A method for constructing a three-dimensional model of a local cloud mass based on radar data

Abstract

This invention discloses a method for constructing a local cloud 3D model based on radar data, belonging to the field of cloud model construction technology. The method involves preprocessing basic information data from radar sites to output a first structured dataset; classifying and processing the first structured dataset to output a second structured dataset; reading and converting the format of the output second structured dataset; using the DBSCAN algorithm to remove isolated noise points based on the converted data structure; extracting boundaries using a distance threshold method; and further filtering false boundary points using geometric constraints from the Alpha Shape method and the local convex hull method to complete the construction of the local cloud 3D model. Through a process from global fitting to local completion and then to filtering and purification, this method solves the core problems of inaccurate shape, loss of detail, and noise interference inherent in traditional single methods, achieving a more realistic, refined, and reliable 3D boundary for cloud formations.

Description

Technical Field

[0001] This invention relates to the field of cloud model construction technology, specifically a method for constructing a local cloud 3D model based on radar data. Background Technology

[0002] As an important carrier of atmospheric motion, cloud clusters directly affect the occurrence and development of weather phenomena such as precipitation and severe convection due to their morphology, structure, physical characteristics, and evolution. Accurate cloud cluster modeling is the core foundation for improving the accuracy of short-term and nowcasting weather forecasts and carrying out severe weather warnings.

[0003] With the continuous upgrading of meteorological detection technology, radar has become the core equipment for cloud monitoring due to its advantages such as high spatiotemporal resolution and ability to penetrate clouds. Among them, X-band radar, with its moderate detection range, high spatial resolution, and strong ability to detect boundary layer atmosphere, has shown unique advantages in monitoring small-scale cloud clusters such as local sudden severe convective weather, and is gradually becoming the core data source for refined cloud cluster modeling.

[0004] X-band radar data offers higher spatiotemporal resolution, enabling precise capture of fine cloud structures and supporting dynamic cloud modeling. However, current research on X-band radar cloud modeling algorithms focuses on how to achieve cloud region division through single clustering or contour extraction algorithms, without considering the impact of actual operational scenarios such as regional clutter interference, multiple cloud cluster superposition, and cloud edge blurring on modeling accuracy. This makes it unsuitable for refined modeling under complex meteorological conditions, reducing the credibility of 3D cloud models. Furthermore, existing operational radar data processing systems do not fully exploit the spatial morphological value of the data, only outputting two-dimensional statistical information, significantly reducing the utilization rate and application value of X-band radar in refined meteorological services.

[0005] Against this backdrop, developing a dedicated cloud cluster modeling method for X-band radar data has become a key technological breakthrough for improving the accuracy of cloud cluster monitoring and forecasting. Summary of the Invention

[0006] The purpose of this invention is to provide a technical solution for constructing a local cloud 3D model based on radar data, in order to solve the technical problem that the existing 3D cloud modeling methods do not consider the impact of actual business scenarios such as regional clutter interference, multiple cloud superposition, and cloud edge blurring on the modeling accuracy, and are not suitable for fine modeling work under complex meteorological conditions, thus reducing the credibility of the cloud 3D model.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a local cloud 3D model based on radar data, comprising the following steps:

[0008] S1. Preprocess the basic information data of the radar site and output the first structured dataset; S2. Classify and process the first structured dataset obtained in S1, and output the second structured dataset; S3. Read and convert the second structured dataset output by S2. Based on the converted data structure, remove isolated noise points using the DBSCAN method, extract the boundary using the distance threshold method, and further filter false boundary points using the geometric constraints of the AlphaShape method and the local convex hull method to complete the construction of the local cloud 3D model.

[0009] According to the above technical solution, step S3 includes the following steps: S301. Define cloud clusters of different shapes using the density connectivity method of DBSCAN; S302. Use the distance threshold method to extract the cloud boundary of the cluster defined by the DBSCAN method, and obtain the discrete point set of the cloud boundary. S303. The boundaries extracted in S302 are filtered using the Alpha Shape method to obtain the preliminary cloud boundary contour. S304. The preliminary boundary contour selected by the Alpha Shape method is divided into local regions using the local convex hull method, and the convex hull is calculated for each local region. Small subsets of points within the local region that satisfy convexity are directly retained. If there are depressions, the Alpha Shape method is used to further filter the data, and S303 is executed again. This process continues until the construction of the local cloud 3D model is completed.

[0010] According to the above technical solution, in S302, the distance threshold method extracts the cloud boundary by setting a fixed or dynamic distance threshold to determine the spatial relationship between the point and the target core area and / or between points. Among them, the threshold is dynamically calculated based on the distance distribution of the discrete point set itself; at the same time, the cloud boundary is extracted through 3D threshold.

[0011] Furthermore, based on the distance distribution of the discrete point set itself, the threshold is dynamically calculated using the k-nearest neighbor distance quantile method. Specifically, the average distance from each point to its k nearest neighbors is calculated, and the median of the k-nearest neighbor distances of all points is taken as the adaptive threshold.

[0012] According to the above technical solution, in S303, the discrete point set obtained by the distance threshold method is triangulated by Delaunay triangulation, and each triangular facet of the Delaunay triangulation is traversed to calculate its circumcircle radius. Triangular facets that meet the conditions are α-valid facets and are retained; those that do not meet the conditions are discarded. Then iterate through all the retained α-valid faces and count the number of times each triangle edge is shared: if an edge is shared by only 1 α-valid face, then the edge is the boundary edge of the Alpha Shape; If an edge is shared by at least two α-effective patches, then the edge is an internal edge and is removed. All boundary edges are connected in sequence to form a closed Alpha Shape contour, thus obtaining the preliminary cloud boundary contour.

[0013] According to the above technical solution, in S304, the local convex hull method sorts the distances from the boundary points to the center of the cloud and retains b% of the outer points, where b% represents the percentage of values. Several boundary points with a distance less than a set threshold form a local cluster, and each local cluster is regarded as a small subset of points.

[0014] According to the above technical solution, S1 specifically includes the following steps: S101. Read and analyze the basic information data of the radar site; S102. Convert the polar coordinates of each data point in the radar volume scan data to station-centered rectangular coordinates; S103. Convert the station-centered rectangular coordinates of each data point obtained in S102 to geocentric and geofixed coordinates; S104. Convert the geocentric coordinates of each data point obtained in S103 to geodetic coordinates; S105. Detect and filter abnormal data, and output the first structured dataset.

[0015] According to the above technical solution, step S2 specifically includes the following steps: S201. Use the data filtering module to classify, filter and lightweight the first structured dataset, and output valid cloud data. S202. Use the cloud clustering module to perform preliminary clustering on the effective cloud data processed by the data filtering module, and output the preliminary clustering results. S203. The density merging optimization module is used to refine the effective cloud clusters by dynamically selecting the K value, performing two-dimensional merging, and multi-dimensional noise reduction on the preliminary clustering results output by the cloud clustering module.

[0016] According to the above technical solution, in S201, the radar data column names from different sources and in different formats are identified by the keyword fuzzy matching algorithm, and the columns containing specific keywords are mapped to standard column names. Valid cloud data is filtered by setting a reflectivity threshold, and spatial consistency checks are performed on the valid cloud data that has passed the reflectivity threshold to remove outliers that are outside the physical range. By retaining core fields and extracting redundant information, effective cloud data optimization and lightweight processing are performed, and a standardized format is output.

[0017] According to the above technical solution, in S202, the cloud reflectivity is truncated by setting a threshold value for the fractional range, and abnormal values ​​of cloud reflectivity are eliminated. The effective cloud data is Z-score standardized, and the standardized feature values ​​are then optimized using K-Medoids coarse clustering. Three-dimensional buffer analysis and overlapping cluster merging were performed on the effective cloud data after K-Medoids coarse clustering optimization. Then, HDBSCAN is used for purification to output preliminary clustering results.

[0018] According to the above technical solution, step S203 includes the following steps: S203-1. Dynamically select the number of clusters K for K-Medoids based on the number N of data points in the preliminary clustering results output by the cloud clustering module; S203-2, Perform a two-dimensional merging of density correlation dimension and spatial proximity dimension; S203-3, Perform multi-dimensional noise filtering; S203-4. Organize the data filtered in S203-3 and output the second structured dataset.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. The core of the synergistic approach of the Alpha Shape method, the local convex hull method, and the density method is "to leverage strengths and avoid weaknesses, and to advance step by step": the Alpha Shape method lays the foundation for the global shape, the local convex hull method completes the local details, and the density method refines the boundary quality, ultimately achieving the extraction of 3D cloud boundaries with "realistic shape + fine details + reliable quality". This not only meets the intuitiveness requirements of visualization, but also supports the accuracy requirements of scientific research analysis, and is especially suitable for application scenarios with complex shapes and easily disturbed data, such as ocean precipitation clouds.

[0020] The synergistic application of the Alpha Shape method, the local convex hull method, and the distance threshold method essentially solves the three core problems of traditional single methods—inaccurate shape, loss of detail, and noise interference—through the process of "global fitting (Alpha Shape) → local completion (local convex hull) → filtering and purification (distance threshold)." Ultimately, it achieves more realistic, detailed, and reliable 3D cloud boundaries, satisfying both the intuitiveness of visualization and the accuracy requirements of subsequent scientific research analysis. It is particularly suitable for application scenarios with complex shapes and easily interfered data, such as ocean precipitation clouds.

[0021] 2. High noise removal efficiency: Through a three-level filtering mechanism (physical threshold filtering, spatial continuity check, and altitude layer optimization), various types of noise such as atmospheric clutter and ground object echoes are systematically removed, resulting in a higher overall noise removal rate. Among them, the accuracy of non-precipitation signal identification is improved. Thorough format standardization: The column name intelligent adaptation algorithm has a higher compatibility with common radar data formats, enabling seamless integration of data from different sources and in different formats, and eliminating processing errors caused by inconsistent data formats; Clustering accuracy is greatly improved: High-quality data input makes the subsequent clustering algorithm more accurate than traditional methods, especially in key steps such as cloud boundary identification and weak signal extraction, and the false judgment rate is reduced.

[0022] The hybrid strategy has obvious advantages: the hybrid strategy of "K-Medoids coarse clustering + HDBSCAN fine purification" effectively overcomes the limitations of a single algorithm. K-Medoids is robust to outliers, and HDBSCAN can identify density changes. The combination of the two makes the algorithm more adaptable to problems such as uneven cloud density and blurred boundaries. Capability to handle complex scenarios: This method achieves higher clustering accuracy in complex meteorological scenarios such as overlapping cloud clusters, multi-layered cloud systems, and fragmented cloud clusters.

[0023] Parameter adaptive optimization: The dynamic K-value selection mechanism and adaptive parameter adjustment enable the algorithm to adapt to the data characteristics of different weather types (such as typhoons, rainstorms, convective precipitation, etc.) without human intervention.

[0024] Superior purification accuracy: Through a two-dimensional merging strategy (density correlation + spatial proximity) and a multi-dimensional noise reduction strategy (size + reflectivity + dispersion), cloud clusters are purified with higher accuracy, enabling accurate separation of spatially adjacent but physically different cloud bodies. High boundary recognition accuracy: The cloud boundary is significantly improved compared to traditional methods, ensuring the accuracy of 3D modeling.

[0025] Seamless adaptation to multiple formats: Supports multiple data formats such as Excel, Parquet, NetCDF, and NPY, with higher compatibility with existing meteorological business systems' data interfaces, and can be integrated into existing business processes without additional development; Wide range of applications: The processing results can be directly applied to multiple business processes such as short-term forecasting, precipitation estimation, cloud physics analysis, and numerical model initialization.

[0026] 3. Through rigorous coordinate transformation and anomaly screening, the raw binary radar data is transformed into a structured dataset with accurate georeferencing and controllable quality. The output is in standard geographic coordinates and a common data format, promoting data sharing and business collaboration. By employing the WGS84 ellipsoidal model and iterative coordinate transformation algorithm, higher positioning accuracy is achieved. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the processing flow of the data filtering module of the present invention; Figure 2 This is a simplified flowchart of the cloud clustering module of the present invention; Figure 3 This is a simplified flowchart of the density merging optimization module of the present invention; Figure 4 This is a diagram showing the cloud structure after the cloud clustering module of this invention performs algorithmic classification. Figure 5 This is a 3D cloud distribution map after cloud density merging by the density merging optimization module of the present invention. Figure 6 This is a comparison chart of the average reflectance after cloud density merging by the density merging optimization module of the present invention. Figure 7 This is a cloud size distribution diagram after cloud density merging is performed by the density merging optimization module of the present invention. Figure 8 This is a three-dimensional model of a local cloud obtained by combining the Alpha shape method and the local convex hull method of this invention. Figure 9 This is a three-dimensional model of a local cloud obtained by combining the Alpha shape method, the local convex hull method, and the distance threshold method of this invention. Figure 10 This is a three-dimensional model of a local cloud obtained by combining the Alpha shape method, the local convex hull method, and the density method of this invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example: This example discloses a method for constructing a local cloud 3D model based on radar data, including the following steps: S1. Preprocess the basic information data of the radar site and output the first structured dataset.

[0030] Specifically, it includes the following steps: Step 1: Read and parse the basic information data of the radar site; In this embodiment, step 1 is specifically implemented in the following manner: 1. Parse the site code, site name, reference longitude λref (unit: degrees), reference latitude φref (unit: degrees), antenna altitude h_antenna (unit: meters), and ground altitude h_ground (unit: meters) data from the basic information data of the radar site (i.e., the radar data file header); 2. Calculate the total altitude H_ref of the station according to the following formula: H_ref = h_antenna + h_ground; 3. Read the radar volume scan data. Specifically, for each elevation angle layer (this embodiment processes the range of 0°~57.75° with a step size of 0.25°), obtain the elevation angle number n_elev, elevation angle value θ (unit: degrees), radial quantity N_radial, range quantity N_bin, azimuth sequence {φ_i} (i=1,...,N_radial, unit: degrees), radial range sequence {r_j} (j=1,...,N_bin, unit: meters), and the corresponding radar reflectivity matrix Z_ij (unit: dBZ).

[0031] Step 2: Convert the polar coordinates of each data point in the radar volume scan data to station-centered rectangular coordinates; In this embodiment, a station-centered rectangular coordinate system is established with the radar station as the origin. Specifically, the polar coordinates (φ_i, θ, r_j) of each data point are converted to station-centered rectangular coordinates (E, N, U). The conversion process is as follows: Horizontal distance R_h =: R_h = r_j·cos(θrad); Vertical height R_v: R_v = r_j·sin(θrad); Eastward component E: E = R_h·sin(φrad); Northward component N: N = R_h·cos(φrad); Celestial component U: U = R_v; Where θrad=(π / 180)·θ, φrad=(π / 180)·φ_i, and θrad is the angle converted to radians.

[0032] Step 3: Convert the station-centered rectangular coordinates of each data point obtained in Step 2 to geocentric coordinates; In this embodiment, the geocentric coordinate system has the Earth's center of mass as the origin, the X-axis pointing to the intersection of the prime meridian and the equator, the Z-axis pointing to the North Pole, and the Y-axis forming a right-handed coordinate system with the X and Z axes. The conversion from station-centered rectangular coordinates (E, N, U) to geocentric and Earth-fixed coordinates (X, Y, Z) is as follows: (X,Y,Z)=R(φref,λref)·(E,N,U)+(X_ref,Y_ref,Z_ref); in: (X_ref, Y_ref, Z_ref) are the geocentric and geofixed coordinates of the radar station, calculated using the WGS84 ellipsoid model. N_ref=a / sqrt(1-e²sin²φref); X_ref=(N_ref+H_ref)cosφref cosλref; Y_ref=(N_ref+H_ref)cosφref sinλref; Z_ref=[N_ref(1-e²)+H_ref]sinφref; R(φref, λref) is the rotation matrix from the station-centered rectangular coordinate system to the geocentric Earth-fixed coordinate system: R = [-sinλref,-sinφref cosλref,cosφref cosλref; cosλref, -sinφref sinλref, cosφref sinλref; 0, cosφref, sinφref]; in: X_ref: The X-coordinate of the radar station in the Earth-centered Earth-fixed coordinate system, in meters; N_ref: Radius of curvature of the tropomorphic circle, in meters; H_ref: The total altitude of the radar site, H_ref = h_antenna + h_ground, in meters; φref: Reference latitude of the radar station, in degrees; λref: Reference longitude of the radar station, in degrees.

[0033] In this embodiment, the semi-major axis of the WGS84 ellipsoid model is a = 6378137.0 m; The first eccentricity square of the WGS84 elliptic model is e² = 6.69437999014 × 10⁻⁶. - ³.

[0034] Step 4: Convert the geocentric coordinates of each data point obtained in Step 3 to geodetic coordinates; In this embodiment, the geocentric coordinates (X,Y,Z) are converted to geodetic coordinates (φ,λ,h) using an iterative method, as follows: p = sqrt(X² + Y²); λ=arctan 2 (Y,X); φ (0) =arctan(Z / p·1 / (1-e²)); Iterate k=0,1,2,... until convergence: N(φ (k) = a / sqrt(1-e²sin²φ) (k) ); h (k) =p / cosφ (k) -N(φ (k) ); φ (k+ ¹ ) =arctan(Z / p·1 / (1-e²·N(φ (k) ) / (N(φ (k) )+h (k) ))); The convergence condition is: |φ (k+ ¹ ) -φ (k) |<10 -12 rad, ultimately yielding the latitude φ=φ (k+ ¹ ) Longitude λ, altitude h=h (k) .

[0035] in: N(φ (k) ): Latitude φ at the k-th iteration (k) The radius of curvature of the elliptic at point 1, in meters; a: The semi-major axis of the WGS84 ellipsoid model, with a value of 6,378,137.0 meters; e²: The square of the first eccentricity of the WGS84 ellipsoid model, with a value of 6.69437999014 × 10⁻⁶. - ³; φ (k) : Latitude value at the k-th iteration, in degrees.

[0036] Step 5: Detect and filter out abnormal data; In this embodiment, step 5 is specifically implemented in the following manner: 1. Invalid value replacement: Replace all padded and missing values ​​in the original data with NaN; 2. Check the rationality of reflectance: Eliminate physical anomalies with reflectance Z < -30dBZ or Z > 70dBZ; 3. Spatial range rationality check: Remove points with radial distance r_j > 300000m and invalid points with altitude h < 0m or h > 15000m; 4. Geographical scope rationality check: The data in this embodiment is limited to the Chinese region, that is, latitude φ∈[18°, 54°], longitude λ∈[73°, 135°].

[0037] Step 6: Output the first structured dataset; In this embodiment, the valid data points are organized into a first structured dataset, with each row containing: longitude, latitude, altitude (unit: m), reflectivity (unit: dBZ), elevation angle (unit: °), azimuth angle (unit: °), radial distance (unit: m), and elevation angle category. The output format supports: Excel (.xlsx), Parquet, NumPy (.npy), and NetCDF (.nc).

[0038] It should be noted that the following verification implementation method is provided in this embodiment: Taking a certain X-band radar data file Z_XBAND_RADAR_20230520010000.bin.bz2 as an example: Data read: Site information successfully parsed: Latitude 30.25°N, Longitude 120.18°E, Antenna height 85m, Ground height 10m, Total height 95m; Coordinate transformation: Processed 232 elevation angle layers from 0° to 57.75°, totaling 1,250,384 raw data points, and completed the transformation from polar coordinate system to geodetic coordinate system; Anomaly screening: 21,307 NaN values, 8,452 abnormal reflectivity points, 35,209 out-of-range spatial points, and 2,143 data points from outside the country were removed. Output: 1,183,273 valid data points were generated and saved as Z_XBAND_RADAR_20230520010000_0_57_corrected.xlsx (85.2 MB). Key statistics are as follows: Valid points: 1,183,273; Latitude range: 29.84°N~31.02°N; Longitude range: 119.76°E~121.37°E; Height range: 95.2 m to 14,832.6 m; Reflectivity range: -28.5dBZ~68.7dBZ.

[0039] S2. Classify and process the first structured dataset obtained from S1, and output the second structured dataset.

[0040] Specifically, S2 classifies and processes the first structured dataset through a data filtering module, a cloud clustering module, and a density merging optimization module.

[0041] In this embodiment, as Figure 1 As shown, the processing flow of the data filtering module is as follows: Step 1: Perform intelligent column name adaptation and format standardization on the first structured dataset; Specifically, the keyword fuzzy matching algorithm automatically identifies radar data column names from different sources and in different formats. Based on Chinese semantic similarity calculation, the keyword fuzzy matching algorithm automatically maps columns containing specific keywords to standard column names. In this embodiment, for example: Words containing "longitude", "lon", or "longitude" are mapped to "longitude". Words containing "latitude", "lat", or "latitude" are mapped to "latitude". Words containing "height", "high", "alt", "height", etc. → are mapped to "height (unit: m)"; Terms containing "reflectivity", "dBZ", or "reflectivity" are mapped to "reflectivity (unit: dBZ)". Furthermore, when a core field is detected as missing, the system automatically records and skips the file, while generating a detailed error report to ensure the robustness of the processing flow. Since skipping the missing data and generating an error report are conventional techniques for those skilled in the art, they will not be elaborated upon in this application.

[0042] Step 2: Perform physical threshold filtering and effective data screening on the data processed in Step 1; Specifically, in this embodiment, based on extensive historical data analysis and verification, the reflectivity threshold is set to 15 dBZ for effective cloud signal filtering. The scientific basis for this reflectivity threshold is as follows: X-band radar signals below 15 dBZ are mostly atmospheric clutter, clear-sky turbulence, and other non-precipitation signals; a reflectivity greater than 15 dBZ corresponds to a precipitation intensity of approximately 0.1 mm / h, possessing actual precipitation potential; this reflectivity threshold eliminates noise while preserving the complete signal of weak precipitation cloud clusters. Subsequently, a spatial consistency check was performed on the data filtered by the reflectivity threshold to remove outliers that clearly exceeded the physical range.

[0043] Step 3: Optimize and lightweight the data structure of the data filtered and selected in Step 2; Specifically, 1. Core fields are retained: Only five core meteorological physical quantity fields are retained, including longitude (geographic coordinates), latitude (geographic coordinates), altitude (meters) (vertical coordinates), reflectance (dBZ) (physical intensity), and cloud cluster number (initial value is 0, used for subsequent clustering labels); 2. Redundant information removal: Auxiliary information such as elevation angle, azimuth angle, and scanning time are removed, reducing the data dimensions from more than 10 to 5, which significantly reduces the complexity of subsequent calculations. 3. Standardized output: The output is uniformly in Excel (.xlsx) format to ensure data readability and cross-platform compatibility.

[0044] In this embodiment, as Figure 2 and Figure 4 As shown, the cloud clustering module performs preliminary clustering on the standardized data processed by the data filtering module and outputs preliminary clustering results. The specific processing flow is as follows: Step 1: Preprocess and enhance the standardized data output by the data filtering module; Specifically, 1. The reflectance is truncated using the 0.01~0.99 quantile range, with the mathematical expression: Q_0.01≤Z≤Q_0.99, where Z is the reflectance, and Q_0.01 and Q_0.99 are the 1% and 99th quantiles of the reflectance distribution, respectively. This can effectively remove extreme outliers in reflectance and avoid their interference with cluster centers. 2. Perform Z-score standardization on the four core feature dimensions, as follows: X'=(X-μX) / σX; Where X is the original feature value, such as longitude, latitude, altitude and reflectivity, μX is the mean of any feature value, σX is the standard deviation of any feature value, and X' is the feature value after standardization. Standardization eliminates the influence of different physical dimensions on distance measurement.

[0045] Step 2: Perform K-Medoids coarse clustering optimization on the standardized feature values; 1. K-Medoids selects actual data points as cluster centers (medoids), which is more robust to outliers than K-Means (which uses virtual mean points); 2. In distance calculation, a weight of 3 is applied to the reflectivity dimension. The formula for the weighted distance d_weighted is: d_weighted=sqrt[(x2-x1)²+(y2-y1)²+(z2-z1)²+3×(z2-z1)²]; Where z represents the value of reflectivity; This weighted strategy highlights the core role of reflectivity in cloud identification and enhances the physical meaning of clustering; 3. Setting a fixed number of clusters K=5 for coarse clustering can effectively cover the number of major cloud clusters in typical weather scenarios.

[0046] Step 3: Perform three-dimensional buffer analysis and overlap cluster merging; 1. Construct a three-dimensional buffer with the centroid of each coarse cluster as the center, and set the buffer radius to 0.03 (normalized distance unit). 2. Perform height normalization. Considering the difference in magnitude between height and horizontal distance, the height is normalized in the buffer distance calculation. The formula for the normalized distance d_3D is: d_3D=sqrt[(Δlon)²+(Δlat)²+(Δh / 20000)²]; Where Δh is the height difference, divided by 20,000 meters (typical tropopause height) to achieve uniformity of magnitude; in: d_3D: 3D normalized distance, unit: dimensionless; Δlon: The difference in longitude between two points, in degrees; Δlat: The difference in latitude between two points, in degrees; Δh: The difference in height between two points, in meters; 20000: Typical tropopause height value, used to normalize height differences to the same order of magnitude as latitude and longitude differences.

[0047] 3. Perform overlap judgment and merging: When the number of overlapping points in the buffers of any two coarse clusters exceeds 50, they are determined to be different parts of the same cloud and the cluster merging operation is performed.

[0048] Step 4: Use HDBSCAN for fine purification and output preliminary clustering results; In this embodiment, specifically, based on the spatial characteristics of cloud data, the key parameters of HDBSCAN are set: min_cluster_size=100: Ensures that the smallest cluster identified is statistically significant; Here, min_cluster_size represents the minimum cluster size parameter of the HDBSCAN algorithm, which is set to 100 to ensure that the smallest cloud cluster identified has statistical significance; This parameter value is derived from big data statistical analysis and can balance clustering accuracy and computational efficiency.

[0049] min_samples=30: Balances noise sensitivity with clustering stability; HDBSCAN can effectively identify points with low density located at the boundary of a cloud cluster and mark them as noise. HDBSCAN's hierarchical clustering feature can identify substructures with different densities, making it suitable for handling density variations that may exist within a cloud cluster. min_samples represents the minimum number of samples parameter of the HDBSCAN algorithm. This parameter is used to balance noise sensitivity and clustering stability. The larger the value, the more likely it is to mark points as noise, and the smaller the value, the easier it is to form clusters.

[0050] In this embodiment, as Figure 3 and Figures 5-7 As shown, the density merging optimization module refines the preliminary clustering results output by the cloud clustering module through dynamic K-value selection, two-dimensional merging, and multi-dimensional noise reduction, as detailed below: Step 1: Dynamically select the K value for the preliminary clustering results output by the cloud clustering module; specifically, dynamically select the number of K-Medoid clusters K based on the number of input data points N: When N < 800, K = 3; When 800 ≤ N < 2000, K = 4; When N≥2000, K=5; This strategy is based on big data statistical analysis to ensure optimal clustering results under different data scales; it achieves fully automated parameter selection, significantly improving the ease of use and stability of the system.

[0051] Step 2: Perform a two-dimensional merging of the density correlation dimension and the spatial proximity dimension; 1. Density association dimension: By analyzing the density clustering results of HDBSCAN, clusters belonging to the same density cluster but separated by K-Medoids are identified; 2. Spatial Proximity Dimension: Calculate the spatial relationships between clusters, with the following merging conditions: centroid Euclidean distance less than 0.03 (normalized units) and concavity / convexity difference less than 0.2; Concavity / convexity calculation: The concavity / convexity index C is defined as: C = A_convex / A_actual; Where A_convex is the convex hull area, A_actual is the actual point cloud projection area, and the closer C is to 1, the closer the shape is to a convex shape.

[0052] Step 3: Perform multi-dimensional noise filtering; 1. Remove small clusters with fewer than 20 points, as these are mostly noise or fragmented clouds; 2. Eliminate weak signal clusters with an average reflectance of less than 10 dBZ, as these may be non-precipitating clouds or residual noise; 3. Spatial dispersion filtering: Calculate the spatial dispersion index D of the cluster: D=(1 / n) Σ||x_i-x ||; in: D: Spatial dispersion index, representing the average distance from each point within the cloud to the centroid, unit: dimensionless; n: The number of points within the cluster; x_i: The coordinate vector of the i-th point within the cluster; x : The centroid coordinate vector of the cluster; ||x_i-x ||:From point x_i to the centroid x The Euclidean distance; When D > 0.1, it is judged as an overly dispersed possible noisy cluster.

[0053] Step 4: Establish a multi-dimensional visualization system; 1. Use the Plotly library to generate interactive 3D visualizations that can be rotated, scaled, and selected, supporting different colors to distinguish different cloud clusters, displaying detailed information on mouse hover, and layer control (showing / hiding specific cloud clusters). 2. Detailed statistical dashboard: includes histograms of cloud point distribution and box plots of reflectance statistics; 3. Cartesian coordinate system multi-view diagram: Provides three orthogonal projection views: longitude-latitude, longitude-altitude, and latitude-altitude, to fully display the spatial structure of the cloud cluster.

[0054] Step 5: Output and quality assessment of the second structured dataset; The output types of the second structured dataset include: NPY file: A NumPy array containing all points and their final cluster labels; JSON statistics file: Detailed physical characteristics statistics for each cloud cluster; Parquet files: columnar storage, facilitating big data analysis; Automated quality report output: The system automatically generates a processing quality assessment report, including the proportion of noise points and cluster stability indicators.

[0055] S3. Process the second structured dataset output by S2 and construct a local cloud 3D model.

[0056] Specifically, the processing of the second structured dataset includes data reading and format conversion. In this implementation, the .npy data file output by S2 is converted into a .nc file. The column order in the .npy file is ['longitude', 'latitude', 'altitude (meters)', 'reflectivity factor (dBZ)', 'cloud cluster number']. After creating the corresponding dictionary, the dataset is constructed to achieve data structure conversion.

[0057] In this embodiment, the synergistic relationship of the four methods is as follows: 1) Non-sequential use: These four methods are not strictly executed in the order of Alpha Shape, local convex hull, DBSCAN, and distance threshold, but rather adopt a strategy of "cross-combination and complementary enhancement"; 2) Method crossover: The DBSCAN concept (density connectivity) is mainly used to remove isolated noise points, providing clean data for subsequent boundary extraction; the distance threshold method is used for preliminary boundary point extraction; the Alpha Shape method and the local convex hull method further filter false boundary points through geometric constraints to form the final boundary; 3) Complementary logic: The shortcoming of the distance threshold method is that it cannot distinguish between "peripheral noise points" and "peripheral boundary points". Therefore, isolated noise points are first removed by DBSCAN, and then the boundary is extracted by the distance threshold method. At the same time, the geometric constraints of Alpha Shape and local convex hull method can further filter false boundary points to ensure the accuracy of the results.

[0058] Subsequently, based on the transformed data structure, the boundary points of the cloud body are extracted. Specifically, the following combination of methods is used to construct the local cloud body 3D model: 1. Alpha Shape Method: Generates Delaunay triangulation. The input discrete point set is triangulated using Delaunay triangulation. The discrete point set refers to the point cloud data in the second structured dataset output after classification and processing in step S2. Each point contains coordinates of four core physical quantities: longitude, latitude, altitude, and reflectivity factor. This results in a triangular mesh composed of triangular facets. The core characteristic of Delaunay triangulation is maximizing the minimum interior angle of the triangles, avoiding "elongated triangles," ensuring the uniformity of the triangular mesh, and providing a high-quality basic structure for subsequent selection. Each triangular facet of the Delaunay triangulation is traversed, its circumcircle radius is calculated, and it is determined whether the condition is met: if the condition is met, the triangle is an α-effective facet and is retained; otherwise, it is discarded. All retained α-effective facets are traversed, and the number of times each triangle's edge is shared is counted: if an edge is shared by only one α-effective facet, then that edge is the boundary edge of the Alpha Shape. If an edge is shared by two α-active faces, then that edge is an internal edge and should be removed. Connecting all boundary edges in sequence will form a closed Alpha Shape profile.

[0059] 2. Local Convex Hull Method: The local convex hull method is an optimized extension of the traditional global convex hull method. Its core is to define local windows or regions, calculate the convex hull for each local region of a discrete point set, and then integrate the results of each local convex hull to capture the local convexity and detailed contours of the target. Based on the boundary points selected by the Alpha Shape algorithm, the program retains 95% of the outer points by sorting them by distance from the boundary points to the center of the cloud. During this process, boundary points with similar distances form "local clusters," each of which can be considered a small subset of points. For each local cluster, the default principle of "convexity priority" is followed: if the point set within the local cluster satisfies convexity (no obvious depressions), it is directly retained; if depressions exist (such as boundary breaks caused by local sparsity), the constraint effect of the local convex hull is indirectly achieved through the α-effective patch selection of Alpha Shape (i.e., no other points interfering locally).

[0060] 3. DBSCAN Method: The DBSCAN method defines clusters based on "density connectivity." A cluster is a continuous set of high-density regions. Any two points within a cluster can be connected through a series of high-density points, and the cluster is separated from other clusters by low-density regions (noise areas). In cloud boundary extraction, this idea can be simply understood as: ① The points (high density) in the core region of the cloud cluster are the "core points," which constitute the main body of the cluster; ② The points at the edge of the cloud cluster (with slightly lower density, but still connected to the core point) are "boundary points", that is, the semantic boundary of the cloud cluster; ③ Isolated points far from all high-density areas (such as atmospheric background noise points) are "noise points" and need to be eliminated.

[0061] 4. Distance Threshold Method: This is a simple and intuitive boundary extraction and filtering algorithm based on the spatial distance relationship between points. The core idea is to determine the spatial relationship between a point and the target core area, and between points themselves, by setting a fixed or dynamic distance threshold. This distinguishes between "target boundary points," "internal points," and "external noise points." To address the problems of "difficult threshold setting and poor dimensionality adaptation" in traditional distance threshold methods, the program has undergone targeted optimization. ① From "fixed threshold" to "adaptive threshold" Traditional distance thresholding methods require manually preset fixed thresholds (e.g., 1 km), but cloud scales vary greatly (small-scale clouds have diameters of a few kilometers, while large-scale clouds have diameters of tens of kilometers), making fixed thresholds unsuitable. The program addresses this by: Thresholds are dynamically calculated based on the distance distribution (quantiles) of the point set itself, without the need for manual intervention; The threshold is automatically adjusted according to the cloud size: the distance quantile of large-scale clouds is naturally larger, and the threshold increases accordingly; the distance quantile of small-scale clouds is smaller, and the threshold decreases accordingly, perfectly adapting to clouds of different sizes.

[0062] In this embodiment, the dynamic calculation of the threshold specifically includes the following steps: 1) Dynamic threshold calculation: Based on the distance distribution of the discrete point set itself, the threshold is dynamically calculated using the k nearest neighbor distance quantile method. Specifically, the average distance from each point to its k nearest neighbors is calculated, and the median of the k nearest neighbor distances of all points is taken as the adaptive threshold. 2) 3D Threshold Setting: Based on the traditional 2D distance threshold, it is extended to 3D space. The distance calculation formula is: d_3D=sqrt[(Δlon)²+(Δlat)²+(Δh / 20000)²], where Δlon and Δlat are the difference in latitude and longitude, and Δh is the difference in height (unit: meters). Dividing by 20000 achieves the unification of the order of magnitude of height and horizontal distance. 3) Adaptive characteristics: Large-scale cloud clusters naturally have larger distance quantiles, and the threshold increases accordingly; small-scale cloud clusters have smaller distance quantiles, and the threshold decreases accordingly, thus adapting to cloud clusters of different scales without human intervention.

[0063] ②From "2D Threshold" to "3D Threshold" Traditional distance thresholding methods are mostly used in 2D planes, but clouds are 3D targets, so the program extends them to a 3D scene: Distance calculation uses 3D Euclidean distance, taking into account both latitude and longitude (horizontal direction) and altitude (vertical direction) spatial differences; Specifically, the traditional 2D distance threshold is extended to 3D space. The distance calculation formula is: d_3D=sqrt[(Δlon)²+(Δlat)²+(Δh / 20000)²], where Δlon and Δlat are the difference in latitude and longitude, and Δh is the difference in height, in meters. Dividing by 20000 unifies the magnitude of height and horizontal distance. For example, the bottom edge of a cloud is close to the center in horizontal distance but low in height; the top edge is close to the center in horizontal distance but high in height. 3D distance calculation can identify both types of boundary points simultaneously, avoiding omission of vertical boundaries.

[0064] It should be noted that in this embodiment, the program circumvents the limitations of the distance threshold method by combining three methods: The limitations of the distance threshold method are: it cannot distinguish between "peripheral noise points" and "peripheral boundary points" (if there is isolated noise at a distance, it may be misjudged as a boundary). In this embodiment, the specific implementation of the DBSCAN approach for removing isolated noise points is as follows: 1) Definition of density connectivity: Clusters are defined by the concept of "density connectivity" in the DBSCAN algorithm. A cluster is a continuous set of high-density regions. Any two points in a cluster can be connected through a series of high-density points, and the cluster is separated from other clusters by low-density regions (noise regions). 2) Noise point removal: a) The points (high density) in the core region of the cloud cluster are the "core points," which constitute the main body of the cluster; b) Points at the edge of the cloud (with slightly lower density, but still connected to the core point) are "boundary points", that is, the semantic boundary of the cloud; c) Isolated points far from all high-density areas (such as atmospheric background noise points) are "noise points" with a label of -1 and must be removed; 3) Implementation process: a) Input point cloud data: Each point contains four dimensions: [longitude, latitude, altitude, reflectivity factor]; b) Adaptive calculation of eps parameter: The k-nearest neighbor median method is used, with k=10. The distance from each point to its 10 nearest neighbors is calculated, and the median of all distances is taken as the eps value. c) Perform DBSCAN clustering: Set min_samples=5 to perform density clustering based on Euclidean distance; d) Extract boundary points: Mark points whose labels are not equal to -1 and are not core samples as boundary points; e) Backup mechanism when there are insufficient boundary points: If fewer than 4 boundary points are extracted, 10 points around the core point are added as boundary points.

[0065] 4) Effect: After removing isolated noise points by density connectivity, boundary extraction can effectively avoid misjudging isolated noise far from the cloud as boundary points, thus improving the accuracy of boundary extraction.

[0066] Complementary logic: The program first uses the DBSCAN concept (implicit in cloud region filtering) to remove isolated noise points, then uses the distance threshold method to extract boundaries. At the same time, the geometric constraints of Alpha Shape and local convex hull method can further filter false boundary points to ensure accurate results, and finally complete the construction of the local cloud 3D model.

[0067] In this embodiment, as Figure 8 As shown, the basic modeling framework of "Alpha Shape + Local Convex Hull Method" adapts to different cloud clusters by dynamically adjusting the alpha parameter, balancing contour fit and boundary smoothness. First, the program uses the Alpha Shape algorithm to calculate the global boundary of the cloud's 3D point set. This method is similar to "stretching an elastic membrane" around the point set, which can capture the overall irregular shape of the cloud.

[0068] Alpha Shape controls the "tightness" of the boundary by adjusting the α parameter: the smaller the α value, the tighter the boundary adheres to the surface of the point set, and the more it can identify depressions and voids; the larger the α value, the smoother the boundary, which is close to the convex hull shape.

[0069] The program adaptively adjusts the α value according to the size of the point set: α=0.02 when the number of points exceeds 1000, α=0.015 when the number of points is between 500 and 1000, and α=0.01 when the number of points is small.

[0070] This adaptive strategy ensures a suitable initial boundary regardless of cloud size. To compensate for AlphaShape's shortcomings in detail capture, the program then employs a local convex hull method for layered processing. This method divides the cloud into multiple horizontal layers along the vertical height direction (usually 4-6 layers, dynamically adjusted according to the cloud height range), and calculates the 2D convex hull for each point in each layer individually (considering only longitude and latitude dimensions).

[0071] The advantage of layered processing is that it can independently capture the horizontal extension characteristics of clouds at different altitudes, especially when clouds exhibit different horizontal ranges at different altitudes.

[0072] By merging the boundary point set obtained through layered calculations with the Alpha Shape results to remove duplicates, a dual effect is achieved: first, it repairs potential vertical breaks in the Alpha Shape, ensuring the continuity and integrity of the cloud's sides; second, it supplements local features missed by the Alpha Shape, such as small-scale protrusions or depressions on the cloud's sides. This method is particularly suitable for convective clouds with obvious vertical layering structures, and can more realistically reflect the three-dimensional morphology of the cloud.

[0073] In another embodiment, such as Figure 10 As shown, using the advanced modeling framework of "Alpha Shape + Local Convex Hull Method + Density Method", the core region of the bottom surface of the cloud is first located by the density method (KDE + quantile), and at the same time, the sparse noise points in the global point set are removed by the density threshold. Contour extraction: For the denoised point set, the Alpha Shape method is used to extract the global irregular contours, and the local convex hull method is used to perform block calculations to supplement local details and repair contour breaks; Boundary merging and simplification: Alpha Shape boundary points, local convex hull boundary points, and core points after density filtering are combined and duplicates are removed; Secondary simplification: Calculate the distance from the boundary point to the center of the cloud after merging, retain 95% of the outer points, and further eliminate redundancy; Backup and optimization: If the merged boundary points are insufficient (<4), the original point set is used as a backup; finally, abnormal height points (points that exceed the height range of the cloud body) are filtered out to obtain the final 3D boundary.

[0074] In another embodiment, such as Figure 9 As shown, by using the "Alpha Shape + Local Convex Hull Method + Distance Threshold Method", the Alpha Shape method is first used to calculate the global irregular boundary of the 3D point set of the cloud, and a preliminary boundary point set that fits the shape of the cloud is obtained. Then, local detail completion is performed. The point set is divided into blocks according to height / space, and the boundary points of each block are extracted using the local convex hull method. These blocks are then merged with global boundary points (to remove duplicates) to repair breaks and add details. Next, a dual distance filtering method is used: a "point-to-center" distance threshold is used to remove noise points outside the cloud body; and a "point-to-point" distance threshold is used to streamline redundant points inside the boundary. Specifically, the program finally uses a distance threshold method to refine and filter boundary points, which includes a two-level filtering mechanism: The first layer is the "point to cloud center" distance threshold filtering.

[0075] First, calculate the geometric center (average of longitude, latitude, and altitude) of all 3D points of the cloud cluster, and then calculate the Euclidean distance from each boundary point to that center.

[0076] The program employs an adaptive threshold strategy: it takes the 70th percentile of all distance values ​​as the threshold (which can be adjusted via parameters) and retains only points whose distance is greater than or equal to the threshold.

[0077] The core function of this screening is to eliminate redundant points inside the cloud and noise points outside: internal points are usually closer to the center and are non-boundary redundancies; external noise points may be farther away, but they often exist in isolation and may be misjudged in the previous boundary extraction.

[0078] The second layer is a secondary distance filtering after merging.

[0079] After merging the boundary points from the Alpha Shape, local convex hull, and distance threshold methods, the program recalculates the distance from these points to the center of the cloud, sorts them in descending order of distance, and retains only the outermost 95% of the points (the retention ratio is adjustable).

[0080] This "survival of the fittest" mechanism implicitly uses "point-to-point" distance filtering: retaining points with large distances naturally eliminates relatively close points, further simplifying the boundary point set, while ensuring that the points that best represent the outer contour of the cloud are retained.

[0081] In this embodiment, different embodiments include: 1) Basic implementation method: such as Figure 8 As shown, a combined framework of "Alpha Shape + Local Convex Hull Method" is adopted. By dynamically adjusting the alpha parameter, different cloud patterns are adapted to balance the contour fit and boundary smoothness. 2) Advanced implementation methods: such as Figure 9 As shown, a combined framework of "Alpha Shape + Local Convex Hull Method + Distance Threshold Method" is adopted. First, the Alpha Shape method is used to extract the global irregular boundary, then the Local Convex Hull method is used to supplement the details, and finally the Distance Threshold Method is used for double filtering. 3) Optimal implementation method: such as Figure 10 As shown, a combined framework of "Alpha Shape + Local Convex Hull Method + Density Method" is adopted. First, the density method is used to locate the core area of ​​the cloud and remove global noise points. Then, the Alpha Shape method and the local convex hull method are used to extract and supplement the boundary. Finally, the boundary is merged and simplified to obtain the final 3D boundary. 4) Implementation method selection criteria: Different combination methods are selected according to the complexity of the cloud. When the cloud shape is simple, the basic framework is used, and when the cloud shape is complex or there is a lot of noise, the advanced or optimal framework is used.

[0082] Final optimization and fallback: Merge the filtered boundary points and calculate the final 3D boundary; if there are insufficient boundary points (<4), degenerate to global convex hull method + distance threshold filtering to ensure program stability; output result: obtain a cloud 3D boundary point set with "global fit + local refinement + noise-free" for subsequent 3D boundary surface drawing.

[0083] In another embodiment, cloud 3D modeling relies on technologies such as lidar and weather radar to acquire point cloud data, creating rich application value in multiple fields: 1. In the fields of transportation infrastructure and geographic surveying; By leveraging point cloud modeling with vehicle-mounted and airborne LiDAR, precise road inspections can be conducted to promptly identify potential hazards such as road surface damage and deformation; 3D scanning and monitoring of bridges can help assess structural safety; in geographic surveying, high-precision 3D models can be built for different terrain scenarios such as urban areas and mountainous areas, making up for the lack of information collection in complex terrain and hidden areas in traditional surveying, improving the accuracy and timeliness of geographic information data, and providing precise spatial data support for transportation planning and infrastructure construction layout.

[0084] 2. In the field of industrial manufacturing and testing; In industrial manufacturing, processing airborne and vehicle-mounted LiDAR point clouds can accurately identify surface defects and component misalignment on equipment, assist in power transmission line inspection, and detect problems such as line wear and tower tilt in advance, ensuring the safety of industrial production and power transmission. By monitoring the entire life cycle of industrial equipment through point cloud data, it can help enterprises optimize maintenance plans, reduce equipment failure losses, and improve production efficiency and product quality control capabilities.

[0085] 3. Within the scope of agricultural and forestry resource management; In agricultural remote sensing, UAV lidar point cloud modeling assists in precise fertilization and pest and disease early warning. Based on crop growth monitoring data, it enables variable input and scientific prevention and control, improving the precision of agricultural production. During forestry resource surveys, it can quickly count the number of trees and tree height / diameter at breast height. Based on airborne point cloud classification, it can accurately grasp the distribution of forest resources and vegetation structure, providing data support for forest cultivation, ecological protection, and forestry industry planning, and helping to promote the coordinated development of ecological and economic benefits.

[0086] 4. In terms of meteorological services and disaster prevention and mitigation; Weather radar point cloud data drives 3D cloud modeling. For the general public, it can improve their daily life and travel experience by reconstructing cloud dynamics (such as predicting the impact of severe weather on travel routes). It can also provide disaster avoidance guidance based on cloud models when disasters occur. For forecasters, it enables accurate reconstruction of 3D models and fusion of multi-source data to improve the accuracy of weather forecasts. For early warning, it accurately locates disaster-affected areas based on 3D cloud dynamic monitoring, buying time for disaster prevention and mitigation decisions and emergency response deployment. It can also provide related meteorological services to other fields such as water conservancy and agriculture, strengthening cross-sectoral disaster prevention and control capabilities.

[0087] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for constructing a local cloud 3D model based on radar data, characterized in that, Includes the following steps: S1. Preprocess the basic information data of the radar site and output the first structured dataset; S2. Classify and process the first structured dataset obtained in S1, and output the second structured dataset; S3. Read and convert the second structured dataset output by S2. Based on the converted data structure, remove isolated noise points using the DBSCAN method, extract the boundary using the distance threshold method, and further filter false boundary points using the geometric constraints of the Alpha Shape method and the local convex hull method to complete the construction of the local cloud 3D model.

2. The method for constructing a local cloud 3D model based on radar data according to claim 1, characterized in that, S3 includes the following steps: S301. Define cloud clusters of different shapes using the density connectivity method of DBSCAN; S302. Use the distance threshold method to extract the cloud boundary of the cluster defined by the DBSCAN method, and obtain the discrete point set of the cloud boundary. S303. The boundaries extracted in S302 are filtered using the Alpha Shape method to obtain the preliminary cloud boundary contour. S304. The preliminary boundary contour selected by the Alpha Shape method is divided into local regions using the local convex hull method, and the convex hull is calculated for each local region. Small subsets of points within the local region that satisfy convexity are directly retained. If there are depressions, the Alpha Shape method is used to further filter the data, and S303 is executed again. This process continues until the construction of the local cloud 3D model is completed.

3. The method for constructing a local cloud 3D model based on radar data according to claim 2, characterized in that, In S302, the distance threshold method extracts cloud boundaries by setting fixed or dynamic distance thresholds to determine the spatial relationship between points and the core area of ​​the target and / or between points. Among them, the threshold is dynamically calculated based on the distance distribution of the discrete point set itself; at the same time, the cloud boundary is extracted through 3D threshold.

4. The method for constructing a local cloud 3D model based on radar data according to claim 2, characterized in that, In S303, the discrete point set obtained by the distance threshold method is triangulated by Delaunay triangulation, and each triangular facet of the Delaunay triangulation is traversed to calculate its circumcircle radius. Triangular facets that meet the conditions are α-valid facets and are retained; If the condition is not met, the triangular facet is discarded. Then iterate through all the retained α-valid faces and count the number of times each triangle edge is shared: if an edge is shared by only 1 α-valid face, then the edge is the boundary edge of the Alpha Shape; If an edge is shared by at least two α-effective patches, then the edge is an internal edge and is removed. All boundary edges are connected in sequence to form a closed Alpha Shape contour, thus obtaining the preliminary cloud boundary contour.

5. The method for constructing a local cloud 3D model based on radar data according to claim 2, characterized in that, In S304, the local convex hull method sorts the distances from the boundary points to the center of the cloud and retains b% of the outer points, where b% represents the percentage of values. Several boundary points with a distance less than a set threshold form a local cluster, and each local cluster is regarded as a small subset of points.

6. The method for constructing a local cloud 3D model based on radar data according to claim 1, characterized in that, S1 specifically includes the following steps: S101. Read and analyze the basic information data of the radar site; S102. Convert the polar coordinates of each data point in the radar volume scan data to station-centered rectangular coordinates; S103. Convert the station-centered rectangular coordinates of each data point obtained in S102 to geocentric and geofixed coordinates; S104. Convert the geocentric coordinates of each data point obtained in S103 to geodetic coordinates; S105. Detect and filter abnormal data, and output the first structured dataset.

7. The method for constructing a local cloud 3D model based on radar data according to claim 1, characterized in that, S2 specifically includes the following steps: S201. Use the data filtering module to classify, filter and lightweight the first structured dataset, and output valid cloud data. S202. Use the cloud clustering module to perform preliminary clustering on the effective cloud data processed by the data filtering module, and output the preliminary clustering results. S203. The density merging optimization module is used to refine the effective cloud clusters by dynamically selecting the K value, performing two-dimensional merging, and multi-dimensional noise reduction on the preliminary clustering results output by the cloud clustering module.

8. The method for constructing a local cloud 3D model based on radar data according to claim 7, characterized in that, In S201, a keyword fuzzy matching algorithm is used to identify radar data column names from different sources and in different formats, and to map columns containing specific keywords to standard column names. Valid cloud data is filtered by setting a reflectivity threshold, and spatial consistency checks are performed on the valid cloud data that has passed the reflectivity threshold to remove outliers that are outside the physical range. By retaining core fields and extracting redundant information, effective cloud data optimization and lightweight processing are performed, and a standardized format is output.

9. The method for constructing a local cloud 3D model based on radar data according to claim 7, characterized in that, In S202, cloud reflectivity is truncated by setting a threshold for the fractional range to remove outlier values. The effective cloud data is Z-score standardized, and the standardized feature values ​​are then optimized using K-Medoids coarse clustering. Three-dimensional buffer analysis and overlapping cluster merging were performed on the effective cloud data after K-Medoids coarse clustering optimization. Then, HDBSCAN is used for purification to output preliminary clustering results.

10. The method for constructing a local cloud 3D model based on radar data according to claim 7, characterized in that, S203 includes the following steps: S203-1. Dynamically select the number of clusters K for K-Medoids based on the number N of data points in the preliminary clustering results output by the cloud clustering module; S203-2, Perform a two-dimensional merging of density correlation dimension and spatial proximity dimension; S203-3, Perform multi-dimensional noise filtering; S203-4. Organize the data filtered in S203-3 and output the second structured dataset.

Images