A weather radar ground object echo recognition method based on XGBoost

Abstract

The application discloses a meteorological radar ground object echo recognition method based on XGBoost, and comprises the following steps: step S1, selecting radar parameters for inputting an XGBoost model, wherein the radar parameters comprise a reflectivity factor, a reflectivity factor texture, a differential reflectivity factor, a change degree of echo intensity along a radial direction, a regional average value of a radial velocity and a correlation coefficient; step S2, initializing the XGBoost model to determine parameters of the initial XGBoost model; step S3, inputting a meteorological radar measured data sample set, and dividing the meteorological radar measured data sample set into a training set and a test set; step S4, performing Bayesian optimization on the initial XGBoost model by using the training set to obtain a final XGBoost optimization model; step S5, inputting the test set into the XGBoost optimization model to perform iteration, and stopping the iteration when a maximum iteration number is reached or a loss function of the XGBoost optimization model reaches an optimal solution of an objective function, thereby completing final recognition and classification. The application is suitable for various weather conditions and can more accurately recognize ground object echoes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of meteorological radar signal processing, and particularly relates to a meteorological radar ground object echo recognition method based on XGBoost. BACKGROUND

[0002] Meteorological radar is an important means of weather detection, and is widely used in the observation of weather phenomena such as clouds, rain, snow, thunderstorms and turbulence, and meteorological targets are analyzed through echo information. There are many types of disaster weather and they occur frequently, which can easily cause high economic losses, and therefore, the research work of monitoring and forecasting disaster weather by using meteorological radar has very important significance. In the radar-detected echo, there is a considerable part of ground object echo, and if it is not identified, it will lead to a decrease in the quality of radar observation data and affect the detection and identification of meteorological targets. Therefore, in order to obtain accurate radar data and ensure the quality of radar products, the research on the identification method of ground object echo is very important.

[0003] Kessinger proposed a radar echo classification algorithm, which first calculated a feature parameter based on radar basic parameters, and identified meteorological echoes based on the idea of fuzzy logic. Cluckie et al. proposed a method combining Bayesian recognition algorithm and fuzzy logic for identification and classification. On the basis of the fuzzy logic method proposed by Kessinger, Liu Liping et al. proposed a distributed ground object recognition method, which used the same feature parameters as the method proposed by Kessinger and improved the ground object echo recognition effect. Zhuang Wei et al. improved the over-suppression problem of Liu Liping's method by adding an echo filling function before processing the echo, and changed the ground object discrimination threshold from a fixed value to a function that changes with distance, reducing the over-suppression of ground object echoes at a long distance. The characteristics of these methods are to find a criterion or discriminant factor that can distinguish meteorological targets from non-meteorological targets as the input parameter of the fuzzy logic method, which relies on a large number of empirical values, cannot automatically adjust the membership function in the method, and has low self-adaptation to weather conditions and is not universal. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a meteorological radar ground object echo recognition method based on XGBoost. The present application avoids the problem that the traditional fuzzy logic cannot adaptively process meteorological data, improves the recognition efficiency, and at the same time optimizes the XGBoost model by using the Bayesian optimization method, further improves the recognition accuracy.

[0005] The present application adopts the following technical solutions to solve the above technical problems:

[0006] The application provides a meteorological radar ground object echo recognition method based on XGBoost.

[0007] Step S1, radar parameters for inputting an XGBoost model are selected, and the radar parameters include a reflectivity factor, reflectivity factor texture, differential reflectivity factor, a change degree of echo intensity along a radial direction, a regional average value of radial velocity, and a correlation coefficient;

[0008] Step S2, an XGBoost model is initialized, and parameters of the initial XGBoost model are determined;

[0009] Step S3, meteorological radar measured data samples [x i ,y i ] are input, wherein x i is the i-th meteorological radar parameter, and the size of x i is an m-dimensional vector, y i is a label of the i-th meteorological data sample, and the meteorological radar measured data samples are divided into a training set and a test set;

[0010] Step S4, the initial XGBoost model is subjected to Bayesian optimization by using the training set, and a final XGBoost optimization model is obtained;

[0011] Step S5, the test set is input into the XGBoost optimization model, iteration is performed, and when a maximum iteration number is reached or a loss function of the XGBoost optimization model reaches an optimal solution of an objective function, the iteration is stopped, and final recognition and classification are completed.

[0012] As a further optimization scheme of the meteorological radar ground object echo recognition method based on XGBoost, the Bayesian optimization in step S4 is specifically as follows.

[0013] Step S41, the Bayesian optimization model is initialized, parameters to be optimized in the XGBoost model are input into the Bayesian model, a population P(0) is initialized, and evolution generation t is set to 0;

[0014] Step S42, a candidate solution S(t) of the t-th iteration is selected from the population P(t) of the t-th iteration;

[0015] Step S43, a Bayesian network B(μ,σ) is established from the candidate solution S(t), and the formula is as follows:

[0016]

[0017] Wherein, x t is a sampling position of the next XGBoost model is a weight parameter, μ t-1 (x) is a predicted mean value, is variance, x is the hyperparameter of the weather radar ground echo recognition XGBoost model, D is the hyperparameter data set of the weather radar ground echo recognition XGBoost model;

[0018] Step S44, generating a new solution O(t) according to the joint distribution function of the Bayesian network B(mu, sigma);

[0019] Step S45, replacing part of the solution in P(t) with O(t) to form a new population, let t=t+1, return to step S42, until the optimal solution of x is found, thereby obtaining the final XGBoost optimization model. t

[0020] As a further optimization scheme of the XGBoost-based weather radar ground echo recognition method according to the application, in step S2, the parameters are set as: maximum tree depth Max_depth=4, minimum sample weight of child node min_child_weight=1, minimum loss reduction gamma=0, learning rate eta=0.1, regularization parameter lamda=0, and aplha=0.

[0021] As a further optimization scheme of the XGBoost-based weather radar ground echo recognition method according to the application, in step S4, the Bayesian optimization method is used to optimize the XGBoost model; the Bayesian optimization method first randomly selects an initial population and its candidate solution using an evolutionary algorithm, constructs a new Bayesian network model, updates the candidate solution and substitutes it into the original population, and repeats the process until the maximum value of the function is found.

[0022] Compared with the prior art, the application has the following technical effects:

[0023] (1) The application provides an XGBoost-based weather radar ground echo recognition method, which can adaptively train data and adapt to various weather conditions, and can more accurately recognize ground echo compared with the traditional fuzzy logic method;

[0024] (2) The application uses the Bayesian optimization algorithm to optimize the XGBoost network model, which does not need manual parameter adjustment and has a higher recognition rate than other optimization algorithms such as grid search and random search;

[0025] (3) In addition, the method of the application has better performance in recognizing ground echo using discrete weather radar parameters, is better than other machine learning methods, and has a higher ground echo recognition rate. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is the framework diagram of the XGboost tree model.​

[0027] Figure 2 is the overall algorithm flowchart.

[0028] Figure 3 is the network training process of the data set.

[0029] Figure 4 is the PPI chart of the KVTX radar ground feature echo recognition result in the United States; wherein (a) is a reference chart of the classification result given by the waterborne object classification algorithm used by the WDR-88D radar in the United States, and (b) is a chart of classification of ground feature echoes and meteorological targets completed by the XGBoost algorithm. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be described in detail below with reference to the drawings and specific embodiments.

[0031] Figure 1 is the framework diagram of the XGboost tree model, Figure 2 is the overall algorithm flowchart. The method of the present application is as follows:

[0032] Step S1, selecting radar parameters input into the XGBoost model, the radar parameters including reflectivity factor, reflectivity factor texture, differential reflectivity factor, the degree of change of echo intensity along the radial direction, the regional average value of radial velocity and the correlation coefficient;

[0033] Step S2, initializing the XGBoost model to determine the parameters of the initial XGBoost model;

[0034] Step S3, inputting the meteorological radar measured data sample set [x i ,y i ], wherein x i is the i th meteorological radar parameter, and the size is an m-dimensional vector, y i is the label of the i th meteorological data sample, and the meteorological radar measured data sample set is divided into a training set and a test set;

[0035] Step S4, using the training set to perform Bayesian optimization on the initial XGBoost model to obtain a final XGBoost optimization model;

[0036] Step S5, inputting the test set into the XGBoost optimization model for iteration, and stopping the iteration when the maximum number of iterations is reached or the loss function of the XGBoost optimization model reaches the optimal solution of the objective function, thereby completing the final recognition and classification.

[0037] The Bayesian optimization in step S4 is specifically as follows:

[0038] Step S41, initialize the Bayesian optimization model, input the parameters to be optimized in XGBoost into the Bayesian model, initialize the population P(0), let the evolution number t=0;

[0039] Step S42, select the candidate solution S(t) of the tth iteration from the population P(t) of the tth iteration;

[0040] Step S43, establish the Bayesian network B(μ,σ) from the candidate solution S(t), the formula is:

[0041]

[0042] Where, x t is the next model sampling position is the weight parameter, μ t-1 (x) is the predicted mean, is the variance, x is the hyperparameter of the weather radar ground echo recognition XGBoost model, and D is the hyperparameter data set of the weather radar ground echo recognition XGBoost model;

[0043] Step S44, generate a new solution O(t) according to the joint distribution function of the Bayesian network B(μ,σ);

[0044] Step S45, replace part of the solution in P(t) with O(t) to form a new population, let t=t+1, return to step S42 until the optimal solution of x t is found, and thus the final XGBoost optimization model is obtained.

[0045] Three basic parameters of dual-polarized radar, reflectivity factor, differential reflectivity factor, and correlation coefficient, and three characteristic parameters, reflectivity factor texture, reflectivity factor along the radial direction Change degree, and regional average of radial velocity, the specific content is:

[0046] Two characteristic parameters are extracted from the reflectivity factor: reflectivity factor texture T DBZ , and the change degree of reflectivity factor along the radial direction S PIN , which are defined as follows:

[0047]

[0048]

[0049]

[0050] Two characteristic parameters are extracted from the radial velocity: regional average of radial velocity M DVE , and variance of radial velocity S DVE , which are defined as follows:

[0051] M DVE = med{VE i,j ,i∈N beams ,j∈N gates} (3)

[0052] Where, N gates , N beams denote the number of calculation points defined in the range of distance and azimuth, Z i,j is the echo intensity of the coordinate point (i,j), T DBZ is the variance between adjacent distance banks of echo intensity, mainly reflecting the change of echo intensity within a certain range; Z low and Z up are the echo intensities of the corresponding low and high elevation PPI scans; Z thresh is the threshold value of adjacent echo intensity change, generally 2-5dB, S PIN reflects the consistency of the change of echo intensity along the radial direction; M DVE denotes the radial velocity value of any point after median filtering processing.

[0053] The classification method based on XGBoost provided in the application constructs a target recognition network architecture with radar feature parameters as input, and the specific content is as follows:

[0054] When XGBoost is used for classification, model training is performed, in order to make the prediction value interval of the classification problem result between [0,1], similar to logistic regression, XGBoost introduces a log(odds) function:

[0055]

[0056] The prediction value is converted into log(odds), and when the default initial prediction value is 0.5, log(odds)=0, in Gradient Boosting, the loss function used by us is:

[0057] L(y i ,p i )=-[y i log(p i )+(1-y i )log(1-p i )] (7)

[0058] The loss function of a single sample of XGBoost becomes:

[0059] L(y i ,log(odds))=-y i log(odds)+log(1+elog(odds) ) (8)

[0060] The first derivative and the second derivative are respectively:

[0061]

[0062] In the classification model, the prediction weight O j is:

[0063]

[0064] In the formula, we can regard the residual as the sum of p i ×(1-p i ) as the product of the predicted value and 1-predicted value, and the weight O value in the classification model can be converted to:

[0065]

[0066] The objective function Obj is:

[0067]

[0068] The prediction value of the current tree is:

[0069]

[0070] According to the sample set selected in step S3, the specific content is as follows:

[0071] The classification target is the ground echo and the weather target, which is a binary classification. The KVTX radar data of the United States is selected as the data set, which contains 20107 samples in total, of which 6028 are non-weather target data and 14079 are weather target data. The data is divided into training set and test set according to the ratio of 8:2. Because it belongs to the case of binary classification, and we pay more attention to the classification of ground echo, so the non-weather target and weather target are respectively recorded as positive class and negative class, that is, 1 and 0. Before the algorithm processing, the data is preprocessed.

[0072] First of all, the default value processing. Statistics are made on all samples, and there are 1222 default data in total for 6 input parameters, of which the default value of the area average value M DVE of the radial velocity is the most, and there is no default value for the parameters related to echo intensity. Because the number of default values accounts for a small part of the total number of samples, the default values are directly removed, which is relatively simple and does not need to be specially processed. Therefore, finally there are 18885 samples left, of which 13508 are weather target data and 5377 are non-weather target data.

[0073] Next is the normalization process. Depending on the observation time or weather conditions, there may be significant differences between the data. In algorithm learning, large differences between different feature values ​​may lead to incorrect prediction directions. To reduce the impact of these differences on training, all data is normalized. This paper uses a linear function for normalization:

[0074]

[0075] In the formula, X max and X min These are the maximum and minimum values ​​of each feature parameter, respectively.

[0076] The training set is used to train the model initialized in step S4 to obtain the optimal training model. Training stops when the optimal solution is found or the number of training iterations reaches the maximum. The trained learning model is then used to classify and identify the test set.

[0077] Implementation Examples

[0078] To verify the effectiveness of the XGBoost-based classification algorithm, experimental verification was conducted using US X-band KVTX weather radar data. Furthermore, to illustrate the advantages of the XGBoost-based weather radar ground object echo identification method, the accuracy of various optimized models was first compared, and then the target identification results obtained by this method were compared with other machine learning identification results.

[0079] Figure 3 The graph shows the network training process for the dataset. The solid line represents the loss function curve for the training set, and the dashed line represents the loss function curve for the test set. The red dot represents the minimum point of the loss function for the test set, which is the optimal solution of the entire model. It can be seen that the minimum value of the loss function occurs when the number of iterations is small, indicating that the network training efficiency is high and it can achieve target recognition and classification well.

[0080] Figure 4 PPI map for ground feature echo identification. Figure 4 The recognition result image in (b) is the same as Figure 4 Compared with the reference results in (a), the recognition effect at close range of radar is good and basically consistent with the reference classification. However, when there are meteorological echoes in the southwest direction at long range of radar, the overall recognition effect is good.

[0081] Table 1. Identification accuracy of different XGBoost optimization methods

[0082] Accuracy Initial model 0.9370 Model after grid search 0.9423 Model after Bayesian optimization 0.9632

[0083] Table 1 compares the recognition rates of different XGBoost optimization methods. The initial model has the lowest accuracy under default parameters because of its high learning rate. Grid search requires manual parameter tuning to reduce the learning rate, and the resulting model needs multiple experiments to find the optimal solution. The Bayesian method, on the other hand, adaptively selects the optimal solution through an optimization function, resulting in more accurate selection and a higher recognition rate than the other two models. Therefore, Bayesian optimization was chosen as the optimization algorithm for the XGBoost model.

[0084] Table 2. Confusion matrix of ground feature echo and meteorological target identification results.

[0085]

[0086] Table 2 shows the confusion matrix for the final binary classification. It can be seen that the overall recognition rate is 95.99%, with ground object echo recognition at 94.92% and meteorological echo recognition at 96.43%. The accuracy of ground object echo recognition is lower than that of meteorological echo recognition. Furthermore, the ground object echo recognition results based on XGBoost were compared with the target recognition results of other methods, as shown in Table 3. Table 3 presents the target recognition rates of different methods. It can be seen from Table 3 that the classification accuracy of the XGBoost network model optimized using Bayesian methods is slightly better than other traditional machine learning methods. At the same time, this method outperforms neural network models in discrete feature training and effectively prevents overfitting.

[0087] Table 3 Comparison of recognition rates of different methods

[0088] Method name Accuracy RF model 0.9532 SVM model 0.9247 ANN model 0.9393 XGBoost model 0.9599

[0089] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying ground object echoes from meteorological radar based on XGBoost, characterized in that, Includes the following steps: Step S1: Select the radar parameters for the input XGBoost model. The radar parameters include reflectivity factor, reflectivity factor texture, differential reflectivity factor, radial variation of echo intensity, regional average of radial velocity, and correlation coefficient. Step S2: Initialize the XGBoost model and determine the parameters of the initial XGBoost model; Step S3: Input the meteorological radar measured data sample set [x] i ,y i ], where x i Let y be the i-th weather radar parameter, which is an m-dimensional vector. i Let i be the label for the i-th meteorological data sample, and divide the meteorological radar measured data sample set into a training set and a test set; Step S4: Use the training set to perform Bayesian optimization on the initial XGBoost model to obtain the final optimized XGBoost model. Step S5: Input the test set into the XGBoost optimization model and iterate. Stop the iteration when the maximum number of iterations is reached or the loss function of the XGBoost optimization model reaches the optimal solution of the objective function, and complete the final recognition and classification. The Bayesian optimization in step S4 is as follows: Step S41: Initialize the Bayesian optimization model by inputting the parameters to be optimized in the XGBoost model into the Bayesian model, initializing the population P(0), and setting the number of generations t = 0. Step S42: Select the candidate solution S(t) for the t-th iteration from the population P(t) of the t-th iteration; Step S43: Construct a Bayesian network B(μ,σ) from the candidate solutions S(t), with the following formula: Where, x t The sampling location for the next XGBoost model μ is the weight parameter. t-1 (x) is the predicted mean. denoted as variance, x represents the hyperparameters of the XGBoost model for ground object echo recognition from weather radar, and D represents the hyperparameter dataset of the XGBoost model for ground object echo recognition from weather radar. Step S44: Generate a new solution O(t) based on the joint distribution function of the Bayesian network B(μ,σ); Step S45: Replace some solutions in P(t) with O(t) to form a new population. Let t = t + 1, and return to step S42 until x is found. t The optimal solution is obtained, thus yielding the final XGBoost optimized model.

2. The method for identifying ground features based on XGBoost meteorological radar according to claim 1, characterized in that, In step S2, the parameters are set as follows: maximum tree depth Max_depth = 4, minimum child weight min_child_weight = 1, minimum loss reduction gamma = 0, learning rate eta = 0.1, regularization parameter lambda = 0, aplha = 0.

3. The method for identifying ground features based on XGBoost meteorological radar according to claim 1, characterized in that, In step S4, the XGBoost model is optimized using the Bayesian optimization method. The Bayesian optimization method first uses an evolutionary algorithm to randomly select an initial population and its candidate solutions, constructs a new Bayesian network model, updates the candidate solutions and substitutes them back into the original population, and repeats this process until the maximum value of the function is found.

Images