A PPI radar echo intelligent filtering method based on UNet

Abstract

The invention discloses a PPI radar echo intelligent filtering method based on UNet, comprising the following steps: generating labels of millimeter wave cloud radar observation data based on morphological processing methods such as multiple threshold filtering and corrosion expansion to construct a high-quality radar echo classification data set; extracting multiple feature parameters from the original observation data according to the statistical characteristics of ground clutter as a supplement to the original data set; splicing these feature parameters with the observation data and inputting them into a UNet-based deep network for model training to finally obtain a radar echo filtering model; the invention is superior to a traditional UNet network under multiple evaluation indicators, and can efficiently and accurately identify meteorological echoes and ground clutter in radar data.

Description

Technical Field

[0001] The present invention relates to the technical field of PPI radar echo data filtering, and in particular to a PPI radar echo intelligent filtering method based on UNet. Background Art

[0002] Millimeter-wave cloud radar is a ground-based active remote sensing device that can provide meteorological products such as echo intensity, vertical velocity, and velocity spectrum width of meteorological targets for atmospheric detection services through the backscattering effect of meteorological particle groups such as cloud droplets on the millimeter-wave electromagnetic waves emitted by the radar. PPI (Plan Position Indicator) is a radar scanning method that generates plane scanning results through horizontal rotation and can obtain echo data in different directions in real time.

[0003] However, within the radar detection range, non-meteorological targets such as ground objects, insects and birds will also generate echo signals. These clutters interfere with the normal detection of radar echoes, seriously reducing the quality and effectiveness of radar echo data, and have a significant impact on the accuracy and reliability of meteorological research. Therefore, designing an effective algorithm to filter out non-meteorological echoes in radar echoes has become an urgent problem to be solved.

[0004] Traditional PPI radar echo processing technology is mainly based on power spectrum analysis, fuzzy logic, and horizontal and vertical reflectivity factor structures to classify radar echo signals. These methods have limited recognition capabilities in complex environments and slow processing speeds. Under the influence of dynamically changing environmental factors, it is often difficult to classify efficiently and accurately. Therefore, how to improve the accuracy of PPI radar filtering in high-interference environments has become an important research direction in the current field of radar signal processing. Summary of the invention

[0005] In order to solve the above problems in the prior art, the present invention proposes a PPI radar echo intelligent filtering method based on UNet to solve the problems of limited recognition ability and slow processing speed in complex environments.

[0006] The PPI radar echo intelligent filtering method based on UNet described in the present invention comprises the following steps:

[0007] S1 obtains the PPI observation data of millimeter-wave cloud radar, generates labels for radar observation data based on multiple threshold filtering and corrosion dilation morphological processing methods, and constructs a radar echo classification dataset;

[0008] S2 builds a deep learning model based on UNet, using residual connections and attention modules to optimize the encoder and decoder structures and capture key features that affect model performance;

[0009] S3 extracts ground clutter feature parameters from the radar echo classification data set based on the statistical characteristics of ground clutter, splices them with the original observation data, and inputs them into the UNet model for training and optimization to obtain the radar echo filtering target model;

[0010] S4 uses the optimized model to intelligently filter the radar observation data in the study area and conducts evaluation and analysis of multiple scoring indicators.

[0011] Further, step S1 includes the following steps:

[0012] S11 distinguishes target and background echoes based on the statistical characteristics of clutter signals from different objects;

[0013] S12 applies the corrosion and dilation morphological processing method to the data labels in the dataset to remove noise, fill holes and optimize the edge contour of the echo signal.

[0014] Furthermore, step S11 is specifically as follows:

[0015] S111 performs background field calibration based on physical characteristics of the background field;

[0016] S112 integrates radar observation data and ground clutter statistical characteristics to set pixel label values ​​for pixels outside the background field at different radial velocities.

[0017] Furthermore, step S12 is specifically as follows: using an erosion operation to reduce the boundary of the target area and remove local small-range noise points; using an expansion operation to repair the eroded target edge area and restore the target part that may be lost.

[0018] Furthermore, step S2 is specifically as follows: an encoder-decoder architecture with residual connection and attention mechanism is adopted, including: stacking the original radar features and the high-order features after secondary extraction in the channel dimension, and inputting the stacked features into the encoder for data processing; the encoder part realizes the gradual extraction and transformation of features through three layers of residual convolution blocks, spatial attention blocks and pooling layers, and each residual convolution block is composed of residual connections of 2 convolution layers, 2 batch normalization layers and 1 activation layer; the decoder part contains three layers of identical residual convolution blocks and upsampling layers; and a channel attention module is added to the jump connection part between the encoder and the decoder.

[0019] Further, step S3 includes the following steps:

[0020] S31 extracts 6 characteristic parameters from the original radar observation data based on the statistical characteristics of ground clutter to supplement the original data set; among them, the 6 characteristic parameters include: reflectivity factor horizontal texture parameter, reflectivity factor radial inter-library average sign change value, reflectivity factor radial inter-library variation degree, radial velocity regional average value, standard deviation and regional average value of velocity spectrum width six characteristic parameters to highlight the ground clutter characteristics.

[0021] S32 randomly divides the radar echo classification after data enhancement into training set, test set and validation set according to 8:1:1 and standardizes them, and then inputs them into the deep network;

[0022] S33 uses binary cross entropy loss function and Adam optimizer for model training;

[0023] Further, step S4 is as follows: based on the binary confusion matrix, five evaluation indicators including bias score, BIAS, hit rate, Probability of detection, POD, false alarm ratio, FAR, missing alarm ratio, MAR and critical success index, CSI are calculated to perform deterministic evaluation on the filtering effect, and the calculation formula is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] Among them, TP represents the number of targets that are actually ground object echoes and are correctly identified as ground object echoes; FP represents the number of targets that are actually non-ground object echoes but are mistakenly identified as ground object echoes; TN represents the number of targets that are actually non-ground object echoes and are correctly identified as non-ground object echoes; FN represents the number of targets that are actually ground object echoes but are mistakenly identified as non-ground object echoes.

[0030] The PPI radar echo intelligent filtering system based on UNet described in the present invention comprises:

[0031] Echo classification dataset module: used to obtain PPI observation data of millimeter-wave cloud radar, generate labels for radar observation data based on multiple threshold filtering and corrosion expansion morphological processing methods, and construct radar echo classification dataset;

[0032] UNet module: used to build a UNet-based deep learning model, using residual connections and attention modules to optimize the encoder and decoder structures and capture key features that affect model performance;

[0033] Training module: It is used to extract the characteristic parameters of ground clutter from the radar echo classification data set based on the statistical characteristics of ground clutter, and then input them into the UNet model for training optimization after splicing with the original observation data to obtain the radar echo filtering target model;

[0034] Evaluation module: It is used to use the optimized model to intelligently filter the radar observation data in the study area and to evaluate and analyze multiple scoring indicators.

[0035] An electronic device described in 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, any one of the UNet-based PPI radar echo intelligent filtering methods is implemented.

[0036] A storage medium described in the present invention stores a computer program, and when the computer program is executed by a processor, any one of the UNet-based PPI radar echo intelligent filtering methods is implemented.

[0037] Beneficial effects: Compared with the prior art, the present invention constructs a high-quality radar echo classification dataset through morphological processing methods such as multiple threshold filtering and corrosion expansion, and can automatically generate a label dataset based on the observation data, reducing manual intervention; the present invention builds a UNet deep learning model suitable for radar echo classification tasks and adds residual connections and attention mechanisms to optimize the network, and then uses the echo classification dataset to train the network to achieve efficient reasoning of echo classification data, thereby simplifying the radar echo recognition and filtering process; the present invention is superior to the traditional UNet network in multiple evaluation indicators, and can efficiently and accurately identify meteorological echoes and ground clutter in PPI radar observation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the present invention;

[0039] Figure 2 The UNet network structure diagram of the present invention with added residual connection and attention mechanism;

[0040] Figure 3 The difference in evaluation results between the UNet network used in the present invention and the traditional UNet network for a PPI radar echo filtering example. DETAILED DESCRIPTION

[0041] The present invention will be further explained and illustrated below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the embodiments are only used to illustrate and explain the present invention, and do not impose any limitation on the scope of implementation of the present invention.

[0042] The embodiment of the present invention provides a PPI radar echo intelligent filtering method based on UNet, and provides a suitable deep learning model design scheme for the specific application scenario of distinguishing ground clutter and meteorological echo in millimeter-wave cloud radar PPI observation data: construct a high-quality radar echo classification data set based on morphological processing methods such as multiple threshold filtering and corrosion expansion, build a UNet model with residual connection and attention mechanism, and perform model training, and finally obtain a filtering model that can efficiently and accurately distinguish ground clutter and meteorological echo. The implementation flow chart is as follows Figure 1 As shown, the specific implementation steps are as follows:

[0043] Step 1: Obtain the PPI observation data of the millimeter-wave cloud radar, and generate labels for the radar observation data based on morphological processing methods such as multiple threshold filtering and erosion dilation to build a high-quality radar echo classification dataset;

[0044] The research object of the embodiment of the present invention is the PPI scanning observation result of the millimeter wave cloud radar. The raw data includes the radar reflectivity factor Z, velocity V, spectrum width W and signal-to-noise ratio SNR. The dimension of each array is 360×500, 360 is the number of azimuth radials, a total of 360 radials, and the azimuth angle range is 0 to 359°; 500 is the number of range bins, that is, a maximum of 500 bins are observed on each radial, the range resolution is 30 meters, and the maximum detection distance is 15 kilometers.

[0045] After obtaining the PPI observation data of the millimeter-wave cloud radar, it is necessary to generate labels for the observation data, so that they can be input into the supervised learning network in pairs to learn the statistical relationship between the observation data and the labels. The label generation technology is mainly based on the statistical characteristics of ground clutter and meteorological echoes, and is implemented using morphological processing methods such as multiple threshold filtering and corrosion expansion. Specifically, it can be divided into the following two main steps:

[0046] Step S1: Combining the statistical characteristics of different ground clutter signals, multiple thresholds are combined to distinguish target and background echoes.

[0047] Step S2: Apply morphological processing methods such as corrosion and dilation to the data labels in the data set to remove noise, fill holes and optimize the edge contour of the echo signal.

[0048] The multiple threshold filtering method of step S1 may include:

[0049] Step S101: background field calibration;

[0050] The background field in the PPI radar observation data may contain some noise information, which may interfere with the model learning process of the deep network. Therefore, the background field needs to be calibrated first. Based on the physical characteristics of the background field, in the specific implementation, the pixels with the radar reflectivity factor of the default value are marked as the background field, and these pixels are excluded in the subsequent preprocessing.

[0051] Step S102: integrating radar observation data and ground clutter statistical features, and setting pixel label values ​​for pixels outside the background field according to different radial velocities;

[0052] When the average radial velocity absolute value is less than 0.2m / s, the condition for the point to be marked as a weather echo is that the default value in the 5×5 window does not exceed 2, and the two directions of the window simultaneously meet the conditions that the maximum absolute value of the velocity exceeds 0.1m / s and the velocity difference is greater than 0.4m / s; when the average radial velocity is greater than 0.2m / s but less than 1m / s, the condition for marking it as a weather echo is that the radar reflectivity factor exceeds 18dBZ or the signal-to-noise ratio exceeds 18; when the average radial velocity is greater than 1m / s, the condition for marking it as a weather echo is that the radar reflectivity factor is less than -5dBZ and the velocity spectrum width is greater than 0.4m / s. When the above conditions are not met, the point is marked as a ground object echo.

[0053] The label image obtained by the multiple threshold filtering method has multiple small-scale noise points, and the image edge is irregular, which does not conform to the characteristics of ground clutter and weather echo radar images. Therefore, morphological processing of corrosion and expansion is also required, that is, step S2.

[0054] The morphological processing method of step S2 may include:

[0055] Step S201: corrosion operation;

[0056] Considering the spatial distribution characteristics of the ground object echo, the erosion operation is first applied to reduce the irrelevant interference in the image by shrinking the target area, removing small-scale noise and isolated noise. In the specific implementation, the classification label data can be regarded as binary data. The erosion operation uses the defined structural element to slide the label pixel by pixel. When the structural element completely matches the target area, the center pixel will be set to 1, that is, the element is retained, otherwise it is 0. The calculation formula can be expressed as:

[0057] ;

[0058] Where A is the input image and B is the structural element.

[0059] Step S202: expansion operation;

[0060] The expansion operation expands the target area, repairs the missing edges caused by corrosion, fills in the original information of the echo, restores the target area to a relatively complete form, and ensures the continuity and integrity of the echo signal. In the specific implementation, the expansion operation also uses the structural element for sliding. As long as the structural element and the target area have an intersection, the central pixel is set to 1 and the element is retained. The calculation formula is expressed as:

[0061] ;

[0062] Applying corrosion and expansion to the data labels after threshold filtering can optimize the label edge contour and improve the signal quality. To ensure the filtering effect, the present invention performs two corrosion operations and one expansion operation with a 3×3 structural element.

[0063] Step 2: Build a deep learning model based on UNet, and use residual connections and attention modules to optimize the encoder and decoder structures to capture the key features that affect model performance;

[0064] like Figure 2 As shown, the UNet-based deep learning model proposed in the present invention adopts an encoder-decoder architecture with residual connections and attention mechanisms to improve the accuracy of radar echo recognition.

[0065] like Figure 2As shown in the figure, the encoder receives a multi-dimensional array of multiple features after stacking as input, and gradually extracts and transforms the features through three layers of residual convolution blocks, spatial attention blocks and pooling layers. Each residual convolution block consists of 2 convolution layers, 2 batch normalization layers and 1 activation layer with residual connections. Specifically, each convolution block uses a 3×3 convolution kernel, with a stride and padding of 1, and a pooling layer window size of 2×2. The spatial attention block assigns different weights to different spatial positions of the image to enhance the model's attention to key areas in the image, helping the model to better focus on key areas. Through the feature extraction of the convolution block and the feature focusing of the spatial attention block, the local features and detail information in the radar echo image can be effectively captured, helping the model to better classify echoes. The jump connection between the encoder and decoder adds a channel attention module, which uses global average pooling and maximum pooling to obtain the corresponding two channel descriptions, and then sends them to the two-layer shared neural network. The two features output by the network are then added and passed through a Sigmoid activation function to obtain the weight coefficient. At this time, the original input features are converted into scaled features. The channel attention mechanism allows the network to automatically select which channels are most important for the current task. In this way, the model can adaptively enhance the features of important channels while suppressing the features of unimportant channels, thereby enhancing the model's attention and extraction capabilities for key features. The decoder corresponds to the encoder and contains three layers of identical residual convolution blocks and upsampling layers to restore the features compressed by the encoder. The last layer of convolution in the decoder is usually a 1×1 convolution, which is used to reduce the number of channels in the decoded feature map to the number of target categories to generate pixel-by-pixel segmentation results.

[0066] Step 3: Based on the statistical characteristics of ground clutter, the ground clutter characteristic parameters are extracted from the radar echo classification data set, spliced ​​with the original observation data, and input into the UNet model for training optimization to obtain the radar echo filtering target model;

[0067] Since the original observation data cannot fully characterize the comprehensive characteristics of ground clutter, the model may not be able to learn the statistical laws of radar echo classification. In this paper, six ground clutter characteristic parameters are extracted from the original data to achieve data enhancement.

[0068] Specifically, feature extraction is a method of characterizing sample features by mapping information in low-dimensional space. In view of many typical statistical characteristics of ground clutter, such as the spatial discontinuity and irregularity of the reflectivity factor, the propagation characteristics of ground clutter, and the range of the average radial velocity and spectral width, the present invention uses the reflectivity factor horizontal texture parameter in addition to the original data. , radial average sign change value of reflectivity factor between libraries , radial variation degree of reflectivity factor between libraries , the regional average of the radial velocity , Standard Deviation and the regional average of the velocity spectrum width Six characteristic parameters are used to highlight the clutter characteristics of ground objects. Specifically:

[0069] Reflectivity factor horizontal texture parameter Reflectivity factor horizontal texture parameter , radial average sign change value of reflectivity factor between libraries , radial variation degree of reflectivity factor between libraries They are characteristic parameters extracted from the radar echo intensity, which respectively characterize the local change degree of the plane scanning display (PPI) of the reflectivity factor, the proportion of the increase of the reflectivity factor in the radial direction, and the reflectivity factor exceeding the threshold between adjacent distance bins in the region. the proportion of

[0070] Regional average of radial velocity , Standard Deviation It is a characteristic parameter extracted from the radial velocity, representing the degree of change within the radial velocity range smoothed by median filtering;

[0071] Regional average of velocity spectrum width It is a feature extracted from the velocity spectrum width, representing the mean value within the range of the velocity spectrum width.

[0072] These characteristic parameters provide more comprehensive echo information for the subsequent classification of ground object echoes and weather echoes. The comprehensive use of multiple statistical features can effectively improve the model's ability to distinguish between ground clutter and weather echoes.

[0073] The extracted feature parameters are concatenated with the original observation data in the channel dimension and then standardized. Z-score standardization is used for standardization. The original sample individuals in the radar echo classification data set after data enhancement are subtracted from the overall mean of the data and then divided by the overall standard deviation, so that the standardized data meets the distribution with a mean of 0 and a standard deviation of 1. The calculation formula is: ,in, represents the overall mean of the data, In order to obtain an unbiased estimate of the overall standard deviation of the data, the standardized value X' of each sample element X is used as the final input of the model. Then, it is randomly divided into training set, test set and validation set according to 8:1:1. This division method can avoid the interference of accidental factors to a certain extent. The binary cross entropy loss function and Adam optimizer are used to train the model on the training set. The training cycle (Epochs) is set to 100, the batch size (Batchsize) is set to 16, and the learning rate (Learning rate) is set to 1e-3. Finally, an efficient and accurate model is obtained for radar echo intelligent filtering.

[0074] Step 4: Use the optimized model to intelligently filter the radar observation data in the study area and evaluate and analyze multiple scoring indicators.

[0075] Based on the binary confusion matrix, five evaluation indicators, including bias score (BIAS), hit rate (POD), false alarm ratio (FAR), missing alarm ratio (MAR) and critical success index (CSI), are calculated to evaluate the filtering effect deterministically. The calculation formula is as follows:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] Among them, TP represents the number of targets that are actually ground object echoes and are correctly identified as ground object echoes; FP represents the number of targets that are actually non-ground object echoes but are mistakenly identified as ground object echoes; TN represents the number of targets that are actually non-ground object echoes and are correctly identified as non-ground object echoes; FN represents the number of targets that are actually ground object echoes but are mistakenly identified as non-ground object echoes.

[0082] Multiple evaluation indicators are used to evaluate and analyze the basic UNet model and the UNet model proposed in this invention. A set of PPI radar observation data is taken as an example to evaluate the recognition effect. The visualization results and evaluation indicators are shown in Figure 2. Figure 3 And as shown in Table 1.

[0083] Table 1 Comparison of evaluation indicators of different UNet models in PPI radar echo filtering example

[0084] ;

[0085] like Figure 3As shown in the figure, the sample input is the radar reflectivity factor of the original input of the model, the traditional UNet network is the reflectivity factor map after filtering using the traditional UNet network, the improved UNet network is the UNet network filtering map used in the present invention, and the sample labels on the right are the labels in the corresponding radar echo classification data set. The visualization results show that the radar echo intelligent recognition network of the present invention can better distinguish between ground clutter and meteorological echoes. At the same time, from the comparison of specific evaluation indicators, the results of Table 1 show that adding residual connections and attention mechanisms can improve the UNet model's ability to extract radar echo features and significantly improve recognition accuracy.

[0086] The embodiment of the present invention also provides a PPI radar echo intelligent filtering system based on UNet, including:

[0087] Echo classification dataset module: used to obtain PPI observation data of millimeter-wave cloud radar, generate labels for radar observation data based on multiple threshold filtering and corrosion expansion morphological processing methods, and construct radar echo classification dataset;

[0088] UNet module: used to build a UNet-based deep learning model, using residual connections and attention modules to optimize the encoder and decoder structures and capture key features that affect model performance;

[0089] Training module: It is used to extract the characteristic parameters of ground clutter from the radar echo classification data set based on the statistical characteristics of ground clutter, and then input them into the UNet model for training optimization after splicing with the original observation data to obtain the radar echo filtering target model;

[0090] Evaluation module: It is used to use the optimized model to intelligently filter the radar observation data in the study area and to evaluate and analyze multiple scoring indicators.

[0091] An embodiment of the present invention also provides an electronic device, including 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, any one of the UNet-based PPI radar echo intelligent filtering methods is implemented.

[0092] An embodiment of the present invention further provides a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, any one of the UNet-based PPI radar echo intelligent filtering methods is implemented.

Claims

1. A PPI radar echo intelligent filtering method based on UNet, characterized in that: include: S1 obtains the PPI observation data of millimeter-wave cloud radar, generates labels for radar observation data based on multiple threshold filtering and corrosion dilation morphological processing methods, and constructs a radar echo classification dataset; S2 builds a deep learning model based on UNet, using residual connections and attention modules to optimize the encoder and decoder structures and capture key features that affect model performance; Specifically, an encoder-decoder architecture with residual connections and attention mechanism is adopted, including: stacking the original radar features and the high-order features after secondary extraction in the channel dimension, and inputting the stacked features into the encoder for data processing; The encoder part uses three layers of residual convolution blocks, spatial attention blocks and pooling layers to gradually extract and transform features. Each residual convolution block consists of a residual connection of two convolution layers, two batch normalization layers and one activation layer. The decoder part contains three identical residual convolution blocks and upsampling layers. A channel attention module is added to the jump connection between the encoder and decoder. S3 extracts ground clutter feature parameters from the radar echo classification data set based on the statistical characteristics of ground clutter, splices them with the original observation data, and inputs them into the UNet model for training and optimization to obtain the radar echo filtering target model; S4 uses the optimized model to intelligently filter the radar observation data in the study area and conducts evaluation and analysis of multiple scoring indicators.

2. According to the UNet-based PPI radar echo intelligent filtering method of claim 1, it is characterized in that: Step S1 includes the following steps: S11 distinguishes target and background echoes based on the statistical characteristics of clutter signals from different objects; S12 applies the corrosion and dilation morphological processing method to the data labels in the dataset to remove noise, fill holes and optimize the edge contour of the echo signal.

3. According to a UNet-based PPI radar echo intelligent filtering method according to claim 2, it is characterized in that: Step S11 is specifically as follows: S111 performs background field calibration based on physical characteristics of the background field; S112 integrates radar observation data and ground clutter statistical characteristics to set pixel label values ​​for pixels outside the background field at different radial velocities.

4. According to a UNet-based PPI radar echo intelligent filtering method according to claim 2, it is characterized in that: Step S12 is specifically as follows: using an erosion operation to reduce the boundary of the target area and remove local small-range noise points; using an expansion operation to repair the eroded target edge area and restore the lost target part.

5. According to a UNet-based PPI radar echo intelligent filtering method according to claim 1, it is characterized in that: Step S3 includes the following steps: S31 extracts 6 characteristic parameters from the original radar observation data based on the statistical characteristics of ground clutter to supplement the original data set; the 6 characteristic parameters include: reflectivity factor horizontal texture parameter, reflectivity factor radial inter-library average sign change value, reflectivity factor radial inter-library variation degree, radial velocity regional average value, standard deviation and velocity spectrum width regional average value six characteristic parameters to highlight the ground clutter characteristics; S32 randomly divides the radar echo classification after data enhancement into training set, test set and validation set according to 8:1:1 and standardizes them, and then inputs them into the deep network; S33 uses binary cross entropy loss function and Adam optimizer for model training.

6. The PPI radar echo intelligent filtering method based on UNet according to claim 1 is characterized in that: Step S4 is as follows: Based on the binary confusion matrix, five evaluation indicators, namely, bias score BIAS, hit rate POD, false alarm rate FAR, missed alarm rate MAR and critical success index CSI, are calculated to perform deterministic evaluation on the filtering effect. The calculation formula is as follows: ； ； ； ； ； Among them, TP represents the number of targets that are actually ground object echoes and are correctly identified as ground object echoes; FP represents the number of targets that are actually non-ground object echoes but are mistakenly identified as ground object echoes; TN represents the number of targets that are actually non-ground object echoes and are correctly identified as non-ground object echoes; FN represents the number of targets that are actually ground object echoes but are mistakenly identified as non-ground object echoes.

7. A PPI radar echo intelligent filtering system based on UNet, characterized in that: include: Echo classification dataset module: used to obtain PPI observation data of millimeter-wave cloud radar, generate labels for radar observation data based on multiple threshold filtering and corrosion expansion morphological processing methods, and construct radar echo classification dataset; UNet module: used to build a UNet-based deep learning model, using residual connections and attention modules to optimize the encoder and decoder structures and capture key features that affect model performance; Specifically, an encoder-decoder architecture with residual connections and attention mechanisms is adopted, including: stacking the original radar features with the high-order features after secondary extraction in the channel dimension, and inputting the stacked features into the encoder for data processing; the encoder part uses three layers of residual convolution blocks, spatial attention blocks and pooling layers to realize the gradual extraction and transformation of features. Each residual convolution block consists of a residual connection of two convolution layers, two batch normalization layers and one activation layer; the decoder part contains three layers of the same residual convolution blocks and upsampling layers; a channel attention module is added to the jump connection between the encoder and the decoder; Training module: It is used to extract the characteristic parameters of ground clutter from the radar echo classification data set based on the statistical characteristics of ground clutter, and then input them into the UNet model for training optimization after splicing with the original observation data to obtain the radar echo filtering target model; Evaluation module: It is used to use the optimized model to intelligently filter the radar observation data in the study area and to evaluate and analyze multiple scoring indicators.

8. 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, a PPI radar echo intelligent filtering method based on UNet is implemented according to any one of claims 1 to 7.

9. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a PPI radar echo intelligent filtering method based on UNet is implemented according to any one of claims 1 to 7.

Images