A method for identifying electromagnetic interference echoes of weather radar based on target detection

Abstract

The present invention relates to a method for identifying electromagnetic interference echoes of weather radars based on target detection, comprising: obtaining radar base data to be identified and generating radar echo images to be identified; inputting the radar echo images to be identified into a preset target detection model to identify electromagnetic interference echoes of weather radars; wherein the target detection model is obtained by training an improved YOLOv8 model with a pre-constructed data set, and the data set includes: an electromagnetic interference echo image of a weather radar and annotated data. The technical solution provided by the present invention improves the data quality of weather radars and can be applied to the data processing of weather radars in existing business operations.

Description

Technical Field

[0001] The present invention relates to the technical field of weather radars, and in particular to a method for identifying electromagnetic interference echoes of weather radars based on target detection. Background Art

[0002] Severe convective weather has the characteristics of suddenness and strong locality, small scale, short life history, and serious disasters. It has always been the focus and difficulty in weather forecasting. As an active remote sensing observation instrument, weather radar is one of the fastest and most effective tools for detecting convective storms due to its high temporal and spatial resolution. It plays an irreplaceable role in severe convective monitoring and early warning. In recent years, with the continuous development of local economy and the continuous expansion of urban and rural construction, the overall situation of meteorological detection environment in various places is not optimistic. In particular, with the rapid development of electronic technology, various electronic products have entered all fields of society and all aspects of people's lives, and various radio transmitting devices have also been widely used. This has led to the possibility of electromagnetic interference of weather radar at any time, posing a serious threat to the accuracy and authenticity of weather radar monitoring. The problem of electromagnetic interference echo in weather radar data has also become increasingly prominent, seriously affecting the forecaster's analysis and judgment of the weather situation, and is a huge challenge to radar support and data application.

[0003] The electromagnetic interference echoes generated by the weather radar under the influence of electromagnetic interference sources often appear as spiral bands, pockmarks, and banner-shaped straight lines. Usually, close-range electromagnetic interference will cause interference pockmarks on the full screen, while long-range electromagnetic interference often has fixed interference echoes in a certain direction. In addition, interference with a certain bandwidth will also cause radar echoes to appear as interference pockmarks on the full screen, while the electromagnetic interference echoes caused by single-frequency interference are mostly in the shape of banners. At present, there are still few related studies on the elimination of electromagnetic interference echoes. Zhou Honggen et al. (2008) analyzed the causes of radar echo anomalies from the aspects of external electromagnetic wave interference, computer system, receiver system, antenna servo system, etc. in the operation of the new generation weather radar (CINRAD / SA), and summarized the troubleshooting methods and techniques. This method requires manual judgment of radar error data and cannot solve the problem of data use in real-time business operation. Jiang Yuan (2013) studied the interference echo identification method of regional filtering and checking the continuity of echo radial and azimuth for isolated and banner-shaped electromagnetic interference by understanding the characteristics of electromagnetic interference echoes themselves. Guo Chunhui et al. (2014) used edge recognition and radial recognition methods to remove electromagnetic interference echoes. Wen Hao et al. (2016) evaluated the electromagnetic interference quality control algorithm used in the new generation weather radar construction business software system development and application (ROSE) system and found that the existing algorithm only identified from the shape characteristics of the echo, and could only identify and remove electromagnetic interference echoes with less than 5 radials and some pockmarks and spirals. Du Yanxia et al. (2017) used the measured data of Quanzhou and Xiamen weather radars to study the radial interference echo removal method based on the power method, and compared the removal effect with the filtering method and interpolation method. Yang Meilin et al. (2018) improved the interference echo recognition algorithm of Jiang Yuan et al. (2013) based on the characteristic differences of different electromagnetic interference echoes of Beijing weather radar. The evaluation results showed that it had a good processing effect on scattered points, narrow strips and unknown interference echoes at high elevation layers, but had a poor processing effect on full-screen spiral bands and pockmarks and wide strips of electromagnetic interference echoes. Wen Hao et al. (2020) constructed characteristic parameters based on radar base data and used fuzzy logic methods to identify radial interference clutter, which greatly improved the recognition accuracy, but there were problems of misidentification and misdeletion of layered cloud precipitation. Su Tianji et al. (2020) analyzed the morphology and characteristics of electromagnetic interference echoes based on radar scanning and data acquisition strategies, and used radial data interference feature judgment and series autocorrelation methods to filter out most of the electromagnetic interference echoes in the base data. Shu Yi et al. (2022) described the signal characteristics, echo characteristics and generation mechanism of the same-frequency electromagnetic interference of weather radar, providing a reference for the rapid and accurate identification and suppression of radar electromagnetic interference echoes.

[0004] In summary, although the existing radar electromagnetic interference echo recognition method has a good recognition effect, there are still problems such as misidentification and missed recognition in business applications, and the recognition accuracy needs to be further improved. In addition, the current radar electromagnetic interference echo recognition method is mainly aimed at radial strip-shaped and scattered point distribution interference echoes. There are relatively few studies on spiral band interference echoes with more complex distribution forms, and the overall recognition effect is poor. To this end, according to the relevant characteristics of radar electromagnetic interference echoes, in-depth research on electromagnetic interference echo recognition methods based on business weather radar data, improving the quality of weather radar data and its application level in severe weather monitoring and early warning is of great significance to improving disaster prevention and mitigation capabilities and reducing the loss of life and property.

[0005] The increasingly complex electromagnetic environment will interfere with the detection of weather radars, seriously affecting the accuracy of radar weather monitoring. In order to effectively make up for the shortcomings of the existing weather radar data quality control methods in processing electromagnetic interference echoes and solve the increasingly significant electromagnetic interference problem in the actual operation of existing business weather radars, a weather radar electromagnetic interference echo recognition method based on target detection for radar data quality control is provided. Summary of the invention

[0006] The purpose of the present invention is to propose a weather radar electromagnetic interference echo recognition method based on target detection to solve the problems existing in the above-mentioned prior art, improve the data quality of weather radar, and can be applied to the weather radar data processing of existing business operations.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A method for identifying electromagnetic interference echoes of weather radar based on target detection, comprising:

[0009] Acquire radar base data to be identified and generate radar echo images to be identified;

[0010] The radar echo image to be identified is input into a preset target detection model to identify the weather radar electromagnetic interference echo; wherein the target detection model is obtained by training an improved YOLOv8 model with a pre-constructed data set, and the data set includes: weather radar electromagnetic interference echo image and annotation data.

[0011] Furthermore, pre-constructing the data set includes:

[0012] Based on weather radar base data, obtain electromagnetic interference echo images and annotation data of weather radars with electromagnetic interference;

[0013] The weather radar electromagnetic interference echo image and the annotated data are preprocessed to obtain the data set.

[0014] Furthermore, obtaining the weather radar electromagnetic interference echo image and annotation data includes:

[0015] Based on weather radar base data, obtain initial radar echo image;

[0016] Based on the initial radar echo image, screening out the weather radar electromagnetic interference echo image with electromagnetic interference;

[0017] The weather radar electromagnetic interference echo image is segmented, the polygonal boxes of the interference echoes are marked, and the marked interference echoes are classified to generate the annotation data.

[0018] Further, preprocessing the weather radar electromagnetic interference echo image and the annotated data includes:

[0019] Performing horizontal flipping processing, adding Gaussian noise processing, and adjusting brightness processing on the weather radar electromagnetic interference echo image and the annotated data;

[0020] The weather radar electromagnetic interference echo image and annotation data are merged with the processed weather radar electromagnetic interference echo image and annotation data to establish the data set.

[0021] Further improvements to the YOLOv8 model include:

[0022] Introduce a switchable dilated convolution module SAConv into the C2f module of the YOLOv8 model to form the SAConv-C2f module of the backbone network;

[0023] Add a large separable kernel attention mechanism LSKA to the spatial pooling pyramid module SPPF of the YOLOv8 model to form an improved LSKA-SPPF module of the backbone network;

[0024] Add a convolutional attention mechanism module CBAM to the head network of the YOLOv8 model;

[0025] The loss of the YOLOv8 model uses a cross loss function with weighted classification loss and bounding box regression loss.

[0026] Furthermore, the SAConv-C2f module includes: a front context perception layer, a switchable hole convolution layer and a rear context perception layer;

[0027] The front context perception layer and the rear context perception layer both include: a global average pooling layer and a convolutional layer;

[0028] In the pre-context perception layer and the post-context perception layer, the input features are firstly compressed by the global average pooling layer, then extracted by the convolution layer, and finally the convolution layer output is added to the input features;

[0029] The switchable hole convolution layer includes: a feature extraction unit, a switching function unit and a weighted combination unit;

[0030] The feature extraction unit includes: two 3×3 convolutional layers Conv1 and Conv2 with different dilated convolutions;

[0031] The switching function unit includes a global average pooling layer and a convolution layer;

[0032] The weighted combination unit is used to perform weighted addition on the outputs of the two convolutional layers of the feature extraction unit to obtain an intermediate feature;

[0033] The intermediate features are:

[0034]

[0035] Where x is the input, S(x) represents the switching function, and Represents the weight of the convolution operation.

[0036] Furthermore, the LSKA-SPPF module includes: SPPF spatial pooling pyramid and LSKA large separable kernel attention mechanism;

[0037] The SPPF module includes two convolutional layers of size 1×1 and maximum pooling layers of sizes 1×1, 2×2, and 3×3 respectively;

[0038] In the SPPF module, the input feature map first passes through an initial convolution layer to extract a preliminary feature map, and then the preliminary feature map is subjected to three maximum pooling operations respectively, and then the preliminary feature map and the three pooling feature maps are spliced ​​and combined into one feature map in the concat module;

[0039] The LSKA attention module includes: a convolutional layer, cascaded horizontal and vertical one-dimensional convolution kernels;

[0040] In the LSKA large separable kernel attention mechanism, the feature map output by the SPPF module first passes through two initialization convolution layers to extract horizontal and vertical features respectively, and then further extracts image features through two cascaded horizontal and vertical 1-D convolution kernels with different expansion rates to obtain a feature map, and then obtains the output feature map of the LSKA-SPPF module through a concat module and a convolution operation;

[0041] Furthermore, the convolutional attention mechanism module CBAM includes: a channel attention module and a spatial attention module;

[0042] The channel attention module includes: a maximum pooling layer, an average pooling layer, a shared fully connected layer and a ReLU activation function;

[0043] In the channel attention module, the input feature map F1 is subjected to two parallel maximum pooling operations and average pooling operations, and then input into the shared fully connected layer submodule, and then input into the ReLU activation function to obtain the channel attention feature map F2;

[0044] The spatial attention module includes: a maximum pooling layer, an average pooling layer and a convolutional layer;

[0045] In the spatial attention module, the feature map F2 output by the channel attention is multiplied with the input F1 first, and then the maximum pooling operation and the average pooling operation are respectively performed, the outputs of the two pooling layers are concat operated, and finally the feature map F3 of the spatial attention is obtained by the convolution operation and the sigmoid activation function;

[0046] In the convolutional attention mechanism module CBAM, the input feature map F1, the channel attention feature map F2 and the spatial attention feature map F3 are multiplied to obtain the feature map F4.

[0047] Furthermore, the weighted cross loss function adopts the Task Aligned Assigner positive sample assignment strategy, which is composed of a weighted combination of two loss functions: classification loss and bounding box regression loss.

[0048] The classification loss adopts a binary cross entropy loss function to solve the problem of sample imbalance:

[0049]

[0050] in, is the predicted label element, is the target element, if It is a positive sample, is the positive sample weight coefficient, if It is a negative sample;

[0051] The bounding box regression loss adopts the bounding box center point position loss and bounding box area loss A weighted combination of

[0052] The bounding box center point position loss for:

[0053]

[0054] in, is the predicted value of the bounding box coordinate point, is the predicted value of the neighboring point of the bounding box coordinate, y is the target value, is the label integral value, is the label integral value of the coordinate neighboring points;

[0055] The bounding box area loss for:

[0056]

[0057] in, is the intersection of the predicted bounding box and the target bounding box, is the union of the predicted bounding box and the target bounding box, To predict the center point of the bounding box, is the center point of the target bounding box;

[0058] The cross loss function is:

[0059]

[0060] in, , , They are , , The weight of .

[0061] Furthermore, training the improved YOLOv8 model with the pre-built dataset includes:

[0062] Preset precision, recall, and mean average precision (mAP) as model performance evaluation indicators;

[0063] Preset relevant training parameters of the model; the relevant training parameters include: training rounds, input image size, and training end conditions;

[0064] The model is trained and verified based on the data set. According to the recognition results and combined with the model performance evaluation indicators, the model parameters are continuously optimized and adjusted. The model training and testing are repeated to obtain the best recognition effect and the best recognition model.

[0065] The beneficial effects of the present invention are:

[0066] The present invention uses the YOLOv8 network as the baseline model, based on the targeted recognition task of weather radar electromagnetic interference echo, by introducing a switchable hole convolution module SAConv in the C2f module, adding an LSKA attention module to the spatial pooling pyramid module, adding a CBAM attention mechanism module to the head network and improving the loss function, etc., to construct a weather radar electromagnetic interference echo recognition method based on target detection. The present invention can effectively make up for the shortcomings of the existing radar data quality control system in processing electromagnetic interference echoes, and improve the application level of business weather radar data in quantitative estimation of precipitation, severe weather monitoring and early warning, and high-resolution numerical model data assimilation. The technical solution provided by the present invention improves the data quality of weather radars and can be used in the data processing of weather radars in existing business operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0068] Figure 1 This is a flow chart of a method for identifying weather radar electromagnetic interference echoes based on target detection according to an embodiment of the present invention;

[0069] Figure 2 A schematic diagram of the structure of an improved YOLOv8 target detection model according to an embodiment of the present invention;

[0070] Figure 3 A schematic diagram of the structure of the backbone network SAConv-C2f module according to an embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram of the structure of the backbone network LSKA-SPPF module according to an embodiment of the present invention;

[0072] Figure 5 A schematic diagram of the structure of adding a CBAM attention mechanism to the head network of an embodiment of the present invention;

[0073] Figure 6 This is a PPI diagram of the echo intensity and electromagnetic interference echo identification result at an elevation angle of 0.5° of Suizhou S-band weather radar at 23:39 (UTC) on June 23, 2019 according to an embodiment of the present invention;

[0074] Figure 7 This is a PPI diagram of the echo intensity and electromagnetic interference echo identification result at an elevation angle of 0.5° of Suizhou S-band weather radar at 19:03 (UTC) on July 16, 2019 according to an embodiment of the present invention;

[0075] Figure 8 This is a PPI diagram of the echo intensity and electromagnetic interference echo identification results at an elevation angle of 0.5° of Suizhou S-band weather radar at 09:55 (UTC) on June 10, 2019 according to an embodiment of the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0078] This embodiment proposes a weather radar electromagnetic interference echo recognition method based on target detection, including:

[0079] Acquire radar base data to be identified and generate radar echo images to be identified;

[0080] The radar echo image to be identified is input into a preset target detection model to identify the weather radar electromagnetic interference echo; wherein the target detection model is obtained by training an improved YOLOv8 model with a pre-constructed data set, and the data set includes: weather radar electromagnetic interference echo image and annotation data.

[0081] Specifically, this embodiment uses the YOLOv8 network as the baseline model, based on the targeted recognition task of weather radar electromagnetic interference echo, by introducing the switchable hole convolution module SAConv in the C2f module, adding the LSKA attention module to the spatial pooling pyramid module, adding the CBAM attention mechanism module to the head network and improving the loss function, etc., to construct a weather radar electromagnetic interference echo recognition method based on target detection. It can effectively make up for the shortcomings of the existing radar data quality control system in processing electromagnetic interference echoes, and improve the application level of business weather radar data in quantitative estimation of precipitation, severe weather monitoring and early warning, and high-resolution numerical model data assimilation. The technical solution provided in this embodiment improves the data quality of weather radar and can be used in the processing of weather radar data in existing business operations.

[0082] Furthermore, the pre-built datasets include:

[0083] Based on weather radar base data, obtain electromagnetic interference echo images and annotation data of weather radars with electromagnetic interference;

[0084] The weather radar electromagnetic interference echo image and annotation data are preprocessed to obtain the data set.

[0085] Furthermore, in the data set construction of this embodiment, radar echo image drawing and interference echo annotation are first performed: weather radar base data is read, echo intensity (reflectivity factor) plane position display (PPI) image file of the first layer elevation angle is drawn and output; classified electromagnetic interference echoes are annotated on the radar echo image to generate an annotation file; specifically, the following steps are performed:

[0086] S11: Use the radar drawing program written in Python to complete batch processing operations such as reading weather radar base data, polar coordinate conversion, PPI image drawing and image file output to obtain a large number of radar echo images;

[0087] S12: Based on the radar echo images obtained in step S11, manually screening out radar echo images with electromagnetic interference;

[0088] S13: Use LabelMe software to segment the electromagnetic interference echo image screened out in step S12, draw a polygonal box of the interference echo, and classify the marked interference echoes to generate a json format annotation file that corresponds one-to-one to the image file.

[0089] Furthermore, in the data set construction of this embodiment, the data set is enhanced: the weather radar electromagnetic interference echo annotated image and the annotated file produced in step S1 are expanded by using a data enhancement method to establish a weather radar electromagnetic interference echo image data set; specifically, the following steps are performed:

[0090] S21: performing a horizontal flip operation on the weather radar electromagnetic interference echo annotation image produced in step S1 by 180 degrees from left to right, and synchronously performing a horizontal flip operation on the polygon frame coordinate data of the corresponding json annotation file;

[0091] S22: adding Gaussian noise to the weather radar electromagnetic interference echo annotation image produced in step S1, and synchronously adding Gaussian offset to the polygonal frame coordinate data of the corresponding json annotation file;

[0092] S23: increasing the brightness of the weather radar electromagnetic interference echo annotated image produced in step S1 by 20% and decreasing the brightness by 20% respectively;

[0093] S24: Merge the weather radar electromagnetic interference echo annotated image and annotated file produced in step S1 with the weather radar electromagnetic interference echo annotated image and annotated file obtained in steps S21 to S23 to establish a weather radar electromagnetic interference echo image dataset.

[0094] Furthermore, an improved target detection model is constructed: an improved YOLOv8 target detection model with optimized model structure and loss function is constructed; the improved YOLOv8 model includes:

[0095] Introduce a switchable dilated convolution module SAConv into the C2f module of the YOLOv8 model to form the SAConv-C2f module of the backbone network;

[0096] Add the large separable kernel attention mechanism LSKA to the SPPF module of the YOLOv8 model to form the improved LSKA-SPPF module of the backbone network;

[0097] Add attention mechanism module CBAM to the head network of YOLOv8 model;

[0098] The weighted cross loss function is used in the classification loss of the YOLOv8 model.

[0099] Further, the data set is preprocessed: the weather radar electromagnetic interference echo image data set obtained in step S2 is divided into a training set and a test set according to a ratio of 80% and 20%;

[0100] Further, model training and testing: input the training set into the target detection model constructed in step S3 for training, and perform performance testing through the test set; continuously optimize and adjust the model parameters, and iterate the training to obtain the best recognition effect and the best target detection model;

[0101] Furthermore, in this embodiment, model training and testing include:

[0102] S51: Determine precision, recall, and mean average precision (mAP) as model performance evaluation indicators;

[0103] S52: Setting relevant training parameters of the model, such as training rounds, input image size, training end conditions, etc.;

[0104] S53: Based on the training set obtained in step S4, the backbone network is trained by using the transfer parameter learning, and the model is verified using the test set obtained in step S4. According to the recognition results and combined with the model performance evaluation indicators, the model parameters are continuously optimized and adjusted, and the model training and testing are repeated to obtain the best recognition effect and the best recognition model.

[0105] Furthermore, interference echo identification and information output: the optimal target detection model obtained in step S5 is used to identify the electromagnetic interference echo of the weather radar, the identified interference echo is marked, and the coordinate information of the interference echo mask is synchronously output.

[0106] Furthermore, in this embodiment, interference echo identification and information output include:

[0107] S61: Read radar base data to be identified and output radar echo image;

[0108] S62: Input the radar echo image to be identified obtained in step S61 into the target detection model, and output the electromagnetic interference echo annotation image and the electromagnetic interference echo mask contour coordinate information file.

[0109] S63: Read the electromagnetic interference echo mask contour coordinate information file outputted in step S62, convert the image coordinates of the interference echo into radar polar coordinates, and set the radar parameters under the polar coordinates to invalid values, so as to realize the recognition and elimination of the electromagnetic interference echo of the weather radar.

[0110] The following is an example of an operational S-band weather radar deployed in Suizhou, Hubei (31.712°N, 113.581°E). Figure 1 As shown, the implementation process of this embodiment is described in detail:

[0111] Step S1: Obtain the historical base data of the operational S-band weather radar deployed in Suizhou, Hubei (31.712°N, 113.581°E), read the weather radar base data, draw and output the echo intensity PPI image file of the first layer elevation angle; mark the classified electromagnetic interference echo on the radar echo image, and generate a marking file;

[0112] Step S2: using a data enhancement method to expand the weather radar electromagnetic interference echo annotated image and annotation file produced in step S1, and establish a weather radar electromagnetic interference echo image dataset;

[0113] Step S3: Build an object detection model based on improved YOLOv8;

[0114] Step S4: Divide the weather radar electromagnetic interference echo image data set obtained in step S2 into a training set and a test set in a ratio of 80% and 20%;

[0115] Step S5: Input the training set into the target detection model constructed in step S3 for training, and perform performance testing through the test set; continuously optimize and adjust the model parameters, and iterate the training to obtain the best recognition effect and the best target detection model;

[0116] Step S6: using the optimal target detection model obtained in step S5 to identify the electromagnetic interference echo of the weather radar, marking the identified interference echo, and synchronously outputting the coordinate information of the interference echo mask.

[0117] Furthermore, in step S1, the specific steps of making the weather radar electromagnetic interference echo image and annotation file are:

[0118] Step S11: using a radar drawing program written in Python language to complete batch processing operations such as reading weather radar base data, polar coordinate conversion, PPI image drawing and image file output, and obtain a large number of radar echo images;

[0119] Step S12: Based on the radar echo images obtained in step S11, manually screening out radar echo images with electromagnetic interference;

[0120] Step S13: Use LabelMe software to segment the electromagnetic interference echo image screened out in step S12, and mark the polygonal contour of the interference echo. The marking process follows the principles of accuracy, consistency, and distinguishability. Each image contains at least one interference echo annotation, and its type is defined as abnormal. Generate a json format annotation file that corresponds to the image file one by one.

[0121] Furthermore, in step S2, the specific steps of using the data enhancement method to establish a weather radar electromagnetic interference echo image data set are:

[0122] Step S21: performing a horizontal flip operation on the weather radar electromagnetic interference echo annotation image produced in step S1 by 180 degrees from left to right, and synchronously performing a horizontal flip operation on the polygon frame coordinate data of the corresponding json annotation file;

[0123] Step S22: adding Gaussian noise to the weather radar electromagnetic interference echo annotation image produced in step S1, and synchronously adding Gaussian offset to the polygonal frame coordinate data of the corresponding json annotation file;

[0124] Step S23: increasing the brightness of the weather radar electromagnetic interference echo annotated image produced in step S1 by 20% and decreasing the brightness by 20% respectively;

[0125] Step S24: Merge the weather radar electromagnetic interference echo annotated image and annotated file produced in step S1 with the weather radar electromagnetic interference echo annotated image and annotated file obtained in steps S21 to S23 to establish a weather radar electromagnetic interference echo image dataset.

[0126] Furthermore, in step S3, an improved YOLOv8 target detection model is constructed (such as Figure 2 The specific steps are as follows:

[0127] Step S31: Introduce a switchable hole convolution module SAConv for the C2f module of YOLOv8 (such as Figure 3As shown in Figure 2), the SAConv-C2f module of the backbone network is formed. This module consists of a front context-aware layer, a switchable hole convolution layer and a rear context-aware layer. Both context-aware layers include a global feature average pooling layer and a 1×1 convolution layer. The input feature F1 first passes through the global average pooling layer for model compression, then passes through the convolution layer, and finally the convolution layer output is added to the input feature. When the output of the front context-aware layer passes through the switchable hole convolution layer, it is divided into three parts: feature extraction, switching function and weighted combination: in the feature extraction part, the input feature passes through two 3×3 convolution layers Conv1 and Conv2 with different hole rates respectively; the switching function part is an average pooling layer with a 5×5 kernel and a 1×1 convolution layer to determine the weights of the two convolution layers in the feature extraction part; the weighted combination part performs weighted addition of the outputs of the two convolution layers in the feature extraction part to obtain the intermediate feature F, that is:

[0128]

[0129] Where x is the input, S(x) represents the switching function, and Represents the weight of the convolution operation. The intermediate feature F passes through the post-context perception layer and outputs the final feature map F2;

[0130] Step S32: Add the LSKA attention mechanism to the SPPF module to form an improved LSKA-SPPF module of the backbone network (such as Figure 4 As shown). In the LSKA-SPPF module, the input feature map first passes through an initial convolution layer to extract a preliminary feature map, and then the preliminary feature map is respectively subjected to three maximum pooling operations, and then the preliminary feature map and the three pooling feature maps are spliced ​​and combined into a feature map in the concat module, and then two initialization convolution layers are passed to extract horizontal and vertical features, and then two cascaded horizontal and vertical 1-D convolution kernels with different expansion rates are used to further extract image features to obtain a feature map, and finally the output feature map of the LSKA-SPPF module is obtained through the concat module and convolution operation;

[0131] Step S33: Add CBAM attention mechanism to the head network (such as Figure 5As shown). The input feature map F1 of size H×W×C enters the channel attention module, first passes through two parallel maximum pooling layers and average pooling layers, then passes through a shared fully connected layer module, and then obtains two activated results through the ReLU activation function. The two output results are added element by element to obtain the feature map F2 for channel attention. After multiplying F1 and F2, they continue to be input into the spatial attention module, first undergoing parallel maximum pooling and average pooling operations to obtain two 1×H×W feature maps, and then completing the concatenation of the two feature maps through the Concat operation, and converting them into a 1-channel feature map F3 for spatial attention through the sigmoid activation function. Finally, F3 is multiplied again with the previous multiplication result of F1 and F2 to obtain an output feature map F4 of size 1×1×C;

[0132] Step S34: The loss function of YOLOv8 adopts the Task Aligned Assigner positive sample assignment strategy, which is a weighted combination of the classification loss and the bounding box regression loss.

[0133] The classification loss uses a binary cross entropy loss function to solve the problem of sample imbalance:

[0134]

[0135] in, is the predicted label element, is the target element, if It is a positive sample, is the positive sample weight coefficient, if It is a negative sample;

[0136] The bounding box regression loss uses the bounding box center point position loss and bounding box area loss A weighted combination of:

[0137]

[0138] in, is the predicted value of the bounding box coordinate point, is the predicted value of the neighboring point of the bounding box coordinate, y is the target value, is the label integral value, is the label integral value of the coordinate neighboring points;

[0139]

[0140] in, is the intersection of the predicted bounding box and the target bounding box, is the union of the predicted bounding box and the target bounding box, To predict the center point of the bounding box, is the center point of the target bounding box;

[0141] The cross loss function is:

[0142]

[0143] in, , , They are , , The weight of .

[0144] Furthermore, in step S5, the specific steps of model training and testing are:

[0145] Step S51: Determine precision (Precision), recall (Recall), and mean average precision (mAP) as model performance evaluation indicators;

[0146] Step S52: setting relevant training parameters of the model, such as training rounds, input image size, training end conditions, etc.;

[0147] Step S53: Based on the training set obtained in step S4, the backbone network is trained by using the transfer parameter learning, and the model is verified using the test set obtained in step S4. According to the recognition results and combined with the model performance evaluation indicators, the model parameters are continuously optimized and adjusted, and the model training and testing are repeated to obtain the best recognition effect and the best recognition model.

[0148] Furthermore, in step S6, the specific steps of interference echo identification and information output are:

[0149] Step S61: reading radar base data to be identified and outputting radar echo images;

[0150] Step S62: input the radar echo image to be identified obtained in step S61 into the target detection model, and output the electromagnetic interference echo annotation image and the electromagnetic interference echo mask contour coordinate information file.

[0151] Step S63: read the electromagnetic interference echo mask contour coordinate information file outputted in step S62, convert the image coordinates of the interference echo into radar polar coordinates, and set the radar parameters under the polar coordinates to invalid values, so as to realize the recognition and elimination of the electromagnetic interference echo of the weather radar.

[0152] The radar base data selected in this embodiment comes from the business S-band weather radar in Suizhou City, Hubei Province. A large-scale wind farm is built around the radar. During the wind power generation process, the rotation of the wind wheel, the flow of current, and the operation of electrical equipment may cause electromagnetic interference to the radar operation. In addition, the electromagnetic environment around the radar station is complex, and there are some military and civilian radio detection equipment. The above-mentioned multiple factors often lead to obvious electromagnetic interference echoes on the radar echo map, which seriously affects the application of radar data in weather monitoring and forecasting services.

[0153] The following is an analysis of the recognition effect of a weather radar electromagnetic interference echo recognition method based on target detection involved in this application in combination with actual observations of this operational S-band weather radar.

[0154] In order to verify the actual effect of the weather radar electromagnetic interference echo recognition method based on target detection, the actual observation data of Suizhou operational S-band weather radar was used to carry out evaluation.

[0155] Figures 6 to 8 They are Suizhou S-band weather radar at 23:39 on June 23, 2019 ( Figure 6 )、July 16, 2019 19:03( Figure 7 ) and 09:55 on June 10, 2019 ( Figure 8 ) Echo intensity observed at an elevation angle of 0.5° and the PPI diagram of the identified electromagnetic interference echo (data time is universal time, and the radar is located at the center of the image).

[0156] from Figure 6 It can be found that there is an obvious arc-shaped echo in the southeast direction of the PPI map, which is the identified electromagnetic interference echo (the box marked area in the figure, the same below), and its echo intensity (reflectivity factor) is about 15~25dBZ, which is basically consistent with the intensity of the precipitation echo. Therefore, if it is not effectively identified and processed, it will be mistaken for a precipitation echo when the radar data is quantitatively applied, resulting in errors. At the same time, in addition to appearing as an isolated arc-shaped echo, the electromagnetic interference echo sometimes also presents a stacked multiple arc-shaped distribution feature in a certain direction. Figure 7 As shown in the figure, the electromagnetic interference echo recognition method based on target detection also recognizes the multi-layer interference echoes appearing in the southwest direction of the PPI image, and a small interference echo in the west can also be successfully identified. In addition, electromagnetic interference echoes sometimes show complex distribution characteristics with multiple directions and multiple forms coexisting. Figure 8As shown in the figure, the interference echo in the south direction of the PPI image shows discontinuous characteristics, while arc-shaped interference echoes and multi-layer interference echoes also appear in the southeast and southwest directions. Although the morphological characteristics of the two are relatively clear, they are adjacent to the precipitation echo. Faced with such a complex distribution, the electromagnetic interference echo recognition method based on target detection also successfully identified the interference echoes of various types and forms in the figure, and the overall recognition effect was good.

[0157] This example is based on the actual detection data of Suizhou's operational S-band weather radar from June to July 2019. First, the radar's base data is read and the radar echo image is output. Then the radar image is input into the trained weather radar electromagnetic interference echo recognition model based on target detection. Finally, the radar echo image for identifying electromagnetic interference echo and the interference echo coordinate information file are output. The following conclusions are initially obtained:

[0158] (1) The electromagnetic interference echo exhibits significant image characteristics, namely, the overall shape of the electromagnetic interference echo presents a spiral or arc-shaped distribution feature, and the arc is sometimes complete and clear, and sometimes intermittent and blurred. In addition, the electromagnetic interference echo sometimes appears in isolation, sometimes appears in multiple directions at the same time, and sometimes appears continuously in the same direction.

[0159] (2) Traditional radar data quality control methods have difficulty in identifying complex and changeable electromagnetic interference echoes, especially those that are discretely distributed or mixed with precipitation echoes, which are even more difficult to successfully identify. The target detection-based weather radar electromagnetic interference echo recognition method involved in this application takes the image features of the interference echo as the key, improves the target detection model and conducts training modeling, establishes a weather radar electromagnetic interference echo recognition model, and realizes the effective recognition of electromagnetic interference echoes. Judging from the recognition results of actual radar observation data, this method can effectively identify various types of electromagnetic interference echoes, effectively improve the quality of radar data, and provide favorable support for the quantitative application of operational weather radar data.

[0160] From the above conclusions, it can be seen that this embodiment obtains a large number of radar image files by batch processing historical radar base data; uses the data enhancement method to expand the electromagnetic interference echo images manually screened out to establish an electromagnetic interference echo image dataset; conducts training and testing based on the improved YOLOv8 target detection model; and finally obtains the best electromagnetic interference echo recognition model by optimizing and adjusting model parameters and iterative training; thereby realizing weather radar electromagnetic interference echo recognition based on target detection.

[0161] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for identifying electromagnetic interference echoes of weather radar based on target detection, characterized in that: include: Acquire radar base data to be identified and generate radar echo images to be identified; Input the radar echo image to be identified into a preset target detection model to identify the weather radar electromagnetic interference echo; wherein the target detection model is obtained by training an improved YOLOv8 model with a pre-constructed data set, and the data set includes: weather radar electromagnetic interference echo image and annotation data; Improvements to the YOLOv8 model include: Introduce a switchable dilated convolution module SAConv into the C2f module of the YOLOv8 model to form the SAConv-C2f module of the backbone network; Add a large separable kernel attention mechanism LSKA to the spatial pooling pyramid module SPPF of the YOLOv8 model to form an improved LSKA-SPPF module of the backbone network; Add a convolutional attention mechanism module CBAM to the head network of the YOLOv8 model; The loss function of the YOLOv8 model consists of a weighted combination of classification loss and bounding box regression loss.

2. The method for identifying weather radar electromagnetic interference echoes based on target detection according to claim 1, characterized in that: The pre-built dataset includes: Based on weather radar base data, obtain electromagnetic interference echo images and annotation data of weather radars with electromagnetic interference; The weather radar electromagnetic interference echo image and the annotated data are preprocessed to obtain the data set.

3. The method for identifying weather radar electromagnetic interference echoes based on target detection according to claim 2 is characterized in that: Acquiring the weather radar electromagnetic interference echo image and annotation data includes: Based on weather radar base data, obtain initial radar echo image; Based on the initial radar echo image, screening out the weather radar electromagnetic interference echo image with electromagnetic interference; The weather radar electromagnetic interference echo image is segmented, the polygonal boxes of the interference echoes are marked, and the marked interference echoes are classified to generate the annotation data.

4. The method for identifying weather radar electromagnetic interference echoes based on target detection according to claim 2, characterized in that: Preprocessing the weather radar electromagnetic interference echo image and the annotated data includes: Performing horizontal flipping processing, adding Gaussian noise processing, and adjusting brightness processing on the weather radar electromagnetic interference echo image and the annotated data; The weather radar electromagnetic interference echo image and annotation data are merged with the processed weather radar electromagnetic interference echo image and annotation data to establish the data set.

5. The method for identifying weather radar electromagnetic interference echoes based on target detection according to claim 1, characterized in that: The SAConv-C2f module includes: a front context perception layer, a switchable hole convolution layer and a rear context perception layer; The front context perception layer and the rear context perception layer both include: a global average pooling layer and a convolutional layer; In the pre-context perception layer and the post-context perception layer, the input features are firstly passed through the global average pooling layer for model compression, and then passed through the convolution layer, and finally the convolution layer output is added to the input features; The switchable hole convolution layer includes: a feature extraction unit, a switching function unit and a weighted combination unit; The feature extraction unit includes: two convolutional layers Conv1 and Conv2 with different dilated convolutions; The switching function unit includes a global average pooling layer and a convolution layer; The weighted combination unit is used to perform weighted addition on the outputs of the two convolutional layers of the feature extraction unit to obtain an intermediate feature; The intermediate features are: Where x is the input, S(x) represents the switching function, and Represents the weight of the convolution operation.

6. The method for identifying weather radar electromagnetic interference echoes based on target detection according to claim 1, characterized in that: The improved LSKA-SPPF module comprises: an SPPF module and an LSKA module; The SPPF module includes two convolutional layers of size 1×1 and three maximum pooling layers of size 1×1, 2×2, and 3×3 respectively; In the SPPF module, the input feature map first passes through an initial convolution layer to extract a preliminary feature map, and then the preliminary feature map is subjected to three maximum pooling operations respectively, and the preliminary feature map and the three pooling feature maps are spliced ​​and combined into one feature map in the concat module; The LSKA module includes: a convolution layer, cascaded horizontal and vertical one-dimensional convolution kernels; In the LSKA large separable kernel attention mechanism, the feature map output by the SPPF module first passes through two initialization convolution layers to extract horizontal and vertical features respectively, and then further extracts image features through two cascaded horizontal and vertical 1-D convolution kernels with different expansion rates to obtain a feature map, and then obtains the output feature map of the LSKA-SPPF module through the concat module and convolution operation.

7. The method for identifying weather radar electromagnetic interference echoes based on target detection according to claim 1, characterized in that: The convolutional attention mechanism module CBAM includes: a channel attention module and a spatial attention module; The channel attention module includes: a maximum pooling layer, an average pooling layer, a shared fully connected layer and a ReLU activation function; In the channel attention module, the input feature map F1 undergoes two parallel maximum pooling operations and average pooling operations, then enters the shared fully connected layer submodule, and then enters the ReLU activation function to obtain the channel attention feature map F2; The spatial attention module includes: a maximum pooling layer, an average pooling layer and a convolutional layer; In the spatial attention module, the feature map F2 output by the channel attention is first subjected to the maximum pooling operation and the average pooling operation respectively, the outputs of the two pooling layers are concat-operated, and finally the feature map F3 of the spatial attention is obtained by the convolution operation and the sigmoid activation function; In the convolutional attention mechanism module CBAM, the input feature map F1, the channel attention feature map F2 and the spatial attention feature map F3 are multiplied to obtain the feature map F4.

8. The method for identifying weather radar electromagnetic interference echoes based on target detection according to claim 1, characterized in that: The loss function of the YOLOv8 model adopts the Task Aligned Assigner positive sample allocation strategy, which is composed of a weighted combination of two loss functions: classification loss and bounding box regression loss. The classification loss adopts a binary cross entropy loss function to solve the problem of sample imbalance: in, is the predicted label element, is the target element, if is a positive sample, α is the positive sample weight coefficient, if It is a negative sample; The bounding box regression loss adopts the bounding box center point position loss and bounding box area loss A weighted combination of The bounding box center point position loss for: in, is the predicted value of the bounding box coordinate point, is the predicted value of the neighboring point of the bounding box coordinate, y is the target value, is the label integral value, is the label integral value of the coordinate neighboring points; The bounding box area loss for: in, is the intersection of the predicted bounding box and the target bounding box, is the union of the predicted bounding box and the target bounding box, To predict the center point of the bounding box, is the center point of the target bounding box; The cross loss function is: in, , , They are , , The weight of .

9. The method for identifying weather radar electromagnetic interference echoes based on target detection according to claim 1, characterized in that: Training the improved YOLOv8 model with a pre-built dataset includes: The preset precision, recall rate, and average accuracy are used as the model performance evaluation indicators; Preset relevant training parameters of the model; the relevant training parameters include: training rounds, input image size, and training end conditions; The model is trained and verified based on the data set. According to the recognition results and combined with the model performance evaluation indicators, the model parameters are continuously optimized and adjusted. The model training and testing are repeated to obtain the best recognition effect and the best recognition model.

Images