A method for retrieving precipitation particle size distribution using dual-polarization radar

Abstract

The application discloses a method for retrieving precipitation particle size distribution by using dual-polarization radar, and belongs to the technical field of meteorological remote sensing and data retrieval. The method introduces multi-source observation data, combines physical models and machine learning techniques, establishes a particle size distribution retrieval model, and couples physical consistency constraints in the retrieval process to realize high-precision retrieval of radar precipitation particle size distribution. At the same time, based on statistical analysis and uncertainty quantification method, the uncertainty of the retrieval result is quantitatively evaluated. The application improves the accuracy and reliability of the radar precipitation particle size distribution retrieval, enhances the physical interpretability of the model, is suitable for quantitative precipitation estimation and hydro-meteorological analysis in complex meteorological environments, and has a wide application prospect.

Description

Technical Field

[0001] This invention belongs to the field of meteorological remote sensing and data inversion technology, and more specifically relates to a method for inverting precipitation particle size distribution using dual-polarization radar. Background Technology

[0002] Quantitative radar detection of precipitation particle size distribution is a core issue for in-depth research into precipitation microphysical processes, improving quantitative precipitation forecasting capabilities, and achieving precise water resource management and meteorological disaster prevention. Traditional methods often rely on simple empirical relationships or physical models with limited parameters for particle size distribution inversion, making it difficult to simultaneously account for the complex characteristics of multi-source observations and the diverse spatial variations of particle size distributions. This results in limited confidence levels, insufficient physical consistency, and inadequate robustness of the inversion results. With the widespread deployment of new-generation observation systems such as "dual-polarization radar + Doppler," it has become possible to acquire richer radar multi-channel parameters and auxiliary meteorological observation data, which places higher demands on improving the accuracy of particle size distribution inversion.

[0003] Meanwhile, deep learning and data-driven methods have demonstrated powerful feature extraction and distribution modeling capabilities in the field of meteorological remote sensing inversion. However, their lack of endogenous integration of physical constraints may lead to "black box" risks and insufficient physical interpretability. Current technologies have not yet effectively addressed how to integrate physical priors, perturbation consistency, and adaptive multi-source observation multi-scale features to achieve high-reliability particle size distribution inversion and uncertainty quantification. Therefore, developing a novel radar precipitation particle size distribution inversion method that integrates physical constraints and data-driven approaches can accurately characterize uncertainties in the inversion process, providing a reliable data foundation for subsequent quantitative precipitation forecasting, meteorological research, and operational applications. This has significant scientific and practical value. Summary of the Invention

[0004] This invention aims to address the technical problems existing in current radar precipitation particle size distribution inversion methods, such as insufficient ability to fuse multi-source observation information, inadequate introduction of physical constraints, and limited ability to assess the confidence and uncertainty of inversion results. It provides a method that can fully explore and utilize multi-modal observation data to achieve high-precision radar precipitation particle size distribution inversion and quantitative assessment of its uncertainty, while ensuring physical consistency.

[0005] To achieve the above objectives, the present invention employs the following technical solution: the method comprises the following:

[0006] Step 1: Data preprocessing and feature space construction. Using the original observation data from multi-source radar and auxiliary meteorological parameters, data cleaning and quality control are performed, and a three-dimensional feature matrix related to particle size inversion is obtained through spatial registration, spatiotemporal interpolation and other methods.

[0007] Step 2: Multimodal feature coupling and encoding. Using a multi-network structure and attention mechanism, features from different sources are fused to obtain a high-dimensional feature representation that can characterize the evolution of particle size structure, and then standardized encoding is performed.

[0008] Step 3: Based on physical constraints and prior-driven particle size inversion modeling, the particle size distribution inversion network is dynamically perturbed, integrating the parameter embedding of the particle size distribution physical model, introducing an adjustable perturbation term at the network output, and combining an innovative perturbation consistency discrimination loss function to improve the accuracy and physical consistency of particle size distribution inversion.

[0009] Step 4: Multi-source mutual verification and adaptive correction of inversion results. A multi-scale virtual observation consistency algorithm is adopted, which uses various theoretical particle size distribution assumptions to simulate and generate radar observations. The results are compared and verified with the output of the neural network, and the weights of the network disturbance terms are adaptively optimized and fed back.

[0010] Step 5: Output particle size distribution results and uncertainty assessment. Based on the data-physical joint uncertainty quantification method, the variance of the output distribution of the integrated model, the perturbation consistency residual and the dispersion of the observed characteristic distribution are used to give the confidence interval of the inverted particle size distribution results and perform uncertainty assessment.

[0011] In one approach, the dynamic perturbation particle size distribution inference network uses generalized gamma distribution parameters as physical priors when outputting the particle size distribution probability density function. By introducing a perturbation consistent discrimination loss function, it forces the gradient changes between the physical principal distribution and the perturbation term learned by the network to remain consistent, thereby achieving high accuracy and stability in particle size distribution inversion.

[0012] In one approach, the multi-scale virtual observation consensus algorithm compares the consistency error between simulated radar observations and actual observations or observations under multiple distribution assumptions, and adaptively adjusts the perturbation weights of the dynamic perturbation particle size distribution inference network based on the error signal, thereby achieving automatic correction and accuracy improvement of the distribution inversion results.

[0013] In one approach, the data-physical joint uncertainty quantification method constructs a joint index by integrating the sample variance output by the integrated model, the consistency residual of the physical principal distribution, and the distribution dispersion of the observed features, and gives the confidence interval for each particle size point accordingly, thereby achieving a quantitative assessment of the uncertainty of the particle size distribution inversion results.

[0014] In one approach, the data preprocessing and feature space construction include quality control, anomaly removal, spatial registration, spatiotemporal interpolation, and normalization of auxiliary meteorological elements on the raw radar observation data to ensure the integrity of the input feature matrix and consistency of spatiotemporal resolution.

[0015] In one approach, the multimodal feature coupling and encoding step employs a multi-branch deep neural network structure to jointly represent multimodal inputs such as radar multi-channel data and meteorological environmental features, and utilizes an attention mechanism to enhance the expressive ability of the correlation between different physical quantities.

[0016] In one approach, the output particle size distribution result and uncertainty assessment step includes estimating the range of values ​​for the particle size distribution probability density function at each particle size point, and outputting a confidence interval covering the main sources of risk based on a joint uncertainty quantification method, so that users can make subsequent meteorological decisions and conduct scientific research based on the confidence level of the results.

[0017] Beneficial effects of this invention:

[0018] This invention fully integrates the advantages of physical constraints and data-driven methods to achieve high-precision inversion of radar precipitation particle size distribution and quantification of its uncertainties, effectively improving the reliability and physical consistency of particle size spectrum inversion. By introducing multi-source observation data and advanced modeling techniques, it not only significantly enhances the adaptability of the inversion results to complex precipitation scenarios but also provides more reliable and abundant basic data support for subsequent applications such as quantitative precipitation forecasting, extreme weather warnings, and hydrometeorological analysis. Therefore, this invention has significant technological advancements and practical application value in improving the efficiency of radar precipitation particle size spectrum inversion, reducing the risk of "black box" errors, and enhancing physical interpretability. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0021] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0022] like Figure 1As shown, a method for inverting precipitation particle size distribution using dual-polarization radar is described, and the specific implementation steps are as follows:

[0023] Step 1: Radar Data Acquisition and Preliminary Preprocessing

[0024] Reflectivity was simultaneously acquired in the same spatiotemporal unit using dual-polarization (horizontal and vertical) radar. , Differential reflectivity Correlation coefficient Observational data, etc.

[0025] First, a dual-polarization Doppler weather radar was used to scan the same spatiotemporal cell of the target weather area in synchronous mode to obtain the horizontal polarization reflectivity. Vertical polarization reflectivity Differential reflectivity (Right now ), correlation coefficient And parameters such as spectral width. Let the time-series radar observation data be denoted as... ,in Let represent the feature vector at the t-th sampling time.

[0026] Raw radar data is often affected by environmental noise, instrument errors, and short-term disturbances. Therefore, an innovative hierarchical rolling filtering algorithm is used for multi-scale preprocessing. This algorithm first combines temporal characteristics to preprocess each group of continuous observation windows. Local rolling average and variance analysis are performed. For the adaptive determination of the window length L, the algorithm analyzes the real-time noise level through a data adaptive clustering module based on a pre-set minimum and maximum length range. Within each window, based on the Euclidean distance of the feature vectors, a clustering method (such as clustering based on dynamic time warping distance) is used to cluster the data into m classes. The filter value at the center of the window is denoted as...

[0027]

[0028] in This indicates that the subsets in this window that are classified into the main cluster have cluster center distances less than the adaptive threshold. .

[0029] A hierarchical rolling filtering algorithm is used to preprocess time-series radar data at multiple scales, suppressing environmental noise while preserving particle size-related microstructure information. The filtering step size and threshold parameters are optimized in real time by a data adaptive clustering module.

[0030] To achieve multi-scale information extraction, hierarchical rolling filtering will apply different window lengths. The process is repeated until a window is selected collaboratively based on the largest particle size sensitivity. Threshold parameter The adjustment of the window length L is controlled by the adaptive clustering module, and the algorithm analyzes the noise level within the current window in real time. If the noise variance increases abnormally, the window length is shortened and the clustering threshold is tightened; conversely, a larger window and looser spacing are allowed. This method effectively suppresses abrupt noise and transient disturbances while preserving microstructural changes related to precipitation particle size. The final output is a filtered dataset. This provides a reliable basis for downstream particle size distribution inversion.

[0031] Step 2: Feature Extraction for Enhanced Physical Consistency

[0032] Differential reflectivity By combining the perturbation characteristics of reflectivity with multi-scale short-time Fourier transform (MS-STFT), a set of highly correlated feature vectors is generated in the time-frequency-amplitude three-dimensional space, reflecting the physical coexistence characteristics of particles of various sizes in precipitation.

[0033] First, the differential reflectance time series after preliminary preprocessing is analyzed. A multi-scale short-time Fourier transform (MS-STFT) is performed to obtain its time-frequency local energy distribution. Specifically, let the window function be... Window length is Then, at each scale, the transformation is expressed as

[0034]

[0035] in It also contains phase information. and amplitude information The amplitude-phase features at different scales are stacked in the third dimension to form a three-dimensional tensor. Each element of this tensor comprehensively records the dynamic characteristics of precipitation particle size at a specific time, frequency, and window. Simultaneously, it records the reflectance... The window perturbation characteristics are normalized, and their mean drift and fluctuation amplitude under each time window are extracted.

[0036]

[0037] It describes its instantaneous perturbation and thus jointly maps it to the time-frequency-amplitude feature space, which significantly improves the separability of particles with coexistence and mutual interference.

[0038] For overlapping particle size signals, a learnable heterogeneous thermodynamic field distribution mapping algorithm is introduced to further fuse spatial and energy information into a weighted feature matrix, thereby improving the discriminative power of small particle size categories.

[0039] The algorithm uses the aforementioned three-dimensional tensor First, the mapping is applied to a spatial grid set, where each grid represents a region to be identified with a dense concentration of certain physical features. To assign different discrimination weights to each spatial unit, each unit is assigned a weight based on energy and spatial location. ,in

[0040]

[0041] in This represents the energy of the i-th grid feature. Its distance from the center of the main thermal characteristics, These are trainable parameters. The algorithm is dynamically optimized through end-to-end learning. This makes the characteristic distribution of small particle size categories more prominent, enhances the ability to distinguish particle size heterogeneity, and the final output weighted feature matrix not only covers the physical commonalities of multiple scales, but also maximizes the discriminative information between particle size categories, laying a solid foundation for the accurate inversion of particle size distribution in the future.

[0042] Step 3: Particle size inversion modeling based on physical constraints and prior knowledge

[0043] A Dynamic Perturbation DSD Inference Network (DP-DSDIN) is proposed: The network structure integrates the parameter embeddings of the physical model of particle size distribution, but instead of directly fitting the principal parameters, it introduces an adjustable perturbation term at the output, adaptively optimizing the weights based on physical priors and data distribution. The network input is the 3D feature matrix generated in the second step, and the output is the particle size distribution probability density function (PDF) for each spatial unit. The perturbation-consistent discrimination loss function achieves stable and refined inversion by forcing the gradient consistency between the physical principal distribution and the perturbation learned by the network.

[0044] First, the three-dimensional weighted feature matrix obtained in the second step... This serves as the input to the Dynamic Perturbation DSD Inference Network (DP-DSDIN). This network, based on a deep neural network structure, outputs the probability density function (PDF) of precipitation particle size distribution for each spatial cell. , where D represents the particle size.

[0045] To ensure the physical reliability and generalization ability of the inversion results, DP-DSDIN integrates parameter information from classical distributions such as the generalized gamma distribution in the output layer. Specifically, the network does not directly handle the main parameters of particle size distribution (such as shape parameters). Scale parameters Location parameters Instead of performing conventional regression fitting, these parameters are estimated based on physical priors. However, the final output particle size distribution is a perturbed composite expression.

[0046]

[0047] in It is the probability density of a gamma-type distribution (or other distribution) under the physical model. These are prior parameters; This is the distribution perturbation term that the neural network adaptively learns based on input features. These are network weights. In this way, the network, based on a physically reasonable particle size distribution, learns small perturbations to compensate for complex details that simplified physical models cannot capture, thus achieving a deep integration of data and physical knowledge.

[0048] To prevent explicit perturbations from deviating excessively from physical priors, a perturbation-consistent discriminative loss function is proposed. This loss term balances the physical consistency of the distribution results with numerical precision. Assume the total network loss is...

[0049]

[0050] in The loss is based on the distribution fitting of observed particle size statistics or labels, commonly using KL divergence or Wasserstein distance; while the innovative perturbation uniformity loss is defined as...

[0051]

[0052] This means that the neural network is directly forced to learn perturbations at the gradient level of the output distribution so that the particle size distribution does not violate the overall change law of physical models such as generalized gamma in the gradient trend. In this way, not only is the fitting accurate, but physical inconsistencies are also effectively limited.

[0053] in, and To adjust the hyperparameters for loss weights, The network outputs the inferred particle size distribution. For the physical model of particle size distribution, The three-dimensional feature matrix generated in step two. These are the parameters of the neural network. This is a perturbation term that is adaptively adjusted based on features.

[0054] Through the above structure, the DP-DSDIN network effectively integrates the physical priors of particle size distribution with the actual observational features. At the same time, it maintains the scientific nature and precision of the distribution fitting through an innovative loss function, significantly enhancing the generalization and inversion ability for complex precipitation particle size distributions.

[0055] Step 4: Multi-source mutual verification and adaptive correction of inversion results

[0056] Simulated radar observations with different particle size distributions (three or more hypothetical particle size distribution models) are generated and compared with the actual particle size distribution output by the network in step three to automatically verify the rationality of the inference. If there is a significant deviation, it is fed back to the DP-DSDIN network to adjust the perturbation weights, achieve adaptive correction, and optimize the overall inversion accuracy.

[0057] First, a multi-scale virtual observation consensus algorithm is introduced to automatically verify and optimize the particle size distribution inversion results output by the aforementioned DP-DSDIN network at both physical and statistical levels. Specifically, the probability density function of the particle size distribution of each spatial unit output in step three is used as the basis for this process. Based on this, three or more mainstream physical hypothesis distribution models are selected, such as the generalized gamma distribution, the exponential distribution, and the log-normal distribution, and their parameter sets are defined respectively. Using "virtual radar observation simulation," each distribution... Radar simulation operator Mapped to the corresponding multipolarized radar observation set Similarly, the inference results from the DP-DSDIN network are mapped to virtual observations. .

[0058] This simulation operator is based on the classical radar particle size reflectivity relationship. For example, for a given particle size distribution, it calculates the virtual radar horizontal reflectivity.

[0059]

[0060] in For particle size distribution, This is the correlation coefficient between wavelength and dielectric constant. Other polarization observations (such as...) , It can be obtained according to the corresponding microphysical scattering formula.

[0061] Then, the virtual observations are compared at various scales (e.g., spatial profiles, time segments, feature windows, etc.). Observation sets generated with three benchmark distributions and actual radar observations The differences. Define the consistency metric as...

[0062]

[0063] in For the consistency error of the k-th type of observations, For spatial units or time slices, and These are virtual inferred observations and actual observations, respectively. For each scale and observation type, error matrices are calculated under multiple distribution assumptions.

[0064] If a certain type of observation consistency error is detected Exceeding the preset tolerance range This indicates that there is a physical or structural anomaly in the corresponding particle size distribution inversion. At this point, the consensus algorithm automatically constructs an error signal vector. The perturbation weight adjustment module of the DP-DSDIN network is fed back to implement the following adaptive correction mechanism: the perturbation weight factor is adjusted... The goal of differential fine-tuning is to minimize the weighted sum of multi-scale consistency errors, i.e.

[0065]

[0066] in The importance weights for various observations are assigned. Based on gradient descent or weighted backpropagation optimization principles, the network dynamically adjusts the ratio of the disturbance term to the physical principal distribution, thereby gradually correcting the distribution inversion and making the new round of virtual observations more consistent with the actual measurements and the reasonable physical range.

[0067] In this way, the particle size inversion was reconstrained, and the network's adaptive immunity and evolutionary ability to physical and observational anomalies was realized through an intelligent feedback correction mechanism, which greatly improved the overall inversion accuracy and robustness.

[0068] Step 5: Output particle size distribution results and uncertainty assessment

[0069] The algorithm ultimately outputs the particle size distribution probability density function for each spatial unit. At the same time, an innovative “data-physical joint uncertainty quantification method” is used to systematically evaluate the uncertainty of the inversion results.

[0070] Specifically, this method comprehensively quantifies the confidence range of the inverted particle size distribution from three key dimensions: first, the sample variance within the model output distribution—specifically, the particle size distribution sample set obtained by the neural network through multiple sampling, parameter perturbation, or Bayesian approximation (such as Monte Carlo dropout, weight perturbation, etc.). Calculate its mean distribution and variance

[0071]

[0072] This reflects the uncertainty in the model's output.

[0073] Second, the perturbation residuals based on physical consistency, which are the final estimated distributions. With physical main distribution The consistency error serves as a quantitative supplement to the risk of physical bias. Specifically, it can be expressed as the perturbation consistency residual.

[0074]

[0075] The total residual is obtained by integrating and normalizing over the particle size range. It is used to characterize the degree of deviation between the model and the physical prior.

[0076] Third, at the macroscopic observation characteristics level, the distribution dispersion between network simulation and actual radar observation results is compared. This is quantified using the variance or mean square error of the differences between multi-scale virtual observations (see step four), for example...

[0077]

[0078] in and These represent the k-th type of observation in the first place. Simulated and measured values ​​of spatial units, where k represents the observation type.

[0079] The above three sources of uncertainty are standardized and weighted to form a total uncertainty measure.

[0080]

[0081] in Weights that can be learned or set based on experience. This joint index targets particle size. Reliability assessments are provided point by point.

[0082] Finally, based on The confidence interval quantile method is used to determine the upper and lower confidence bounds for each particle size point. For example, for confidence levels... Output range

[0083]

[0084] in The corresponding standard scores are used. In this way, not only are the inferred values ​​of particle size distribution quantitatively given, but the reliable bandwidth and potential sources of risk at each particle size are also fully revealed, so that the final particle size distribution inversion results have a scientific, interpretable and engineering-usable confidence description.

[0085] Example:

[0086] I. Experimental Background and Data Acquisition

[0087] This embodiment assumes that the observation focuses on a regional rainstorm event in a certain city on June 12, 2024, and selects data from the local C-band dual-polarization radar (station number JS001) and surrounding meteorological stations. The observation period is 06:00–07:00, and volume scan radar data and ground meteorological elements (temperature, humidity, wind speed, etc.) are collected synchronously every 10 minutes.

[0088] Table 1. Original observation data:

[0089]

[0090] II. Step 1: Data Preprocessing and Feature Space Construction

[0091] This step focuses on improving data quality and feature alignment:

[0092] 1. Quality control and outlier removal

[0093] Examine all radar observation data, setting the effective range of reflectivity ZH (0–55 dBZ), differential reflectivity ZDR (-0.5–3.0 dB), and correlation coefficient ρhv > 0.93. Remove abnormal data such as clutter, sidelobes, and strong attenuation to ensure physical plausibility.

[0094] 2. Spatial registration and interpolation

[0095] Radar data from different elevation angles and sampling distances are uniformly projected onto a 1km×1km grid, and missing measurement points are reconstructed through three-dimensional interpolation to form a spatially consistent feature field.

[0096] 3. Normalization of meteorological elements

[0097] z-score (standard score) normalization is performed on parameters such as temperature, humidity, and wind speed from ground meteorological stations to eliminate dimensional differences and improve the effectiveness of model fusion.

[0098] 4. Feature Matrix Construction

[0099] After the above processing, a multi-dimensional feature matrix including radar reflectivity, dual polarization parameters, and meteorological elements is constructed as input for subsequent modeling.

[0100] Table 2. Processed Feature Data

[0101]

[0102] III. Step 2: Multimodal Feature Coupling and Encoding

[0103] 1. Multi-branch deep network architecture design

[0104] A three-branch neural network is established to process radar reflectivity (ZH), dual polarization parameters (ZDR, ρhv), and meteorological environmental characteristics (normalized temperature, humidity, and wind speed), respectively.

[0105] 2. Feature fusion and attention mechanisms

[0106] The outputs of each branch are integrated by a feature fusion layer, and the self-attention mechanism is used to highlight physical quantities that are sensitive to particle size (such as high ZH and large ZDR range).

[0107] 3. Coding and Standardization

[0108] All features are encoded and standardized to become the input of the particle size inversion network, taking into account the spatial and physical consistency of multi-source information.

[0109] IV. Step 3: Particle size inversion modeling based on physical constraints and prior knowledge

[0110] 1. Integration of physical prior parameters

[0111] The main particle size distribution adopts a generalized gamma distribution, and the physical parameters α (shape), β (scale), and λ (offset) are predicted end-to-end by the neural network. These parameters are embedded in the network structure as physical constraints on the particle size distribution.

[0112] 2. Introduction of dynamic disturbance terms

[0113] Add a disturbance branch to output the disturbance term. (D). In the network loss function, the "perturbation-consistent discrimination loss" is adopted, which forces the gradients of the physical principal distribution and the perturbation distribution to be consistent, thereby improving the physical interpretability and stability of the inversion.

[0114] 3. Inversion Modeling Process

[0115] The network input is encoded and outputs a particle size probability density N(D). The loss function integrates MSE (mean squared error), physical consistency loss, and perturbation discrimination loss to optimize inversion accuracy and consistency through multi-objective optimization.

[0116] Table 3. Output of particle size distribution parameters:

[0117]

[0118] V. Step 4: Multi-source mutual verification and adaptive correction of inversion results

[0119] 1. Virtual Observation Consistency Algorithm

[0120] Based on the inverted particle size distribution (N(D)), various theoretical particle size assumptions (generalized gamma, log-normal, etc.) are used to back-calculate simulated radar observations (ZHsim, ZDRsim).

[0121] 2. Error Calculation and Adaptive Feedback

[0122] If the consistency error exceeds the 2σ threshold when comparing simulated radar measurements with actual observations, the neural network perturbation weights are adaptively adjusted in a timely manner to enhance the fit of the distribution inversion.

[0123] 3. Correction Mechanism

[0124] By using observation-simulation error signals as perturbation branch backpropagation training, the network can automatically correct the inversion effect and continuously approximate the physical measurement.

[0125] VI. Step 5: Particle size distribution output and uncertainty quantification

[0126] 1. Particle size distribution probability density output

[0127] Output the particle size distribution N(D) for each grid point and give the probability density function values ​​at 0.2 mm intervals.

[0128] 2. Joint Uncertainty Quantification

[0129] The statistical model outputs the sample variance σ, the physical distribution consistency residual R, and the dispersion of the observed characteristic distribution V. A comprehensive uncertainty evaluation index UQ is constructed, which outputs a confidence interval for each particle size point, achieving highly reliable particle size inversion.

[0130] Table 4 Output Example

[0131]

[0132] This embodiment, based on high-quality dual-polarization radar data and multi-source meteorological elements, utilizes multi-modal fusion, physical constraints, perturbation consistency mechanisms, and multi-source virtual observation cross-verification to achieve high-precision inversion of precipitation particle size distribution and quantification of uncertainty. This method not only overcomes the limitations of traditional empirical formulas and improves the self-consistency and physical reliability of the inversion, but also provides solid scientific evidence and input data support for fields such as quantitative precipitation estimation, meteorological disaster forecasting, and hydrological runoff simulation.

[0133] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0134] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for inverting precipitation particle size distribution using dual-polarization radar, characterized in that: The method includes the following: Step 1: Data preprocessing and feature space construction. Using the original observation data from multi-source radar and auxiliary meteorological parameters, data cleaning and quality control are performed, and a three-dimensional feature matrix related to particle size inversion is obtained through spatial registration and spatiotemporal interpolation methods. Step 2: Multimodal feature coupling and encoding. Using a multi-network structure and attention mechanism, features from different sources are fused to obtain a high-dimensional feature representation that can characterize the evolution of particle size structure, and then standardized encoding is performed. Step 3: Based on physical constraints and prior-driven particle size inversion modeling, the particle size distribution inference network is dynamically perturbed, the parameters of the particle size distribution physical model are integrated, an adjustable perturbation term is introduced at the network output, and the perturbation consistency discrimination loss function is combined to improve the particle size distribution inversion accuracy and physical consistency. Step 4: Multi-source mutual verification and adaptive correction of inversion results. A multi-scale virtual observation consistency algorithm is adopted, which uses various theoretical particle size distribution assumptions to simulate and generate radar observations. The results are compared and verified with the output of the neural network, and the weights of the network disturbance terms are adaptively optimized and fed back. Step 5: Output particle size distribution results and uncertainty assessment. Based on the data-physical joint uncertainty quantification method, the variance of the output distribution of the integrated model, the perturbation consistency residual and the dispersion of the observed characteristic distribution are used to give the confidence interval of the inverted particle size distribution results and perform uncertainty assessment.

2. The method for inverting precipitation particle size distribution using dual-polarization radar according to claim 1, characterized in that: The dynamic perturbation particle size distribution inference network uses generalized gamma distribution parameters as physical priors in the process of outputting the particle size distribution probability density function. By introducing a perturbation consistent discrimination loss function, it forces the gradient changes between the physical main distribution and the perturbation term learned by the network to be consistent, so as to achieve high accuracy and stability of particle size distribution inversion.

3. The method for inverting precipitation particle size distribution using dual-polarization radar according to claim 1, characterized in that: The multi-scale virtual observation consistency algorithm compares the consistency error between simulated radar observations and actual observations or observations under multiple distribution assumptions, and adaptively adjusts the perturbation weights of the dynamic perturbation particle size distribution inference network based on the error signal, thereby achieving automatic correction and accuracy improvement of distribution inversion results.

4. The method for inverting precipitation particle size distribution using dual-polarization radar according to claim 1, characterized in that: The data-physics joint uncertainty quantification method constructs a joint index by integrating the sample variance output by the integrated model, the consistency residual of the physical principal distribution, and the distribution dispersion of the observed features, and gives the confidence interval for each particle size point accordingly, thereby realizing the quantitative assessment of the uncertainty of the particle size distribution inversion results.

5. The method for inverting precipitation particle size distribution using dual-polarization radar according to claim 1, characterized in that: The data preprocessing and feature space construction include quality control, anomaly removal, spatial registration, spatiotemporal interpolation, and normalization of auxiliary meteorological elements on the raw radar observation data to ensure the integrity of the input feature matrix and the consistency of spatiotemporal resolution.

6. The method for inverting precipitation particle size distribution using dual-polarization radar according to claim 1, characterized in that: The multimodal feature coupling and encoding step adopts a multi-branch deep neural network structure to jointly represent radar multi-channel data and meteorological environmental feature multimodal inputs, and uses an attention mechanism to improve the ability to express the correlation between different physical quantities.

7. The method for inverting precipitation particle size distribution using dual-polarization radar according to claim 1, characterized in that: The output particle size distribution results and uncertainty assessment steps include estimating the range of values ​​of the particle size distribution probability density function at each particle size point, and using a data-physical joint uncertainty quantification method to output a confidence interval covering the main sources of risk.

Images