A meteorological radar classification and identification method combining meteorological numerical models

Abstract

This invention discloses a meteorological radar classification and identification method combining meteorological numerical models, relating to the field of intelligent meteorological sensing technology. The method includes: generating a unified initial state vector using statistical information from historical typhoon similar samples within a preset calibration area to achieve cold-start calibration under radar data-free conditions; subsequently, allocating weights between the radar extrapolation field within two hours and the numerical model field after two hours based on a time-gating function, calculating the observation residuals and observation error covariance matrices respectively, and using Kalman-type updates to jointly correct the amplitude and deformation components; obtaining a fused precipitation field through large-scale extrapolation of the model CLDAS field, spatial resolution refinement, terrain factor correction, and high-precision fusion of the radar extrapolation field, and generating a classification map based on an adaptive quantile threshold to complete iterative updates; this method achieves stability of precipitation field initialization under radar-free conditions by generating a unified initial state vector from statistical information from historical typhoon similar samples.

Description

Technical Field

[0001] This invention relates to the field of intelligent meteorological sensing technology, and in particular to a meteorological radar hierarchical identification method that combines meteorological numerical models. Background Technology

[0002] With the continuous development of radar observation technology, numerical weather prediction models, and data assimilation methods, the spatiotemporal resolution and real-time performance of meteorological precipitation forecasts have been significantly improved. Currently, meteorological numerical models can provide the evolution trend of precipitation fields over the next few hours to days on a large spatial scale, while radar observations can achieve fine monitoring of precipitation intensity and structure information at the minute level.

[0003] However, existing methods still have two limitations: First, existing multi-source fusion methods fail to distinguish the dominant role of radar extrapolation data within two hours and numerical model data after two hours in the time dimension, lack a dynamic weight switching mechanism, and find it difficult to maintain the continuity of prediction accuracy; Second, existing methods rely on a single model field or simple interpolation to construct initial values ​​under conditions without radar observation, and lack the ability to generate a unified initial state vector by using similar samples of historical typhoons and calibration of multiple trajectories, making it difficult to guarantee the stability and physical consistency of the initial value field. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a meteorological radar classification and identification method that combines meteorological numerical models to solve the problem that existing fusion methods of meteorological numerical models and radar extrapolation lack a dynamic adjustment mechanism for the weight allocation of short-term nowcasts and mesoscale forecasts.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a meteorological radar classification and identification method combining meteorological numerical models, which includes,

[0008] Collect multi-source meteorological data, establish a unified time axis and spatial grid, statistically analyze the information of preset calibration areas and historical typhoon samples, and obtain multi-source datasets;

[0009] By using historical typhoon sample information, a unified initial state vector is obtained in a preset calibration area, and a unified state vector is obtained by calibrating the second route trajectory anchor point.

[0010] The large-scale fields of meteorological numerical models in multi-source datasets are extrapolated and corrected using a unified state vector, and the background field is obtained through a one-time consistency constraint.

[0011] The background field is spatially refined, local corrections are made by combining terrain factors, and mass conservation and edge smoothing are performed to obtain a high-resolution refined field.

[0012] Extrapolate the radar observation field from the multi-source dataset to obtain the radar extrapolated field. Then, fuse the high-resolution refined field with the radar extrapolated field through time gating to obtain the fused precipitation field.

[0013] Obtain a single composite classification index of the merged precipitation field, classify it according to the quantile threshold, and obtain a classification map;

[0014] By fusing precipitation field and graded map, the unified state vector is jointly corrected for amplitude and deformation based on Kalman-type update to obtain the updated unified state vector.

[0015] As a preferred embodiment of the meteorological radar hierarchical identification method combining meteorological numerical models described in this invention, the steps of collecting multi-source meteorological data, establishing a unified time axis and spatial grid, statistically analyzing preset calibration area and historical typhoon sample information, and obtaining multi-source datasets and regional information are as follows.

[0016] Collect meteorological numerical model data, radar observation data and ground observation data, and perform time and space unification processing to establish a unified time axis and spatial grid.

[0017] Determine the preset calibration area and historical typhoon sample information, and obtain standardized multi-source datasets and regional information for initialization.

[0018] As a preferred embodiment of the meteorological radar hierarchical identification method combining meteorological numerical models described in this invention, the steps of obtaining a unified initial state vector in a preset calibration area using historical typhoon sample information, and obtaining a deformation-corrected unified state vector through second route trajectory anchor point calibration, are as follows:

[0019] Based on the historical typhoon sample information, a unified initial state vector consisting of amplitude components and deformation components is constructed within a preset calibration area.

[0020] The unified initial state vector is calibrated using the second route trajectory anchor point to obtain a deformation-corrected unified state vector.

[0021] As a preferred embodiment of the meteorological radar hierarchical identification method combining meteorological numerical models described in this invention, the steps of extrapolating and correcting the large-scale field of the meteorological numerical models in the multi-source dataset using a unified state vector, and obtaining the background field through a one-time consistency constraint, are as follows:

[0022] Spatial deformation alignment and amplitude bias correction are performed on the large-scale fields of meteorological numerical models in multi-source datasets using a unified state vector, and the background field is obtained through a one-time consistency constraint.

[0023] As a preferred embodiment of the meteorological radar hierarchical identification method combining meteorological numerical models described in this invention, the steps of refining the spatial resolution of the background field, performing local corrections based on terrain factors, and implementing mass conservation and edge-preserving smoothing processing to obtain a high-resolution refined field are as follows.

[0024] The spatial resolution of the background field is refined to obtain a preliminary refined field.

[0025] Based on the preliminary refined field combined with topographic factors, local feature corrections are performed to form a topographically corrected refined field;

[0026] Mass conservation constraints are applied to the terrain-corrected refined field, and edge smoothing is performed to obtain the final high-resolution refined field.

[0027] As a preferred embodiment of the meteorological radar hierarchical identification method combining meteorological numerical models described in this invention, the specific steps for extrapolating the radar observation fields from the multi-source dataset to obtain the extrapolated radar field are as follows:

[0028] Radar observation fields are extracted from multi-source datasets, and quality control and dimension unification are performed to obtain radar reflectivity, which is then converted into an equivalent precipitation intensity field.

[0029] Based on the equivalent precipitation intensity field of adjacent time intervals, the motion vector field is estimated using the optical flow method to obtain the moving speed and direction information of precipitation echoes.

[0030] By extrapolating the equivalent precipitation intensity field at the current moment through the motion vector field, the radar extrapolation field at the future moment can be obtained.

[0031] As a preferred embodiment of the meteorological radar hierarchical identification method combining meteorological numerical models described in this invention, the step of fusing the high-resolution refined field with the radar extrapolation field through time-gated inverse variance weighting to obtain the fused precipitation field includes the following specific steps.

[0032] Based on the preset time gating function, the error variances of the high-resolution refined field and the radar extrapolation field under different forecast lead times are calculated respectively, and the mode weights and radar weights that change with time are obtained.

[0033] By applying the inverse variance weighting principle, the mode weights and radar weights are normalized to generate a fusion coefficient field.

[0034] Based on the fusion coefficient field, the high-resolution refined field and the radar extrapolation field are weighted and fused at each grid point to obtain the fused precipitation field.

[0035] As a preferred embodiment of the meteorological radar classification and identification method combining meteorological numerical models described in this invention, the specific steps for obtaining the single composite classification index of the fused precipitation field are as follows:

[0036] Precipitation intensity is calculated point by point on a unified spatial grid and a unified time axis by merging precipitation fields, and the precipitation intensity field is output.

[0037] Based on the intensity threshold, connected components are extracted and the area and shape index of the connected components are calculated to obtain the spatial structure feature field.

[0038] The trend feature field is obtained by calculating the short-term rate of change on a unified time axis, and a single composite classification index is obtained by combining the precipitation intensity field, spatial structure feature field, and trend feature field.

[0039] As a preferred embodiment of the meteorological radar classification and identification method combining meteorological numerical models described in this invention, the specific steps for classifying and obtaining a classification map by using quantile thresholds are as follows:

[0040] The empirical distribution is calculated for each time segment by a single composite grading index, and the value of the single composite grading index at the corresponding quantile position is read at the target quantile position to generate the quantile threshold.

[0041] The classification level of each grid point is determined by comparing a single composite classification index with a quantile threshold, and the classification levels of all grid points are combined in a unified spatial grid to form an initial classification map.

[0042] The hierarchical graph is obtained by performing eight-neighbor connectivity cleanup on the initial hierarchical graph.

[0043] As a preferred embodiment of the meteorological radar classification and identification method combining meteorological numerical models described in this invention, the following steps are taken: By fusing precipitation fields and classification maps, and performing joint amplitude and deformation correction on the unified state vector based on Kalman spectroscopy to obtain the updated unified state vector.

[0044] Based on the fused precipitation field and graded map, the observation residuals and uncertainties are calculated, and the observation vector and observation error covariance matrix are constructed.

[0045] The predicted state vector is generated by extrapolating the observation vector, the observation error covariance matrix and the unified state vector updated in the previous round through large-scale field. The Kalman-type update method is used to jointly correct the magnitude component and deformation component in the unified state vector to obtain the updated unified state vector.

[0046] The fused precipitation field and classification map are dynamically corrected based on the updated unified state vector to obtain the final classification recognition result.

[0047] The beneficial effects of this invention are as follows: by generating a unified initial state vector based on statistical analysis of similar historical typhoon samples, stability of the precipitation field initialization without radar is achieved; by using time-division weighting based on time-gated functions, the radar extrapolation field and the numerical model field are dynamically weighted and fused according to the inverse of the normalized error variance, achieving the optimal combination of information at different time scales and significantly reducing short-term forecast errors; and by using model CLDAS field extrapolation, spatial resolution refinement, and radar extrapolation field fusion combined with Kalman update, high-precision forecasts of precipitation field intensity, structure, and trend are achieved. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a meteorological radar classification and identification method that combines meteorological numerical models.

[0050] Figure 2 This is a flowchart for round-by-round fusion.

[0051] Figure 3 This is a flowchart of time gating and weight calculation.

[0052] Figure 4 The flowchart for the Kalman-type update closed loop. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0056] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a meteorological radar classification identification method combining meteorological numerical models, including the following steps:

[0057] S1. Collect multi-source meteorological data, establish a unified time axis and spatial grid, statistically analyze the information of similar samples of the preset calibration area and historical typhoons, and obtain multi-source datasets and regional information.

[0058] Collect meteorological numerical model data, radar observation data, and ground observation data, perform time and space unification processing on the data, and establish a unified time axis and spatial grid.

[0059] Furthermore, meteorological numerical model data is acquired periodically through the data interface of the meteorological forecasting center, radar observation data is collected through the standard data interface of the weather radar observation network, and ground observation data is collected through the data transmission links of automatic weather stations, rain gauges, and manual observation stations. The meteorological numerical model data, radar observation data, and ground observation data are timestamped to extract the time identifiers from each data source. The time intersection of the meteorological numerical model data, radar observation data, and ground observation data containing time identifiers is calculated to generate a unified time series. The time resolution is set to a fixed time (e.g., 10 minutes), and time interpolation is performed on the unified time series to obtain time-aligned meteorological numerical model data, radar observation data, and ground observation data. The spatial location information of the time-aligned meteorological numerical model data, radar observation data, and ground observation data is extracted, and the spatial range intersection of all data is calculated to generate a unified spatial grid. The time-aligned meteorological numerical model data, radar observation data, and ground observation data are spatially interpolated on the unified spatial grid to generate meteorological numerical model data, radar observation data, and ground observation data that are both time-aligned and spatially aligned. The time-aligned and spatially aligned data are uniformly mapped to obtain a unified time axis and spatial grid.

[0060] Determine the preset calibration area and statistical information of similar samples from historical typhoons, and obtain standardized multi-source datasets and regional information for initialization.

[0061] Furthermore, based on the collected meteorological numerical model data, radar observation data, and ground observation data, the latitude and longitude coordinates of all data points are extracted. The minimum and maximum latitude and longitude values ​​are combined to form a rectangular area that can cover all data points as the spatial boundary information of the observation range. By combining the spatial boundary information of the observation range with the historical typhoon path distribution, the historical typhoon path distribution is superimposed within the spatial boundary information of the observation range. Based on the area where the path coverage probability is greater than the typhoon path coverage probability threshold (e.g., 0.7), a preset calibration area is determined. Meteorological numerical model data, radar observation data, and ground observation data are extracted within the preset calibration area to obtain a multi-source dataset within the preset calibration area. Based on the multi-source dataset within the preset calibration area and the historical typhoon path distribution, the similarity between the historical typhoon path and the current typhoon path is calculated using Euclidean distance to obtain the statistical information of historical typhoon similarity samples. The expression is:

[0062] ;

[0063] ;

[0064] in, Indicates the first The similarity between a sample of historical typhoons and the current typhoon path. Indicates the first Euclidean distance between a sample of historical typhoons and the current typhoon path. , Indicates the current typhoon path at the [number]th [location]. Latitude and longitude at each point in time , Indicates the first The historical typhoon path in the first Latitude and longitude at each point in time Indicates the number of time points.

[0065] It should be noted that the typhoon path coverage probability threshold is set to 0.7 because in the spatial coverage statistics of historical typhoon path samples, areas with a coverage probability higher than 0.7 often correspond to the main path zone where the typhoon center has a higher probability of passing through. This can ensure the representativeness of the sample while avoiding over-expansion into the marginal areas where the probability of typhoon impact is low, thus taking into account both the stability and physical rationality of the calibration area.

[0066] The multi-source dataset within the preset calibration area is time-aligned and spatially interpolated to obtain a standardized multi-source dataset. Combined with the preset calibration area, a standardized multi-source dataset and regional information for initialization are generated.

[0067] S2. By using statistical information of similar samples from historical typhoons, a unified initial state vector is obtained in the preset calibration area, and a deformation-corrected unified state vector is obtained by calibrating the second route trajectory anchor point.

[0068] Based on the statistical information of similar historical typhoon samples, a unified initial state vector consisting of amplitude deviation and deformation field is constructed within a preset calibration area.

[0069] Based on historical typhoon similarity sample statistical information and a preset calibration region, the amplitude deviation statistics and deformation field statistics are extracted from the historical typhoon similarity sample statistical information and restricted to the preset calibration region to obtain prior statistical parameters, including the prior mean vector and the prior covariance matrix. Within the preset calibration region, Cholesky decomposition is performed on the prior covariance matrix, and the initial values ​​of amplitude deviation and deformation field are obtained by linear transformation of the standard normal random vector. The initial values ​​of amplitude deviation and deformation field satisfy physical range constraints at each grid point (e.g., non-negative precipitation intensity and displacement not exceeding a certain limit). For example, (200 km), based on the statistical information of historical typhoon similar samples, for each historical typhoon path sequence, the latitude and longitude points of the typhoon center are extracted as path anchor points at fixed intervals on a unified time axis. The latitude and longitude points of the typhoon center corresponding to all time points are arranged in chronological order to form a path anchor point sequence. Through the initial values ​​of amplitude deviation and deformation field, and the path anchor point sequence in the statistical information of historical typhoon similar samples, combined with meteorological observation data collected from ground observation stations and meteorological satellites, and the numerical model analysis field obtained based on numerical models (such as CLDAS), the path anchor points are extracted at fixed intervals. The latitude and longitude coordinates of the current typhoon center are taken and arranged in chronological order to form the current typhoon path sequence. Within a preset calibration area, the hourly time points of the current typhoon path sequence are matched one-to-one with the path anchor point sequences in the statistical information of similar historical typhoon samples according to a unified time axis. The latitude and longitude of the current typhoon path sequence and the path anchor point sequence at the same time point are uniformly projected into equidistant plane coordinates. The east-west coordinate difference and the north-south coordinate difference are calculated and combined to form a two-dimensional error vector. The Euclidean length of the two-dimensional error vector is taken as the path anchor point error. The path anchor point error is used to construct a path... The Gaussian radial basis function matrix centered at the anchor point is used to solve for the radial basis function coefficient vector using the regularized least squares method. The initial value correction of the deformation field is calculated at all grid points by weighting the elements of the Gaussian radial basis function matrix and the radial basis function coefficient vector. The initial value of the amplitude deviation and the initial value correction of the deformation field are used to form the amplitude component and the deformation component in the preset calibration area. The amplitude component is arranged in the order of the grid points, then the east-west component of the deformation component is arranged in the order, and finally the north-north component of the deformation component is arranged in the order. The three parts are spliced ​​together in sequence to obtain a unified initial state vector.

[0070] The unified initial state vector is calibrated using the second route trajectory anchor point to obtain a deformation-corrected unified state vector.

[0071] Furthermore, time registration pairs are established between the latitude and longitude of the typhoon center at the corresponding time point and the current typhoon center, based on the statistical information of similar historical typhoon samples. All time registration pairs are then arranged in chronological order to form a second route trajectory anchor point sequence. By unifying the initial state vector and the second route trajectory anchor point sequence, the second route trajectory anchor point error is calculated. A radial basis function matrix is ​​constructed within a preset calibration region using the unified initial state vector and the second route trajectory anchor point error, and the weight vector is solved using a regularized least squares method. The expression is:

[0072] ;

[0073] ;

[0074] in, Indicates the first The coordinates of each grid point Indicates the first Path anchor point coordinates Indicates the scale parameter. Represents the radial basis function matrix. Represents the weight vector. This represents a column vector formed by stacking the anchor point errors of the second route trajectory in chronological order. This represents the regularization coefficient.

[0075] It should be noted that, This involves calculating the set of Euclidean distances between all pairwise path anchor points within a predefined calibration area, and then multiplying the sample percentile of the Euclidean distance set by the empirical coefficient. , The noise variance is estimated by calculating the variance of the path anchor point error vector at all time points. Then, K-fold cross-validation is performed on one hundred candidate regularization coefficients to minimize the validation set mean square error. Finally, the regularization coefficient that minimizes the validation set mean square error is selected as the regularization coefficient. , The value range is limited to the product interval of the noise variance. , ], To prevent overfitting and numerical instability, It is to avoid underfitting caused by overregulation. This represents the variance of the path anchor point error vector at all time points.

[0076] The initial value correction of the deformation field and the deformation vector in the unified initial state vector are obtained at each grid point by using the weight vector. The unified initial state vector after deformation is obtained, and the unified state vector after deformation correction is obtained by applying continuity constraints and boundary smoothing within the preset calibration area.

[0077] S3. Extrapolate and correct the large-scale field of meteorological numerical models in multi-source datasets using a unified state vector, and obtain the background field through a one-time consistency constraint.

[0078] Spatial deformation alignment and amplitude bias correction are performed on the large-scale fields of meteorological numerical models in multi-source datasets using a unified state vector, and the background field after large-scale correction is obtained through a one-time consistency constraint.

[0079] Furthermore, by performing spatial coordinate transformation on the large-scale fields of meteorological numerical models in the multi-source dataset using the deformation component in the unified state vector, a spatially deformed aligned large-scale field is obtained. Then, by superimposing the spatially deformed aligned large-scale field grid-by-grid using the amplitude component in the unified state vector, a large-scale field corrected for amplitude bias is obtained. Finally, a one-time Helmholtz constraint with terrain weights is applied to the large-scale field corrected for amplitude bias within a preset calibration area to obtain the background field, expressed as:

[0080] ;

[0081] ;

[0082] ;

[0083] in, Represents spatial location coordinates, This represents the large-scale corrected background field defined within the preset calibration area. This represents the large-scale field defined within the preset calibration region, after spatial deformation alignment and amplitude deviation correction. Indicates the smoothness intensity coefficient. Indicates terrain weight, Represents constant-type Lagrange multipliers. This represents the total large-scale field of meteorological numerical models in the multi-source dataset before deformation within the preset calibration area. Represents the identity operator, Indicates the zero normal flux boundary condition. This represents the spatial gradient operator.

[0084] It should be noted that, The method involves calculating the spatial gradient variance of the large-scale field after spatial deformation alignment and amplitude deviation correction within a preset calibration region, and performing cross-validation among multiple candidate smoothing intensity coefficients to minimize the mean square error between the smoothed field and the observed field. The smoothing intensity coefficient with the smallest mean square error is selected as the value of μ. By using coordinates at each location on a uniform spatial grid The terrain elevation data is read, the terrain gradient is calculated, and the result is normalized to the [0,1] interval. Within the preset calibration region, the intermediate field without constant terms is first solved to satisfy... Then the integral of the intermediate field over the preset calibration region is combined with the given total amount. Dividing the difference by the area of ​​the preset calibration region yields the constant-type Lagrange multiplier. It is an identity operator obtained by constructing an identity matrix within a preset calibration region, preserving the values ​​of all grid points as one, and multiplying it by any field to output an equal value to the original field. Indicates the boundary of the preset calibration area The background field, after large-scale correction, has no flux leakage in the normal direction. It uses the spatial coordinates and grid resolution of a regular grid as input within a preset calibration area, and constructs a pair at each grid point according to the finite difference method. The partial derivative operators in three directions are stacked in columns to form a unified spatial gradient operator.

[0085] S4. The large-scale corrected background field is spatially refined, local correction is performed in combination with terrain factors, and mass conservation and edge smoothing processing are performed to obtain a high-resolution refined field.

[0086] The spatial resolution of the large-scale corrected background field is refined, and local features are corrected by combining topographic factors to obtain a refined field after topographic correction.

[0087] Furthermore, within the preset calibration area, the coarse-resolution grid and high-resolution grid of the large-scale calibration background field are uniformly projected onto an equidistant plane. The overlap area between each high-resolution grid and the coarse-resolution grid is calculated. The normalized ratio of the overlap area to the high-resolution grid area is used as the spatial interpolation weighting coefficient matrix of the large-scale calibration background field on the high-resolution grid. The interpolation weighting coefficient matrix is ​​constructed using the bilinear interpolation method, with each interpolation weighting coefficient limited to between zero and one, and the sum of the interpolation weighting coefficients of four adjacent grid points equal to one. Bilinear interpolation is performed on the large-scale calibration background field grid by grid point using the spatial interpolation weighting coefficient matrix, increasing the spatial resolution of the large-scale calibration background field from ten kilometers to one kilometer, resulting in the spatially interpolated large-scale calibration background field. Topographic elevation data is extracted from publicly available geographic information data sources at a spatial resolution of one kilometer, and the topographic elevation data is interpolated onto the same grid as the large-scale field of the meteorological numerical model. Finite difference operations are performed on the east-west and north-south directions respectively to obtain the east-west topographic slope field and the north-south topographic slope field. The terrain slope field is calculated, and the square root of the sum of the squares of the slopes in both directions is used as the terrain gradient field at each grid point. The mean and standard deviation of the terrain gradient field are calculated at all grid points. Grid points with terrain gradient fields greater than the mean plus the standard deviation are marked as areas with significant terrain undulation, and the remaining grid points are marked as areas with gentle terrain. A mask for areas with significant terrain undulation is obtained. In areas with significant terrain undulation, a local feature adjustment coefficient is calculated using the normalized terrain gradient field. The local feature adjustment coefficient is obtained by linearly mapping the difference between 1 and the normalized terrain gradient value to an interval of, for example, 0.7 to 1. Limiting the local feature adjustment coefficient to between 0.7 and 1 is to weaken the large-scale corrected background field value after spatial interpolation within an amplitude of no more than 30%, thereby avoiding physical quantity distortion and maintaining numerical stability. The large-scale corrected background field after spatial interpolation is multiplicatively scaled grid-by-grid in areas with significant terrain undulation using the local feature adjustment coefficient. In areas with gentle terrain, the large-scale corrected background field after spatial interpolation is directly used to obtain a refined field after terrain correction.

[0088] Within each coarse grid, mass conservation constraints are applied to the terrain-corrected refined field, and edge smoothing is performed to obtain the final high-resolution refined field.

[0089] Furthermore, all the terrain-corrected refined field values ​​belonging to the coarse grid are weighted and integrated according to the corresponding high-resolution grid area to obtain the spatial total of the corresponding coarse grid. The product of the terrain-corrected refined field and the time step, where the terrain-corrected refined field represents precipitation intensity, is converted into the total precipitation. The total amount of the terrain-corrected refined field is composed of the spatial total and the total precipitation. The total amount of the large-scale corrected background field on the same coarse grid is recorded as the total amount of the large-scale corrected background field. The ratio of the total amount of the large-scale corrected background field to the total amount of the terrain-corrected refined field is obtained within each coarse grid. A small constant (e.g., ...) is added when the total amount of the terrain-corrected refined field approaches zero. Furthermore, the ratio of the total large-scale corrected background field to the total topographically corrected refined field is limited to, for example, a range of 0.5 to 1 as the mass conservation scaling factor. If the mass conservation scaling factor is less than 0.5, it may severely weaken the original local precipitation or wind field characteristics. If the mass conservation scaling factor is greater than 1, it may amplify the noise in the observation or model, producing non-physical extrema. Within each coarse grid, the product of the topographically corrected refined field at each grid point and the mass conservation scaling factor is obtained to ensure that the total amount of each coarse grid is consistent with the total amount of the large-scale corrected background field in that coarse grid. In the neighborhood of each coarse grid boundary, an edge-preserving smoothing method is used to limit the numerical difference between adjacent grid points to no more than five millimeters to avoid discontinuities at the boundary. The limit of no more than five millimeters is a reasonable upper limit determined based on historical observation resolution and numerical model stability experience to avoid non-physical abrupt changes at the boundary, thus obtaining the final high-resolution refined field.

[0090] S5. Extrapolate the radar observation field from the multi-source dataset to obtain the radar extrapolated field. Then, fuse the high-resolution refined field with the radar extrapolated field by time-gated inverse variance weighting to obtain the fused precipitation field.

[0091] Radar observation fields are extracted from multi-source datasets, and quality control and dimensional unification are performed to convert radar reflectivity into an equivalent precipitation intensity field.

[0092] Furthermore, radar observation data was acquired through multi-source datasets. Clutter removal, ground object obstruction correction, and weak echo filtering were performed on the radar observation data at each scanning elevation angle to obtain quality-controlled radar observation data. The radar reflectivity values ​​at all time points were converted to a unified unit (dBZ), and the observation data from different radar scanning times were resampled to a unified time interval (e.g., every five minutes) using a time interpolation method. Simultaneously, a weighted average method was used to generate grid values ​​with a unified resolution for overlapping areas of different radars in the spatial dimension, obtaining radar observation data that is completely consistent in both dimensions and temporal resolution. Using the empirical relationship between radar reflectivity and precipitation intensity, the radar reflectivity was converted into an equivalent precipitation intensity field, expressed as:

[0093] ;

[0094] ;

[0095] in, Indicates linear reflectivity. Indicates precipitation intensity. and This represents the empirical coefficient.

[0096] It should be noted that, It is the linear reflectivity converted from radar reflectivity in dBZ. and The least squares method is used to fit the scatter distribution of radar reflectivity and precipitation intensity in logarithmic coordinates, and the parameters that minimize the sum of squared residuals are solved. and The results were obtained by performing cross-validation on at least thirty paired samples to avoid overfitting.

[0097] Based on the equivalent precipitation intensity field of adjacent time intervals, the motion vector field is estimated using the optical flow method to obtain the moving speed and direction information of precipitation echoes.

[0098] Furthermore, based on the equivalent precipitation intensity field of adjacent time intervals, ground clutter removal and weak echo filtering are performed on the data to obtain the quality-controlled equivalent precipitation intensity field of adjacent time intervals. Pixels with pixel values ​​below -10 decibels are marked as invalid pixels. Pixels below -10 decibels are marked as invalid pixels because the echoes in this intensity range are usually close to the instrument noise limit and are difficult to correspond to the actual precipitation signal, which would interfere with the stability of optical flow calculation. Other pixels are marked as valid pixels, forming an effective pixel mask for optical flow. Under the constraint of the effective pixel mask for optical flow, the equivalent precipitation intensity field of adjacent time intervals after quality control is blurred with a 5×5 Gaussian kernel and constructed from coarse to fine using a 2x downsampling to form, for example, a three-layer pyramid equivalent precipitation intensity field. At the same time, the pyramid equivalent precipitation intensity is filtered at each layer. The gradient field is calculated using a 3×3 Sobel operator to calculate the east-west gradient field and the north-south gradient field respectively, and the zero value is maintained at the position of the effective pixel mask of optical flow. The pyramid equivalent precipitation intensity field and the pyramid gradient field are obtained. The pixel displacement is solved iteratively from coarse to fine using the pyramid Lucas-Kanade optical flow method to obtain the initial motion vector field. The edge smoothing and sparse extrapolation filling are performed by using the initial motion vector field and the effective pixel mask of optical flow. The speed is converted according to the time interval between adjacent time intervals (e.g., six minutes) to obtain the motion vector field. The time interval is six minutes because the operational meteorological radar CINRAD usually completes a volume scan every 5-6 minutes. The speed magnitude and azimuth of the motion vector field are calculated grid by grid to obtain the moving speed and direction information of precipitation echo.

[0099] By extrapolating the radar observation field at the current moment through the motion vector field, the radar extrapolation field at the future moment can be obtained.

[0100] Furthermore, to calculate the reverse tracing coordinates for each future grid point on the unified spatial grid, the reverse tracing coordinates are obtained by performing coordinate difference calculations on the combination of the future grid point coordinates, the velocity vector of the motion vector field at the future grid point location, and the extrapolated time step. The extrapolated time step is six minutes. Using the radar observation field at the current moment and the reverse tracing coordinates, bilinear interpolation is used to extract values ​​point-by-point on the radar observation field at the current moment to form the extrapolated value field. Invalid pixels in the extrapolated value field are masked using an optical flow effective pixel mask, and then the numerical values ​​are processed. Physical quantity ranges were clipped, with reflectivity limited to -10 dBZ to 60 dBZ and equivalent precipitation intensity limited to 0 to 100 mm / h. The reflectivity was limited to -10 dBZ to 60 dBZ because echoes below -10 dBZ are usually close to radar noise and cannot correspond to actual precipitation. Echoes above 60 dBZ in mid-latitude regions generally correspond to extreme convection or hail. Direct extrapolation would lead to non-physical maxima, which could easily cause numerical divergence in subsequent fusion or update steps. The equivalent precipitation intensity was limited to 0 to 100 mm / h because 100 mm / h corresponds to an echo of about 60 dBZ. Precipitation exceeding this value is extremely rare in actual observations and is often noise or local extreme pixels. Direct extrapolation would make subsequent quality control difficult. The mask voids were filled by interpolation using a neighborhood-preserving method to obtain the radar extrapolation field at future times.

[0101] Based on a preset time gating function, the error variances of the high-resolution refined field and the radar extrapolation field under different forecast lead times are calculated respectively, and the mode weights and radar weights that change with time are obtained.

[0102] Furthermore, the preset time gating function is expressed as:

[0103] ;

[0104] in, Indicates the time-gated function at time t. The value of , This indicates the maximum forecast lead time, with a value of 120 minutes.

[0105] It should be noted that, The value of 120 minutes is chosen because the radar extrapolation method is relatively reliable in predicting the translation of precipitation echoes within 0-2 hours, but after 2 hours, the generation, dissipation, and intensity changes of precipitation echoes will cause the extrapolation error to increase rapidly.

[0106] The time weighting factor for each forecast lead time is calculated using a time-gated function. For each forecast lead time, the difference field (the difference between the high-resolution refined field and the radar extrapolation field) is first calculated, and the difference is obtained grid-by-grid based on the grid area. The product of the difference field and the time weighting factor is used as the observation bias of the high-resolution refined field. The complementarity between the time weighting factor and constant 1 is defined as the complementarity factor. The quantity in the opposite direction after combining the difference field and the complementarity factor is used as the observation bias of the radar extrapolation field. For each forecast lead time, the observation bias is adjusted by removing the mean according to the grid area and the weight of the time weighting factor. The mean-reduced observation bias is then... The deviation is squared at each grid point and weighted and averaged with equal weights. Extremely large values ​​are truncated, for example, to the 99th percentile. The obtained weighted average squared values ​​are used as the high-resolution refined field error variance and the radar extrapolation field error variance, respectively. At each forecast lead time, the reciprocals of the high-resolution refined field error variance and the radar extrapolation field error variance are calculated and a small positive number is added to avoid the value being zero. The sum of the two is then used as the model weight and radar weight for a unified constraint. Finally, the model weight and radar weight are unified according to the reciprocal variance weighting principle to form the fusion coefficient field.

[0107] Based on the fusion coefficient field, the high-resolution refined field and the radar extrapolation field are weighted and fused at each grid point to obtain the fused precipitation field.

[0108] Furthermore, by fusing the coefficient field, the high-resolution refined field, and the radar extrapolation field, the values ​​of the fusing coefficient field, the high-resolution refined field, and the radar extrapolation field are read point by point on a unified spatial grid and a unified time axis. Then, at each grid point, the value of the high-resolution refined field at that grid point is proportionally adjusted according to the value of the fusing coefficient field at that grid point, and the value of the radar extrapolation field at that grid point is adjusted accordingly. The adjusted values ​​of the high-resolution refined field and the adjusted values ​​of the radar extrapolation field at that grid point are then linearly combined at that grid point. Finally, the combined results of all grid points are gathered on a unified spatial grid and a unified time axis to obtain the fused precipitation field.

[0109] Furthermore, by fusing the coefficient field, the high-resolution refined field, and the radar extrapolation field, the values ​​of the fusing coefficient field, the high-resolution refined field, and the radar extrapolation field are read point by point on a unified spatial grid and a unified time axis. Then, at each grid point, the product of the value of the fusing coefficient field at that grid point and the value of the high-resolution refined field at that grid point is obtained, and the sum of this product and the value of the fusing coefficient field at that grid point is subtracted as an intermediate value. The product of the intermediate value and the value of the radar extrapolation field at that grid point is calculated as a weighted sum. Finally, the weighted sum of all grid points is used to reconstruct the spatial distribution on a unified spatial grid and a unified time axis to obtain the fused precipitation field.

[0110] S6. Collect and integrate the construction intensity, structure and trend of precipitation fields, obtain a single composite classification index, classify it through adaptive quantile threshold, and obtain a classification map.

[0111] The numerical units of the fused precipitation field at each grid point are unified to millimeters per hour. The corresponding precipitation is then calculated at a time step of six minutes, and the ratio of precipitation to time step is converted into precipitation intensity. The precipitation intensity results of all grid points are output as a precipitation intensity field on a unified spatial grid and a unified time axis. An eight-neighbor precipitation connectivity mask is generated for the fused precipitation field according to an intensity threshold (e.g., not less than one millimeter per hour). The intensity threshold is set to not less than one millimeter per hour because precipitation less than one millimeter per hour is close to the lower limit of instrument measurement accuracy in many rain gauge records and is easily affected by noise and evaporation. Connectivity domain marking is then performed on the precipitation connectivity mask, and the area of ​​each connected domain is obtained as the number of connected pixels, the area of ​​the grid cell, and the perimeter as the product of the number of boundary pixels and the grid edge length. Finally, the spatial structure characteristics of each grid point are defined as the connected domain area and the shape index, where the shape index is the perimeter. The spatial structure feature field of the fused precipitation field is obtained by taking the square root of the ratio of the square of the product of 4π and the area of ​​the connected domain. The temporal variation trend of the fused precipitation field is calculated point by point on a unified time axis. The temporal variation trend field is obtained by using the slope of time series linear regression. At each grid point, the precipitation intensity field, spatial structure feature field and temporal variation trend field are summed by weighted normalization to calculate a single composite classification index. The single composite classification index values ​​of all grid points are collected at each time section. They are sorted from smallest to largest and the cumulative frequency of each value in all samples is assigned to each value position to obtain the empirical distribution of the single composite classification index at each time section. The values ​​are read at the corresponding quantile positions as classification thresholds to ensure that the three-level classification is divided into low, medium and high levels at the 0.6 and 0.9 quantiles. The adaptive quantile threshold is obtained point by point. The classification level is mapped point by point and eight-neighbor connectivity cleaning is performed to obtain the classification map.

[0112] It should be noted that the grading threshold is set at 0.6 in the lower quantile to ensure that precipitation that is above average but not extreme can be identified as medium-level, while it is set at 0.9 in the higher quantile because it corresponds to the top 10% of extremely high intensity precipitation, which is usually used to mark high-risk or heavy precipitation areas.

[0113] S7. By fusing precipitation field and graded map, the unified state vector is jointly corrected for amplitude and deformation based on Kalman-type update, and the updated unified state vector is obtained and passed to the next round of large-scale extrapolation.

[0114] Based on the fused precipitation field and graded map, the observation residuals and their uncertainties are calculated, and the observation vector and observation error covariance matrix are constructed.

[0115] Furthermore, based on the fused precipitation field and the classification map, the quantile representative values ​​of the fused precipitation field are statistically analyzed for each level on a unified spatial grid. The quantile representative value is the value at the median position after sorting the fused precipitation field values ​​for each level. The quantile representative value is written into the corresponding grid point to form the level representative intensity field. The difference between the fused precipitation field and the level representative intensity field is calculated point by point on the unified spatial grid to obtain the observation residual field. By using the observation residual field and the classification map, the observation residual field values ​​corresponding to all grid points on the classification map that are equal to the current level are collected for each time section and each level. The median of the observation residual values ​​of all grid points at that time section is calculated as the median value. Then, the absolute value of the difference between each value and the median value is calculated, and the median is taken as the median absolute deviation. Subsequently, the robust standard deviation is obtained by multiplying the median absolute deviation by 1.4812. The square of the robust standard deviation is then multiplied by... Take the larger value as the robust variance estimate of the time section and the level, assign the robust variance estimate back to all grid points of the classification map that are equal to the level, obtain the observation uncertainty field, and stack the values ​​of the fused precipitation field at all grid points in row priority order to form the observation vector. Fill the diagonal matrix with the variance of the observation uncertainty field at each grid point to obtain the observation error covariance matrix, thereby obtaining the observation vector and the observation error covariance matrix.

[0116] By using the observation vector, the observation error covariance matrix, and the predicted state vector, a Kalman-type update method is employed to jointly correct the magnitude component and deformation component in the unified state vector.

[0117] Furthermore, based on the fixed rule of mapping the amplitude component of the unified state vector to the observed precipitation intensity and the deformation component to the observed spatial pixel displacement, an observation operator matrix is ​​constructed. This matrix is ​​applied to the predicted state vector to obtain simulated observations. The difference between the simulated observations and the observed vectors is used to form the updated observation vector. The predicted state error covariance matrix is ​​obtained by extrapolating the state error covariance matrix from the previous update through Jacobian propagation and superimposing the process noise covariance matrix. Within a unified time axis and a unified spatial grid, the product of the observation operator matrix, the predicted state error covariance matrix, and the transpose of the observation operator matrix is ​​used as the intermediate observation value. The sum of the intermediate observation value and the observation error covariance matrix is ​​then obtained. Symmetry and minor regularization are performed to obtain the observation covariance matrix, which is used to characterize the uncertainty of updating the observation vector. The observation covariance matrix and minor regularization terms (e.g., ...) are then obtained. The sum of the products of the identity operators is used to obtain the lower triangular factors through Cholesky decomposition. Then, the inverse action of the observation covariance matrix is ​​obtained by solving the trigonometric equations twice. The Kalman gain matrix is ​​calculated by first performing matrix operations on the transpose of the predicted state error covariance matrix and the observation operator matrix, and then performing matrix operations on the inverse action of the resulting matrix and the observation covariance matrix in sequence. The Kalman gain matrix is ​​then multiplied by the updated observation vector to perform a joint additive correction on the magnitude and deformation components of the unified state vector, resulting in the updated unified state vector. The predicted state error covariance matrix is ​​updated according to the Kalman-type update method, and symmetry and non-negative qualitative corrections are applied to obtain the updated state error covariance matrix. The intermediate covariance is updated according to the intermediate formula, expressed as:

[0118] ;

[0119] in, This represents the updated state error covariance matrix. This represents the covariance matrix of the predicted state error. Represents the Kalman gain matrix. Represents the observation operator matrix, Represents the observation error covariance matrix. This represents the transpose of a matrix.

[0120] The updated state error covariance matrix is ​​symmetrically processed and expressed as follows:

[0121] ;

[0122] The symmetric state error covariance matrix is ​​subjected to eigenvalue decomposition, and all negative eigenvalues ​​are truncated to zero before reconstruction to obtain the updated state error covariance matrix.

[0123] The unified state vector updated by the Kalman algorithm is used as the initial state for the next round of large-scale extrapolation, and is used for subsequent iterative cycles.

[0124] Furthermore, firstly, the starting time of the next round of large-scale extrapolation is calculated on a unified time axis, and the unified state vector updated by the Kalman algorithm is directly assigned as the initial state vector of the next round of large-scale extrapolation. At the same time, the updated state error covariance matrix is ​​directly assigned as the initial state error covariance matrix of the next round of large-scale extrapolation. Then, based on the extrapolation operator of spatial deformation alignment and amplitude deviation correction, linearized forward propagation is performed on the extrapolation time step. The Jacobian matrix of the extrapolation operator is used to perform a two-sided matrix action on the initial state error covariance matrix and the process noise covariance matrix is ​​superimposed to obtain the predicted state error of the next round. The covariance matrix is ​​obtained, and the next round of predicted state vectors is obtained. The large-scale field after spatial deformation alignment and the background field after large-scale correction are generated sequentially. Spatial resolution refinement, terrain correction and mass conservation constraints are completed. Then, the radar extrapolation field and the model field are weighted and fused according to the fusion coefficient field to obtain the fused precipitation field. The fused precipitation field and the classification map are used to construct the next round of observation vector and observation error covariance matrix. Thus, in the next round of Kalman-type update, the next round of predicted state vector and the next round of predicted state error covariance matrix are used for closed-loop iteration. The classification map generated in each round is used as the classification recognition result.

[0125] This embodiment also provides a computer device applicable to the meteorological radar classification and identification method combined with meteorological numerical models, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the meteorological radar classification and identification method combined with meteorological numerical models as proposed in the above embodiment.

[0126] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0127] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the meteorological radar classification and identification method combining meteorological numerical models as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0128] In summary, this invention achieves stability in the initialization of precipitation fields without radar by generating a unified initial state vector from statistical analysis of similar historical typhoon samples; it also achieves optimal combination of information at different time scales by dynamically weighting and fusing the radar extrapolation field within 2 hours and the numerical model field 2 hours later according to the normalized inverse of the error variance through time-gated weight allocation based on a time-gated function, significantly reducing short-term forecast errors; and it achieves high-precision forecasting of precipitation field intensity, structure, and trend by extrapolating the model CLDAS field, refining the spatial resolution, fusing the radar extrapolation field, and combining it with Kalman updates.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A hierarchical identification method of a weather radar combined with a weather numerical model, characterized in that: comprising, Collecting multi-source meteorological data, establishing a unified time axis and spatial grid, and statistically pre-setting calibration regions and historical typhoon sample information to obtain a multi-source data set; Through the historical typhoon sample information, a unified initial state vector is obtained in the pre-set calibration region, and a unified state vector is obtained through second route trajectory anchor calibration; Through the unified state vector, the large-scale field of the meteorological numerical model in the multi-source data set is extrapolated and corrected, and a background field is obtained through one-time consistency constraint; The background field is spatially refined, locally corrected in combination with a terrain factor, and quality conservation and edge preservation smoothing are performed to obtain a high-resolution refined field; The radar observation field in the multi-source data set is extrapolated to obtain a radar extrapolation field, and the high-resolution refined field and the radar extrapolation field are fused through time-gated inverse variance weighting to obtain a fused precipitation field; A single composite grading index of the fused precipitation field is obtained, and grading is performed through a quantile threshold to obtain a grading map; Through the fused precipitation field and the grading map, the unified state vector is jointly corrected in amplitude and deformation based on Kalman-type updating to obtain an updated unified state vector; Based on the updated unified state vector, the fused precipitation field and the grading map are dynamically corrected to obtain a final grading recognition result; Through the corresponding time point typhoon center longitude and latitude points in the historical typhoon sample information and the current typhoon center longitude and latitude points, a time registration pair is established, and all time registration pairs are arranged in time sequence to form a second route trajectory anchor sequence.

2. The hierarchical identification method of weather radar combined with weather numerical model according to claim 1, characterized in that: The collecting multi-source meteorological data, establishing a unified time axis and spatial grid, and statistically pre-setting calibration regions and historical typhoon sample information to obtain a multi-source data set and region information, and the specific steps are, Collecting meteorological numerical model data, radar observation data, and ground observation data and performing time and space unification processing to establish a unified time axis and spatial grid; Determine the pre-set calibration region and the historical typhoon sample information to obtain the standardized multi-source data set and the region information for initialization.

3. The hierarchical identification method of weather radar combined with weather numerical model according to claim 2, characterized in that: The through the historical typhoon sample information, a unified initial state vector is obtained in the pre-set calibration region, and a unified state vector is obtained through second route trajectory anchor calibration, and the specific steps are, Based on the historical typhoon sample information, a unified initial state vector composed of amplitude components and deformation components is constructed in the pre-set calibration region; The unified initial state vector is calibrated through the second route trajectory anchor to obtain a unified state vector that has been deformed.

4. The hierarchical identification method of weather radar combined with weather numerical model according to claim 3, characterized in that: The through the unified state vector, the large-scale field of the meteorological numerical model in the multi-source data set is extrapolated and corrected, and a background field is obtained through one-time consistency constraint, and the specific steps are, Through the unified state vector, the large-scale field of the meteorological numerical model in the multi-source data set is spatially deformed and aligned, and the amplitude deviation is corrected, and the background field is obtained through one-time consistency constraint.

5. The method for hierarchical identification of weather radar according to weather numerical model as claimed in claim 4, characterized in that: The background field is spatially refined, locally corrected in combination with a terrain factor, and quality conservation and edge preservation smoothing are performed to obtain a high-resolution refined field, and the specific steps are, The background field is spatially refined to obtain a preliminary refined field; The terrain-corrected refined field is obtained by correcting the preliminary refined field based on terrain factors; The terrain-corrected refined field is subjected to a mass conservation constraint, and edge-keeping smoothing processing is performed to obtain a final high-resolution refined field.

6. The method for hierarchical recognition of weather radar combined with weather numerical model according to claim 5, characterized in that: The radar extrapolation field is obtained by extrapolating the radar observation field in the multi-source data set, and the specific steps are as follows, The radar observation field is extracted from the multi-source data set, and quality control and dimension unification are performed to obtain the radar reflectivity, and the radar reflectivity is converted into an equivalent precipitation intensity field; Based on the equivalent precipitation intensity field of the adjacent time, the motion vector field is estimated by using an optical flow method to obtain the moving speed and direction information of the precipitation echo; The radar extrapolation field at the future time is obtained by performing reverse tracking extrapolation on the equivalent precipitation intensity field at the current time through the motion vector field.

7. The method for hierarchical recognition of weather radar combined with weather numerical model according to claim 6, characterized in that: The fusion precipitation field is obtained by fusing the high-resolution refined field and the radar extrapolation field through time-gated inverse variance weighting, and the specific steps are as follows, Based on the preset time-gated function, the error variances of the high-resolution refined field and the radar extrapolation field under different prediction time lengths are calculated to obtain the model weight and the radar weight varying with time; The model weight and the radar weight are normalized by the inverse variance weighting principle to generate a fusion coefficient field; Based on the fusion coefficient field, the high-resolution refined field and the radar extrapolation field are weighted and fused at each grid point to obtain the fusion precipitation field.

8. The hierarchical identification method of weather radar integrated with weather numerical model according to claim 7, characterized in that: The single composite hierarchical index of the fusion precipitation field is obtained, and the specific steps are as follows, The precipitation intensity is calculated and the precipitation intensity field is output through the fusion precipitation field at each grid point on the unified spatial grid and the unified time axis; The spatial structure feature field is obtained by extracting the connected domain based on the intensity threshold and calculating the connected domain area and the shape index; The trend feature field is calculated on the unified time axis, and the single composite hierarchical index is obtained through the precipitation intensity field, the spatial structure feature field and the trend feature field.

9. The hierarchical identification method of weather radar integrated with weather numerical model according to claim 8, characterized in that: The hierarchical graph is obtained by hierarchical classification through the quantile threshold, and the specific steps are as follows, The quantile threshold is generated by calculating the empirical distribution of the single composite hierarchical index at each time section and reading the value of the single composite hierarchical index at the corresponding quantile position at the target quantile position; The hierarchical level of each grid point is determined by comparing the single composite hierarchical index and the quantile threshold at each grid point, and the hierarchical levels of all grid points are combined to form an initial hierarchical graph in the unified spatial grid; The hierarchical graph is obtained by performing eight-neighbor connectedness cleaning on the initial hierarchical graph.

10. The method for hierarchical identification of weather radar in combination with a weather numerical model according to claim 9, characterized in that: The updated unified state vector is obtained by jointly correcting the amplitude and deformation components of the unified state vector based on the Kalman-type update of the fusion precipitation field and the hierarchical graph, and the specific steps are as follows, Based on the fusion precipitation field and the hierarchical graph, the observation residual and the uncertainty are calculated, and the observation vector and the observation error covariance matrix are constructed; The amplitude component and the deformation component of the unified state vector are jointly corrected by using the Kalman-type update method based on the observation vector, the observation error covariance matrix and the predicted state vector generated by large-scale field extrapolation of the updated unified state vector in the last round to obtain the updated unified state vector. The final hierarchical recognition result is obtained by dynamically correcting the fused precipitation field and the hierarchical map based on the updated unified state vector. The final hierarchical recognition result is obtained by dynamically correcting the fused precipitation field and the hierarchical map based on the updated unified state vector.

Images