Radar rainfall real-time estimation system based on double-method fusion and ground verification

Abstract

The invention provides a radar rainfall real-time estimation system based on double-method fusion and ground verification, belongs to the technical field of meteorological radar rainfall estimation, and adopts double-method fusion of a physical model estimation module and a driving model estimation module to cooperate with a causal inference fusion engine to improve the precision and adaptability of rainfall estimation and improve the rainfall estimation accuracy. The physical model module adaptively selects a Z-R, ZDR-R or KDP-R relation according to the quality of radar-based data and rainfall scenes, drives the model module to capture space-time nonlinear evolution characteristics of rainfall echoes and predicts short-time rainfall dynamics, and the causal inference engine identifies differentiated characteristics of two estimation fields, inferes hail, vertical airflow and other physical hybrid factors, and performs real-time prediction on the hail, the vertical airflow and other physical hybrid factors. A comprehensive rainfall estimation field is generated, a continuous error correction field converted from ground station observation data is corrected through a space-time error atlas network, and finally rainfall estimation information is generated for real-time rainfall estimation, so that the adaptability to a complex weather field and the precision of the estimation field are improved.

Description

Technical Field

[0001] This invention relates to the field of meteorological radar precipitation estimation technology, and more specifically, to a real-time radar precipitation estimation system based on dual-method fusion and ground verification. Background Technology

[0002] In natural meteorological monitoring, real-time precipitation detection is essential, and radar is a key tool for monitoring precipitation. Radar inverts precipitation intensity by emitting electromagnetic waves into the atmosphere and receiving the echoes scattered by precipitation particles. Traditional methods mainly rely on physical models such as the Zr relationship to convert radar data into precipitation amounts. However, due to limitations such as electromagnetic wave attenuation, non-meteorological echo interference, and the complexity of precipitation types (e.g., liquid rain, hail, or mixed precipitation), the estimation accuracy is often unsatisfactory. Furthermore, the spatiotemporal nonlinear evolution of precipitation processes makes it difficult for a single model to capture dynamic changes, and the sparsity of ground observation stations further limits calibration capabilities. In recent years, data-driven methods such as machine learning have been applied, but they lack physical interpretability and struggle to adapt to changing weather scenarios in real time, leading to biased estimation results under complex conditions.

[0003] The following technical problems exist in the existing technology: 1. Traditional radar precipitation estimation is limited by the limitations of a single physical model or data-driven model, making it difficult to adapt to the diversity of precipitation types and complex weather scenarios, resulting in insufficient estimation accuracy. 2. Traditional precipitation estimation systems lack real-time dynamic calibration capabilities, making it difficult to quickly adapt to the spatiotemporal changes in precipitation processes, resulting in delayed or significantly biased estimation results. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a real-time radar precipitation estimation system based on dual-method fusion and ground verification, the system comprising: The data acquisition module is used to acquire radar base data and ground station observation data; The physical model estimation module is used to filter the quality of different parameters in the radar base data and adaptively select the optimal precipitation estimation formula according to the preset physical rules to generate the first precipitation estimation field. The driving model estimation module is used to collect radar data fields from N consecutive past moments as sequence input, extract the nonlinear evolution characteristics of precipitation echoes in the spatiotemporal dimension, and generate a second precipitation estimation field. A causal inference fusion engine is used to identify the differential features between the first precipitation estimation field and the second precipitation estimation field. Based on the differential features and the physical properties of the radar base data, a physical confounding factor is inferred, and a comprehensive precipitation estimation field is generated by combining the first precipitation estimation field and the second precipitation estimation field. The spatiotemporal error mapping network module is used to calculate the deviation between the comprehensive precipitation estimation field and the ground station observation data at the ground station locations, and to generate a continuous error correction field for correcting the comprehensive precipitation estimation field. The meta-learning self-evolutionary calibration module is used to set meta-learning tasks and learn, generate update strategies for adjusting the internal parameters of each module in the system, and make online adjustments based on real-time feedback signals.

[0005] In another aspect, embodiments of the present invention also provide a radar precipitation real-time estimation system based on dual-method fusion and ground verification. The hardware of the system includes a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or code. The processor is used to execute the programs, instructions or code in the machine-readable storage medium to implement the above-mentioned method.

[0006] Based on the above, the beneficial effects of the radar precipitation real-time estimation system based on dual-method fusion and ground verification of the present invention are as follows: This invention employs a dual-method fusion of a physical model estimation module and a driving model estimation module, in conjunction with a causal inference fusion engine, significantly improving the accuracy and adaptability of precipitation estimation. The physical model module, based on the quality of radar-based data and the precipitation scenario, adaptively selects ZR, ZDR-R, or KDP-R relationships to generate a highly physically interpretable first precipitation estimation field. The driving model module, by capturing the spatiotemporal nonlinear evolution characteristics of precipitation echoes, predicts short-term precipitation dynamics, compensating for the limitations of the physical model in complex weather conditions. The causal inference engine identifies the differential characteristics of the two estimation fields, inferring physical confounding factors such as hail and vertical airflow, and generates a comprehensive precipitation estimation field through dynamic weighting. The spatiotemporal error mapping network further utilizes graph neural networks to transform ground station observation data into a continuous error correction field, optimizing overall accuracy and achieving a balance and adaptation between physical parameters and complex scenarios, thus improving adaptability to complex weather fields and the accuracy of the estimation field.

[0007] This invention achieves real-time dynamic optimization through a meta-learning self-evolutionary calibration module, ensuring the real-time performance and robustness of precipitation estimation. It utilizes historical and real-time precipitation data to set up a meta-learning task, generating a general parameter initialization strategy. The meta-optimizer dynamically adjusts the parameters of the physical model, driving model, and error correction module based on deviation feedback from ground station observation data, rapidly adapting to the spatiotemporal variations of different precipitation scenarios. The precipitation information generation and distribution module formats the corrected comprehensive precipitation estimation field, generating precipitation intensity distribution maps, total precipitation maps, and warning area maps, which are efficiently distributed via API or file interfaces to meet real-time application requirements. The entire system autonomously switches its calibration parameters based on environmental backgrounds such as season and weather type. The system ultimately generates a parameter optimization function, which, in the face of new weather processes, achieves a highly customized and efficient online adaptive mechanism by calling the function, ensuring high accuracy and stability during complex and variable precipitation events. Attached Figure Description

[0008] Figure 1 This is a structural diagram of a radar precipitation real-time estimation system based on dual-method fusion and ground verification provided in an embodiment of the present invention.

[0009] Figure 2 This is a data transmission relationship diagram of a radar precipitation real-time estimation system based on dual-method fusion and ground verification provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1: like Figures 1 to 2 As shown, a real-time radar precipitation estimation system based on dual-method fusion and ground verification is presented. The system includes: The data acquisition module is used to acquire radar base data and ground station observation data. It acquires the original radar base data collected by radar volume scan and the ground station observation data collected by station observation, and converts them into highly reliable grid fields that can be directly used in other modules. This ensures the quality and consistency of all input data. The radar base data includes reflectivity (Z), differential reflectivity (ZDR), differential phase shift rate (KDP), correlation coefficient (ρhv), radial velocity, beam height information, and volume scan sequence. Among them, reflectivity is used to characterize the overall echo intensity of precipitation particles, differential reflectivity is used to characterize particle morphology, differential phase shift rate is used to reflect the integral effect of raindrop morphology with distance, correlation coefficient is used to identify clutter interference and non-meteorological echoes, radial velocity is used to analyze airflow motion characteristics, and beam height and volume scan sequence are used to supplement the characterization of the three-dimensional structure of precipitation. The ground station observation data includes precipitation from minute-level automatic rain gauges, hourly observed precipitation, and gridded ground precipitation fields formed by interpolation between stations, which are used for comparison and calibration with radar estimation results.

[0012] During radar base data processing, the system receives the raw radar volume scan data stream in real time through the data acquisition module. This stream includes elevation angles, timestamps, and raw radar base data forming a raw binary file. The raw binary file is then unpacked into multi-parameter data cubes, each corresponding to a complete volume scan, with dimensions of [elevation angle, azimuth angle, radial distance]. A correlation coefficient (ρhv) threshold is then set, such as ρhv < 0.8, in conjunction with ZR correlation for joint discrimination. The system filters out non-precipitation particles such as insects, birds, diurnal variation, and ground clutter with low correlation coefficients. Specifically, if the radial velocity of a pixel is close to zero and the correlation coefficient is low, it is identified as ground clutter and removed. A radial velocity absolute value less than 1 m / s indicates almost no airflow movement. The low correlation coefficient is limited to ρhv < 0.8. This threshold is based on standard meteorological radar practices because the ρhv value reflects the coherence of the echo signal; high values ​​(close to 1) are considered less likely to cause interference. The value represents meteorological targets such as precipitation particles, while a low value represents non-meteorological interference such as scattering from insects, birds, or ground objects. The ρhv value range is usually from 0 to 1. Pixels with a radial velocity below 0.8 are considered static clutter (such as reflections from buildings or vegetation) under near-zero radial velocity conditions. The system applies this rule pixel by pixel to remove outliers in the corresponding data cube, while maintaining a log to record filtering events to ensure the physical authenticity and consistency of the input data. The system maintains a fuzzy logic classifier based on Z, ZDR, and KDP parameters. For example, if Z is very high (>55dBZ) but ZDR is close to 0, the echo is likely hail. The system will mark it for subsequent module processing and use an adaptive ZPHI algorithm to correct the attenuation of Z and ZDR library by library based on the cumulative KDP value along the radial direction. It identifies paths where the signal decays rapidly, i.e., the echo intensity drops sharply in the radial direction. These paths are marked with "[Possible Attenuation]" for the physical model estimation module to use for fusion, ensuring the authenticity of its physical values.

[0013] When processing ground station observation data, the system acquires minute-level automatic rain gauge data in real time via API, converting the ground station observation data from the original format (such as XML or JSON) into a unified internal data structure. Due to the uneven distribution of stations, the system uses the inverse distance weighted (IDW) interpolation method to interpolate the discrete station data onto a regular 1 km × 1 km grid, which is the same as the radar grid, and outputs the uncertainty estimate of the station difference. The interpolated ground precipitation field is used as the gold standard for subsequent calibration.

[0014] The physical model estimation module is used to filter the quality of different parameters in radar base data and adaptively select the optimal precipitation estimation relation according to preset physical rules to generate a first precipitation estimation field. Based on the quality-filtered radar base data, it adaptively selects or weights different empirical relations according to meteorological scenarios and parameter reliability to generate a precipitation rate field with physical interpretability and provides the uncertainty. The physical model in the physical model estimation module refers to a model that estimates precipitation rate based on the physical characteristics of radar base data through preset physical rules and empirical relations (such as ZR, ZDR-R, KDP-R relations). The data acquisition module provides processed radar base data, and the rule sources are obtained based on meteorological empirical relations and fuzzy rule generators to provide precipitation estimation based on physical rules. It emphasizes physical interpretability and the direct correlation of parameters, and is applicable to various precipitation scenarios such as weak precipitation, heavy precipitation, and hail. The module improves estimation accuracy by adaptively selecting relations, providing highly reliable and physically meaningful benchmark data for subsequent fusion and calibration. The specific content of the physical model estimation module includes: The radar base data is acquired and filtered and quality controlled according to the quality of different parameters, including clutter removal, non-meteorological echo identification and attenuation correction. The precipitation estimation formulas include ZR formulas, ZDR-R formulas and KDP-R formulas. Each precipitation estimation formula can be retrieved using existing technologies, and will not be elaborated here.

[0015] When the reflectance Z value is in a weak echo region, the KDP-R relationship is used; when a strong echo region appears, a mixture of ZR and ZDR-R relationships is combined to distinguish large raindrops from hail particles; under stratified cloud precipitation conditions, the linear relationship of KDP is preferentially used to enhance the sensitivity to small raindrop groups; under mixed rain conditions, a dynamic weight adjustment method is used to switch between ZR, KDP-R, and ZDR-R relationships to generate an adaptive first precipitation estimation field. Large raindrops specifically refer to raindrops with a liquid precipitation particle diameter greater than 2 mm, typically exhibiting Z > 40 dB in strong echo regions. ZDR > 1.5 dB (indicating elongated particle shape) indicates convective precipitation or warm cloud processes without ice phase interference. This is used to correct the nonlinear influence of large droplets on the ZR relationship in the ZR and ZDR-R mixing relation. Hail particles refer to solid ice spherical particles with a diameter greater than 5 mm. In the mixing phase determination, the parameters are limited to Z > 45 dB and ZDR ≈ 0 dB (close to spherical shape, low differential reflectivity). The environment is limited to severe convective weather with convective cloud tops exceeding 6 km. The system reduces the ZDR-R weight and labels these particles to suppress overestimation and avoid the negative impact of hail on the liquid phase assumption. Stratified cloud precipitation bars... The parameters for stable precipitation generated by stratiform clouds (such as warm front cloud systems) are limited to Z values ​​between 20-35 dBZ, KDP increasing linearly (<0.2° / km), ZDR <1 dB (uniform distribution of droplets), low vertical velocity (<2 m / s), and large-scale cloud thickness <3 km. A linear KDP-R relationship is preferred to enhance sensitivity to small raindrops. Small raindrops specifically refer to tiny liquid particles with a diameter less than 1 mm. In weak echo regions, the parameters are limited to Z <30 dBZ, low but reliable KDP (noise <0.1° / km), and ρhv >0.95, with environmental constraints. For stratified or weak convection processes, the KDP-R relationship is used to capture weak signals and avoid underestimation of ZR under low reflectivity. Rainy mixed conditions refer to complex scenarios where liquid and solid or different intensities of precipitation coexist. The parameters are limited to Z 30-50dB, large ZDR fluctuation (0-3dB), and nonlinear KDP variation. For transitional weather such as multiphase clouds under the influence of cold fronts or topography, the system switches between ZR, KDP-R and ZDR-R through dynamic weight adjustment (probability output by fuzzy logic classifier) ​​to ensure adaptive fusion to distinguish the above conditions and generate a physically interpretable precipitation rate field.

[0016] The system first evaluates the quality of each parameter after radar base data processing through the physical model estimation module. If the KDP noise in a certain area is too high or unusable, the system will avoid using KDP-dependent relationships. The system runs a set of fast determination rules for each pixel to generate local scene labels, including weak echo, layering, convection kernel, hail, blind zone, etc. The local scene labels are determined based on the collected values ​​of each parameter of the physical base data. The determination rules are as follows: Weak precipitation determination: If Z < 30 dBZ in a certain area, the system determines it to be a weak echo area and will prioritize the KDP-R relationship; Mixed phase determination: When Z>45dBZ and the ZDR value is close to 0, the system infers that there is hail in the area, selects the mixed relationship of ZR and ZDR-R, and assigns a lower weight to the ZDR-R relationship to reduce the negative impact of hail on the estimation; Then, the reliability of Z, KDP, and ZDR is evaluated respectively, and the parameter reliability scores (high / medium / low) of the three are output. Based on the parameter reliability scores, the precipitation estimation formula is selected: If KDP is reliable and Z has low reliability, KDP estimation should be used first. If Z is high and ZDR indicates large liquid raindrops, then ZR is the primary indicator, and ZDR is used for correction.

[0017] If hail is suspected, reduce the confidence level of ZR based on the liquid assumption, use ZDR or ρhv as auxiliary labels, and suppress overestimation of precipitation.

[0018] In complex precipitation scenarios, the system employs a fuzzy logic classifier. This classifier takes parameters such as Z, ZDR, and ρhv as input and outputs the probability that each grid point belongs to the stratiform cloud / rain, convective rain, or hail type. The system then uses these probabilities to weightedly fuse the relationships between different precipitation types. ; ; Where R is the final estimated precipitation rate; , , Dynamic weights; The precipitation rate is obtained from the ZR relationship; The precipitation rate is obtained from the KDP-R relationship; The precipitation rate is obtained from the ZDR-R relationship; Finally, the statistical consistency factor of the local scene label within the local window (e.g., 5×5) is used to smooth the weights and prevent isolated mutations from being misselected. The pixel estimate is generated according to the selected precipitation estimation formula or weighted combination, and the source (e.g., Z-dominated / KDP-dominated / mixed) is written into the metadata. Based on parameter confidence and local spatial consistency, it is attached to the pixel to generate the first precipitation estimation field and the uncertainty and weight information of each pixel. Among them, the pixel estimate is the precipitation rate of each grid point (1mm×1mm), which is calculated by ZR, KDP-R, ZDR-R relationship or their weighted combination. Each pixel corresponds to a precipitation intensity value, providing a precipitation intensity estimate for each grid point, forming a spatially continuous precipitation rate field, which is used to construct the first precipitation estimation field and output to the causal inference engine as the basic data of the comprehensive precipitation estimation field. Metadata is descriptive information attached to each cell, recording the source and context information of precipitation rate estimation, including specific precipitation estimation formulas, parameter confidence scores, local scenario labels, etc., and is attached to the first precipitation estimation field for analysis and calibration of the spatiotemporal error map network module. Local spatial consistency is a statistical consistency factor for local scene labels within a local window (e.g., a 5×5 grid) of a pixel. It is used to smooth weights and reduce isolated abrupt changes. The consistency factor is calculated based on the distribution consistency of local scene labels within the 5×5 window.

[0019] In this embodiment, when evaluating the credibility of Z, KDP, and ZDR respectively, firstly, the radar base data (Z, ZDR, KDP, ρhv, etc.) after quality control processing by the data acquisition module is acquired, including the results of clutter removal, non-meteorological echo identification, and attenuation correction. Then, a quality check is performed on each parameter. The Z check is based on its dynamic range (e.g., 10-60 dBZ is a reasonable range; exceeding this range may indicate noise or ground clutter, reducing credibility). The KDP check is based on noise level (assessed through the KDP standard deviation; if the standard deviation > 0.5° / km, the noise is considered too high, resulting in low credibility). The ZDR check is based on calibration deviation and physical consistency (e.g., ZDR < -1 or > 6 dB is abnormal, reducing credibility). Then, a joint discrimination is performed using the correlation coefficient (ρhv). For example, when ρhv < 0.8 and the Z or KDP value is abnormal, it is determined to be a non-precipitation echo (such as insects or ground features), further reducing credibility. Subsequently, a fuzzy logic classifier is applied to... Z, ZDR, KDP, and ρhv are inputs. Local scenario labels (such as weak precipitation, hail, and convection nuclei) are generated based on preset rules. For example, if Z > 45 dBZ and ZDR ≈ 0, hail is inferred, and the reliability of liquid precipitation estimation for Z is reduced. The high stability of KDP in the weak echo (Z < 30 dBZ) region increases its reliability. Then, the reliability score of each parameter is calculated by weighting the comprehensive quality check results (weight 50%), ρhv value (weight 30%), and scenario label (weight 20%). The score is mapped to high (>0.8), medium (0.5-0.8), and low (<0.5). Finally, the reliability score is recorded in the metadata and output to the precipitation estimation relation selection step (e.g., KDP-R relation is used first when KDP reliability is high, and ZR weight is reduced when Z reliability is low). It is also passed to the causal inference fusion engine for subsequent fusion and calibration. The whole process ensures the accuracy of parameter reliability through multi-dimensional analysis and optimizes the physical rationality and accuracy of precipitation estimation.

[0020] The driving model estimation module collects radar data fields from N consecutive past moments as sequential input, extracts the nonlinear evolution characteristics of precipitation echoes in the spatiotemporal dimension, and generates a second precipitation estimation field. The driving model of this module refers to a data-driven machine learning method (Convolutional Long Short-Term Memory Network, ConvLSTM) that extracts the nonlinear evolution characteristics of precipitation echoes in the spatiotemporal dimension from historical radar base data to predict future short-term precipitation fields. It emphasizes dynamic prediction capabilities, compensates for the shortcomings of physical models in predicting dynamic changes, enhances the system's short-term forecasting capabilities, provides a data-driven second precipitation estimation field, and improves overall estimation accuracy by combining with the physical model. The specific content of the driving model estimation module includes: The radar data field from the radar base data of the past N consecutive moments is obtained, and time series normalization and spatial noise reduction processing are performed. The radar data field is decomposed into intensity gradient and motion vector layer by layer to capture the spatiotemporal dynamic changes of precipitation echo. The radar data field consists of reflectivity, differential reflectivity and differential phase shift rate. Nonlinear evolution characteristics of precipitation echoes in the spatiotemporal dimensions are extracted, and a nonlinear dynamic model is constructed based on this to predict the intensity evolution and spatial translation of precipitation echoes on short time scales. This generates the radar data field for the next time step, and a second precipitation estimation field reflecting the future short-term precipitation development is generated based on the radar data field at the next time step. The nonlinear evolution characteristics include spatial and temporal features. Spatial features are obtained through moving vector field analysis, and temporal features are obtained by identifying the development and dissipation trends of strong echoes using a recursive residual method. Precipitation echoes refer to the electromagnetic wave echo signals reflecting precipitation particles in the radar base data, used to characterize the intensity, distribution, and spatiotemporal evolution of precipitation. These echoes originate from the radar base data provided by the data acquisition module, and are unpacked from the original binary file into a multi-parameter... A data cube, with dimensions [elevation angle, azimuth angle, radial distance], is used to extract nonlinear evolution features and generate a second precipitation estimation field, which is then output to the causal inference fusion engine. The short-term scale is limited to a real-time prediction window of 1-15 minutes, generating a radar data field for the next moment, typically within the next 1-5 minutes. Short-term precipitation is defined as precipitation development within the next 5-30 minutes, with specific parameters including predicted precipitation rate R0-100 mm / h, uncertainty <20%, and spatiotemporal consistency constraints (difference between adjacent grids <10%). The second precipitation estimation field reflects the short-term precipitation intensity distribution, movement trajectory, and total accumulation (e.g., accumulated rainfall <5 mm within 5 minutes), ensuring prediction coverage of convective outbursts or weak precipitation evolution, and serves as a dynamic supplement to the causal inference fusion engine. Nonlinear evolution characteristics refer to the intensity changes and movement trends of precipitation echoes in the spatiotemporal dimension, reflecting the dynamic evolution law of precipitation systems. This is achieved by first acquiring radar data fields from the past 10 consecutive moments, resampling them to a uniform spatial grid and temporal resolution, applying quantile or median standardization to each frame of data to eliminate elevation angle and calibration differences, applying spatial filtering (such as median filtering) to remove noise and non-meteorological echoes, recording time intervals and data integrity, marking missing frames to avoid interference, dividing the radar data field into intensity gradients and motion vectors, calculating the Z, ZDR, and KDP gradients of each grid point in the spatial dimension to reflect local changes in precipitation intensity, and extracting the spatial movement direction and velocity of the echoes through optical flow or correlation analysis. When extracting spatial features, moving vector field analysis is used to calculate the spatial translation characteristics of the echo. Based on radar data fields at adjacent time points, correlation tracking or optical flow algorithms are used to generate a vector field representing the horizontal movement direction and velocity of the echo. When extracting temporal features, a recursive residual method is used to identify the development and dissipation trends of strong echoes. The time difference is calculated for the Z values ​​at consecutive time points to extract the trend of intensity change (enhancement, weakening, or stabilization). Through time series analysis, periodic or sudden changes in the echo are detected. The spatial features (moving vector field) and temporal features (intensity change trend) are integrated into a multidimensional feature tensor. The dimensions of the multidimensional feature tensor are time, spatial grid, parameters, and feature type. The multidimensional feature tensor is input into a convolutional long short-term memory network to construct a nonlinear dynamic model. The model architecture and training of the convolutional long short-term memory network can be found in the descriptions in existing technologies, and will not be elaborated here.

[0021] The core model uses a convolutional long short-term memory network. The input layer is a multi-dimensional feature tensor with a dimension of 10. The 10 input radar data fields are standardized to ensure that the values ​​are within the effective input range of the model. The output is the radar data field at the next time step. Then, it is transformed into a second precipitation estimation field through a pre-trained precipitation inversion network. Compared with the traditional long short-term memory network, the introduction of convolutional operation enables it to capture spatial features and temporal series dependencies at the same time.

[0022] The precipitation inversion network converts the radar data field at the next moment into a precipitation intensity field using a simplified but stable ZR relationship. Post-processing optimization is performed by adding an adjustment factor based on prediction reliability. Spatiotemporal consistency constraints are applied for physical rationality checks to generate a smooth precipitation estimation field output. The output precipitation estimation data sequence provides prediction reliability indicators, recording the model's running status and performance metrics. The reliability adjustment factor is a correction parameter used in the precipitation inversion network to optimize the precipitation intensity field prediction. The ZR relationship conversion result is adjusted based on the credibility of the predicted radar field. The accuracy of the second precipitation estimation field is optimized by generating the prediction reliability and historical data through a convolutional long short-term memory network.

[0023] The causal inference fusion engine identifies the differential features between the first and second precipitation estimation fields. Based on these differential features and the physical properties of the radar-based data, it infers the potential physical confounding factors that cause the differences between the first and second precipitation estimation fields. Finally, it combines the first and second precipitation estimation fields to generate a comprehensive precipitation estimation field that eliminates the influence of confounding factors. Specifically, the causal inference fusion engine includes: Data from the first and second precipitation estimation fields are acquired, and the differences between them are compared based on spatial correlation analysis. A dual-field difference matrix is ​​calculated to identify regions of significant difference and to recognize discrepancies in spatial structure, intensity level, and movement trend. For example, if the physical model shows abnormally high values ​​in a certain area due to hail, while the driving model's predicted values ​​are smooth and consistent with the surrounding area, the system will identify this difference. The dual-field difference matrix is ​​calculated using the pixel-by-pixel absolute difference, diff = |R|. 物理 -R 驱动 The matrix is ​​generated (unit: mm / h), and then the criteria for identifying significantly different regions are multi-dimensional threshold judgments. Significantly different regions include intensity difference thresholds (diff > 5 mm / h or > twice the standard deviation of the average diff), spatial structure differences (gradient inconsistency within a 5×5 local window > 20%, such as isolated high peaks in the physical model), intensity level categories (physical model > 50 mm / h but driving force < 20 mm / h, marked as abnormally overestimated, possibly due to hail interference), and movement trend differences (vector field direction deviation > 30° or velocity difference > 5 m / s, categorized as dynamic). The system identifies regions with significant differences (e.g., mismatches in physical properties, such as deviations in convective core movement) and statistical significance. For example, if the physical model generates abnormally high hail values ​​(>100 mm / h) due to Z>45 dBZ, while the smoothing values ​​(<30 mm / h) of the driving model are consistent with the surrounding area, then it is determined to be a region of significant difference (category: dominated by confounding factors, such as hail or attenuation). The system further confirms the region boundaries through spatial correlation analysis (Pearson coefficient <0.7), ensuring that the identified regions of significant difference (coverage >10% of the grid) are used to infer physical confounding factors and dynamically weighted and fused to improve the accuracy of the comprehensive estimation. Based on the differential characteristics and physical properties in radar base data, physical confounding factors that may cause differences are inferred, and a causal relationship map is established to characterize the influence path of physical confounding factors on estimation accuracy. Physical confounding factors include, but are not limited to, hail, vertical airflow, electromagnetic wave attenuation and non-meteorological echoes. The credibility weights of the first and second precipitation estimation fields in different regions are determined based on the physical confounding factor. The reliability of the two estimation fields is dynamically weighted and combined according to the physical confounding factor to generate a comprehensive precipitation estimation field that takes into account both physical interpretability and data-driven prediction capability. Among them, the reliability of the two estimated fields is dynamically weighted and combined according to the physical confounding factor, and the fusion formula is: R 综合 =w 物理 ×R 物理 +w 驱动 ×R 驱动 , where w 物理 and w 驱动 The weights are dynamically calculated based on the probability of the physical confounding factor, satisfying w 物理 +w 驱动 =1 For example, when the probability of hail is high, w 物理 It will decrease.

[0024] The inference process for the physical confounding factors causing the differences is as follows: First, data from the first and second precipitation estimation fields are acquired, and the dual-field difference matrix is ​​calculated. Significant difference regions are identified through spatial correlation analysis, analyzing differences in spatial structure, intensity level, and movement trend. For example, if a physical model infers abnormally high precipitation rates in a certain region due to high Z-values, and the driving model's predicted values ​​are smooth and consistent with the surrounding areas, then that region is marked as significantly different. Next, based on the physical characteristics of the radar base data (such as Z, ZDR, KDP, ρhv), combined with attenuation, phase delay, and ground object scattering signals, a causal relationship map is constructed. Specifically, for Z>4... Regions with a ZDR of 5 dBZ and ZDR ≈ 0 are marked as potentially experiencing hail, regions with rapidly changing KDP are marked as electromagnetic wave attenuation, and regions with ρhv < 0.8 are marked as non-meteorological echoes. The probabilities of these factors are quantified using a fuzzy logic classifier. Then, the influence paths of confounding factors on estimation accuracy are analyzed using a causal relationship graph. For example, hail causes the physical model to overestimate precipitation rate, and vertical airflow causes smoothing bias in the driving model. Finally, based on the differential features and factor probabilities, the physical confounding factors (such as hail and attenuation) of each region are determined, and their influence weights on the first and second estimation fields are recorded. These weights are then output to the subsequent fusion step for dynamically weighted generation of a comprehensive precipitation estimation field.

[0025] In this embodiment, the first precipitation estimation field is a precipitation rate field generated by the physical model estimation module based on radar base data (Z, ZDR, KDP, ρhv, etc.) using physical rules and empirical relationships (such as ZR, KDP-R, ZDR-R). It includes precipitation intensity (mm / h), uncertainty, weight information (such as Z-dominant, KDP-dominant, or mixed) and metadata (scenario labels such as weak precipitation, hail, etc.) for each 1km×1km grid point. The data comes from quality-controlled radar base data and is output to the causal inference fusion engine to provide a physical interpretability benchmark for comprehensive precipitation estimation, reflecting the physical characteristics of precipitation particles. It is applicable to various precipitation scenarios and ensures estimation accuracy through adaptive relational selection and fuzzy logic classification.

[0026] The second precipitation estimation field is a short-term precipitation intensity field predicted by the driving model estimation module based on the Convolutional Long Short-Term Memory (ConvLSTM) network. It includes the predicted precipitation rate field (1km×1km grid), prediction reliability index and spatiotemporal consistency information. The data comes from radar data fields (Z, ZDR, KDP) from the past 10 consecutive time moments. After spatiotemporal feature extraction and nonlinear dynamic modeling, it is output to the causal inference fusion engine to capture the nonlinear spatiotemporal evolution of precipitation echoes (such as intensity changes and movement trends), providing data-driven short-term prediction capabilities and making up for the shortcomings of physical models in dynamically changing scenarios.

[0027] The integrated precipitation estimation field is a precipitation rate field generated by the causal inference fusion engine through dynamic weighted fusion of the first and second precipitation estimation fields. It includes the fused precipitation intensity (mm / h) and confidence weight (w). 物理 w 驱动 The data, including metadata (probabilities of confounding factors such as hail and attenuation), originates from the physical property analysis of the first and second precipitation estimation fields and radar-based data. It is output to the spatiotemporal error map network module to balance the interpretability of the physical model and drive the predictive ability of the model. By identifying physical confounding factors (such as hail and vertical airflow), the weights are dynamically adjusted to generate a more accurate precipitation field for subsequent correction and information generation.

[0028] The spatiotemporal error mapping network module is used to calculate the deviation between the comprehensive precipitation estimation field and the ground station observation data at the ground station locations, and to generate a continuous error correction field for correcting the comprehensive precipitation estimation field. Acquire integrated precipitation estimation field and ground station observation data, perform spatiotemporal resolution matching on the ground station observation data, align the spatial grid of the integrated precipitation estimation field, interpolate the ground station observation data to the same spatial grid structure as the integrated precipitation estimation field using the inverse distance weighted interpolation method, and calculate the deviation between the interpolated ground station observation data and the integrated precipitation estimation field at each ground station location. A ground station map is constructed based on ground station observation data. Ground stations serve as nodes in a dynamic map, and the edges of the dynamic map are determined by the physical correlation of spatiotemporal changes between stations. A graph neural network is used to learn the propagation law of deviation on the dynamic map. Dynamic spatial grid expansion and boundary smoothing are applied to generate a continuous error correction field covering the radar estimation range. The continuous error correction field is a gridded error field generated by the spatiotemporal error map network module based on the deviation between the comprehensive precipitation estimation field and the ground station observation data. It includes the deviation value (mm / h) of each 1km×1km grid point and the correction coefficient after spatial smoothing. The data comes from the comprehensive precipitation estimation field and the ground station observation data (interpolated to the same grid through IDW). The data is output to the precipitation information generation and distribution module. The graph neural network learns the deviation propagation law and extends it to the entire domain to correct the systematic error of the comprehensive precipitation estimation field, ensuring the accuracy and spatial consistency of the final precipitation estimation information and meeting the needs of meteorological forecasting and disaster prevention.

[0029] The continuous error correction field is superimposed onto the comprehensive precipitation estimation field to perform global correction on the estimation results. The continuous error correction field is superimposed pixel by pixel onto the comprehensive precipitation estimation field (R correction = R comprehensive + error correction field) to perform global correction on the estimation results. The corrected comprehensive precipitation estimation field is output to the precipitation information generation and distribution module to generate the final real-time precipitation estimation information.

[0030] The meta-learning self-evolutionary calibration module is used to set meta-learning tasks and learn, generate update strategies for adjusting the internal parameters of each module in the system, and make online adjustments based on real-time feedback signals.

[0031] Historical and real-time precipitation data from radar data fields, comprehensive precipitation estimation fields, and continuous error correction fields are acquired to train a defined meta-learning task. This task involves learning across different precipitation events and updating a common initialization parameter set. This allows the model to quickly converge to optimal parameters with only a small amount of data and gradient step size when facing new meta-tasks. The initialization parameter set, generated through cross-task learning across different precipitation events, is used to initialize the models of various system modules, enabling them to quickly adapt and optimize for new precipitation scenarios. Specifically, this includes the fuzzy logic classifier thresholds for the physical model estimation module (e.g., hail thresholds for Z>45dBZ, and miscellaneous thresholds for ρhv<0.8). The empirical coefficients of the ZR, KDP-R, and ZDR-R relationships (wave filtering threshold), the initial weights and learning rate of the ConvLSTM network driving the model estimation module, and the initialization of the graph neural network edge weights and the bias propagation learning rate of the spatiotemporal error map network module are used to form a low-dimensional parameter vector or tensor (the dimension depends on the total number of modules, for example, the total parameter space is compressed to 100-500 dimensions). The meta-learning task set is constructed from historical and real-time precipitation process data (including radar data fields Z / ZDR / KDP, integrated precipitation estimation fields, continuous error correction fields, and bias feedback from ground station observation data). The set is divided according to precipitation scenarios and multi-task training is performed. The meta-optimizer is used to aggregate the gradients of the inner loop in the outer loop to update the initial parameters. When a new precipitation event occurs, the new data is used as the input of the meta-task. Starting from the general initialization parameters, a small number of internal loops are performed for fine-tuning, which quickly converges to the task-specific optimal parameters. These parameters are then applied to the module update strategy to improve the system's adaptability and robustness in various precipitation scenarios, ensure parameter synchronization, and ultimately improve the overall precipitation estimation accuracy and real-time response capability. At the same time, the general parameters are continuously iterated through real-time deviation feedback to achieve the system's self-evolution calibration. The meta-optimizer is used to generate update strategies for adjusting the internal parameters of the physical model estimation module, the driving model estimation module, and the spatiotemporal error mapping network module in the system. The meta-optimizer uses the real-time deviation feedback between ground station observation data and the corrected comprehensive precipitation estimation field as a new meta-learning task, fine-tunes the update strategy using common initialization parameters, and performs online adjustments to the system's self-evolution calibration.

[0032] The meta-learning self-evolutionary calibration module includes multi-dimensional feature extraction and precipitation scene classification of data from meta-learning tasks. The extraction process divides historical and real-time data into spatiotemporal patterns. When generating update strategies, it prioritizes adjusting the parameter synchronization mechanism between modules. During online adjustment, it performs iterative parameter updates through dynamic analysis of real-time deviation signals and multi-module linkage optimization.

[0033] The meta-learning task is a task within the meta-learning self-evolutionary calibration module that learns universal initialization parameters across different precipitation processes to quickly adapt to new precipitation scenarios and optimize system module parameters. Specifically, it involves using historical and real-time precipitation process data as input to train a meta-learning model to generate universal initialization parameters. This allows the system to quickly converge to optimal parameters with only a small amount of data and gradient step size when facing new precipitation data. The data comes from radar data fields (Z, ZDR, KDP, etc.), integrated precipitation estimation fields, continuous error correction fields, and ground station observation data. A task set is constructed through multi-dimensional feature extraction and precipitation scenario classification (such as weak precipitation, convective rain, hail). The data is output to the meta-optimizer to generate parameter update strategies, improving the system's adaptability and calibration accuracy under different precipitation scenarios, ensuring that the system quickly adapts to new scenarios and reduces errors.

[0034] The meta-optimizer is an optimization algorithm used in the meta-learning self-evolutionary calibration module to generate parameter update strategies for system modules. Based on feedback from the meta-learning task, it dynamically adjusts the internal parameters of the physical model estimation module, the driving model estimation module, and the spatiotemporal error mapping network module. Specifically, it includes gradient descent-based meta-optimization algorithms (such as MAML or Reptile), which generate parameter adjustment strategies based on general initialization parameters through cross-task gradient updates. The data comes from the input of the meta-learning task (historical and real-time precipitation process data, including radar data fields, comprehensive precipitation estimation fields, and continuous error correction fields) and real-time bias feedback (the deviation between the corrected comprehensive precipitation estimation field and the ground station observation data). The data is output to the parameter update interface of each module to optimize module parameters, improve precipitation estimation accuracy and adaptability, and achieve multi-module linkage optimization by dynamically analyzing bias signals to ensure the robustness and accuracy of the system in complex precipitation scenarios.

[0035] The generation process of the update strategy in the meta-learning self-evolutionary calibration module is achieved through a meta-optimizer that dynamically optimizes module parameters based on cross-task learning and real-time feedback. Specifically, the process is as follows: First, historical and real-time precipitation data are collected, including radar data fields (Z, ZDR, KDP), comprehensive precipitation estimation fields, and continuous error correction fields. Multi-dimensional feature extraction (such as intensity gradient and spatiotemporal pattern) is performed in conjunction with ground station observation data, and meta-learning tasks are divided according to precipitation scenarios (such as weak precipitation and convection kernels). Next, the meta-learning model (such as MAML) is initialized, setting common initial parameters for the physical model estimation module, the driving model estimation module, and the spatiotemporal error mapping network module. Then, inner-loop optimization is performed on each meta-learning task, quickly adapting to task-specific data using a small gradient step size, and calculating... The module parameters (such as ConvLSTM weights and fuzzy logic classifier thresholds) are processed using task-level gradients. Subsequently, in the outer loop, a meta-optimizer (such as Reptile) updates general parameters based on multi-task gradient aggregation, generating a cross-task general update strategy. Next, combined with real-time bias feedback (the deviation between the comprehensive precipitation estimation field and ground station observation data), the update strategy is fine-tuned online by dynamically analyzing the bias signal, prioritizing the adjustment of parameter synchronization mechanisms between modules (such as the weight allocation of the physical model and the driving model). Finally, the update strategy is applied to each module to update its internal parameters (such as ZR relationship coefficients and ConvLSTM hidden layer weights), and the optimization performance indicators are recorded and output to subsequent calibration loops to ensure that the system quickly converges to the optimal parameter configuration under different precipitation scenarios.

[0036] The meta-learning self-evolutionary calibration module provides structured input for the meta-learning task through multi-dimensional feature extraction and precipitation scene classification. First, it collects historical and real-time precipitation data, including radar data fields (Z, ZDR, KDP, ρhv), integrated precipitation estimation fields, continuous error correction fields, and ground station observation data. Next, it performs multi-dimensional feature extraction, which includes temporal features (calculating the intensity trends of Z, ZDR, and KDP using recursive residual methods, such as enhancement or dissipation), spatial features (extracting the spatial distribution and movement trajectory of echoes through gradient analysis and moving vector fields), physical features (identifying hail based on Z>45dB and ZDR≈0, or identifying non-meteorological echoes based on ρhv<0.8), and error features (the spatial distribution of the deviation between the integrated precipitation estimation field and ground station data). These features are integrated into a multi-dimensional feature tensor with dimensions of [time, grid, parameters, feature type]. Then, precipitation scene classification is performed. The classification process uses a pre-trained classifier (such as a random forest or neural network) as input, and outputs precipitation scene labels (such as weak precipitation, stratified clouds, convective nuclei, hail) as input. The features are weighted by fuzzy logic rules or machine learning models. For example, Z < 30dBZ is labeled as weak precipitation, and Z > 45dBZ and ZDR ≈ 0 is labeled as hail. At the same time, the spatiotemporal pattern (such as echo movement speed) is combined to divide the dynamic scene. The classification results are used for group meta-learning tasks to ensure that the tasks are organized according to precipitation type and spatiotemporal characteristics. The results are output to the meta-optimizer to generate update strategies. The whole process enhances the structured representation of the task by extracting multidimensional features and scene classification, and improves the adaptability of the meta-learning model to different precipitation scenes.

[0037] The system also includes a precipitation information generation and distribution module, which receives the comprehensive precipitation estimation field corrected by the meta-learning self-evolutionary calibration module, performs data formatting and visualization processing, and generates the final real-time precipitation estimation information, including a precipitation intensity distribution map, a total precipitation map, and a precipitation warning area map. This information is then distributed to external systems via API or file service interfaces. These external systems include meteorological service agencies (such as the National Meteorological Administration), disaster prevention and mitigation departments (such as water resources departments and emergency management departments), agricultural management departments, aviation transportation systems, and public weather service platforms. The distribution steps are as follows: First, the corrected comprehensive precipitation estimation field is obtained, containing gridded precipitation intensity data and uncertainty information; then, data formatting processing is performed, converting the precipitation data into a standard format (such as NetCDF, GRIB2) to generate the precipitation intensity distribution... The system generates a precipitation rate field (1km×1km grid), a total precipitation map (cumulative precipitation over a specified time period), and a precipitation warning area map (high-risk areas are marked based on intensity thresholds). Then, it performs visualization processing, using graphics libraries (such as Matplotlib or GIS tools) to generate two-dimensional or three-dimensional precipitation distribution images, complete with color coding and uncertainty annotations. Subsequently, the formatted data and visualization information are pushed to external systems in real time via API services (such as RESTful API), supporting dynamic queries in JSON or XML formats, or periodically uploading data files to designated servers via file service interfaces (such as FTP, SFTP). Finally, a distribution log is recorded, the data transmission status is monitored, and the integrity of external system reception is verified to ensure timely and accurate information delivery. The entire process meets the needs of meteorological forecasting, disaster prevention decision-making, and public services. Specific Implementation Example 2: like Figures 1 to 2 As shown, based on the content of the above specific embodiments, the following content is further disclosed: This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code that, when executed by the one or more processors, can perform the radar precipitation real-time estimation system based on dual-method fusion and ground verification as described above.

[0039] The methods or systems according to embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as a ROM or hard disk, may store the radar precipitation real-time estimation system based on dual-method fusion and ground verification provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components in the electronic device shown in this application may be omitted according to actual needs.

[0040] Example 3 One embodiment of this application discloses a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the real-time radar precipitation estimation system based on dual-method fusion and ground verification according to an embodiment of this application, as described with reference to the above figures, can be executed. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0041] Furthermore, according to embodiments of this application, the processes described in the above-described flowcharts can be implemented as computer software programs. For example, this application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided in this application, such as a radar precipitation real-time estimation system based on dual-method fusion and ground verification. When this computer program is executed by a central processing unit (CPU), it performs the functions defined in the method of this application.

[0042] The system hardware includes a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the methods described above.

[0043] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0044] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0045] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

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

Claims

1. A radar precipitation real-time estimation system based on dual-method fusion and ground verification, characterized in that, The system includes: The data acquisition module is used to acquire radar base data and ground station observation data; The physical model estimation module is used to filter the quality of different parameters in the radar base data and adaptively select the optimal precipitation estimation formula according to the preset physical rules to generate the first precipitation estimation field. The driving model estimation module is used to collect radar data fields from N consecutive past moments as sequence input, extract the nonlinear evolution characteristics of precipitation echoes in the spatiotemporal dimension, and generate a second precipitation estimation field. A causal inference fusion engine is used to identify the differential features between the first precipitation estimation field and the second precipitation estimation field. Based on the differential features and the physical properties of the radar base data, a physical confounding factor is inferred, and a comprehensive precipitation estimation field is generated by combining the first precipitation estimation field and the second precipitation estimation field. The spatiotemporal error mapping network module is used to calculate the deviation between the comprehensive precipitation estimation field and the ground station observation data at the ground station locations, and to generate a continuous error correction field for correcting the comprehensive precipitation estimation field. The meta-learning self-evolutionary calibration module is used to set meta-learning tasks and learn, generate update strategies for adjusting the internal parameters of each module in the system, and make online adjustments based on real-time feedback signals.

2. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 1, characterized in that, The radar base data includes reflectivity, differential reflectivity, differential phase shift rate, correlation coefficient, radial velocity, beam height information, and volume scan sequence. The ground station observation data includes precipitation from minute-level automatic rain gauges, hourly observed precipitation, and gridded ground precipitation fields formed by interpolation between stations.

3. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 1, characterized in that, The process of filtering different parameters in radar base data and adaptively selecting the optimal precipitation estimation formula according to preset physical rules to generate the first precipitation estimation field includes: The radar base data is acquired and filtered and quality controlled according to the quality of different parameters, including clutter removal, non-meteorological echo identification and attenuation correction. The precipitation estimation formulas include ZR relationship, ZDR-R relationship and KDP-R relationship. When reflectance values ​​are in weak echo regions, the KDP-R relationship is used; when strong echo regions are present, a mixture of ZR and ZDR-R relationships is used to distinguish between large raindrops and hail particles; under stratified cloud precipitation conditions, the linear relationship of KDP is given priority; under mixed rainy conditions, the ZR, KDP-R, and ZDR-R relationships are switched through dynamic weight adjustment to generate an adaptive first precipitation estimation field.

4. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 1, characterized in that, The process of collecting radar data fields from N consecutive past moments as sequence input, extracting the nonlinear evolution characteristics of precipitation echoes in the spatiotemporal dimension, and generating a second precipitation estimation field includes: The radar data field from the radar base data of the past N consecutive moments is obtained and subjected to time series normalization and spatial noise reduction processing. The radar data field is decomposed layer by layer to capture the spatiotemporal dynamic changes of precipitation echoes. The spatiotemporal dynamic changes include intensity gradient and motion vector. The radar data field is composed of reflectivity, differential reflectivity and differential phase shift rate. Nonlinear evolution characteristics of precipitation echoes in the spatiotemporal dimension are extracted, and a nonlinear dynamic model is constructed based on this. The intensity evolution and spatial translation of precipitation echoes on the time scale are predicted, generating the radar data field at the next moment. Based on the radar data field at the next moment, a second precipitation estimation field reflecting the future precipitation development is generated. The nonlinear evolution characteristics include spatial characteristics and temporal characteristics. Spatial characteristics are obtained through moving vector field analysis, and temporal characteristics are obtained by identifying the development and dissipation trends of strong echoes through the recursive residual method.

5. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 1, characterized in that, The method for identifying the differential characteristics between the first and second precipitation estimation fields, inferring potential physical confounding factors that cause the differences between the first and second precipitation estimation fields based on the differential characteristics and the physical properties of the radar-based data, and generating a comprehensive precipitation estimation field that eliminates the influence of confounding factors by combining the first and second precipitation estimation fields, includes: Data from the first and second precipitation estimation fields were obtained, and the differences between the two precipitation estimation fields were compared based on spatial correlation analysis to identify the differentiated features in spatial structure, intensity level and movement trend. Based on the aforementioned differential characteristics and the physical properties of radar base data, physical confounding factors are inferred, and a causal relationship map is established to characterize the influence path of physical confounding factors on estimation accuracy. The physical confounding factors include hail, vertical airflow, electromagnetic wave attenuation, and non-meteorological echoes. The credibility weights of the first and second precipitation estimation fields in different regions are determined based on the physical confounding factor. The reliability of the two estimation fields is dynamically weighted and combined according to the physical confounding factor to generate a comprehensive precipitation estimation field that takes into account both physical interpretability and data-driven prediction capability.

6. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 1, characterized in that, The method for calculating the deviation between the integrated precipitation estimation field and the ground station observation data at the ground station locations, and generating a continuous error correction field for correcting the integrated precipitation estimation field, includes: Acquire integrated precipitation estimation field and ground station observation data, perform spatiotemporal resolution matching on the ground station observation data, align the grid structure of the integrated precipitation estimation field, interpolate the ground station observation data to the same spatial grid as the integrated precipitation estimation field using the inverse distance weighted interpolation method, and calculate the deviation between the interpolated ground station observation data and the integrated precipitation estimation field at each ground station location. A ground station map is constructed based on ground station observation data. The ground stations are used as nodes of the dynamic map. The edges of the dynamic map are determined by the physical correlation of spatiotemporal changes between stations. The propagation law of deviation on the dynamic map is learned through graph neural network. Dynamic grid expansion and boundary smoothing are applied to generate a continuous error correction field covering the radar estimation range. The continuous error correction field is superimposed on the comprehensive precipitation estimation field to perform global correction on the estimation results.

7. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 1, characterized in that, The method for setting meta-learning tasks and performing learning, generating update strategies for adjusting the internal parameters of each module in the system, and making online adjustments based on real-time feedback signals includes: Historical and real-time precipitation process data from radar data field, comprehensive precipitation estimation field, and continuous error correction field are acquired to train a set meta-learning task. The meta-learning task is set to learn between different precipitation processes and update a general initialization parameter. An update strategy for adjusting the internal parameters of the physical model estimation module, the driving model estimation module, and the spatiotemporal error mapping network module in the system is generated using a meta-optimizer. The meta-optimizer uses the real-time deviation feedback between ground station observation data and the corrected comprehensive precipitation estimation field as a new meta-learning task, fine-tunes the update strategy using common initialization parameters, and performs online adjustments to the system's self-evolution calibration.

8. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 7, characterized in that, The meta-learning self-evolutionary calibration module includes multi-dimensional feature extraction and precipitation scene classification of data from meta-learning tasks. The extraction process divides historical and real-time data into spatiotemporal patterns. When generating update strategies, it prioritizes adjusting the parameter synchronization mechanism between modules. During online adjustment, it performs iterative parameter updates through dynamic analysis of real-time deviation signals and multi-module linkage optimization.

9. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 8, characterized in that, The system also includes a precipitation information generation and distribution module, which receives the comprehensive precipitation estimation field corrected by the meta-learning self-evolution calibration module, performs data formatting and visualization processing, and generates the final real-time precipitation estimation information. The precipitation estimation information includes a precipitation intensity distribution map, a total precipitation map, and a precipitation warning area map, and is distributed to external systems through API or file service interfaces.

10. The radar precipitation real-time estimation system based on dual-method fusion and ground verification according to claim 1, characterized in that, The hardware of the system includes a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the content of any one of claims 1-9.

Images