A method and device for classifying and identifying meteorological radar clutter

Abstract

The present invention discloses a method and device for classifying and identifying meteorological radar clutter. The method includes obtaining the tested meteorological radar data and performing preprocessing; inputting the preprocessed tested meteorological radar data into a trained SegNet network model to obtain a classification and identification result. Among them, the training process of the SegNet network model includes: obtaining the trained meteorological radar data and performing preprocessing; integrating the preprocessed meteorological radar data into corresponding training data sets based on data types; respectively calculating the standard deviations of the training data sets corresponding to radar reflectivity and differential phase as the texture data of radar reflectivity and differential phase; using a fuzzy logic algorithm to process the training data and texture data in the training data set to obtain a label set of particle phase types; initializing the SegNet network model, and performing iterative training based on the training data set and the label set. The present invention can effectively classify and identify meteorological radar clutter, thereby obtaining meteorological radar data with higher quality.

Description

Technical Field

[0001] The present invention relates to a method and device for classifying and identifying meteorological radar clutter, belonging to the technical field of meteorological radar. Background Art

[0002] An important prerequisite for analyzing and processing meteorological radar data is to establish an intelligent meteorological radar quality control model. The identification and classification of radar clutter have important application values in multiple fields. In the aviation field, it not only has a warning effect on aviation hazards brought by complex weather, but also can provide a decision-making basis for route planning; in the field of artificial weather modification, it can not only improve the accuracy of quantitative precipitation detection, but also provide an important reference basis for the operation decision-making and evaluation of artificial weather modification.

[0003] In order to obtain meteorological radar data with higher quality, a primary problem is to reasonably distinguish meteorological echoes from non-meteorological echoes. The classification and identification methods of radar clutter include: statistical decision methods, decision diagram methods, etc. The statistical decision method combines the polarization characteristics of precipitation particles and clutter particles and the research experience of predecessors, and realizes the classification of precipitation particles and clutter particles by setting different polarization parameter thresholds for precipitation particles and clutter particles. However, the threshold values of this method are generally fixed. Therefore, when the environment in the target area under study changes, the classification accuracy will be greatly affected. The decision diagram method classifies precipitation particles and clutter particles according to the pre-determined type boundaries. However, since the covariance matrices of different precipitation particles and clutter particles received by the meteorological radar are not independent of each other, the classification accuracy of the decision diagram method will be affected to a certain extent. For the classification and identification task of precipitation particles and clutter particles, it is necessary to solve the problem of low intelligence of traditional algorithms and manual detection. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method and device for classifying and identifying meteorological radar clutter, which can effectively classify and identify meteorological radar clutter, so as to obtain meteorological radar data with higher quality.

[0005] To achieve the above purpose, the present invention is implemented by adopting the following technical solutions:

[0006] In the first aspect, the present invention provides a method for classifying and identifying meteorological radar clutter, including:

[0007] Obtain the tested meteorological radar data and perform preprocessing;

[0008] Input the preprocessed tested meteorological radar data into the trained SegNet network model to obtain the classification and identification results;

[0009] Among them, the training process of the SegNet network model includes:

[0010] Obtain the trained meteorological radar data and perform preprocessing; the meteorological radar data includes radar reflectivity, differential reflectivity, differential propagation phase shift rate, correlation coefficient, and differential phase;

[0011] Integrate the preprocessed meteorological radar data into corresponding training data sets based on the data types;

[0012] Calculate the standard deviations of the training data sets corresponding to the radar reflectivity and differential phase respectively as the texture data of the radar reflectivity and differential phase;

[0013] Adopt a fuzzy logic algorithm to process the training data and texture data in the training data sets to obtain a label set of particle phase types;

[0014] Initialize the SegNet network model, and perform iterative training based on the training data sets and the label set until the preset maximum number of iterations is reached or the weight parameters of the SegNet network model converge.

[0015] Optionally, the preprocessing includes data cleaning and scale expansion;

[0016] The data cleaning includes finding the NaN data values in the meteorological radar data and the radar reflectivity less than zero, and setting them to zero respectively;

[0017] The scale expansion includes expanding the meteorological radar data in the way of aligning the upper left corner and filling zeros in the lower right corner.

[0018] Optionally, the calculating the standard deviations of the training data sets corresponding to the radar reflectivity and differential phase respectively includes:

[0019] Traverse the training data set corresponding to the radar reflectivity, and calculate the standard deviation S of the radar reflectivity within 1 km D (Z H )

[0020]

[0021]

[0022] In the formula, m(Z H ) is the average value of the radar reflectivity Z H within 1 km, and n Z is the number of data points of the radar reflectivity within 1 km;

[0023] Traverse the training data set corresponding to the differential phase, and calculate the standard deviation of the differential phase within 2 km

[0024]

[0025]

[0026] Wherein, m(φ DP ) is the average value of the differential phase φ within 2 km, and n DP is the number of data points of the differential phase within 2 km. φ is the number of data points of the differential phase within 2 km.

[0027] Optionally, the method for obtaining the label set of the particle phase state type by processing the training data set and the texture data using the fuzzy logic algorithm includes:

[0028] Using the trapezoidal function as the membership function of the fuzzy logic algorithm, performing fuzzy logic operations on the training data in the training data set and the texture data as input parameters, and obtaining the membership degree corresponding to the particle phase state type as the fuzzy logic output; the trapezoidal function is:

[0029]

[0030] Wherein, X1, X2, X3, X4 are the threshold parameters of the trapezoidal function, and P (j) (x i ) is the membership degree of the jth input parameter to the ith particle phase state type;

[0031] Performing defuzzification on the fuzzy logic output using the weighted average decision method to obtain the integrated membership value corresponding to the particle phase state type:

[0032]

[0033] Wherein, S i is the integrated membership value of the jth input parameter to the ith particle phase state type, W ij is the weight coefficient of the jth input parameter of the ith particle phase state type, and J is the number of input parameters;

[0034] If the integrated membership value of the input parameter is 0, the particle phase state type corresponding to the input parameter is no meteorological echo, and its label is set to 0;

[0035] If the integrated membership value of the input parameter is greater than the set threshold, the particle phase state type corresponding to the input parameter is clutter, and its label is set to 1;

[0036] If the integrated membership value of the input parameter is less than the set threshold and greater than 0, the particle phase state type corresponding to the input parameter is other, and its label is set to 2;

[0037] Construct a label set of the particle phase state type according to the labels of the input parameters.

[0038] Optionally, the iterative training based on the training data set and the label set includes:

[0039] The training data is downsampled by the Encoder network based on the SegNet network model to obtain data features;

[0040] The feature data is upsampled by the Decoder network based on the SegNet network model to obtain the prediction result;

[0041] Based on the cross-entropy loss function, the loss between the prediction result and its label is calculated, and the weight parameters of the SegNet network model are updated by backpropagation according to the loss.

[0042] Optionally, the Encoder network includes:

[0043] The first layer consists of two 64×3×3 convolutions and max-pooling downsampling in sequence;

[0044] The second layer consists of two 128×3×3 convolutions and max-pooling downsampling in sequence;

[0045] The third layer consists of three 256×3×3 convolutions and max-pooling downsampling in sequence;

[0046] The fourth layer consists of three 512×3×3 convolutions and max-pooling downsampling in sequence;

[0047] The fifth layer consists of three 5212×3×3 convolutions and max-pooling downsampling in sequence;

[0048] The Decoder network includes:

[0049] The first layer consists of anti-max-pooling upsampling and three 521×3×3 convolutions in sequence;

[0050] The second layer consists of anti-max-pooling upsampling and three 521×3×3 convolutions in sequence;

[0051] The third layer consists of anti-max-pooling upsampling and two 256×3×3 convolutions in sequence;

[0052] The fourth layer consists of anti-max-pooling upsampling and two 128×3×3 convolutions in sequence;

[0053] The fifth layer consists of anti-max-pooling upsampling and two 64×3×3 convolutions in sequence.

[0054] Optionally, during the iterative training process, Pixel Accuracy is used as the accuracy evaluation index to obtain the closeness between the prediction result and the label.

[0055] In a second aspect, the present invention provides a meteorological radar clutter classification and recognition device, and the device includes:

[0056] A preprocessing module for obtaining and preprocessing test meteorological radar data;

[0057] A classification and recognition module for inputting the preprocessed test meteorological radar data into a trained SegNet network model to obtain classification and recognition results;

[0058] A model training module for:

[0059] Obtaining and preprocessing training meteorological radar data; the meteorological radar data includes radar reflectivity, differential reflectivity, differential propagation phase shift rate, correlation coefficient, and differential phase;

[0060] Integrating the preprocessed meteorological radar data into corresponding training data sets based on data types;

[0061] Respectively calculating the standard deviations of the training data sets corresponding to radar reflectivity and differential phase as the texture data of radar reflectivity and differential phase;

[0062] Using a fuzzy logic algorithm to process the training data and texture data in the training data set to obtain a label set of particle phase types;

[0063] Initializing the SegNet network model and performing iterative training based on the training data set and the label set until the preset maximum number of iterations is reached or the weight parameters of the SegNet network model converge.

[0064] In a third aspect, the present invention provides a meteorological radar clutter classification and recognition device, including a processor and a storage medium;

[0065] The storage medium is used to store instructions;

[0066] The processor is used to operate according to the instructions to execute the steps of the above method.

[0067] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the above method are implemented.

[0068] Compared with the prior art, the beneficial effects achieved by the present invention:

[0069] A meteorological radar clutter classification and recognition method and device provided by the present invention can effectively classify and recognize meteorological radar clutter by constructing a five-channel training data set through obtaining radar reflectivity, differential reflectivity, differential propagation phase shift rate, correlation coefficient, and differential phase, obtaining a label set of training data through a fuzzy logic algorithm, training a SegNet network model through the training data set and the label set, and classifying and recognizing meteorological radar clutter through the trained SegNet network model, so as to obtain meteorological radar data with higher quality. Brief Description of the Drawings

[0070] Figure 1 FIG. is a flowchart of a method for classifying and identifying weather radar clutter provided in Embodiment 1 of the present invention. Detailed Description of the Invention

[0071] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0072] Embodiment 1:

[0073] As Figure 1 shown, the present invention provides a method for classifying and identifying weather radar clutter, including:

[0074] 1. Obtain the test weather radar data and perform preprocessing;

[0075] 2. Input the preprocessed test weather radar data into the trained SegNet network model to obtain the classification and identification result;

[0076] Among them, the training process of the SegNet network model includes:

[0077] S101. Obtain the training weather radar data and perform preprocessing;

[0078] S102. Integrate the preprocessed weather radar data into corresponding training data sets based on the data type;

[0079] S103. Calculate the standard deviations of the training data sets corresponding to the radar reflectivity and differential phase respectively as the texture data of the radar reflectivity and differential phase;

[0080] S104. Use the fuzzy logic algorithm to process the training data and texture data in the training data set to obtain the label set of the particle phase state type;

[0081] S105. Initialize the SegNet network model and perform iterative training based on the training data set and the label set until the preset maximum number of iterations is reached or the weight parameters of the SegNet network model converge.

[0082] Among them, the iterative training based on the training data set and the label set includes:

[0083] S201. Downsample the training data based on the Encoder network of the SegNet network model to obtain data features;

[0084] S202. Upsample the feature data based on the Decoder network of the SegNet network model to obtain the prediction result;

[0085] S203. Calculate the loss between the prediction result and its label based on the cross - entropy loss function, and update the weight parameters of the SegNet network model through backpropagation according to the loss.

[0086] (1) The meteorological radar data provided in this embodiment includes radar reflectivity, differential reflectivity, differential propagation phase shift rate, correlation coefficient, and differential phase, and is saved in a numpy file.

[0087] (2) The pre - processing includes data cleaning and scale expansion.

[0088] Data cleaning includes finding the NaN (null data) values in the meteorological radar data and the radar reflectivity less than zero, and setting them to zero respectively.

[0089] Scale expansion includes expanding the meteorological radar data in the way of aligning the upper left corner and filling zeros in the lower right corner. Generally, the size of the collected data is 336×920, and in order to meet the requirements of the max - pooling operation of the SegNet network model, it needs to be expanded to 384×1088.

[0090] (3) Calculating the standard deviations of the training data sets corresponding to the radar reflectivity and differential phase respectively includes:

[0091] Traverse the training data set corresponding to the radar reflectivity, and calculate the standard deviation S D (Z H ) of the radar reflectivity within 1 km:

[0092]

[0093]

[0094] In the formula, m(Z H ) is the average value of the radar reflectivity Z H within 1 km, and n Z is the number of data points of the radar reflectivity within 1 km.

[0095] Traverse the training data set corresponding to the differential phase, and calculate the standard deviation of the differential phase within 2 km

[0096]

[0097]

[0098] In the formula, m(φ DP ) is the average value of the differential phase φ DP within 2 km, and n φ is the number of data points of the differential phase within 2 km.

[0099] (4) The use of the fuzzy logic algorithm to process the training data set and the texture data to obtain the label set of the particle phase type includes:

[0100] Using the trapezoidal function as the membership function of the fuzzy logic algorithm, taking the training data in the training data set and the texture data as input parameters for fuzzy logic operations, and obtaining the membership degree corresponding to the particle phase type as the fuzzy logic output; the trapezoidal function is:

[0101]

[0102] In the formula, X1, X2, X3, and X4 are the threshold parameters of the trapezoidal function, and P (j) (x i ) is the membership degree of the jth input parameter to the ith particle phase type;

[0103] Using the weighted average decision method to defuzzify the fuzzy logic output and obtain the integrated membership value corresponding to the particle phase type:

[0104]

[0105] In the formula, S i is the integrated membership value of the jth input parameter to the ith particle phase type, W ij is the weight coefficient of the jth input parameter of the ith particle phase type, and J is the number of input parameters;

[0106] If the integrated membership value of the input parameter is 0, the particle phase type corresponding to the input parameter is no meteorological echo, and its label is set to 0;

[0107] If the integrated membership value of the input parameter is greater than the set threshold, the particle phase type corresponding to the input parameter is clutter, and its label is set to 1;

[0108] If the integrated membership value of the input parameter is less than the set threshold and greater than 0, the particle phase type corresponding to the input parameter is other, and its label is set to 2;

[0109] Construct the label set of the particle phase type according to the labels of the input parameters.

[0110] (5) Convert the training data set of numpy type into tensor type data of (5×384×1088), and combine the training data and its corresponding label into a dataset. Set the batch of the training data set to 1, and use the DataLoader method in pytorch to load the tensor of size (1×channel×384×1088), where the channel of the training data is 5 (the training data of 5 data types are used as 5 input channel data), and the channel of the label is 1.

[0111] The structure of the designed Encoder network is as follows:

[0112] The first layer consists of two 64×3×3 convolutions and max-pooling downsampling in sequence;

[0113] The second layer consists of two 128×3×3 convolutions and max-pooling downsampling in sequence;

[0114] The third layer consists of three 256×3×3 convolutions and max-pooling downsampling in sequence;

[0115] The fourth layer consists of three 512×3×3 convolutions and max-pooling downsampling in sequence;

[0116] The fifth layer consists of three 5212×3×3 convolutions and max-pooling downsampling in sequence;

[0117] According to this network structure, training data of (1×5×384×1088) is processed to obtain feature data of (1×512×6×17), where 512 represents the number of feature channels.

[0118] The structure of the designed Decoder network is as follows:

[0119] The first layer consists of inverse max-pooling upsampling and three 521×3×3 convolutions in sequence;

[0120] The second layer consists of inverse max-pooling upsampling and three 521×3×3 convolutions in sequence;

[0121] The third layer consists of inverse max-pooling upsampling and two 256×3×3 convolutions in sequence;

[0122] The fourth layer consists of inverse max-pooling upsampling and two 128×3×3 convolutions in sequence;

[0123] The fifth layer consists of inverse max-pooling upsampling and two 64×3×3 convolutions in sequence.

[0124] According to this network structure, feature data of (1×512×6×17) is processed to obtain prediction data of (1×3×384×1088), where 3 represents the three categories corresponding to the number of channels of the labels.

[0125] Among them, the Encoder network also includes:

[0126] Set the parameter size of the convolution kernel in the Encoder network of the SegNet network model to 3, the stride to 1, and the padding parameter to 1, indicating the use of a square convolution kernel with a size of 3×3 and a stride of 1. When the convolution operation exceeds the data edge, a circle of zeros is padded around the data to prevent out-of-bounds exceptions. The same convolution is used for different convolution layers to extract the features of the training data. The expression of the convolution operation is as follows:

[0127]

[0128]

[0129] In the formula, H in represents the length of the input data, W in represents the width of the input data, K H represents the length of the convolution kernel, K W represents the width of the convolution kernel, P represents the number of circles of zero padding, and S represents the stride of the convolution kernel.

[0130] Since the data distribution characteristics after each convolution are inconsistent, the subsequent convolution layer needs to continuously adapt to the output changes of the previous convolution layer, reducing the decoupling between layers in the network. Use batch normalization to normalize the data before convolution, fix its distribution characteristics, and make the network learning more stable. The expression of batch normalization is as follows:

[0131]

[0132] In the formula, (x i ) (b) represents the value of the i th -th input node of this layer when sampling the b th samples of the current batch, μ(x i ) is the mean of the input data, σ(x i ) is the standard deviation of the input data, ∈ is a very small value to prevent division by zero, and γ and β are scale and shift parameters to be learned.

[0133] Use the Relu activation function to increase the non-linear relationship between convolution layers and avoid the situation of gradient disappearance. Relu will make a part of the convolution output zero, which results in the sparsity of the network, and reduces the interdependence of parameters, alleviating the occurrence of overfitting problems. The expression of Relu is as follows:

[0134] f(x) = max(0, x)

[0135] For the data after ReLU, perform max pooling downsampling to pick out these data points with more information, effectively removing redundant information and achieving feature extraction. While downsampling, save the index information of the maximum value. The expression for max pooling is as follows:

[0136]

[0137] In the formula, f represents the size of the pooling kernel, and s represents the stride of the pooling kernel movement.

[0138] The Decoder network also includes:

[0139] For the feature data after the unpooling convolution operation, gradually restore the size of the data. Since max pooling records the index of the maximum value, during unpooling, directly restore the data at the index position and fill 0 at other positions, thereby gradually restoring the original size of the data. The expression for unpooling is as follows:

[0140] X out =(X in -1)×s+f

[0141] The Decoder network uses same convolution to reduce the number of data feature channels after unpooling. The parameters of the convolution kernel are size 3, stride 1, and padding parameter 1. The length and width dimensions of the data before and after convolution in different convolution layers do not change, only the number of feature channels is changed. For the convolved data, use batch normalization to stabilize the distribution characteristics of the data and improve the learning ability of the network. Secondly, use the ReLU activation function to increase the non-linearity between convolution layers and avoid the situation of gradient disappearance.

[0142] (6) Calculating the loss between the prediction result and its label based on the cross-entropy loss function includes:

[0143] Use the softmax function to calculate the probabilities of various categories of the prediction data. The sum of the probabilities at the corresponding positions on each channel is 1. The index of the channel with the maximum probability is the predicted label of the data point. The expression for Softmax is as follows:

[0144]

[0145] In the formula, X i is the data point on the i th channel.

[0146] Calculate the cross-entropy loss between the prediction data and the label as the loss between the predicted value and the true value. Use the backpropagation mechanism in pytorch to update the gradients and parameter values of the weights in the SegNet network model to improve the prediction accuracy of SegNet. The expression for the cross-entropy loss function is as follows:

[0147]

[0148] Wherein, p i is the true label value, and q i is the predicted data value after softmax calculation.

[0149] (7) During the iterative training process, the Pixel Accuracy is used as the accuracy evaluation index to obtain the closeness between the prediction result and the label;

[0150] According to the value of the hyperparameter learning rate, select an appropriate number of training times, and continuously train the SegNet network model. During the training iteration process, save the model parameters with the current minimum loss. After the network training is completed, load the optimal network model, input the test data set, obtain the predicted data, and use the Pixel Accuracy as the accuracy evaluation index to check the closeness between the predicted data and the true label value. The expression for calculating the Pixel Accuracy is as follows:

[0151]

[0152] Wherein, k is the number of target classifications, and P ii is the number of correctly predicted pixels, and P ij is the number of pixels that belong to class i but are classified as class j.

[0153] Example Two:

[0154] The embodiment of the present invention provides a meteorological radar clutter classification and recognition device, and the device includes:

[0155] A preprocessing module, which is used to obtain the test meteorological radar data and perform preprocessing;

[0156] A classification and recognition module, which is used to input the preprocessed test meteorological radar data into the trained SegNet network model to obtain the classification and recognition result;

[0157] A model training module, which is used for:

[0158] Obtain the training meteorological radar data and perform preprocessing; the meteorological radar data includes radar reflectivity, differential reflectivity, differential propagation phase shift rate, correlation coefficient, and differential phase;

[0159] Integrate the preprocessed meteorological radar data into the corresponding training data set based on the data type;

[0160] Calculate the standard deviations of the training data sets corresponding to the radar reflectivity and differential phase respectively as the texture data of the radar reflectivity and differential phase;

[0161] Use the fuzzy logic algorithm to process the training data and texture data in the training dataset to obtain the label set of the particle phase type;

[0162] Initialize the SegNet network model, and perform iterative training based on the training dataset and the label set until the preset maximum number of iterations is reached or the weight parameters of the SegNet network model converge.

[0163] Example Three:

[0164] Based on Example One, an embodiment of the present invention provides a weather radar clutter classification and recognition device, including a processor and a storage medium;

[0165] The storage medium is used to store instructions;

[0166] The processor is used to operate according to the instructions to execute the steps of the above method.

[0167] Example Four:

[0168] Based on Example One, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the above method are implemented.

[0169] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product 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.

[0170] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for realizing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0171] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction device that implements the functions specified in one or more of the processes and / or blocks Figure 1 one or more of the processes and / or blocks Figure 1 specified in the block or blocks.

[0172] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the processes and / or blocks Figure 1 one or more of the processes and / or blocks Figure 1 specified in the block or blocks.

[0173] The foregoing is only a preferred embodiment of the present invention, and it should be noted that for those of ordinary skill in the art, several improvements and modifications can be made without departing from the technical principle of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for classifying and identifying meteorological radar clutter, characterized in that, Including: Obtain the measured meteorological radar data and perform preprocessing; Input the preprocessed measured meteorological radar data into the trained SegNet network model to obtain the classification and recognition results; Among them, the training process of the SegNet network model includes: Obtain the training meteorological radar data and perform preprocessing; the meteorological radar data includes radar reflectivity, differential reflectivity, differential propagation phase shift rate, correlation coefficient, and differential phase; Integrate the preprocessed meteorological radar data into corresponding training data sets based on the data type; Calculate the standard deviations of the training data sets corresponding to the radar reflectivity and differential phase respectively as the texture data of the radar reflectivity and differential phase; Use the fuzzy logic algorithm to process the training data and texture data in the training data set to obtain the label set of the particle phase state type; Initialize the SegNet network model, and perform iterative training based on the training data set and the label set until the preset maximum number of iterations is reached or the weight parameters of the SegNet network model converge.

2. The method for classifying and identifying clutter of a meteorological radar according to claim 1, characterized in that The preprocessing includes data cleaning and scale augmentation; The data cleaning includes finding the NaN data values and radar reflectivities less than zero in the meteorological radar data, and setting them to zero respectively; The scale augmentation includes augmenting the meteorological radar data in the way of aligning the upper left corner and filling zeros in the lower right corner.

3. A method for classifying and identifying clutter in a weather radar according to claim 1, characterized in that, The calculating the standard deviations of the training data sets corresponding to the radar reflectivity and differential phase respectively includes: Traverse the training dataset corresponding to the radar reflectivity and calculate the standard deviation S of the radar reflectivity within 1 km D (Z H ): Wherein, m(Z H ) is the average value of the radar reflectivity Z H within 1 km, and n Z is the number of data points of the radar reflectivity within 1 km; Traverse the training data set corresponding to the differential phase, and calculate the standard deviation of the differential phase within 2 km; where m(φ DP ) is the average value of the differential phase φ DP within 2 km, and n φ is the number of data points of the differential phase within 2 km.

4. A clutter classification and recognition method for a weather radar according to claim 1, characterized in that The using the fuzzy logic algorithm to process the training data set and the texture data to obtain the label set of the particle phase state type includes: Use the trapezoidal function as the membership function of the fuzzy logic algorithm, and use the training data and texture data in the training data set as input parameters for fuzzy logic operation to obtain the membership degree corresponding to the particle phase state type as the fuzzy logic output; the trapezoidal function is: where X1, X2, X3, and X4 are the threshold parameters of the trapezoidal function, and P (j) (x i ) is the membership degree of the j-th input parameter to the i-th particle phase state type; Use the weighted average decision method to defuzzify the fuzzy logic output to obtain the integrated membership value corresponding to the particle phase state type: where S i is the membership degree integration value of the j-th input parameter for the i-th particle phase state type, and W ij is the weight coefficient of the j-th input parameter for the i-th particle phase state type, and J is the number of input parameters; If the integrated membership value of the input parameter is 0, the particle phase state type corresponding to the input parameter is no meteorological echo, and its label is set to 0; If the integrated membership value of the input parameter is greater than the set threshold, the particle phase state type corresponding to the input parameter is clutter, and its label is set to 1; If the integrated membership value of the input parameter is less than the set threshold and greater than 0, the particle phase state type corresponding to the input parameter is other, and its label is set to 2; Construct the label set of the particle phase state type according to the labels of the input parameters.

5. A method for classifying and identifying clutter in a weather radar according to claim 1, characterized in that, The iterative training based on the training data set and the label set includes: Downsample the training data based on the Encoder network of the SegNet network model to obtain data features; Upsample the feature data based on the Decoder network of the SegNet network model to obtain the prediction results; Calculate the loss between the prediction result and its label based on the cross-entropy loss function, and update the weight parameters of the SegNet network model according to the loss by backpropagation.

6. A method for classifying and identifying clutter of a weather radar according to claim 5, characterized in that The Encoder network includes: The first layer is composed of two convolutions of 64×3×3 and max-pooling downsampling in sequence; The second layer consists of two 128×3×3 convolutions and max-pooling downsampling in sequence; The third layer consists of three 256×3×3 convolutions and max-pooling downsampling in sequence; The fourth layer consists of three 512×3×3 convolutions and max-pooling downsampling in sequence; There is an error in this item. It should be corrected to a reasonable value before translation. Assuming it is "512" instead of "5212", the fifth layer consists of three 512×3×3 convolutions and max-pooling downsampling in sequence; The Decoder network includes: The first layer consists of inverse max-pooling upsampling and three 512×3×3 convolutions in sequence; The second layer consists of inverse max-pooling upsampling and three 512×3×3 convolutions in sequence; The third layer consists of inverse max-pooling upsampling and two 256×3×3 convolutions in sequence; The fourth layer consists of inverse max-pooling upsampling and two 128×3×3 convolutions in sequence; The fifth layer consists of inverse max-pooling upsampling and two 64×3×3 convolutions in sequence.

7. A method for classifying and identifying weather radar clutter according to claim 5, characterized in that During the iterative training process, Pixel Accuracy is used as the accuracy evaluation metric to obtain the closeness between the prediction result and the label.

8. A clutter classification and recognition device for a meteorological radar, characterized in that The device includes: A preprocessing module for obtaining and preprocessing the test meteorological radar data; A classification and recognition module for inputting the preprocessed test meteorological radar data into the trained SegNet network model to obtain the classification and recognition result; A model training module for: Obtaining and preprocessing the training meteorological radar data; the meteorological radar data includes radar reflectivity, differential reflectivity, differential propagation phase shift rate, correlation coefficient, and differential phase; Integrating the preprocessed meteorological radar data into corresponding training data sets based on the data type; Calculating the standard deviations of the training data sets corresponding to radar reflectivity and differential phase respectively as the texture data of radar reflectivity and differential phase; Using a fuzzy logic algorithm to process the training data and texture data in the training data set to obtain the label set of particle phase states; Initializing the SegNet network model and performing iterative training based on the training data set and the label set until the preset maximum number of iterations is reached or the weight parameters of the SegNet network model converge.

9. A clutter classification and recognition device for a meteorological radar, characterized in that, It includes a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method according to any one of claims 1-7.

Images