Doppler Radar Clutter Recognition and Classification Method Based on SE-Res2Net-101

Abstract

The present invention discloses a method for identifying and classifying Doppler radar clutter based on SE-Res2Net-101, including extracting radar reflectivity Z in the base data H , differential reflectivity Z DR , correlation coefficient ρ HV and differential propagation phase shift rate K DP numerical data; preprocessing the extracted numerical data to generate a data set; constructing an SE-Res2Net-101 model; inputting the training set data into SE-Res2Net-101 for model training; inputting the test data set into the trained model, and making clutter judgments through the prediction probability given by the model. Solve the technical problem of current clutter detection technology that requires digital simulation of ground terrain and the provision of DEM, and establish a clutter prediction model from the variables of Z H , Z DR , ρ HV and K DP A method for identifying clutter in a Doppler radar and classifying biological clutter and ground clutter based on a residual convolutional network with an SE attention mechanism, which improves accuracy and speeds up the prediction speed

Description

Technical Field

[0001] The present invention relates to the field of Doppler radar, and specifically to a method for identifying and classifying Doppler radar clutter based on SE-Res2Net-101. Background Art

[0002] In the existing technology, clutter is recognized based on the fact that clutter is usually distributed within a certain spatial range and its physical size is much larger than the radar resolution unit. It is often divided into two categories: surface clutter and volume clutter. Of course, there are also "point" or discrete clutter, such as special structures like TV towers and buildings. Since clutter is similar to target radar, physical methods such as radar wavelength, polarization characteristics, illumination area, illumination direction, and even ground shape are required to identify it. Several radar clutter statistical models have been established successively, including Rayleigh distribution, lognormal distribution, Weibull distribution, and K distribution, etc. Analyzing clutter and establishing accurate clutter statistical models and corresponding simulation methods can, on the one hand, provide a realistic clutter environment model for radar simulators; on the other hand, it also helps the design and implementation of radar clutter filters, improves the ability to suppress clutter, and enhances radar detection performance.

[0003] In the weather radar echo reflectivity image, problems such as clutter are often caused by obstacles such as insects, birds, trees, buildings, and mountains. Therefore, the research on radar environmental characteristics is of great significance for improving radar performance, especially in the face of modern target stealth technology and the threat of ultra-low altitude penetration, which becomes even more important. Therefore, a method for identifying and classifying Doppler radar clutter based on SE-Res2Net-101 is proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for identifying and classifying Doppler radar clutter based on SE-Res2Net-101, to solve the technical problem of current clutter detection technology that requires digital simulation of the ground terrain and the provision of DEM, and to establish a clutter prediction model from these variables of radar reflectivity Z H , differential reflectivity Z DR , correlation coefficient ρ HV , differential propagation phase shift rate K DP to help reduce the cost and improve the accuracy of clutter identification, and a method for identifying clutter in a Doppler radar and classifying biological clutter and ground clutter based on a residual convolutional network with SE attention mechanism, realizing a clutter detection method with higher accuracy, faster prediction speed, and simpler implementation.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] Doppler Radar Clutter Recognition and Classification Method Based on SE-Res2Net-101, the clutter recognition and classification method comprising the following steps:

[0007] Step 1, extract the radar reflectivity Z H , differential reflectivity Z DR , correlation coefficient ρ HV and differential propagation phase shift rate K DP numerical data.

[0008] Step 2, preprocess the Z H , Z DR , ρ HV , K DP numerical data in Step 1 to generate a data set.

[0009] Step 3, construct the SE-Res2Net-101 model.

[0010] Step 4, input the training set data into SE-Res2Net-101 for model training.

[0011] Step 5, input the test data set into the trained model, and make clutter judgment through the prediction probability given by the model.

[0012] Furthermore, for the numerical data preprocessing in Step 2, the preprocessing comprises the following steps:

[0013] S1. Verify the dimensions of the read radar reflectivity Z H , differential reflectivity Z DR , correlation coefficient ρ HV , differential propagation phase shift rate K DP , and delete the files that do not conform to the aforementioned data dimensions, where the dimensions of N0H, N0C, N0K, and N0X are all 360*1200, and the dimension of N0R is 360*230.

[0014] S2. Process the data dimensions of the read N0R.

[0015] S3. Extract the required clutter.

[0016] S4. Split the N0H data to generate a matrix with dimensions of 30*30, take the most frequent radar echo classification in N0H as the label, and count the number of times this classification appears in the 30*30 matrix. Discard the matrices with the number of appearances less than 2.5%.

[0017] S5. For the N0X, N0C, N0K, and N0R data, only take the dimensions extracted from N0H, and fill the values of N0X, N0C, N0K, and N0R that are not the corresponding clutter labels with the average value of the label corresponding to the current matrix.

[0018] S6. Structurally store the valid data extracted from the above-mentioned NOX, NOC, NOK, NOR, and NOH, and record the corresponding file names in text format respectively. For clutter, the value extracted from the NOH file is used as a label, and for non-clutter, all labels are recorded as 0.

[0019] S7. Read the corresponding NOC, NOK, NOR, and NOX data from the text files of the corresponding labels, and stack them in the first dimension to obtain a 120*30 matrix stacked by four 30*30 matrices. The matrix contains the data involved in the training data.

[0020] S8. Divide the data provided in S7 into a training set in a ratio of 7:2:1, use Python to dynamically generate a Bash script, and use batch processing for data division.

[0021] Furthermore, the construction of the SE-ResNet-101 model is specifically as follows:

[0022] 1) SE-Res2Net-101 is improved on the basis of the previous ResNet, and the receptive field is expanded compared to large convolutional kernels by stacking layers.

[0023] 2) Construct a Res2Net_module. The 3×3 convolutional kernel inside the Bottleneckblock in ResNet is improved into a multi-scale non-homogeneous convolutional combination, and residual settings are made for different layers of the layer to prevent gradient dispersion and gradient explosion during training. This structure is called a Res2Net_module.

[0024] 3) Construct an SE_Res2Net_module. Since four variables Z H 、Z DR 、ρ HV 、K DP need to be weighted when judging clutter, a squeeze-and-excitation attention module se_module is added after each Res2Net_module to allow the model to self-learn the weight relationship of the four variables, making the model more accurate when judging clutter. The combination of the Res2Net_module and the se_module is called an SE_Res2Net_module.

[0025] 4) Initialize the training parameters of the SE-Res2Net-101 neural network model. Set the input as a matrix with a batch size of 256, 4 channels, and lengths and widths of 30 and 30 respectively, that is, the input dimension is 256 * 4 * 30 * 30. The output is 256 * the number of clutter classifications, where 256 represents the batch size, and the learning rate lr is 0.001. The convolutional kernel sizes of the hidden layers are of two types, 1 * 1 and 3 * 3 respectively. Given the repetition times of the SE_Res2Net_module as [3, 4, 23, 3], construct a 101-layer network, and set the final outfeatures of the network to the number of clutter types.

[0026] Further, after the first 1×1 convolutional layer of the Res2Net_module, divide the input into N sub-feature sets, defined as X i , i ∈ 1, 2, 3, …, N, and the input feature of each channel is and the channel feature dimension is all 3 * 3. In the module that performs convolutional stacking on N input subsets, a residual structure is used to prevent the situation of gradient dispersion and gradient explosion during training due to excessive stacking layers. Except for X1, all N - 1 sub-feature sets contain a 3 * 3 convolutional kernel K i (), and the X i after passing through the convolutional kernel K i is y i = K i (X i ), and the sum of X i and y i-1 is given to y i as the input. The specific formula is as follows:

[0027]

[0028] Further, the se_module includes compression, excitation, and scaling.

[0029] The compression operation: Perform feature compression along the spatial dimension, convert each two-dimensional feature channel into a real number. The real number has a global receptive field to a certain extent, and the dimensions of the input feature and the output feature match each other. The compression operation enables the global feature distribution situation to be obtained on the channel feature and enables the first ResNet Block to obtain the global feature. The key formula of the compression operation is as follows:

[0030] X → U, X ∈ R W′×H′×C′ , U ∈ R W×H×C

[0031]

[0032]

[0033] Among them, the first formula represents the conversion from X to U, and the number of channels is changed through convolution, where v c represents the c-th Convolutional kernel, and x s represents the value output from the (s - 1)-th convolution. Multiplying it with v c can obtain

[0034] The excitation operation: Use the W parameter to generate corresponding weights for each channel feature, where the parameter is automatically learned to explicitly construct the relationship between channel features. The key formula for the excitation operation is as follows:

[0035] s1 = F ex (z, W) = σ(g(z, W)) = σ(W2δ(W1z))

[0036] Among them, W1z represents the dot product operation between W1 and z, which is an operation of the fully connected layer. The dimension of W1 is C / S * C, where S in C / S * C is the scaling factor. S is selected as 8 according to the training samples, and the purpose of scaling is to reduce the computational amount by reducing the number of channels. Then, the obtained result passes through the ReLU activation function and is then multiplied by W2 and undergoes the operation of the fully connected layer. Finally, the result is limited within the interval (0, 1) through the Sigmoid activation function to obtain s1, where the dimension of s1 is 1 * 1 * C, and C represents the number of channels.

[0037] The scaling: Consider the weights output by the excitation as the importance of each feature channel after feature selection, and then multiply them channel by channel to weight the output features obtained after passing through the Res2Net_module, completing the recalibration of the original features in the channel dimension. The key formula for the scaling operation is as follows:

[0038]

[0039] Among them, the scale operation is the channel-wise multiplication operation, which performs a matrix dot product operation on u with dimensions W×H×C c and s with dimensions 1×1×C c , where s c is the weight corresponding to each channel.

[0040] The compression operation is implemented using the global average method. Subsequently, two fully connected layers are constructed to form a Bottleneck architecture to establish the relationship between the overall channels and output weights with the same number as the input feature. The input feature is reduced to 1 / 8 of the original. After that, the ReLu activation function is used, and then the reduced 1 / 8 dimension is restored to the original dimension through a fully connected layer. Then, the Sigmoid activation function is used to obtain values of the weights in the range of [0,1]. Finally, through the Scale operation, the values obtained through the activation function are weighted to each channel feature.

[0041] Furthermore, the training of the SE-Res2Net-101 specifically includes the following steps:

[0042] S1. Use the training set data divided in step 2 to train the SE-Res2Net-101. The data is sliced in the first dimension. Slicing with a size of 30 can obtain four 30*30 matrices, and the four matrices are stacked in the channel layer and input into the SE-Res2Net-101.

[0043] S2. Select the CrossEntropy loss function. The formula of the CrossEntropy loss function is:

[0044]

[0045] where y i,k represents the true label k of the i-th specimen, there are a total of k label values and N samples, and p i,k represents the probability that the i-th specimen is predicted as the k-th label value.

[0046] S3. Set the epoch to 600. For each epoch, calculate the current accuracy. If the accuracy is higher than the highest value, save the model trained in this epoch. Use the SGD optimizer to optimize the W parameter and cooperate with the Momentum mechanism to make the gradient smoother so as to try to ensure finding the global optimal solution. Among them, the key formula of SGD is as follows:

[0047]

[0048] where θ represents the weight parameter of the feature, η represents the learning rate, represents the step size required to move in the direction of the current gradient descent.

[0049] The key formula of the Momentum is as follows:

[0050]

[0051] θ = θ - v t

[0052] γ represents the learning rate, and v t represents the momentum of the t-th movement.

[0053] Furthermore, the specific process of testing in step 5 using the SE-Res2Net-101 test set is as follows:

[0054] (1) The test data is sliced in the first dimension and sliced into 4 matrices of 30*30 with a size of 30. The 4 matrices are stacked in the channel layer and input into SE-Res2Net-101.

[0055] (2) Take the maximum value in the classification prediction values, and determine the classification to which the obtained prediction value belongs. If the label is 0, it represents non-clutter, and other labels represent their corresponding clutter classifications.

[0056] Advantages of the present invention:

[0057] 1. The clutter recognition and classification method of the present invention only needs to provide the radar reflectivity Z H , differential reflectivity Z DR , correlation coefficient ρ HV , differential propagation phase shift rate K DP These four variables can be used to determine whether it is clutter and classify it. This method only needs to train one model and does not require any other equipment to judge clutter, which improves the speed of clutter recognition and conforms to physical rules;

[0058] 2. The clutter recognition and classification method of the present invention solves the technical problem of current clutter detection that requires digital simulation of ground terrain and the provision of DEM. Based on the radar reflectivity Z H , differential reflectivity Z DR , correlation coefficient ρ HV , differential propagation phase shift rate K DP These variables are used to establish a clutter prediction model to help reduce the cost and improve the accuracy of clutter recognition;

[0059] 3. The clutter recognition and classification method of the present invention is a method for recognizing clutter in a Doppler radar and classifying biological clutter and ground clutter based on a residual convolutional network with SE attention mechanism, realizing a clutter detection method with higher accuracy, faster prediction speed, and simpler implementation;

[0060] 4. The clutter recognition and classification method of the present invention replaces the large 3×3 convolutional kernel in the block of ResNet with multiple small 3×3 convolutional kernels, and uses a residual structure in the module for convolutional stacking of N input subsets to solve the problems of gradient dispersion and gradient explosion during training caused by excessive stacking layers, making the model easier to train and improving the robustness of the model. Description of the Drawings

[0061] The present invention will be further described below with reference to the accompanying drawings.

[0062] Figure 1 is a flowchart of the clutter recognition and classification method of the present invention;

[0063] Figure 2 is the overall network framework diagram of the clutter recognition and classification method of the present invention. Detailed Embodiments

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0065] The Doppler radar clutter recognition and classification method based on SE-Res2Net-101, as Figure 1 、 Figure 2 shown, the clutter recognition and classification method includes the following steps:

[0066] Step 1: Extract the radar reflectivity Z H 、differential reflectivity Z DR 、correlation coefficient ρ HV 、differential propagation phase shift rate K DP from the relevant files in the basic data

[0067] Specifically, extracting the relevant files in the basic data means that traditionally, parameters such as differential reflectivity, differential propagation phase shift rate, correlation coefficient, and radar reflectivity are only applied to precipitation measurement. However, different features of meteorology and clutter can also be extracted from these parameters. Write a Bash script to operate on the data set, and extract the radar reflectivity Z H 、differential reflectivity Z DR 、correlation coefficient ρ HV 、differential propagation phase shift rate K DP . The file names of their corresponding files contain the characters N0R, N0X, N0C, and N0K respectively. The extracted files are assembled into a preliminary data set and the corresponding file names are recorded in text format respectively

[0068] Step 2: Extract Z in Step 1 H 、Z DR 、ρ HV 、K DP from the relevant files of numerical data, and superimpose and integrate the four kinds of numerical data of Z H 、Z DR 、ρ HV 、K DP to generate several 120*30 matrix data

[0069] Read the file information extracted in Step 1, and preprocess the extracted data. The data preprocessing steps include:

[0070] (1) Verify the dimensions of the radar reflectivity Z H 、differential reflectivity Z DR 、correlation coefficient ρ HV 、differential propagation phase shift rate K DP in terms of dimensions, and delete the files that do not meet the aforementioned data dimensions. The dimensions of N0H, N0C, N0K, and N0X are all 360*1200, while the dimension of N0R is 360*230.

[0071] (2) Process the data dimension read from N0R: Since the data dimension read from N0R is different from other dimensions, if the dimension of N0R is not processed, the four-layer channel dimensions of the model will be inconsistent and the model cannot be trained. Therefore, it is necessary to repeat and expand N0R five times. At this time, the second dimension changes from 230 to 1150. For the missing 50 dimensions, the value obtained by taking the modulus of 1150 by 23 can be used as the dimension sequence. Read the data of the dimension sequence obtained by taking the modulus of 360*1150-dimensional N0R data sequence and store it. At this time, taking the modulus can obtain 360*50-dimensional data. Concatenating this 360*50 with 360*1150 can obtain 360*1200-sized N0R data.

[0072] (3) Extract the required clutter: Among them, the data corresponding to the clutter with values of 10 (biological) and 20 (ambiguous echo, ground clutter) in NOH, etc., and mask the data of non-clutter. And compress the obtained N0H matrix after masking into one dimension. If the unmasked data in the obtained one-dimensional matrix is less than 5%, discard this matrix. In addition to extracting the required clutter, it is also necessary to extract non-clutter data as the counterexample data of the clutter data. The method of extracting data is the same as the above method of extracting clutter data. After masking the clutter data, compress it into one dimension. However, the non-clutter data contains many categories, such as small raindrops, large raindrops, hail, drizzle, etc. In the experiment, they are all merged into the 0 label to represent the non-clutter data.

[0073] (4) Split the N0H data to generate a 30*30 dimensional matrix. Classify the most frequent radar echoes of N0H as labels, and count the number of times this classification appears in the 30*30 matrix. Matrices with an appearance frequency less than 2.5% are directly discarded.

[0074] (5) For the N0X, N0C, N0K, and N0R data, only extract the dimensions obtained from N0H, and fill the values of N0X, N0C, N0K, and N0R that are not the corresponding clutter labels with the average value of the labels corresponding to the current matrix.

[0075] (6) Structurally store the valid data extracted from the above N0X, N0C, N0K, N0R, and N0H, and record the corresponding file names in text format. For clutter, the values extracted from the N0H file are used as labels. For non-clutter, all labels are recorded as 0.

[0076] (7) Read the corresponding N0C, N0K, N0R, and N0X data from the text files of the corresponding labels and stack them along the first dimension. A 120*30 matrix can be obtained by stacking 4 30*30 matrices, and this matrix contains the data involved in the training data.

[0077] (8) Divide the data provided above into a training set in a ratio of 7:2:1. Use Python to dynamically generate a Bash script and perform data division using batch processing.

[0078] Step 3: Build the SE-Res2Net-101 model and initialize all parameters, including the batch size batch_size, learning rate lr, the number of stacked layers layer of the network, the SE model, and the size k and number n of the convolutional kernels in the hidden layer

[0079] The steps to build the SE-ResNet-101 model include:

[0080] 1) SE-Res2Net-101 makes improvements in many aspects based on the previous ResNet. By stacking layers, it can more effectively expand the receptive field than large convolutional kernels. Therefore, the 3×3 convolutional kernels inside the Bottleneck block in ResNet are improved into a multi-scale non-homogeneous convolutional combination, and residual settings are made for different layers of layer to prevent gradient dispersion and gradient explosion during training. This structure is called the Res2Net_module; since the use of Z H 、Z DR 、ρ HV 、K DPFour variables need to be weighted when judging clutter. Therefore, a squeeze-and-excitation attention module se_module is added after each Res2Net_module to enable the model to self-learn the weight relationship of the four variables, so that the model has a better clutter judgment effect. The combination of the Res2Net_module and the se_module is called the SE_Res2Net_module.

[0081] 2) Construct the Res2Net_module. After the first 1×1 convolutional layer of the Res2Net_module, the input is divided into N sub-feature sets, defined as X i , i ∈ 1, 2, 3, …, N, and the input feature of each channel is and the channel feature dimension is all 3*3; a residual structure is used in the module that performs convolutional stacking on the N input subsets to prevent the occurrence of gradient dispersion and gradient explosion during training due to excessive stacking layers; except for X1, all N - 1 sub-feature sets contain a 3*3 convolutional kernel K i (), and X i () after passing through the convolutional kernel K i is y i = K i (X i ), and the sum of X i and y i-1 is given to y i as the input. The specific formula is as follows:

[0082]

[0083] 3) Construct the SE_Res2Net_module. Since the four variables Z H , Z DR , ρ HV , K DP need to be weighted when judging clutter, a squeeze-and-excitation attention module se_module is added after each Res2Net_module to enable the model to self-learn the weight relationship of the four variables, so that the model is more accurate when judging clutter.

[0084] The se_module includes squeezing, excitation, and scaling;

[0085] Compression operation: Perform feature compression along the spatial dimension to convert each two-dimensional feature channel into a real number. The real number has a global receptive field to a certain extent, and the dimensions of the input feature and the output feature also match each other. The compression operation can not only obtain the global feature distribution of the channel feature, but also enable the first ResNet Block to obtain the global feature.

[0086] The key formula for the compression operation is as follows:

[0087] X→U, X∈R W′×H′×C′ , U∈R W×H×C

[0088]

[0089]

[0090] Among them, the first formula represents the conversion from X to U, changing the number of channels through convolution, and v c represents the c-th Convolutional kernel, and x s represents the value output from the (s - 1)-th convolution. Multiplying it by v c can obtain

[0091] Activation operation: Similar to the gate mechanism in RNN, mainly to control the numerical size and play a role in current limiting. Generate corresponding weights for each channel feature through the W parameter, where the parameter is automatically learned to explicitly construct the relationship between channel features. The key formula for the activation operation is as follows:

[0092] s1 = F ex (z, W) = σ(g(z, W)) = σ(W2δ(W1z))

[0093] Among them, W1z represents the dot product operation of W1 and z, which is an operation of the fully connected layer. The dimension of W1 is C / S * C, and S in C / S * C is the scaling factor. S is selected as 8 according to the training samples, and the purpose of scaling is to reduce the computational amount by reducing the number of channels. Then, the obtained result passes through the ReLU activation function and then multiplies with W2 and performs the operation of the fully connected layer. Finally, the obtained result is limited within the interval (0, 1) through the Sigmoid activation function to obtain s1, where the dimension of s1 is 1 * 1 * C, and C represents the number of channels.

[0094] Scaling: Consider the Weight of the excitation output as the importance of each feature channel after feature selection, and then multiply each channel by this weight to the output features obtained after the Res2Net_module to complete the recalibration of the original features in the channel dimension. The key formula for the scaling operation is as follows:

[0095]

[0096] Among them, the scale operation is the channel-wise multiplication operation, which performs a matrix dot product operation on the \(W\times H\times C\) \(u\) c and the \(1\times1\times C\) \(s\) c , where \(s\) c is the weight corresponding to each channel.

[0097] The compression operation is implemented using the global average method. Subsequently, two fully connected layers are constructed to form a Bottleneck architecture to establish the relationship between the overall channels and output weights with the same number as the input feature; the input feature is reduced to 1 / 8 of the original, and after using the ReLu activation function, the dimension of the reduced 1 / 8 is restored to the original dimension through a fully connected layer; then the Sigmoid activation function is used to obtain the value of the weight in the range of [0,1], and finally, through the Scale operation, the values obtained by the activation function before are weighted to each channel feature.

[0098] 4) Initialize the training parameters of the SE-Res2Net-101 neural network model. Set the input as a matrix with a batch_size of 256, 4 channels, and lengths and widths of 30 and 30 respectively, that is, the input dimension is 256*4*30*30; the output is 256*the number of clutter classifications, where 256 represents the batch_size, the learning rate lr is 0.001; the convolutional kernel sizes of the hidden layers are of two types, 1*1 and 3*3 respectively. Given the number of repetitions of the SE_Res2Net_module as [3,4,23,3], a 101-layer network is constructed, and the final out features of the network are set to the number of clutter types.

[0099] Step 4: Input the training set data into SE-Res2Net-101 for model training. The specific training of SE-Res2Net-101 is as follows:

[0100] (1) Use the training set data divided in step 2 to train SE-Res2Net-101. Split the data along the first dimension. When splitting with a size of 30, four matrices of 30*30 can be obtained. Stack the four matrices along the channel dimension and input them into SE-Res2Net-101.

[0101] (2) Select the CrossEntropy loss function. The formula for the CrossEntropy loss function is:

[0102]

[0103] where y i,k represents the true label k of the i-th specimen. There are a total of k label values and N samples. p i,k represents the probability that the i-th specimen is predicted as the k-th label value.

[0104] (3) Set the number of epochs to 600. For each epoch, calculate the current accuracy. If the accuracy is higher than the highest value, save the model trained in this epoch. Use the SGD optimizer to optimize the W parameter and combine it with the Momentum mechanism to make the gradient smoother to try to ensure finding the global optimal solution.

[0105] The key formula for SGD is as follows:

[0106]

[0107] where θ represents the weight parameter of the feature, η represents the learning rate, represents the step size required to move in the direction of the current gradient descent.

[0108] The key formula for Momentum is as follows:

[0109]

[0110] θ = θ - v t

[0111] γ represents the learning rate, and v t represents the momentum of the t-th movement.

[0112] Step 5: Input the test data set into the trained model. Determine whether it belongs to clutter and which type of clutter it belongs to based on the predicted probability given by the model. The specific test using the SE-Res2Net-101 test set is as follows:

[0113] (1) Split the test data in the first dimension, split it into 4 matrices of 30*30 with a size of 30, and stack the 4 matrices and input them into SE-Res2Net-101 in the channel layer.

[0114] (2) Take the maximum value in the classification prediction value, judge the category to which the obtained prediction value belongs. If it is the 0 label, it represents non-clutter, and other labels represent their corresponding clutter classifications.

[0115] In the description of this specification, the description with reference to terms such as "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0116] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed.

Claims

1. A method for Doppler radar clutter recognition and classification based on SE-Res2Net-101, characterized in that, The clutter recognition and classification method includes the following steps: Step 1: Extract the radar reflectivity Z, differential reflectivity Z, correlation coefficient ρ, and differential propagation phase shift rate K numerical data from the basic data H , differential reflectivity Z DR , correlation coefficient ρ HV and differential propagation phase shift rate K DP numerical data; Step 2: Preprocess the Z in Step 1 H Z DR ρ HV K DP Generate a dataset by preprocessing numerical data Step 3: Construct an SE-Res2Net-101 model; Step 4: Input the training set data into SE-Res2Net-101 for model training; Step 5: Input the test data set into the trained model, and make clutter judgments based on the prediction probabilities given by the model.

2. The method for Doppler radar clutter recognition and classification based on SE-Res2Net-101 according to claim 1, characterized in that In the numerical data preprocessing in Step 2, the preprocessing includes the following steps: S1. Verify the radar reflectivity Z read out H , differential reflectivity Z DR , correlation coefficient ρ HV , differential propagation phase shift rate K DP Perform verification in terms of dimensions, and delete files that do not conform to the aforementioned data dimensions. The dimensions of N0H, N0C, N0K, and N0X are all 360 * 1200, and the dimension of N0R is 360 * 230; S2: Process the data dimension read from N0R; S3: Extract the required clutter; S4: Split the N0H data to generate a 30*30 dimensional matrix. Classify the most frequently occurring radar echoes in N0H as labels, and count the number of times this classification appears in the 30*30 matrix. Matrices with an appearance frequency less than 2.5% are directly discarded; S5: For the N0X, N0C, N0K, and N0R data, only take the dimensions extracted from N0H, and fill the values of N0X, N0C, N0K, and N0R that are not the corresponding clutter labels with the average value of the corresponding label in the current matrix; S6: Structurally store the valid data extracted from the above N0X, N0C, N0K, N0R, and N0H, and record the corresponding file names in text format. For clutter, the numerical values extracted from the N0H file are used as labels. For non-clutter, all labels are recorded as 0; S7: Read the corresponding N0C, N0K, N0R, and N0X data from the text files of the corresponding labels, and stack them in the first dimension to obtain a 120*30 matrix stacked by 4 30*30 matrices. The matrix contains the data involved in the training data; S8: Divide the data provided in S7 into a training set in a ratio of 7:2:

1. Use Python to dynamically generate a Bash script and use batch processing for data division.

3. The method for Doppler radar clutter recognition and classification based on SE-Res2Net-101 according to claim 1, wherein The construction of the SE-Res2Net-101 model is specifically as follows: 1) SE-Res2Net-101 is improved based on the previous ResNet. By stacking layers, it realizes an enlarged receptive field compared to large convolutional kernels; 2) Construct a Res2Net_module. The 3×3 convolutional kernel inside the Bottleneck block in ResNet is improved into a multi-scale non-hierarchical convolutional combination, and residual settings are made for different layers to prevent gradient dispersion and gradient explosion during training. This structure is called Res2Net_module; 3) Construct the SE_Res2Net_module. Since the use of Z H and Z DR , ρ HV , and K DP need to consider weights when judging clutter, add a squeeze-and-excitation attention module se_module after each Res2Net_module to enable the model to self-learn the weight relationship of the four variables, making the model more accurate in judging clutter. Combine the two modules of Res2Net_module and se_module and call it SE_Res2Net_module; 4) Initialize the training parameters of the SE-Res2Net-101 neural network model. Set the input as a matrix with a batch size of 256, 4 channels, and lengths and widths of 30 and 30 respectively, that is, the input dimension is 256 * 4 * 30 * 30; the output is 256 * the number of clutter classifications, where 256 represents the batch size, the learning rate lr is 0.001; the convolutional kernel sizes of the hidden layers are of two types, 1 * 1 and 3 * 3 respectively. Give the number of repetitions of the SE_Res2Net_module as [3, 4, 23, 3], construct 101 layers, and set the final out features of the network to the number of clutter types.

4. The method for Doppler radar clutter recognition and classification based on SE-Res2Net-101 according to claim 3, wherein After the first 1×1 convolutional layer of the constructed Res2Net_module, the input is divided into N sub-feature sets, defined as X i , i ∈ 1, 2, 3, …, N, and the input feature of each channel is and the channel feature dimension is all 3*3; in the module where convolutional superposition is performed on N input subsets, a residual structure is used to prevent the occurrence of gradient dispersion and gradient explosion during training due to excessive stacking of layers; except for X1, all N - 1 sub-feature sets contain a 3*3 convolutional kernel K i (), after the convolutional kernel K i () of X i is y i = K i (X i ), X i and y i-1 are added to y i as the input, and the specific formula is as follows:

5. The method for Doppler radar clutter recognition and classification based on SE-Res2Net-101 according to claim 3, characterized in that The se_module includes compression, excitation, and scaling; The compression operation: Perform feature compression along the spatial dimension, convert each two-dimensional feature channel into a real number. The real number has a global receptive field to a certain extent, and the dimensions of the input feature and the output feature match each other; the compression operation enables the global feature distribution obtained on the channel feature and allows the first ResNetBlock to obtain the global feature. The key formula for the compression operation is as follows: X → U, X ∈ R W′×H′×C′ , U ∈ R W×H×C Among them, the first formula represents the conversion from X to U, changing the number of channels through convolution, and v c represents the c-th Convolutional kernel, and x s represents the value output from the (s - 1)-th convolution. Multiplying it with v c can obtain The excitation operation: Generate corresponding weights for each channel feature through the W parameter, where the parameter is automatically learned to explicitly construct the relationship between channel features. The key formula for the excitation operation is as follows: s1 = F ex (z, W) = σ(g(z, W)) = σ(W2δ(W1z)) Among them, W1z represents the dot product operation of W1 and z, which is the operation of the fully connected layer. The dimension of W1 is C / S * C, and S in C / S * C is the scaling factor. S is selected as 8 according to the training samples. The purpose of scaling is to reduce the amount of computation by reducing the number of channels; then the obtained result is passed through the ReLU activation function and then multiplied by W2 and then the operation of the fully connected layer is performed. Finally, the obtained result is limited within the interval (0, 1) through the Sigmoid activation function to obtain s1, where the dimension of s1 is 1 * 1 * C, and C represents the number of channels; The scaling: Regard the weights output by the excitation as the importance of each feature channel after feature selection, and then weight each channel by multiplication to the output features obtained after passing through the Res2Net_module, completing the recalibration of the original features in the channel dimension. The key formula for the scaling operation is as follows: Among them, the scale operation is the channel-wise multiplication operation, which multiplies the u of W×H×C c with the s of 1×1×C c by matrix dot product operation, and s c is the weight corresponding to each channel; The compression operation is implemented using the global average method. Subsequently, two fully connected layers are constructed to form a Bottleneck architecture to build the relationship between the overall channels and output weights with the same number as the input feature; the input feature is reduced to 1 / 8 of the original, and after using the ReLu activation function, the reduced 1 / 8 dimension is restored to the original dimension through a fully connected layer; then the Sigmoid activation function is used to obtain the value of the weight in the range of [0, 1], and finally, each channel feature is weighted by the value obtained through the activation function through the Scale operation.

6. The method for Doppler radar clutter recognition and classification based on SE-Res2Net-101 according to claim 3, characterized in that, The training of SE-Res2Net-101 specifically includes the following steps: S1. Use the training set data divided in step 2 to train SE-Res2Net-101. The data is sliced in the first dimension. Slicing with a size of 30 can obtain 4 matrices of 30*30, and the four matrices are stacked in the channel layer and input into SE-Res2Net-101. S2. The loss function selects the CrossEntropy loss function, and the formula of the CrossEntropy loss function is: Among them, y i,k represents the true label k of the i-th specimen, with a total of k label values and N samples, p i,k represents the probability that the i-th specimen is predicted as the k-th label value; S3. Set the epoch to 600. For each epoch, calculate the current accuracy. If the accuracy is higher than the highest value, save the model trained in this epoch. Use the SGD optimizer to optimize the W parameter and cooperate with the Momentum mechanism to make the gradient smoother to try to ensure finding the global optimal solution. Among them, the key formula of SGD is as follows: where θ represents the weight parameter of the feature and η represents the learning rate, which represents the step size required to move in the direction of the current gradient descent; The key formula of the Momentum is as follows: θ = θ - v t γ represents the learning rate, and v t represents the momentum of the t-th movement.

7. The method for Doppler radar clutter recognition and classification based on SE-Res2Net-101 according to claim 1, wherein The specific operation of using the SE-Res2Net-101 test set in step 5 is as follows: (1) Slice the test data in the first dimension. Slicing with a size of 30 can obtain 4 matrices of 30*30, and the 4 matrices are stacked in the channel layer and input into SE-Res2Net-101. (2) Take out the maximum value in the classification prediction values, and judge the category to which the obtained prediction value belongs. If the label is 0, it represents non-clutter, and other labels represent their corresponding clutter classifications.

Images