A multi-scale phase space-time encoding and multi-source fusion hail probability prediction method constrained by missing inversion

Abstract

The application discloses a multi-source fusion hail probability prediction method based on multi-scale phase space-time coding and missing data inversion constraints, performs quality control on multi-source data, aligns a unified time axis and maps a space grid, generates a usable state, a quality weight and an alignment feature set, constructs an event level and a grid point level hail label, controls sample training contribution by using label confidence, searches similar samples according to space-time embedding, performs weighted inversion completion on missing elements, constructs a gate constraint according to a quality weight and completion confidence, inputs environmental field features, radar structure features and space-time embedding into a double-branch fusion model, and outputs a hail occurrence probability field, a risk classification product and corresponding trace information.

Description

Technical Field

[0001] This invention belongs to the field of meteorological disaster weather monitoring and forecasting technology, specifically involving a hail probability forecasting method. Addressing the needs of multi-source meteorological data access and processing in operational needs, it constructs phase-time codes and geographic location codes encompassing seasonal and intra-diurnal scales through unified spatiotemporal alignment and observability representation. Under conditions of missing data, it performs inversion and completion based on spatiotemporal similarity and introduces confidence constraints. It integrates radar echo structure information and environmental field element information for probabilistic inference, outputting a calibrated hail occurrence probability field and risk classification product for short-term forecasting and imminent warning operational applications. Background Technology

[0002] Hail is an important type of severe convective weather, characterized by its small scale of occurrence, rapid development and evolution, and localized impact. Hail forecasting requires not only clearly defined impact areas and time windows, but also reliable probability information and risk level conclusions to support early warning issuance, dispatching, and disaster prevention and mitigation decision-making. Because hail events are often accompanied by the rapid development of mesoscale convective systems, environmental field conditions and echo structures can change significantly in a short period of time. Therefore, forecasting systems need to simultaneously characterize the triggering environment, convective evolution, and spatial distribution features to produce usable products within a limited timeframe.

[0003] Existing hail forecasting technologies mainly include: discrimination methods based on radar feature thresholds, classification and regression methods based on statistical learning, and data-driven prediction methods represented by deep learning. Threshold discrimination methods rely on empirical rules and have strong interpretability, but are sensitive to regional and seasonal differences, and the thresholds show insufficient stability when migrating across regions. Statistical learning methods have certain effects when the sample size and feature quality are good, but their ability to characterize nonlinear structures is limited, and they cannot simultaneously take into account multi-scale information of the environmental field and echo structure. Data-driven methods have advantages in fusing multi-source information and learning complex relationships, but their performance is highly dependent on data alignment quality, label reliability, and the stability of the input distribution on the operational side.

[0004] In application scenarios, multi-source data exhibits significant heterogeneity. Different data sources have inconsistent time update frequencies, observation delays, spatial projections, and resolutions, and complex missing data patterns. Without a unified spatiotemporal alignment process and observability representation mechanism, the input distribution during training and runtime can easily shift, leading to unstable model output. Multi-source data also suffers from quality differences and local occlusion issues. Directly inputting unusable information into the fusion process can easily cause false alarms or missed detections in critical areas, impacting product reliability and business decisions.

[0005] The total number of hail samples is relatively limited, and their spatiotemporal distribution is uneven. Label sources may contain incomplete records and spatial positioning errors. Without label confidence control and noise sample management, the training process is easily influenced by erroneous labels, causing the model to overfit to a few strong samples and exhibit insufficient generalization ability across different regions and weather conditions. For probabilistic forecast products, label noise also affects probability reliability, making the same probability value incomparable across different regions and seasons, reducing the stability of risk classification thresholds, and making hail label construction a significant factor limiting the effectiveness of the technology.

[0006] Severe convective processes exhibit significant seasonality and diurnal variation. The probability of hail occurrence shows regular differences across different periods of the year and within different times of day. Different regions also exhibit similarities and differences in topographic conditions, underlying surface, and climatic background. These spatiotemporal patterns significantly constrain forecasting. Existing technologies mostly use static temporal features or simple geographical features, making it difficult to structurally characterize periodic and spatial similarities. Furthermore, it is challenging to utilize historical similarity processes to invert and complete key elements and constrain confidence levels when data is missing. Under conditions of incomplete or degraded input, model output is prone to significant fluctuations, failing to meet the requirements of continuous operational use.

[0007] Hailstorm probability products typically need to be directly integrated with risk grading on the business side. Without an effective probability calibration and grading generation mechanism, the probability output may be overly confident or overly conservative, leading to inconsistent risk levels across different regions or seasons. During operation, factors such as data access latency, computing resource constraints, and product update frequency must also be considered. Without a clear modular process and deployable implementation path, it is difficult to achieve stable business delivery capabilities.

[0008] In summary, existing technologies have shortcomings in terms of spatiotemporal consistency of multi-source data, construction of stable inputs under missing data conditions, label confidence management, structured expression of spatiotemporal patterns, and availability of probability calibration and risk classification. Summary of the Invention

[0009] The technical problems to be solved by this invention include: Hail is a severe convective weather disaster characterized by strong locality, high suddenness, and rapid evolution. Operational forecasts not only need to provide the spatial location and time window of hail occurrence, but also need to provide comparable and actionable probability results and risk levels.

[0010] Existing hail forecasting methods, when used in operational settings, generally suffer from problems such as difficulty in unifying and coordinating multi-source data, unstable output under conditions of missing data, insufficient utilization of spatiotemporal patterns, and weak consistency of probability products. These problems make it difficult to meet the requirements of continuous operation and operational applications, and can be manifested in four aspects.

[0011] First, multi-source meteorological data exhibits strong heterogeneity, making spatiotemporal alignment difficult. Different data sources show significant differences in update time, transmission delay, spatial projection, resolution, and quality control standards. Without a unified spatiotemporal alignment mechanism and consistent data organization, inconsistencies in input distribution between the training and operational phases can easily occur, leading to large fluctuations and insufficient stability in forecast results.

[0012] Second, data gaps and quality degradation are widespread, and there is a lack of stable and controllable mechanisms for handling data gaps. Operational issues frequently arise, including radar obstruction, observation blind spots, missing station reports, and model field delays. Simply discarding missing data elements results in incomplete input information. Using unconstrained data completion, on the other hand, can easily introduce low-reliability information into the inference process, leading to false alarms or missed reports in key areas, thus affecting the reliability of forecast results.

[0013] Third, hail samples are scarce, label quality is uneven, and the use of noisy samples and spatiotemporal patterns is insufficient, resulting in problems such as limited total quantity, uneven spatiotemporal distribution, insufficient record completeness, and inadequate positioning accuracy. Without label confidence control and sample governance mechanisms, the model is easily affected by noisy labels, reducing its generalization ability under different regions and weather conditions. Hail events exhibit significant seasonality, diurnal variation characteristics, and regional differences, and the expression of periodic and spatial similarities is relatively coarse, making it difficult to effectively improve the continuity and stability of forecasts.

[0014] Fourth, there is a lack of a unified and stable linkage mechanism between probability forecast results and risk classification. After the probability is output, there is a lack of effective calibration and classification processing, which can easily lead to inconsistent meanings of the same probability value in different regions and seasons. This results in inconsistent risk level definitions, affecting business release, early warning decision-making and practical application effectiveness.

[0015] To address the problems existing in the prior art, this invention proposes a multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints. It constructs a unified spatiotemporal alignment and observability expression mechanism, introduces multi-scale phase spatiotemporal coding, label confidence control, and missing measurement inversion completion constraints, and fuses multi-source environmental field information and radar structure information. The output probability is calibrated and risk-classified, forming a complete technical chain from data alignment, sample and label construction, fusion inference to calibration and graded output. This improves the stability, reliability, and availability of hail probability forecasts under complex operational conditions.

[0016] The methods include unified spatiotemporal alignment, observability constraints, multi-scale phase spatiotemporal coding, missing measurement inversion and completion, multi-source fusion inference, probability calibration and risk classification, to achieve multi-source fusion hail probability forecasting and risk output for operational purposes.

[0017] Step 1: Access and unified organization of multi-source meteorological data.

[0018] The system acquires multi-source meteorological data for hail forecasting, including radar echo data, satellite remote sensing data, ground meteorological observation data, and numerical weather prediction data. It uses a unified data interface for receiving, storing, and managing the data, and records the source, timestamp, spatial range, resolution, update frequency, and quality status of each type of data, providing a unified input basis for subsequent alignment processing, feature construction, and inference calculations.

[0019] Step two: unify spatiotemporal alignment and observability representation.

[0020] A unified time axis and a unified spatial grid are established for multi-source meteorological data. The observation time, transmission delay, spatial projection, and resolution differences of various types of data are corrected, mapped, and resampled to ensure that various types of data are consistent with the same operational time and the same spatial reference. For the missing data, occlusion, boundary crossing, and quality degradation of various types of data in different regions and at different times, observability representation information is constructed, including availability status and quality weight, which serves as the basis for subsequent missing data identification, completion triggering, and fusion weighting.

[0021] Step 3: Sample construction and label confidence control.

[0022] Training and prediction samples are constructed using a unified spatiotemporal benchmark. An event-level and grid-level consistent hail labeling system is established to unify the training and business output standards. A label confidence control mechanism is introduced to convert the label reliability into sample weights or loss constraints, thereby reducing the impact of noisy samples on model training.

[0023] Step 3 addresses the issues of scarce hail samples, uneven spatiotemporal distribution, incomplete label sources, and positioning bias, thereby improving the model's stability and generalization ability in different regions and under different weather conditions.

[0024] Step 4: Construction of multi-scale phase spatiotemporal coding.

[0025] Multi-scale phase spatiotemporal coding is constructed for the seasonal, intra-day, and regional differences of hail, including intra-annual cycle coding, intra-day cycle coding, and geographic location coding. Monthly cycle coding is extended according to business needs. The time coding expresses the similarity of seasonal stages and the daily variation pattern, and the location coding expresses the spatial climate background and topographic constraints. The spatiotemporal coding serves as the structured input of the fusion inference model and the basis for measuring the spatiotemporal similarity of the missing data inversion process.

[0026] Step 5: Missing test trigger determination and similarity inversion completion.

[0027] If a data source is unavailable at a certain time or in a certain region, or if its quality weight is lower than a preset threshold, a missing data inversion and completion is triggered. Based on the spatiotemporal similarity measure, time periods and regions similar to the current process are retrieved from the historical sample library or the neighborhood sample set. The missing data elements are then weighted and inverted to complete the missing data, and the corresponding completion confidence is output to characterize the reliability of the completion result. This serves as constraint information for subsequent fusion inference, preventing low-confidence completion results from affecting the final forecast.

[0028] Step six: dual-branch fusion reasoning of environmental field and radar structure information.

[0029] After completing input alignment, quality constraints, and missing measurement completion, a multi-source fusion reasoning framework combining environmental field branch and radar structure branch is constructed. The environmental field branch extracts background feature elements related to hail triggering and maintenance, while the radar structure branch extracts real-time features related to convective cell development, echo intensity structure, and evolution trend. The two branches perform joint reasoning at the fusion layer, combining observable information, quality weights, and completion confidence to constrain the contribution of each source information and output the grid-level hail occurrence probability.

[0030] Step 7: Probability calibration processing.

[0031] Temperature scaling calibration and reliability correction are used to perform consistency calibration of hail occurrence probability, providing a basis for unified threshold release and risk assessment.

[0032] Step 7 enables the same probability value to have a more stable and clear practical meaning in business applications, and improves the comparability and reliability of probability values ​​in different regions, seasons and weather backgrounds.

[0033] Step 8: Risk classification generation and business product output.

[0034] Based on the calibrated hail occurrence probability, the system generates corresponding risk level results according to a preset threshold system, outputting a hail occurrence probability field, risk classification map, and early warning information for key areas. It retains the correspondence between the input data status, alignment results, missing data completion status, quality weights, and output results for subsequent result traceability, error analysis, model evaluation, and continuous optimization.

[0035] Compared with the prior art, this invention proposes four innovations and improvements.

[0036] First, a unified spatiotemporal alignment, observability representation, and label confidence control mechanism for business operations was proposed, which solved the problems of inconsistent input standards for multi-source data and unstable training supervision.

[0037] Existing hail forecasting methods commonly suffer from problems such as varying update frequencies of multi-source data, different observation delays, inconsistent spatial projections and resolutions, and inconsistent quality control standards. These issues can easily lead to inconsistencies in input distribution between the training and runtime phases, causing fluctuations in model output. Furthermore, the limited total number of hail samples, uneven spatiotemporal distribution, incomplete label records, and spatial positioning biases, coupled with a lack of quantitative constraints on label reliability, make the model susceptible to noise-driven data, affecting the stability and generalization ability of probabilistic forecasts.

[0038] To address the aforementioned issues, this invention first establishes a unified timeline and a unified spatial grid to align multi-source data from radar, satellite, ground observations, and numerical weather predictions, constructing observability representation information and incorporating the availability status, quality weights, and anomalies of each data source into a unified input chain. Secondly, it constructs a labeling system with consistent standards at both the event and grid level, introducing a label confidence control mechanism to convert label reliability into sample weights or loss weights, suppressing the adverse effects of noisy and uncertain samples on the training process. Through these designs, this invention unifies data alignment, quality representation, and label governance within a single technical framework, ensuring consistency between training and operation, and providing a stable foundation for subsequent missing data completion, fusion inference, and probabilistic output.

[0039] Compared with existing technologies, the improvements of this invention are: it focuses on the access and fusion of multi-source data, and incorporates observability representation and label confidence control as core constraints, making it more suitable for continuous applications in complex business operation scenarios.

[0040] Second, a multi-scale phase spatiotemporal coding mechanism was proposed, which enhanced the ability to structurally express the seasonality, diurnal variation, and regional climate background patterns of hail.

[0041] Hailstorms exhibit significant seasonal variations and intraday evolution patterns, and different regions are simultaneously influenced by differences in topography, underlying surface environment, and climatic background. Existing technologies, which only use static time fields or simple latitude and longitude information as additional features, struggle to reliably represent the statistical relationships between the same seasonal stage, similar intraday times, and similar spatial backgrounds. This is particularly problematic when applied across regions and seasons, leading to discontinuous representations, insufficient utilization of patterns, and unstable output.

[0042] To address the aforementioned issues, this invention constructs a multi-scale phase spatiotemporal coding mechanism, uniformly mapping intra-annual, intra-day, and optional intra-monthly cycles to a phase space. This, along with latitude-longitude cycle expressions, latitudinal gradient expressions, and terrain components such as altitude, forms a unified spatiotemporal embedding vector. This spatiotemporal embedding vector serves two purposes: firstly, as a structured input to the fusion inference model, constraining the model's learning of environmental field features and radar structural features by temporal and spatial background patterns; and secondly, as a basis for similarity measurement during the missing data inversion and completion process, used to retrieve historically similar time periods and regions, thus unifying spatiotemporal constraints between the training and operational phases.

[0043] Compared with the prior art, the improvement of this invention is that time and space are not only used as ordinary index variables, but are also transformed into a unified constraint representation that participates in reasoning and completion, so that the model can more stably utilize the similarity patterns across years, days, and regions, and improve the consistency and transferability of hail probability forecasts.

[0044] Third, a missing measurement inversion and completion mechanism based on spatiotemporal similarity and confidence refeeding mechanism is proposed to solve the problems of input mutation and output instability under missing measurement conditions.

[0045] In actual operational use, issues such as radar obstruction, observation blind spots, unstable satellite inversion, missing site data, and delayed model field updates are difficult to avoid. Existing technologies address missing data by either directly discarding missing features or using simple interpolation to fill in the gaps. The former creates structural gaps in the model input, leading to significant jumps in the output results. The latter easily introduces low-reliability information directly into the inference link, causing false alarms to spread or missed detections in critical areas.

[0046] To address the aforementioned issues, this invention employs a missing data inversion and completion mechanism based on spatiotemporal similarity, along with a confidence level reinjection. When a data source becomes unavailable at a certain time or in a certain region, or its quality weight falls below a preset threshold, the missing data completion process is automatically triggered. A similarity metric is constructed using a unified spatiotemporal embedding representation. Similar samples with similar temporal phases, spatial backgrounds, and high quality are retrieved from a historical sample library or neighborhood sample set, and weighted inversion and completion are performed on the missing data elements. A corresponding completion confidence level is output for each completion result, quantifying the reliability of the completion result. Instead of directly treating the completion result as equivalent to the measured information, the completion confidence level is reinjected into the subsequent fusion and inference process. Gating or weight adjustment methods are used to control the contribution of the completion features to the output probability, ensuring that low-confidence completion results only serve as weak reference signals in the inference process.

[0047] Compared with the prior art, the improvement of this invention is that it upgrades the missing test handling from simple compensation to a complete closed-loop mechanism with similarity inversion completion, credibility constraints, and controlled contribution. This maintains the integrity of the input structure, controls the additional risks caused by missing tests, and significantly improves the output stability under complex operating conditions.

[0048] Fourth, a dual-branch fusion reasoning mechanism integrating environmental field and radar structure, and an integrated output mechanism for probability calibration and risk classification were proposed, which improved the operational availability of probability products.

[0049] Hailfall occurrence is constrained by environmental background conditions and is closely related to the echo structure and evolution of convective systems. Existing technologies, relying solely on environmental field elements, easily overlook the immediate signals brought about by rapid changes in convective structure, making it difficult to respond promptly to the rapid development of strong convection. Simply relying on radar structural characteristics easily ignores environmental triggering and sustaining conditions, resulting in insufficient model adaptability across different weather patterns and regional backgrounds. Even if the model can output probability results, without subsequent probability calibration and risk classification mechanisms, the same probability value may lack consistent meaning across different regions and seasons, making it difficult to directly serve operational deployments.

[0050] To address the aforementioned issues, this invention proposes a dual-branch fusion reasoning framework combining an environmental field branch and a radar structure branch. The environmental field branch characterizes background potential information such as stratification stability, water vapor conditions, wind shear, and uplift conditions, while the radar structure branch characterizes immediate evidence such as echo intensity structure, vertical development, strong echo volume, and evolution trends. Both types of features participate in probabilistic reasoning in the fusion layer along with spatiotemporal coding. Gating constraints are implemented using quality weights and completion confidence levels to suppress the adverse effects of low-quality inputs on the output. After obtaining the uncalibrated probability, probabilistic calibration processing is introduced to give the same probability value a more consistent and realistic meaning under different regions and seasons. Risk grading results are generated according to a unified threshold system, forming business products such as a hail probability field, risk grading map, and related traceability fields.

[0051] Compared with existing technologies, the improvements of this invention are as follows: it incorporates environmental background information, radar structure information, input credibility constraints, probability calibration, and risk classification output into the same link, thereby improving the model's adaptability to different weather types and rapid evolution stages, and enhancing the comparability, interpretability, and direct publication capability of probability results.

[0052] Compared with the prior art, the present invention has five new technical effects.

[0053] First, it improves the consistency of multi-source data input and the stability of operational processes. This invention establishes a unified timeline, a unified spatial grid, and a unified data organization method, aligning multi-source data from radar, satellite, ground observations, and numerical weather prediction in a consistent manner. It also synchronously introduces a usable state and quality weighting expression mechanism, ensuring consistency in input caliber between the training and operational phases. This effectively reduces input distribution offsets caused by differences in update time, transmission delays, spatial resolution differences, and local quality fluctuations, minimizing result jumps during operational processes and improving the continuity and stability of hail probability forecasts. Existing technologies focus on simple overlay or conventional fusion of multi-source data, lacking unified modeling for spatiotemporal consistency and observability constraints, making long-term stable operation under complex operational conditions difficult.

[0054] Second, it enhances robustness under conditions of missing data and reduces the risk of false alarms and missed alarms. When data is missing, occluded, or degraded, this invention constructs a similarity metric based on multi-scale phase spatiotemporal coding, retrieves historical or neighborhood similar samples, performs inversion and completion on missing elements, and outputs completion confidence. This completion confidence is fed back into the fusion inference stage, combined with quality weights, to construct gating constraints, ensuring that low-confidence completion results do not have an excessive impact on the final output. While maintaining the integrity of the input structure, it avoids forecast instability problems caused by simple deletion or unconstrained completion, improving the system's adaptability to missing data and abnormal disturbances in real-world operational environments. Existing technologies use direct discarding of missing features or conventional interpolation for completion, lacking a complete closed-loop mechanism for completion, confidence assessment, and controlled contribution, making it difficult to balance input integrity and output reliability.

[0055] Third, this invention enhances the ability to express the spatiotemporal evolution patterns of hail and its cross-regional applicability, transforming spatiotemporal patterns into a unified constraint representation that directly participates in inference and completion. This invention constructs intra-annual periodic phase codes, intra-diurnal periodic phase codes, and geographic location codes, incorporating spatial background elements such as topography into a unified spatiotemporal representation. This allows the model to better learn the regular constraints of hail processes at seasonal, diurnal, and regional climatic scales. As structured inputs to the fusion inference model and similarity bases in the inversion and completion process, this improves the model's ability to characterize hail occurrence patterns under different regions, seasons, and weather conditions, enhancing the consistency and transferability of the results. Existing technologies use static time fields or simple latitude and longitude information, making it difficult to fully utilize periodic and spatial similarities.

[0056] Fourth, it improves the comparability, interpretability, and direct business application capabilities of hail probability products, forming a more complete and implementable business product system. This invention employs a dual-branch fusion reasoning framework combining environmental and radar structure branches, simultaneously characterizing real-time evidence of background environmental occurrence and convective echo structure, thus enhancing the ability to identify rapidly developing hail processes. The output probability is calibrated, and risk grading results are generated based on a preset threshold system, ensuring that the same probability value has consistent business meaning under different regions and seasons, forming a unified and stable early warning calibrator. The output hail probability products, risk grading products, and traceability information directly serve short-term forecasts, near-term warnings, and disaster prevention and mitigation decision-making. Existing technologies are limited to single probability output, lacking an integrated business chain from fusion reasoning to probability calibration, risk grading, and result traceability.

[0057] Fifth, this invention improves the engineering deployability and result traceability of the method, balancing forecast performance with engineering application requirements. From multi-source data access, unified alignment, quality control, missing data completion, fusion inference, probabilistic calibration to risk-level output, this invention forms a complete technical chain with clear modules, well-defined steps, and consistent connections. It simultaneously outputs traceability information such as model version, data version, quality statistics, and calibration parameter configuration. This facilitates engineering implementation, version iteration, and effect evaluation, and is beneficial for business departments to conduct result backtracking, error analysis, and continuous optimization, demonstrating strong practical application value. Existing technologies focus on the model's performance itself, lacking systematic consideration for business deployment, operation and maintenance, and result tracking. Attached Figure Description

[0058] Figure 1 It is a flowchart of the method. Detailed Implementation

[0059] The embodiments of this invention are applicable to existing meteorological data processing systems, primarily deployed on high-performance computing platforms or cloud computing environments, supporting large-scale data processing and efficient inference. They rely on powerful computing capabilities, especially parallel processing capabilities, to meet the operational needs of real-time forecasting and nowcasting warnings. Hardware configurations include computing clusters, storage clusters, and network communication infrastructure to ensure rapid response during data access, processing, storage, and inference.

[0060] Historical and real-time observational data are collected from multiple meteorological sources, including radar echo data, satellite remote sensing data, surface meteorological observation data, and numerical weather prediction data. Radar echo data provides information reflecting the structure, evolution, and intensity of meteorological systems, typically volumetric radar data. Satellite remote sensing data provides information on cloud cover, temperature, and humidity in the atmosphere, helping to predict the dynamic changes of weather systems. Surface meteorological observation data provides local temperature, humidity, wind speed, and air pressure. Numerical weather prediction data, through predictions of future meteorological conditions provided by meteorological models, provides background information for hail forecasting.

[0061] All data is transmitted to the data processing platform through a unified interface for quality control, timeline alignment to ensure consistent time granularity, and resampling of each data source at the same spatial resolution, mapping them to a unified grid for subsequent feature extraction, spatiotemporal analysis, and model inference.

[0062] The output products include a hail occurrence probability field and a risk classification. The hail occurrence probability field represents the probability of hail occurring at each grid point, with values ​​ranging from 0 to 1. The closer the value is to 1, the greater the likelihood of hail occurring. The risk classification product, based on the probability value of hail occurrence, classifies risks into different levels according to preset thresholds, such as high risk, medium risk, and low risk, for use in meteorological early warning and disaster prevention and mitigation decision support.

[0063] Based on business needs, the product generation process is updated in real time, supporting subsequent decision analysis, report generation, and visualization. The input data and generated results for each prediction are recorded for subsequent backtracking analysis and model optimization, ensuring product stability and reliability.

[0064] The present invention will be further described in detail below with reference to the accompanying drawings. The overall process from data input to product output is as follows: Figure 1 As shown, the data source is connected to the system through a unified interface. After spatiotemporal alignment, quality control, and feature extraction, multi-source fusion is performed. Following model inference and risk calibration, a hail probability and risk classification product is generated. All steps are performed using a unified time and space benchmark to ensure the quality and reliability of the output product.

[0065] Step 1: Receive real-time observation data, historical data, and numerical weather predictions from multiple meteorological data sources in parallel, including radar echo data, satellite remote sensing data, ground meteorological station data, and numerical weather prediction data.

[0066] Radar echo data includes reflectivity and radial velocity parameters, capturing the dynamic characteristics of weather systems, such as changes in severe convective weather. It is volumetric data and is normalized according to time and space.

[0067] Satellite remote sensing data, which includes atmospheric remote sensing images such as cloud cover, temperature, and humidity, provides large-scale weather system information and is used for analyzing cloud development in hail warnings.

[0068] Ground weather station data consists of real-time measurements from weather stations, including temperature, humidity, wind speed, and air pressure, which are used to supplement detailed information about local convective systems.

[0069] Numerical weather forecast data is obtained from meteorological models and provides background information on the probability of hail. It belongs to large-scale weather forecasting and provides long-term forecast trends.

[0070] All data sources are aligned and mapped according to a unified spatiotemporal reference and processed at the same temporal granularity and spatial resolution.

[0071] Step two involves quality control of the data, including cleaning the raw data, handling missing data, identifying and removing outliers, and quality assessment and weighting to ensure the data is valid and reliable.

[0072] Handling missing data: Meteorological data comes from a wide range of sources and has inconsistent collection frequencies. Data may be missing for some time periods or spatial points. Interpolation and filling methods are used to complete the data. Interpolation is performed based on data from nearby times or adjacent spatial points, or based on time-series linear interpolation, or based on nearest neighbor interpolation within the sample.

[0073] Outlier Identification and Removal: Sensor malfunctions or data transmission errors generate outliers. Statistical outlier detection methods are used to identify these outliers, including preset threshold ranges or pattern-based methods, such as 3... The criteria are used to screen out values ​​that significantly deviate from the normal range, and combined with conventional meteorological models, anomalies in radar and satellite data are detected.

[0074] Quality assessment and weighting: The quality level of each data source is assessed, and the assessment indicators include the timeliness, reliability, accuracy, and consistency with other data sources. Low-quality data is assigned low quality weights to reduce its impact on forecast results. When fusing multi-source data, the quality weights are dynamically adjusted based on the historical performance and real-time data quality of each data source. The quality assessment results determine the weight of each data source in the fusion process.

[0075] The formula for calculating weights is: ,in For the first Data sources at time The weight, For this data source at time The quality assessment score is based on the data's reliability, missing percentage, and anomaly level.

[0076] Data quality control includes an observability mask, which is used to identify whether a data source is available at a certain time or spatial location. When a data source is missing at a certain time or in a certain area, the corresponding mask value is 0, and when the data source is valid, the mask value is 1.

[0077] By combining observability masks with quality weights, a dynamic control mechanism is provided for subsequent missing data inversion and completion, spatiotemporal coding, and multi-source fusion, enabling the model to maintain high reliability when dealing with low-quality or missing data.

[0078] Masking and weighting information play a crucial role in the missing value completion process, including missing value completion, similarity retrieval, and weighted completion.

[0079] Unifying data from different sources onto the same time axis and spatial grid ensures that subsequent feature fusion and model inference maintain consistency in the input caliber.

[0080] Differences among different data sources include different update times, different arrival delays, different observation time representativeness, and different spatial resolutions.

[0081] Define business time as a rolling timeline The step size is determined by the business update frequency. The original observation timestamp of each data source is mapped to the most recent business time index to correct for delay and representativeness errors. The time mapping follows the principle of alignment before fusion to avoid time series mismatch caused by mixing data with different time calibers during the fusion stage.

[0082] Time alignment is abstracted into a mapping operator, which transforms the original sequence of each data source... Transformed into a business timeline sequence ,in Represents the original timestamp. This represents a business time index, introducing a delay parameter for each data source. The time delay between the observation and the data entry is represented by the effective timestamp after delay correction. Perform alignment mapping.

[0083] Choose nearest neighbor or bounded window aggregation to align to the business time index. Nearest neighbor is used for data with high update frequency and stable latency, while window aggregation is used for data with low observation frequency or data that is jittery.

[0084] Bounded window aggregation will fall into the window The observations are weighted and synthesized according to quality weights to form representative values ​​for each operational time. It means that among them For the set of observations within the window, For quality weights.

[0085] Spatial alignment maps data from various sources to a unified grid. The location of the grid points is determined by the projected coordinate system, spatial resolution, and coverage area. Spatial alignment for each data source includes projection transformation and resampling. Projection transformation transforms the original coordinate system data to a unified coordinate system, and resampling maps the original resolution to a unified grid resolution.

[0086] An aggregation strategy is used to map high-resolution data to low-resolution grids, and an interpolation strategy is used to map low-resolution data to high-resolution grids, retaining uncertainty information related to interpolation for subsequent weight constraints.

[0087] Spatial mapping is abstracted into operators Data after time alignment Fields mapped to a unified grid ,use express.

[0088] A consistent caliber constraint spatial resampling strategy is adopted, and area or distance-weighted aggregation mapping is used to reduce the impact of single-point anomalies on grid values. Bilinear or distance-based interpolation mapping is used to generate interpolation confidence or effective neighborhood number after interpolation, which, together with quality weight, constrains subsequent fusion.

[0089] During the spatiotemporal alignment process, observability information is output synchronously to provide a basis for missing data identification, missing data completion, and fusion weighting. An observability mask for each data source is constructed on a unified grid. ,use This means that if the data is missing, exceeds the coverage area, is abnormal and is removed, or is of lower quality than the threshold, the mask is set to 0; otherwise, it is set to 1.

[0090] Based on the quality control output, and considering factors such as timeliness, missing test rate, anomaly severity, and cross-source consistency, a quality weight is constructed. The normalized weights are obtained by normalizing the available data to characterize its reliability. ,use This indicates that the weights are comparable when multiple sources are fused.

[0091] The data structure of the aligned output remains consistent throughout the training and runtime phases, at each business time step. The output consists of three types of fields: a set of multi-source feature fields with a unified grid. The corresponding set of observable masks The corresponding set of quality weights The organization is a fusion input package ,use express.

[0092] Alignment enables subsequent processes to achieve two things: First, training inputs and running inputs are consistent in terms of spatiotemporal scope, field organization, and observability representation, reducing output fluctuations caused by distribution offsets; Second, observability masks and quality weights are used throughout subsequent missing test inversion and fusion inference, automatically adjusting information contributions under missing test conditions or quality degradation, thereby improving the stability and traceability of continuous operation.

[0093] Based on the unified and aligned output, the results of missing test judgment, anomaly removal, and quality evaluation formed in the quality control stage are solidified into constraint fields for subsequent use. This allows the training and operation stages to use the same set of rules and generate the same type of fields, avoiding probability output fluctuations caused by caliber drift.

[0094] Missing tests and anomalies are used as prerequisites for generating mask fields. Each data source, each business time, and each grid position is checked to see if the data has arrived, if it is within the spatial coverage, if the key fields are complete, and if the quality control rejection rules have been triggered. If any condition is not met, the data is marked as unusable to avoid the misuse of invalid data in subsequent feature fusion and model inference. The quality degradation information of the data that has arrived but has not triggered rejection is retained. No binary processing is performed, and the contribution is constrained according to weights to avoid input breakage under boundary conditions.

[0095] By taking availability and reliability as the core objectives, multiple verifiable quality factors are combined to generate a quality weight field. Instead of directly taking a single indicator, the factors cover timeliness, completeness, and consistency. The judgment results of the quality control stage are transformed into continuously adjustable contribution constraints, so that the fusion inference maintains stable output when data quality fluctuates.

[0096] The timeliness factor reflects the deviation between the representativeness of the data time and the business time. The greater the deviation, the lower the weight. The completeness factor reflects the proportion of missing test blocks, the number of effective neighborhoods, and the situation of missing key elements. The more serious the missing, the lower the weight. The consistency factor reflects the degree of consistency with other data sources or historical statistical characteristics. The worse the consistency, the lower the weight.

[0097] A multiplicative structure is used to combine mass weights, with the mask acting as a gating term, so that the mass weights of unusable locations are automatically set to zero. It means that among them As a time-sensitive factor, As an integrity factor, As a consistency factor, its value range is normalized to... This enables review and verification.

[0098] In the engineering implementation, the rule version number, threshold configuration, and statistical summary of each training and business operation are recorded, including the availability rate of each data source, average weight, distribution of low-weight areas, and number of abnormal triggers. The source of output fluctuations is explained, which is used to screen high-value samples during subsequent closed-loop optimization and avoid low-quality samples from entering the retraining process, causing performance degradation.

[0099] The mask and quality weights are written as fixed fields into the aligned output structure, serving as common constraint inputs for subsequent missing test completion, fusion inference, and probability calibration. The mask is read first to determine whether completion is needed, and the quality weights are read to determine the contribution ratio of the retrieved samples and the completion results. Multi-source features are weighted or gated to reduce the impact of low-quality inputs on the output and reduce the risk of false alarm spread and missed detections in key areas.

[0100] Step 3: Adopt a label organization method with consistent caliber at both the event level and grid level. Event-level labels provide a continuous description of the process, while grid-level labels ensure spatial consistency of the output probability field. The label system transforms hail event information into a supervisory signal consistent with a unified time axis and grid, maintaining training stability under conditions of scarce samples and uncertain labels. It aligns the training objective with the business product form, avoiding inconsistencies between the label caliber during the training phase and the output caliber during the operation phase.

[0101] When event-level tags are stored in the database, they include the occurrence time window, description of the affected area, and identification of the credible source. A unified grid is used as the carrier to map event-level tags. Grid-level tags are generated in a rolling manner according to business time. A buffering strategy is introduced at the boundary of the time window to reduce the confidence of samples near the boundary participating in training, so as to reduce the interference of event boundary error on training. A transition strategy is introduced at the spatial boundary to distinguish between the core influence area and the transition area in terms of weight, so as to maintain the spatial smoothness and interpretability of the probability field without spreading noise.

[0102] Label confidence is used to quantify label reliability, which is then converted into loss weights during training. The confidence level is not subjectively assigned but is formed by verifiable source reliability, spatiotemporal matching consistency, and cross-source consistency. Source reliability reflects the integrity of the record and the strength of data support. Spatiotemporal matching consistency reflects the degree of matching between the label and key observation signals and observable states. Cross-source consistency reflects the degree of common support from multiple independent sources for the same process. The stronger the consistency, the higher the confidence level. If the label indicates an event but there is a long period of missing data, the confidence level of that region is reduced to prevent the model from being swayed by uncertain samples.

[0103] During the training phase, a weighted binary cross-entropy loss is used, with confidence scores introduced as grid point weights into the optimization process. The loss form is as follows: ,in For the total weight, For grid labels, The model outputs probabilities, and the total weights are composed of class imbalance weights and label confidence scores. Class imbalance weights mitigate the bias caused by the low proportion of hail samples, while label confidence scores suppress the influence of noisy labels on training. The total weights are... It means that among them For class imbalance weights, The confidence level of the label.

[0104] In sample governance, a strategy combining screening and weighting is adopted to remove samples with obvious errors or lack of supporting evidence to prevent contamination of the training set, while retaining uncertain but valuable samples with reduced confidence levels for training, thus maintaining sample coverage and controlling risk. Versioned management of label generation rules, confidence calculation rules, sample screening rules, and class weighting strategies is employed, with version numbers and parameter summaries permanently embedded in the training records. This makes model iterations traceable and performance changes interpretable, reviewable, and rollbackable.

[0105] Step four: The time and space of multi-source fusion forecasts carry stable patterns such as seasonal stages, intraday variations, and regional climate backgrounds. They are not simply index variables. If linear year, month, day, hour, or latitude and longitude values ​​are used, the model is prone to discontinuous expressions at the cycle boundaries, making it difficult to stably utilize the statistical similarity corresponding to the same seasonal stage, the same intraday time, and similar geographical backgrounds.

[0106] A multi-scale phase spatiotemporal encoding is constructed to uniformly map the intra-year cycle, intra-day cycle, and optional intra-month cycle to the phase space. This, along with the position encoding, forms a spatiotemporal embedding vector, which serves as the input feature of the fusion inference model and the basis for similarity measurement for missing test inversion and completion. This forms a consistent constraint caliber on both the training and operation sides, enhancing the model's ability to transfer and utilize periodic patterns and spatial similarities.

[0107] Time coding uses phase-based representation, assuming the service time is... The corresponding day number is Hourly rate The phase within the year is calculated using the formula. This indicates that the similarity between seasonal stages is expressed, among which Configure the year to 365 days or according to the actual number of days in the year. The periodic space is represented by sine and cosine pairs to avoid discontinuities caused by year-end boundaries. The corresponding intra-year phase encoding is as follows: It expresses the daily variation pattern, mapping the 24 hours of a day to... The periodic space is represented by sine and cosine pairs, and the intraday phase is represented by the formula. It means that among them Take 24. If a finer granularity is used, replace the hour with the minute. The corresponding intraday phase code is: .

[0108] Extending the monthly phases to characterize the cyclical structure of dates within the month, assuming the maximum number of days in the month is... On that day The phase within the month is The corresponding intra-month phase code is If the business does not require a monthly cycle, then this component should be left empty or not concatenated to keep the input dimension stable.

[0109] By relying on the combination of monthly and daily phases, without introducing additional manual rule terms, the model automatically learns the statistical association between this combination and hail risk during training. In engineering implementation, this is fixed as a configuration item and remains consistent during training and operation.

[0110] Location coding expresses spatial similarity and differences in terrain background, including the periodicity of latitude and longitude and latitudinal gradient characteristics. Let longitude be... Latitude is The altitude is Mapping longitude, which has a natural periodicity, to The phase space retains non-periodic latitudinal gradient information to distinguish north-south differences. Elevation is used as a key terrain component without additional encoding or normalization. The absolute value of elevation is used as an input parameter to make regions with similar elevations comparable in the feature space, providing terrain background constraints for subsequent similarity retrieval and fusion reasoning.

[0111] Longitude and latitude phases are defined as follows: , The corresponding position phase code is The absolute value of the altitude component is As continuous numerical feature input, without constructing normalized variables, the spatiotemporal embedding vector is obtained by concatenating the time code and the location code. ,in This indicates vector concatenation; if intra-month phase is enabled, it changes to... .

[0112] The usage includes two paths, sharing the same encoding caliber to ensure that training and execution remain consistent in spatiotemporal constraints, reducing output drift caused by changes in input distribution: First, ... As part of the model input, it is fused and inferred together with environmental field features and radar structural features, so that the model is constrained by periodic laws and spatial background when learning complex nonlinear relationships; secondly, in the missing measurement inversion completion, As the basis for similarity retrieval, it searches for similar historical time periods and similar regions to construct reliable alternative information when data is missing.

[0113] The system unifies the constraints on coding dimensions and parameter settings. The intra-year phase and intra-day phase are enabled by default, while the intra-month phase is an optional component. The latitude and longitude phase and the altitude normalized value are enabled by default. The scale processing method, constant term values, and whether they are enabled for all components are fixed in the configuration file. The version number is recorded during training and operation to maintain traceability in subsequent evaluation and iteration, ensuring that the project implementation is controllable. Through multi-scale phase spatiotemporal coding, the system stably expresses the similarity relationships across years, days, and regions, providing a unified constraint basis for subsequent missing data inversion and fusion inference.

[0114] Step 5: The business operation suffers from multiple sources of missing data and quality fluctuations. Missing data includes data that has not arrived at all, local occlusion, insufficient coverage, and gaps formed after anomaly removal. The quality degrades to below the unusable threshold. If missing data sources or missing data areas are discarded during the inference stage, the input structure will change over time, and the probability output will jump. In the rapid development stage of strong convection, this will cause non-physical fluctuations in the risk level.

[0115] Based on the unified alignment output and spatiotemporal embedding constraints, a similarity-based missing test inversion completion method is adopted. During the completion process, the completion confidence is output and fed back into the subsequent fusion inference and probability calibration stages, so as to achieve closed-loop control that makes the completion usable, the contribution controllable, and the results traceable.

[0116] If the status is unavailable or the quality assessment result is of low confidence, a missing test inversion completion is triggered. When a data source is unavailable at a grid point or the quality score is lower than a preset threshold, the completion process is initiated using a formula. It means that among them For the threshold, For quality weight, It is in an available state.

[0117] when At the same time, alternative estimates of the corresponding data source grid points are generated, and the missing elements are inverted using available samples from similar time periods and similar regions. The similarity measure takes the spatiotemporal embedding vector as the main body and superimposes a small amount of background state constraints to avoid cross-weather mismatches.

[0118] Let the target point be Candidate samples are The corresponding spatiotemporal embeddings are respectively and The base distance is the Euclidean distance. Similarity weights are constructed based on distance. The subsequent and quality terms determine the final contribution. As a similarity term, an exponential kernel is used to map the distance to similarity. This is the bandwidth parameter.

[0119] First, define the scope, then sort the samples to select a candidate sample set, making the calculation controllable and avoiding the introduction of irrelevant samples. Finally, limit the historical window in the time dimension. Alternatively, within the same seasonal window, a neighborhood or climate zone set can be defined in the spatial dimension, and the available state and quality weight of each candidate sample can be introduced to avoid using low-quality samples to fill in missing data.

[0120] Let the available states of the candidate samples and the normalized quality weights be respectively... and The combined weights used for completion are: The missing feature values ​​are generated using a weighted average method. Let the data source features to be filled be... Then in The completion result is ,in It is a stable term.

[0121] The completion confidence score expresses the reliability of the completion result, using a normalized form, and its value ranges from [value range missing]. Let the size of the candidate set be... To avoid dimensional differences between different grid points, the effective strength of the combined weights is defined as follows: This reflects whether there are enough reliable similar samples to support the completion.

[0122] When a missing test is triggered, the features in the fused input are replaced with the completed values, and the completed confidence level is written as an additional constraint field into the input packet, using the formula... This indicates that it is controlled by a test failure trigger switch for subsequent use.

[0123] Step 6: Feedback the completion confidence level to limit the contribution of the completion result to subsequent fusion inference. If the completion confidence level is higher than the preset threshold, it is considered a reliable input. If it is lower than the preset threshold, it is used as a reference signal and does not affect the probability output.

[0124] Complete the confidence level and quality weights to form the fusion gating coefficient. Adjusting the contribution of features in fusion, scaling features before they enter the model or adjusting weights in the fusion layer, explicitly passing the uncertainty of missing test completion to the probability output link, so that the output probability adapts to changes in input credibility, thereby improving the reliability and interpretability of probability products.

[0125] Record the candidate set range, bandwidth parameters, number of valid samples, completion confidence statistics, and trigger ratio during the completion process to quickly locate the cause when the product is abnormal and provide reusable analytical evidence for subsequent iterations.

[0126] By inverting missing data to complete the data and reinjecting confidence, the input structure is stabilized under conditions of missing data and quality fluctuations, reducing probability output jumps and avoiding the introduction of uncontrollable risks. Historical similarity patterns are used to compensate for insufficient observations, providing a reliable input basis for subsequent bi-branch fusion inference and calibration hierarchical output.

[0127] In the fusion reasoning and product generation stages, based on environmental background conditions and convection structure evidence, the grid-level hail occurrence probability is output. Through calibration and grading rules, risk products for business release are generated. Gating constraints are introduced to make the contribution of multi-source features adaptively change with the input credibility, avoiding non-physical influence on the results of a certain type of input under conditions of quality degradation or lack of testing.

[0128] The reasoning process is organized using a dual-branch structure. The two branches converge at the fusion layer and participate in probability estimation together with the spatiotemporal embedding. This allows the model to maintain a consistent output caliber across regions and seasons by utilizing the stable patterns of seasonal stages, intraday times, and spatial backgrounds. One branch characterizes the occurrence of environmental background, inputting environmental field elements and covering the thermodynamic and dynamic conditions related to convection triggering and maintenance, and outputting an environmental branch representation vector. The other branch characterizes the immediate evidence of convection structure, inputting radar structural features and covering echo intensity structure, vertical development, and evolution trends, and outputting a radar branch representation vector.

[0129] Before fusion, gating coefficients are constructed for each type of input to suppress the impact of low-quality or low-confidence completion results on inference. The gating coefficients are determined by normalized quality weights and completion confidence. ,in To normalize the quality weights, To supplement the confidence level, when the original data is available and of stable quality, Take 1 when using the missing test result completion method. The confidence level of the completion is taken, so that the completion is usable but its contribution is limited. The features are then gated and scaled before entering the model to form the final input features. This prevents the fusion layer from being amplified by low-confidence inputs.

[0130] A reproducible standard network structure is used for bi-branch feature extraction, ensuring fixed input and output dimensions, controllable inference overhead, and sliding window operation. The environment branch encodes the environmental field features of a unified grid at each business time step to obtain a representation vector. The radar branch encodes the radar structural features of the unified grid to obtain a representation vector. The spatiotemporal embedding vector is .

[0131] The fusion layer concatenates the inputs to the probability head and outputs the uncalibrated log odds. The probability output is obtained through the Sigmoid function. ,in The fusion and output combination function is implemented by several fully connected layers or lightweight attention layers. The fusion layer explicitly retains sensitivity to gating information. When the overall gating level is low, the output probability tends to be conservative, reducing the risk of false alarm propagation. A gating statistic is added to the fusion input or a gating modulation term is added to the probability header to adjust the gating mean. Introducing logarithmic odds adjustment, using the formula It means that among them This is a configurable coefficient.

[0132] Step 7: Use probability calibration to improve the comparability and reliability across regions and seasons without changing the ranking relationship, so that the actual occurrence frequency corresponding to the same probability value is consistent, and the threshold caliber and risk classification are unified.

[0133] Temperature parameters are obtained by minimizing calibration loss on the validation set using a temperature scaling method. If it is fixed in a versioned manner, then the probability after calibration is If the service sets different temperature parameters according to region or season, the parameter selection rules should be clearly defined in the configuration and the version number should be recorded.

[0134] Step 8: The calibration output generates a risk grading product. The grading rules adopt a fixed threshold system, and the thresholds are determined based on historical verification statistics and business cost constraints. Three threshold levels are set: low, medium, and high. The risk level is .

[0135] Output products include gridded probability field products. As a product for assessing the probability and risk of hail occurrence. As early warning support, it also outputs key traceability fields and quality summaries, including model version number, data version number, availability of each source, gating statistics, missing test completion trigger ratio, calibration parameters, threshold configuration, traceability fields, quality summaries, and product archives. When an anomaly occurs, it can quickly locate the cause and provide reusable evidence for subsequent iteration and optimization.

[0136] The operation organization is updated on a rolling basis according to the business rhythm. At each business moment, alignment, gating, dual-branch inference, calibration and classification are executed sequentially to output probability and risk products. The full-link statistics and key parameters are recorded in the log. The inference stage adopts batch grid inference and block parallel strategy to output stably in real time over a large grid area, avoiding the boundary effect of parallelism. The block strategy retains the boundary overlap area. In the output stage, splicing consistency processing is performed to make the probability field space continuous and avoid human splicing traces.

[0137] Dual-branch fusion and gating constraints complement each other between environmental background and radar structural evidence, adapting to different weather types and different development stages. Through probability calibration and risk classification, the output products have a unified caliber, maintainable thresholds, and a traceable evidence chain, directly serving business early warning and risk assessment scenarios.

[0138] In summary, a representative hail event is selected, and inputs and outputs are organized according to the rolling timeline of the operation. Data from multiple sources, including radar, satellite, ground observations, and numerical weather prediction, are aggregated using a unified timeline and grid to generate an aligned input packet. Quality weights and missing data trigger information are obtained, and similarity completion is performed on the missing data areas to obtain the completed values ​​and completion confidence. Gated scaling is then applied, and the data is input into a two-branch fusion model to output the uncalibrated probability. Temperature scaling calibration is performed, and the data is graded according to thresholds to form a probability field and risk classification product. Finally, traceability fields and quality summaries are archived.

[0139] The effectiveness is evaluated using a combination of process-level and grid-level approaches. The process-level approach verifies the reasonableness of the early warning lead time and coverage area, while the grid-level approach verifies probabilistic reliability and spatial consistency. Hit rate, false alarm rate, and comprehensive score indicators are recorded. Combined with the missing data trigger ratio and gating statistics, output discrepancies are explained. If significant missing data exists, the probabilistic performance of completing low-confidence regions is examined to verify whether gating constraints suppress the spread of false alarms caused by low-confidence inputs, thus validating the stability advantage of this invention under incomplete data conditions.

[0140] The above embodiments do not limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are included within the scope of protection.

Claims

1. A multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints, characterized in that, include: Step 1: Access and unified organization of multi-source meteorological data: Acquire multi-source meteorological data for hail forecasting, including radar echo data, satellite remote sensing data, ground meteorological observation data, and numerical weather prediction data. Use a unified data interface for receiving, storing, and managing the data, and record the source, timestamp, spatial range, resolution, update frequency, and quality status of each type of data to provide a unified input basis for subsequent alignment processing, feature construction, and inference calculation. Step 2, unifying spatiotemporal alignment and observability representation: A unified time axis and a unified spatial grid are established for multi-source meteorological data. The observation time, transmission delay, spatial projection, and resolution differences of various types of data are corrected, mapped, and resampled to ensure that various types of data are consistent with the same operational time and the same spatial reference. For the missing data, occlusion, boundary crossing, and quality degradation of various types of data in different regions and at different times, observability expression information is constructed, including availability status and quality weight, as the basis for subsequent missing data identification, completion triggering, and fusion weighting. Step 3, Sample Construction and Label Confidence Control: Training and prediction samples are constructed using a unified spatiotemporal benchmark. An event-level and grid-level consistent hail labeling system is established to unify the training and business output standards. A label confidence control mechanism is introduced to convert the label reliability into sample weights or loss constraints, thereby reducing the impact of noisy samples on model training. Step 4, Construction of multi-scale phase spatiotemporal coding: Multi-scale phase spatiotemporal coding is constructed for the seasonal, intra-day, and regional differences of hail, including intra-annual cycle coding, intra-day cycle coding, and geographic location coding. Monthly cycle coding is extended according to business needs. The time coding expresses the similarity of seasonal stages and the daily variation pattern, and the location coding expresses the spatial climate background and terrain constraints. The spatiotemporal coding serves as the structured input of the fusion inference model and the basis for the spatiotemporal similarity measurement in the missing data inversion process. Step 5: Missing Test Trigger Detection and Similarity Inversion Completion: If a data source is unavailable at a certain time or in a certain region, or if its quality weight is lower than a preset threshold, a missing data inversion and completion is triggered. Based on the spatiotemporal similarity measure, time periods and regions similar to the current process are retrieved from the historical sample library or neighborhood sample set. The missing data elements are then weighted and inverted to complete the missing data, and the corresponding completion confidence is output to characterize the reliability of the completion result. This serves as constraint information for subsequent fusion inference, preventing low-confidence completion results from affecting the final forecast. Step Six: Two-branch fusion reasoning of environmental field and radar structure information: After completing input alignment, quality constraints, and missing measurement completion, a multi-source fusion reasoning framework combining environmental field branch and radar structure branch is constructed. The environmental field branch extracts background feature elements related to hail triggering and maintenance, while the radar structure branch extracts real-time features related to convective cell development, echo intensity structure, and evolution trend. The two branches perform joint reasoning at the fusion layer, combining observable information, quality weights, and completion confidence to constrain the contribution of each source information and output the grid-level hail occurrence probability. Step 7, Probability Calibration Processing: Temperature scaling calibration and reliability correction are used to perform consistency calibration of hail occurrence probability, providing a basis for unified threshold release and risk assessment. Step 8: Risk classification generation and business product output: Based on the calibrated hail occurrence probability, the system generates corresponding risk level results according to a preset threshold system, outputting a hail occurrence probability field, risk classification map, and early warning information for key areas. It retains the correspondence between the input data status, alignment results, missing data completion status, quality weights, and output results for subsequent result traceability, error analysis, model evaluation, and continuous optimization.

2. The multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints according to claim 1, characterized in that, Step one includes: receiving real-time observation data, historical data, and numerical weather prediction data in parallel. The data includes radar echo data, satellite remote sensing data, ground meteorological station data, and numerical weather prediction data. Among them, radar echo data includes at least reflectivity and radial velocity, satellite remote sensing data includes at least cloud cover, temperature, and humidity, and ground meteorological station data includes at least temperature, humidity, wind speed, and air pressure. Data parsing, field standardization, and metadata registration are completed through a unified data interface to record the source, timestamp, spatial range, spatial resolution, update frequency, and quality status of each data source, forming a multi-source meteorological data input set.

3. The multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints according to claim 1, characterized in that, Step two includes: performing missing data identification, outlier identification and removal, and quality assessment on the multi-source meteorological data generated in step one; constructing a rolling operational timeline, and including the first... Original observation timestamps from each data source According to delay parameters Corrected to valid timestamp It also employs nearest neighbor mapping or bounded window quality-weighted aggregation to map the original sequence to the business time. Perform projective coordinate transformation and spatial resampling on the time-aligned data to map each data source to a unified spatial grid. Generate an observable mask for each data source, business time, and grid location. and normalized quality weights The observability mask is set to 0 when data is missing, exceeds coverage, is removed due to anomalies, or has a quality below the usable threshold; otherwise, it is set to 1. The multi-source feature fields, observability masks, and quality weights on the unified grid are organized into a fusion input package. .

4. The multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints according to claim 1, characterized in that, Step three includes: establishing event-level labels and grid-level labels, wherein the event-level labels include at least the hail occurrence time window, the affected area, and a reliable source identifier; mapping the event-level labels to grid-level labels according to a unified business timeline and a unified spatial grid, and setting buffer zones at the time window boundaries and core affected areas and transition zones at the spatial boundaries; calculating the label confidence level based on source reliability, spatiotemporal matching consistency, and cross-source consistency. The model was trained using weighted binary cross-entropy loss. ,in , For class imbalance weights, For grid labels, Output probabilities for the model; remove erroneous samples without supporting evidence, and use samples with uncertainty in training according to the reduced label confidence.

5. The multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints according to claim 1, characterized in that, Step four includes: setting the service time as The corresponding day number is The time of day is Constructing intra-annual periodic phase encoding and intraday cycle phase encoding Optionally, the maximum number of days in the month can be set as follows: The serial number of the day is Constructing intra-month periodic phase encoding Let the longitude be... latitude is The absolute value of altitude is Constructing geolocation codes and Concatenate the various codes into a spatiotemporal embedding vector. When monthly cycle phase coding is enabled, They are concatenated together; the spatiotemporal embedding vector is simultaneously input into the fusion inference model and used for similarity calculation in step five.

6. The multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints according to claim 1, characterized in that, Step five includes: when the first Each data source at business time and grid position The corresponding observability mask is 0, or the quality weight is lower than a preset threshold. At that time, according to Trigger missing test inversion and completion; construct a candidate sample set within a preset time window and spatial range. Calculate the target position With candidate samples Spatiotemporal embedding distance between and similarity weight The similarity weight is combined with the observability mask and quality weight of the candidate samples to form a... And calculate the completion value. and complete confidence level When missing test inversion completion is triggered, the corresponding input feature is replaced with the completion value, and the completion confidence is written into the fusion input packet.

7. The multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints according to claim 1, characterized in that, Step six includes: constructing gating coefficients based on the missing test trigger state, quality weights, and completion confidence. And perform gated scaling on the input features. ,in The original available features or the completed features generated in step five are used as input; the environmental field branch encodes the gated environmental field features to obtain the environmental representation. The radar structure branch encodes the gated radar structure features to obtain the radar characterization. By inputting environmental characterization, radar characterization, and spatiotemporal embedding vectors into the fusion function, the uncalibrated log probability is obtained. The probability of hail occurrence is obtained using the Sigmoid function. Optionally, the gating average can be... By introducing log-probability modulation, we obtain .

8. The multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints according to claim 1, characterized in that, Step seven includes: calibrating the uncalibrated log-probability obtained in step six using a temperature scaling method, and determining the temperature parameter on the validation set. and in accordance with Calculate the calibrated probability of hail occurrence; when gated mean modulation is not enabled, use Alternative The temperature parameters are configured according to uniform parameters, regional parameters, or seasonal parameters, and the corresponding versions are recorded.

9. The multi-source fusion hail probability forecasting method with multi-scale phase spatiotemporal coding and missing measurement inversion constraints according to claim 1, characterized in that, Step eight includes: setting two incremental risk grading thresholds to satisfy... According to the calibrated probability of hail occurrence Generate high-risk, medium-risk, and low-risk levels; output a gridded hail occurrence probability field. Risk-rated products It also outputs the model version, data version, calibration parameters, and threshold configuration simultaneously.

Images