A ResNeXt network meteorological target fine-grained recognition method fused with LightGBM

By combining LightGBM and ResNeXt networks, high-precision and fine-grained recognition of meteorological radar signals are achieved, solving the problems of insufficient recognition accuracy and insufficient granularity in the prior art, and providing more refined and accurate meteorological target recognition results.

CN117849803BActive Publication Date: 2024-11-22NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311759416.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-20
Publication Date
2024-11-22
Estimated Expiration
2043-12-20

AI Technical Summary

Technical Problem

The existing meteorological radar technology has problems such as insufficient recognition accuracy and insufficient granularity in the recognition of meteorological targets and non-meteorological targets. In particular, the block recognition results of deep learning algorithms lead to insecure distance unit types, resulting in identification deviations.

Method used

Using a method of integrating LightGBM and ResNeXt networks, coarse-grained identification is performed through the ResNeXt network, and refined identification is performed using LightGBM to obtain high-precision and fine-grained meteorological target recognition results.

Benefits of technology

The identification accuracy and particle size of meteorological targets and non-meteorological targets are improved, and the identification results are consistent with the particle size of the radar sampling unit, which improves the accuracy and meteorological target classification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117849803B_ABST
    Figure CN117849803B_ABST
Patent Text Reader

Abstract

The present invention discloses a fine-grained meteorological target recognition method of ResNeXt network integrated with LightGBM, including: making a data set; building a ResNeXt network model, training an optimal model and testing; inputting the difference radar distance unit data of the coarse-grained recognition result of the data set in the ResNeXt network optimal model and the reference result into the LightGBM classifier for training and testing; using measured data, first obtaining the coarse-grained recognition result with the ResNeXt network optimal model, if the recognition is correct, saving the recognition result of the point, if the judgment is wrong, inputting the difference distance unit data into the LightGBM classifier to obtain the fine-grained meteorological target recognition result. The present invention can finely and accurately identify meteorological targets and clutter, and its result is finer than the recognition result of the ResNeXt network model, and the accuracy is better than that of the LightGBM classifier, which lays a strong foundation for the classification of meteorological targets.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a weather radar signal processing technology, and in particular to a ResNeXt network weather target fine-grained recognition method integrating LightGBM. Background Art

[0002] Meteorological radar is the infrastructure for meteorological target identification, and meteorological target classification is of great significance in human life and action. Conventional Doppler meteorological radar relies on the emission of electromagnetic waves in a single polarization direction, and can only obtain single polarization direction parameters such as reflectivity factor and velocity spectrum width through echoes. However, it is difficult to understand the specific information of meteorological targets in terms of geometric size, shape, spatial orientation, etc. by relying solely on parameters in a single polarization direction, which often leads to misjudgment of meteorological targets. In order to obtain more information about meteorological targets, dual-polarization radar was born. Dual-polarization meteorological radar is a further development of conventional Doppler meteorological radar. It uses electromagnetic waves in horizontal and vertical directions to emit simultaneously or alternately, and can additionally obtain differential polarization parameters such as differential reflectivity, correlation coefficient, and differential phase shift rate between electromagnetic waves in two different polarization directions. These polarization parameters enable dual-polarization meteorological radar to have the ability to quickly sense information such as the phase, type, and geometric shape of meteorological targets. The emergence of dual-polarization meteorological radar has greatly promoted the development of meteorological radar in meteorological target identification.

[0003] In actual radar work, when the radar scans and illuminates ground objects or aerial organisms, its echo contains non-meteorological echoes in addition to meteorological echoes. Non-meteorological echoes (mainly biological clutter and ground clutter) are mixed with meteorological echoes, affecting the quality of meteorological echo data and the classification of meteorological targets. Therefore, the study of the discrimination between non-meteorological targets and meteorological targets has always been an important part of meteorological radar data quality control.

[0004] At present, many achievements have been made in the identification of meteorological targets and non-meteorological targets. In 2019, GOLBON-HAGHIGHI et al. proposed a new three-dimensional discriminant function as part of their proposed improved ground clutter identification algorithm, which effectively distinguishes the probability density function of weather signals and clutter signals, and then selects a simple Bayesian classifier. Finally, the best decision based on the 3D discriminant function is made to distinguish clutter from weather signals. In 2020, WANG et al. proposed an airborne meteorological radar meteorological target detection algorithm based on pitch angle spatial processing. According to the difference in the interference phase data between the clutter signal and the weather signal, the interference phase data is used as a test statistic to analyze the statistical characteristics of the interference phase between the clutter signal and the weather signal, and a constant false alarm rate detector is designed to realize weather signal detection. In 2021, JATAU et al. used the high sensitivity characteristics of the S-band radar and the proposed bird (insect) features as data-driven to train a mature ridge regression classifier, which was combined with the hydrometeor classification algorithm to effectively distinguish bird echoes from insect echoes. In 2023, SASIDHARAN et al. used meteorological radar polarization parameters and polarization parameter spatial texture parameters to develop a new fuzzy logic algorithm to distinguish and eliminate ground clutter and abnormal propagation. In addition, some researchers began to use convolutional neural networks to distinguish meteorological targets from non-meteorological targets. In 2020, YU et al. applied the basic CNN network to the field of airborne meteorological radar echo recognition and completed the recognition of weather signals in meteorological radar echoes. In 2022, Gao Yongxing et al. proposed a dual-polarization meteorological radar weather signal detection algorithm based on residual convolutional neural network, which effectively distinguishes weather signals from clutter signals.

[0005] Deep learning algorithms have powerful data recognition and classification capabilities. Just like the meteorological target recognition method based on residual convolutional neural network used by Gao Yongxing et al., this deep learning algorithm can achieve extremely good recognition rate. However, this recognition method needs to be analogous to image classification. The data sample is a block matrix, and the recognition result is also based on blocks as the basic unit and distributed in blocks, so the recognition result is relatively rough. In reality, radar obtains information in radar distance units as the basic unit. The block sample unit contains multiple distance units, and the classification distribution of meteorological targets is often discrete and uncertain, resulting in non-unique distance unit types in the sample block. Therefore, this type of algorithm is prone to recognition bias in small areas within the block, and the recognition accuracy is insufficient.

[0006] Based on the above analysis, in order to improve the accuracy and granularity of existing technologies in identifying meteorological targets and non-meteorological targets, it is very important to study new meteorological target classification algorithms. Summary of the invention

[0007] Purpose of the invention: The purpose of the present invention is to provide a fine-grained recognition method for meteorological targets in a ResNeXt network integrated with LightGBM.

[0008] Technical solution: A fine-grained meteorological target recognition method of ResNeXt network integrated with LightGBM of the present invention comprises the following steps:

[0009] S1. Select NOAA weather radar observation data to create a data set;

[0010] S2. Build a ResNeXt network model, use the data set to train the optimal model and test it;

[0011] S3, the difference radar distance unit data between the prediction results and the reference results in the optimal model of the ResNeXt network is used as the input data of the LightGBM classifier, and the LightGBM classifier is trained and tested;

[0012] S4. Using measured data, the coarse-grained recognition results are first obtained through the ResNeXt network optimal model, and the coarse-grained recognition results are compared point-to-point with the NOAA original reference target Target. If the recognition is correct, the point recognition result is saved as the ResNeXt network optimal model recognition result CH; if the judgment is wrong, the difference radar distance unit data between the coarse-grained recognition result and the NOAA original reference target Target is input into the LightGBM classifier, and the LightGBM classifier is used to refine the recognition, optimize the coarse-grained recognition result, and save the difference radar distance unit point recognition result as the LightGBM recognition result LH; finally, the high-precision and fine-grained meteorological target recognition result that is consistent with the actual meteorological distribution is obtained through integration.

[0013] Furthermore, the dataset preparation method is:

[0014] First, the reflectivity factor Z in the NOAA weather radar observation data is h The two-dimensional matrix data is revised and calibrated so that it is consistent with the differential reflectivity factor Z in terms of matrix size and orientation. dr , differential phase shift rate K dp and the correlation coefficient ρ hv The reflectivity factor Z h , differential reflectivity factor Z dr , differential phase shift rate K dp and the correlation coefficient ρ hv The two-dimensional data matrix is ​​stacked to obtain a three-dimensional data array;

[0015] Secondly, the three-dimensional data array is cut and divided into blocks to obtain three-dimensional matrix blocks;

[0016] Then, the three-dimensional matrix blocks with invalid values ​​exceeding 80% are discarded, and the invalid values ​​are further removed from the remaining three-dimensional matrix blocks, and the categories with a proportion exceeding 50% in each three-dimensional matrix block are recorded as this matrix block category, and the matrix blocks with categories that do not account for more than 50% in the remaining three-dimensional matrix blocks are regarded as invalid data and discarded; then, the average of the valid values ​​of each channel matrix block in each of the remaining three-dimensional matrix blocks is selected to replace the invalid values ​​contained in each matrix block;

[0017] Finally, the data of each channel of each three-dimensional matrix block after removing invalid values ​​are normalized.

[0018] Furthermore, the ResNeXt network model includes a convolution, a maximum pooling, four series-connected residual blocks, an average pooling and a linear layer. The input data is processed by convolution, then subjected to a maximum pooling operation, and then processed by four residual blocks in turn. Finally, the coarse-grained recognition result is output after an average pooling operation and a linear operation.

[0019] Furthermore, the formula for the size transformation of the input and output feature maps of the convolutional layer and pooling layer in the ResNeXt network model is:

[0020]

[0021] Among them, Size in Size out They represent the size of the input and output feature maps respectively, Kernel_size represents the convolution kernel size or the pooling window size, Padding represents the padding size, and Stride represents the step size.

[0022] Furthermore, the residual block includes three convolutions and one group convolution. The input data of the residual block passes through convolution, group convolution and convolution in sequence in one way, and passes through a convolution in the other way. The outputs of the two ways are added as the output of the entire residual block.

[0023] Furthermore, the loss function in ResNeXt network model training is the cross entropy loss function.

[0024] Furthermore, during the training of the ResNeXt network model, the SGD optimizer is used to adjust the network weights and update the network parameters using gradient descent.

[0025] Furthermore, the classification method of LightGBM classifier is:

[0026] (1) Initialize sample x ij The predicted state F in the kth class k,0 (x ij ), the meteorological target classification label is k = 1, 2, ..., K; the execution formula is:

[0027] Fk,0 (x ij )=0

[0028] Among them, k is the meteorological target classification label, and K is the total number of categories;

[0029] (2) Iterative calculation, number of iterations m = 1, 2, ..., M;

[0030] a) Calculate the sample x ij The probability p of each category k,m -1(x ij ), the calculation formula is:

[0031]

[0032] Among them, F k,m-1 (x ij ) is the integrated prediction model of the first m-1 tree models;

[0033] b) Calculate the sample x ij The negative gradient Polarization parameter x ij The true probability is y i,k , we get the formula:

[0034]

[0035] c) Get the leaf node value γ after the leaf node is split h,k,m , the sample set on the leaf node is R h,k,m , the number of leaf nodes h=1,2,…,H, the leaf node value after the leaf node splits γ h,k,m The calculation formula is:

[0036]

[0037] d) Model optimization refresh, the learning rate is set to η, and the model optimization refresh formula is:

[0038]

[0039] Among them, F k,m-1 (x ij ) is the integrated prediction model of the first m-1 tree models, For the newly generated tree model in the mth round of iteration, I is an indicator function, indicating that the sample x ij The conditions for being assigned to a node;

[0040] (3) Obtain the final model, the formula is:

[0041]

[0042] F k(X) is the integration of all tree model functions.

[0043] Based on the same inventive concept, a ResNeXt network meteorological target fine-grained recognition system integrating LightGBM of the present invention includes:

[0044] The data set production unit is used to select NOAA weather radar observation data to produce data sets;

[0045] ResNeXt network model training and optimization unit, used to build the ResNeXt network model, and use the data set to train the optimal model and test it;

[0046] LightGBM classifier training and testing unit, used to use the difference radar distance unit data between the prediction results and the reference results of the dataset in the ResNeXt network optimal model as the input data of the LightGBM classifier, train the LightGBM classifier and test it;

[0047] The meteorological target recognition unit is used to detect the measured data. The coarse-grained recognition result is first obtained through the ResNeXt network optimal model, and the coarse-grained recognition result is compared point-to-point with the NOAA original reference target Target. If the recognition is correct, the point recognition result is saved as the ResNeXt network optimal model recognition result CH; if the judgment is wrong, the difference radar distance unit data between the coarse-grained recognition result and the NOAA original reference target Target is input into the LightGBM classifier, and the LightGBM classifier is used to refine the recognition and optimize the coarse-grained recognition result, and the difference radar distance unit point recognition result is saved as the LightGBM recognition result LH; finally, the high-precision and fine-grained meteorological target recognition result that is consistent with the actual meteorological distribution is obtained through integration.

[0048] Based on the same inventive concept, an electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the fine-grained identification method of meteorological targets in a ResNeXt network integrating LightGBM are implemented according to any one of claims 1 to 8.

[0049] Beneficial effects: Compared with the prior art, the method of the present invention aims at the problem of identifying meteorological targets and non-meteorological targets of dual-polarization radar, integrates the fine-grained recognition of LightGBM and the high recognition rate characteristics of ResNeXt network, combines ResNeXt network with LightGBM algorithm, and proposes LGBM-ResNeXt meteorological target recognition method to achieve high-precision and fine-grained recognition of meteorological targets. The ResNeXt network with fast and accurate selection ability is selected as the first-layer algorithm to obtain high-precision coarse-grained recognition results, and then the difference between this result and the actual label is input into the second-layer LightGBM classifier to obtain fine-grained recognition results that meet actual requirements. Experiments show that the ResNeXt network has a high recognition rate for meteorological target recognition. After integrating the LightGBM classifier, the recognition granularity is highly improved, which is consistent with the granularity of the radar sampling unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2 is a flow chart of data set preparation of the present invention;

[0052] Figure 3 It is a structural diagram of the ResNeXt network model of the present invention;

[0053] Figure 4 It is a diagram of the grouped residual convolution structure of the present invention;

[0054] Figure 5 It is a loss function transformation curve diagram of the training set and the validation set during the training process of the present invention;

[0055] Figure 6 It is a curve diagram of the accuracy conversion between the training set and the verification set during the training process of the present invention;

[0056] Figure 7 It is a confusion matrix diagram of the test set of the present invention on the ResNeXt network model;

[0057] Figure 8 The figure is a verification result diagram of the radar 0.5 degree elevation angle data of the KTLX dual-polarization weather radar on the NOAA website of the present invention at 2:56 on April 24, 2021 on the ResNeXt network model; wherein (a) is a reference classification target result diagram, (b) is a reference result diagram after data cutting, and (c) is a ResNeXt meteorological target recognition result diagram;

[0058] Fig. 9 It is a loss function transformation curve diagram of the training set and the validation set during the LightGBM classifier training process of the present invention;

[0059] Fig.10It is a curve diagram of the accuracy conversion between the training set and the verification set during the training process of the LightGBM classifier of the present invention;

[0060] Fig.11 This is a recognition result diagram of the radar 0.5 degree elevation angle data of the KTLX dual-polarization weather radar on the NOAA website of the present invention at 2:56 on April 24, 2021 on the LightGBM classifier;

[0061] Fig.12 This is a graph of the recognition results of the radar 0.5 degree elevation angle data of the KTLX dual-polarization weather radar on the NOAA website at 2:56 on April 24, 2021 on the proposed ResNeXt network meteorological target fine-grained recognition method integrated with LightGBM. DETAILED DESCRIPTION

[0062] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] The embodiment of the present invention discloses a fine-grained meteorological target recognition method of a ResNeXt network integrated with LightGBM. First, NOAA meteorological radar observation data is selected to produce a data set; secondly, a LGBM-ResNeXt meteorological target recognition model is built, and trained and optimized, including: building a ResNeXt network model and training and optimizing to obtain an optimal model and testing it, using the difference radar distance unit data between the prediction result and the reference result of the data set in the ResNeXt network optimal model as the input data of the LightGBM classifier, training the LightGBM classifier and testing it; then, using the measured meteorological data, the coarse-grained recognition result is first obtained through the ResNeXt network optimal model, and then the difference distance unit data between the coarse-grained recognition result and the reference result is input into the LightGBM classifier to obtain the fine-grained meteorological target recognition result. The present invention integrates the fine-grained recognition capability of LightGBM and the high-precision recognition characteristics of ResNeXt, and can finely and accurately identify meteorological targets and clutter. The result is finer than the ResNeXt network recognition result, and the recognition accuracy is better than the LightGBM classifier, which lays a strong foundation for meteorological target classification. Figure 1 As shown, the specific steps include:

[0064] Step S1, select NOAA weather radar observation data to make a data set;

[0065] Meteorological radar observation data specifically refers to the reflectivity factor, differential reflectivity factor, differential phase shift rate, correlation coefficient and NOAA classification results of the meteorological observation data of the National Oceanic and Atmospheric Administration (NOAA), denoted as Z h , Z dr , K dp , hv , Cls. As a feedforward multi-layer network with one-way information propagation, the convolutional neural network is driven by the image GRB channel information. It can automatically learn image features through layer-by-layer transformation and ultimately achieve classification. The four polarization parameters of reflectivity factor, differential reflectivity, correlation coefficient, and differential phase shift rate are used as identification input data, combined with NOAA classification labels to make data set labels.

[0066] There is a gap between these polarization parameters and image channel information. In order to classify and process the polarization parameters in NOAA weather radar observation data using the ResNeXt network, it is necessary to pre-process the polarization parameters such as reflectivity factor, differential reflectivity factor, differential phase shift rate, and correlation coefficient. While ensuring the universality of these polarization parameters, these polarization parameters are equivalent to image channel information. The present invention stacks the two-dimensional data matrix of the four polarization parameters in the spatial latitude to form a three-dimensional data array. Finally, these three-dimensional data arrays are used as data sets. Analogous to image classification, the ResNeXt network is used to complete the distinction between meteorological echoes and non-meteorological echoes. Specifically:

[0067] The reflectivity factor Z in the weather radar observation data provided by NOAA h is a 360×460 two-dimensional matrix, and the rest of the data are 360×1200 two-dimensional matrices, so the reflectivity factor Z h Correct the data so that it corresponds to other data in terms of matrix size and orientation to ensure data accuracy. The correction method is to select the reflectivity factor Z h The first 400 columns of data are used as the basic data, and each column is copied into 4 copies, so that the reflectivity factor Z h It corresponds to other data in both range and azimuth. After processing, the size of each data matrix is ​​360×1200, which does not match the actual image size. In addition, within a single scan range, since the classification distribution of meteorological targets is often discrete and uncertain, a large number of clutter targets and some invalid values ​​are mixed in the meteorological targets. Therefore, in order to ensure the quality of data samples in a single data matrix, it is necessary to avoid interference from mixed echoes as much as possible and process invalid value data. In summary, the radar polarization parameters need to be processed as follows to generate a data set suitable for ResNeXt network training. Figure 2As shown in the figure, first, the original polarization parameter matrix is ​​calibrated and stacked; secondly, it is cut and divided into blocks so that the original 360×1200 large-size two-dimensional matrix is ​​split into 750 24×24 small-size two-dimensional matrix blocks; then, it is judged whether the invalid values ​​in each split three-dimensional matrix block exceed 80%, and the matrix blocks with invalid value points exceeding 80% are discarded; the remaining matrix blocks still contain meteorological targets, clutter and some invalid values. Therefore, it is necessary to further judge whether the proportion of main category data in the remaining matrix blocks exceeds 50%. If not, the data in the matrix block is regarded as invalid data and the matrix block is discarded. Otherwise, the category with a proportion of more than 50% is recorded as the category of this matrix block. After all matrix blocks are screened, the mean of the valid values ​​in the remaining matrix blocks is further selected to replace the invalid values ​​contained in the matrix block, and each matrix block after processing the invalid value points is obtained; finally, the polarization parameters in each remaining matrix block are normalized to make the network model and data fitting more efficient. To ensure the balance of the two types of data samples, the sample sizes of meteorological targets and clutter need to be kept the same, and then the data set is randomly cut into training set, validation set and test set.

[0068] In the embodiment of the present invention, the KVNX dual-polarization weather radar level-III observation data for the whole year of 2021 is selected, and the polarization parameters and classification result data with a radar elevation angle of 0.5° are extracted from the downloaded compressed files: NOQ, NOX, NOC, NOK, and NOH files. These data files correspond to the reflectivity factor, differential reflectivity, correlation coefficient, differential phase shift rate, and NOAA classification results, respectively. And press Figure 2 The data set is produced by the process shown in the figure. The recognition of meteorological targets and clutter (biological clutter and ground clutter) is equivalent to image recognition. The size of the data set sample matrix is ​​preset to be 4×24×24. To ensure the balance of the two types of data samples, the number of meteorological target and clutter samples is 15,000. The data set is then randomly cut into a training set, a validation set, and a test set with a ratio of 7:2:1.

[0069] S2. Build a ResNeXt network model, use the data set to train the optimal model and test it;

[0070] With the continuous optimization of computers, convolutional neural networks (CNNs) have achieved excellent results in the field of image classification. Studies have shown that the deepening and expansion of convolutional neural networks can optimize the accuracy of traditional models, but as the number of network layers increases, it will lead to problems such as gradient explosion, gradient disappearance, and even network degradation. To solve these problems, the ResNet network proposed to replace the basic convolution structure with a residual convolution structure, thereby solving the network degradation caused by network deepening. Then the researchers proposed the ResNeXt network. The ResNeXt network model combines the stacking concept of the VGG network, the repetition strategy of the ResNet network, and the split-conversion-merge idea of ​​the Inception network. It not only solves the problems of gradient disappearance caused by network deepening, but also maintains the basic network structure of ResNet. It only needs to widen the residual convolution module to improve the recognition accuracy without increasing the complexity of the parameters, and to reduce the number of hyperparameters to a certain extent.

[0071] The main idea of ​​the ResNeXt network model is to perform group convolution on the basis of the ResNet model, which improves the model accuracy while keeping the complexity almost unchanged. By using the group residual convolution structure, the model accuracy can be further improved while keeping the complexity almost unchanged.

[0072] The recognition of meteorological targets and clutter (biological clutter and ground clutter) is equivalent to image recognition. The size of the data set sample matrix is ​​predetermined to be 4×24×24. In order to use the ResNeXt network model to classify the data set obtained in step S1, the present invention modifies the network parameters based on the ResNeXt original network model, and designs a ResNeXt network model suitable for the data set in step S1, thereby finally completing the recognition of meteorological targets and clutter.

[0073] like Figure 3 As shown in Figure 1, the ResNeXt network model includes a 3×3 convolution, a 3×3 maximum pooling, four serial residual blocks, a 1×1 average pooling, and a linear layer. The residual block includes three 1×1 convolutions and a 3×3 group convolution. The 3×3 group convolution structure is as follows: Figure 4 As shown in the figure, the input data of the residual block passes through 1×1 convolution, 3×3 group convolution and 1×1 convolution in sequence, and the other way passes through a 1×1 convolution. The two outputs are fused as the output of the entire residual block. The input data of the ResNeXt network model is the data set produced in step S1, which is the reflectivity factor Z h , differential reflectivity factor Z dr , differential phase shift rate K dp and the correlation coefficient ρ hvThe input data is processed by a 3×3 convolution, a 3×3 maximum pooling operation, and then processed by four residual blocks in sequence. Finally, the classification result is output after a 1×1 average pooling operation and a linear operation. The specific parameters of each layer module are shown in Table 1. By using the grouped residual convolution structure, the model accuracy can be further improved while ensuring that the complexity remains almost unchanged.

[0074] Table 1 Specific parameters of each module of ResNeXt

[0075]

[0076] In the ResNeXt network model, the formula for the transformation relationship between the input and output feature map sizes of the convolution layer and the pooling layer is:

[0077]

[0078] Among them, Size in Size out Indicates the size of the input and output feature maps, Kernel_size indicates the size of the convolution kernel or the pooling window, Padding indicates the padding size, and Stride indicates the step size. In addition to the convolution layer, pooling layer, and linear layer, the ResNeXt network model also uses functional layers such as the batch normalization layer (Batch Normalization, BN) and ReLu activation function in the ResNet network.

[0079] Using Pytorch as the deep learning framework, the ResNeXt network meteorological target recognition model and the LightGBM subclassifier were created. The kernel of this study is classification, so the cross entropy loss function is selected as the loss function of the ResNeXt network model, and then the SGD optimizer is used to adjust the network weights using gradient descent. The optimizer parameter settings are shown in Table 2. During the training process, the batch size is set to 128 and the number of iterations is set to 400 to ensure successful model training.

[0080] Table 2 SGD optimizer parameters

[0081]

[0082] An important feature of the deep learning process is to compare the gap between the learning results and the sample labels. The smaller the gap between the two, the stronger the model learning effect. In deep learning, the loss function is generally used to represent this gap. In theory, the smaller the loss function, the better, but it cannot be zero and can only be infinitely close to zero. When the loss function converges, it can be judged that the network model has learned successfully. In addition, classification accuracy is a powerful tool to describe the performance of classification algorithms in classification problems. Therefore, the loss function and accuracy are used to illustrate the performance of the model. Figure 5and Figure 6 It is the loss function and accuracy conversion curve of the training set and the validation set during the training process.

[0083] from Figure 5 It can be seen that with the increase in the number of iterations, the loss function of the ResNeXt network model training set and validation set rapidly decreases and stabilizes, and there is no overfitting phenomenon, which shows that the model training is successfully completed. Figure 6 It can be seen that the accuracy of the ResNeXt network model in identifying the training set and the validation set is at least more than 99.6%. To further prove the expressiveness of the model, the test set is used to conduct a preliminary test on the model. Figure 7 The confusion matrix of the test set on the trained ResNeXt network model shows that the ResNeXt network model has excellent recognition performance for both meteorological echoes and non-meteorological echoes, with the recognition accuracy of both targets exceeding 99.5%, and the overall recognition effect is excellent, which is consistent with the performance of the model during training.

[0084] In order to verify the model's predictive ability for actual radar observation data, the present invention selects non-training data set data: the radar 0.5 degree elevation angle observation echo data of the KTLX dual-polarization weather radar on the NOAA website at 2:56 on April 24, 2021 for testing.

[0085] The radar 0.5 degree elevation reference classification target results of the KTLX dual-polarization weather radar at 02:56 on April 24, 2021 on the NOAA website are as follows: Figure 8 As shown in (a), the reference result after data segmentation is as follows Figure 8 As shown in (b), the meteorological target recognition results of the ResNeXt network model are as follows Figure 8 As shown in (c), the ResNeXt network model identifies one clutter sample block as a meteorological target and two meteorological target sample blocks as clutter. Figure 8 Comparison between (b) and (c) shows that after data segmentation, the ResNeXt network model performs very well in meteorological target recognition overall, which is basically consistent with the segmentation reference image. The main misidentification of the model is that the meteorological target is identified as clutter, which is consistent with the confusion matrix of the test set. Figure 8 In (a) and (c), it can be seen that compared with the actual reference results, the recognition results of the ResNeXt network model are not detailed enough, and the coarse-grained recognition results CH are obtained. Therefore, in order to improve the recognition accuracy, further fine-grained classification and recognition are needed.

[0086] S3, the difference radar distance unit data between the coarse-grained recognition results of the dataset in the optimal model of the ResNeXt network and the NOAA reference results is used as the input data of the LightGBM classifier, and the LightGBM classifier is trained and tested;

[0087] LightGBM is developed from the traditional Gradient Boosting Decision Tree (GBDT). The LightGBM algorithm uses discrete attribute data as input data to represent the relationship between parameters and targets, so that point-to-point classification of data to targets can be completed based on radar range units.

[0088] The coarse-grained recognition result CH output by the ResNeXt network meteorological target recognition model is compared point-to-point with the original reference target label Target of NOAA. The difference radar distance unit point data between the coarse-grained recognition result output by the ResNeXt network and the original reference target label Target of NOAA is used as the input data of the LightGBM refinement classifier. The LightGBM classifier is trained with this data, and then the difference radar distance unit points are identified by the LightGBM classifier to refine the ResNeXt network recognition result. When the input data is i=1,2,3,…,n,j=0,1,2,3,where i represents the i-th data sample, j represents the j-th polarization parameter, n is the number of samples, C i represents the sample meteorological target category label, x ij is the sample data composed of four polarization parameters. The specific implementation of the algorithm is:

[0089] (1) Initialize sample x ij The predicted state F in the kth class k,0 (x ij ), the meteorological target classification label is k = 1, 2, ..., K. The execution formula is:

[0090] F k,0 (x ij )=0 (2)

[0091] Among them, k is the classification label of meteorological targets, and K has a total number of categories.

[0092] (2) Iterative calculation, the number of iterations m = 1, 2, …, M.

[0093] a) Calculate the sample x ij The probability of each category is calculated as follows:

[0094]

[0095] Among them, F k,m-1 (x ij ) is the integrated prediction model of the first m-1 tree models.

[0096] b) Calculate the sample x ijThe negative gradient Polarization parameter x ij The true probability is y i,k , we get the formula:

[0097]

[0098] c) Get the leaf node value γ after the leaf node is split h,k,m , the sample set on the leaf node is R h,k,m , the number of leaf nodes h=1,2,…,H, the leaf node value after the leaf node splits γ h,k,m The calculation formula is:

[0099]

[0100] d) Model optimization refresh, the learning rate is set to η, and the model optimization refresh formula is:

[0101]

[0102] Among them, F k,m-1 (x ij ) is the integrated prediction model of the first m-1 tree models, is the newly generated tree model function for the mth round of iteration, I is an indicator function, indicating that the sample x ij The condition to be assigned to the node.

[0103] (3) Obtain the final model. The formula is:

[0104]

[0105] Among them, F k (X) is the integration of all tree model functions.

[0106] Fig. 9 , Fig.10 This is the LightGBM classifier training process curve. Fig. 9 , Fig.10 As shown in the figure, with the increase of the number of iterations, the loss function and accuracy of the LightGBM classifier in the training set and the validation set quickly stabilized, and there was no overfitting phenomenon. The classifier training was successful. However, the recognition accuracy of the LightGBM classifier for meteorological targets and non-meteorological targets was only more than 98.6%, and the recognition ability was lower than that of the ResNeXt network model. Fig.11 The recognition results using only the LightGBM classifier are given, which is similar to Figure 8 The comparison in (a) shows that the LightGBM classifier has many misclassifications.

[0107] S4. Use the measured data to make a data set according to the method in step S1, and first obtain the coarse-grained recognition result through the optimal model of the ResNeXt network. If the coarse-grained recognition result CH obtained by the ResNeXt network model is established, then save the correct recognition result CH. If CH=Target is not established, then input the difference distance unit data between the coarse-grained recognition result and the reference data into the trained LightGBM classifier to obtain the fine-grained meteorological target recognition result LH;

[0108] First, use the measured data to make a matrix block sample data set, and then input the matrix block data set into the trained ResNeXt network optimal model to obtain the coarse-grained recognition result; compare the coarse-grained recognition result CH output by the ResNeXt meteorological target recognition model with the NOAA original reference target Target point-to-point. If the recognition is correct, the point recognition result is saved as the ResNeXt network optimal model recognition result CH; if the judgment is wrong, the difference radar distance unit data between the ResNeXt network optimal model judgment result and the NOAA original reference target Target is input into the trained LightGBM classifier, and then the LightGBM classifier performs refined recognition to optimize the coarse-grained recognition result, and saves the difference radar distance unit point recognition result as the LightGBM refined result LH; finally, the integration is used to obtain a high-precision and fine-grained recognition result that is consistent with the actual meteorological distribution. Fig.12 As shown by Figure 8 , Fig.12 By comparison, the non-meteorological echo recognition method of ResNeXt network fused with LightGBM can effectively identify meteorological targets and clutter, and the recognition effect is excellent. Compared with a single ResNeXt network, the recognition of meteorological targets is basically the same as the NOAA reference data, with significantly improved accuracy and finer recognition strength. Fig.11 , Fig.12 It can be seen that compared with the fusion algorithm, the LightGBM algorithm has more misjudgments, such as Fig.11 Shown in the box and ellipse.

[0109] The present invention selects reflectivity factor, differential reflectivity, correlation coefficient, differential phase shift rate, and NOAA classification results, and uses these parameters to produce a matrix block data set required for meteorological target recognition and classification suitable for the ResNeXt network model. The produced data set is then used to train the ResNeXt network model to obtain the recognition accuracy and loss function of the ResNeXt network model for the data set, and the measured data is used to verify the accurate performance of the ResNeXt network model. It is found that the ResNeXt network model has extremely high recognition accuracy but insufficient recognition granularity. Then a lightGBM classifier is built, and the lightGBM classifier is trained using radar distance unit data of the difference between the ResNeXt network optimal model recognition result and the reference result. It is found that the lightGBM classifier has excellent recognition granularity but lower recognition accuracy than the ResNeXt network model. Finally, the lightGBM classifier is integrated with the ResNeXt network model to obtain a recognition result with high accuracy and fine granularity. Starting from the accurate and fine recognition of actual data, the present invention proposes a ResNeXt network meteorological target fine-grained recognition method integrating LightGBM, which effectively utilizes the echo parameter information of the meteorological radar, makes full use of the high-precision recognition characteristics of the ResNeXt network model and the superiority of the LightGBM model in recognizing discrete information, improves the accuracy of meteorological target recognition, and can effectively cope with the fine-grained recognition of measured data, and can be effectively used in civil aviation route planning, weather forecasting and other fields.

[0110] Based on the same inventive concept, a ResNeXt network meteorological target fine-grained recognition system integrating LightGBM of the present invention includes:

[0111] The data set production unit is used to select NOAA weather radar observation data to produce data sets;

[0112] ResNeXt network model training and optimization unit, used to build the ResNeXt network model, and use the data set to train the optimal model and test it;

[0113] LightGBM classifier training and testing unit, used to use the difference radar distance unit data between the prediction results and the reference results of the dataset in the ResNeXt network optimal model as the input data of the LightGBM classifier, train the LightGBM classifier and test it;

[0114] The meteorological target recognition unit is used to detect the measured data. The coarse-grained recognition result is first obtained through the ResNeXt network optimal model, and the coarse-grained recognition result is compared point-to-point with the NOAA original reference target Target. If the recognition is correct, the point recognition result is saved as the ResNeXt network optimal model recognition result CH; if the judgment is wrong, the difference radar distance unit data between the coarse-grained recognition result and the NOAA original reference target Target is input into the LightGBM classifier, and the LightGBM classifier is used to refine the recognition and optimize the coarse-grained recognition result, and the difference radar distance unit point recognition result is saved as the LightGBM recognition result LH; finally, the high-precision and fine-grained meteorological target recognition result that is consistent with the actual meteorological distribution is obtained through integration.

[0115] Based on the same inventive concept, an electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the fine-grained meteorological target recognition method based on the above-mentioned ResNeXt network integrated with LightGBM are implemented.

[0116] Those skilled in the art will appreciate 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 complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. 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.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0117] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0120] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0121] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A ResNeXt network meteorological target fine-grained recognition method integrating LightGBM, characterized in that: The following steps are involved: S1. Select NOAA weather radar observation data to create a data set; S2. Build a ResNeXt network model, use the data set to train the optimal model and test it; S3, the difference radar distance unit data between the prediction results and the reference results in the optimal model of the ResNeXt network is used as the input data of the LightGBM classifier, and the LightGBM classifier is trained and tested; S4. Using measured data, the coarse-grained recognition results are first obtained through the ResNeXt network optimal model, and the coarse-grained recognition results are compared point-to-point with the NOAA original reference target Target. If the recognition is correct, the point recognition result is saved as the ResNeXt network optimal model recognition result CH; if the judgment is wrong, the difference radar distance unit data between the coarse-grained recognition result and the NOAA original reference target Target is input into the LightGBM classifier, and the LightGBM classifier is used to refine the recognition, optimize the coarse-grained recognition result, and save the difference radar distance unit point recognition result as the LightGBM recognition result LH; finally, the high-precision and fine-grained meteorological target recognition result that is consistent with the actual meteorological distribution is obtained through integration.

2. According to claim 1, a ResNeXt network meteorological target fine-grained recognition method integrating LightGBM is characterized in that: The dataset preparation method is: First, the reflectivity factor Z in the NOAA weather radar observation data is h The two-dimensional matrix data is revised and calibrated so that it is consistent with the differential reflectivity factor Z in terms of matrix size and orientation. dr , differential phase shift rate K dp and the correlation coefficient ρ hv The reflectivity factor Z h , differential reflectivity factor Z dr , differential phase shift rate K dp and the correlation coefficient ρ hv The two-dimensional data matrix is ​​stacked to obtain a three-dimensional data array; Secondly, the three-dimensional data array is cut and divided into blocks to obtain three-dimensional matrix blocks; Then, the three-dimensional matrix blocks with invalid values ​​exceeding 80% are discarded, and the invalid values ​​are further removed from the remaining three-dimensional matrix blocks, and the categories with a proportion exceeding 50% in each three-dimensional matrix block are recorded as this matrix block category, and the matrix blocks with categories that do not account for more than 50% in the remaining three-dimensional matrix blocks are regarded as invalid data and discarded; then, the average of the valid values ​​of each channel matrix block in each of the remaining three-dimensional matrix blocks is selected to replace the invalid values ​​contained in each matrix block; Finally, the data of each channel of each three-dimensional matrix block after removing invalid values ​​are normalized.

3. According to claim 1, a ResNeXt network meteorological target fine-grained recognition method integrating LightGBM is characterized in that: The ResNeXt network model includes a convolution, a maximum pooling, four series-connected residual blocks, an average pooling and a linear layer. The input data is processed by convolution, then by the maximum pooling operation, and then by the four residual blocks in turn. Finally, the coarse-grained recognition result is output after the average pooling operation and the linear operation.

4. According to claim 3, a ResNeXt network meteorological target fine-grained recognition method integrating LightGBM is characterized in that: The formula for the size transformation of the input and output feature maps of the convolutional layer and pooling layer in the ResNeXt network model is: Among them, Size in Size out They represent the size of the input and output feature maps respectively, Kernel_size represents the convolution kernel size or the pooling window size, Padding represents the padding size, and Stride represents the step size.

5. According to claim 3, a ResNeXt network meteorological target fine-grained recognition method integrating LightGBM is characterized in that: The residual block consists of three convolutions and one group convolution. The input data of the residual block passes through convolution, group convolution and convolution in sequence, and the other way passes through a convolution. The two outputs are added together as the output of the entire residual block.

6. According to claim 1, a ResNeXt network meteorological target fine-grained recognition method integrating LightGBM is characterized in that: The loss function in ResNeXt network model training is the cross entropy loss function.

7. According to claim 1, a ResNeXt network meteorological target fine-grained recognition method integrating LightGBM is characterized in that: During ResNeXt network model training, the SGD optimizer is used to adjust the network weights and update the network parameters using gradient descent.

8. According to claim 1, a ResNeXt network meteorological target fine-grained recognition method integrating LightGBM is characterized in that: The classification method of LightGBM classifier is: (1) Initialize sample x ij The predicted state F in the kth class k,0 (x ij ), the meteorological target classification label is k = 1, 2, ..., K; the execution formula is: F k,0 (x ij )=0 Among them, k is the meteorological target classification label, and K is the total number of categories; (2) Iterative calculation, number of iterations m = 1, 2, ..., M; a) Calculate the sample x ij The probability p of each category k,m-1 (x ij ), the calculation formula is: Among them, F k,m-1 (x ij ) is the integrated prediction model of the first m-1 tree models; b) Calculate the sample x ij The negative gradient Polarization parameter x ij The true probability is y i,k , we get the formula: c) Get the leaf node value γ after the leaf node is split h,k,m , the sample set on the leaf node is R h,k,m , the number of leaf nodes h=1,2,…,H, the leaf node value after the leaf node splits γ h,k,m The calculation formula is: d) Model optimization refresh, the learning rate is set to η, and the model optimization refresh formula is: Among them, F k,m-1 (x ij ) is the integrated prediction model of the first m-1 tree models, For the newly generated tree model in the mth round of iteration, I is an indicator function, indicating that the sample x ij The conditions for being assigned to a node; (3) Obtain the final model, the formula is: F k (X) is the integration of all tree model functions.

9. A ResNeXt network meteorological target fine-grained recognition system integrating LightGBM, characterized in that: include: The data set production unit is used to select NOAA weather radar observation data to produce data sets; ResNeXt network model training and optimization unit, used to build the ResNeXt network model, and use the data set to train the optimal model and test it; LightGBM classifier training and testing unit, used to use the difference radar distance unit data between the prediction results and the reference results of the dataset in the ResNeXt network optimal model as the input data of the LightGBM classifier, train the LightGBM classifier and test it; The meteorological target recognition unit is used to detect the measured data. The coarse-grained recognition result is first obtained through the ResNeXt network optimal model, and the coarse-grained recognition result is compared point-to-point with the NOAA original reference target Target. If the recognition is correct, the point recognition result is saved as the ResNeXt network optimal model recognition result CH; if the judgment is wrong, the difference radar distance unit data between the coarse-grained recognition result and the NOAA original reference target Target is input into the LightGBM classifier, and the LightGBM classifier is used to refine the recognition and optimize the coarse-grained recognition result, and the difference radar distance unit point recognition result is saved as the LightGBM recognition result LH; finally, the high-precision and fine-grained meteorological target recognition result that is consistent with the actual meteorological distribution is obtained through integration.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the steps of the fine-grained recognition method of meteorological targets in the ResNeXt network integrated with LightGBM according to any one of claims 1 to 8 are implemented.