to generate the echo object set for the current time. Since the online phase no longer relies on data from future time periods, object generation and subsequent inference processing can be completed directly within the current service time period.

[0143] This step completely decouples the training phase from the online inference phase, enabling the model to complete real-time false alarm identification using only a single frame of data in an online business environment, thereby meeting the real-time business requirements of weather radar.

[0144] Extract the spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features of the current echo object to construct the current object feature vector;

[0145] Perform multi-scale spatial window partitioning on the current echo object to generate local scale region, mesoscale region and global scale region;

[0146] Calculate the echo persistence index, spatial expansion index, and boundary stability index in local scale region, mesoscale region, and global scale region respectively;

[0147] A multi-scale stability score is generated based on echo persistence, spatial expansion, and boundary stability indices corresponding to different scale regions.

[0148] The object feature vector is enhanced based on the multi-scale stability score to generate a multi-scale enhanced feature vector.

[0149] The multi-scale enhanced feature vector is used as the enhanced input of the object feature vector, and the updated object retention confidence is output.

[0150] A pseudo-evolutionary sliding window is constructed based on the combined reflectivity data of the current single frame; spatial diffusion simulation processing is performed on the decay state data, original state data and diffusion state data in the pseudo-evolutionary sliding window to generate pseudo-temporal evolution data;

[0151] The persistence probability of an object is calculated based on the continuity of object spatial expansion and the continuity of reflectivity change in pseudo-temporal evolution data.

[0152] Perform background echo statistical analysis based on the current time-series single-frame combined reflectivity data to generate local clutter density;

[0153] Based on the spatial expansion rate of echo objects in the pseudo-evolutionary sliding window, the regional echo diffusion rate is generated.

[0154] The object retention threshold is dynamically adjusted based on local clutter density, regional echo spread rate, and object persistence probability.

[0155] The adjusted object retention threshold is used to perform probabilistic discrimination processing on the current echo object;

[0156] The radial angle feature is calculated based on the principal axis direction angle and radar radial azimuth angle of the current echo object; the radial consistency index is calculated based on the radial angle feature; the radial conflict score is calculated based on the radial consistency index, neighborhood connectivity strength, and boundary continuity; the echo objects with radial conflict scores exceeding a preset threshold are subjected to weight reduction processing to generate object suppression weights; the object suppression weights are used to perform weighted correction processing on the retention confidence of the corresponding objects to generate the corrected object retention confidence.

[0157] In this embodiment, multi-scale spatial window partitioning is performed on the echo object of the current time interval to generate local-scale, meso-scale, and global-scale regions. Echo persistence, spatial spread, and boundary stability indices are calculated at each scale to generate a multi-scale stability score. Then, the object feature vector is enhanced based on the multi-scale stability score. Simultaneously, a pseudo-evolutionary sliding window is constructed based on the combined reflectivity data of the current single frame. Spatial diffusion simulation is performed on the attenuation state data, original state data, and diffusion state data to generate pseudo-temporal evolution data. The object persistence probability is calculated based on the object's spatial spread continuity and reflectivity change continuity. Furthermore, the object retention threshold is dynamically adjusted based on local clutter density, regional echo spread rate, and object persistence probability. In addition, radial angle features can be calculated based on the principal axis direction angle and radar radial azimuth angle. Radial conflict scores are calculated based on radial consistency indices, neighborhood connectivity strength, and boundary continuity, and weighting is applied to objects with abnormal radial structures.

[0158] In this embodiment, due to the significant differences in local clutter density and object diffusion characteristics across different weather scenarios, a fixed retention threshold is difficult to adapt to complex operational scenarios. Based on the dynamic mapping relationship between local clutter density, regional echo spread rate, and object persistence probability, a dynamic threshold adjustment function is constructed, thereby obtaining the object retention threshold equation:

[0159]

[0160] Furthermore, to enhance the suppression capability of anomalous radial structures, a radial weighting function is constructed based on the radial angle deviation relationship:

[0161]

[0162] Finally, based on the dynamic retention threshold and the radial weighting results, a joint correction process is performed on the object retention confidence.

[0163] in, A threshold is reserved for dynamically adjusted objects, which is used to perform probability discrimination processing on the current echo objects;

[0164] The threshold is reserved for basic objects, and is a preset basic threshold parameter for the system initialization phase;

[0165] , as well as These are the local clutter density adjustment coefficient, the regional echo spread rate adjustment coefficient, and the object persistence probability adjustment coefficient, all of which are preset weighting parameters.

[0166] Local clutter density is used to characterize the spatial distribution density of low-stability echo objects within the neighborhood of the current echo object. Its data comes from the neighborhood echo statistics in the current time frame combined reflectivity data.

[0167] The regional echo spread rate is used to characterize the degree of spatial expansion of the current echo object in the pseudo-time series evolution data. It is obtained by performing spatial diffusion simulation processing on the diffusion state data in the pseudo-evolution sliding window.

[0168] The object persistence probability characterizes the stability of the current echo object's continuous existence during the pseudo-temporal evolution process. It is calculated based on the continuity of object spatial expansion and the continuity of reflectivity changes. By calculating the dynamically adjusted object retention threshold, adaptive changes in object retention conditions under complex weather conditions can be achieved, thereby improving the ability to retain weak true echoes and suppress false alarms.

[0169] To enhance the identification and suppression capabilities of anomalous radial structures and avoid the impact of radial stripe interference, sidelobe echoes, and anomalous propagation echoes on the probabilistic classification results, a radial weighting function is further constructed. By establishing a nonlinear mapping relationship between radial angle deviation and object suppression weights, anomalous structures whose principal axis is highly aligned with the radar radial direction receive a higher weighting magnitude, thereby improving the model's stability in identifying anomalous radial structures.

[0170] The output value of the radial weighting function is used to generate object suppression weights and perform weighted correction processing on the object retention confidence.

[0171] The radial angle deviation is used to characterize the angle difference between the principal axis direction angle of the current echo object and the radar radial azimuth angle. Its data comes from the calculation results of the principal axis direction angle of the current echo object and the radar radial azimuth angle.

[0172] The radial deviation smoothing coefficient is used to adjust the sensitivity of the radial angle deviation to the weighting amplitude, and is a preset parameter. By calculating the radial weighting function, the radial structural anomaly degree corresponding to the current echo object can be obtained, and the object retention confidence can be further corrected, thereby reducing the impact of radial interference structure on quality control results and improving the spatial structural stability of the combined reflectivity product.

[0173] The multi-scale enhancement mechanism can improve the model's ability to identify weak convection boundary regions; the pseudo-evolutionary sliding window mechanism can simulate the evolution trend of objects without future time data; and the radial consistency correction mechanism can effectively suppress radial stripe interference, sidelobe echoes, and anomalous propagation echoes, thereby improving the overall recognition stability.

[0174] Input the feature vector of the current object into the trained probability classification model, and output the object while retaining its confidence score.

[0175] In this embodiment, the current object feature vector is input into the trained probability classification model, and the object retention confidence score is output. The object retention confidence score is used to characterize the probability level that the current echo object belongs to a real meteorological echo.

[0176] Compared to traditional hard threshold classification, probabilistic output can preserve the uncertainty information of objects, making subsequent quality control processing more flexible and improving recognition stability in complex boundary scenarios.

[0177] Based on the object retention confidence level, the corresponding echo object is subjected to retention processing, attenuation processing, or deletion processing, and the processing results are backfilled into the current time-series combined reflectivity grid to generate a quality-controlled combined reflectivity product. The process of performing retention processing, attenuation processing, or deletion processing on the corresponding echo object based on the object retention confidence level includes: comparing the object retention confidence level with a first retention threshold and a second retention threshold; performing original reflectivity retention processing on the corresponding echo object when the object retention confidence level is greater than or equal to the first retention threshold; performing reflectivity attenuation processing on the corresponding echo object when the object retention confidence level is less than the first retention threshold but greater than or equal to the second retention threshold; and performing deletion processing on the corresponding echo object when the object retention confidence level is less than the second retention threshold, wherein deletion processing includes replacing the reflectivity value of the corresponding echo object with a background value; and remapping the object region after performing retention processing, attenuation processing, or background value replacement processing back into the current time-series combined reflectivity grid to generate a quality-controlled combined reflectivity product.

[0178] In this embodiment, the object retention confidence level is compared with a first retention threshold and a second retention threshold. When the object retention confidence level is higher than the first retention threshold, the object is subjected to original reflectance retention processing. When the object retention confidence level is between the first retention threshold and the second retention threshold, the object is subjected to reflectance attenuation processing. When the object retention confidence level is lower than the second retention threshold, the object is subjected to background value replacement processing. Then, the processing results are remapped to the combined reflectance grid to generate a quality-controlled combined reflectance product.

[0179] By using a hierarchical probability control method, the problem of false deletion of real weak echoes caused by the traditional "one-size-fits-all" deletion method can be avoided, while taking into account both the false alarm suppression effect and the ability to retain real echoes.

[0180] Obtain the combined reflectance product after historical quality control and the corresponding real-time business output results; perform difference statistical processing on the combined reflectance product after historical quality control and the corresponding real-time business output results to generate error feedback data; calculate the object misclassification rate, regional residual rate, and structural deviation rate based on the error feedback data; update the weight parameters in the joint loss function based on the object misclassification rate, regional residual rate, and structural deviation rate; perform incremental iterative training on the probabilistic classification model based on the updated joint loss function to generate the updated probabilistic classification model.

[0181] In this embodiment, the combined reflectance products after historical quality control and the corresponding real-time business output results are obtained, and difference statistical processing is performed on the two to generate error feedback data. Then, based on the error feedback data, the object misclassification rate, regional residual rate, and structural deviation rate are calculated, and the weight parameters in the joint loss function are updated according to the above indicators. Finally, based on the updated joint loss function, incremental iterative training processing is performed on the probabilistic classification model to realize online model updating and continuous optimization.

[0182] Through the incremental feedback training mechanism, the model can continuously adapt to changes in false alarm characteristics under different seasons, terrains, and radar operating conditions, thereby improving long-term operational stability and scenario adaptability.

[0183] Example 2: Specific Application Scenarios

[0184] Taking a coastal severe convective weather monitoring system as an example, multiple S-band weather radar stations upload combined reflectance data to the regional radar service platform at uniform time intervals. The system first acquires historical combined reflectance data from multiple consecutive time periods and generates a historical echo object set through object-level connected region segmentation. Then, based on the spatial overlap, morphological continuity, and reflectance persistence of objects in consecutive time periods, a temporal persistence score is constructed, and high-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels are automatically generated for training the probabilistic classification model. During the real-time service phase, the system only acquires single-frame combined reflectance data for the current time period and generates an echo object set for that time period. Then, it extracts the spatial structure features, reflectance statistical features, spatial connectivity features, and radar geometric features of the objects and inputs them into the trained probabilistic classification model to output object retention confidence. For objects with radially consistent height and abnormal boundary continuity, the system further performs radial conflict scoring processing and reduces the weight of the corresponding objects. Finally, based on the object retention confidence, background value replacement processing is performed on false alarm objects, reflectance attenuation processing is performed on weakly uncertain objects, and a quality-controlled combined reflectance product is output. Compared with traditional methods that rely on tracking trajectories at future times, this solution can complete online identification without waiting for future time data. Therefore, it can meet the requirements of real-time weather services for low-latency processing, and at the same time, it can effectively reduce the interference of short-term isolated false alarm echoes on the monitoring results of severe convection.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation methods of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A networked weather radar combined reflectivity false alarm discrimination and quality control method, characterized by, include: Acquire historical, continuous, multi-timetime combined reflectivity data from the networked weather radar; Perform object-level connected region segmentation on historical continuous multi-time combined reflectivity data to generate a set of historical echo objects; A temporal persistence score is constructed based on the spatial overlap, morphological continuity, and reflectivity persistence relationships between historical echo objects in adjacent time intervals. High-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels are generated based on the temporal persistence score. Extract the spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features of historical echo objects to construct feature vectors for historical objects; A probabilistic classification model is trained based on the feature vectors of historical objects, high-confidence retained pseudo-labels, and high-confidence false alarm pseudo-labels, generating a trained probabilistic classification model. Obtain the combined reflectance data of the current single frame; Perform object-level connected component segmentation on the current time frame combined reflectivity data to generate the current time echo object set; Extract the spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features of the current echo object to construct the current object feature vector; Input the current object feature vector into the trained probability classification model and output the object retention confidence. The object retention confidence is the final object retention confidence after radial correction and multi-scale enhancement. Based on the object retention confidence level, perform retention processing, attenuation processing or deletion processing on the corresponding echo object, and backfill the processing results into the current time-combined reflectivity grid to generate a quality-controlled combined reflectivity product.

2. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, The calculation of time series persistence score includes: Perform centroid position matching processing on adjacent time-series echo objects to generate a set of candidate associated objects; The rate of change of object area, the rate of change of principal axis direction, the rate of change of average reflectance, and the boundary overlap ratio are calculated based on the candidate associated object set. The temporal stability index of the object is generated based on the rate of change of area, the rate of change of principal axis direction, the rate of change of average reflectivity, and the boundary overlap ratio. Calculate the time series persistence score based on the object's time series stability index; Based on the time-series persistence score, persistent clustering is performed on historical continuous multi-time echo objects to generate a stable object set and an unstable object set; Sets of stable objects are marked as high-confidence retained pseudo-labels, while sets of unstable objects are marked as high-confidence false alarm pseudo-labels.

3. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, Extract the spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features of the current echo object, including: Calculate the area, perimeter, morphological compactness, and boundary fragmentation parameters of the current echo object to generate spatial structural features; Statistical features of reflectivity are generated by calculating the maximum reflectivity, average reflectivity, reflectivity gradient, and local variance of the echo object at the current time. Calculate the number of connected regions, neighborhood connectivity strength, and local clustering density of the current echo object to generate spatial connectivity features; Obtain the centroid coordinates, radar radial distance, radial azimuth angle, and principal axis direction angle of the current echo object, and generate radar geometric features; Normalization and fusion processing is performed on spatial structure features, reflectivity statistical features, spatial connectivity features, and radar geometric features to generate the feature vector of the current object.

4. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, Training a probabilistic classification model includes: A supervised training sample set is constructed based on high-confidence retained pseudo-labels and high-confidence false alarm pseudo-labels; The historical object feature vectors are input into the feature mapping network to generate object mapping features, which are the latent representation vectors of objects output from the intermediate layers of the feature mapping network. The confidence level of the output object is preserved through the probability mapping layer of the probability classification model; Perform classification boundary fitting on object mapping features using a supervised training sample set; Temporal consistency constraints are generated based on the temporal variation relationship between adjacent echo objects, spatial morphology consistency constraints are generated based on the spatial structural characteristics of the echo objects, and radial structure consistency constraints are generated based on the relationship between the principal axis direction of the echo objects and the radial azimuth of the radar. A joint loss function is constructed based on the temporal consistency constraints, spatial morphology consistency constraints and radial structure consistency constraints. Based on the joint loss function, the feature mapping network and the object-preserving confidence output layer are iteratively updated to generate a trained probabilistic classification model.

5. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, Perform retention, attenuation, or deletion processing on the corresponding echo object, including: Compare the object retention confidence level with the first retention threshold and the second retention threshold; When the object retention confidence is greater than or equal to the first retention threshold, the original reflectivity retention process is performed on the corresponding echo object; When the object retention confidence is less than the first retention threshold and greater than or equal to the second retention threshold, reflectivity attenuation processing is performed on the corresponding echo object; When the object retention confidence is less than the second retention threshold, the corresponding echo object is deleted. The deletion process includes replacing the reflectivity value of the corresponding echo object with the background value. The object region after performing retention processing, attenuation processing, or background value replacement processing is remapped into the current time-combined reflectance grid to generate a quality-controlled combined reflectance product.

6. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, The method also includes: Perform multi-scale spatial window partitioning on the current echo object to generate local scale region, mesoscale region and global scale region; Calculate the echo persistence index, spatial expansion index, and boundary stability index in local scale region, mesoscale region, and global scale region respectively; A multi-scale stability score is generated based on echo persistence, spatial expansion, and boundary stability indices corresponding to different scale regions. The object feature vector is enhanced based on the multi-scale stability score to generate a multi-scale enhanced feature vector. The multi-scale enhanced feature vector is used as the enhanced input of the object feature vector, and the updated object retention confidence is output to generate the final object retention confidence.

7. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, The method also includes: A pseudo-evolutionary sliding window is constructed based on the combined reflectance data of the current single frame. Spatial diffusion simulation processing is performed on the decay state data, original state data, and diffusion state data in the pseudo-evolutionary sliding window to generate pseudo-temporal evolution data; The persistence probability of an object is calculated based on the continuity of object spatial expansion and the continuity of reflectivity change in pseudo-temporal evolution data. Perform background echo statistical analysis based on the current time-series single-frame combined reflectivity data to generate local clutter density; Based on the spatial expansion rate of echo objects in the pseudo-evolutionary sliding window, the regional echo diffusion rate is generated. The object retention threshold is dynamically adjusted based on local clutter density, regional echo spread rate, and object persistence probability. The adjusted object retention threshold is used to perform probabilistic discrimination processing on the current echo object.

8. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, The method also includes: Calculate the radial angle characteristics based on the principal axis direction angle and radar radial azimuth angle of the current echo object; Calculate the radial consistency index based on the radial angle characteristics; Radial conflict score is calculated based on radial consistency index, neighborhood connectivity strength, and boundary continuity. For echo objects whose radial conflict scores exceed a preset threshold, a weight reduction process is performed to generate object suppression weights; The object retention confidence is weighted and corrected using object suppression weights to generate a corrected object retention confidence, which is then used as a component input of the final object retention confidence.

9. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, Object-level connected component segmentation includes: Perform a preset reflectivity threshold filtering process on the current single-frame combined reflectivity data to generate candidate echo regions; Perform neighborhood connectivity analysis on the candidate echo regions to generate initial connectivity objects; Perform small region culling and hole filling on the initial connected objects to generate standardized echo objects; Perform boundary smoothing and morphology uniformity processing on standardized echo objects to generate a set of object-level echo candidate regions; The set of object-level echo candidate regions is used as input data for subsequent object feature extraction processing.

10. A method for identifying and controlling false alarms based on the reflectivity of a networked weather radar system according to claim 1, characterized in that, The method also includes: Obtain the combined reflectance products after historical quality control and the corresponding real-time business output results; Perform statistical processing on the difference between the combined reflectance products after historical quality control and the corresponding real-time business output results to generate error feedback data; The error feedback data is used to calculate the object misjudgment rate, regional residual rate, and structural deviation rate. The weight parameters in the joint loss function are updated based on the object misclassification rate, region residual rate, and structural deviation rate. Incremental iterative training is performed on the probabilistic classification model based on the updated joint loss function to generate an updated probabilistic classification model.

Images