Method and product of urban area radar rainfall correction by fusing raindrop dynamic process simulation and machine learning

Abstract

The embodiment of the application discloses a city area radar rainfall correction method and product fusing raindrop dynamic process simulation and machine learning, relates to the meteorological radar data processing technical field, and provides the scheme in the application, on one hand, aiming at the complex underlying surface and atmospheric environment in the city, three-dimensional atmospheric field data and dual-polarization radar data are used, the drift process of raindrops in the atmosphere is simulated, high-resolution three-dimensional atmospheric field and raindrop size distribution (DSD) data are combined, and the radar-gauge rainfall deviation caused by raindrop drift is corrected; on the other hand, the random forest (RF) method of machine learning is introduced, radar observation data, gauge measurement values and related atmospheric influence factors are taken as inputs, a non-linear mapping model is constructed to mine the complex functional relationship between the two, and the deviation of the radar-gauge pair under the city environment is directly calculated and reduced. Finally, the results of the physical correction method and the random forest method are weighted and fused, and the rainfall difference between the radar and the gauge is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of meteorological radar data processing technology, and in particular to a method and product for urban area radar rainfall correction that integrates raindrop dynamic process simulation and machine learning. Background Technology

[0002] Modern weather radars are capable of estimating instantaneous precipitation over large areas with high spatial and temporal resolution and are widely used in hydrology and meteorology. However, due to the fluctuating atmospheric environment, radar precipitation estimates are subject to significant uncertainties.

[0003] Widely recognized errors in radar rainfall measurement include ground clutter and anomalous propagation, signal attenuation, beam blocking, ZR relationship parameterization, and vertical variations in the measured radar reflectivity. These errors can be physically adjusted step-by-step using different radar quality control algorithms. Furthermore, leveraging the advantages of rain gauge measurements to perform additional statistical adjustments to radar rainfall data can achieve a more reliable "ground reality" for radar rainfall estimation. In radar-rain gauge rainfall comparisons, discrepancies are often attributed to radar rainfall measurements, neglecting the fact that radar observes high-altitude rainfall while rain gauges capture ground-level raindrops. In other words, rainfall observed by high-altitude radar is assumed to fall vertically to the ground and should correspond to ground-level rainfall below the volume captured by the radar beam. In reality, raindrops drift in the atmosphere due to wind as they fall. Based on this, raindrop drift has a significant impact on the quality of radar data after correction in rain gauge measurements. This is mainly reflected in the fact that wind causes raindrop drift, resulting in inconsistent spatial correlation between the radar and the rain gauge. This is known as the wind drift effect. Raindrop drift can cause large errors over a considerable distance, even under special circumstances (even with a resolution of 2 km).

[0004] In existing technologies, radar precipitation correction primarily relies on statistical methods, neglecting the microphysical processes of raindrops. While some studies have proposed methods that consider wind drift, these methods typically only account for a single factor and lack accurate simulations of three-dimensional wind fields and raindrop size distributions. Only a few studies have simulated raindrop evolution in spatially varying atmospheric environments. More importantly, a unified framework is still lacking to simulate the overall processes of wind drift and precipitation, and to achieve quantitative precipitation estimation using radar polarization methods. Summary of the Invention

[0005] This invention provides a method and product for urban area radar rainfall correction that integrates raindrop dynamics simulation and machine learning, in order to at least partially solve the above-mentioned problems.

[0006] The first aspect of this invention provides a method for correcting urban area radar rainfall by integrating raindrop dynamics simulation and machine learning, the method comprising:

[0007] Construct a multi-source fusion dataset of urban rainfall and preprocess the dataset.

[0008] The space between the radar scanning layer and the ground layer is divided into multiple vertical layers, and each layer is divided into multiple horizontal square grids, resulting in a space composed of multiple three-dimensional subspaces. This allows us to set the initial diameter and initial position of the raindrops as the initial conditions for simulating the raindrop trajectory.

[0009] Based on the division of multiple subspaces and the initial condition settings, the trajectory of raindrops in the current subspace between the ground and radar waves is simulated, and the displacement within each time step is accumulated in real time until the raindrops move out of the current subspace.

[0010] After the raindrop enters the new subspace, the simulation process of the raindrop's trajectory continues to iterate in the new subspace until the raindrop reaches the ground. The spatial relationship between the radar and the rain gauge is obtained, and the error caused by the wind drift effect is corrected. This is recorded as the physical correction result.

[0011] Based on a pre-built random forest model, radar rainfall data at each time step, and three-dimensional atmospheric field data, the predicted rain gauge rainfall at each time step is determined. The three-dimensional atmospheric field data includes atmospheric data at the altitude where the radar scan is located and atmospheric data at the near-surface level. The random forest model has pre-learned the correspondence between radar rainfall data and rain gauge measurement data, which is denoted as the random forest correction result.

[0012] The physical correction results and the random forest correction results are weighted and fused to obtain the comprehensive correction coefficient;

[0013] A spatial distribution surface of rainfall based on rain gauge and radar observation data is constructed using interpolation methods to obtain the rainfall measured by rain gauges and radar. Based on the spatial relationship, the rainfall measured by rain gauges, the rainfall measured by radar, the radar-rain gauge deviation, and the comprehensive correction coefficient, a radar rainfall deviation function is constructed. The radar rainfall measured by radar is adjusted based on this function, ultimately achieving high-precision correction of radar rainfall in urban areas.

[0014] Optionally, a multi-source fusion dataset of urban rainfall is constructed, and the dataset is preprocessed, including:

[0015] A WRF model was constructed by selecting multiple rainfall events with different time ranges, rainfall intensities, and seasons.

[0016] Acquire radar observation data, rain gauge observation data, and meteorological data to construct a multi-source fusion dataset of urban rainfall;

[0017] The urban rainfall multi-source fusion dataset was subjected to quality control, consistency checks, and correction of outliers.

[0018] Optionally, the space between the radar scanning layer and the ground layer is divided into multiple vertical layers, each layer being divided into multiple horizontal square grids, resulting in a space composed of multiple three-dimensional subspaces, thereby setting the initial diameter and initial position of the raindrops, including:

[0019] The space between the ground and radar waves is divided into multiple vertical layers, and each vertical layer is divided into multiple horizontal square grids. The horizontal and vertical resolutions of the subspaces are configured to be consistent with the resolution of the atmospheric field data used, ensuring that the atmospheric parameters are accurately matched with the subspace scale.

[0020] The initial diameter of the raindrop is obtained by estimating the diameter of a single radar pixel using a normalized gamma distribution.

[0021] The original radar pixels of the raindrop are divided into four sub-grids. The initial horizontal coordinate of the raindrop is taken as the center point, and the radar beam height corresponding to that point is taken as the initial height. The initial position of the raindrop is then obtained.

[0022] Optionally, based on the division of multiple subspaces and the initial condition settings, the trajectory of raindrops within the current subspace between the ground and radar waves is simulated, and the displacement within each time step is accumulated in real time until the raindrops move out of the current subspace, including:

[0023] The trajectory of raindrops in each layer and subspace of the ground and radar wave space is simulated by solving the particle motion equations.

[0024] Simulate the trajectory of raindrops in each layer and subspace of the ground and radar wave space within each time step. At the end of the time step, obtain the new position and velocity of the raindrops for the next time step.

[0025] All time steps x and y The displacements in each direction are accumulated and iterated repeatedly until the raindrop moves out of the subspace.

[0026] Optionally, after the raindrop enters the new subspace, the simulation process of the raindrop's trajectory continues iteratively in the new subspace until the raindrop reaches the ground, obtaining the spatial relationship between the radar and the rain gauge, and correcting errors caused by wind drift effects, including:

[0027] The trajectory of the raindrop is simulated in a new subspace and iterated repeatedly until the raindrop reaches the ground. The final position coordinates of the raindrop on the ground observed by radar are obtained, and the final position coordinates are used as the corrected raindrop point.

[0028] For each rain gauge, a group of corrected raindrop points around it is searched and the corrected raindrop point closest to the rain gauge is selected. The original radar pixel of the corrected raindrop point is paired with the rain gauge to obtain an adjusted radar-rain gauge pair. The radar pixel constructed based on radar beam projection is paired with the rain gauge to obtain an unadjusted radar-rain gauge pair.

[0029] Optionally, the random forest model is constructed and trained based on the following steps:

[0030] Radar rainfall data and three-dimensional atmospheric field data from rainfall events are selected as input features, and rain gauge measurement data at the corresponding time are used as output labels.

[0031] The input features are standardized, and the training set and validation set are divided using the 5-fold cross-validation method. The core parameters of the random forest model are set as follows: the number of decision trees is 100-500, the maximum depth of the decision trees is 10-30 layers, the minimum number of split samples for a node is 5-20, and the minimum number of samples for a leaf node is 1-5.

[0032] The random forest model is trained using the training set, and its performance is evaluated and parameters are optimized using the validation set until the model's prediction error meets a preset threshold.

[0033] Optionally, the physical correction results and the random forest correction results are weighted and fused to obtain a comprehensive correction coefficient, including:

[0034] The trajectory of raindrops is simulated in a new subspace, iterated repeatedly until the raindrops reach the ground, obtaining the final position coordinates of the raindrops on the ground as observed by radar. These final position coordinates are used as the corrected raindrop point for physical correction, and the physical correction rainfall corresponding to the original radar pixel at this corrected raindrop point is extracted. P raw Simultaneously calculate the physical correction coefficient. C phys ;

[0035] Based on the random forest model, the radar rainfall data and three-dimensional atmospheric field data of the corresponding original radar pixel are input, the predicted rainfall from the rain gauge is output, the measured rainfall from the rain gauge is obtained, and the random forest correction coefficient is calculated. C rf ;

[0036] For physical correction coefficients C phys With Random Forest Correction Coefficient C rf The comprehensive correction coefficient corresponding to the corrected raindrop point is obtained by direct averaging. C :

[0037]

[0038] in, C The range of values ​​varies C phys , C rf Synchronous changes ensure that the correction intensity is consistent with the trend of the results from both methods.

[0039] Optionally, based on the spatial relationship, rainfall measured by rain gauge, rainfall measured by radar, and radar-rain gauge deviation and comprehensive correction coefficient, a radar rainfall deviation function is constructed, including:

[0040] For each rain gauge, search for the surrounding group of corrected raindrop points and select the corrected raindrop point closest to the rain gauge. Pair the original radar pixel of the corrected raindrop point with the rain gauge to obtain an adjusted radar-rain gauge pair. Pair the radar pixel constructed based on radar beam projection with the rain gauge to obtain an unadjusted radar-rain gauge pair.

[0041] The rainfall amount of raindrops in each pixel is adjusted with a three-dimensional matrix in projection space and time. Based on the function for calculating radar rainfall deviation, the radar measured rainfall amount and location are adjusted, resulting in the corresponding corrected radar rainfall amount. P adj for: .

[0042] A second aspect of this invention provides an urban area radar rainfall correction device that integrates raindrop dynamics simulation and machine learning, the device comprising:

[0043] The dataset construction module is used to build a multi-source fusion dataset of urban rainfall and to preprocess the multi-source fusion dataset of urban rainfall.

[0044] The initial condition setting module is used to divide the space between the radar scanning layer and the ground layer into multiple vertical layers, each layer into multiple horizontal square grids, resulting in a space composed of multiple three-dimensional subspaces, thereby setting the initial diameter and initial position of the raindrops as the initial condition settings for the simulation of the raindrop trajectory.

[0045] The simulation module is used to simulate the trajectory of raindrops in the current subspace between the ground and radar waves based on the division of multiple subspaces and the initial condition settings, and to accumulate the displacement in each time step in real time until the raindrops move out of the current subspace.

[0046] The physical correction module is used to continue iterating the simulation process of raindrop trajectory in the new subspace after the raindrop enters the new subspace until the raindrop reaches the ground, obtain the spatial relationship between the radar and the rain gauge, correct the error caused by the wind drift effect, and record it as the physical correction result.

[0047] The random forest correction module is used to determine the predicted rain gauge rainfall at each time step based on a pre-built random forest model, radar rainfall data at each time step, and three-dimensional atmospheric field data. The three-dimensional atmospheric field data includes atmospheric data at the altitude where the radar scan is located and atmospheric data at the near-surface level. The random forest model has pre-learned the correspondence between radar rainfall data and rain gauge measurement data, which is denoted as the random forest correction result.

[0048] The fusion module is used to perform weighted fusion of physical correction results and random forest correction results to obtain a comprehensive correction coefficient.

[0049] The adjustment module is used to construct a spatial distribution surface of rainfall based on rain gauge observation data and radar observation data through interpolation methods, thereby obtaining the rainfall measured by the rain gauge and the rainfall measured by the radar. Based on the spatial relationship, the rainfall measured by the rain gauge, the rainfall measured by the radar, the radar-rain gauge deviation, and the comprehensive correction coefficient, a radar rainfall deviation function is constructed. Based on this function, the radar rainfall measurement is adjusted to achieve high-precision correction of radar rainfall in urban areas.

[0050] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes an urban area radar rainfall correction method that integrates raindrop dynamics simulation and machine learning as described in the first aspect of the present invention.

[0051] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the urban area radar rainfall correction method that integrates raindrop dynamics simulation and machine learning as described in the first aspect of the present invention.

[0052] A fifth aspect of the present invention provides a computer program product, including a computer program / instructions, which are implemented by a processor as the steps in the urban area radar rainfall correction method that integrates raindrop dynamics simulation and machine learning as described in the first aspect of the present invention.

[0053] This invention discloses a method for adjusting the radar-rain gauge precipitation deviation caused by raindrop drift using three-dimensional wind field data and dual-polarized radar data. This method simulates the drift process of raindrops in the atmosphere and combines high-resolution three-dimensional wind field and raindrop size distribution (DSD) data to correct the precipitation observed by radar, thereby reducing the difference in precipitation between the radar and the rain gauge.

[0054] Specifically, this invention, by introducing three-dimensional wind field data, accurately simulates the drift phenomenon of raindrops in the atmosphere, significantly reducing the difference in rainfall between radar and rain gauges (under strong wind conditions, the wind drift effect can cause raindrops to drift several kilometers). By correcting for this phenomenon, this invention significantly improves the accuracy of rainfall estimation.

[0055] In this embodiment of the invention, high-resolution three-dimensional wind field data can be generated using weather research and forecasting models, and combined with raindrop size distribution data acquired by dual-polarization radar, to accurately simulate the trajectory and size evolution of raindrops in the atmosphere. This method is particularly effective under complex terrain and strong wind conditions, and can more accurately reflect the actual movement process of raindrops, thereby significantly improving the reliability of radar rainfall estimation.

[0056] This invention presents, for the first time, a unified correction method that takes into account wind drift. By integrating measured raindrop spectrometer data and dual-polarization radar data, it achieves polarization correction for radar quantitative precipitation estimation. This method not only simplifies the correction process but also significantly improves correction efficiency. It is applicable to various meteorological conditions and has broad application prospects.

[0057] This invention also utilizes a machine learning random forest method to construct a radar-rain gauge rainfall data model. It studies the complex functional relationship between radar data and rain gauge measurements, calculating the bias between the radar and rain gauge pairs. Finally, it performs a weighted fusion of the results obtained from the physical method and the random forest method, thereby reducing the rainfall discrepancy between radar and rain gauge readings.

[0058] In summary, in this embodiment of the invention, on the one hand, considering the complex underlying surface and atmospheric environment of cities, three-dimensional atmospheric field data and dual-polarized radar data are used to simulate the drift process of raindrops in the atmosphere. This is combined with high-resolution three-dimensional atmospheric field and raindrop size distribution (DSD) data to correct the radar-rain gauge precipitation deviation caused by raindrop drift. On the other hand, a machine learning random forest (RF) method is introduced. Radar observation data, rain gauge measurements, and relevant atmospheric influencing factors are used as inputs to construct a nonlinear mapping model to uncover the complex functional relationship between the two, directly calculating and reducing the deviation of the radar-rain gauge pair in urban environments. Finally, the results of the physical correction method and the random forest method are weighted and fused to effectively reduce the precipitation difference between the radar and rain gauge. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. 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.

[0060] Figure 1 This is a flowchart illustrating the steps of the urban area radar rainfall correction method that integrates raindrop dynamics simulation and machine learning, as provided in this embodiment of the invention.

[0061] Figure 2 This is a schematic diagram of the surface wind and rainfall results before and after correction for a rainfall event in an exemplary embodiment of the urban area radar rainfall correction method that integrates raindrop dynamic process simulation and machine learning provided in this invention.

[0062] Figure 3 This is a schematic diagram of the surface wind and rainfall results before and after correction for another rainfall event in the urban area radar rainfall correction method that integrates raindrop dynamic process simulation and machine learning provided in the embodiments of the present invention. Detailed Implementation

[0063] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] like Figure 1 The diagram illustrates a flowchart of the urban area radar rainfall correction method that integrates raindrop dynamics simulation and machine learning, as provided in an embodiment of the present invention. The method includes the following steps:

[0065] S101, Construct a multi-source fusion dataset of urban rainfall and preprocess the multi-source fusion dataset of urban rainfall.

[0066] In this embodiment of the invention, the data in the urban rainfall multi-source fusion dataset comes from radar observations, measured surface rainfall data, and meteorological data.

[0067] In this invention, radar observation data obtained from radar scanning is used to acquire state data of raindrops in the upper atmosphere; measured surface rainfall data is obtained through rain gauges and used for comparison with radar rainfall calculation results. Furthermore, meteorological data is used to improve the consistency between radar and instrument data and to support the supplementation of the three-dimensional atmospheric field. This invention comprehensively considers the process of raindrops drifting from the upper atmosphere to the ground, establishing a more realistic radar rainfall correction method, which is an important component of processing radar data and handling fundamental rainfall-related processes.

[0068] Preprocessing operations include: quality control, consistency checks, and correction of outliers.

[0069] S102 divides the space between the radar scanning layer and the ground layer into multiple vertical layers, each layer into multiple horizontal square grids, resulting in a space composed of multiple three-dimensional subspaces, thereby setting the initial diameter and initial position of the raindrops as the initial conditions for simulating the raindrop trajectory.

[0070] In this embodiment of the invention, when dividing the space between radar waves and the ground, it can be divided according to the horizontal direction and the vertical direction respectively to obtain multiple grid subspaces.

[0071] S103, based on the division of multiple subspaces and the initial condition settings, simulate the trajectory of raindrops in the current subspace between the ground and radar waves, and accumulate the displacement in each time step in real time until the raindrops move out of the current subspace.

[0072] In this embodiment of the invention, the trajectory of raindrops in each layer and subspace of the ground and radar wave space is simulated by solving the particle motion equation.

[0073] S104. After the raindrop enters the new subspace, the simulation process of the raindrop's trajectory continues to iterate in the new subspace until the raindrop reaches the ground. The spatial relationship between the radar and the rain gauge is obtained, and the error caused by the wind drift effect is corrected. This is recorded as the physical correction result.

[0074] In this embodiment of the invention, the rain gauge closest to the final position of the raindrop reaching the ground is paired with the radar pixel of the raindrop. The radar pixel and the paired rain gauge are regarded as an adjusted radar-rain gauge pair, thereby achieving spatial matching of the radar-rain gauge pair.

[0075] S105. Based on the pre-built random forest model, radar rainfall data at each time step, and three-dimensional atmospheric field data, determine the predicted rain gauge rainfall at each time step. The three-dimensional atmospheric field data includes atmospheric data at the altitude of the radar scan and atmospheric data at the near-surface level. The random forest model has pre-learned the correspondence between radar rainfall data and rain gauge measurement data, which is denoted as the physical correction result.

[0076] In this embodiment of the invention, considering that atmospheric fields such as pressure, temperature and humidity fields may cause changes in the mass of raindrops, random forest machine learning is used to discover the complex functional relationship between radar data and rain gauge measurements.

[0077] In this embodiment of the invention, in step S105, the random forest model is constructed and trained based on the following steps:

[0078] S1 selects radar rainfall data and three-dimensional atmospheric field data in the rainfall event as input features, and rain gauge measurement data at the corresponding time as output labels.

[0079] S2. Standardize the input features and use a 5-fold cross-validation method to divide the training and validation sets. Set the core parameters of the random forest model: the number of decision trees is 100-500, the maximum depth of the decision trees is 10-30 layers, the minimum number of split samples for a node is 5-20, and the minimum number of samples for a leaf node is 1-5.

[0080] S3 uses the training set to train the random forest model, evaluates the model performance using the validation set, and optimizes the parameters until the model's prediction error meets the preset threshold.

[0081] Specifically, the model was built using radar and raindrop spectral rainfall data:

[0082] (1) The correlation was analyzed using rainfall events at each rain gauge point, and the Spearman rank correlation coefficient was used to evaluate the ordinal association between the differences and the atmospheric field. The atmospheric field values ​​included pressure, temperature, relative humidity, and horizontal and vertical wind values.

[0083] (2) A random forest model is adopted, with radar rainfall and three-dimensional atmospheric field (pressure, temperature, relative humidity, horizontal and vertical wind) at each time step as input. Through ensemble learning of multiple decision trees (randomly sampled samples and features, single-tree rule prediction and then averaged), the rain gauge rainfall at the corresponding time step is output. In order to simplify the model structure and improve the computational efficiency, the embodiments of the present invention do not use complete three-dimensional atmospheric field data, but only select the atmospheric state at two key levels, namely the height layer where the radar scan is located and the near-surface layer, which is consistent with the physical method.

[0084] (3) Error correction: The performance of the proposed random forest model was evaluated using test point data that were not involved in the modeling.

[0085] The random forest model used in this embodiment of the invention consists of an ensemble of multiple decision trees. The model inputs radar precipitation data and three-dimensional atmospheric field features for each time step, and the output is the instrument precipitation for the corresponding time step. The model uses a bootstrap sampling method to randomly generate multiple sample subsets from the training set, constructing an independent decision tree for each subset. During the splitting process, each tree only randomly selects a subset of features (such as radar reflectivity and atmospheric temperature vertical gradient) to calculate the optimal split point, thereby reducing the risk of overfitting and improving generalization ability. The specific formulas related to the random forest are as follows:

[0086] 1. Bootstrap Sample Sampling Formula

[0087] Let the original training set be D={(X1,y1),(X2,y2),……,(X N ,y N )}, where X i={(X i1 ,y i1 ),(X i2 ,y i2 ),……,(X iN ,y iN )} is the feature vector of the i-th sample (containing radar data and three-dimensional atmospheric field features), y i Let be the instrument rainfall label for the i-th sample, N be the total number of samples, and d be the feature dimension. K sample subsets D are generated through sampling with replacement. K (Corresponding to K decision trees), the number of samples in each subset is the same as the original training set: D k ={(X i1 ,y i1 ),(X i2 ,y i2 ),……,(X iN ,y iN )}, ij∈[1,N], k=1,2,……,K, where: K is the number of decision trees, and ij is the sample index of random sampling (repetition is allowed).

[0088] 2. Decision Tree Feature Splitting Rules

[0089] For the k-th decision tree, in the feature subset F k In {1,2,…,d}(|Fk|=m, m=0.6d), select feature f F k The current node's sample set S is compared with the splitting threshold t. D k Divide into left child node S L , with right child node S R The splitting objective is to minimize the mean square error (MSE) of rainfall in samples within a node:

[0090]

[0091] The mean square error of nodal rainfall is defined as follows:

[0092]

[0093] In the formula: Let |S| be the average rainfall of the samples within node S. L |、|S R | represents the number of samples for the left and right child nodes, respectively.

[0094] 3. Random Forest Ensemble Prediction Formula

[0095] Let h be the rainfall prediction value of the k-th decision tree for the input sample X. kIf (X), then the final predicted value H(X) of the random forest model (i.e., the output metered rainfall) is the arithmetic mean of all decision tree predictions:

[0096]

[0097] In the formula: H(X) is the ensemble prediction result (target output) of the random forest, h k (X) represents the predicted value of a single decision tree, and K represents the total number of decision trees (optimally set to 200).

[0098] The prediction results of the random forest are obtained by integrating the outputs of all decision trees through voting (classification task) or averaging (regression task). For the rainfall regression problem in this embodiment of the invention, a single decision tree generates prediction rules by recursively partitioning the input feature space. For example, when the radar reflectivity ZH > 45 dBZ and the near-surface humidity > 85%, the rainfall at that time step is determined to fall into a certain interval. Finally, the average of the predictions from all decision trees is taken as the metered rainfall output by the model. In the model structure, hyperparameters such as the number of decision trees, maximum depth, and feature sampling ratio are optimized and determined through 5-fold cross-validation.

[0099] S106. The physical correction results and the random forest correction results are weighted and fused to obtain the comprehensive correction coefficient.

[0100] In this embodiment of the invention, a radar-rain gauge rainfall data model is constructed using the machine learning random forest (RF) method. The complex functional relationship between radar data and rain gauge measurements is studied, and the bias of the radar-rain gauge pair is calculated and reduced. Finally, the bias of the radar-rain gauge pair obtained based on the machine learning random forest and the bias obtained based on physical methods (numerical simulation of wind drift to reduce discrepancies) are weighted and fused to obtain a comprehensive correction coefficient for correcting radar rainfall.

[0101] S107. A spatial distribution surface of rainfall based on rain gauge observation data and radar observation data is constructed by interpolation method to obtain the rainfall measured by rain gauge and the rainfall measured by radar. Based on the spatial relationship, the rainfall measured by rain gauge, the rainfall measured by radar, the radar-rain gauge deviation and the comprehensive correction coefficient, a function of radar rainfall deviation is constructed. Based on this function, the radar measured rainfall is adjusted to achieve high-precision correction of radar rainfall in urban areas.

[0102] In this embodiment of the invention, the rain gauge observation data and radar observation data are processed separately. The rainfall measured by the rain gauge and the rainfall measured by the radar can be obtained. Further, based on the spatial relationship between the radar and the rain gauge obtained in step S104, a function for calculating the radar rainfall deviation is calculated; this function can also be called the radar correction coefficient, to adjust the radar-measured rainfall.

[0103] In this embodiment of the invention, step S101 includes the following sub-steps:

[0104] S1011, select multiple rainfall events with different time ranges, different precipitation intensities and different seasons to construct a WRF model;

[0105] S1012: Acquire radar observation data, rain gauge observation data, and meteorological data to construct a multi-source fusion dataset of urban rainfall;

[0106] S1013, perform quality control, consistency check and correction of outliers on the multi-source fusion dataset of urban rainfall.

[0107] In this embodiment of the invention, meteorological data is used to improve the consistency between radar and instrument data and to support the construction of a three-dimensional atmospheric field. Quality control and consistency checks are performed on the radar and rain gauge data, and rainfall case events are selected to ensure the representativeness of the analysis and accurate matching with atmospheric parameters.

[0108] In this embodiment of the invention, radar observation data, rain gauge observation data, and meteorological data all fully cover the selected rainfall events.

[0109] In this embodiment of the invention, step S102 includes the following sub-steps:

[0110] S1021 divides the space between the ground and radar waves into multiple vertical layers, each vertical layer into multiple horizontal grids, and the horizontal and vertical resolutions of the subspaces are configured to be consistent with the resolution of the atmospheric field data used.

[0111] In this embodiment of the invention, the horizontal and vertical resolutions of the subspace are configured to be consistent with the resolution of the atmospheric field data used, namely, a three-dimensional instantaneous wind field with a spatial resolution of 3.3 km, a time resolution of 1 h, and a vertical resolution of 28 layers, to ensure that the atmospheric parameters are accurately matched with the subspace scale.

[0112] S1022, the raindrop diameter of a single radar pixel is estimated using a normalized gamma distribution to obtain the initial diameter of the raindrop.

[0113] In this embodiment of the invention, the average raindrop diameter of a single radar pixel is estimated using a normalized gamma distribution. To more accurately describe the raindrop distribution at different times and locations, a dual-polarization radar inversion method is further employed to retrieve DSD parameters.

[0114] S1023, the original radar pixel is divided into four sub-grids, with its center point as the initial horizontal coordinate and the radar beam height corresponding to that point as the initial height, to obtain the initial position of the raindrop.

[0115] In this embodiment of the invention, step S103 includes the following sub-steps:

[0116] S1031 simulates the trajectory of raindrops in each layer and subspace of the ground and radar wave space by solving the particle motion equations:

[0117]

[0118] in, U , V , W It is in the wind field x , y and η Quantity; m For raindrop mass; R e It is the Reynolds number; ρ a and ρ w These are the densities of air and water, respectively. μ air viscosity; D The diameter of the raindrop; C d denoted as the drag coefficient of the raindrop.

[0119] S1032 simulates the trajectory of raindrops in each layer and subspace of the ground and radar wave space within each time step. At the end of the time step, the new position and velocity of the raindrops are obtained for the next time step.

[0120] In this embodiment of the invention, the numerical simulation shown in the above formula (1) is performed in each time step. At the end of the time step, the new position and velocity of the raindrop are obtained for the next time step.

[0121] S1033, all time steps x and y The displacements in each direction are accumulated and iterated repeatedly until the raindrop moves out of the subspace.

[0122] The specific calculation process is shown in the following formula:

[0123]

[0124] In the formula: , as well as They are grids i ,layer k and time step t exist x , y , η Displacement in the direction. tn k refer to kThe time step of the layer is obtained by solving equations (1) and (4).

[0125] In this embodiment of the invention, the raindrop exit position and velocity at the boundary are used as inputs to the next subspace, and the determination of the next subspace provides the position of the raindrop exit.

[0126] In this embodiment of the invention, step S104 includes the following sub-steps:

[0127] S1041, simulate the trajectory of the raindrop in the new subspace, iterate repeatedly until the raindrop reaches the ground, obtain the final position coordinates of the raindrop on the ground as observed by radar, and use the final position coordinates as the corrected raindrop point.

[0128] In this embodiment of the invention, the above formula (1) is simulated in a new subspace and iterated repeatedly until the raindrops reach the ground (the final position coordinates of the raindrops reaching the ground are called the corrected raindrop point group RRPs), and the final position coordinates of the raindrops on the ground observed by radar are obtained. X i,R , Y i,R (Formulas (5) and (6)) are used to name the modified raindrop points (RRPs) to achieve rain gauge position transformation.

[0129]

[0130] in,( X i,R , Y i,R ) represents the coordinates of grid i; X i,O ,Y i,O The original horizontal coordinates of the raindrop are called the original raindrop point (ORP). kn The number of floors.

[0131] S1042, for each rain gauge, search for a group of corrected raindrop points around it and select the corrected raindrop point closest to the rain gauge, and pair the original radar pixel of the corrected raindrop point with the rain gauge to obtain an adjusted radar-rain gauge pair.

[0132] In this embodiment of the invention, for each rain gauge, a group of corrected raindrop points (RRPs) around it is searched, and the corrected raindrop point RRP closest to the rain gauge is selected to achieve spatial matching of the radar-rain gauge pair. The radar cell whose final position is RRP and its paired rain gauge are regarded as an adjusted radar-rain gauge pair.

[0133] In this embodiment of the invention, the spatial relationship between the radar and the rain gauge is corrected based on the location coordinates of the corrected raindrop point (i.e., the final location coordinates of the raindrop on the ground as observed by the radar considering the raindrop drift trajectory).

[0134] In this embodiment of the invention, step S106 includes:

[0135] S1061, the trajectory of the raindrop is simulated in a new subspace, iterated repeatedly until the raindrop reaches the ground, and the final position coordinates of the raindrop on the ground observed by the radar are obtained. The final position coordinates are used as the corrected raindrop point for physical correction, and the physical correction rainfall corresponding to the original radar pixel of the corrected raindrop point is extracted. P raw Simultaneously calculate the physical correction coefficient. C phys ;

[0136] S1062, Based on the random forest model, input the radar rainfall data and three-dimensional atmospheric field data of the corresponding original radar pixel, output the predicted rain gauge rainfall of that pixel, obtain the measured rainfall of the rain gauge, and calculate the random forest correction coefficient. C rf ;

[0137] S1063, Physical correction coefficient C phys With Random Forest Correction Coefficient C rf The comprehensive correction coefficient corresponding to the corrected raindrop point is obtained by direct averaging. C :

[0138]

[0139] in, C The range of values ​​varies C phys , C rf Synchronous changes ensure that the correction intensity is consistent with the trend of the results from both methods.

[0140] In this embodiment of the invention, step S107 includes:

[0141] S1071, for each rain gauge, search for the surrounding group of corrected raindrop points and select the corrected raindrop point closest to the rain gauge, pair the original radar pixel of the corrected raindrop point with the rain gauge to obtain an adjusted radar-rain gauge pair, and pair the radar pixel constructed based on radar beam projection with the rain gauge to obtain an unadjusted radar-rain gauge pair.

[0142] S1072, the rainfall amount of raindrops in each pixel is adjusted with the three-dimensional matrix in the projection space and time, and the radar measured rainfall amount and location are adjusted based on the function for calculating radar rainfall deviation, and the corresponding corrected radar rainfall amount is obtained. P adj for: .

[0143] In step S107, interpolation techniques can be used to obtain the rainfall surface estimated by the rain gauge, and its mean can be calculated to obtain the rainfall measured by the rain gauge. Correspondingly, by performing bicubic spline interpolation on the radar rainfall, the rainfall surface estimated by the radar measurement can be obtained, and the mean rainfall of this surface can be calculated to obtain the radar measured rainfall.

[0144] Based on the spatial relationship, rain gauge measurement of rainfall, radar measurement of rainfall, and radar-rain gauge deviation, a function for calculating radar rainfall deviation is obtained, including the rainfall of raindrops in each pixel and the adjustment of a three-dimensional matrix in projection space and time. The radar measurement of rainfall and location are adjusted based on the function for calculating radar rainfall deviation.

[0145] For ease of understanding, this embodiment of the invention also provides a specific implementation scheme, in which...

[0146] The data used included radar observations, field observations, and meteorological data. Radar observations were from the Hameldon Hill radar located in Lancashire, UK (53°45′17”N, 2°17′19”W). This surface radar performs a series of scans at different angles every 5 minutes. Ground observation data came from 11 rain gauge stations within a 50 km radius of the Hameldon Hill radar, with a time resolution of 1 hour. Additionally, meteorological data used the ERA-Interim reanalysis dataset (2015–2017) to improve the consistency between radar and instrument data and support the construction of a three-dimensional atmospheric field. Quality control and consistency checks were performed on the radar and rain gauge data, and rainfall case events were selected to ensure the representativeness of the analysis and its compatibility with the WRF model. This example uses two rainfall events, May 8, 2015, and November 21, 2016, for experimental simulation and provides a graphical demonstration.

[0147] The implementation scheme uses the ground-based radar dataset as input to the WRF model. According to step S201, the space between the ground and the radar wave is divided, and the horizontal and vertical resolutions are configured accordingly. The diameter and initial position of the raindrops are set to initialize the simulation of the raindrop trajectory. In this example, the diameter of the raindrops is set to 0.2 mm, 1 mm, and 5 mm, and the initial position is set to 51.11 N and 22.47 W.

[0148] The raindrop trajectory simulation process in this implementation plan includes:

[0149] Step 1: Simulate the trajectory of each raindrop in the ground and radar wave space according to the above formula (1);

[0150] Step 2: Perform the above numerical simulation within each time step. At the end of the time step, the new position and velocity of the raindrops will be used for the next step. In this example, the time step is set to 30 seconds.

[0151] Step 3: Accumulate the displacements in the x and y directions of all time steps respectively (Formulas (2) and (3)), and iterate repeatedly until the raindrop moves out of the subspace;

[0152] Step 4: Use the raindrop exit position and velocity at the boundary as input for the next subspace. Determining the next subspace provides the location of the raindrop exit.

[0153] In this implementation scheme, the simulation of raindrop trajectories is iteratively repeated in a new subspace until the raindrops reach the ground, in order to obtain the correct spatial relationship between the radar and the rain gauge, thereby correcting the error caused by wind drift. Ground winds of 0.2, 1, and 5 mm at different time steps during the May 8, 2015 event are shown below. Figure 2 The first row shows the surface wind and raindrop drift of 0.2, 1, and 5 mm at different time steps during the November 21, 2016 event. Figure 3 As shown in the first row.

[0154] Interpolation methods are used to calculate the rainfall surface estimated by the rain gauge and the rainfall surface estimated by the radar. Based on this, the correction coefficient for the radar bias is calculated, and finally the radar rainfall estimate is adjusted. Figure 3 This study compares rainfall data before and after correction from the Hamel Mountain Radar, located in Lancashire, UK (53.75°N, 2.29°S) and its surrounding areas, using different time steps for the May 8, 2015 event. Figure 2 The second and third rows show a comparison of the original and corrected radar images at different time steps for the November 21, 2016 event. Figure 3 The second and third rows are shown.

[0155] Based on the same inventive concept, embodiments of the present invention also provide an urban area radar rainfall correction device that integrates raindrop dynamics simulation and machine learning, the device comprising:

[0156] The dataset construction module is used to build a multi-source fusion dataset of urban rainfall and to preprocess the multi-source fusion dataset of urban rainfall.

[0157] The initial condition setting module is used to divide the space between the radar scanning layer and the ground layer into multiple vertical layers, each layer into multiple horizontal square grids, resulting in a space composed of multiple three-dimensional subspaces, thereby setting the initial diameter and initial position of the raindrops as the initial condition settings for the simulation of the raindrop trajectory.

[0158] The simulation module is used to simulate the trajectory of raindrops in the current subspace between the ground and radar waves based on the division of multiple subspaces and the initial condition settings, and to accumulate the displacement in each time step in real time until the raindrops move out of the current subspace.

[0159] The physical correction module is used to continue iterating the simulation process of raindrop trajectory in the new subspace after the raindrop enters the new subspace until the raindrop reaches the ground, obtain the spatial relationship between the radar and the rain gauge, correct the error caused by the wind drift effect, and record it as the physical correction result.

[0160] The random forest correction module is used to determine the predicted rain gauge rainfall at each time step based on a pre-built random forest model, radar rainfall data at each time step, and three-dimensional atmospheric field data. The three-dimensional atmospheric field data includes atmospheric data at the altitude where the radar scan is located and atmospheric data at the near-surface level. The random forest model has pre-learned the correspondence between radar rainfall data and rain gauge measurement data, which is denoted as the random forest correction result.

[0161] The fusion module is used to perform weighted fusion of physical correction results and random forest correction results to obtain a comprehensive correction coefficient.

[0162] The adjustment module is used to construct a spatial distribution surface of rainfall based on rain gauge observation data and radar observation data through interpolation methods, thereby obtaining the rainfall measured by the rain gauge and the rainfall measured by the radar. Based on the spatial relationship, the rainfall measured by the rain gauge, the rainfall measured by the radar, the radar-rain gauge deviation, and the comprehensive correction coefficient, a radar rainfall deviation function is constructed. Based on this function, the radar rainfall measurement is adjusted to achieve high-precision correction of radar rainfall in urban areas.

[0163] Optionally, the dataset construction module is used for:

[0164] A WRF model was constructed by selecting multiple rainfall events with different time ranges, rainfall intensities, and seasons.

[0165] Acquire radar observation data, rain gauge observation data, and meteorological data to construct a multi-source fusion dataset of urban rainfall;

[0166] The urban rainfall multi-source fusion dataset was subjected to quality control, consistency checks, and correction of outliers.

[0167] Optionally, the initial condition setting module is used to:

[0168] The space between the ground and radar waves is divided into multiple vertical layers, and each vertical layer is divided into multiple horizontal square grids. The horizontal and vertical resolutions of the subspaces are configured to be consistent with the resolution of the atmospheric field data used.

[0169] The initial diameter of the raindrop is obtained by estimating the diameter of a single radar pixel using a normalized gamma distribution.

[0170] The original radar pixel is divided into four sub-grids. The initial horizontal coordinate of the sub-grid is taken as the center point of the sub-grid, and the initial height of the radar beam corresponding to the center point is taken as the initial height. The initial position of the raindrop is then obtained.

[0171] Optionally, the simulation module is used for:

[0172] The trajectory of raindrops in each layer and subspace of the ground and radar wave space is simulated by solving the particle motion equations.

[0173] Simulate the trajectory of raindrops in each layer and subspace of the ground and radar wave space within each time step. At the end of the time step, obtain the new position and velocity of the raindrops for the next time step.

[0174] All time steps x and y The displacements in each direction are accumulated and iterated repeatedly until the raindrop moves out of the subspace.

[0175] Optionally, the physical correction module is used for:

[0176] The trajectory of the raindrop is simulated in a new subspace and iterated repeatedly until the raindrop reaches the ground. The final position coordinates of the raindrop on the ground observed by radar are obtained, and the final position coordinates are used as the corrected raindrop point.

[0177] For each rain gauge, a group of corrected raindrop points around it is searched and the corrected raindrop point closest to the rain gauge is selected. The original radar pixel of the corrected raindrop point is paired with the rain gauge to obtain an adjusted radar-rain gauge pair. The radar pixel constructed based on radar beam projection is paired with the rain gauge to obtain an unadjusted radar-rain gauge pair.

[0178] Optionally, the random forest model is constructed and trained based on the following steps:

[0179] Radar rainfall data and three-dimensional atmospheric field data from rainfall events are selected as input features, and rain gauge measurement data at the corresponding time are used as output labels.

[0180] The input features are standardized, and the training set and validation set are divided using the 5-fold cross-validation method. The core parameters of the random forest model are set as follows: the number of decision trees is 100-500, the maximum depth of the decision trees is 10-30 layers, the minimum number of split samples for a node is 5-20, and the minimum number of samples for a leaf node is 1-5.

[0181] The random forest model is trained using the training set, and its performance is evaluated and parameters are optimized using the validation set until the model's prediction error meets a preset threshold.

[0182] Optionally, the fusion module is used for:

[0183] The trajectory of raindrops is simulated in a new subspace, iterated repeatedly until the raindrops reach the ground, obtaining the final position coordinates of the raindrops on the ground as observed by radar. These final position coordinates are used as the corrected raindrop point for physical correction, and the physical correction rainfall corresponding to the original radar pixel at this corrected raindrop point is extracted. P raw Simultaneously calculate the physical correction coefficient. C phys ;

[0184] Based on the random forest model, the radar rainfall data and three-dimensional atmospheric field data of the corresponding original radar pixel are input, the predicted rainfall from the rain gauge is output, the measured rainfall from the rain gauge is obtained, and the random forest correction coefficient is calculated. C rf ;

[0185] For physical correction coefficients C phys With Random Forest Correction Coefficient C rf The comprehensive correction coefficient corresponding to the corrected raindrop point is obtained by direct averaging. C :

[0186]

[0187] in, C The range of values ​​varies C phys , C rf Synchronous changes ensure that the correction intensity is consistent with the trend of the results from both methods.

[0188] Optionally, the adjustment module is used to:

[0189] For each rain gauge, search for the surrounding group of corrected raindrop points and select the corrected raindrop point closest to the rain gauge. Pair the original radar pixel of the corrected raindrop point with the rain gauge to obtain an adjusted radar-rain gauge pair. Pair the radar pixel constructed based on radar beam projection with the rain gauge to obtain an unadjusted radar-rain gauge pair.

[0190] The rainfall amount of raindrops in each pixel is adjusted with a three-dimensional matrix in projection space and time. Based on the function for calculating radar rainfall deviation, the radar measured rainfall amount and location are adjusted, resulting in the corresponding corrected radar rainfall amount. P adj for: .

[0191] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the steps of the urban area radar rainfall correction method that integrates raindrop dynamic process simulation and machine learning as described in any of the above embodiments.

[0192] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the urban area radar rainfall correction method that integrates raindrop dynamics simulation and machine learning as described in any of the above embodiments.

[0193] Based on the same inventive concept, embodiments of the present invention provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the urban area radar rainfall correction method that integrates raindrop dynamics simulation and machine learning as described in any of the above embodiments.

[0194] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0195] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0196] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0197] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable terminal device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0198] These computer program instructions can also be loaded onto a computer or other programmable terminal device to cause a series of operational steps to be performed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0199] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0200] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0201] The above provides a detailed description of the urban area radar rainfall correction method that integrates raindrop dynamic process simulation and machine learning provided by the present invention. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for radar rainfall correction in urban areas that integrates raindrop dynamics simulation and machine learning, characterized in that, The method includes: Construct a multi-source fusion dataset of urban rainfall and preprocess the dataset. The space between the radar scanning layer and the ground layer is divided into multiple vertical layers, and each layer is divided into multiple horizontal square grids, resulting in a space composed of multiple three-dimensional subspaces. This allows us to set the initial diameter and initial position of the raindrops as the initial conditions for simulating the raindrop trajectory. Based on the division of multiple subspaces and the initial condition settings, the trajectory of raindrops in the current subspace between the ground and radar waves is simulated, and the displacement within each time step is accumulated in real time until the raindrops move out of the current subspace. After the raindrop enters the new subspace, the simulation process of the raindrop's trajectory continues to iterate in the new subspace until the raindrop reaches the ground. The spatial relationship between the radar and the rain gauge is obtained, and the error caused by the wind drift effect is corrected. This is recorded as the physical correction result. Based on a pre-built random forest model, radar rainfall data at each time step, and three-dimensional atmospheric field data, the predicted rain gauge rainfall at each time step is determined. The three-dimensional atmospheric field data includes atmospheric data at the altitude where the radar scan is located and atmospheric data at the near-surface level. The random forest model has pre-learned the correspondence between radar rainfall data and rain gauge measurement data, which is denoted as the random forest correction result. The physical correction results and the random forest correction results are weighted and fused to obtain the comprehensive correction coefficient; A spatial distribution surface of rainfall based on rain gauge and radar observation data is constructed using interpolation methods to obtain the rainfall measured by rain gauges and radar. Based on the spatial relationship, the rainfall measured by rain gauges, the rainfall measured by radar, the radar-rain gauge deviation, and the comprehensive correction coefficient, a radar rainfall deviation function is constructed. The radar rainfall measurement is adjusted based on this function to achieve high-precision correction of radar rainfall in urban areas.

2. The urban area radar rainfall correction method integrating raindrop dynamics simulation and machine learning as described in claim 1, characterized in that, Construct a multi-source fusion dataset of urban rainfall, and preprocess the dataset, including: A WRF model was constructed by selecting multiple rainfall events with different time ranges, rainfall intensities, and seasons. Acquire radar observation data, rain gauge observation data, and meteorological data to construct a multi-source fusion dataset of urban rainfall; The urban rainfall multi-source fusion dataset was subjected to quality control, consistency checks, and correction of outliers.

3. The urban area radar rainfall correction method integrating raindrop dynamics simulation and machine learning as described in claim 1, characterized in that, The space between the radar scanning layer and the ground layer is divided into multiple vertical layers, each layer is further divided into multiple horizontal square grids, resulting in a space composed of multiple three-dimensional subspaces. This allows for the setting of the initial diameter and initial position of raindrops, including: The space between the ground and radar waves is divided into multiple vertical layers, and each vertical layer is divided into multiple horizontal square grids. The horizontal and vertical resolutions of the subspaces are configured to be consistent with the resolution of the atmospheric field data used. The initial diameter of the raindrop is obtained by estimating the diameter of a single radar pixel using a normalized gamma distribution. The original radar pixel is divided into four sub-grids. The initial horizontal coordinate of the sub-grid is taken as the center point of the sub-grid, and the initial height of the radar beam corresponding to the center point is taken as the initial height. The initial position of the raindrop is then obtained.

4. The urban area radar rainfall correction method according to claim 1, which integrates raindrop dynamics simulation and machine learning, simulates the trajectory of raindrops in the current subspace between the ground and radar waves based on the division of multiple subspaces and the initial condition settings, and accumulates the displacement in each time step in real time until the raindrops move out of the current subspace, including: The trajectory of raindrops in each layer and subspace of the ground and radar wave space is simulated by solving the particle motion equations. Simulate the trajectory of raindrops in each layer and subspace of the ground and radar wave space within each time step. At the end of the time step, obtain the new position and velocity of the raindrops for the next time step. All time steps x and y The displacements in each direction are accumulated and iterated repeatedly until the raindrop moves out of the subspace.

5. The urban area radar rainfall correction method integrating raindrop dynamics simulation and machine learning according to claim 4, characterized in that, After the raindrop enters the new subspace, the simulation process of the raindrop's trajectory continues iteratively in the new subspace until the raindrop reaches the ground. This yields the spatial relationship between the radar and the rain gauge, correcting for errors caused by wind drift, including: The trajectory of the raindrop is simulated in a new subspace and iterated repeatedly until the raindrop reaches the ground. The final position coordinates of the raindrop on the ground observed by radar are obtained, and the final position coordinates are used as the corrected raindrop point. For each rain gauge, a group of corrected raindrop points around it is searched and the corrected raindrop point closest to the rain gauge is selected. The original radar pixel of the corrected raindrop point is paired with the rain gauge to obtain an adjusted radar-rain gauge pair. The radar pixel constructed based on radar beam projection is paired with the rain gauge to obtain an unadjusted radar-rain gauge pair.

6. The urban area radar rainfall correction method integrating raindrop dynamics simulation and machine learning according to claim 1, characterized in that, The random forest model is constructed and trained based on the following steps: Radar rainfall data and three-dimensional atmospheric field data from rainfall events are selected as input features, and rain gauge measurement data at the corresponding time are used as output labels. The input features are standardized, and the training set and validation set are divided using the 5-fold cross-validation method. The core parameters of the random forest model are set as follows: the number of decision trees is 100-500, the maximum depth of the decision trees is 10-30 layers, the minimum number of split samples for a node is 5-20, and the minimum number of samples for a leaf node is 1-5. The random forest model is trained using the training set, and its performance is evaluated and parameters are optimized using the validation set until the model's prediction error meets a preset threshold.

7. The urban area radar rainfall correction method integrating raindrop dynamics simulation and machine learning according to claim 4 or 5, characterized in that, The physical correction results and the random forest correction results are weighted and fused to obtain the comprehensive correction coefficients, including: The trajectory of raindrops is simulated in a new subspace, iterated repeatedly until the raindrops reach the ground, obtaining the final position coordinates of the raindrops on the ground as observed by radar. These final position coordinates are used as the corrected raindrop point for physical correction, and the physical correction rainfall corresponding to the original radar pixel at this corrected raindrop point is extracted. P raw Simultaneously calculate the physical correction coefficient. C phys ; Based on the random forest model, the radar rainfall data and three-dimensional atmospheric field data of the corresponding original radar pixel are input, the predicted rainfall from the rain gauge for that pixel is output, the measured rainfall from the rain gauge is obtained, and the random forest correction coefficient is calculated. C rf ; For physical correction coefficients C phys With Random Forest Correction Coefficient C rf The comprehensive correction coefficient corresponding to the corrected raindrop point is obtained by direct averaging. C : in, C The range of values ​​varies C phys , C rf Synchronous changes ensure that the correction intensity is consistent with the trend of the results from both methods.

8. The urban area radar rainfall correction method integrating raindrop dynamics simulation and machine learning according to claim 7, characterized in that, Based on the aforementioned spatial relationships, rainfall measured by rain gauges, rainfall measured by radar, and radar-rain gauge deviation and comprehensive correction coefficients, a radar rainfall deviation function is constructed, including: For each rain gauge, search for the surrounding group of corrected raindrop points and select the corrected raindrop point closest to the rain gauge. Pair the original radar pixel of the corrected raindrop point with the rain gauge to obtain an adjusted radar-rain gauge pair. Pair the radar pixel constructed based on radar beam projection with the rain gauge to obtain an unadjusted radar-rain gauge pair. The rainfall amount of raindrops in each pixel is adjusted with a three-dimensional matrix in projection space and time. Based on the function for calculating radar rainfall deviation, the radar measured rainfall amount and location are adjusted, resulting in the corresponding corrected radar rainfall amount. P adj for: .

9. A radar rainfall correction device for urban areas that integrates raindrop dynamics simulation and machine learning, characterized in that, The device includes: The dataset construction module is used to build a multi-source fusion dataset of urban rainfall and to preprocess the multi-source fusion dataset of urban rainfall. The initial condition setting module is used to divide the space between the radar scanning layer and the ground layer into multiple vertical layers, each layer into multiple horizontal square grids, resulting in a space composed of multiple three-dimensional subspaces, thereby setting the initial diameter and initial position of the raindrops as the initial condition settings for the simulation of the raindrop trajectory. The simulation module is used to simulate the trajectory of raindrops in the current subspace between the ground and radar waves based on the division of multiple subspaces and the initial condition settings, and to accumulate the displacement in each time step in real time until the raindrops move out of the current subspace. The physical correction module is used to continue iterating the simulation process of raindrop trajectory in the new subspace after the raindrop enters the new subspace until the raindrop reaches the ground, obtain the spatial relationship between the radar and the rain gauge, correct the error caused by the wind drift effect, and record it as the physical correction result. The random forest correction module is used to determine the predicted rain gauge rainfall at each time step based on a pre-built random forest model, radar rainfall data at each time step, and three-dimensional atmospheric field data. The three-dimensional atmospheric field data includes atmospheric data at the altitude where the radar scan is located and atmospheric data at the near-surface level. The random forest model has pre-learned the correspondence between radar rainfall data and rain gauge measurement data, which is denoted as the random forest correction result. The fusion module is used to perform weighted fusion of physical correction results and random forest correction results to obtain a comprehensive correction coefficient. The adjustment module is used to construct a spatial distribution surface of rainfall based on rain gauge observation data and radar observation data through interpolation methods, thereby obtaining the rainfall measured by the rain gauge and the rainfall measured by the radar. Based on the spatial relationship, the rainfall measured by the rain gauge, the rainfall measured by the radar, the radar-rain gauge deviation, and the comprehensive correction coefficient, a radar rainfall deviation function is constructed. Based on this function, the radar rainfall measurement is adjusted to achieve high-precision correction of radar rainfall in urban areas.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the urban area radar rainfall correction method according to any one of claims 1-8, which integrates raindrop dynamics simulation and machine learning.

Images